Chris Biemann

Debayan Banerjee, Pranav Ajit Nair, Jivat Neet Kaur et al.

In this work, we focus on the task of generating SPARQL queries from natural language questions, which can then be executed on Knowledge Graphs (KGs). We assume that gold entity and relations have been provided, and the remaining task is to arrange them in the right order along with SPARQL vocabulary, and input tokens to produce the correct SPARQL query. Pre-trained Language Models (PLMs) have not been explored in depth on this task so far, so we experiment with BART, T5 and PGNs (Pointer Generator Networks) with BERT embeddings, looking for new baselines in the PLM era for this task, on DBpedia and Wikidata KGs. We show that T5 requires special input tokenisation, but produces state of the art performance on LC-QuAD 1.0 and LC-QuAD 2.0 datasets, and outperforms task-specific models from previous works. Moreover, the methods enable semantic parsing for questions where a part of the input needs to be copied to the output query, thus enabling a new paradigm in KG semantic parsing.

Debayan Banerjee, Sushil Awale, Ricardo Usbeck et al.

In this work we create a question answering dataset over the DBLP scholarly knowledge graph (KG). DBLP is an on-line reference for bibliographic information on major computer science publications that indexes over 4.4 million publications published by more than 2.2 million authors. Our dataset consists of 10,000 question answer pairs with the corresponding SPARQL queries which can be executed over the DBLP KG to fetch the correct answer. DBLP-QuAD is the largest scholarly question answering dataset.

Debayan Banerjee, Arefa, Ricardo Usbeck et al.

In this work, we present a web application named DBLPLink, which performs entity linking over the DBLP scholarly knowledge graph. DBLPLink uses text-to-text pre-trained language models, such as T5, to produce entity label spans from an input text question. Entity candidates are fetched from a database based on the labels, and an entity re-ranker sorts them based on entity embeddings, such as TransE, DistMult and ComplEx. The results are displayed so that users may compare and contrast the results between T5-small, T5-base and the different KG embeddings used. The demo can be accessed at https://ltdemos.informatik.uni-hamburg.de/dblplink/.

Debayan Banerjee, Pranav Ajit Nair, Ricardo Usbeck et al.

In this work, we present an end-to-end Knowledge Graph Question Answering (KGQA) system named GETT-QA. GETT-QA uses T5, a popular text-to-text pre-trained language model. The model takes a question in natural language as input and produces a simpler form of the intended SPARQL query. In the simpler form, the model does not directly produce entity and relation IDs. Instead, it produces corresponding entity and relation labels. The labels are grounded to KG entity and relation IDs in a subsequent step. To further improve the results, we instruct the model to produce a truncated version of the KG embedding for each entity. The truncated KG embedding enables a finer search for disambiguation purposes. We find that T5 is able to learn the truncated KG embeddings without any change of loss function, improving KGQA performance. As a result, we report strong results for LC-QuAD 2.0 and SimpleQuestions-Wikidata datasets on end-to-end KGQA over Wikidata.

Usman Naseem, Robert Geislinger, Juan Ren et al.

We present SemEval-2026 Task 9, a shared task on online polarization detection, covering 22 languages and comprising over 110K annotated instances. Each data instance is multi-labeled with the presence of polarization, polarization type, and polarization manifestation. Participants were asked to predict labels in three sub-tasks: (1) detecting the presence of polarization, (2) identifying the type of polarization, and (3) recognizing the polarization manifestation. The three tasks attracted over 1,000 participants worldwide and more than 10k submission on Codabench. We received final submissions from 67 teams and 73 system description papers. We report the baseline results and analyze the performance of the best-performing systems, highlighting the most common approaches and the most effective methods across different subtasks and languages. The dataset of this task is publicly available.

Christian Haase, Saba Anwar, Seid Muhie Yimam et al.

We present Sense Clustering over Time (SCoT), a novel network-based tool for analysing lexical change. SCoT represents the meanings of a word as clusters of similar words. It visualises their formation, change, and demise. There are two main approaches to the exploration of dynamic networks: the discrete one compares a series of clustered graphs from separate points in time. The continuous one analyses the changes of one dynamic network over a time-span. SCoT offers a new hybrid solution. First, it aggregates time-stamped documents into intervals and calculates one sense graph per discrete interval. Then, it merges the static graphs to a new type of dynamic semantic neighbourhood graph over time. The resulting sense clusters offer uniquely detailed insights into lexical change over continuous intervals with model transparency and provenance. SCoT has been successfully used in a European study on the changing meaning of `crisis'.

Debayan Banerjee, Mathis Poser, Christina Wiethof et al.

AI enabled chat bots have recently been put to use to answer customer service queries, however it is a common feedback of users that bots lack a personal touch and are often unable to understand the real intent of the user's question. To this end, it is desirable to have human involvement in the customer servicing process. In this work, we present a system where a human support agent collaborates in real-time with an AI agent to satisfactorily answer customer queries. We describe the user interaction elements of the solution, along with the machine learning techniques involved in the AI agent.

Xintong Wang, Xiaoyu Li, Xingshan Li et al.

Large Language Models (LLMs) have emerged as dominant foundational models in modern NLP. However, the understanding of their prediction processes and internal mechanisms, such as feed-forward networks (FFN) and multi-head self-attention (MHSA), remains largely unexplored. In this work, we probe LLMs from a human behavioral perspective, correlating values from LLMs with eye-tracking measures, which are widely recognized as meaningful indicators of human reading patterns. Our findings reveal that LLMs exhibit a similar prediction pattern with humans but distinct from that of Shallow Language Models (SLMs). Moreover, with the escalation of LLM layers from the middle layers, the correlation coefficients also increase in FFN and MHSA, indicating that the logits within FFN increasingly encapsulate word semantics suitable for predicting tokens from the vocabulary.

Debayan Banerjee, Seid Muhie Yimam, Sushil Awale et al.

In this work, we present ARDIAS, a web-based application that aims to provide researchers with a full suite of discovery and collaboration tools. ARDIAS currently allows searching for authors and articles by name and gaining insights into the research topics of a particular researcher. With the aid of AI-based tools, ARDIAS aims to recommend potential collaborators and topics to researchers. In the near future, we aim to add tools that allow researchers to communicate with each other and start new projects.

Narges Baba Ahmadi, Jan Strich, Martin Semmann et al.

Large language models (LLMs) are increasingly used to access legal information. Yet, their deployment in multilingual legal settings is constrained by unreliable retrieval and the lack of domain-adapted, open-embedding models. In particular, existing multilingual legal corpora are not designed for semantic retrieval, and PDF-based legislative sources introduce substantial noise due to imperfect text extraction. To address these challenges, we introduce LEMUR, a large-scale multilingual corpus of EU environmental legislation constructed from 24,953 official EUR-Lex PDF documents covering 25 languages. We quantify the fidelity of PDF-to-text conversion by measuring lexical consistency against authoritative HTML versions using the Lexical Content Score (LCS). Building on LEMUR, we fine-tune three state-of-the-art multilingual embedding models using contrastive objectives in both monolingual and bilingual settings, reflecting realistic legal-retrieval scenarios. Experiments across low- and high-resource languages demonstrate that legal-domain fine-tuning consistently improves Top-k retrieval accuracy relative to strong baselines, with particularly pronounced gains for low-resource languages. Cross-lingual evaluations show that these improvements transfer to unseen languages, indicating that fine-tuning primarily enhances language-independent, content-level legal representations rather than language-specific cues. We publish code\footnote{\href{https://github.com/nargesbh/eur_lex}{GitHub Repository}} and data\footnote{\href{https://huggingface.co/datasets/G4KMU/LEMUR}{Hugging Face Dataset}}.

Klejda Alushi, Jan Strich, Chris Biemann et al.

Conversational question answering increasingly relies on retrieval-augmented generation (RAG) to ground large language models (LLMs) in external knowledge. Yet, most existing studies evaluate RAG methods in isolation and primarily focus on single-turn settings. This paper addresses the lack of a systematic comparison of RAG methods for multi-turn conversational QA, where dialogue history, coreference, and shifting user intent substantially complicate retrieval. We present a comprehensive empirical study of vanilla and advanced RAG methods across eight diverse conversational QA datasets spanning multiple domains. Using a unified experimental setup, we evaluate retrieval quality and answer generation using generator and retrieval metrics, and analyze how performance evolves across conversation turns. Our results show that robust yet straightforward methods, such as reranking, hybrid BM25, and HyDE, consistently outperform vanilla RAG. In contrast, several advanced techniques fail to yield gains and can even degrade performance below the No-RAG baseline. We further demonstrate that dataset characteristics and dialogue length strongly influence retrieval effectiveness, explaining why no single RAG strategy dominates across settings. Overall, our findings indicate that effective conversational RAG depends less on method complexity than on alignment between the retrieval strategy and the dataset structure. We publish the code used.\footnote{\href{https://github.com/Klejda-A/exp-rag.git}{GitHub Repository}}

Anshul Singh, Chris Biemann, Jan Strich

Vision-Language Models (VLMs) have demonstrated remarkable capabilities in interpreting visual layouts and text. However, a significant challenge remains in their ability to interpret robustly and reason over multi-tabular data presented as images, a common occurrence in real-world scenarios like web pages and digital documents. Existing benchmarks typically address single tables or non-visual data (text/structured). This leaves a critical gap: they don't assess the ability to parse diverse table images, correlate information across them, and perform multi-hop reasoning on the combined visual data. We introduce MTabVQA, a novel benchmark specifically designed for multi-tabular visual question answering to bridge that gap. MTabVQA comprises 3,745 complex question-answer pairs that necessitate multi-hop reasoning across several visually rendered table images. We provide extensive benchmark results for state-of-the-art VLMs on MTabVQA, revealing significant performance limitations. We further investigate post-training techniques to enhance these reasoning abilities and release MTabVQA-Instruct, a large-scale instruction-tuning dataset. Our experiments show that fine-tuning VLMs with MTabVQA-Instruct substantially improves their performance on visual multi-tabular reasoning. Code and dataset (https://huggingface.co/datasets/mtabvqa/MTabVQA-Eval) are available online (https://anonymous.4open.science/r/MTabVQA-EMNLP-B16E).

Xintong Wang, Jingheng Pan, Liang Ding et al.

Large Vision-Language Models (LVLMs) are increasingly adept at generating contextually detailed and coherent responses from visual inputs. However, their application in multimodal decision-making and open-ended generation is hindered by a notable rate of hallucinations, where generated text inaccurately represents the visual contents. To address this issue, this paper introduces the Instruction Contrastive Decoding (ICD) method, a novel approach designed to reduce hallucinations during LVLM inference. Our method is inspired by our observation that what we call disturbance instructions significantly exacerbate hallucinations in multimodal fusion modules. ICD contrasts distributions from standard and instruction disturbance, thereby increasing alignment uncertainty and effectively subtracting hallucinated concepts from the original distribution. Through comprehensive experiments on discriminative benchmarks (POPE and MME) and a generative benchmark (LLaVa-Bench), we demonstrate that ICD significantly mitigates both object-level and attribute-level hallucinations. Moreover, our method not only addresses hallucinations but also significantly enhances the general perception and recognition capabilities of LVLMs.

Thennal D K, Chris Biemann, Hans Ole Hatzel

Label projection is an effective technique for cross-lingual transfer, extending span-annotated datasets from a high-resource language to low-resource ones. Most approaches perform label projection as a separate step after machine translation, and prior work that combines the two reports degraded translation quality. We re-evaluate this claim with LabelPigeon, a novel framework that jointly performs translation and label projection via XML tags. We design a direct evaluation scheme for label projection, and find that LabelPigeon outperforms baselines and actively improves translation quality in 11 languages. We further assess translation quality across 203 languages and varying annotation complexity, finding consistent improvement attributed to additional fine-tuning. Finally, across 27 languages and three downstream tasks, we report substantial gains in cross-lingual transfer over comparable work, up to +39.9 F1 on NER. Overall, our results demonstrate that XML-tagged label projection provides effective and efficient label transfer without compromising translation quality.

Xintong Wang, Jingheng Pan, Yixiao Liu et al.

Vision-Language Translation (VLT) is a challenging task that requires accurately recognizing multilingual text embedded in images and translating it into the target language with the support of visual context. While recent Large Vision-Language Models (LVLMs) have demonstrated strong multilingual and visual understanding capabilities, there is a lack of systematic evaluation and understanding of their performance on VLT. In this work, we present a comprehensive study of VLT from three key perspectives: data quality, model architecture, and evaluation metrics. (1) We identify critical limitations in existing datasets, particularly in semantic and cultural fidelity, and introduce AibTrans -- a multilingual, parallel, human-verified dataset with OCR-corrected annotations. (2) We benchmark 11 commercial LVLMs/LLMs and 6 state-of-the-art open-source models across end-to-end and cascaded architectures, revealing their OCR dependency and contrasting generation versus reasoning behaviors. (3) We propose Density-Aware Evaluation to address metric reliability issues under varying contextual complexity, introducing the DA Score as a more robust measure of translation quality. Building upon these findings, we establish a new evaluation benchmark for VLT. Notably, we observe that fine-tuning LVLMs on high-resource language pairs degrades cross-lingual performance, and we propose a balanced multilingual fine-tuning strategy that effectively adapts LVLMs to VLT without sacrificing their generalization ability.

Xintong Wang, Yixiao Liu, Jingheng Pan et al.

Detoxifying offensive language while preserving the speaker's original intent is a challenging yet critical goal for improving the quality of online interactions. Although large language models (LLMs) show promise in rewriting toxic content, they often default to overly polite rewrites, distorting the emotional tone and communicative intent. This problem is especially acute in Chinese, where toxicity often arises implicitly through emojis, homophones, or discourse context. We present ToxiRewriteCN, the first Chinese detoxification dataset explicitly designed to preserve sentiment polarity. The dataset comprises 1,556 carefully annotated triplets, each containing a toxic sentence, a sentiment-aligned non-toxic rewrite, and labeled toxic spans. It covers five real-world scenarios: standard expressions, emoji-induced and homophonic toxicity, as well as single-turn and multi-turn dialogues. We evaluate 17 LLMs, including commercial and open-source models with variant architectures, across four dimensions: detoxification accuracy, fluency, content preservation, and sentiment polarity. Results show that while commercial and MoE models perform best overall, all models struggle to balance safety with emotional fidelity in more subtle or context-heavy settings such as emoji, homophone, and dialogue-based inputs. We release ToxiRewriteCN to support future research on controllable, sentiment-aware detoxification for Chinese.

Viktor Moskvoretskii, Nazarii Tupitsa, Chris Biemann et al.

We present a new approach called MeritOpt based on the Personalized Federated Learning algorithm MeritFed that can be applied to Natural Language Tasks with heterogeneous data. We evaluate it on the Low-Resource Machine Translation task, using the datasets of South East Asian and Finno-Ugric languages. In addition to its effectiveness, MeritOpt is also highly interpretable, as it can be applied to track the impact of each language used for training. Our analysis reveals that target dataset size affects weight distribution across auxiliary languages, that unrelated languages do not interfere with the training, and auxiliary optimizer parameters have minimal impact. Our approach is easy to apply with a few lines of code, and we provide scripts for reproducing the experiments at https://github.com/VityaVitalich/MeritOpt.

Binny Mathew, Punyajoy Saha, Seid Muhie Yimam et al.

Hate speech is a challenging issue plaguing the online social media. While better models for hate speech detection are continuously being developed, there is little research on the bias and interpretability aspects of hate speech. In this paper, we introduce HateXplain, the first benchmark hate speech dataset covering multiple aspects of the issue. Each post in our dataset is annotated from three different perspectives: the basic, commonly used 3-class classification (i.e., hate, offensive or normal), the target community (i.e., the community that has been the victim of hate speech/offensive speech in the post), and the rationales, i.e., the portions of the post on which their labelling decision (as hate, offensive or normal) is based. We utilize existing state-of-the-art models and observe that even models that perform very well in classification do not score high on explainability metrics like model plausibility and faithfulness. We also observe that models, which utilize the human rationales for training, perform better in reducing unintended bias towards target communities. We have made our code and dataset public at https://github.com/punyajoy/HateXplain

Gregor Wiedemann, Seid Muhie Yimam, Chris Biemann

Investigative journalism in recent years is confronted with two major challenges: 1) vast amounts of unstructured data originating from large text collections such as leaks or answers to Freedom of Information requests, and 2) multi-lingual data due to intensified global cooperation and communication in politics, business and civil society. Faced with these challenges, journalists are increasingly cooperating in international networks. To support such collaborations, we present the new version of new/s/leak 2.0, our open-source software for content-based searching of leaks. It includes three novel main features: 1) automatic language detection and language-dependent information extraction for 40 languages, 2) entity and keyword visualization for efficient exploration, and 3) decentral deployment for analysis of confidential data from various formats. We illustrate the new analysis capabilities with an exemplary case study.

Benjamin Milde, Chris Biemann

We introduce "Unspeech" embeddings, which are based on unsupervised learning of context feature representations for spoken language. The embeddings were trained on up to 9500 hours of crawled English speech data without transcriptions or speaker information, by using a straightforward learning objective based on context and non-context discrimination with negative sampling. We use a Siamese convolutional neural network architecture to train Unspeech embeddings and evaluate them on speaker comparison, utterance clustering and as a context feature in TDNN-HMM acoustic models trained on TED-LIUM, comparing it to i-vector baselines. Particularly decoding out-of-domain speech data from the recently released Common Voice corpus shows consistent WER reductions. We release our source code and pre-trained Unspeech models under a permissive open source license.

Mohamed Rissal Hedna, Jan Strich, Martin Semmann et al.

Large language models (LLMs) are systematically overconfident: they routinely express high certainty on questions they often answer incorrectly. Existing calibration methods either require labeled validation data, degrade under distribution shifts, or incur substantial inference costs. Recent work has shown that LLMs already contain a better-calibrated signal than the one they verbalize: the token probability of "True" when the model is asked "Is this answer correct?" ($P(\text{True})$) consistently outperforms their stated confidence, a gap that is theoretically grounded as generative error is lower-bounded by roughly twice the corresponding discriminative error. We introduce $\textbf{SECL}$ ($\textbf{SE}$lf-$\textbf{C}$alibrating $\textbf{L}$anguage Models), a test-time training (TTT) pipeline that exploits this gap as label-free self-supervision, requiring no labeled data or human supervision. SECL adapts only when the input distribution shifts, training on just 6--26% of the question stream at lower cost than the baseline it distills from. Across four small language models from three model families and four diverse domains, SECL reduces Expected Calibration Error (ECE) by 56--78%, outperforming its own supervision signal and matching or outperforming recent inference-time methods. SECL is the first method to apply TTT to calibration; seven ablations covering signal quality, gating strategy, weight accumulation, loss design, domain ordering, hyperparameter sensitivity, and layer selection confirm that each component is crucial and robust across configurations. Code: https://anonymous.4open.science/r/secl-emnlp26-submission-C890

Nedjma Ousidhoum, Shamsuddeen Hassan Muhammad, Mohamed Abdalla et al.

Exploring and quantifying semantic relatedness is central to representing language and holds significant implications across various NLP tasks. While earlier NLP research primarily focused on semantic similarity, often within the English language context, we instead investigate the broader phenomenon of semantic relatedness. In this paper, we present \textit{SemRel}, a new semantic relatedness dataset collection annotated by native speakers across 13 languages: \textit{Afrikaans, Algerian Arabic, Amharic, English, Hausa, Hindi, Indonesian, Kinyarwanda, Marathi, Moroccan Arabic, Modern Standard Arabic, Spanish,} and \textit{Telugu}. These languages originate from five distinct language families and are predominantly spoken in Africa and Asia -- regions characterised by a relatively limited availability of NLP resources. Each instance in the SemRel datasets is a sentence pair associated with a score that represents the degree of semantic textual relatedness between the two sentences. The scores are obtained using a comparative annotation framework. We describe the data collection and annotation processes, challenges when building the datasets, baseline experiments, and their impact and utility in NLP.

Tim Fischer, Chris Biemann

This paper introduces Perspectives, an interactive extension of the Discourse Analysis Tool Suite designed to empower Digital Humanities (DH) scholars to explore and organize large, unstructured document collections. Perspectives implements a flexible, aspect-focused document clustering pipeline with human-in-the-loop refinement capabilities. We showcase how this process can be initially steered by defining analytical lenses through document rewriting prompts and instruction-based embeddings, and further aligned with user intent through tools for refining clusters and mechanisms for fine-tuning the embedding model. The demonstration highlights a typical workflow, illustrating how DH researchers can leverage Perspectives's interactive document map to uncover topics, sentiments, or other relevant categories, thereby gaining insights and preparing their data for subsequent in-depth analysis.

Jan Strich, Adeline Scharfenberg, Chris Biemann et al.

We introduce EncouRAGe, a comprehensive Python framework designed to streamline the development and evaluation of Retrieval-Augmented Generation (RAG) systems using Large Language Models (LLMs) and Embedding Models. EncouRAGe comprises five modular and extensible components: Type Manifest, RAG Factory, Inference, Vector Store, and Metrics, facilitating flexible experimentation and extensible development. The framework emphasizes scientific reproducibility, diverse evaluation metrics, and local deployment, enabling researchers to efficiently assess datasets within RAG workflows. This paper presents implementation details and an extensive evaluation across multiple benchmark datasets, including 25k QA pairs and over 51k documents. Our results show that RAG still underperforms compared to the Oracle Context, while Hybrid BM25 consistently achieves the best results across all four datasets. We further examine the effects of reranking, observing only marginal performance improvements accompanied by higher response latency.

Punyajoy Saha, Aalok Agrawal, Abhik Jana et al.

With the emergence of numerous Large Language Models (LLM), the usage of such models in various Natural Language Processing (NLP) applications is increasing extensively. Counterspeech generation is one such key task where efforts are made to develop generative models by fine-tuning LLMs with hatespeech - counterspeech pairs, but none of these attempts explores the intrinsic properties of large language models in zero-shot settings. In this work, we present a comprehensive analysis of the performances of four LLMs namely GPT-2, DialoGPT, ChatGPT and FlanT5 in zero-shot settings for counterspeech generation, which is the first of its kind. For GPT-2 and DialoGPT, we further investigate the deviation in performance with respect to the sizes (small, medium, large) of the models. On the other hand, we propose three different prompting strategies for generating different types of counterspeech and analyse the impact of such strategies on the performance of the models. Our analysis shows that there is an improvement in generation quality for two datasets (17%), however the toxicity increase (25%) with increase in model size. Considering type of model, GPT-2 and FlanT5 models are significantly better in terms of counterspeech quality but also have high toxicity as compared to DialoGPT. ChatGPT are much better at generating counter speech than other models across all metrics. In terms of prompting, we find that our proposed strategies help in improving counter speech generation across all the models.

Jingheng Pan, Xintong Wang, Longyue Wang et al.

Ambiguity resolution is a key challenge in multimodal machine translation (MMT), where models must genuinely leverage visual input to map an ambiguous expression to its intended meaning. Although prior work has proposed disambiguation-oriented benchmarks that provide supportive evidence for the role of vision, we observe substantial issues in data quality and a mismatch with translation scenarios. Moreover, existing ambiguity-oriented evaluations are not well suited to broader ambiguity types in open-ended translation. To address these limitations, we present VIDA (Visually-Dependent Ambiguity), a dataset of 2,500 carefully curated instances in which resolving an annotated ambiguous source span requires visual evidence. We further propose Disambiguation-Centric Metrics that use an LLM-as-a-judge classifier to verify whether annotated ambiguous expressions are resolved correctly at the span level. Experiments with two state-of-the-art Large Vision Language Models under vanilla inference, supervised fine-tuning (SFT), and our chain-of-thought SFT (CoT-SFT) show that while SFT improves overall translation quality, CoT-SFT yields more consistent gains in disambiguation accuracy, especially on out-of-distribution subsets, indicating a stronger generalization for resolving diverse ambiguity types.

Hans Ole Hatzel, Ekaterina Artemova, Haimo Paul Stiemer et al.

We present the shared task on narrative similarity and narrative representation learning - NSNRL (pronounced "nass-na-rel"). The task operationalizes narrative similarity as a binary classification problem: determining which of two stories is more similar to an anchor story. We introduce a novel definition of narrative similarity, compatible with both narrative theory and intuitive judgment. Based on the similarity judgments collected under this concept, we also evaluate narrative embedding representations. We collected at least two annotations each for more than 1,000 story summary triples, with each annotation being backed by at least two annotators in agreement. This paper describes the sampling and annotation process for the dataset; further, we give an overview of the submitted systems and the techniques they employ. We received a total of 71 final submissions from 46 teams across our two tracks. In our triple-based classification setup, LLM ensembles make up many of the top-scoring systems, while in the embedding setup, systems with pre- and post-processing on pretrained embedding models perform about on par with custom fine-tuned solutions. Our analysis identifies potential headroom for improvement of automated systems in both tracks. The task website includes visualizations of embeddings alongside instance-level classification results for all teams.

Florian Schneider, Carolin Holtermann, Chris Biemann et al.

Large Vision-Language Models (LVLMs) have recently gained attention due to their distinctive performance and broad applicability. While it has been previously shown that their efficacy in usage scenarios involving non-Western contexts falls short, existing studies are limited in scope, covering just a narrow range of cultures, focusing exclusively on a small number of cultural aspects, or evaluating a limited selection of models on a single task only. Towards globally inclusive LVLM research, we introduce GIMMICK, an extensive multimodal benchmark designed to assess a broad spectrum of cultural knowledge across 144 countries representing six global macro-regions. GIMMICK comprises six tasks built upon three new datasets that span 728 unique cultural events or facets on which we evaluated 20 LVLMs and 11 LLMs, including five proprietary and 26 open-weight models of all sizes. We systematically examine (1) regional cultural biases, (2) the influence of model size, (3) input modalities, and (4) external cues. Our analyses reveal strong biases toward Western cultures across models and tasks and highlight strong correlations between model size and performance, as well as the effectiveness of multimodal input and external geographic cues. We further find that models have more knowledge of tangible than intangible aspects (e.g., food vs. rituals) and that they excel in recognizing broad cultural origins but struggle with a more nuanced understanding.

Florian Schneider, Narges Baba Ahmadi, Niloufar Baba Ahmadi et al.

In this paper, we introduce CollEx, an innovative multimodal agentic Retrieval-Augmented Generation (RAG) system designed to enhance interactive exploration of extensive scientific collections. Given the overwhelming volume and inherent complexity of scientific collections, conventional search systems often lack necessary intuitiveness and interactivity, presenting substantial barriers for learners, educators, and researchers. CollEx addresses these limitations by employing state-of-the-art Large Vision-Language Models (LVLMs) as multimodal agents accessible through an intuitive chat interface. By abstracting complex interactions via specialized agents equipped with advanced tools, CollEx facilitates curiosity-driven exploration, significantly simplifying access to diverse scientific collections and records therein. Our system integrates textual and visual modalities, supporting educational scenarios that are helpful for teachers, pupils, students, and researchers by fostering independent exploration as well as scientific excitement and curiosity. Furthermore, CollEx serves the research community by discovering interdisciplinary connections and complementing visual data. We illustrate the effectiveness of our system through a proof-of-concept application containing over 64,000 unique records across 32 collections from a local scientific collection from a public university.

Abinew Ali Ayele, Esubalew Alemneh Jalew, Adem Chanie Ali et al.

The prevalence of digital media and evolving sociopolitical dynamics have significantly amplified the dissemination of hateful content. Existing studies mainly focus on classifying texts into binary categories, often overlooking the continuous spectrum of offensiveness and hatefulness inherent in the text. In this research, we present an extensive benchmark dataset for Amharic, comprising 8,258 tweets annotated for three distinct tasks: category classification, identification of hate targets, and rating offensiveness and hatefulness intensities. Our study highlights that a considerable majority of tweets belong to the less offensive and less hate intensity levels, underscoring the need for early interventions by stakeholders. The prevalence of ethnic and political hatred targets, with significant overlaps in our dataset, emphasizes the complex relationships within Ethiopia's sociopolitical landscape. We build classification and regression models and investigate the efficacy of models in handling these tasks. Our results reveal that hate and offensive speech can not be addressed by a simplistic binary classification, instead manifesting as variables across a continuous range of values. The Afro-XLMR-large model exhibits the best performances achieving F1-scores of 75.30%, 70.59%, and 29.42% for the category, target, and regression tasks, respectively. The 80.22% correlation coefficient of the Afro-XLMR-large model indicates strong alignments.

Thennal D K, Tim Fischer, Chris Biemann

Large language models (LLMs) demonstrate strong performance as text embedding models when finetuned with supervised contrastive training. However, their large size balloons inference time and memory requirements. In this paper, we show that by pruning the last $p\%$ layers of an LLM before supervised training for only 1000 steps, we can achieve a proportional reduction in memory and inference time. We evaluate four different state-of-the-art LLMs on text embedding tasks and find that our method can prune up to 30\% of layers with negligible impact on performance and up to 80\% with only a modest drop. With only three lines of code, our method is easily implemented in any pipeline for transforming LLMs to text encoders. We also propose $\text{L}^3 \text{Prune}$, a novel layer-pruning strategy based on the model's initial loss that provides two optimal pruning configurations: a large variant with negligible performance loss and a small variant for resource-constrained settings. On average, the large variant prunes 21\% of the parameters with a $-0.3$ performance drop, and the small variant only suffers from a $-5.1$ decrease while pruning 74\% of the model. We consider these results strong evidence that LLMs are overparameterized for text embedding tasks, and can be easily pruned.

Viktor Moskvoretskii, Chris Biemann, Irina Nikishina

Although large language models (LLMs) have achieved remarkable performance across various tasks, they remain prone to errors. A key challenge is enabling them to self-correct. While prior research has relied on external tools or large proprietary models, this work explores self-correction in small language models (SLMs) through iterative fine-tuning using solely self-generated data. We introduce the Self-Taught Self-Correction (STaSC) algorithm, which incorporates multiple algorithmic design choices. Experimental results on a question-answering task demonstrate that STaSC effectively learns self-correction, leading to significant performance improvements. Our analysis further provides insights into the mechanisms of self-correction and the impact of different design choices on learning dynamics and overall performance. To support future research, we release our user-friendly codebase and lightweight models.

Fynn Petersen-Frey, Chris Biemann

Extracting who says what to whom is a crucial part in analyzing human communication in today's abundance of data such as online news articles. Yet, the lack of annotated data for this task in German news articles severely limits the quality and usability of possible systems. To remedy this, we present a new, freely available, creative-commons-licensed dataset for quotation attribution in German news articles based on WIKINEWS. The dataset provides curated, high-quality annotations across 1000 documents (250,000 tokens) in a fine-grained annotation schema enabling various downstream uses for the dataset. The annotations not only specify who said what but also how, in which context, to whom and define the type of quotation. We specify our annotation schema, describe the creation of the dataset and provide a quantitative analysis. Further, we describe suitable evaluation metrics, apply two existing systems for quotation attribution, discuss their results to evaluate the utility of our dataset and outline use cases of our dataset in downstream tasks.

Naquee Rizwan, Seid Muhie Yimam, Daryna Dementieva et al.

Despite regulations imposed by nations and social media platforms, e.g. (Government of India, 2021; European Parliament and Council of the European Union, 2022), inter alia, hateful content persists as a significant challenge. Existing approaches primarily rely on reactive measures such as blocking or suspending offensive messages, with emerging strategies focusing on proactive measurements like detoxification and counterspeech. In our work, which we call HatePRISM, we conduct a comprehensive examination of hate speech regulations and strategies from three perspectives: country regulations, social platform policies, and NLP research datasets. Our findings reveal significant inconsistencies in hate speech definitions and moderation practices across jurisdictions and platforms, alongside a lack of alignment with research efforts. Based on these insights, we suggest ideas and research direction for further exploration of a unified framework for automated hate speech moderation incorporating diverse strategies.

Gregor Geigle, Florian Schneider, Carolin Holtermann et al.

Most Large Vision-Language Models (LVLMs) to date are trained predominantly on English data, which makes them struggle to understand non-English input and fail to generate output in the desired target language. Existing efforts mitigate these issues by adding multilingual training data, but do so in a largely ad-hoc manner, lacking insight into how different training mixes tip the scale for different groups of languages. In this work, we present a comprehensive investigation into the training strategies for massively multilingual LVLMs. First, we conduct a series of multi-stage experiments spanning 13 downstream vision-language tasks and 43 languages, systematically examining: (1) the number of training languages that can be included without degrading English performance and (2) optimal language distributions of pre-training as well as (3) instruction-tuning data. Further, we (4) investigate how to improve multilingual text-in-image understanding, and introduce a new benchmark for the task. Surprisingly, our analysis reveals that one can (i) include as many as 100 training languages simultaneously (ii) with as little as 25-50\% of non-English data, to greatly improve multilingual performance while retaining strong English performance. We further find that (iii) including non-English OCR data in pre-training and instruction-tuning is paramount for improving multilingual text-in-image understanding. Finally, we put all our findings together and train Centurio, a 100-language LVLM, offering state-of-the-art performance in an evaluation covering 14 tasks and 56 languages.

Fabian David Schmidt, Florian Schneider, Chris Biemann et al.

Existing multilingual vision-language (VL) benchmarks often only cover a handful of languages. Consequently, evaluations of large vision-language models (LVLMs) predominantly target high-resource languages, underscoring the need for evaluation data for low-resource languages. To address this limitation, we introduce MVL-SIB, a massively multilingual vision-language benchmark that evaluates both cross-modal and text-only topical matching across 205 languages -- over 100 more than the most multilingual existing VL benchmarks encompass. We then benchmark a range of of open-weight LVLMs together with GPT-4o(-mini) on MVL-SIB. Our results reveal that LVLMs struggle in cross-modal topic matching in lower-resource languages, performing no better than chance on languages like N'Koo. Our analysis further reveals that VL support in LVLMs declines disproportionately relative to textual support for lower-resource languages, as evidenced by comparison of cross-modal and text-only topical matching performance. We further observe that open-weight LVLMs do not benefit from representing a topic with more than one image, suggesting that these models are not yet fully effective at handling multi-image tasks. By correlating performance on MVL-SIB with other multilingual VL benchmarks, we highlight that MVL-SIB serves as a comprehensive probe of multilingual VL understanding in LVLMs.

Xintong Wang, Jingheng Pan, Liang Ding et al.

Large Language Models (LLMs) achieve remarkable performance through pretraining on extensive data. This enables efficient adaptation to diverse downstream tasks. However, the lack of interpretability in their underlying mechanisms limits the ability to effectively steer LLMs for specific applications. In this work, we investigate the intrinsic mechanisms of LLMs from a cognitive perspective using eye movement measures. Specifically, we analyze the layer-wise correlation between human cognitive indicators and LLM representations. Building on these insights, we propose a heuristic approach for selecting the optimal steering layer to modulate LLM semantics. To this end, we introduce an efficient selective layer intervention based on prominent parameter-efficient fine-tuning methods, which conventionally adjust either all layers or only the final layer. Additionally, we present an implicit layer contrastive intervention during inference to steer LLMs away from toxic outputs. Extensive experiments on natural language understanding, reasoning, and generation tasks, conducted on GPT-2, Llama2-7B, and Mistral-7B, demonstrate the effectiveness and efficiency of our approach. As a model-agnostic framework, it enhances the interpretability of LLMs while improving efficiency for safe deployment.

Rudy Alexandro Garrido Veliz, Martin Semmann, Chris Biemann et al.

Neo-fascism is a political and societal ideology that has been having remarkable growth in the last decade in the United States of America (USA), as well as in other Western societies. It poses a grave danger to democracy and the minorities it targets, and it requires active actions against it to avoid escalation. This work presents the first-of-its-kind neo-fascist coding scheme for digital discourse in the USA societal context, overseen by political science researchers. Our work bridges the gap between Natural Language Processing (NLP) and political science against this phenomena. Furthermore, to test the coding scheme, we collect a tremendous amount of activity on the internet from notable neo-fascist groups (the forums of Iron March and Stormfront.org), and the guidelines are applied to a subset of the collected posts. Through crowdsourcing, we annotate a total of a thousand posts that are labeled as neo-fascist or non-neo-fascist. With this labeled data set, we fine-tune and test both Small Language Models (SLMs) and Large Language Models (LLMs), obtaining the very first classification models for neo-fascist discourse. We find that the prevalence of neo-fascist rhetoric in this kind of forum is ever-present, making them a good target for future research. The societal context is a key consideration for neo-fascist speech when conducting NLP research. Finally, the work against this kind of political movement must be pressed upon and continued for the well-being of a democratic society. Disclaimer: This study focuses on detecting neo-fascist content in text, similar to other hate speech analyses, without labeling individuals or organizations.

Viktor Moskvoretskii, Alina Lobanova, Ekaterina Neminova et al.

This paper explores the feasibility of using text-to-image models in a zero-shot setup to generate images for taxonomy concepts. While text-based methods for taxonomy enrichment are well-established, the potential of the visual dimension remains unexplored. To address this, we propose a comprehensive benchmark for Taxonomy Image Generation that assesses models' abilities to understand taxonomy concepts and generate relevant, high-quality images. The benchmark includes common-sense and randomly sampled WordNet concepts, alongside the LLM generated predictions. The 12 models are evaluated using 9 novel taxonomy-related text-to-image metrics and human feedback. Moreover, we pioneer the use of pairwise evaluation with GPT-4 feedback for image generation. Experimental results show that the ranking of models differs significantly from standard T2I tasks. Playground-v2 and FLUX consistently outperform across metrics and subsets and the retrieval-based approach performs poorly. These findings highlight the potential for automating the curation of structured data resources.

Irina Nikishina, Saba Anwar, Nikolay Dolgov et al.

In this paper, we aim to solve the problems standing in the way of automatic comparative question answering. To this end, we propose an evaluation framework to assess the quality of comparative question answering summaries. We formulate 15 criteria for assessing comparative answers created using manual annotation and annotation from 6 large language models and two comparative question asnwering datasets. We perform our tests using several LLMs and manual annotation under different settings and demonstrate the constituency of both evaluations. Our results demonstrate that the Llama-3 70B Instruct model demonstrates the best results for summary evaluation, while GPT-4 is the best for answering comparative questions. All used data, code, and evaluation results are publicly available\footnote{\url{https://anonymous.4open.science/r/cqa-evaluation-benchmark-4561/README.md}}.

Seid Muhie Yimam, Daryna Dementieva, Tim Fischer et al.

Despite regulations imposed by nations and social media platforms, such as recent EU regulations targeting digital violence, abusive content persists as a significant challenge. Existing approaches primarily rely on binary solutions, such as outright blocking or banning, yet fail to address the complex nature of abusive speech. In this work, we propose a more comprehensive approach called Demarcation scoring abusive speech based on four aspect -- (i) severity scale; (ii) presence of a target; (iii) context scale; (iv) legal scale -- and suggesting more options of actions like detoxification, counter speech generation, blocking, or, as a final measure, human intervention. Through a thorough analysis of abusive speech regulations across diverse jurisdictions, platforms, and research papers we highlight the gap in preventing measures and advocate for tailored proactive steps to combat its multifaceted manifestations. Our work aims to inform future strategies for effectively addressing abusive speech online.

Debayan Banerjee, Pranav Ajit Nair, Ricardo Usbeck et al.

In this work, we analyse the role of output vocabulary for text-to-text (T2T) models on the task of SPARQL semantic parsing. We perform experiments within the the context of knowledge graph question answering (KGQA), where the task is to convert questions in natural language to the SPARQL query language. We observe that the query vocabulary is distinct from human vocabulary. Language Models (LMs) are pre-dominantly trained for human language tasks, and hence, if the query vocabulary is replaced with a vocabulary more attuned to the LM tokenizer, the performance of models may improve. We carry out carefully selected vocabulary substitutions on the queries and find absolute gains in the range of 17% on the GrailQA dataset.

Markus J. Hofmann, Steffen Remus, Chris Biemann et al.

Though there is a strong consensus that word length and frequency are the most important single-word features determining visual-orthographic access to the mental lexicon, there is less agreement as how to best capture syntactic and semantic factors. The traditional approach in cognitive reading research assumes that word predictability from sentence context is best captured by cloze completion probability (CCP) derived from human performance data. We review recent research suggesting that probabilistic language models provide deeper explanations for syntactic and semantic effects than CCP. Then we compare CCP with (1) Symbolic n-gram models consolidate syntactic and semantic short-range relations by computing the probability of a word to occur, given two preceding words. (2) Topic models rely on subsymbolic representations to capture long-range semantic similarity by word co-occurrence counts in documents. (3) In recurrent neural networks (RNNs), the subsymbolic units are trained to predict the next word, given all preceding words in the sentences. To examine lexical retrieval, these models were used to predict single fixation durations and gaze durations to capture rapidly successful and standard lexical access, and total viewing time to capture late semantic integration. The linear item-level analyses showed greater correlations of all language models with all eye-movement measures than CCP. Then we examined non-linear relations between the different types of predictability and the reading times using generalized additive models. N-gram and RNN probabilities of the present word more consistently predicted reading performance compared with topic models or CCP.

Niklas von Boguszewski, Sana Moin, Anirban Bhowmick et al.

In this research, we investigate techniques to detect hate speech in movies. We introduce a new dataset collected from the subtitles of six movies, where each utterance is annotated either as hate, offensive or normal. We apply transfer learning techniques of domain adaptation and fine-tuning on existing social media datasets, namely from Twitter and Fox News. We evaluate different representations, i.e., Bag of Words (BoW), Bi-directional Long short-term memory (Bi-LSTM), and Bidirectional Encoder Representations from Transformers (BERT) on 11k movie subtitles. The BERT model obtained the best macro-averaged F1-score of 77%. Hence, we show that transfer learning from the social media domain is efficacious in classifying hate and offensive speech in movies through subtitles.

Dirk Johannßen, Chris Biemann

The COVID-19 pandemic has caused international social tension and unrest. Besides the crisis itself, there are growing signs of rising conflict potential of societies around the world. Indicators of global mood changes are hard to detect and direct questionnaires suffer from social desirability biases. However, so-called implicit methods can reveal humans intrinsic desires from e.g. social media texts. We present psychologically validated social unrest predictors and replicate scalable and automated predictions, setting a new state of the art on a recent German shared task dataset. We employ this model to investigate a change of language towards social unrest during the COVID-19 pandemic by comparing established psychological predictors on samples of tweets from spring 2019 with spring 2020. The results show a significant increase of the conflict indicating psychometrics. With this work, we demonstrate the applicability of automated NLP-based approaches to quantitative psychological research.

Seid Muhie Yimam, Abinew Ali Ayele, Gopalakrishnan Venkatesh et al.

The availability of different pre-trained semantic models enabled the quick development of machine learning components for downstream applications. Despite the availability of abundant text data for low resource languages, only a few semantic models are publicly available. Publicly available pre-trained models are usually built as a multilingual version of semantic models that can not fit well for each language due to context variations. In this work, we introduce different semantic models for Amharic. After we experiment with the existing pre-trained semantic models, we trained and fine-tuned nine new different models using a monolingual text corpus. The models are build using word2Vec embeddings, distributional thesaurus (DT), contextual embeddings, and DT embeddings obtained via network embedding algorithms. Moreover, we employ these models for different NLP tasks and investigate their impact. We find that newly trained models perform better than pre-trained multilingual models. Furthermore, models based on contextual embeddings from RoBERTA perform better than the word2Vec models.

Markus J. Hofmann, Lara Müller, Andre Rölke et al.

The corpus, from which a predictive language model is trained, can be considered the experience of a semantic system. We recorded everyday reading of two participants for two months on a tablet, generating individual corpus samples of 300/500K tokens. Then we trained word2vec models from individual corpora and a 70 million-sentence newspaper corpus to obtain individual and norm-based long-term memory structure. To test whether individual corpora can make better predictions for a cognitive task of long-term memory retrieval, we generated stimulus materials consisting of 134 sentences with uncorrelated individual and norm-based word probabilities. For the subsequent eye tracking study 1-2 months later, our regression analyses revealed that individual, but not norm-corpus-based word probabilities can account for first-fixation duration and first-pass gaze duration. Word length additionally affected gaze duration and total viewing duration. The results suggest that corpora representative for an individual's longterm memory structure can better explain reading performance than a norm corpus, and that recently acquired information is lexically accessed rapidly.

Ozge Sevgili, Artem Shelmanov, Mikhail Arkhipov et al.

This survey presents a comprehensive description of recent neural entity linking (EL) systems developed since 2015 as a result of the "deep learning revolution" in natural language processing. Its goal is to systemize design features of neural entity linking systems and compare their performance to the remarkable classic methods on common benchmarks. This work distills a generic architecture of a neural EL system and discusses its components, such as candidate generation, mention-context encoding, and entity ranking, summarizing prominent methods for each of them. The vast variety of modifications of this general architecture are grouped by several common themes: joint entity mention detection and disambiguation, models for global linking, domain-independent techniques including zero-shot and distant supervision methods, and cross-lingual approaches. Since many neural models take advantage of entity and mention/context embeddings to represent their meaning, this work also overviews prominent entity embedding techniques. Finally, the survey touches on applications of entity linking, focusing on the recently emerged use-case of enhancing deep pre-trained masked language models based on the Transformer architecture.

Benjamin Milde, Chris Biemann

The Sparsespeech model is an unsupervised acoustic model that can generate discrete pseudo-labels for untranscribed speech. We extend the Sparsespeech model to allow for sampling over a random discrete variable, yielding pseudo-posteriorgrams. The degree of sparsity in this posteriorgram can be fully controlled after the model has been trained. We use the Gumbel-Softmax trick to approximately sample from a discrete distribution in the neural network and this allows us to train the network efficiently with standard backpropagation. The new and improved model is trained and evaluated on the Libri-Light corpus, a benchmark for ASR with limited or no supervision. The model is trained on 600h and 6000h of English read speech. We evaluate the improved model using the ABX error measure and a semi-supervised setting with 10h of transcribed speech. We observe a relative improvement of up to 31.4% on ABX error rates across speakers on the test set with the improved Sparsespeech model on 600h of speech data and further improvements when we scale the model to 6000h.

Gregor Wiedemann, Seid Muhie Yimam, Chris Biemann

Fine-tuning of pre-trained transformer networks such as BERT yield state-of-the-art results for text classification tasks. Typically, fine-tuning is performed on task-specific training datasets in a supervised manner. One can also fine-tune in unsupervised manner beforehand by further pre-training the masked language modeling (MLM) task. Hereby, in-domain data for unsupervised MLM resembling the actual classification target dataset allows for domain adaptation of the model. In this paper, we compare current pre-trained transformer networks with and without MLM fine-tuning on their performance for offensive language detection. Our MLM fine-tuned RoBERTa-based classifier officially ranks 1st in the SemEval 2020 Shared Task~12 for the English language. Further experiments with the ALBERT model even surpass this result.