publications

2026

1. The Hidden Bias: A Study on Explicit and Implicit Political Stereotypes in Large Language Models

   Konrad Löhr, Shuzhou Yuan, and Michael Färber

   In EACL, 2026

   For insiders: We measure political leanings and stereotypes in major LLMs and show that language variation can reveal stronger implicit bias than persona prompts.

   For everyone: We run chatbots through a “political compass” quiz and then take a close look at the stereotypes they repeat, especially when we ask the same things in another language.

   Abstract PDF

   Large Language Models (LLMs) are increasingly integral to information dissemination and decision-making processes. Given their growing societal influence, understanding potential biases, particularly within the political domain, is crucial to prevent undue influence on public opinion and democratic processes. This work investigates political bias and stereotype propagation across eight prominent LLMs using the two-dimensional Political Compass Test (PCT). Initially, the PCT is employed to assess the inherent political leanings of these models. Subsequently, persona prompting with the PCT is used to explore explicit stereotypes across various social dimensions. In a final step, implicit stereotypes are uncovered by evaluating models with multilingual versions of the PCT. Key findings reveal a consistent left-leaning political alignment across all investigated models. Furthermore, while the nature and extent of stereotypes vary considerably between models, implicit stereotypes elicited through language variation are more pronounced than those identified via explicit persona prompting. Interestingly, for most models, implicit and explicit stereotypes show a notable alignment, suggesting a degree of transparency or "awareness" regarding their inherent biases. This study underscores the complex interplay of political bias and stereotypes in LLMs.

2. PoeTone: A Framework for Constrained Generation of Structured Chinese Songci with LLMs

   Zhan Qu, Shuzhou Yuan, and Michael Färber

   In AAAI, Singapore, 2026

   For insiders: We build a framework for generating classical Chinese Songci that satisfies strict tone and rhyme constraints.

   For everyone: Songci poetry has rules like sheet music, and we teach an AI to write it while staying on the right tones and rhymes, with a strict checker that keeps it honest.

   Abstract PDF

   This paper presents a systematic investigation into the constrained generation capabilities of large language models (LLMs) in producing Songci, a classical Chinese poetry form characterized by strict structural, tonal, and rhyme constraints defined by Cipai templates. We first develop a comprehensive, multi-faceted evaluation framework that includes: (i) a formal conformity score, (ii) automated quality assessment using LLMs, (iii) human evaluation, and (iv) classification-based probing tasks. Using this framework, we evaluate the generative performance of 18 LLMs, including 3 proprietary models and 15 open-source models across four families, under five prompting strategies: zero-shot, one-shot, completion-based, instruction-tuned, and chain-of-thought. Finally, we propose a Generate-Critic architecture in which the evaluation framework functions as an automated critic. Leveraging the critic’s feedback as a reward signal, we fine-tune three lightweight open-source LLMs via supervised fine-tuning (SFT), resulting in improvements of up to 5.88% in formal conformity. Our findings offer new insights into the generative strengths and limitations of LLMs in producing culturally significant and formally constrained literary texts.

2025

1. Incorporating Content-based Features into Quantum Knowledge Graph Embeddings

   Jonas Hendl, and Michael Färber

   In AIQxQIA@ECAI, 2025

   For insiders: We introduce the first quantum link prediction framework that fuses knowledge-graph structure with text features, using amplitude and angle encoding to inject reduced text embeddings into quantum circuits and improve prediction accuracy.

   For everyone: We teach a quantum model to “read the labels” in a knowledge graph—not just follow the links. By turning short texts into simple signals for a quantum circuit, it predicts missing connections more accurately.

   Abstract PDF

   Link prediction in relational data structures, such as knowledge graphs, plays a crucial role in maintaining up-to-date and accurate information. While classical approaches typically leverage either the graph’s connectivity or associated textual descriptions (e.g., labels, definitions), recent quantum models have focused predominantly on structural aspects, leaving the potential of textual information largely unexplored. This paper presents the first quantum link prediction framework that combines both structural and textual modalities. We first generate classical text embeddings, apply dimensionality reduction, and encode them into quantum circuits using two complementary strategies: an Amplitude Encoding Model for high-dimensional fidelity and an Angle Encoding Model optimized for gate efficiency. Experimental results on standard benchmark datasets demonstrate that incorporating textual features in quantum architectures is not only feasible but also enhances the predictive performance of quantum link prediction models.

2. From Monolingual to Bilingual: Investigating Language Conditioning in Large Language Models for Psycholinguistic Tasks

   Shuzhou Yuan, Zhan Qu, Mario Tawfelis, and 1 more author

   In AACL, Mumbai, India, 2025

   For insiders: We show that prompting an LLM with a specific language identity changes its behavior on psycholinguistic tasks and shifts what its internal layers encode across languages.

   For everyone: We show that an AI can “put on” another language like a costume and then answers differently, and even its inner signals shift when it is prompted in a different language.

   Abstract PDF

   Large Language Models (LLMs) exhibit strong linguistic capabilities, but little is known about how they encode psycholinguistic knowledge across languages. We investigate whether and how LLMs exhibit human-like psycholinguistic responses under different linguistic identities using two tasks: sound symbolism and word valence. We evaluate two models, Llama-3.3-70B-Instruct and Qwen2.5-72B-Instruct, under monolingual and bilingual prompting in English, Dutch, and Chinese. Behaviorally, both models adjust their outputs based on prompted language identity, with Qwen showing greater sensitivity and sharper distinctions between Dutch and Chinese. Probing analysis reveals that psycholinguistic signals become more decodable in deeper layers, with Chinese prompts yielding stronger and more stable valence representations than Dutch. Our results demonstrate that language identity conditions both output behavior and internal representations in LLMs, providing new insights into their application as models of cross-linguistic cognition.

3. LLM in the Loop: Creating the ParaDeHate Dataset for Hate Speech Detoxification

   Shuzhou Yuan, Ercong Nie, Lukas Kouba, and 4 more authors

   In AACL, 2025

   For insiders: We create ParaDeHate, an LLM-in-the-loop dataset for hate-speech detoxification, enabling systematic evaluation of meaning-preserving toxic-text rewriting.

   For everyone: We build a large training set that rewrites toxic text into cleaner language without changing the meaning, like running dirty water through a filter and checking the result.

   Abstract

   Detoxification, the task of rewriting harmful language into non-toxic text, has become increasingly important amid the growing prevalence of toxic content online. However, high-quality parallel datasets for detoxification, especially for hate speech, remain scarce due to the cost and sensitivity of human annotation. In this paper, we propose a novel LLM-in-the-loop pipeline leveraging GPT-4o-mini for automated detoxification. We first replicate the ParaDetox pipeline by replacing human annotators with an LLM and show that the LLM performs comparably to human annotation. Building on this, we construct ParaDeHate, a large-scale parallel dataset specifically for hate-speech detoxification. We release ParaDeHate as a benchmark of over 8K hate/non-hate text pairs and evaluate a wide range of baseline methods. Experimental results show that models such as BART, fine-tuned on ParaDeHate, achieve better performance in style accuracy, content preservation, and fluency, demonstrating the effectiveness of LLM-generated detoxification text as a scalable alternative to human annotation.

4. Decomposed Prompting: Probing Multilingual Linguistic Structure Knowledge in Large Language Models

   Ercong Nie, Shuzhou Yuan, Bolei Ma, and 4 more authors

   In AACL, 2025

   For insiders: We introduce token-wise decomposed prompting for multilingual sequence labeling, improving accuracy and efficiency for POS tagging across 38 languages.

   For everyone: Instead of asking an AI to label a whole sentence in one go, we ask about each word step by step, like a teacher calling roll, and this makes the labels more reliable across languages.

   Abstract

   Probing the multilingual knowledge of linguistic structure in large language models (LLMs), often characterized as sequence labeling, faces challenges with maintaining output templates in current text-to-text prompting strategies. To address this, we introduce a decomposed prompting approach for sequence labeling tasks. Instead of relying on a single text-to-text prompt, our method generates an individual prompt for each token in the input sentence, querying its linguistic label. We evaluate our approach on Universal Dependencies part-of-speech tagging across 38 languages, using both English-centric and multilingual LLMs. Our results show that decomposed prompting outperforms iterative prompting baselines in both effectiveness and efficiency under zero- and few-shot settings. Moreover, our analysis of multilingual performance in English-centric LLMs provides insights into the transferability of linguistic knowledge via multilingual prompting.

5. Context-Based URL Classification for Open Access Datasets and Software in Scholarly Documents

   Lamia Salsabil, Rochana R. Obadage, Bipasha Banerjee, and 6 more authors

   In JCDL, 2025

   For insiders: We classify URLs in scholarly documents to identify open datasets and software links, enabling large-scale tracking of research artifacts and openness.

   For everyone: We build a “link bouncer” for research papers that checks whether a URL really leads to open data or usable software, so open resources are easier to find and count.

   Abstract

   This study presents a novel framework for automatically classifying open-access datasets and software (OADS) URLs in scholarly documents. Accurate classification of OADS-URLs is the first step in investigating the availability and preservability of OADS, a crucial step toward open science and computational reproducibility. Our framework, EnSU, leverages an ensemble-based approach to classify OADS-URLs by their citation contexts. The ensemble integrates three models: a Supervised Contrastive Learning model, a SciBERT-based model, and a BertGCN model. Our framework distinguishes the resource types (dataset vs. software) and providers (author vs. third-party). To train and evaluate EnSU, we compiled a dataset, OADS-1K, comprising 1,129 manually annotated sentences containing URLs along with their expanded contexts. Our model outperforms all baseline classifiers, including a large language model-based approach, with the best F1-score of 90%. The dataset and source code are publicly available at: https://github.com/lamps-lab/EnSU/tree/main.

6. Revisiting Projection-based Data Transfer for Cross-Lingual Named Entity Recognition in Low-Resource Languages

   Andrei Politov, Oleh Shkalikov, René Jäkel, and 1 more author

   In NoDaLiDa, Tallinn, Estonia, 2025

   For insiders: We improve projection-based cross-lingual NER using back-translation and a formal matching step, outperforming baselines across 57 low-resource languages.

   For everyone: We improve name detection in languages where training material is scarce by transferring labels from other languages and checking again via translation, like carbon paper plus a second copy to catch errors.

   Abstract PDF

   Cross-lingual Named Entity Recognition (NER) exploits knowledge transfer across languages to identify and classify named entities, making it particularly valuable for low-resource languages. In this paper, we show that data-based cross-lingual transfer is an effective strategy for cross-lingual NER and can outperform multilingual language models in low-resource settings. We introduce two key enhancements to the annotation projection step. First, we refine word alignments by leveraging back-translation to improve projection accuracy. Second, we propose a novel, formalized projection method that matches source-language entities with candidate entities extracted in the target language. Through extensive experiments on two datasets covering 57 languages, we demonstrate that our approach consistently outperforms existing projection-based methods for low-resource languages. These results highlight the robustness and effectiveness of projection-based data transfer as a strong alternative to model-based approaches for cross-lingual named entity recognition in low-resource scenarios.

7. Claim2Source at CheckThat! 2025: Zero-Shot Style Transfer for Scientific Claim-Source Retrieval

   Tobias Schreieder, and Michael Färber

   In CLEF, 2025

   For insiders: We test whether rewriting informal scientific claims in a more formal style improves claim-to-paper retrieval, and analyze when hybrid retrieval helps most.

   For everyone: We show that rewriting informal claims into a cleaner, paper-like wording helps find the right scientific source, like using the correct catalog terms when searching a library.

   Abstract PDF

   In this paper, we present our participation in the CheckThat! 2025 Task 4b on scientific claim–source retrieval. We systematically investigate the impact of style transfer on retrieving the scientific publication referenced by COVID-19-related tweets. To this end, we apply seven distinct style transfer methods to both claims and sources and evaluate their effects on retrieval performance. The style-transferred variants are assessed across 15 retrieval systems, comprising one sparse, seven dense, and seven hybrid models, by testing all combinations of claim and source styles. The style transfer process is guided by a modular zero-shot prompting template with detailed instructions using a large language model. Our results show that GritLM-7B achieves the best performance without style transfer, indicating strong robustness to informal text. In contrast, most retrieval models, particularly sparse and hybrid approaches, benefit from transforming claims into a more formal writing style. Furthermore, we find that hybrid retrieval models generally outperform dense-only models, highlighting the advantages of combining sparse and dense retrieval paradigms for scientific claim–source retrieval.

8. Optical-access networks for smart sustainable cities: from network architecture to fiber deployment

   Md Mosaddek Hossain Adib, Patrick Matalla, Christoph Füllner, and 14 more authors

   J. Opt. Commun. Netw., Mar 2025

   For insiders: We outline optical access-network designs for smart cities and discuss AI-based monitoring and cost-aware fiber deployment for sustainable connectivity.

   For everyone: This work sketches the “roads and water pipes” of smart cities, explaining how fiber networks and AI monitoring can keep future connectivity fast, stable, and energy-aware.

   Abstract DOI PDF

   With the steadily progressing digitization of our society and the migration into urban areas, digitized and highly connected smart cities have attracted much attention from the research community due to their impact on everyday life, potential for new innovations, and ability to reduce carbon footprints. The versatile applications, which are intended to improve life in cities in various aspects, have one thing in common—they rely on widespread, reliable, and high-performing communication networks. Therefore, optical-access networks will be a crucial part of the smart cities’ network infrastructure as they provide cost-effective and high-speed connectivity to antenna sites, residents, enterprises, businesses, and regional data centers in a point-to-multipoint topology. In this article, we address the overall impact of this urban transformation on such networks. We outline our vision of the future smart sustainable city, which will leverage advanced optical-access networks. Subsequently, the physical layer design of optical-access networks is analyzed in the context of point-to-multipoint network topology. This includes a 100-Gbit/s intensity-modulation and direct-detection passive optical network (PON) and a 200-Gbit/s coherent PON utilizing eight-digital subcarrier-based time- and wavelength-division multiplexing and coherent detection. We discuss artificial intelligence-based network monitoring and resource allocation. Next, we provide a techno-economical study for sustainable fiber deployment strategies. Finally, we report the results of a network demonstration for the remote assistance of a connected autonomous vehicle.

9. SQuAI: Scientific Question-Answering with Multi-Agent Retrieval-Augmented Generation

   Ines Besrour, Jingbo He, Tobias Schreieder, and 1 more author

   In CIKM, Seoul, South Korea, Mar 2025

   For insiders: SQuAI answers scientific questions via multi-agent retrieval over millions of papers and produces claim-level citations with supporting evidence sentences.

   For everyone: SQuAI works like a team of librarians by splitting a hard question into smaller ones, pulling evidence from millions of papers, and answering with sources that readers can check.

   Abstract PDF

   We present SQuAI, a scalable and trustworthy multi-agent retrieval-augmented generation (RAG) framework for scientific question answering (QA) with large language models (LLMs). SQuAI addresses key limitations of existing RAG systems in the scholarly domain, where complex, open-domain questions demand accurate answers, explicit claims with citations, and retrieval across millions of scientific documents. Built on over 2.3 million full-text papers from arXiv.org, SQuAI employs four collaborative agents to decompose complex questions into sub-questions, retrieve targeted evidence via hybrid sparse-dense retrieval, and adaptively filter documents to improve contextual relevance. To ensure faithfulness and traceability, SQuAI integrates in-line citations for each generated claim and provides supporting sentences from the source documents. Our system improves faithfulness, answer relevance, and contextual relevance by up to +0.088 (12%) over a strong RAG baseline. We further release a benchmark of 1,000 scientific question-answer-evidence triplets to support reproducibility. With transparent reasoning, verifiable citations, and domain-wide scalability, SQuAI demonstrates how multi-agent RAG enables more trustworthy scientific QA with LLMs.

10. Real-E: A Foundation Benchmark for Advancing Robust and Generalizable Electricity Forecasting

    Chen Shao, Yue Wang, Zhenyi Zhu, and 5 more authors

    In CIKM, Seoul, South Korea, Mar 2025

    For insiders: We release Real-E, a large multi-country electricity-forecasting benchmark, and show current methods struggle under non-stationary correlation shifts.

    For everyone: We build a “weather station” for electricity, release a large real-world testbed, and show why models stumble when the grid’s patterns change over time.

    Abstract PDF

    Energy forecasting is vital for grid reliability and operational efficiency. Although recent advances in time series forecasting have led to progress, existing benchmarks remain limited in spatial and temporal scope and lack multi-energy features. This raises concerns about their reliability and applicability in real-world deployment. To address this, we present the Real-E dataset, covering over 74 power stations across 30+ European countries over a 10-year span with rich metadata. Using Real-E, we conduct an extensive data analysis and benchmark over 20 baselines across various model types. We introduce a new metric to quantify shifts in correlation structures and show that existing methods struggle on our dataset, which exhibits more complex and non-stationary correlation dynamics. Our findings highlight key limitations of current methods and offer a strong empirical basis for building more robust forecasting models.

11. Hateful Person or Hateful Model? Investigating the Role of Personas in Hate Speech Detection by Large Language Models

    Shuzhou Yuan, Ercong Nie, Mario Tawfelis, and 3 more authors

    In PALS@EMNLP, Mar 2025

    For insiders: We show that persona prompts can substantially change LLM hate-speech labels, raising fairness and reliability concerns for LLM-based moderation and annotation.

    For everyone: We show that switching a chatbot’s “persona mask” can change its hate-speech judgments, like getting different verdicts from the same judge in different costumes.

    Abstract PDF

    Hate speech detection is a socially sensitive and inherently subjective task, with judgments often varying based on personal traits. While prior work has examined how socio-demographic factors influence annotation, the impact of personality traits on Large Language Models (LLMs) remains largely unexplored. In this paper, we present the first comprehensive study on the role of persona prompts in hate speech classification, focusing on MBTI-based traits. A human annotation survey confirms that MBTI dimensions significantly affect labeling behavior. Extending this to LLMs, we prompt four open-source models with MBTI personas and evaluate their outputs across three hate speech datasets. Our analysis uncovers substantial persona-driven variation, including inconsistencies with ground truth, inter-persona disagreement, and logit-level biases. These findings highlight the need to carefully define persona prompts in LLM-based annotation workflows, with implications for fairness and alignment with human values.

12. ComplexTempQA: A 100m Dataset for Complex Temporal Question Answering

    Raphael Gruber, Abdelrahman Abdallah, Michael Färber, and 1 more author

    In EMNLP, Suzhou, China, Mar 2025

    For insiders: ComplexTempQA provides 100 million temporal QA pairs that require multi-hop reasoning, enabling realistic evaluation of temporal reasoning in language models.

    For everyone: We create a giant time-travel quiz for AI with 100 million questions about “before”, “after”, and “during”, to see whether models truly keep track of time.

    Abstract PDF

    We introduce ComplexTempQA, a large-scale dataset consisting of over 100 million question-answer pairs designed to tackle the challenges in temporal question answering. ComplexTempQA significantly surpasses existing benchmarks in scale and scope. Utilizing Wikipedia and Wikidata, the dataset covers questions spanning over two decades and offers an unmatched scale. We introduce a new taxonomy that categorizes questions as attributes, comparisons, and counting questions, revolving around events, entities, and time periods, respectively. A standout feature of ComplexTempQA is the high complexity of its questions, which demand reasoning capabilities for answering such as across-time comparison, temporal aggregation, and multi-hop reasoning involving temporal event ordering and entity recognition. Additionally, each question is accompanied by detailed metadata, including specific time scopes, allowing for comprehensive evaluation of temporal reasoning abilities of large language models.

13. Extractive Fact Decomposition for Interpretable Natural Language Inference in One Forward Pass

    Nicholas Popovic, and Michael Färber

    In EMNLP, Suzhou, China, Mar 2025

    For insiders: JEDI performs atomic fact decomposition and interpretable NLI in a single encoder-only forward pass, reducing reliance on expensive generative inference at test time.

    For everyone: We break long statements into small fact “building blocks” and check them efficiently, like snapping sentences into LEGO pieces so the system can justify true or false.

    Abstract PDF

    Recent works in Natural Language Inference (NLI) and related tasks, such as automated fact-checking, employ atomic fact decomposition to enhance interpretability and robustness. For this, existing methods rely on resource-intensive generative large language models (LLMs) to perform decomposition. We propose JEDI, an encoder-only architecture that jointly performs extractive atomic fact decomposition and interpretable inference without requiring generative models during inference. To facilitate training, we produce a large corpus of synthetic rationales covering multiple NLI benchmarks. Experimental results demonstrate that JEDI achieves competitive accuracy in distribution and significantly improves robustness out of distribution and in adversarial settings over models based solely on extractive rationale supervision. Our findings show that interpretability and robust generalization in NLI can be realized using encoder-only architectures and synthetic rationales.

14. Graph-Guided Textual Explanation Generation Framework

    Shuzhou Yuan, Jingyi Sun, Ran Zhang, and 4 more authors

    In EMNLP, Mar 2025

    For insiders: We generate free-text explanations guided by faithful highlight cues encoded with a graph layer, improving explanation faithfulness across reasoning datasets.

    For everyone: We make explanations more like a guided tour by having the model point to evidence first and then connect it step by step, instead of telling a nice story after the fact.

    Abstract PDF

    Natural language explanations (NLEs) are commonly used to provide plausible free-text explanations of a model’s reasoning about its predictions. However, recent work has questioned their faithfulness, as they may not accurately reflect the model’s internal reasoning process regarding its predicted answer. In contrast, highlight explanations–input fragments critical for the model’s predicted answers–exhibit measurable faithfulness. Building on this foundation, we propose G-TEx, a Graph-Guided Textual Explanation Generation framework designed to enhance the faithfulness of NLEs. Specifically, highlight explanations are first extracted as faithful cues reflecting the model’s reasoning logic toward answer prediction. They are subsequently encoded through a graph neural network layer to guide the NLE generation, which aligns the generated explanations with the model’s underlying reasoning toward the predicted answer. Experiments on both encoder-decoder and decoder-only models across three reasoning datasets demonstrate that G-TEx improves NLE faithfulness by up to 12.18% compared to baseline methods. Additionally, G-TEx generates NLEs with greater semantic and lexical similarity to human-written ones. Human evaluations show that G-TEx can decrease redundant content and enhance the overall quality of NLEs. Our work presents a novel method for explicitly guiding NLE generation to enhance faithfulness, serving as a foundation for addressing broader criteria in NLE and generated text.

15. In-Context Learning for Information Extraction using Fully Synthetic Demonstrations

    Nicholas Popovic, Ashish Kangen, Tim Schopf, and 1 more author

    In XLLM@ACL, Vienna, Austria, Mar 2025

    For insiders: We generate synthetic demonstrations and retrieve them at inference to improve document-level entity and relation extraction in low- or zero-shot settings.

    For everyone: We let the AI write its own practice questions and then pull the most useful ones at the right moment, like learning with flashcards you made yourself and picking the cards that really help.

    Abstract PDF

    Large, high-quality annotated corpora remain scarce in document-level entity and relation extraction in zero-shot or few-shot settings. In this paper, we present a fully automatic, LLM-based pipeline for synthetic data generation and in-context learning for document-level entity and relation extraction. In contrast to existing approaches that rely on manually annotated demonstrations or direct zero-shot inference, our method combines synthetic data generation with retrieval-based in-context learning, using a reasoning-optimized language model. This allows us to build a high-quality demonstration database without manual annotation and to dynamically retrieve relevant examples at inference time. Based on our approach we produce a synthetic dataset of over 5k Wikipedia abstracts with approximately 59k entities and 30k relation triples. Finally, we evaluate in-context learning performance on the DocIE shared task, extracting entities and relations from long documents in a zero-shot setting. We find that in-context joint entity and relation extraction at document-level remains a challenging task, even for state-of-the-art large language models.

16. Natural Language Inference Fine-tuning for Scientific Hallucination Detection

    In SDP@ACL, Vienna, Austria, Mar 2025

    For insiders: We fine-tune NLI models to detect scientific hallucinations by checking claims against references, achieving top performance in a shared task.

    For everyone: We build a science fact-checker that compares a model’s statements with the sources it cites, like a receipt check that flags items that do not match.

    Abstract PDF

    Modern generative Large Language Models (LLMs) are capable of generating text that sounds coherent and convincing, but are also prone to producing hallucinations, facts that contradict the world knowledge. Even in the case of Retrieval-Augmented Generation (RAG) systems, where relevant context is first retrieved and passed in the input, the generated facts can contradict or not be verifiable by the provided references. This has motivated SciHal 2025, a shared task that focuses on the detection of hallucinations for scientific content. The two subtasks focused on: (1) predicting whether a claim from a generated LLM answer is entailed, contradicted, or unverifiable by the used references; (2) predicting a fine-grained category of erroneous claims. Our best performing approach used an ensemble of fine-tuned encoder-only ModernBERT and DeBERTa-v3 models for classification. Out of nine competing teams, our approach achieved the first place in sub-task 1 and the second place in sub-task 2.

17. Paths to Causality: Finding Informative Subgraphs within Knowledge Graphs for Knowledge-Based Causal Discovery

    Yuni Susanti, and Michael Färber

    In KDD, Mar 2025

    For insiders: We propose a neurosymbolic method for knowledge-based causal discovery that selects relevant knowledge graph subgraphs to ground LLM prompting.

    For everyone: We help AI reason about cause and effect by highlighting the most informative routes through a knowledge map, like handing a detective the best trail of clues.

    Abstract DOI PDF

    Inferring causal relationships between variable pairs is crucial for understanding multivariate interactions in complex systems. Knowledge-based causal discovery—which involves inferring causal relationships by reasoning over the metadata of variables (e.g., names or textual context)—offers a compelling alternative to traditional methods that rely on observational data. However, existing methods using Large Language Models (LLMs) often produce unstable and inconsistent results, compromising their reliability for causal inference. To address this, we introduce a novel approach that integrates Knowledge Graphs (KGs) with LLMs to enhance knowledge-based causal discovery. Our approach identifies informative metapath-based subgraphs within KGs and further refines their selection using Learning-to-Rank models. The top-ranked subgraphs are then incorporated into zero-shot prompts, improving the effectiveness of LLMs in inferring the causal relationship. Extensive experiments on biomedical and open-domain datasets demonstrate that our method outperforms most baselines by up to 44.8 points in F1 scores, evaluated across diverse LLMs and KGs. Our code and datasets are available on GitHub.

18. Revisiting Projection-based Data Transfer for Cross-Lingual Named Entity Recognition in Low-Resource Languages

    Andrei Politov, Oleh Shkalikov, René Jäkel, and 1 more author

    In NoDaLiDa/Baltic-HLT, Tallinn, Estonia, Mar 2025

    For insiders: We refine projection-based cross-lingual NER transfer and show data-based transfer can outperform multilingual-model baselines in low-resource settings.

    For everyone: We revisit cross-language label transfer for name detection, like tracing a drawing onto another sheet, and show how to make the tracing far more accurate.

    Abstract PDF

    Cross-lingual Named Entity Recognition (NER) leverages knowledge transfer between languages to identify and classify named entities, making it particularly useful for low-resource languages. We show that the data-based cross-lingual transfer method is an effective technique for cross-lingual NER and can outperform multilingual language models for low-resource languages. This paper introduces two key enhancements to the annotation projection step in cross-lingual NER for low-resource languages. First, we explore refining word alignments using back-translation to improve accuracy. Second, we present a novel formalized projection approach of matching source entities with extracted target candidates. Through extensive experiments on two datasets spanning 57 languages, we demonstrated that our approach surpasses existing projection-based methods in low-resource settings. These findings highlight the robustness of projection-based data transfer as an alternative to model-based methods for cross-lingual named entity recognition in low-resource languages.

19. Predicting Company ESG Ratings from News Articles Using Multivariate Time-series Analysis

    Tanja Aue, Adam Jatowt, and Michael Färber

    In TempWeb@WWW, Sydney, Australia, Mar 2025

    For insiders: We predict company ESG ratings from large news streams using multivariate time-series models, enabling more data-driven sustainability signals at scale.

    For everyone: We predict ESG ratings from large streams of news, like building a sustainability dashboard that learns from what companies do and what is reported about them over time.

    Abstract PDF

    In recent years, corporate environmental, social, and governance (ESG) engagement has received significant public attention. As mandatory ESG reporting is increasingly adopted and investors place greater emphasis on sustainability in their decisions, the demand for transparent and reliable ESG ratings is growing. However, existing automatic approaches to ESG rating prediction remain limited. Many rely on traditional machine learning methods like random forests or social network analysis, rather than leveraging incoming news article streams and large multivariate time series data, which, for the first time, enables capturing the dynamic relationships between topics, sentiments, and events. In this paper, we propose a novel approach to predicting ESG ratings from news articles by uniquely combining multivariate time series construction with advanced deep learning techniques. We create an extensive dataset of 3.7 million news articles spanning three years and covering 3,000 U.S. companies, providing a robust foundation for training and evaluating our approach. Our approach achieves high accuracy and outperforms existing approaches, underscoring its potential as a scalable, data-driven solution for ESG rating prediction.

20. SimplifyMyText: An LLM-Based System for Inclusive Plain Language Text Simplification

    Michael Färber, Parisa Aghdam, Kyuri Im, and 2 more authors

    In ECIR, Lucca,Italy, Mar 2025

    For insiders: We build and evaluate an LLM system that turns complex documents into inclusive plain language, improving accessibility for readers with comprehension barriers.

    For everyone: SimplifyMyText turns dense documents into plain language, like clearing bureaucratic fog so more people can understand what matters.

    Abstract PDF

    Text simplification is essential for making complex content accessible to diverse audiences who face comprehension challenges. Yet, the limited availability of simplified materials creates significant barriers to personal and professional growth and hinders social inclusion. Although researchers have explored various methods for automatic text simplification, none fully leverage large language models (LLMs) to offer tailored customization for different target groups and varying levels of simplicity. Moreover, despite its proven benefits for both consumers and organizations, the well-established practice of plain language remains underutilized. In this paper, we introduce https://simplifymytext.org, the first system designed to produce plain language content from multiple input formats, including typed text and file uploads, with flexible customization options for diverse audiences. We employ GPT-4 and Llama-3 and evaluate outputs across multiple metrics. Overall, our work contributes to research on automatic text simplification and highlights the importance of tailored communication in promoting inclusivity.

21. Explainable LiDAR 3D Point Cloud Segmentation and Clustering for Detecting Airplane-Generated Wind Turbulence

    Zhan Qu, Shuzhou Yuan, Michael Färber, and 3 more authors

    In KDD, Toronto, Canada, Mar 2025

    For insiders: We detect aircraft wake turbulence from LiDAR point clouds with explainable ML, providing transparent evidence for aviation safety decisions.

    For everyone: We detect dangerous wake turbulence with LiDAR and explain the model’s decision, like an alarm that not only rings but also points to the evidence in the air.

    Abstract PDF

    Wake vortices—strong, coherent air turbulences created by aircrafts—pose a significant risk to aviation safety and therefore require accurate and reliable detection methods. In this paper, we present an advanced, explainable machine learning method that utilizes Light Detection and Ranging (LiDAR) data for effective wake vortex detection. Our method leverages a dynamic graph CNN (DGCNN) with semantic segmentation to partition a 3D LiDAR point cloud into meaningful segments. Further refinement is achieved through clustering techniques. A novel feature of our research is the use of a perturbation-based explanation technique, which clarifies the model’s decision-making processes for air traffic regulators and controllers, increasing transparency and building trust. Our experimental results, based on measured and simulated LiDAR scans compared against four baseline methods, underscore the effectiveness and reliability of our approach. This combination of semantic segmentation and clustering for real-time wake vortex tracking significantly advances aviation safety measures, ensuring that these are both effective and comprehensible.

22. CoDy: Counterfactual Explainers for Dynamic Graphs

    Zhan Qu, Daniel Gomm, and Michael Färber

    In ICML, Vancouver, Canada, Mar 2025

    For insiders: CoDy generates counterfactual explanations for dynamic graph models, identifying minimal changes that flip predictions and improving interpretability of temporal GNNs.

    For everyone: CoDy explains changing network predictions with “what if” examples, like showing the smallest tweak that would flip a decision so humans can follow the logic.

    Abstract PDF

    Temporal Graph Neural Networks (TGNNs) are widely used to model dynamic systems where relationships and features evolve over time. Although TGNNs demonstrate strong predictive capabilities in these domains, their complex architectures pose significant challenges for explainability. Counterfactual explanation methods provide a promising solution by illustrating how modifications to input graphs can influence model predictions. To address this challenge, we present CoDy—Counterfactual Explainer for Dynamic Graphs—a model-agnostic, instance-level explanation approach that identifies counterfactual subgraphs to interpret TGNN predictions. CoDy employs a search algorithm that combines Monte Carlo Tree Search with heuristic selection policies, efficiently exploring a vast search space of potential explanatory subgraphs by leveraging spatial, temporal, and local event impact information. Extensive experiments against state-of-the-art factual and counterfactual baselines demonstrate CoDy’s effectiveness, with improvements of 16% in AUFSC+ over the strongest baseline. Our code is available at: https://github.com/daniel-gomm/CoDy

23. Hallucinations Can Improve Large Language Models in Drug Discovery

    Shuzhou Yuan, and Michael Färber

    Mar 2025

    For insiders: We show that some hallucinated LLM descriptions of molecules can improve downstream property prediction, suggesting ways to leverage noisy text in drug discovery.

    For everyone: We show that even some AI made-up descriptions can still carry useful signals for drug discovery, like rough sketches that are not perfect but still reveal patterns.

    Abstract PDF

    Hallucinations in large language models (LLMs), plausible but factually inaccurate text, are often viewed as undesirable. However, recent work suggests that such outputs may hold creative potential. In this paper, we investigate whether hallucinations can improve LLMs on molecule property prediction, a key task in early-stage drug discovery. We prompt LLMs to generate natural language descriptions from molecular SMILES strings and incorporate these often hallucinated descriptions into downstream classification tasks. Evaluating seven instruction-tuned LLMs across five datasets, we find that hallucinations significantly improve predictive accuracy for some models. Notably, Falcon3-Mamba-7B outperforms all baselines when hallucinated text is included, while hallucinations generated by GPT-4o consistently yield the greatest gains between models. We further identify and categorize over 18,000 beneficial hallucinations, with structural misdescriptions emerging as the most impactful type, suggesting that hallucinated statements about molecular structure may increase model confidence. Ablation studies show that larger models benefit more from hallucinations, while temperature has a limited effect. Our findings challenge conventional views of hallucination as purely problematic and suggest new directions for leveraging hallucinations as a useful signal in scientific modeling tasks like drug discovery.

24. Optimizing Small Transformer-Based Language Models for Multi-Label Sentiment Analysis in Short Texts

    Julius Neumann, Robert Lange, Yuni Susanti, and 1 more author

    CoRR, Mar 2025

    For insiders: We study how to optimize small transformer models for short-text multi-label sentiment and show when augmentation helps and when it adds harmful noise.

    For everyone: We study how small language models can handle short, messy texts well, like making a compact tool do a big job without losing accuracy.

    Abstract DOI PDF

    Sentiment classification on short-text datasets poses substantial challenges, including class imbalance, limited training data, and the inherent subjectivity of sentiment labels. These issues are further amplified by the lack of contextual information in short texts, which increases ambiguity and data sparsity and complicates effective model learning. In this paper, we evaluate the effectiveness of small Transformer-based models, specifically BERT and RoBERTa with fewer than one billion parameters, for multi-label sentiment classification in short-text scenarios. We investigate three key factors influencing performance: (1) continued domain-specific pre-training, (2) data augmentation via automatically generated examples, with a focus on generative data augmentation, and (3) architectural variations of the classification head. Our experimental results show that data augmentation consistently improves classification performance, whereas continued pre-training on augmented data can introduce noise and degrade accuracy. In addition, we find that modifications to the classification head provide only marginal performance gains. Overall, our findings offer practical guidance for optimizing BERT-based models under resource constraints and for refining modeling strategies for sentiment classification on short-text datasets.

25. Large Physics Models: Towards a Collaborative Approach with Large Language Models and Foundation Models

    Kristian G. Barman, Sascha Caron, Emily Sullivan, and 19 more authors

    Mar 2025

    For insiders: We propose a roadmap for physics-specific foundation models, emphasizing collaborative development, rigorous evaluation, and implications for scientific discovery.

    For everyone: We propose a roadmap for “Large Physics Models”, like planning a new AI lab assistant that can read physics papers, handle equations, and be tested properly before it is trusted.

    Abstract PDF

    This paper explores ideas and provides a potential roadmap for the development and evaluation of physics-specific large-scale AI models, which we call Large Physics Models (LPMs). These models, based on foundation models such as Large Language Models (LLMs) - trained on broad data - are tailored to address the demands of physics research. LPMs can function independently or as part of an integrated framework. This framework can incorporate specialized tools, including symbolic reasoning modules for mathematical manipulations, frameworks to analyse specific experimental and simulated data, and mechanisms for synthesizing theories and scientific literature. We begin by examining whether the physics community should actively develop and refine dedicated models, rather than relying solely on commercial LLMs. We then outline how LPMs can be realized through interdisciplinary collaboration among experts in physics, computer science, and philosophy of science. To integrate these models effectively, we identify three key pillars: Development, Evaluation, and Philosophical Reflection. Development focuses on constructing models capable of processing physics texts, mathematical formulations, and diverse physical data. Evaluation assesses accuracy and reliability by testing and benchmarking. Finally, Philosophical Reflection encompasses the analysis of broader implications of LLMs in physics, including their potential to generate new scientific understanding and what novel collaboration dynamics might arise in research. Inspired by the organizational structure of experimental collaborations in particle physics, we propose a similarly interdisciplinary and collaborative approach to building and refining Large Physics Models. This roadmap provides specific objectives, defines pathways to achieve them, and identifies challenges that must be addressed to realise physics-specific large scale AI models.

26. Implementation of an Open Chemistry Knowledge Base with a Semantic Wiki

    Charlotte Neidiger, Tarek Saier, Kai Kühn, and 13 more authors

    J. Cheminformatics, Mar 2025

    For insiders: We build an open chemistry knowledge base with a semantic wiki and structured forms, enabling collaborative capture and comparison of chemical research results.

    For everyone: We build an open chemistry knowledge base like a shared lab notebook on the web, so results can be entered in a structured way and reused by others.

    Abstract DOI PDF

    In this work, we present a concept for an open chemistry knowledge base aimed at integrating chemical research results into a collaboratively usable platform. To this end, we extend Semantic MediaWiki to support the structured collection and summarization of chemical data reported in scientific publications. We implement tools for capturing chemical structures in machine-readable formats and design dedicated data forms and a data model to ensure standardized input and organization of research results. These extensions enable effective data comparison and contextual analysis within an extensible wiki-based environment. The applicability of the platform is demonstrated through a case study on research related to CO_2 reduction in homogeneous photocatalytic systems, highlighting its potential to substantially enhance the collaborative collection and comparison of research outcomes. Overall, this work demonstrates how subject-specific knowledge in chemistry can be collaboratively collected and managed in an open database by integrating cheminformatics workflows into an established knowledge management technology, thereby helping to close gaps in the information infrastructure of scientists without relying on commercial systems.

2024

1. Assessing Privacy Policies with AI: Ethical, Legal, and Technical Challenges

   Irem Aydin, Hermann Diebel-Fischer, Vincent Freiberger, and 5 more authors

   In AISyS, Mar 2024

   For insiders: We analyze technical feasibility plus ethical and legal risks of using LLMs to assess privacy policies.

   For everyone: We ask whether AI can read privacy policies like a magnifying glass for fine print and we lay out the technical, ethical, and legal pitfalls.

   Abstract PDF

   The growing use of Machine Learning and Artificial Intelligence (AI), particularly Large Language Models (LLMs) like OpenAI’s GPT series, leads to disruptive changes across organizations. At the same time, there is a growing concern about how organizations handle personal data. Thus, privacy policies are essential for transparency in data processing practices, enabling users to assess privacy risks. However, these policies are often long and complex. This might lead to user confusion and consent fatigue, where users accept data practices against their interests, and abusive or unfair practices might go unnoticed. LLMss can be used to assess privacy policies for users automatically. In this interdisciplinary work, we explore the challenges of this approach in three pillars, namely technical feasibility, ethical implications, and legal compatibility of using LLMs to assess privacy policies. Our findings aim to identify potential for future research, and to foster a discussion on the use of LLM technologies for enabling users to fulfil their important role as decision-makers in a constantly developing AI-driven digital economy.

2. ML approaches for OTDR diagnoses in passive optical networks - event detection and classification: ways for ODN branch assignment

   Michael Straub, Johannes Reber, Tarek Saier, and 7 more authors

   J. Opt. Commun. Netw., Mar 2024

   For insiders: We improve ML-based OTDR diagnostics using measured and synthetic traces and map events to network branches for practical fiber monitoring.

   For everyone: We improve fiber fault diagnosis by training on both real and simulated signals, like training a mechanic on real cars and realistic simulators.

   Abstract DOI PDF

   We introduce and demonstrate an ML-supported diagnostic concept for detecting and classifying events on OTDR traces in a PON optical distribution network. By incorporating PON deployment data, events can also be associated with specific ODN branches. We analyze an ensemble classifier and neural networks, and evaluate the use of synthetic OTDR-like traces alongside measured data for training. In our proof-of-concept, an ensemble classifier achieves 98% precision and 95% recall on measured OTDR traces, including successful mapping to ODN branches or groups of branches. On emulated data, we obtain an average precision of 70% and an average recall of 91%.

3. Benefits of international collaboration in computer science: a case study of China, the European Union, and the United States

   Alberto Gómez-Espés, Michael Färber, and Adam Jatowt

   Scientometrics, Mar 2024

   For insiders: We quantify how international collaboration in computer science relates to productivity and citation impact across China, the EU, and the United States.

   For everyone: We map scientific collaboration like flight routes and study how cross-border co-authoring relates to productivity and impact over decades.

   Abstract DOI PDF

   Co-authored publications can yield benefits such as additional expertise, increased funding opportunities, and enhanced research impact. China, the European Union, and the United States have collaborated extensively in Computer Science over the past decades, though the scale and effects of these collaborations have changed over time. In this paper, we analyze 31 years of Computer Science publications from these three regions, examining co-authorship patterns, citation impact, and research topics. Our results show that China’s growing emphasis on Computer Science has made it the most productive region in recent years, and collaborations with the EU and the US have increased its citation rates. Conversely, the EU and the US have benefited from China’s expanding research output, resulting in a higher volume of joint publications.

4. Future Timelines: Extraction and Visualization of Future-related Content From News Articles

   Juwal Regev, Adam Jatowt, and Michael Färber

   In WSDM, Mar 2024

   For insiders: We extract future-oriented statements from news and visualize them on timelines, helping users anticipate developments under daily information overload.

   For everyone: We pull future-looking statements from news and place them on a timeline, like turning daily headlines into a simple roadmap of what might come next.

   Abstract DOI PDF

   In today’s rapidly evolving world, maintaining a comprehensive overview of the future landscape is essential for staying competitive and making informed decisions. However, given the largevolume of daily news, manually obtaining a thorough overview ofan entity’s future prospects is quite challenging. To address this, wepresent a system designed to automatically extract and summarizefuture-related information of a queried entity from news articles. Our approach utilizes a novel and publicly accessible multi-sourcedataset comprising 6,800 annotated sentences to fine-tune a language model to identify future-related sentences. We then use topicmodeling to extract the main topics from the data and rank themby relevance as well as present them on an interactive timeline. User evaluations have shown that the timelines and summariesour system produces are useful. The system is available as a webapplication at: https://chronicle2050.regevson.com.

5. Embedded Named Entity Recognition using Probing Classifiers

   Nicholas Popovic, and Michael Färber

   In EMNLP, Miami, FL, USA, Mar 2024

   For insiders: EMBER enables fast NER in decoder-only language models via probing, adding minimal overhead and avoiding destructive fine-tuning.

   For everyone: We add “live labels” to a chatbot as it writes, like sticky notes appearing while you type instead of only after the text is finished.

   Abstract PDF

   Streaming text generation has become a common way of increasing the responsiveness of language model powered applications, such as chat assistants. At the same time, extracting semantic information from generated text is a useful tool for applications such as automated fact checking or retrieval augmented generation. Currently, this requires either separate models during inference, which increases computational cost, or destructive fine-tuning of the language model. Instead, we propose an approach called EMBER which enables streaming named entity recognition in decoder-only language models without fine-tuning them and while incurring minimal additional computational cost at inference time. Specifically, our experiments show that EMBER maintains high token generation rates, with only a negligible decrease in speed of around 1% compared to a 43.64% slowdown measured for a baseline. We make our code and data available online1, including a toolkit2 for training, testing, and deploying efficient token classification models optimized for streaming text generation.

6. The Effects of Hallucinations in Synthetic Training Data for Relation Extraction

   Steven Rogulsky, Nicholas Popovic, and Michael Färber

   In KBC-LM@ISWC, Baltimore, USA, Mar 2024

   For insiders: We show that hallucinated synthetic training data can seriously reduce relation extraction quality and propose detectors to filter hallucinations.

   For everyone: We show that synthetic training data can be polluted by made-up details, like feeding a model mislabeled food, and we propose ways to detect and clean it.

   Abstract PDF

   Relation extraction is crucial for constructing knowledge graphs, with large high-quality datasets serving as the foundation for training, fine-tuning, and evaluating models. Generative data augmentation (GDA) is a common approach to expand such datasets. However, this approach often introduces hallucinations, such as spurious facts, whose impact on relation extraction remains underexplored. In this paper, we examine the effects of hallucinations on the performance of relation extraction on the document and sentence levels. Our empirical study reveals that hallucinations considerably compromise the ability of models to extract relations from text, with recall reductions between 19.1% and 39.2%. We identify that relevant hallucinations impair the model’s performance, while irrelevant hallucinations have a minimal impact. Additionally, we develop methods for the detection of hallucinations to improve data quality and model performance. Our approaches successfully classify texts as either ’hallucinated’ or ’clean,’ achieving high F1-scores of 83.8% and 92.2%. These methods not only assist in removing hallucinations but also help in estimating their prevalence within datasets, which is crucial for selecting high-quality data. Overall, our work confirms the profound impact of relevant hallucinations on the effectiveness of relation extraction models.

7. Knowledge Graph Structure as Prompt: Improving Small Language Models Capabilities for Knowledge-Based Causal Discovery

   Yuni Susanti, and Michael Färber

   In ISWC, Baltimore, MD, USA, Mar 2024

   For insiders: We inject knowledge graph structure into prompts to improve small language models for causal discovery, showing strong few-shot gains over baselines.

   For everyone: We help smaller language models think about causes by giving them a structured knowledge map, like signposts that reduce guessing in the dark.

   Abstract DOI PDF

   Causal discovery aims to estimate causal structures among variables based on observational data. Large Language Models (LLMs) offer a fresh perspective to tackle the causal discovery problem by reasoning on the metadata associated with variables rather than their actual data values, an approach referred to as knowledge-based causal discovery. In this paper, we investigate the capabilities of Small Language Models (SLMs, defined as LLMs with fewer than 1 billion parameters) with prompt-based learning for knowledge-based causal discovery. Specifically, we present “KG Structure as Prompt”, a novel approach for integrating structural information from a knowledge graph, such as common neighbor nodes and metapaths, into prompt-based learning to enhance the capabilities of SLMs. Experimental results on three types of biomedical and open-domain datasets under few-shot settings demonstrate the effectiveness of our approach, surpassing most baselines and even conventional fine-tuning approaches trained on full datasets. Our findings further highlight the strong capabilities of SLMs: in combination with knowledge graphs and prompt-based learning, SLMs demonstrate the potential to surpass LLMs with larger number of parameters. Our code and datasets are available on GitHub.

8. AutoRDF2GML: Facilitating RDF Integration in Graph Machine Learning

   Michael Färber, David Lamprecht, and Yuni Susanti

   In ISWC, Baltimore, MD, USA, Mar 2024

   For insiders: AutoRDF2GML turns RDF knowledge graphs into graph-ML-ready datasets with content and topology features, bridging semantic web data and graph learning.

   For everyone: AutoRDF2GML converts knowledge graphs into machine learning datasets, like turning a city map into a format a navigation system can train on.

   Abstract DOI PDF

   In this paper, we introduce AutoRDF2GML, a framework designed to convert RDF data into data representations tailored for graph machine learning tasks. AutoRDF2GML enables, for the first time, the creation of both content-based features—i.e., features based on RDF datatype properties—and topology-based features—i.e., features based on RDF object properties. Characterized by automated feature extraction, AutoRDF2GML makes it possible even for users less familiar with RDF and SPARQL to generate data representations ready for graph machine learning tasks, such as link prediction, node classification, and graph classification. Furthermore, we present four new benchmark datasets for graph machine learning, created from large RDF knowledge graphs using our framework. These datasets serve as valuable resources for evaluating graph machine learning approaches, such as graph neural networks. Overall, our framework effectively bridges the gap between the Graph Machine Learning and Semantic Web communities, paving the way for RDF-based machine learning applications.

9. GNNavi: Navigating the Information Flow in Large Language Models by Graph Neural Network

   Shuzhou Yuan, Ercong Nie, Michael Färber, and 2 more authors

   In ACL, Bangkok, Thailand, Mar 2024

   For insiders: GNNavi guides information flow in prompt-based fine-tuning via a graph neural layer, improving few-shot learning while updating only a small fraction of parameters.

   For everyone: GNNavi steers how information flows during prompting, like a traffic controller that routes signals to the right places so the model learns better.

   Abstract DOI PDF

   Large Language Models (LLMs) exhibit strong In-Context Learning (ICL) capabilities when prompts with demonstrations are used. However, fine-tuning still remains crucial to further enhance their adaptability. Prompt-based fine-tuning proves to be an effective fine-tuning method in low-data scenarios, but high demands on computing resources limit its practicality. We address this issue by introducing a prompt-based parameter-efficient finetuning (PEFT) approach. GNNAVI leverages insights into ICL’s information flow dynamics, which indicates that label words act in prompts as anchors for information propagation. GNNAVI employs a Graph Neural Network (GNN) layer to precisely guide the aggregation and distribution of information flow during the processing of prompts by hardwiring the desired information flow into the GNN. Our experiments on text classification tasks with GPT-2 and Llama2 show GNNAVI surpasses standard prompt-based fine-tuning methods in few-shot settings by updating just 0.2% to 0.5% of parameters. We compare GNNAVI with prevalent PEFT approaches, such as prefix tuning, LoRA and Adapter in terms of performance and efficiency. Our analysis reveals that GNNAVI enhances information flow and ensures a clear aggregation process.

10. GraSAME: Injecting Token-Level Structural Information to Pretrained Language Models via Graph-guided Self-Attention Mechanism

    Shuzhou Yuan, and Michael Färber

    In NAACL, Mexico City, Mexico, Mar 2024

    For insiders: GraSAME injects token-level graph structure into pretrained language models via graph-guided self-attention for efficient graph-to-text generation.

    For everyone: GraSAME weaves relationship structure into how a language model reads, like stitching a network into its attention so it can describe connected data more accurately.

    Abstract DOI PDF

    Pretrained Language Models (PLMs) benefit from external knowledge stored in graph structures for various downstream tasks. However, bridging the modality gap between graph structures and text remains a significant challenge. Traditional methods like linearizing graphs for PLMs lose vital graph connectivity, whereas Graph Neural Networks (GNNs) require cumbersome processes for integration into PLMs. In this work, we propose a novel graphguided self-attention mechanism, GraSAME. GraSAME seamlessly incorporates token-level structural information into PLMs without necessitating additional alignment or concatenation efforts. As an end-to-end, lightweight multimodal module, GraSAME follows a multitask learning strategy and effectively bridges the gap between graph and textual modalities, facilitating dynamic interactions between GNNs and PLMs. Our experiments on the graph-to-text generation task demonstrate that GraSAME outperforms baseline models and achieves results comparable to state-of-the-art (SOTA) models on WebNLG datasets. Furthermore, compared to SOTA models, GraSAME eliminates the need for extra pre-training tasks to adjust graph inputs and reduces the number of trainable parameters by over 100 million

11. KITspotlight: A System for Spotlighting Researchers in the Media

    Michael Färber, Benjamin Zagoruiko, and Markus Wambach

    In ICWE, Tampere, Finland, Mar 2024

    For insiders: KITspotlight identifies and aggregates researchers mentioned in news articles, enabling scalable media monitoring and visibility analytics for institutions.

    For everyone: KITspotlight scans news for mentions of researchers, like a radar that turns scattered articles into a clear visibility dashboard.

    Abstract DOI PDF

    Academic institutions, such as universities, heavily rely on public relations and are therefore interested in media monitoring. However, tracking mentions of their researchers in news articles presents challenges, such as identifying affiliated personnel and their departments, and aggregating the extracted information. In this paper, we introduce KITspotlight, a novel system that automatically identifies researchers of the Karlsruhe Institute of Technology (KIT) in newspaper articles, associates these individuals with their departments, and presents this information visually. KITspotlight is tailored for department heads, administrative staff, and individual researchers, focusing either on the institution’s overall public visibility or on an individual researcher’s media presence. Based on an analysis of 2,280 articles over 12 months, our system provides a model for monitoring academic personnel at any research institution.

12. ToPro: Token-Level Prompt Decomposition for Cross-Lingual Sequence Labeling Tasks

    Bolei Ma, Ercong Nie, Shuzhou Yuan, and 4 more authors

    In EACL, Malta, Mar 2024

    For insiders: ToPro decomposes prompts token-by-token for sequence labeling, improving cross-lingual transfer for tasks like NER and POS tagging.

    For everyone: ToPro breaks prompts into tiny word-level questions, like asking about each word in turn, which makes labeling across languages less brittle.

    Abstract PDF

    Prompt-based methods have been successfully applied to multilingual pretrained language models for zero-shot cross-lingual understanding. However, most previous studies primarily focused on sentence-level classification tasks, and only a few considered token-level labeling tasks such as Named Entity Recognition (NER) and Part-of-Speech (POS) tagging. In this paper, we propose Token-Level Prompt Decomposition (TOPRO), which facilitates the prompt-based method for token-level sequence labeling tasks. The TOPRO method decomposes an input sentence into single tokens and applies one prompt template to each token. Our experiments on multilingual NER and POS tagging datasets demonstrate that TOPRO-based fine-tuning outperforms Vanilla fine-tuning and Prompt-Tuning in zero-shot cross-lingual transfer, especially for languages that are typologically different from the source language English. Our method also attains state-of-theart performance when employed with the mT5 model. Besides, our exploratory study in multilingual large language models shows that TOPRO performs much better than the current in-context learning method. Overall, the performance improvements show that TOPRO could potentially serve as a novel and simple benchmarking method for sequence labeling tasks

13. HyperPIE: Hyperparameter Information Extraction from Scientific Publications

    Tarek Saier, Mayumi Ohta, Takuto Asakura, and 1 more author

    In ECIR, Glasgow, UK, Mar 2024

    For insiders: HyperPIE extracts hyperparameters from scientific papers, enabling machine-readable experiment settings for reproducibility, search, and meta-analysis.

    For everyone: HyperPIE reads papers to extract exact experiment settings, like pulling oven temperature and timing from thousands of recipes so results can be compared.

    Abstract DOI PDF

    Automatic extraction of information from publications is key to making scientific knowledge machine readable at a large scale. The extracted information can, for example, facilitate academic search, decision making, and knowledge graph construction. An important type of information not covered by existing approaches is hyperparameters. In this paper, we formalize and tackle hyperparameter information extraction (HyperPIE) as an entity recognition and relation extraction task. We create a labeled data set covering publications from a variety of computer science disciplines. Using this data set, we train and evaluate BERT-based fine-tuned models as well as five large language models: GPT-3.5, GALACTICA, Falcon, Vicuna, and WizardLM. For fine-tuned models, we develop a relation extraction approach that achieves an improvement of 29% F1 over a state-of-the-art baseline. For large language models, we develop an approach leveraging YAML output for structured data extraction, which achieves an average improvement of 5.5% F1 in entity recognition over using JSON. With our best performing model we extract hyperparameter information from a large number of unannotated papers, and analyze patterns across disciplines. All our data and source code is publicly available at \urlhttps://github.com/IllDepence/hyperpie.

14. A Novel Machine Learning-based Equalizer for a Downstream 100G PAM-4 PON

    Chen Shao, Elias G. Giacoumidis, Shi Li, and 4 more authors

    In OFC, San Diego, CA, USA, Mar 2024

    For insiders: We propose an ML-based equalizer for downstream 100G PAM-4 PON and show large BER improvements with lower complexity than baselines.

    For everyone: We use AI to clean up fiber signals, like noise-cancelling headphones for optical networks, so high-speed links work with fewer errors.

    Abstract PDF

    A frequency-calibrated SCINet (FC-SCINet) equalizer is proposed for downstream 100G PON with 28.7 dB path loss. At 5 km, FC-SCINet improves the BER by 88.87% compared to FFE and a 3-layer DNN with 10.57% lower complexity.

15. Advanced Equalization in 112 Gb/s Upstream PON Using a Novel Fourier Convolution-based Network

    Chen Shao, Elias Giacoumidis, Patrick Matalla, and 6 more authors

    In ECOC, Frankfurt, Germany, Mar 2024

    For everyone: We design a smarter signal-cleaning model for fiber links, like tuning a radio with a better filter so high-speed internet arrives with fewer glitches.

    Abstract PDF

    This work presents FConvNet, a Fourier Convolution-based Network for equalization in upstream PON. It outperforms traditional equalizers in bit-error-rate and power efficiency, demonstrating potential for high-speed optical networks.

16. Machine Learning in Short-Reach Optical Systems: A Comprehensive Survey

    Chen Shao, Elias Giacoumidis, Syed Moktacim Billah, and 6 more authors

    Photonics, Mar 2024

    For everyone: This survey is a field guide to AI in optical networks, explaining what works where and why efficiency and reliability matter for critical infrastructure.

    Abstract DOI PDF

    This survey reviews the application of machine learning in short-reach optical systems. It categorizes ML techniques into traditional, convolutional, transformer-based, and Fourier-based models, highlighting their potential in tasks like equalization, fault detection, and bandwidth prediction, with a focus on PONs and hardware efficiency.

17. Incorporating Quantum Advantage in Quantum Circuit Generation through Genetic Programming

    Christoph Stein, and Michael Färber

    Mar 2024

    For insiders: We use quantum advantage in genetic programming for circuit generation, improving convergence and producing circuits competitive with expert designs on standard tasks.

    For everyone: We evolve quantum circuits by trial and reward, like selective breeding for better designs that solve test problems faster.

    Abstract PDF

    Designing efficient quantum circuits that leverage quantum advantage compared to classical computing has become increasingly critical. Genetic algorithms have shown potential in generating such circuits through artificial evolution. However, integrating quantum advantage into the fitness function of these algorithms remains unexplored. In this paper, we aim to enhance the efficiency of quantum circuit design by proposing two novel approaches for incorporating quantum advantage metrics into the fitness function of genetic algorithms.1 We evaluate our approaches based on the Bernstein-Vazirani Problem and the Unstructured Database Search Problem as test cases. The results demonstrate that our approaches not only improve the convergence speed of the genetic algorithm but also produce circuits comparable to expert-designed solutions. Our findings suggest that automated quantum circuit design using genetic algorithms that incorporate a measure of quantum advantage is a promising approach to accelerating the development of quantum algorithms.

18. Why Lift So Heavy? Slimming Large Language Models by Cutting Off the Layers

    Shuzhou Yuan, Ercong Nie, Bolei Ma, and 1 more author

    In IJCNN, Rome, Italy, Mar 2024

    For insiders: We show that truncating layers can shrink large language models while retaining strong performance on some tasks.

    For everyone: We show that some large language models can go on a diet by removing layers without collapsing on certain tasks, pointing to more efficient and greener AI.

    Abstract PDF

    Large Language Models (LLMs) exhibit exceptional capabilities across a wide range of natural language processing tasks. However, their immense size poses challenges for storage, training, and inference due to the billions of parameters introduced through deep layer stacking. While methods such as pruning and distillation can reduce model size, they frequently compromise performance. In this work, we systematically investigate the impact of reducing the number of layers in LLMs. Surprisingly, we find that even significantly shallower models maintain comparable or sometimes superior performance, particularly in prompt-based fine-tuning for text classification. In several cases, single-layer models outperform their fully layered counterparts. These results provide important insights for designing more efficient LLMs that mitigate size constraints while preserving performance, opening new avenues for resource-efficient large-scale language modeling.

2023

1. Linked Papers With Code: The Latest in Machine Learning as an RDF Knowledge Graph

   Michael Färber, and David Lamprecht

   In ISWC, Athens, Greece, Mar 2023

   For insiders: LPWC turns the ML landscape into a machine-queryable knowledge graph linking papers to tasks, datasets, methods, and results for semantic search and analysis.

   For everyone: LPWC turns machine-learning research into a structured “periodic table” of papers, tasks, datasets, and results that can be searched and connected.

   Abstract PDF

   In this paper, we introduce Linked Papers With Code (LPWC), an RDF knowledge graph that provides comprehensive, current information about almost 400,000 machine learning publications. This includes the tasks addressed, the datasets utilized, the methods implemented, and the evaluations conducted, along with their results. Compared to its non-RDF-based counterpart Papers With Code, LPWC not only translates the latest advancements in machine learning into RDF format, but also enables novel ways for scientific impact quantification and scholarly key content recommendation. LPWC is openly accessible at this URL and is licensed under CC-BY-SA 4.0. As a knowledge graph in the Linked Open Data cloud, we offer LPWC in multiple formats, from RDF dump files to a SPARQL endpoint for direct web queries, as well as a data source with resolvable URIs and links to the data sources SemOpenAlex, Wikidata, and DBLP. Additionally, we supply knowledge graph embeddings, enabling LPWC to be readily applied in machine learning applications.

2. SemOpenAlex: The Scientific Landscape in 26 Billion RDF Triples

   Michael Färber, David Lamprecht, Johan Krause, and 2 more authors

   In ISWC, Mar 2023

   For insiders: We release SemOpenAlex, a scholarly knowledge graph with 26 billion triples, dumps, SPARQL access, and embeddings, enabling large-scale semantic science analytics and search.

   For everyone: SemOpenAlex is an open “Google Maps for science”, built as a connected map of papers and authors so others can navigate research at web scale.

   Abstract PDF

   We present SemOpenAlex, an extensive RDF knowledge graph that contains over 26 billion triples about scientific publications and their associated entities, such as authors, institutions, journals, and concepts. SemOpenAlex is licensed under CC0, providing free and open access to the data. We offer the data through multiple channels, including RDF dump files, a SPARQL endpoint, and as a data source in the Linked Open Data cloud, complete with resolvable URIs and links to other data sources. Moreover, we provide embeddings for knowledge graph entities using high-performance computing. SemOpenAlex enables a broad range of use-case scenarios, such as exploratory semantic search via our website, large-scale scientific impact quantification, and other forms of scholarly big data analytics within and across scientific disciplines. Additionally, it enables academic recommender systems, such as recommending collaborators, publications, and venues, including explainability capabilities. Finally, SemOpenAlex can serve for RDF query optimization benchmarks, creating scholarly knowledge-guided language models, and as a hub for semantic scientific publishing.

3. Biases in Scholarly Recommender Systems: Impact, Prevalence, and Mitigation

   Michael Färber, Melissa Coutinho, and Shuzhou Yuan

   Scientometrics, Mar 2023

   For insiders: We measure biases in scholarly recommender systems and discuss mitigation strategies to support fairer literature discovery under severe paper overload.

   For everyone: We show that literature recommenders can have blind spots, like a librarian who keeps pointing to the same shelves, and we outline how to measure and reduce this effect.

   Abstract DOI PDF

   With the remarkable increase in the number of scientific entities such as publications, researchers, and scientific topics, and the associated information overload in science, academic recommender systems have become increasingly important for millions of researchers and science enthusiasts. However, it is often overlooked that these systems are subject to various biases. In this article, we first break down the biases of academic recommender systems and characterize them according to their impact and prevalence. In doing so, we distinguish between biases originally caused by humans and biases induced by the recommender system. Second, we provide an overview of methods that have been used to mitigate these biases in the scholarly domain. Based on this, third, we present a framework that can be used by researchers and developers to mitigate biases in scholarly recommender systems and to evaluate recommender systems fairly. Finally, we discuss open challenges and possible research directions related to scholarly biases.

4. AI-Based OTDR Event Detection, Classification and Assignment to ODN Branches in Passive Optical Networks

   Michael Straub, T. Saier, J. Reber, and 6 more authors

   Journal of Optical Communications and Networking, Mar 2023

   For insiders: We detect and classify events in optical networks from OTDR traces and assign faults to network branches, improving reliability of passive optical networks.

   For everyone: We listen to echoes in fiber cables like sonar and use machine learning to spot and locate faults, improving how networks can be monitored.

   Abstract DOI PDF

   An AI-supported monitoring concept is demonstrated allowing detection and classification of events on OTDR traces with high precision and recall for application on a PON optical distribution network. We can also associate events with ODN branches by using deployment data of the PON topology.

5. Analyzing the Impact of Companies on AI Research Based on Publications

   Michael Färber, and Lazaros Tampakis

   Scientometrics, Mar 2023

   For insiders: We measure how companies influence AI research via publishing and show differences in citation impact and online attention between industry and academia.

   For everyone: We measure how strongly companies shape AI research through publishing, like tracking who is rowing and who is steering in a shared boat.

   Abstract DOI PDF

   Artificial Intelligence (AI) is one of the most momentous technologies of our time. Thus, it is of major importance to know which stakeholders influence AI research. Besides researchers at universities and colleges, researchers in companies have hardly been considered in this context. In this article, we consider how the influence of companies on AI research can be made measurable on the basis of scientific publishing activities. We compare academic- and company-authored AI publications published in the last decade and use scientometric data from multiple scholarly databases to look for differences across these groups and to disclose the top contributing organizations. While the vast majority of publications is still produced by academia, we find that the citation count an individual publication receives is significantly higher when it is (co-)authored by a company. Furthermore, using a variety of altmetric indicators, we notice that publications with company participation receive considerably more attention online. Finally, we place our analysis results in a broader context and present targeted recommendations to safeguard a harmonious balance between academia and industry in the realm of AI research.

6. A Full-Fledged Framework for Combining Entity Linking Systems and Components

   Kristian Noullet, Ayoub Ourgani, and Michael Färber

   In K-CAP, Mar 2023

   For insiders: We present a full framework for modular entity-linking pipelines, enabling easier combination of components and fairer comparison across systems.

   For everyone: We offer a framework to combine entity-linking systems, like a mixing console where you can plug in different components and compare results fairly.

   Abstract DOI PDF

   Named entity recognition and disambiguation, often referred to as entity linking systems, refers to the task of automatically identifying knowledge graph entities in text documents. While a variety of entity linking systems based on very different approaches exist, these systems implicitly share certain processing steps in their pipeline. Despite this fact, they have been mainly used as stand-alone solutions. In this paper, we propose a framework for combining entity linking methods. This allows multiple entity linking systems and especially their components to be used in combination to an unlimited extent, thus allowing to achieve the best possible performance. In addition, the framework allows user-developed entity linking systems or components to be easily tested and automatically evaluated against other systems without having to set up other systems first. Essentially, our framework is knowledge graph agnostic and entity linking systems can be compared across knowledge graphs. Furthermore, our framework enables entity linking method or component recommendation, supporting the goal of achieving the best performance in a given context. We demonstrate that non-domain-expert users are able to deploy the framework within minutes and integrate unknown homebrew systems into it in less than an hour. Our framework is fully open source and available on GitHub along with Docker containers and tutorials (including Jupyter Notebooks).

7. A Full-fledged Commit Message Quality Checker Based on Machine Learning

   David Faragó, Michael Färber, and Christian Petrov

   In COMPSAC, Torino, Italy, Mar 2023

   For insiders: We build an ML-based tool that checks commit-message quality using established rules, supporting maintainability and better software engineering practice.

   For everyone: We build a quality checker for commit messages, like a grammar checker for software teams so code changes stay understandable.

   Abstract DOI PDF

   Commit messages (CMs) are an essential part of version control. By providing important context regarding what has changed and why, they strongly support software maintenance and evolution. However, writing good CMs is difficult and often neglected by developers. So far, there is no tool suitable for practice that automatically assesses how well a CM is written, including its meaning and context. Since this task is challenging, we ask the research question: how well can commit message quality, including semantics and context, be measured with machine learning methods? By considering all rules from the most popular CM quality guideline, creating datasets for those rules, and training and evaluating state-of-the-art machine learning models to check these rules, we can answer the research question with: sufficiently well for practice, with the lowest F1 score of 82.9% for the most challenging task. We develop a full-fledged open-source framework that checks all these CM quality rules. It is useful for research (e.g., automatic CM generation), but most importantly for software practitioners to raise the quality of CMs and thus the maintainability and evolution speed of their software.

8. Ablesbarkeitsmesser: A System for Assessing the Readability of German Text

   Florian Pickelmann, Michael Färber, and Adam Jatowt

   In ECIR, Dublin, Ireland, Mar 2023

   For insiders: We provide an online service to estimate readability of German texts, supporting clearer writing for education and public communication.

   For everyone: We build a “readability thermometer” for German, so people can quickly see whether a text is easy or hard to read and adjust it for the audience.

   Abstract DOI PDF

   While several approaches have been proposed for estimating the readability of English texts, there is much less work for other languages. In this paper, we present an online service, available at https://readability-check.org/, that provides five well-established statistical methods and two machine learning models for measuring the readability of texts in German. For the machine learning methods, we train two BERT models. To bring all the measures together, we provide an interactive website that allows users to evaluate the readability of German texts at the sentence level. Our research can be useful for anyone who wants to know whether the text content at hand is easy or difficult and therefore can be used in certain situations or rather needs to be adapted and improved. In education, for example, it can help to assess the suitability of a particular teaching material for a particular grade.

9. CoCon: A Data Set on Combined Contextualized Research Artifact Use

   Tarek Saier, Youxiang Dong, and Michael Färber

   In JCDL, Mar 2023

   For insiders: We release CoCon, a dataset capturing combined use of research artifacts in papers, enabling new prediction and recommendation tasks beyond paper-level analysis.

   For everyone: CoCon shows which tools, datasets, and methods tend to appear together in papers, like revealing which ingredients often show up in the same scientific recipes.

   Abstract PDF

   In the wake of information overload in academia, methodologies and systems for search, recommendation, and prediction to aid researchers in identifying relevant research are actively studied and developed. Existing work, however, is limited in terms of granularity, focusing only on the level of papers or a single type of artifact, such as data sets. To enable more holistic analyses and systems dealing with academic publications and their content, we propose CoCon, a large scholarly data set reflecting the combined use of research artifacts, contextualized in academic publications’ full-text. Our data set comprises 35 k artifacts (data sets, methods, models, and tasks) and 340 k publications. We additionally formalize a link prediction task for "combined research artifact use prediction" and provide code to utilize analyses of and the development of ML applications on our data. All data and code is publicly available at https://github.com/IllDepence/contextgraph.

10. Evaluating Generative Models for Graph-to-Text Generation

    Shuzhou Yuan, and Michael Färber

    In RANLP, Varna, Bulgaria, Mar 2023

    For insiders: We benchmark GPT-style models for graph-to-text generation in zero-shot settings and analyze when they hallucinate versus fine-tuned baselines.

    For everyone: We test whether language models can describe a network diagram faithfully, like asking a narrator to tell a story from a map and checking where it invents details.

    Abstract PDF

    Large language models (LLMs) have been widely employed for graph-to-text generation tasks. However, the process of finetuning LLMs requires significant training resources and annotation work. In this paper, we explore the capability of generative models to generate descriptive text from graph data in a zero-shot setting. Specifically, we evaluate GPT-3 and ChatGPT on two graph-to-text datasets and compare their performance with that of finetuned LLM models such as T5 and BART. Our results demonstrate that generative models are capable of generating fluent and coherent text, achieving BLEU scores of 10.57 and 11.08 for the AGENDA and WebNLG datasets, respectively. However, our error analysis reveals that generative models still struggle with understanding the semantic relations between entities, and they also tend to generate text with hallucinations or irrelevant information. As part of the error analysis, we utilize BERT to detect machine-generated text and achieve high macro-F1 scores. We have made the text generated by generative models publicly available.

11. Automatic Hint Generation

    Adam Jatowt, Calvin Gehrer, and Michael Färber

    In ICTIR, Taipei, Taiwan, Mar 2023

    For insiders: We generate hints from Wikipedia for question-answer pairs so learners can reach answers themselves, supporting reasoning without giving solutions away.

    For everyone: We generate helpful hints instead of full answers, like giving a nudge in the right direction so people learn and reason rather than being handed the solution.

    Abstract DOI PDF

    At times when answers to user questions are readily and easily available (at essentially zero cost), it is important for humans to maintain their knowledge and strong reasoning capabilities. We believe that in many cases providing hints rather than final answers should be sufficient and beneficial for users as it requires thinking and stimulates learning as well as remembering processes. We propose in this paper a novel task of automatic hint generation that supports users in finding the correct answers to their questions without the need of looking the answers up. As the first attempt towards this new task, we design and implement an approach that uses Wikipedia to automatically provide hints for any input question-answer pair. We then evaluate our approach with a user group of 10 persons and demonstrate that the generated hints help users successfully answer more questions than when provided with baseline hints.

12. Impact, Attention, Influence: Early Assessment of Autonomous Driving Datasets

    Daniel Bogdoll, Jonas Hendl, Felix Schreyer, and 3 more authors

    In ICCRE, Mar 2023

    For insiders: We analyze early impact signals of autonomous-driving datasets and propose an influence score to estimate dataset value before citations mature.

    For everyone: We estimate early which self-driving datasets will matter later, like spotting which seedlings are likely to become the strongest plants.

    Abstract PDF

    Autonomous Driving (AD), the area of robotics with the greatest potential impact on society, has gained a lot of momentum in the last decade. As a result, the number of datasets in AD has increased rapidly. Creators and users of datasets can benefit from a better understanding of developments in the field. While scientometric analysis has been conducted in other fields, it rarely revolves around datasets. Thus, the impact, attention, and influence of datasets on autonomous driving remains a rarely investigated field. In this work, we provide a scientometric analysis for over 200 datasets in AD. We perform a rigorous evaluation of relations between available metadata and citation counts based on linear regression. Subsequently, we propose an Influence Score to assess a dataset already early on without the need for a track-record of citations, which is only available with a certain delay.

13. Measuring Variety, Balance, and Disparity: An Analysis of Media Coverage of the 2021 German Federal Election

    Michael Färber, Jannik Schwade, and Adam Jatowt

    Mar 2023

    For insiders: We measure news diversity in German election coverage using variety, balance, and disparity metrics to support transparent analysis of media pluralism.

    For everyone: We measure how diverse news coverage is, like checking whether the media menu offers many perspectives or keeps serving the same dish.

    Abstract PDF

    This paper introduces a framework to assess news diversity based on variety, balance, and disparity. Using over 26,000 headlines from the 2021 German election, it applies topic modeling and proposes a novel metric for disparity, contributing to automated media diversity analysis.

14. Theories About World Representations for the Internet of Things

    Michael Färber, Yulia Svetashova, and Andreas Harth

    In CARLA, Mar 2023

    For insiders: We connect theories of meaning and knowledge representation to IoT practice, highlighting challenges of grounding, intersubjectivity, and modeling change over time.

    For everyone: We compare how devices and AI systems build pictures of the world, like comparing different mental models to see which ones support shared understanding.

    Abstract PDF

    This paper discusses intersubjectivity, symbol grounding, and dynamic ontologies as foundational components for modeling knowledge in IoT settings. It connects philosophical theories with practical implications in computer science.

15. Vocab-Expander: A System for Creating Domain-Specific Vocabularies Based on Word Embeddings

    Michael Färber, and Nicholas Popovic

    In RANLP, Mar 2023

    For insiders: Vocab-Expander builds domain vocabularies from embeddings and knowledge bases, helping users discover terms for search, innovation, and interdisciplinary work.

    For everyone: Vocab-Expander grows a domain vocabulary like expanding a map legend, suggesting useful terms and links so people can search and communicate more easily.

    Abstract DOI PDF

    VOCAB-EXPANDER is an online tool for building domain-specific vocabularies using word embeddings and ConceptNet. It supports tasks in innovation management, education, and interdisciplinary communication through term expansion and graph-based visualization.

2022

1. The Microsoft Academic Knowledge Graph enhanced: Author name disambiguation, publication classification, and embeddings

   Michael Färber, and Lin Ao

   Quant. Science Studies, Mar 2022

   For insiders: We enhance a massive scholarly knowledge graph with improved author disambiguation, publication classification, and embeddings to improve search and analytics at scale.

   For everyone: We clean and enrich a huge catalog of scientific papers, like fixing name tags and adding smart labels so large-scale search works more reliably.

   Abstract DOI PDF

   Although several large knowledge graphs have been proposed in the scholarly field, such graphs are limited with respect to several data quality dimensions such as accuracy and coverage. In this article, we present methods for enhancing the Microsoft Academic Knowledge Graph (MAKG), a recently published large-scale knowledge graph containing metadata about scientific publications and associated authors, venues, and affiliations. Based on a qualitative analysis of the MAKG, we address three aspects. First, we adopt and evaluate unsupervised approaches for large-scale author name disambiguation. Second, we develop and evaluate methods for tagging publications by their discipline and by keywords, facilitating enhanced search and recommendation of publications and associated entities. Third, we compute and evaluate embeddings for all 239 million publications, 243 million authors, 49,000 journals, and 16,000 conference entities in the MAKG based on several state-of-the-art embedding techniques. Finally, we provide statistics for the updated MAKG. Our final MAKG is publicly available at https://makg.org and can be used for the search or recommendation of scholarly entities, as well as enhanced scientific impact quantification.

2. The Green AI Ontology: An Ontology for Modeling the Energy Consumption of AI Models

   Michael Färber, and David Lamprecht

   In ISWC, Virtual Conference, Hangzhou, China, Mar 2022

   For insiders: We introduce the Green AI Ontology to represent energy use and sustainability properties of AI models, enabling transparent reporting and comparison.

   For everyone: We build a “nutrition label” for AI models that records energy use and sustainability, so systems can be compared with clear facts.

   Abstract PDF

   Modeling AI systems’ characteristics of energy consumption and their sustainability level as an extension of the FAIR data principles has been considered only rudimentarily. In this paper, we propose the Green AI Ontology for modeling the energy consumption and other environmental aspects of AI models. We evaluate our ontology based on competency questions. Our ontology is available at https://w3id.org/Green-AI-Ontology and can be used in a variety of scenarios, ranging from comprehensive research data management to strategic controlling of institutions and environmental efforts in politics.

3. A Blocking-Based Approach to Enhance Large-Scale Reference Linking

   Tarek Saier, Meng Luan, and Michael Färber

   In LITE@JCDL, Mar 2022

   For insiders: We improve large-scale reference linking using a blocking strategy over reference fields.

   For everyone: We speed up linking references by sorting them into smart bins first, like organizing mail before matching addresses so more links are made correctly.

   Abstract PDF

   Analyses and applications based on bibliographic references are of ever increasing importance. However, reference linking methods described in the literature are only able to link around half of the references in papers. To improve the quality of reference linking in large scholarly data sets, we propose a blocking-based reference linking approach that utilizes a rich set of reference fields (title, author, journal, year, etc.) and is independent of a target collection of paper records to be linked to. We evaluate our approach on a corpus of 300,000 references. Relative to the original data, we achieve a 90% increase in papers linked through references, a five-fold increase in bibliographic coupling, and a nine-fold increase in in-text citations covered. The newly established links are of high quality (85% F1).

4. How Does Author Affiliation Affect Preprint Citation Count? Analyzing Citation Bias at the Institution and Country Level

   Chifumi Nishioka, Michael Färber, and Tarek Saier

   In JCDL, Cologne, Germany, Mar 2022

   For insiders: We quantify affiliation-based citation inequality for bioRxiv preprints and show stronger bias signals than in publisher versions.

   For everyone: We show that an author’s origin can influence citation outcomes, like a nameplate effect that raises fairness questions in how science is evaluated.

   Abstract DOI PDF

   Citing is an important aspect of scientific discourse and essential for quantifying the scientific impact of researchers. Previous works observed that citations are made not only based on scholarly contributions but also based on non-scholarly attributes, such as the affiliation or gender of authors, thereby producing citation bias. Existing research, however, has not analyzed preprints with respect to citation bias, even though they play an increasingly important role in modern scholarly communication. In this paper, we investigate whether preprints are affected by citation bias with respect to author affiliation. We measure citation bias for bioRxiv preprints and their publisher versions at the institution and country levels, using the Lorenz curve and Gini coefficient. This allows us to mitigate the effects of confounding factors and examine whether affiliation-related citation biases have an increased effect on preprint citations. We observe consistently higher Gini coefficients for preprints than for publisher versions. Thus, we can confirm that citation bias exists and is more severe for preprints. As preprints continue to rise in importance, affiliation-based citation bias becomes a relevant topic not only for authors (e.g., when deciding what to cite) but also for stakeholders who use citations for scientific impact quantification, such as funding agencies evaluating researchers based on citation counts.

5. Few-Shot Document-Level Relation Extraction

   Nicholas Popovic, and Michael Färber

   In NAACL, Seattle, WA, USA, Mar 2022

   For insiders: We introduce a few-shot benchmark for document-level relation extraction and reveal challenges beyond sentence-level settings, including realistic NOTA behavior.

   For everyone: We create a benchmark for extracting relationships from whole documents with only a few examples, like testing whether a student understood the full story and not just one line.

   Abstract DOI PDF

   We present FREDo, a few-shot document-level relation extraction (FSDLRE) benchmark. As opposed to existing benchmarks which are built on sentence-level relation extraction corpora, we argue that document-level corpora provide more realism, particularly regarding none-of-the-above (NOTA) distributions. Therefore, we propose a set of FSDLRE tasks and construct a benchmark based on two existing supervised learning data sets, DocRED and sciERC. We adapt the state-of-the-art sentence-level method MNAV to the document-level and develop it further for improved domain adaptation. We find FSDLRE to be a challenging setting with interesting new characteristics such as the ability to sample NOTA instances from the support set. The data, code, and trained models are available online.

6. AIFB-WebScience at SemEval-2022 Task 12: Relation Extraction First - Using Relation Extraction to Identify Entities

   Nicholas Popovic, Walter Laurito, and Michael Färber

   In SemEval@NAACL, Mar 2022

   For insiders: We build an end-to-end pipeline where relation extraction supports entity identification, improving extraction from LaTeX documents in a shared-task setting.

   For everyone: We help find names in technical text by learning relationships first.

   Abstract DOI PDF

   In this paper, we present an end-to-end joint entity and relation extraction approach based on transformer-based language models. We apply the model to the task of linking mathematical symbols to their descriptions in LaTeX documents. In contrast to existing approaches, which perform entity and relation extraction in sequence, our system incorporates information from relation extraction into entity extraction. This means that the system can be trained even on data sets where only a subset of all valid entity spans is annotated. We provide an extensive evaluation of the proposed system and its strengths and weaknesses. Our approach, which can be scaled dynamically in computational complexity at inference time, produces predictions with high precision and reaches 3rd place in the leaderboard of SemEval-2022 Task 12. For inputs in the domain of physics and math, it achieves high relation extraction macro F1 scores of 95.43% and 79.17%, respectively. The code used for training and evaluating our models is available at: https://github.com/nicpopovic/RE1st.

7. Which Publications’ Metadata Are in Which Bibliographic Databases? A System for Exploration

   Michael Färber, Christoph Braun, Nicholas Popovic, and 2 more authors

   In BIR@ECIR, Stavanger, Norway, Mar 2022

   For insiders: RefBee helps researchers check which publications are missing from major bibliographic databases, improving visibility and metadata completeness.

   For everyone: RefBee is a checklist for your publication record, like spotting missing books on a shelf before others do.

   Abstract PDF

   The choice of bibliographic databases containing publication metadata determines the available publication list of any author and thus their public appearance and evaluation. Ensuring that all publications are listed across databases is therefore important for researchers. However, as the average number of publications per researcher steadily rises, it becomes labor-intensive and time-consuming to verify whether all publications appear in all major databases. In this paper, we present RefBee, an online system that retrieves publication metadata for a given author from multiple bibliographic databases and indicates which publications are missing in which database. RefBee, available at http://refbee.org/, currently supports Wikidata, ORCID, Google Scholar, VIAF, DBLP, Dimensions, Microsoft Academic, Semantic Scholar, and DNB/GNB. The system can serve as an assistance tool for more than 4.7 million researchers across disciplines and languages, while also incentivizing the use and population of Wikidata in the scholarly domain.

8. Safety Aware Reinforcement Learning by Identifying Comprehensible Constraints in Expert Demonstrations

   Leopold Müller, Lars Böcking, and Michael Färber

   In SafeAI@AAAI, Mar 2022

   For insiders: We learn human-readable safety constraints from expert demonstrations to monitor RL agents, improving safety without losing interpretability.

   For everyone: We turn expert demonstrations into clear safety rules, like writing down a skilled driver’s habits so an AI can avoid dangerous moves.

   Abstract PDF

   When used in real-world environments, agents must meet high safety requirements as errors have direct consequences. Besides safety, the explainability of the systems is particularly important: errors should not only be avoided during learning, but the decision process must also be transparent. Existing approaches are limited to solving a single problem, whereas real-world scenarios require satisfying several criteria simultaneously. In this paper, we derive comprehensible rules from expert demonstrations that can be used to monitor the agent. The approach uses state-of-the-art classification and regression trees to derive safety rules, combined with concepts from association rule mining. The result is a compact and interpretable rule set that explains expert behavior and ensures safety. We evaluate our framework in common OpenAI environments, showing that it identifies safety-relevant rules and effectively imitates expert behavior, especially in edge cases. Further evaluations in higher-dimensional observation spaces and continuous action spaces demonstrate that the approach transfers well to new tasks while maintaining the compactness and comprehensibility of the rule set.

9. When to Use Which Neural Network? Finding the Right Neural Network Architecture for a Research Problem

   Michael Färber, and Nicolas Weber

   In SDU@AAAI, Mar 2022

   For insiders: We recommend neural-network architectures from textual problem descriptions and show that simpler models can be strong baselines for this task.

   For everyone: We recommend neural-network designs from a short problem description, like a tool clerk suggesting the right instrument once you describe the job.

   Abstract PDF

   Considering the increasing rate of scientific papers published in recent years, researchers across all disciplines face the challenge of keeping track of which latest scientific methods are suitable for which applications. In particular, an unmanageable number of neural network architectures has been introduced. In this paper, we propose the task of recommending neural network architectures based on textual problem descriptions. We frame this recommendation as a text classification task and develop appropriate text classification models for it. In experiments based on three datasets, we find that an SVM classifier outperforms a more complex model based on BERT. Overall, we provide evidence that neural network architecture recommendation is a nontrivial yet promising research topic.

10. Explaining Convolutional Neural Networks by Tagging Filters

    Anna Nguyen, Daniel Hagenmayer, Tobias Weller, and 1 more author

    In AIMLAI@CIKM, Mar 2022

    For insiders: We explain CNN decisions by tagging filters with human-readable labels, producing intuitive explanations and supporting error analysis.

    For everyone: We explain image classifiers by naming what their hidden filters react to, like labeling the internal sensors so people can understand why a decision happened.

    Abstract PDF

    Convolutional neural networks (CNNs) have achieved astonishing performance on various image classification tasks, but it is difficult for humans to understand how a classification comes about. Recent literature proposes methods to explain the classification process to humans, which mostly focus on visualizing feature maps and filter weights and are not very intuitive for non-experts. In this paper, we propose \textscFilTag, an approach to effectively explain CNNs even to non-experts. The idea is that if images of a class frequently activate a convolutional filter, that filter will be tagged with that class. Based on this tagging, individual image classifications can then be intuitively explained using the tags of the filters that the input image activates. Finally, we show that the tags are useful in analyzing classification errors caused by noisy input images and that the tags can be further processed by machines.

11. Sequence Labeling for Citation Field Extraction from Cyrillic Script References

    Igor Shapiro, Tarek Saier, and Michael Färber

    In SDU@AAAI, Mar 2022

    For insiders: We create a large multilingual dataset of Cyrillic references and train accurate sequence labelers for citation-field extraction beyond Latin scripts.

    For everyone: We teach models to parse Cyrillic references, like sorting foreign-language mail into the right slots so citation data can be extracted reliably.

    Abstract PDF

    Extracting structured data from bibliographic references is a crucial task for the creation of scholarly databases. While approaches, tools, and evaluation data sets for this task exist, there is a distinct lack of support for languages other than English and scripts other than the Latin alphabet, excluding a significant portion of scientific literature written in Cyrillic script languages. To address this problem, we introduce a new multilingual and multidisciplinary data set of over 100,000 labeled reference strings covering multiple Cyrillic languages, including over 700 manually labeled references with the remainder generated synthetically. Using random samples of varying size from this data, we train multiple high-performing sequence labeling BERT models and demonstrate the usability of the proposed data set. We further showcase a multilingual BERT model trained on the synthetic data and evaluated on the manually labeled references, achieving an F1 score of 0.93 and significantly outperforming a retrained state-of-the-art model evaluated on the same data.

2021

1. The Data Set Knowledge Graph: Creating a Linked Open Data Source for Data Sets

   Michael Färber, and David Lamprecht

   Quant. Science Studies, Mar 2021

   For insiders: We introduce the Data Set Knowledge Graph (DSKG), the first large-scale linked open dataset that connects datasets to the scholarly papers that mention them, enabling better dataset discovery and making data contributions transparent and measurable.

   For everyone: We link datasets to the papers that mention them, creating a map that helps people find data faster and gives data creators clearer credit.

   Abstract PDF

   Several scholarly knowledge graphs have been proposed to model and analyze the academic landscape; however, despite the remarkable increase in available data sets in recent years, existing knowledge graphs primarily focus on associated entities such as publications rather than on data sets themselves, and publicly available data set knowledge graphs do not systematically contain links to the publications in which data sets are mentioned. In this paper, we present an approach for constructing an RDF knowledge graph that fulfills these criteria. Our data set knowledge graph, DSKG, is publicly available and contains metadata of data sets across all scientific disciplines. To ensure high data quality, we identify suitable raw data set collections, establish links between data sets and publications modeled in the Microsoft Academic Knowledge Graph that mention these data sets, and address author name ambiguity by developing and evaluating an author name disambiguation method that enriches the knowledge graph with links to ORCID. Overall, the knowledge graph contains more than 2,000 data sets with associated properties and 814,000 links to 635,000 scientific publications, enabling a variety of use cases including advanced data set search systems and new ways of measuring and rewarding data set provisioning.

2. Cross-Lingual Citations in English Papers: A Large-Scale Analysis of Prevalence, Usage, and Impact

   Tarek Saier, Michael Färber, and Tornike Tsereteli

   International Journal on Digital Libraries, Mar 2021

   For insiders: We analyze cross-lingual citations at scale to quantify prevalence and impact, showing how language affects scholarly visibility and citation behavior.

   For everyone: We trace citations across languages like trade routes on a map, showing which languages get noticed and cited in English scientific writing.

   Abstract PDF

   Citation information in scholarly data is an important source of insight into the reception of publications and scholarly discourse, yet the outcomes of citation analyses and the applicability of citation-based machine learning approaches heavily depend on data completeness. A key shortcoming of current scholarly data is the frequent exclusion of non-English publications or the absence of language metadata, which has limited the study of cross-lingual citations. In this paper, we present an analysis of cross-lingual citations based on over one million English-language papers spanning three scientific disciplines and three decades, examining differences across cited languages and disciplines, temporal trends, and the usage characteristics and impact of cross-lingual citations. Our findings include an increasing rate of citations to Chinese-language publications, a predominance of citations to local non-English languages, and consistency in citation intent between cross-lingual and monolingual citations. To facilitate further research, we make the collected data and source code publicly available.

3. Improving Question Answering for Event-focused Questions in Temporal Collections of News Articles

   Jiexin Wang, Adam Jatowt, Michael Färber, and 1 more author

   Information Retrieval Journal, Mar 2021

   For insiders: We improve question answering over long-term news archives with temporal reranking, helping users retrieve relevant evidence for event-centric questions.

   For everyone: We make question answering over news archives time-aware, like a historian who knows which year to search when the question forgets to say it.

   Abstract PDF

   Temporal collections of news articles contain large numbers of accurate and time-aligned documents that are valuable for understanding past events, yet access to such archives remains difficult for average users due to their size and complexity. This work addresses the task of machine reading at scale on long-term news archives, motivated by the observation that questions over news collections are typically event-centric and exhibit strong temporal characteristics. We propose a large-scale question answering model specifically designed for long-term news article collections, incorporating a dedicated module for re-ranking articles using temporal information from multiple perspectives. Experimental results demonstrate that the proposed model outperforms existing question answering systems, primarily due to the temporal re-ranking module’s ability to identify more relevant documents.

4. Towards Full-Fledged Argument Search: A Framework for Extracting and Clustering Arguments from Unstructured Text

   Michael Färber, and Anna Steyer

   CoRR, Mar 2021

   For insiders: We propose a unified argument-search framework that extracts and clusters arguments from text for more comprehensive debate analysis.

   For everyone: We build a machine that mines arguments from messy text and sorts them into clusters, like turning a chaotic debate into neat stacks of note cards.

   Abstract PDF

   Argument search aims at identifying arguments in natural language texts, a task that has traditionally been addressed through a combination of keyword search and sentence- or document-level argument identification. However, existing frameworks often cover only specific components of argument search and fail to address key challenges, including argument–query matching for differently framed topics, identification of multi-sentence arguments, and topical clustering of retrieved arguments. In this paper, we propose a unified framework to overcome these limitations by combining keyword search with precomputed topic clusters for argument–query matching, introducing a novel sentence-level sequence-labeling approach for multi-sentence argument identification, and presenting aggregated arguments through topic-aware argument clustering. Experiments on several real-world debate data sets show that density-based clustering algorithms such as HDBSCAN are well suited for argument–query matching, that our BiLSTM-based sequence-labeling approach achieves a macro F1 score of 0.71, and that while fine-grained subtopic clustering of arguments remains challenging, it is a promising direction for further exploration.

5. Are Investors Biased Against Women? Analyzing How Gender Affects Startup Funding in Europe

   Michael Färber, and Alexander Klein

   CoRR, Mar 2021

   For insiders: We analyze European startup funding data and quantify how team gender composition relates to funding outcomes, providing evidence of structural bias risks.

   For everyone: We analyze European startup funding like checking a pipeline, asking whether team gender composition relates to how money flows and where bias may appear.

   Abstract PDF

   Raising capital from investors is a central challenge for startups, making it crucial for founders to understand whether and how gender bias affects early-stage funding decisions. While prior gender studies have predominantly focused on the US market, this paper provides a more comprehensive analysis of gender bias in European early-stage startup funding. We examine European startups listed on Crunchbase using Semantic Web technologies and analyze the relationship between the share of female founders in a founding team and the amount of funding raised. Our results show that a higher relative share of female founders has a negative impact on funding amounts and that founder characteristics influence funding outcomes differently based on gender. Furthermore, we find that gender bias is less pronounced for serial founders with prior entrepreneurial experience, as female founders benefit three times more than male founders from having previously founded a startup. Overall, our findings indicate that gender bias exists and should be considered in the context of startup funding.

6. Recommending Datasets for Scientific Problem Descriptions

   Michael Färber, and Ann-Kathrin Leisinger

   In CIKM, Virtual Event, Queensland, Australia, Mar 2021

   For insiders: We propose dataset recommendation from natural-language scientific problem descriptions, cutting dataset search time and improving dataset reuse at scale.

   For everyone: You describe your research problem in plain text and we suggest suitable datasets, like a matchmaking service that connects questions with the data that can answer them.

   Abstract DOI PDF

   The rapidly growing number of available datasets makes it increasingly difficult for researchers and practitioners to identify the most relevant datasets for a given research problem, a challenge that existing keyword-based dataset search engines address only imperfectly. In this paper, we propose a system that recommends suitable datasets directly from research problem descriptions, formulating the task as a domain-specific text classification problem. In a comprehensive offline evaluation using state-of-the-art models and large-scale data comprising 88,000 paper abstracts and 265,000 citation contexts as problem descriptions, our approach achieves an F1 score of 0.75. Additionally, a user study in real-world settings shows that users are satisfied in 88% of all test cases. These results indicate promising future directions for dataset recommendation systems.

7. DataHunter: A System for Finding Datasets Based on Scientific Problem Descriptions

   Michael Färber, and Ann-Kathrin Leisinger

   In RecSys, Amsterdam, The Netherlands, Mar 2021

   For insiders: DataHunter finds datasets from problem descriptions using scholarly text and the Data Set Knowledge Graph, supporting FAIR data discovery workflows.

   For everyone: DataHunter searches for datasets from a problem description, like a treasure map that points to the right data by reading what you want to study.

   Abstract DOI PDF

   The steadily increasing number of datasets makes it difficult for researchers and practitioners across scientific disciplines to identify the most relevant datasets for a given research problem, a limitation that existing keyword-based dataset search engines struggle to overcome for complex queries. In this paper, we present the system at http://data-hunter.io, which recommends suitable datasets based on research problem descriptions. The system leverages fastText for text representation and classification, the Data Set Knowledge Graph (DSKG) containing metadata for nearly 1,700 unique datasets, and 88,000 paper abstracts as research problem descriptions for model training. Overall, our results demonstrate that dataset recommendation supports data provisioning and reuse in line with the FAIR principles and that dataset recommendation constitutes a promising direction for future research.

8. Quantifying Explanations of Neural Networks in E-Commerce Based on LRP

   Anna Nguyen, Franz Krause, Daniel Hagenmayer, and 1 more author

   In ECML-PKDD, Bilbao, Spain, Mar 2021

   For insiders: We quantify explanation quality for neural networks in e-commerce settings, helping practitioners assess trust and compliance for recommendation models.

   For everyone: We measure how good an AI explanation in online shopping really is, like using a ruler to check whether it stays stable, useful, and trustworthy.

   Abstract DOI PDF

   Neural networks are widely used in e-commerce, particularly for product recommendation, yet their black-box nature makes it difficult to understand how recommendations are generated, which hinders trust and reliability. In this paper, we present TransPer, an explanation framework for neural networks that introduces novel explanation measures based on Layer-Wise Relevance Propagation and supports heterogeneous data as well as complex architectures composed of multiple interconnected neural networks. We apply and evaluate TransPer in two real-world online shop scenarios and show that the resulting explanations help to assess prediction quality, derive ideas for improving the neural network, gain deeper insights into customer behavior, and satisfy legal requirements such as those mandated by the GDPR.

9. Combining Linking Techniques for Everyone

   Kristian Noullet, Samuel Printz, and Michael Färber

   In ESWC, Mar 2021

   For insiders: We provide a framework for combining and comparing entity-linking components, improving reproducibility and fair evaluation across pipelines.

   For everyone: We make entity linking plug-and-play, like LEGO bricks that can be swapped and combined so users can compare systems without rebuilding everything.

   Abstract DOI PDF

   While the path in the field of Entity Linking (EL) has been long and has brought forth a plethora of approaches over the years, many of these are exceedingly difficult to execute for purposes of detailed analysis. In many cases, implementations are available, but are far from being a plug-and-play experience. We present Combining Linking Techniques (CLiT), a framework aimed at executing individual linking techniques as well as complex combinations thereof, with a strong focus on reusability, reproducibility, and comparability of existing systems. Furthermore, we introduce protocols for exchanging sub-pipeline-level information with existing and novel systems, enabling improved out-of-the-box compatibility. Among other use cases, our framework can be employed to consolidate multiple systems in combination with meta-learning approaches and to increase support for backward compatibility of existing benchmark annotation systems.

10. Media Bias Everywhere? A Vision for Dealing with the Manipulation of Public Opinion

    Michael Färber, and Frederic Bartscherer

    In BIAS@ECIR, Lucca, Italy, Mar 2021

    For insiders: We outline how AI can detect and communicate media bias, supporting transparency, media literacy, and more resilient public discourse.

    For everyone: We sketch a blueprint for fighting manipulation in the news, showing how AI can detect bias and present it in ways people can actually understand.

    Abstract DOI PDF

    This paper addresses the question of how artificial intelligence can be used to detect media bias within the broader context of manipulation and mood-making. We identify three fields of action that emerge from applying machine learning to the analysis of media bias: evaluation principles for media bias, information presentation of media bias, and transparency in media bias evaluation. Practical applications of our findings arise both in professional settings for journalists and publishers and in the everyday lives of citizens. First, automated analysis can be used to analyze texts in real time and promote more balanced reporting. Second, an intuitive web browser application can reveal existing bias in news texts in a way that is accessible to citizens. Finally, in educational contexts, pupils can experience media bias and the use of artificial intelligence in practice, thereby fostering media literacy.

11. C-Rex: A Comprehensive System for Recommending In-Text Citations with Explanations

    Michael Färber, Vinzenz Zinecker, Isabela Bragaglia Cartus, and 2 more authors

    In Sci-K@WWW, Virtual Event, Ljubljana, Slovenia, Mar 2021

    For insiders: C-Rex is an online system that recommends in-text citations and provides explanations, helping authors write more efficiently and responsibly.

    For everyone: C-Rex is an online citation assistant, like a helpful co-author that suggests supporting papers and briefly explains why they fit.

    Abstract DOI PDF

    Finding suitable citations for scientific publications can be challenging and time-consuming. To address this issue, context-aware citation recommendation approaches have been developed that recommend publications as candidates for in-text citations. In this paper, we present C-Rex, a web-based demonstration system available at http://c-rex.org for context-aware citation recommendation based on the Neural Citation Network and millions of publications from the Microsoft Academic Graph. Our system is one of the first online context-aware citation recommendation systems and the first to incorporate not only a deep learning–based recommendation approach but also explanation components to help users better understand why papers were recommended. In our offline evaluation, our model performs similarly to the one presented in the original paper and can serve as a basic framework for further implementations. In our online evaluation, we found that providing explanations for recommendations increased users’ satisfaction.

12. Exploding TV Sets and Disappointing Laptops: Suggesting Interesting Content in News Archives Based on Surprise Estimation

    Adam Jatowt, I-Chen Hung, Michael Färber, and 2 more authors

    In ECIR, Mar 2021

    For insiders: We recommend surprisingly interesting content from news archives by contrasting past and present, making digital heritage collections more engaging.

    For everyone: We find “surprising gems” in news archives by comparing past statements with today, like opening a time capsule and highlighting what changed.

    Abstract DOI PDF

    Many archival collections have recently been digitized and made available to a wide audience. However, the documents they contain often have limited appeal for ordinary users, as the content may appear obsolete or uninteresting. Archival document collections can become more attractive if suitable content is recommended to users. This work proposes a new research direction, termed Archival Content Suggestion, which aims to discover interesting content from long-term document archives that preserve information about societal history and heritage. To achieve this goal, we introduce two unsupervised approaches for automatically identifying interesting sentences from news article archives. Our methods detect interesting content by comparing information written in the past with information created in the present, thereby exploiting a surprise effect. Experiments on the New York Times corpus demonstrate that our approaches effectively retrieve interesting content.

13. Right for the Right Reasons: Making Image Classification Intuitively Explainable

    Anna Nguyen, Adrian Oberföll, and Michael Färber

    In ECIR, Virtual Event, Mar 2021

    For insiders: We propose a metric for object-aligned explanations in image classification and show training can make models rely on the right evidence.

    For everyone: We check whether image AIs look at the right evidence, like asking a student to point to the exact part of a picture that supports the answer.

    Abstract DOI PDF

    The effectiveness of Convolutional Neural Networks (CNNs) in image classification has been well established. To make classification decisions understandable to humans, various methods for visualizing classification evidence have been developed in recent years. These explanations reveal that images are sometimes classified correctly but for the wrong reasons, that is, based on incidental evidence rather than the actual objects of interest. Ideally, images should be classified correctly for the right reasons, namely based on the relevant evidence. To this end, we propose a new explanation quality metric, termed ObAlEx, which measures object-aligned explanations in image classification. By combining object detection approaches, explanation methods, and the ObAlEx metric, we quantify the extent to which CNNs focus on the actual evidence. Furthermore, we demonstrate that additional training of CNNs can improve this focus without reducing classification accuracy.

14. Identifying Used Methods and Datasets in Scientific Publications

    Michael Färber, Alexander Albers, and Felix Schüber

    In SDU@AAAI, Mar 2021

    For insiders: We identify automatically which methods and datasets were used in papers via extraction and context classification.

    For everyone: We extract which methods and datasets papers really used.

    Abstract PDF

    Although it has become common to assess publications and researchers based on citation counts (e.g., using the h-index), measuring the impact of scientific methods and datasets has so far received limited attention. This is largely due to the fact that information about the usage of methods and datasets is typically not stated explicitly by authors but is instead embedded within the text of publications. In this paper, we propose an approach for identifying methods and datasets in scientific texts that have actually been used by the authors. Our approach first detects dataset and method mentions using a domain-specific named entity recognition technique with minimal human interaction. It then classifies these mentions as used or non-used based on their textual context. The resulting labels are aggregated at the document level and integrated into the Microsoft Academic Knowledge Graph to enrich publication metadata. Experiments on the Microsoft Academic Graph demonstrate that both method and dataset mentions can be accurately identified and correctly classified with respect to their usage. Overall, our approach supports method and dataset recommendation, enhanced paper recommendation, and scientific impact quantification, and can be extended to identify mentions of arbitrary entity types (e.g., tasks).

15. Bootstrapping Multilingual Metadata Extraction: A Showcase in Cyrillic

    Johan Krause, Igor Shapiro, Tarek Saier, and 1 more author

    In SDP@NAACL, Mar 2021

    For insiders: We bootstrap multilingual metadata extraction for Cyrillic scholarly data, reducing barriers for non-Latin research communities and digital libraries.

    For everyone: We help computers read Cyrillic scholarly metadata, like teaching a librarian to catalog books in another script so they do not stay invisible.

    Abstract DOI PDF

    Applications based on scholarly data are of ever increasing importance. This results in disadvantages for areas where high-quality data and compatible systems are not available, such as Cyrillic-script regions. This work showcases techniques for bootstrapping multilingual metadata extraction.

16. AWARE: An Ontology for Situational Awareness of Autonomous Vehicles in Manufacturing

    Boulos El Asmar, Syrine Chelly, and Michael Färber

    In CSKG@AAAI, Mar 2021

    For insiders: We develop an ontology for situational awareness of autonomous factory vehicles, enabling structured reasoning about perception and safer operation.

    For everyone: We build a shared “mental map” for factory robots, like a common language that helps vehicles describe what they sense and choose safe actions.

    Abstract PDF

    With the development of autonomous vehicles, recent research has increasingly focused on the semantic representation of robotic proprioceptive and exteroceptive perceptions, that is, the perception of the robot’s own body and of the external world. Such semantic representations can be queried by reasoning systems to enable what may be referred to as machine awareness. While this aligns with the general goal of artificial intelligence to leverage commonsense knowledge, it particularly relates to challenges in knowledge representation and knowledge elicitation. In this work, we present an ontology for representing background knowledge that grounds the cognitive capabilities of autonomous transport robots. Existing ontologies from the domains of robotics and the Internet of Things are integrated into our approach. We further demonstrate the applicability and extensibility of the proposed ontology in the context of autonomous and automated vehicles deployed in an automobile manufacturing plant. Our robotic situational awareness ontology provides a foundation for organizing and controlling robots in future smart factories and illustrates how situational awareness supports the coexistence of smart autonomous agents.

17. Theories of Meaning for the Internet of Things

    Michael Färber, Yulia Svetashova, and Andreas Harth

    In Language, Cognition, and Mind, Mar 2021

    For insiders: We compare semantic theories for representing meaning in IoT and show why grounding, aligned perspectives, and dynamics must be addressed together.

    For everyone: We compare ways machines can connect words to the real world.

    Abstract DOI PDF

    In this chapter, we consider the theoretical foundations for representing knowledge in the Internet of Things context. Specifically, we consider (1) the model-theoretic semantics (i.e., extensional semantics), (2) the possible-world semantics (i.e., intensional semantics), (3) the situation semantics, and (4) the cognitive/distributional semantics. Given the peculiarities of the Internet of Things, we pay particular attention to (a) perception (i.e., how to establish a connection to the world), (b) intersubjectivity (i.e., how to align world representations), and (c) the dynamics of world knowledge (i.e., how to model events). We come to the conclusion that each of the semantic theories helps in modeling specific aspects, but does not sufficiently address all three aspects simultaneously.

2020

1. Citation Recommendation: Approaches and Datasets

   Michael Färber, and Adam Jatowt

   Int. J. Digit. Libr., Mar 2020

   For insiders: We survey citation recommendation methods and datasets and highlight evaluation pitfalls and open challenges for assisting scientific writing.

   For everyone: This survey explains citation recommendation, like a GPS for references, and summarizes which datasets and tests are needed before such tools deserve trust.

   Abstract DOI PDF

   Citation recommendation addresses the task of recommending relevant citations for a given text. Owing to the rapid growth of published scientific literature on the one hand and the necessity of citing the most appropriate works when authoring scientific texts on the other hand, citation recommendation has emerged as an important research area. In recent years, numerous approaches and evaluation datasets have been proposed. However, to the best of our knowledge, no comprehensive literature survey has explicitly focused on citation recommendation. In this article, we provide a thorough introduction to automatic citation recommendation research. We then present an overview of existing approaches and datasets, identifying their differences and commonalities across multiple dimensions. Finally, we discuss evaluation methods, highlight general challenges in evaluation, and outline ways to address them. While we restrict our survey to citation recommendation for scientific publications, as this document type has been studied most extensively, many of the observations and discussions are also applicable to other text types, such as news articles and encyclopedic content.

2. unarXive: A Large Scholarly Dataset With Publications’ Full Text, Annotated In-Text Citations, and Links to Metadata

   Tarek Saier, and Michael Färber

   Scientometrics, Mar 2020

   For insiders: We present unarXive, a large-scale arXiv full-text corpus with precisely linked in-text citations and rich metadata, empowering citation-aware, transparent, and reproducible scholarly NLP.

   For everyone: We build a large “paper warehouse” where citations inside arXiv texts are tagged and linked, so citation analyses need far less manual cleanup.

   Abstract DOI PDF

   In recent years, scholarly datasets have been employed for a wide range of tasks, including paper recommendation, citation recommendation, citation context analysis, and citation context-based document summarization. The evaluation of methods for these tasks, as well as their applicability in real-world settings, critically depends on the underlying datasets. However, existing scholarly datasets exhibit several limitations. In this paper, we introduce a new dataset comprising all publications from all scientific disciplines available on arXiv.org. In addition to providing the full plain text of papers, in-text citations are annotated using global identifiers. Moreover, both citing and cited publications are linked to the Microsoft Academic Graph, enabling access to rich bibliographic metadata. The resulting dataset contains over one million documents and 29.2 million citation contexts. Freely available for research purposes, this dataset supports more robust evaluation of paper-based and citation context-based approaches and also enables novel analyses of in-text citations, as demonstrated prototypically in this work.

3. Analyzing the GitHub Repositories of Research Papers

   Michael Färber

   In JCDL, Virtual Event, China, Mar 2020

   For insiders: We analyze GitHub repositories linked from research papers to study maintenance and documentation patterns, providing signals for research reproducibility.

   For everyone: We analyze the GitHub code behind research papers, like checking the health of the software that scientific results depend on in practice.

   Abstract DOI PDF

   Linking to code repositories, such as those hosted on GitHub, has become increasingly common in scientific publications, particularly in computer science. However, the actual quality, maintenance, and usage of these repositories have so far remained largely unexplored. In this paper, we present the first comprehensive analysis of all GitHub code repositories referenced in scientific papers, using the Microsoft Academic Graph as the underlying data source. We analyze both the repositories and their associated publications across multiple dimensions. Our results show that the numbers of stars and forks across repositories follow a power-law distribution. In most cases, only a single author of a paper actively contributes to the corresponding repository. Repository documentation is typically concise, consisting of only a few sentences, and the source code is predominantly written in Python. The publications linking to code repositories, as well as their authors, are primarily associated with the artificial intelligence research community.

4. HybridCite: A Hybrid Model for Context-Aware Citation Recommendation

   Michael Färber, and Ashwath Sampath

   In JCDL, Virtual Event, China, Mar 2020

   For insiders: HybridCite combines embedding and IR signals to improve context-aware citation recommendation beyond single-method baselines.

   For everyone: HybridCite recommends citations by combining several signals, like using both a compass and a map so suggestions fit the writing context better.

   Abstract DOI PDF

   Citation recommendation systems seek to suggest relevant references either for an entire paper or for a specific text passage, referred to as a citation context. Recommending citations at the level of citation contexts, known as local citation recommendation, is the focus of this paper. We first develop citation recommendation approaches based on embeddings, topic modeling, and information retrieval techniques. To the best of our knowledge, we then combine the best-performing methods for the first time into a semi-genetic hybrid citation recommender system. We evaluate both the individual approaches and the hybrid model in offline experiments using several datasets, including the Microsoft Academic Graph (MAG) as well as MAG combined with arXiv and ACL. In addition, we conduct an online user study to assess the proposed methods. Our results demonstrate that a hybrid model integrating embedding-based and information retrieval–based components substantially outperforms its individual components as well as other baseline approaches.

5. Who’s Behind That Website? Classifying Websites by the Degree of Commercial Intent

   Michael Färber, Benjamin Scheer, and Frederic Bartscherer

   In ICWE, Helsinki, Finland, Mar 2020

   For insiders: We classify websites by commercial intent with ML, helping distinguish individuals, companies, non-profits, and public institutions from sparse web content.

   For everyone: We sort websites by whether they are commercial or not, like a quick “smell test” that works even when there is little text.

   Abstract DOI PDF

   Web hosting companies aim to provide customized customer services and therefore seek to identify the commercial intent of websites. Determining whether a website is operated by an individual, a company, a non-profit organization, or a public institution is a challenging classification task, as website content is often sparse or limited. In this paper, we present a novel approach for identifying the commercial intent of websites using a combination of supervised and unsupervised machine learning techniques. Based on a large real-world dataset, we evaluate our model in terms of effectiveness and efficiency and find that a multilayer perceptron achieves the best performance.

6. Annotating and Analyzing Biased Sentences in News Articles using Crowdsourcing

   Sora Lim, Adam Jatowt, Michael Färber, and 1 more author

   In LREC, Marseille, France, Mar 2020

   For insiders: We create and analyze a crowdsourced dataset of biased sentence annotations, enabling fine-grained study of subtle news bias.

   For everyone: We crowdsource labels of subtle bias in news, like using many small magnifying glasses to spot slanted wording sentence by sentence.

   Abstract PDF

   The spread and consumption of biased news has become a significant societal concern. Researchers from various fields, including social sciences and media studies, have proposed methods to address media bias, often leveraging techniques from natural language processing and machine learning to automatically identify biased content. However, progress in this area has been hindered by the lack of publicly available datasets, particularly those providing fine-grained bias annotations at the sentence level. In this paper, we introduce a novel news bias dataset designed to support the development and evaluation of methods for detecting subtle bias in news articles and for analyzing the characteristics of biased sentences. The dataset comprises 966 sentences from 46 English-language news articles covering four distinct events and includes sentence-level bias annotations. To ensure scalability, the annotations were obtained via crowdsourcing. Our dataset enables detailed analyses of news bias and serves as a benchmark for developing and evaluating bias detection methods, as well as a resource for related research areas such as fake news detection.

7. KORE 50^DYWC: An Evaluation Data Set for Entity Linking Based on DBpedia, YAGO, Wikidata, and Crunchbase

   Kristian Noullet, Rico Mix, and Michael Färber

   In LREC, Marseille, France, Mar 2020

   For insiders: We extend the KORE 50 entity-linking benchmark to multiple knowledge graphs, enabling broader evaluation across DBpedia, Wikidata, YAGO, and Crunchbase.

   For everyone: We take a well-known test set for linking names to real-world entities and expand it to several different knowledge bases, like checking whether a navigation system still works when you switch from one city map to another.

   Abstract PDF

   Named entity recognition and disambiguation (NERD) constitutes a major research area in natural language processing. A common approach to addressing this task relies on Semantic Web technologies and structured data formats, which facilitate information extraction and enable the construction of knowledge graphs. To properly evaluate NERD systems, high-quality gold standard datasets are required. While numerous evaluation datasets exist, they are predominantly based on Wikipedia or DBpedia. In this work, we extend the widely used gold standard dataset KORE 50 to support NERD evaluation not only for DBpedia, but also for YAGO, Wikidata, and Crunchbase. The resulting dataset, KORE 50^DYWC, enables a substantially broader evaluation landscape. In particular, it allows for the assessment of the knowledge-graph agnosticity of NERD systems across multiple knowledge graphs, which, to the best of our knowledge, has not previously been possible at this scale.

8. Answering Event-Related Questions over Long-Term News Article Archives

   Jiexin Wang, Adam Jatowt, Michael Färber, and 1 more author

   In ECIR, Lisbon, Portugal, Mar 2020

   For insiders: We answer event-related questions over long-term news archives using temporal reranking, improving retrieval when questions have implicit time scopes.

   For everyone: We answer event questions in huge news archives using time clues, like flipping to the right chapter of history even when the question forgets to name the year.

   Abstract DOI PDF

   Long-term news article archives constitute valuable resources for understanding past events, as they provide detailed information about occurrences at specific points in time. To better exploit such heritage collections, this work addresses the task of large-scale question answering over long-term news archives. Questions posed to these archives are often event-related and typically exhibit strong temporal characteristics. They can be broadly categorized into two types: those containing explicit temporal expressions and those that are only implicitly associated with particular time periods. We focus on the latter category, as such questions are more challenging to answer, and propose a retriever–reader model augmented with an additional module that reranks articles by leveraging temporal information from multiple perspectives. Experimental results on a carefully constructed test set demonstrate that our approach outperforms existing question answering systems, primarily due to the enhanced document retrieval enabled by the temporal reranking module.

9. Semantic Modelling of Citation Contexts for Context-Aware Citation Recommendation

   Tarek Saier, and Michael Färber

   In ECIR, Lisbon, Portugal, Mar 2020

   For insiders: We model citation contexts using entities and claim structure and show improved context-aware citation recommendation at scale.

   For everyone: We turn citation contexts into structured signals, like extracting the ingredients and purpose of a reference so models understand more than nearby words.

   Abstract DOI PDF

   The rapid pace at which new research is being published makes it increasingly infeasible for scholars to read and assess all potentially relevant work. To address this challenge, automated processing of scientific publications—such as semantic modeling to support digital handling and the development of information filtering systems—has become an active research area. In this paper, we investigate the benefits of semantically modeling citation contexts for citation recommendation. To this end, we develop semantic representations of citation contexts based on entities and claim structures. We evaluate the effectiveness and conceptual validity of the proposed models through a large-scale offline evaluation on multiple datasets and complement this analysis with a user study. Our results show that the semantic models outperform a non-semantic baseline and effectively capture the types of information they are designed to represent.

10. Neural Citation Recommendation: A Reproducibility Study

    Michael Färber, Timo Klein, and Joan Sigloch

    In BIR@ECIR, Lisbon, Portugal, Mar 2020

    For insiders: We reimplement a neural citation recommender and analyze hyperparameter effects, improving reproducibility and providing practical tuning guidance.

    For everyone: We rebuild a well-known citation recommender from scratch, like re-running a recipe to check whether others can truly reproduce the same result.

    Abstract PDF

    Context-aware citation recommendation aims to reduce the manual effort of searching for relevant references by automatically suggesting suitable papers as citations for a given input text. In this paper, we investigate the reproducibility of a state-of-the-art approach to context-aware citation recommendation, namely the Neural Citation Network (NCN) proposed by Ebesu and Fang. We reimplement the network and evaluate it on both RefSeer, the dataset used in the original work, and arXiv CS as an additional dataset. Furthermore, we analyze the impact of different neural network hyperparameters on the performance of the NCN and demonstrate how they can be tuned to improve results. Through this study, we contribute to increased transparency in citation recommendation research and support the development of more effective neural network–based models and evaluation practices.

2019

1. Linked Crunchbase: A Linked Data API and RDF Data Set About Innovative Companies

   Michael Färber

   CoRR, Mar 2019

   For insiders: We provide a Linked Data API and RDF dataset for Crunchbase, enabling machine-readable analytics of startups, organizations, people, and investments.

   For everyone: We publish Crunchbase as a web-friendly data stream, like adding a spout so machines can “drink” structured startup information.

   Abstract DOI PDF

   Crunchbase is an online platform that aggregates information about startups and technology companies, including attributes and relationships among companies, people, and investments. A large portion of the data contained in Crunchbase is not available from other sources, making it a unique resource. In this paper, we present an approach for bringing Crunchbase data to the Web of Data, enabling its use in a machine-readable RDF format. We first describe the development and hosting of a Linked Data API for Crunchbase, including the integration of \textttsameAs links to external data sources. We then outline our method for crawling RDF data via this API to construct a custom Crunchbase RDF knowledge graph. The resulting dataset comprises over 347 million triples, including 781,000 people, 659,000 organizations, and 343,000 investments. The Crunchbase Linked Data API is publicly available at \urlhttp://linked-crunchbase.org/.

2. The Microsoft Academic Knowledge Graph: A Linked Data Source with 8 Billion Triples of Scholarly Data

   Michael Färber

   In ISWC, Auckland, New Zealand, Mar 2019

   For insiders: We release the Microsoft Academic Knowledge Graph in RDF with billions of triples and embeddings, enabling large-scale scholarly integration and analytics.

   For everyone: We publish a huge linked dataset of scholarly metadata, like turning a library card catalog into a connected web map that machines can navigate.

   Abstract DOI PDF

   In this paper, we present the Microsoft Academic Knowledge Graph (MAKG), a large-scale RDF dataset comprising over eight billion triples that describe scientific publications and related entities, including authors, institutions, journals, and fields of study. The dataset is released under the Open Data Commons Attribution License (ODC-By). By providing the data both as RDF dump files and as a data source within the Linked Open Data cloud—featuring resolvable URIs and links to external data sources—we make a substantial body of scholarly data available on the Web of Data. In addition, we provide entity embeddings for all 210 million publications represented in the graph. The MAKG supports a wide range of use cases, particularly in the domain of digital libraries, such as (1) entity-centric exploration of publications, researchers, and affiliations; (2) data integration tasks using RDF as a common data model and interlinking with other datasets; and (3) large-scale data analysis and knowledge discovery over scholarly data.

3. Identifying Twitter Bots Using a Convolutional Neural Network

   Michael Färber, Agon Qurdina, and Lule Ahmedi

   In CLEF, Lugano, Switzerland, Mar 2019

   For insiders: We detect Twitter bots from tweet text with CNNs and achieve strong accuracy on a shared evaluation task for automated bot profiling.

   For everyone: We detect Twitter bots from their writing style, like using a metal detector to separate automated noise from real people.

   Abstract PDF

   In this paper, we present an approach for identifying Twitter bots based on the textual content of their tweets using a convolutional neural network. We experiment with various word embedding methods, including both pretrained embeddings and embeddings learned directly from the training data, as well as with different convolutional neural network architectures, and systematically compare their performance. When evaluated on the official test set of the CLEF 2019 Bots Profiling Subtask for the English language, our best-performing model achieves an accuracy of 90.34%. These results indicate that convolutional neural networks constitute a promising machine learning technique for Twitter bot detection.

4. Determining How Citations Are Used in Citation Contexts

   Michael Färber, and Ashwath Sampath

   In TPDL, Oslo, Norway, Mar 2019

   For insiders: We define fine-grained categories of citation usage and train models to classify them, improving citation analysis and recommendation.

   For everyone: We label how citations are used inside sentences, like tagging whether a reference backs a claim or just explains a concept.

   Abstract DOI PDF

   Citations have previously been classified based on their textual contexts with respect to criteria such as worthiness, function, polarity, and importance. To the best of our knowledge, however, citations have not yet been automatically classified according to their grammatical role, that is, whether a citation (1) is grammatically integrated into the sentence, (2) is placed directly after an occurrence of author names, (3) supports a concept, (4) supports a claim, or (5) is inappropriate due to incomplete or noisy context. We argue that identifying such classes of citation contexts is beneficial for various tasks, including improved citation recommendation and scientific impact quantification. In this paper, we introduce this classification scheme and propose a machine-learning-based approach for automatically assigning citation contexts to these classes. Our evaluation shows that classification performance differs substantially across the various citation types.

5. Evaluating the Availability of Open Citation Data

   Chifumi Nishioka, and Michael Färber

   In BIRNDL@SIGIR, Paris, France, Mar 2019

   For insiders: We quantify how much citation data is openly available and how openness evolves over time, informing open-science policy and research evaluation.

   For everyone: We take a census of open citation data and measure how much of the scientific reference network is visible to the public and how it changes over time.

   Abstract PDF

   Citation data of scientific publications is essential for a variety of purposes, including research evaluation and the construction of digital library collections. In this paper, we analyze the extent to which citation data of scholarly publications is openly available. Our analysis is based on the intersection of Crossref metadata and an Unpaywall snapshot as the publication dataset, and the COCI dataset as the source of open citation data. We show that citation data is openly available for 24.2% of publications, while it is closed for 16.6%. Furthermore, we find that the proportion of publications with open citation data has increased over time. Our results indicate that publishing a paper as open access does not influence whether its citation data is openly available. However, publications appearing in journals listed in the Directory of Open Access Journals (DOAJ) tend to have a higher share of openly available citation data compared to publications from other journals.

6. Relational schemata for distributed SPARQL query processing

   Victor Anthony Arrascue Ayala, Polina Koleva, Anas Alzogbi, and 6 more authors

   In SBD@SIGMOD, Amsterdam, The Netherlands, Mar 2019

   For insiders: We compare relational schema designs for distributed SPARQL on Spark and show how schema choice impacts performance on standard RDF benchmarks.

   For everyone: We compare different storage “blueprints” for linked data on big clusters, like testing how shelf layouts change how fast you find the right book.

   Abstract DOI PDF

   To leverage mature database technologies, RDF stores are often built on top of relational databases, with SPARQL queries translated into SQL. Scalability can be achieved by executing query processing over large RDF datasets in a distributed manner using shared-nothing computer clusters. In this paper, we investigate the impact of relational schema design when SPARQL queries are mapped to Apache Spark SQL. We consider several schema designs, including a single triple table, predicate-partitioned tables, a single wide table covering all properties, and a set of tables derived from an application model specification, referred to as a domain-dependent schema. For each approach, the corresponding tables are stored in the distributed file system HDFS using the columnar storage format Parquet. Experiments conducted with standard benchmarks show that the single wide property table approach, despite its simplicity, outperforms the other designs. Additional experiments further demonstrate that this approach remains advantageous even when repartitioning by key (RDF subject) is applied prior to query execution.

7. ScholarSight: Visualizing Temporal Trends of Scientific Concepts

   Michael Färber, Chifumi Nishioka, and Adam Jatowt

   In JCDL, Champaign, IL, USA, Mar 2019

   For insiders: ScholarSight visualizes temporal trends of scientific concepts extracted from papers, helping researchers explore emerging and fading topics.

   For everyone: ScholarSight works like a trend telescope for science, showing which ideas rise or fade so researchers can see what is heating up and what is cooling down.

   Abstract DOI PDF

   In this paper, we present a system for exploring temporal trends of scientific concepts. The concepts are extracted by identifying noun phrases and entities from all computer science publications available on arXiv.org. The system enables users to examine time series for a large number of concepts and to identify both positively and negatively trending concepts over time. By applying clustering techniques and visualizations for cluster analysis, the system further reveals groups of concepts that exhibit similar usage patterns across different periods. This system supports exploratory analyses for researchers across scientific disciplines and provides a useful tool for scholars in bibliometrics and scientometrics to study the evolution of scientific concepts.

8. Team Peter Brinkmann at SemEval-2019 Task 4: Detecting Biased News Articles Using Convolutional Neural Networks

   Michael Färber, Agon Qurdina, and Lule Ahmedi

   In SemEval@NAACL-HLT 2019, Minneapolis, MN, USA, Mar 2019

   For insiders: We detect hyperpartisan biased news with neural text models and analyze trade-offs in precision and recall on a large shared-task benchmark.

   For everyone: We build systems that flag extremely one-sided news, like an early warning label for strongly biased reporting, and we study where these systems fail.

   Abstract DOI PDF

   In this paper, we present an approach for classifying news articles as biased, that is, hyperpartisan, or unbiased using a convolutional neural network. We experiment with various word embedding methods, including both pretrained embeddings and embeddings learned directly from the training data, as well as with different convolutional neural network architectures, and compare their performance. When evaluated on the official test set of SemEval 2019 Task 4, our best-performing model achieves relatively low precision and accuracy, but attains the highest recall among all 42 participating teams.

9. Bibliometric-Enhanced arXiv: A Data Set for Paper-Based and Citation-Based Tasks

   Tarek Saier, and Michael Färber

   In BIR@ECIR, Cologne, Germany, Mar 2019

   For insiders: We provide a bibliometric-enhanced arXiv dataset for paper- and citation-based tasks.

   For everyone: We add citation links and metadata to arXiv papers, like attaching extra labels to millions of documents so recommendation tools can be tested fairly.

   Abstract PDF

   In recent years, a variety of research paper–based tasks, such as paper recommendation, as well as citation-based tasks, including citation recommendation and citation context-based document summarization, have been proposed. The evaluation of methods for these tasks and their applicability in real-world settings strongly depend on the underlying datasets. However, existing datasets suffer from several limitations. In this paper, we introduce a new dataset derived from all publications across all scientific disciplines available on arXiv.org. In addition to providing the full plain text of papers, in-text citations are annotated using global identifiers, and cited publications are linked to the Microsoft Academic Graph wherever possible. The resulting dataset comprises over one million documents and 29.2 million citation contexts. Freely available for research purposes, this dataset supports more reliable evaluation of research paper–based and citation context–based approaches and also serves as a foundation for novel methods to analyze scientific publications.

10. Finding Temporal Trends of Scientific Concepts

    Michael Färber, and Adam Jatowt

    In BIR@ECIR, Cologne, Germany, Mar 2019

    For insiders: We track fine-grained scientific concept trends in arXiv full texts using statistical trend tests, helping understand how research topics evolve.

    For everyone: We track fine-grained scientific ideas over time, like a seismograph for concepts that shows which ones steadily grow and which ones fade.

    Abstract PDF

    Science evolves rapidly, and prior work has primarily investigated the evolution of coarse-grained research topics. To the best of our knowledge, however, no large-scale analysis of temporal trends of fine-grained scientific concepts based on full-text publications has been conducted. In this paper, we extract noun phrases as fine-grained concepts from the full texts of all computer science papers available on arXiv.org. We then identify positive and negative temporal trends using simple linear regression, the Mann–Kendall test, and the Theil–Sen estimator. Our experimental results show that the Mann–Kendall test yields particularly meaningful trend detection outcomes, whereas the Theil–Sen estimator and simple linear regression often identify non-scientific concepts as trending. The presented findings are relevant for both general researchers and scholars in bibliometrics and scientometrics who are interested in analyzing the evolution of scientific concepts.

11. PaperHunter: A System for Exploring Papers and Citation Contexts

    Michael Färber, Ashwath Sampath, and Adam Jatowt

    In ECIR, Cologne, Germany, Mar 2019

    For insiders: PaperHunter lets users search citation contexts and polarity, enabling qualitative exploration of how papers are cited and discussed.

    For everyone: PaperHunter lets you search the exact sentences where a paper is cited, like finding every time a name is mentioned and reading the surrounding context.

    Abstract DOI PDF

    In this paper, we present a system that enables researchers to search for scientific papers and in-text citations in a novel manner. In particular, the system allows users to retrieve the textual contexts in which publications are cited, referred to as citation contexts, by specifying either the cited paper’s title or the cited author’s name. To support qualitative assessment of citations, the system provides information on citation polarity, indicating whether a cited work is discussed in a positive, neutral, or negative manner. The system is built on the full corpus of computer science papers available on arXiv.org and can be used both by computer science researchers to reflect on how their work is perceived within the scientific community and by scholars conducting citation analysis research.

2018

1. A Linked Data Wrapper for CrunchBase

   Michael Färber, Carsten Menne, and Andreas Harth

   Semantic Web, Mar 2018

   For insiders: We provide Crunchbase as Linked Data, enabling richer integration and querying of startup, people, and investment data with web standards.

   For everyone: We publish Crunchbase as Linked Data, turning startup and investment information into a connected map that is easier to combine, query, and analyze.

   Abstract DOI PDF

   CrunchBase is a database of startups and technology companies that can be searched, browsed, and edited via a web interface and is also accessible through an entity-centric HTTP API providing data in JSON format. In this paper, we present a wrapper for the CrunchBase API that exposes the data as Linked Data. The wrapper provides schema-level links to established vocabularies such as schema.org, Friend-of-a-Friend, and Vocabulary-of-a-Friend, as well as entity-level links to DBpedia for organization entities. We further describe how the RDF data can be harvested to create a local copy for advanced processing and querying beyond the capabilities of the original CrunchBase API. Finally, we outline several use cases in which both the Linked Data API for CrunchBase and the crawled CrunchBase RDF dataset have been applied in subsequent research.

2. Linked Data Quality of DBpedia, Freebase, OpenCyc, Wikidata, and YAGO

   Michael Färber, Frederic Bartscherer, Carsten Menne, and 1 more author

   Semantic Web, Mar 2018

   For insiders: We compare major knowledge graphs using a systematic data-quality framework and help practitioners choose the right graph for their application needs.

   For everyone: We create a consumer-style test report for major knowledge graphs, comparing data quality so developers can choose the right “map of the world” for their use case.

   Abstract DOI PDF

   In recent years, several large, cross-domain, and openly available knowledge graphs (KGs) have been created, including DBpedia, Freebase, OpenCyc, Wikidata, and YAGO. Despite their widespread use, these knowledge graphs have not yet been subjected to a comprehensive comparative analysis. In this survey, we introduce a set of data quality criteria for systematically analyzing knowledge graphs and apply these criteria to compare the aforementioned KGs. In addition, we propose a framework to support the selection of the most suitable knowledge graph for a given application scenario.

3. Which Knowledge Graph Is Best for Me?

   Michael Färber, and Achim Rettinger

   CoRR, Mar 2018

   For insiders: We translate knowledge graph quality analysis into practical selection rules so users can pick a knowledge graph that matches their data quality requirements.

   For everyone: We offer a practical “which map should I buy?” guide for knowledge graphs, turning quality checks into simple rules that help people choose.

   Abstract PDF

   In recent years, DBpedia, Freebase, OpenCyc, Wikidata, and YAGO have emerged as prominent large, cross-domain, and freely available knowledge graphs. Despite their widespread adoption, these knowledge graphs are difficult to compare within specific application settings, making it challenging for researchers and developers to select the most suitable one for their needs. In a prior in-depth survey, we defined and applied a set of data quality criteria to systematically analyze these knowledge graphs and proposed a decision-support framework for identifying the most appropriate KG for a given context. In this paper, we aim to lower the barrier to accessing these results by presenting simplified rules that map individual data quality requirements to specific knowledge graphs. This work is intended as a complement to, rather than a replacement for, the previously introduced decision-support framework, to which we refer readers seeking a well-informed selection process.

4. A High-Quality Gold Standard for Citation-based Tasks

   Michael Färber, Alexander Thiemann, and Adam Jatowt

   In LREC, Miyazaki, Japan, Mar 2018

   For insiders: We create a high-quality arXiv-based gold standard with correctly linked citations, enabling reliable evaluation of citation-centric NLP tasks.

   For everyone: We build a carefully checked reference set where citations in arXiv full texts are linked correctly.

   Abstract PDF

   Analyzing and recommending citations within their specific citation contexts has gained increasing attention due to the rapid growth of scientific literature. Although datasets such as CiteSeerX have been used to evaluate approaches for citation-dependent tasks, they suffer from substantial limitations, stemming from the challenges of information extraction, entity linking, and entity resolution. In this paper, we introduce a new evaluation dataset for citation-dependent tasks based on publications from arXiv.org. The dataset is distinguished by an almost noise-free extraction process and by the fact that all citations are correctly linked to their referenced publications. In addition to providing sentence-level textual content, cited publications are annotated directly within the text using global identifiers and, where possible, are further linked to the DBLP Computer Science Bibliography. The resulting dataset comprises over 15 million sentences and is freely available for research purposes. It supports the training and evaluation of a wide range of citation-based tasks, including citation recommendation, citation function and importance classification, and citation-based document summarization.

5. PRoST: Distributed Execution of SPARQL Queries Using Mixed Partitioning Strategies

   Matteo Cossu, Michael Färber, and Georg Lausen

   In EDBT, Vienna, Austria, Mar 2018

   For insiders: PRoST is a Spark-based RDF store that speeds up SPARQL by combining vertical partitioning and property tables for scalable graph querying.

   For everyone: PRoST speeds up queries on linked data by changing the “warehouse layout”, so answers can be found faster on big computing clusters.

   Abstract DOI PDF

   The rapidly increasing size of RDF graphs in recent years has created a strong need for distributed storage and parallel query processing strategies. To enable efficient query processing on computer clusters, a wide range of approaches has been proposed, including systems built on top of Hadoop HDFS using technologies such as Apache Accumulo or Apache Spark. In this paper, we introduce a new RDF store called PRoST (Partitioned RDF on Spark Tables) that is based on Apache Spark. PRoST employs a novel storage strategy that combines vertical partitioning with property tables, two established models for RDF data storage. We show that PRoST outperforms state-of-the-art RDF stores in terms of query execution time across a broad spectrum of query types, while requiring no extensive precomputation phase.

6. CITEWERTs: A System Combining Cite-Worthiness with Citation Recommendation

   Michael Färber, Alexander Thiemann, and Adam Jatowt

   In ECIR, Grenoble, France, Mar 2018

   For insiders: CITEWERTs combines cite-worthiness detection with citation recommendation to guide writers to sentences that need citations and suggest references.

   For everyone: CITEWERT first checks whether a sentence really needs a citation, like a spell-checker for missing references, and then suggests suitable papers.

   Abstract DOI PDF

   The vast and continuously growing number of scientific publications has created a need for automatically recommending citations for specific text segments in scholarly documents. Despite this demand, only a limited number of citation-based recommender system demonstrations have been proposed so far. Existing solutions often either ignore the raw textual context, recommend citations only for predefined citation contexts, or operate at the document level. In contrast, we propose a novel two-step architecture for citation recommendation. First, given an input text, the system determines for each potential citation context—typically at the sentence level—whether it is actually citation-worthy. If so, the system then recommends suitable citations for that context. In our demonstration, we illustrate how this architecture guides users directly to sentences that require citations and presents relevant recommendations for individual sentences, thereby reducing the effort needed to review large numbers of sentences and suggested references.

7. To Cite, or Not to Cite? Detecting Citation Contexts in Text

   Michael Färber, Alexander Thiemann, and Adam Jatowt

   In ECIR, Grenoble, France, Mar 2018

   For insiders: We detect which sentences truly require citations, reducing unnecessary recommendations and better matching real scientific writing behavior.

   For everyone: We teach a system to spot which sentences should cite sources.

   Abstract DOI PDF

   Recommending citations for scientific texts and other document types, such as news articles, has recently received considerable attention. Existing citation recommendation approaches, however, typically do not explicitly address whether a given text segment, for example a sentence, actually deserves a citation. Introducing a prior step that determines the cite-worthiness of potential citation contexts is advantageous, as it both reduces the number of computationally expensive recommendation operations and better reflects human citation behavior by avoiding excessive or insufficient recommendations. In this paper, we propose a convolutional recurrent neural network–based approach for classifying potential citation contexts with respect to their cite-worthiness. Our experimental results show that the proposed method significantly outperforms a baseline approach and reduces the number of citation recommendations to approximately one tenth.

8. Annotating Domain-Specific Texts with Babelfy: A Case Study

   Michael Färber, Kristian Noullet, and Boulos El Asmar

   In EYRE, Mar 2018

   For insiders: We evaluate entity linking on German customer feedback using Babelfy and show strong annotation accuracy for practical domain text workflows.

   For everyone: We test how well a knowledge tool tags messy, specialized text, like putting accurate name tags on customer feedback so it becomes searchable.

   Abstract PDF

   Freely available large knowledge graphs, such as DBpedia, Wikidata, and YAGO, provide comprehensive representations of general knowledge and therefore form a solid basis for text annotation. However, their application to domain-specific text documents requires careful consideration, and reports on real-world use cases of entity linking based on such knowledge graphs remain scarce. In this paper, we describe a practical use case of annotating German-language customer feedback texts using Babelfy as the text annotation service. We conduct a manual evaluation of the resulting annotations and show that approximately 85% of the annotations are correct. These results indicate that Babelfy as an annotation approach and BabelNet as its underlying knowledge graph constitute a suitable baseline for the development of domain-specific knowledge graphs and customized entity linking methods.

2017

1. The xLiMe system: Cross-lingual and cross-modal semantic annotation, search and recommendation over live-TV, news and social media streams

   Lei Zhang, Andreas Thalhammer, Achim Rettinger, and 3 more authors

   Web Semantics, Mar 2017

   For insiders: We enable cross-lingual and cross-media semantic annotation, search, and recommendation over live TV, news, and social media streams.

   For everyone: We build a kind of “universal remote” for information streams, linking TV, news, and social media across languages so topics can be followed beyond language barriers.

   Abstract DOI PDF

   Modern web search engines still exhibit several limitations, including the lack of search term disambiguation, the inability to handle multilingual search terms within a single query, restrictions that require retrieved media items to be in the same language as the query, and limited integration of search results across live streams of heterogeneous media sources such as television, online news, and social media. The system presented in this paper addresses these challenges by combining a media stream processing architecture with cross-lingual and cross-modal semantic annotation, search, and recommendation capabilities. All components of the system were developed within the xLiMe project.

2. Semantic Search for Novel Information

   Michael Färber

   Mar 2017

   For insiders: This thesis develops semantic search methods to detect and extract novel information from text streams, supporting monitoring, knowledge graphs, and timely decision-making.

   For everyone: This thesis is about finding truly new facts in a flood of text, like building a semantic metal detector that spots novel information worth saving.

   Abstract

   This book presents novel approaches for detecting and extracting information from unstructured text documents that is both relevant and novel. A central contribution of the proposed approaches is the semantic modeling of both previously available information and newly extracted content. This semantic representation yields several advantages: it enables the resolution of linguistic ambiguities, allows precise specification of information needs with respect to relevance and novelty, and facilitates the integration of knowledge graphs. Specifically, the book makes three main scientific contributions. First, it assesses the suitability of existing large-scale knowledge graphs—namely DBpedia, Freebase, OpenCyc, Wikidata, and YAGO—for detecting novel information in text. Second, it introduces an approach for identifying emerging entities that are missing from a knowledge graph within streams of text documents. Third, it proposes a method for extracting novel, relevant, and semantically structured statements from text. The developed approaches support applications such as recommending emerging entities and novel statements, populating knowledge graphs, and assisting users who rely on timely novel information, including journalists and technology scouts.

2016

1. XKnowSearch! Exploiting Knowledge Bases for Entity-based Cross-lingual Information Retrieval

   Lei Zhang, Michael Färber, and Achim Rettinger

   In CIKM, Indianapolis, Indiana, USA, Mar 2016

   For insiders: XKnowSearch uses knowledge-base entities to disambiguate queries and retrieve documents across languages for entity-based cross-lingual information retrieval.

   For everyone: We build a cross-language search tool that uses structured background knowledge like a bilingual map, so you can find relevant texts even when the query and documents use different languages.

   Abstract DOI PDF

   In recent years, the amount of entities in large knowledge bases available on the Web has been increasing rapidly, making it possible to propose new ways of intelligent information access. Within the context of globalization, there is a clear need for techniques and systems that can enable multilingual and cross-lingual information access. In this paper, we present XKnowSearch!, a novel entity-based system for multilingual and cross-lingual information retrieval, which supports keyword search and also allows users to influence the search process according to their search intents. By leveraging the multilingual knowledge base on the Web, keyword queries and documents can be represented in their semantic forms, which can facilitate query disambiguation and expansion, and can also overcome the language barrier between queries and documents in different languages.

2. Exploiting Knowledge Bases for Multilingual and Cross-lingual Semantic Annotation and Search

   Lei Zhang, Michael Färber, Andreas Thalhammer, and 2 more authors

   In ISWC, Bethlehem, PA, USA, Mar 2016

   For insiders: We build infrastructure for multilingual semantic annotation and demonstrate cross-lingual semantic search that links text to knowledge base entities.

   For everyone: We build a system that highlights the same people and places across different languages, like a multilingual marker that lets readers spot and connect information even when texts are written in different languages.

   Abstract PDF

   The number of entities contained in large knowledge bases has been growing rapidly, enabling new forms of intelligent information access. At the same time, there is an increasing demand for systems that support multilingual and cross-lingual access to information. In this work, we first present X-LiSA, an infrastructure for multilingual and cross-lingual semantic annotation that provides interfaces for annotating unstructured text in multiple languages using knowledge base resources. Building on X-LiSA, we then demonstrate XKnowSearch!, a novel system for multilingual and cross-lingual semantic search that enables keyword-based retrieval over textual data by leveraging semantic annotations.

2015

1. Using a semantic wiki for technology forecast and technology monitoring

   Michael Färber

   Program, Mar 2015

   For insiders: We extend Semantic MediaWiki with forecasting and monitoring features so organizations can track emerging technologies using structured knowledge and practical workflows.

   For everyone: We turn a semantic wiki into a radar for emerging technologies, so teams can collect and compare new signals instead of losing them in scattered notes.

   Abstract DOI PDF

   The purpose of this paper is to present extensions of Semantic MediaWiki that support technology forecasting and technology monitoring. Based on requirements elicited from potential end users, namely technology experts, new visualization and analysis components were designed and implemented. The usability and applicability of these components are evaluated through task-based user studies that assess their effectiveness in realistic scenarios. The results show that, although semantic wikis are well suited for knowledge management in industrial contexts due to their user-friendly mechanisms for storing and retrieving information, they have so far rarely been applied to technology forecasting and monitoring. This work demonstrates that the additional requirements of these tasks can be fulfilled and that established techniques for technology analysis can be integrated into Semantic MediaWiki. By doing so, a new application area for Semantic MediaWiki is introduced. The presented tools and techniques provide practical value for technology and innovation management and offer open-source alternatives that can be applied in real-world settings where previously expensive dedicated software solutions were required.

2014

1. Kuphi - an Investigation Tool for Searching for and via Semantic Relations

   Michael Färber, Lei Zhang, and Achim Rettinger

   In ESWC, Anissaras, Crete, Greece, Mar 2014

   For insiders: Kuphi enables entity-centric search and relation-based exploration, letting users investigate documents by following semantic relations, not only keywords.

   For everyone: Kuphi helps investigate texts by following who is connected to whom, like pulling a thread from one name to the next until a hidden story appears.

   Abstract DOI PDF

   In this work, we present Kuphi, a novel process-oriented approach to information retrieval designed for investigating entities and their semantic relations within text documents. Kuphi extends traditional bag-of-words–based search by enabling entity-centric exploration. Starting from a keyword search for a specific entity, users can not only retrieve occurrences of that entity in documents, but also search for related entities by specifying semantic relations of interest. This allows users to indirectly identify textual manifestations of relationships between entities. Through cross-lingual semantic annotation, the query language can differ from the language of the underlying documents. We demonstrate the applicability of Kuphi using DBpedia as the knowledge base and news articles collected from RSS feeds.

2. xLiD-Lexica: Cross-lingual Linked Data Lexica

   Lei Zhang, Michael Färber, and Achim Rettinger

   In LREC, Reykjavik, Iceland, Mar 2014

   For insiders: We release cross-lingual linked-data lexica that map surface forms to entities across languages, supporting multilingual annotation and information access.

   For everyone: We build a cross-language lexicon that links words to the same real-world things, like a multilingual dictionary where “Paris” always means the same place.

   Abstract PDF

   In this paper, we introduce xLiD-Lexica, a set of cross-lingual linked data lexica constructed by exploiting the multilingual content of Wikipedia and resources from the Linked Open Data (LOD) cloud. The lexica provide cross-lingual groundings of linked data resources as RDF, enabling seamless integration into existing LOD datasets. In addition, we offer a SPARQL endpoint over xLiD-Lexica that allows users to conveniently access the lexica using the SPARQL query language. The availability of such multilingual and cross-lingual lexica facilitates cross-lingual information access, for example by enabling the mapping of natural language expressions in different languages to linked data resources. A wide range of natural language processing tasks, including natural language generation, cross-lingual entity linking, text annotation, and question answering, can benefit from xLiD-Lexica.

3. Exploiting Semantic Annotations for Entity-based Information Retrieval

   Lei Zhang, Michael Färber, Thanh Tran, and 1 more author

   In ISWC, Riva del Garda, Italy, Mar 2014

   For insiders: We show how semantic annotations support entity-based information retrieval and interactive query refinement to reduce ambiguity and improve relevance.

   For everyone: We show that tagging text with real entities improves search by meaning, like swapping fuzzy keyword hunting for clear labels you can click and refine.

   Abstract PDF

   In this paper, we propose a novel approach to entity-based information retrieval that exploits semantic annotations of documents. With the growing availability of structured knowledge bases and semantic annotation techniques, documents and queries can be represented at a semantic level, thereby reducing term ambiguity and bridging language differences between queries and documents. Based on multiple semantic interpretations, users can iteratively refine their queries to better reflect their information needs. By leveraging the semantics of entities and their relations encoded in knowledge bases, we introduce a new ranking scheme designed to more effectively satisfy users’ information needs.

2013

1. A Comparative Evaluation of Cross-Lingual Text Annotation Techniques

   Lei Zhang, Achim Rettinger, Michael Färber, and 1 more author

   In CLEF, Mar 2013

   For insiders: We compare cross-lingual text annotation techniques and introduce an evaluation framework for systematic multilingual knowledge extraction.

   For everyone: We compare ways to connect texts with background knowledge across languages, like labeling the same story in different tongues, and we set up a fair “test track” to compare systems.

   Abstract DOI PDF

   In this paper, we address the problem of extracting knowledge from multilingual textual documents by annotating text using a cross-lingual knowledge base, namely Wikipedia. Our contribution is twofold. First, we introduce a novel evaluation framework for cross-lingual text annotation techniques that is based on annotating a parallel corpus with respect to a hub language in a cross-lingual knowledge base. Second, we analyze the performance of different cross-lingual text annotation methods within this framework. We conduct an empirical comparison of three approaches: Cross-lingual Named Entity Annotation (CL-NEA), Cross-lingual Wikifier Annotation (CL-WIFI), and Cross-lingual Explicit Semantic Analysis (CL-ESA). The results highlight the strengths and limitations of the investigated approaches and demonstrate the usefulness of the proposed evaluation framework for systematic comparison.

2. Ontology-Supported Document Ranking for Novelty Search

   Michael Färber

   In ESWC, Montpellier, France, Mar 2013

   For insiders: We rank documents for ontology-supported novelty search to surface novel, relevant information for knowledge-base population and monitoring.

   For everyone: We use a knowledge-based compass to rank documents by what is genuinely new, like a treasure hunter looking for fresh finds rather than repeated facts.

   Abstract DOI PDF

   Within specific domains, users often face the challenge of populating ontologies according to their particular needs. This challenge is especially pronounced in scenarios involving novelty detection and forecasting, where users aim to integrate new information from unstructured text documents into their ontologies for subsequent knowledge-based analysis. In this paper, we propose a semantic document ranking approach that serves as a prerequisite for ontology population. By leveraging the underlying ontology for both query generation and document ranking, the approach introduces semantic structure into the retrieval process and promises improved ranking performance with respect to relevance and novelty compared to non-semantic methods.

3. A Semantic Wiki for Novelty Search on Documents

   Michael Färber, and Achim Rettinger

   In BIR, Delft, The Netherlands, Mar 2013

   For insiders: We present a semantic-wiki workflow for novelty search that continuously ingests and structures new web information for ongoing monitoring tasks.

   For everyone: We build a semantic wiki as a watchtower that keeps pulling fresh web information into a structured store, so new signals do not get lost in the noise.

   Abstract PDF

   Technology-oriented companies are typically interested in continuously monitoring developments related to their core technologies. However, many organizations, particularly small and medium-sized enterprises, lack efficient and systematic processes for this purpose. Existing efforts are often limited to uncoordinated keyword-based searches over web resources. In this paper, we present a semi-automatic approach that enables the structured and continuous detection of relevant, novel, and domain-specific documents on the Web. Our system is built on a semantic wiki that allows domain experts to (i) store relevant information in a structured knowledge base supporting monitoring and trend mining, and (ii) continuously import newly detected items, such as emerging technologies and their properties, into the knowledge base. Novel items are identified by generating structured queries based on the user’s contextual knowledge and by representing retrieved documents as semantic graphs. This approach facilitates the discovery of novel information in a more effective and semi-automatic manner.