Sebastian Möller

Qianli Wang, Tatiana Anikina, Nils Feldhus et al.

Natural language explanations (NLEs) are vital for elucidating the reasoning behind large language model (LLM) decisions. Many techniques have been developed to generate NLEs using LLMs. However, like humans, LLMs might not always produce optimal NLEs on first attempt. Inspired by human learning processes, we introduce Cross-Refine, which employs role modeling by deploying two LLMs as generator and critic, respectively. The generator outputs a first NLE and then refines this initial explanation using feedback and suggestions provided by the critic. Cross-Refine does not require any supervised training data or additional training. We validate Cross-Refine across three NLP tasks using three state-of-the-art open-source LLMs through automatic and human evaluation. We select Self-Refine (Madaan et al., 2023) as the baseline, which only utilizes self-feedback to refine the explanations. Our findings from automatic evaluation and a user study indicate that Cross-Refine outperforms Self-Refine. Meanwhile, Cross-Refine can perform effectively with less powerful LLMs, whereas Self-Refine only yields strong results with ChatGPT. Additionally, we conduct an ablation study to assess the importance of feedback and suggestions. Both of them play an important role in refining explanations. We further evaluate Cross-Refine on a bilingual dataset in English and German.

Roland Roller, Klemens Budde, Aljoscha Burchardt et al.

Scientific publications about machine learning in healthcare are often about implementing novel methods and boosting the performance - at least from a computer science perspective. However, beyond such often short-lived improvements, much more needs to be taken into consideration if we want to arrive at a sustainable progress in healthcare. What does it take to actually implement such a system, make it usable for the domain expert, and possibly bring it into practical usage? Targeted at Computer Scientists, this work presents a multidisciplinary view on machine learning in medical decision support systems and covers information technology, medical, as well as ethical aspects. Along with an implemented risk prediction system in nephrology, challenges and lessons learned in a pilot project are presented.

Roland Roller, Laura Seiffe, Ammer Ayach et al.

Background: In the information extraction and natural language processing domain, accessible datasets are crucial to reproduce and compare results. Publicly available implementations and tools can serve as benchmark and facilitate the development of more complex applications. However, in the context of clinical text processing the number of accessible datasets is scarce -- and so is the number of existing tools. One of the main reasons is the sensitivity of the data. This problem is even more evident for non-English languages. Approach: In order to address this situation, we introduce a workbench: a collection of German clinical text processing models. The models are trained on a de-identified corpus of German nephrology reports. Result: The presented models provide promising results on in-domain data. Moreover, we show that our models can be also successfully applied to other biomedical text in German. Our workbench is made publicly available so it can be used out of the box, as a benchmark or transferred to related problems.

Lisa Raithel, Philippe Thomas, Roland Roller et al.

In this work, we present the first corpus for German Adverse Drug Reaction (ADR) detection in patient-generated content. The data consists of 4,169 binary annotated documents from a German patient forum, where users talk about health issues and get advice from medical doctors. As is common in social media data in this domain, the class labels of the corpus are very imbalanced. This and a high topic imbalance make it a very challenging dataset, since often, the same symptom can have several causes and is not always related to a medication intake. We aim to encourage further multi-lingual efforts in the domain of ADR detection and provide preliminary experiments for binary classification using different methods of zero- and few-shot learning based on a multi-lingual model. When fine-tuning XLM-RoBERTa first on English patient forum data and then on the new German data, we achieve an F1-score of 37.52 for the positive class. We make the dataset and models publicly available for the community.

Nils Feldhus, Leonhard Hennig, Maximilian Dustin Nasert et al.

Saliency maps can explain a neural model's predictions by identifying important input features. They are difficult to interpret for laypeople, especially for instances with many features. In order to make them more accessible, we formalize the underexplored task of translating saliency maps into natural language and compare methods that address two key challenges of this approach -- what and how to verbalize. In both automatic and human evaluation setups, using token-level attributions from text classification tasks, we compare two novel methods (search-based and instruction-based verbalizations) against conventional feature importance representations (heatmap visualizations and extractive rationales), measuring simulatability, faithfulness, helpfulness and ease of understanding. Instructing GPT-3.5 to generate saliency map verbalizations yields plausible explanations which include associations, abstractive summarization and commonsense reasoning, achieving by far the highest human ratings, but they are not faithfully capturing numeric information and are inconsistent in their interpretation of the task. In comparison, our search-based, model-free verbalization approach efficiently completes templated verbalizations, is faithful by design, but falls short in helpfulness and simulatability. Our results suggest that saliency map verbalization makes feature attribution explanations more comprehensible and less cognitively challenging to humans than conventional representations.

Salar Mohtaj, Babak Naderi, Sebastian Möller et al.

Text readability assessment has a wide range of applications for different target people, from language learners to people with disabilities. The fast pace of textual content production on the web makes it impossible to measure text complexity without the benefit of machine learning and natural language processing techniques. Although various research addressed the readability assessment of English text in recent years, there is still room for improvement of the models for other languages. In this paper, we proposed a new model for text complexity assessment for German text based on transfer learning. Our results show that the model outperforms more classical solutions based on linguistic features extraction from input text. The best model is based on the BERT pre-trained language model achieved the Root Mean Square Error (RMSE) of 0.483.

Nils Feldhus, Qianli Wang, Tatiana Anikina et al.

While recently developed NLP explainability methods let us open the black box in various ways (Madsen et al., 2022), a missing ingredient in this endeavor is an interactive tool offering a conversational interface. Such a dialogue system can help users explore datasets and models with explanations in a contextualized manner, e.g. via clarification or follow-up questions, and through a natural language interface. We adapt the conversational explanation framework TalkToModel (Slack et al., 2022) to the NLP domain, add new NLP-specific operations such as free-text rationalization, and illustrate its generalizability on three NLP tasks (dialogue act classification, question answering, hate speech detection). To recognize user queries for explanations, we evaluate fine-tuned and few-shot prompting models and implement a novel Adapter-based approach. We then conduct two user studies on (1) the perceived correctness and helpfulness of the dialogues, and (2) the simulatability, i.e. how objectively helpful dialogical explanations are for humans in figuring out the model's predicted label when it's not shown. We found rationalization and feature attribution were helpful in explaining the model behavior. Moreover, users could more reliably predict the model outcome based on an explanation dialogue rather than one-off explanations.

Nils Feldhus, Ajay Madhavan Ravichandran, Sebastian Möller

The human-centric explainable artificial intelligence (HCXAI) community has raised the need for framing the explanation process as a conversation between human and machine. In this position paper, we establish desiderata for Mediators, text-based conversational agents which are capable of explaining the behavior of neural models interactively using natural language. From the perspective of natural language processing (NLP) research, we engineer a blueprint of such a Mediator for the task of sentiment analysis and assess how far along current research is on the path towards dialogue-based explanations.

Qianli Wang, Nils Feldhus, Pepa Atanasova et al.

Quantization is widely used to accelerate inference and streamline the deployment of large language models (LLMs), yet its effects on self-explanations (SEs) remain unexplored. SEs, generated by LLMs to justify their own outputs, require reasoning about the model's own decision-making process, a capability that may exhibit particular sensitivity to quantization. As SEs are increasingly relied upon for transparency in high-stakes applications, understanding whether and to what extent quantization degrades SE quality and faithfulness is critical. To address this gap, we examine two types of SEs: natural language explanations (NLEs) and counterfactual examples, generated by LLMs quantized using three common techniques at distinct bit widths. Our findings indicate that quantization typically leads to moderate declines in both SE quality (up to 4.4\%) and faithfulness (up to 2.38\%). The user study further demonstrates that quantization diminishes both the coherence and trustworthiness of SEs (up to 8.5\%). Compared to smaller models, larger models show limited resilience to quantization in terms of SE quality but better maintain faithfulness. Moreover, no quantization technique consistently excels across task accuracy, SE quality, and faithfulness. Given that quantization's impact varies by context, we recommend validating SE quality for specific use cases, especially for NLEs, which show greater sensitivity. Nonetheless, the relatively minor deterioration in SE quality and faithfulness does not undermine quantization's effectiveness as a model compression technique.

Qianli Wang, Van Bach Nguyen, Yihong Liu et al.

Counterfactuals refer to minimally edited inputs that cause a model's prediction to change, serving as a promising approach to explaining the model's behavior. Large language models (LLMs) excel at generating English counterfactuals and demonstrate multilingual proficiency. However, their effectiveness in generating multilingual counterfactuals remains unclear. To this end, we conduct a comprehensive study on multilingual counterfactuals. We first conduct automatic evaluations on both directly generated counterfactuals in the target languages and those derived via English translation across six languages. Although translation-based counterfactuals offer higher validity than their directly generated counterparts, they demand substantially more modifications and still fall short of matching the quality of the original English counterfactuals. Second, we find the patterns of edits applied to high-resource European-language counterfactuals to be remarkably similar, suggesting that cross-lingual perturbations follow common strategic principles. Third, we identify and categorize four main types of errors that consistently appear in the generated counterfactuals across languages. Finally, we reveal that multilingual counterfactual data augmentation (CDA) yields larger model performance improvements than cross-lingual CDA, especially for lower-resource languages. Yet, the imperfections of the generated counterfactuals limit gains in model performance and robustness.

Jing Yang, Nils Feldhus, Salar Mohtaj et al.

Despite advances in Natural Language Generation (NLG), evaluation remains challenging. Although various new metrics and LLM-as-a-judge (LaaJ) methods are proposed, human judgment persists as the gold standard. To systematically review how NLG evaluation has evolved, we employ an automatic information extraction scheme to gather key information from NLG papers, focusing on different evaluation methods (metrics, LaaJ and human evaluation). With extracted metadata from 14,171 papers across four major conferences (ACL, EMNLP, NAACL, and INLG) over the past six years, we reveal several critical findings: (1) Task Divergence: While Dialogue Generation demonstrates a rapid shift toward LaaJ (>40% in 2025), Machine Translation remains locked into n-gram metrics, and Question Answering exhibits a substantial decline in the proportion of studies conducting human evaluation. (2) Metric Inertia: Despite the development of semantic metrics, general-purpose metrics (e.g., BLEU, ROUGE) continue to be widely used across tasks without empirical justification, often lacking the discriminative power to distinguish between specific quality criteria. (3) Human-LaaJ Divergence: Our association analysis challenges the assumption that LLMs act as mere proxies for humans; LaaJ and human evaluations prioritize very different signals, and explicit validation is scarce (<8% of papers comparing the two), with only moderate to low correlation. Based on these observations, we derive practical recommendations to improve the rigor of future NLG evaluation.

Dorothea MacPhail, David Harbecke, Lisa Raithel et al.

An adverse drug effect (ADE) is any harmful event resulting from medical drug treatment. Despite their importance, ADEs are often under-reported in official channels. Some research has therefore turned to detecting discussions of ADEs in social media. Impressive results have been achieved in various attempts to detect ADEs. In a high-stakes domain such as medicine, however, an in-depth evaluation of a model's abilities is crucial. We address the issue of thorough performance evaluation in English-language ADE detection with hand-crafted templates for four capabilities: Temporal order, negation, sentiment, and beneficial effect. We find that models with similar performance on held-out test sets have varying results on these capabilities.

Nils Feldhus, Tanja Baeumel, Elena Golimblevskaia et al.

LLM-as-a-judge has become the dominant paradigm for grading model outputs at scale, yet the same model assigns systematically different scores when its output format changes (e.g., a 1-5 rating vs. a True/False label). Existing diagnoses of these format-induced inconsistencies stop at the input-output level. Using Position-aware Edge Attribution Patching (PEAP), we causally investigate the internal mechanism in Gemma-3, Qwen2.5, and Llama-3. We find that judgments across structured understanding and open-ended preference tasks share a sparse, generalized Latent Evaluator sub-graph in the mid-to-late multi-layer perceptrons (MLPs); zero-ablating it collapses judgment while preserving world knowledge in architecturally modular models. By structurally decoupling abstract judging from output formatting, we provide a mechanistic account of format-induced inconsistency on the open-weight models we study: a continuous judgment signal computed in the shared trunk is mapped through fragile, format-specific terminal branches, enabling format-independent preference to be isolated downstream of the requested output format. Our findings imply that benchmark-level reliability comparisons across formats are partially measuring formatter geometry rather than evaluation quality.

Lukas Stähelin, Veronika Solopova, Max Upravitelev et al.

Propaganda detection in social media is challenging due to noisy, short texts and low annotation agreements. We introduce a new intent-focused taxonomy of propaganda techniques and compare it against an established, higher-agreement schema. Along three dimensions (model portfolio, schema effects, and prompting strategy) we evaluate the taxonomies as a classification task with the help of four language models (GPT-4.1-nano, Phi-4 14B, Qwen2.5-14B, Qwen3-14B). Our results show that fine-tuning is essential, since it transforms weak zero-shot baselines into competitive systems and reveals methodological differences that are hidden using base models. Across schemas, the Qwen models achieve the strongest overall performance, and Phi-4 14B consistently outperforms GPT-4.1-nano. Our hierarchical prompting method (HiPP), which predicts fine-grained techniques before aggregating them, is especially beneficial after fine-tuning and on the more ambiguous, low-agreement taxonomy, while remaining competitive on the simpler schema. The HQP dataset, annotated with the new intent-based labels, provides a richer lens on propaganda's strategic goals and a challenging benchmark for future work on robust, real-world detection.

Adnan Ahmad, Philine Kowol, Stefan Hillmann et al.

In this paper, we provide an extensive analysis of multi-label intent classification using Large Language Models (LLMs) that are open-source, publicly available, and can be run in consumer hardware. We use the MultiWOZ 2.1 dataset, a benchmark in the dialogue system domain, to investigate the efficacy of three popular open-source pre-trained LLMs, namely LLama2-7B-hf, Mistral-7B-v0.1, and Yi-6B. We perform the classification task in a few-shot setup, giving 20 examples in the prompt with some instructions. Our approach focuses on the differences in performance of these models across several performance metrics by methodically assessing these models on multi-label intent classification tasks. Additionally, we compare the performance of the instruction-based fine-tuning approach with supervised learning using the smaller transformer model BertForSequenceClassification as a baseline. To evaluate the performance of the models, we use evaluation metrics like accuracy, precision, and recall as well as micro, macro, and weighted F1 score. We also report the inference time, VRAM requirements, etc. The Mistral-7B-v0.1 outperforms two other generative models on 11 intent classes out of 14 in terms of F-Score, with a weighted average of 0.50. It also has relatively lower Humming Loss and higher Jaccard Similarity, making it the winning model in the few-shot setting. We find BERT based supervised classifier having superior performance compared to the best performing few-shot generative LLM. The study provides a framework for small open-source LLMs in detecting complex multi-intent dialogues, enhancing the Natural Language Understanding aspect of task-oriented chatbots.

Nils Feldhus, Robert Schwarzenberg, Sebastian Möller

In the language domain, as in other domains, neural explainability takes an ever more important role, with feature attribution methods on the forefront. Many such methods require considerable computational resources and expert knowledge about implementation details and parameter choices. To facilitate research, we present Thermostat which consists of a large collection of model explanations and accompanying analysis tools. Thermostat allows easy access to over 200k explanations for the decisions of prominent state-of-the-art models spanning across different NLP tasks, generated with multiple explainers. The dataset took over 10k GPU hours (> one year) to compile; compute time that the community now saves. The accompanying software tools allow to analyse explanations instance-wise but also accumulatively on corpus level. Users can investigate and compare models, datasets and explainers without the need to orchestrate implementation details. Thermostat is fully open source, democratizes explainability research in the language domain, circumvents redundant computations and increases comparability and replicability.

Gabriel Mittag, Babak Naderi, Assmaa Chehadi et al.

In this paper, we present an update to the NISQA speech quality prediction model that is focused on distortions that occur in communication networks. In contrast to the previous version, the model is trained end-to-end and the time-dependency modelling and time-pooling is achieved through a Self-Attention mechanism. Besides overall speech quality, the model also predicts the four speech quality dimensions Noisiness, Coloration, Discontinuity, and Loudness, and in this way gives more insight into the cause of a quality degradation. Furthermore, new datasets with over 13,000 speech files were created for training and validation of the model. The model was finally tested on a new, live-talking test dataset that contains recordings of real telephone calls. Overall, NISQA was trained and evaluated on 81 datasets from different sources and showed to provide reliable predictions also for unknown speech samples. The code, model weights, and datasets are open-sourced.

Babak Naderi, Sebastian Möller

The rank correlation coefficients and the ranked-based statistical tests (as a subset of non-parametric techniques) might be misleading when they are applied to subjectively collected opinion scores. Those techniques assume that the data is measured at least at an ordinal level and define a sequence of scores to represent a tied rank when they have precisely an equal numeric value. In this paper, we show that the definition of tied rank, as mentioned above, is not suitable for Mean Opinion Scores (MOS) and might be misleading conclusions of rank-based statistical techniques. Furthermore, we introduce a method to overcome this issue by transforming the MOS values considering their $95\%$ Confidence Intervals. The rank correlation coefficients and ranked-based statistical tests can then be safely applied to the transformed values. We also provide open-source software packages in different programming languages to utilize the application of our transformation method in the quality of experience domain.

Robert Schwarzenberg, David Harbecke, Vivien Macketanz et al.

Evaluating translation models is a trade-off between effort and detail. On the one end of the spectrum there are automatic count-based methods such as BLEU, on the other end linguistic evaluations by humans, which arguably are more informative but also require a disproportionately high effort. To narrow the spectrum, we propose a general approach on how to automatically expose systematic differences between human and machine translations to human experts. Inspired by adversarial settings, we train a neural text classifier to distinguish human from machine translations. A classifier that performs and generalizes well after training should recognize systematic differences between the two classes, which we uncover with neural explainability methods. Our proof-of-concept implementation, DiaMaT, is open source. Applied to a dataset translated by a state-of-the-art neural Transformer model, DiaMaT achieves a classification accuracy of 75% and exposes meaningful differences between humans and the Transformer, amidst the current discussion about human parity.

Premtim Sahitaj, Jawan Kolanowski, Ariana Sahitaj et al.

Fact-checking articles encode rich supporting evidence and reasoning, yet this evidence remains largely inaccessible to automated verification systems due to unstructured presentation. We introduce PrimeFacts, a methodology and resource for extracting fine-grained evidence from full fact-checking articles. We compile 13,106 PolitiFact articles with claims, verdicts, and all referenced sources, and we identify 49,718 in-article hyperlinks as natural anchors to pinpoint key evidence. Our framework leverages large language models (LLMs) to rewrite these anchor sentences into stand-alone, context-independent premises and investigates the extraction of additional implicit evidence. In evaluations on cross-article evidence retrieval and claim verification, the extracted premises substantially improve performance. Decontextualized evidence yields higher retrievability, achieving up to a 30 percent relative gain in Mean Reciprocal Rank over verbatim sentences, and using the evidence for verdict prediction raises Macro-F1 by 10-20 points over the baseline. These gains are consistent across different verdict granularities (2-class vs. 5-class) and model architectures. A qualitative analysis indicates that the decontextualized premises remain faithful to the original sources. Our work highlights the promise of reusing fact-checkers' evidence for automation and provides a large-scale resource of structured evidence from real-world fact-checks.

Qianli Wang, Tatiana Anikina, Nils Feldhus et al.

Interpretability tools that offer explanations in the form of a dialogue have demonstrated their efficacy in enhancing users' understanding (Slack et al., 2023; Shen et al., 2023), as one-off explanations may fall short in providing sufficient information to the user. Current solutions for dialogue-based explanations, however, often require external tools and modules and are not easily transferable to tasks they were not designed for. With LLMCheckup, we present an easily accessible tool that allows users to chat with any state-of-the-art large language model (LLM) about its behavior. We enable LLMs to generate explanations and perform user intent recognition without fine-tuning, by connecting them with a broad spectrum of Explainable AI (XAI) methods, including white-box explainability tools such as feature attributions, and self-explanations (e.g., for rationale generation). LLM-based (self-)explanations are presented as an interactive dialogue that supports follow-up questions and generates suggestions. LLMCheckupprovides tutorials for operations available in the system, catering to individuals with varying levels of expertise in XAI and supporting multiple input modalities. We introduce a new parsing strategy that substantially enhances the user intent recognition accuracy of the LLM. Finally, we showcase LLMCheckup for the tasks of fact checking and commonsense question answering.

Navid Ashrafi, Vera Schmitt, Robert P. Spang et al.

Preservation of private user data is of paramount importance for high Quality of Experience (QoE) and acceptability, particularly with services treating sensitive data, such as IT-based health services. Whereas anonymization techniques were shown to be prone to data re-identification, synthetic data generation has gradually replaced anonymization since it is relatively less time and resource-consuming and more robust to data leakage. Generative Adversarial Networks (GANs) have been used for generating synthetic datasets, especially GAN frameworks adhering to the differential privacy phenomena. This research compares state-of-the-art GAN-based models for synthetic data generation to generate time-series synthetic medical records of dementia patients which can be distributed without privacy concerns. Predictive modeling, autocorrelation, and distribution analysis are used to assess the Quality of Generating (QoG) of the generated data. The privacy preservation of the respective models is assessed by applying membership inference attacks to determine potential data leakage risks. Our experiments indicate the superiority of the privacy-preserving GAN (PPGAN) model over other models regarding privacy preservation while maintaining an acceptable level of QoG. The presented results can support better data protection for medical use cases in the future.

Premtim Sahitaj, Iffat Maab, Junichi Yamagishi et al.

Fact-checking is necessary to address the increasing volume of misinformation. Traditional fact-checking relies on manual analysis to verify claims, but it is slow and resource-intensive. This study establishes baseline comparisons for Automated Fact-Checking (AFC) using Large Language Models (LLMs) across multiple labeling schemes (binary, three-class, five-class) and extends traditional claim verification by incorporating analysis, verdict classification, and explanation in a structured setup to provide comprehensive justifications for real-world claims. We evaluate Llama-3 models of varying sizes (3B, 8B, 70B) on 17,856 claims collected from PolitiFact (2007-2024) using evidence retrieved via restricted web searches. We utilize TIGERScore as a reference-free evaluation metric to score the justifications. Our results show that larger LLMs consistently outperform smaller LLMs in classification accuracy and justification quality without fine-tuning. We find that smaller LLMs in a one-shot scenario provide comparable task performance to fine-tuned Small Language Models (SLMs) with large context sizes, while larger LLMs consistently surpass them. Evidence integration improves performance across all models, with larger LLMs benefiting most. Distinguishing between nuanced labels remains challenging, emphasizing the need for further exploration of labeling schemes and alignment with evidences. Our findings demonstrate the potential of retrieval-augmented AFC with LLMs.

Lisa Raithel, Hui-Syuan Yeh, Shuntaro Yada et al.

User-generated data sources have gained significance in uncovering Adverse Drug Reactions (ADRs), with an increasing number of discussions occurring in the digital world. However, the existing clinical corpora predominantly revolve around scientific articles in English. This work presents a multilingual corpus of texts concerning ADRs gathered from diverse sources, including patient fora, social media, and clinical reports in German, French, and Japanese. Our corpus contains annotations covering 12 entity types, four attribute types, and 13 relation types. It contributes to the development of real-world multilingual language models for healthcare. We provide statistics to highlight certain challenges associated with the corpus and conduct preliminary experiments resulting in strong baselines for extracting entities and relations between these entities, both within and across languages.

Qianli Wang, Mingyang Wang, Nils Feldhus et al.

Quantization methods are widely used to accelerate inference and streamline the deployment of large language models (LLMs). While prior research has extensively investigated the degradation of various LLM capabilities due to quantization, its effects on model explainability and interpretability, which are crucial for understanding decision-making processes, remain unexplored. To address this gap, we conduct comprehensive experiments using three common quantization techniques at distinct bit widths, in conjunction with two explainability methods, counterfactual examples and natural language explanations, as well as two interpretability approaches, knowledge memorization analysis and latent multi-hop reasoning analysis. We complement our analysis with a thorough user study, evaluating selected explainability methods. Our findings reveal that, depending on the configuration, quantization can significantly impact model explainability and interpretability. Notably, the direction of this effect is not consistent, as it strongly depends on (1) the quantization method, (2) the explainability or interpretability approach, and (3) the evaluation protocol. In some settings, human evaluation shows that quantization degrades explainability, while in others, it even leads to improvements. Our work serves as a cautionary tale, demonstrating that quantization can unpredictably affect model transparency. This insight has important implications for deploying LLMs in applications where transparency is a critical requirement.

Qianli Wang, Nils Feldhus, Simon Ostermann et al.

Counterfactual examples are widely used in natural language processing (NLP) as valuable data to improve models, and in explainable artificial intelligence (XAI) to understand model behavior. The automated generation of counterfactual examples remains a challenging task even for large language models (LLMs), despite their impressive performance on many tasks. In this paper, we first introduce ZeroCF, a faithful approach for leveraging important words derived from feature attribution methods to generate counterfactual examples in a zero-shot setting. Second, we present a new framework, FitCF, which further verifies aforementioned counterfactuals by label flip verification and then inserts them as demonstrations for few-shot prompting, outperforming two state-of-the-art baselines. Through ablation studies, we identify the importance of each of FitCF's core components in improving the quality of counterfactuals, as assessed through flip rate, perplexity, and similarity measures. Furthermore, we show the effectiveness of LIME and Integrated Gradients as backbone attribution methods for FitCF and find that the number of demonstrations has the largest effect on performance. Finally, we reveal a strong correlation between the faithfulness of feature attribution scores and the quality of generated counterfactuals, which we hope will serve as an important finding for future research in this direction.

Arnab Das, Yassine El Kheir, Carlos Franzreb et al.

Rapid advancements in generative modeling have made synthetic audio generation easy, making speech-based services vulnerable to spoofing attacks. Consequently, there is a dire need for robust countermeasures more than ever. Existing solutions for deepfake detection are often criticized for lacking generalizability and fail drastically when applied to real-world data. This study proposes a novel method for generalizable spoofing detection leveraging non-semantic universal audio representations. Extensive experiments have been performed to find suitable non-semantic features using TRILL and TRILLsson models. The results indicate that the proposed method achieves comparable performance on the in-domain test set while significantly outperforming state-of-the-art approaches on out-of-domain test sets. Notably, it demonstrates superior generalization on public-domain data, surpassing methods based on hand-crafted features, semantic embeddings, and end-to-end architectures.

Ariana Sahitaj, Premtim Sahitaj, Veronika Solopova et al.

Propaganda detection on social media remains challenging due to task complexity and limited high-quality labeled data. This paper introduces a novel framework that combines human expertise with Large Language Model (LLM) assistance to improve both annotation consistency and scalability. We propose a hierarchical taxonomy that organizes 14 fine-grained propaganda techniques into three broader categories, conduct a human annotation study on the HQP dataset that reveals low inter-annotator agreement for fine-grained labels, and implement an LLM-assisted pre-annotation pipeline that extracts propagandistic spans, generates concise explanations, and assigns local labels as well as a global label. A secondary human verification study shows significant improvements in both agreement and time-efficiency. Building on this, we fine-tune smaller language models (SLMs) to perform structured annotation. Instead of fine-tuning on human annotations, we train on high-quality LLM-generated data, allowing a large model to produce these annotations and a smaller model to learn to generate them via knowledge distillation. Our work contributes towards the development of scalable and robust propaganda detection systems, supporting the idea of transparent and accountable media ecosystems in line with SDG 16. The code is publicly available at our GitHub repository.

Ibrahim Baroud, Lisa Raithel, Sebastian Möller et al.

Sharing sensitive texts for scientific purposes requires appropriate techniques to protect the privacy of patients and healthcare personnel. Anonymizing textual data is particularly challenging due to the presence of diverse unstructured direct and indirect identifiers. To mitigate the risk of re-identification, this work introduces a schema of nine categories of indirect identifiers designed to account for different potential adversaries, including acquaintances, family members and medical staff. Using this schema, we annotate 100 MIMIC-III discharge summaries and propose baseline models for identifying indirect identifiers. We will release the annotation guidelines, annotation spans (6,199 annotations in total) and the corresponding MIMIC-III document IDs to support further research in this area.

Wafaa Wardah, Tuğçe Melike Koçak Büyüktaş, Kirill Shchegelskiy et al.

Objective speech quality models aim to predict human-perceived speech quality using automated methods. However, cross-lingual generalization remains a major challenge, as Mean Opinion Scores (MOS) vary across languages due to linguistic, perceptual, and dataset-specific differences. A model trained primarily on English data may struggle to generalize to languages with different phonetic, tonal, and prosodic characteristics, leading to inconsistencies in objective assessments. This study investigates the cross-lingual performance of two speech quality models: NISQA, a CNN-based model, and a Transformer-based Audio Spectrogram Transformer (AST) model. Both models were trained exclusively on English datasets containing over 49,000 speech samples and subsequently evaluated on speech in German, French, Mandarin, Swedish, and Dutch. We analyze model performance using Pearson Correlation Coefficient (PCC) and Root Mean Square Error (RMSE) across five speech quality dimensions: coloration, discontinuity, loudness, noise, and MOS. Our findings show that while AST achieves a more stable cross-lingual performance, both models exhibit noticeable biases. Notably, Mandarin speech quality predictions correlate highly with human MOS scores, whereas Swedish and Dutch present greater prediction challenges. Discontinuities remain difficult to model across all languages. These results highlight the need for more balanced multilingual datasets and architecture-specific adaptations to improve cross-lingual generalization.

Adnan Ahmad, Stefan Hillmann, Sebastian Möller

In this study, we explore the application of Large Language Models (LLMs) for generating synthetic users and simulating user conversations with a task-oriented dialogue system and present detailed results and their analysis. We propose a comprehensive novel approach to user simulation technique that uses LLMs to create diverse user profiles, set goals, engage in multi-turn dialogues, and evaluate the conversation success. We employ two proprietary LLMs, namely GPT-4o and GPT-o1 (Achiam et al., 2023), to generate a heterogeneous base of user profiles, characterized by varied demographics, multiple user goals, different conversational styles, initial knowledge levels, interests, and conversational objectives. We perform a detailed analysis of the user profiles generated by LLMs to assess the diversity, consistency, and potential biases inherent in these LLM-generated user simulations. We find that GPT-o1 generates more heterogeneous user distribution across most user attributes, while GPT-4o generates more skewed user attributes. The generated set of user profiles are then utilized to simulate dialogue sessions by interacting with a task-oriented dialogue system.

Luis Felipe Villa-Arenas, Ata Nizamoglu, Qianli Wang et al.

In this work, we introduce a methodology for alignment designed to enhance the ability of large language models (LLMs) to articulate their reasoning (self-explanation) even in the absence of annotated rationale explanations. Our alignment methodology comprises three key components: explanation quality assessment, self-instruction dataset generation, and model alignment. Additionally, we present a novel technique called Alignment with Anchor Preference Pairs, which improves the selection of preference pairs by categorizing model outputs into three groups: consistently correct, consistently incorrect, and variable. By applying tailored strategies to each category, we enhance the effectiveness of Direct Preference Optimization (DPO). Our experimental results demonstrate that this approach significantly improves explanation quality while maintaining accuracy compared to other fine-tuning strategies.

Carlos Franzreb, Francisco Teixeira, Ben Luks et al.

Automatic detection of Parkinson's disease (PD) from speech is a promising non-invasive diagnostic tool, but it raises significant privacy concerns. Speaker anonymization mitigates these risks, but it may suppress the pathological information necessary for PD detection. We assess the trade-off between privacy and PD detection for two anonymizers (STT-TTS and kNN-VC) using two Spanish datasets. STT-TTS provides better privacy but severely degrades PD detection by eradicating prosodic information. kNN-VC preserves macro-prosodic features such as duration and F0 contours, achieving F1 scores only 3-7\% lower than original baselines, demonstrating that privacy-preserving PD detection is viable when using appropriate anonymization. Finally, an acoustic distortion analysis characterizes specific weaknesses in kNN-VC, offering insights for designing anonymizers that better preserve PD information.

Qianli Wang, Tatiana Anikina, Nils Feldhus et al.

Conversational explainable artificial intelligence (ConvXAI) systems based on large language models (LLMs) have garnered considerable attention for their ability to enhance user comprehension through dialogue-based explanations. Current ConvXAI systems often are based on intent recognition to accurately identify the user's desired intention and map it to an explainability method. While such methods offer great precision and reliability in discerning users' underlying intentions for English, a significant challenge in the scarcity of training data persists, which impedes multilingual generalization. Besides, the support for free-form custom inputs, which are user-defined data distinct from pre-configured dataset instances, remains largely limited. To bridge these gaps, we first introduce MultiCoXQL, a multilingual extension of the CoXQL dataset spanning five typologically diverse languages, including one low-resource language. Subsequently, we propose a new parsing approach aimed at enhancing multilingual parsing performance, and evaluate three LLMs on MultiCoXQL using various parsing strategies. Furthermore, we present Compass, a new multilingual dataset designed for custom input extraction in ConvXAI systems, encompassing 11 intents across the same five languages as MultiCoXQL. We conduct monolingual, cross-lingual, and multilingual evaluations on Compass, employing three LLMs of varying sizes alongside BERT-type models.

Carlos Franzreb, Arnab Das, Tim Polzehl et al.

The current privacy evaluation for speaker anonymization often overestimates privacy when a same-gender target selection algorithm (TSA) is used, although this TSA leaks the speaker's gender and should hence be more vulnerable. We hypothesize that this occurs because the evaluation does not account for the fact that anonymized speech contains information from both the source and target speakers. To address this, we propose to add a target classifier that measures the influence of target speaker information in the evaluation, which can also be removed with adversarial learning. Experiments demonstrate that this approach is effective for multiple anonymizers, particularly when using a same-gender TSA, leading to a more reliable assessment.

Eleftherios Avramidis, Vera Czehmann, Fabian Deckert et al.

We present a collection of parallel corpora of 12 sign languages in video format, together with subtitles in the dominant spoken languages of the corresponding countries. The entire collection includes more than 1,300 hours in 4,381 video files, accompanied by 1,3~M subtitles containing 14~M tokens. Most notably, it includes the first consistent parallel corpora for 8 Latin American sign languages, whereas the size of the German Sign Language corpora is ten times the size of the previously available corpora. The collection was created by collecting and processing videos of multiple sign languages from various online sources, mainly broadcast material of news shows, governmental bodies and educational channels. The preparation involved several stages, including data collection, informing the content creators and seeking usage approvals, scraping, and cropping. The paper provides statistics on the collection and an overview of the methods used to collect the data.

Fenya Wasserroth, Eleftherios Avramidis, Vera Czehmann et al.

This paper presents an investigation into the impact of adding adjustment features to an existing sign language (SL) avatar on a Microsoft Hololens 2 device. Through a detailed analysis of interactions of expert German Sign Language (DGS) users with both adjustable and non-adjustable avatars in a specific use case, this study identifies the key factors influencing the comprehensibility, the user experience (UX), and the acceptability of such a system. Despite user preference for adjustable settings, no significant improvements in UX or comprehensibility were observed, which remained at low levels, amid missing SL elements (mouthings and facial expressions) and implementation issues (indistinct hand shapes, lack of feedback and menu positioning). Hedonic quality was rated higher than pragmatic quality, indicating that users found the system more emotionally or aesthetically pleasing than functionally useful. Stress levels were higher for the adjustable avatar, reflecting lower performance, greater effort and more frustration. Additionally, concerns were raised about whether the Hololens adjustment gestures are intuitive and easy to familiarise oneself with. While acceptability of the concept of adjustability was generally positive, it was strongly dependent on usability and animation quality. This study highlights that personalisation alone is insufficient, and that SL avatars must be comprehensible by default. Key recommendations include enhancing mouthing and facial animation, improving interaction interfaces, and applying participatory design.

Ekaterina Borisova, Fabio Barth, Nils Feldhus et al.

Tables are among the most widely used tools for representing structured data in research, business, medicine, and education. Although LLMs demonstrate strong performance in downstream tasks, their efficiency in processing tabular data remains underexplored. In this paper, we investigate the effectiveness of both text-based and multimodal LLMs on table understanding tasks through a cross-domain and cross-modality evaluation. Specifically, we compare their performance on tables from scientific vs. non-scientific contexts and examine their robustness on tables represented as images vs. text. Additionally, we conduct an interpretability analysis to measure context usage and input relevance. We also introduce the TableEval benchmark, comprising 3017 tables from scholarly publications, Wikipedia, and financial reports, where each table is provided in five different formats: Image, Dictionary, HTML, XML, and LaTeX. Our findings indicate that while LLMs maintain robustness across table modalities, they face significant challenges when processing scientific tables.

Roland Roller, Michael Hahn, Ajay Madhavan Ravichandran et al.

Machine learning (ML) models are increasingly used to support clinical decision-making. However, real-world medical datasets are often noisy, incomplete, and imbalanced, leading to performance disparities across patient subgroups. These differences raise fairness concerns, particularly when they reinforce existing disadvantages for marginalized groups. In this work, we analyze several medical prediction tasks and demonstrate how model performance varies with patient characteristics. While ML models may demonstrate good overall performance, we argue that subgroup-level evaluation is essential before integrating them into clinical workflows. By conducting a performance analysis at the subgroup level, differences can be clearly identified-allowing, on the one hand, for performance disparities to be considered in clinical practice, and on the other hand, for these insights to inform the responsible development of more effective models. Thereby, our work contributes to a practical discussion around the subgroup-sensitive development and deployment of medical ML models and the interconnectedness of fairness and transparency.

Qianli Wang, Van Bach Nguyen, Nils Feldhus et al.

Counterfactual examples are widely employed to enhance the performance and robustness of large language models (LLMs) through counterfactual data augmentation (CDA). However, the selection of the judge model used to evaluate label flipping, the primary metric for assessing the validity of generated counterfactuals for CDA, yields inconsistent results. To decipher this, we define four types of relationships between the counterfactual generator and judge models: being the same model, belonging to the same model family, being independent models, and having an distillation relationship. Through extensive experiments involving two state-of-the-art LLM-based methods, three datasets, four generator models, and 15 judge models, complemented by a user study (n = 90), we demonstrate that judge models with an independent, non-fine-tuned relationship to the generator model provide the most reliable label flipping evaluations. Relationships between the generator and judge models, which are closely aligned with the user study for CDA, result in better model performance and robustness. Nevertheless, we find that the gap between the most effective judge models and the results obtained from the user study remains considerably large. This suggests that a fully automated pipeline for CDA may be inadequate and requires human intervention.

Sebastian Möller, Pia Knoeferle, Britta Schulte et al.

Machine learning techniques have conquered many different tasks in speech and natural language processing, such as speech recognition, information extraction, text and speech generation, and human machine interaction using natural language or speech (chatbots). Modern techniques typically rely on large models for representing general knowledge of one or several languages (Large Language Models, LLMs), or for representing speech and general audio characteristics. These models have been trained with large amounts of speech and language data, typically including web content. When humans interact with such technologies, the effectiveness of the interaction will be influenced by how far humans make use of the same type of language the models have been trained on or, in other words, if the models are able to generalize to the language used by humans when interacting with the technology. This may lead to some gradual forms of adaptation in human speech and language production, and users who do not adapt may be excluded from efficient use of such technologies. On top of this, as commercial model development follows market needs, under-represented languages and dialects/sociolects may decrease in terms of priorities. Furthermore, for many lesser spoken languages the necessary data is not available, which will worsen a digital divide in speech and language technology usage. The workshop sets out to discuss this problem based on scientific contributions from the perspective of computer science and linguistics (including computational linguistics and NLP).

Pia Knoeferle, Sebastian Möller, Dorothea Kolossa et al.

We discuss how desirable it is that Large Language Models (LLMs) be able to adapt or align their language behavior with users who may be diverse in their language use. User diversity may come about among others due to i) age differences; ii) gender characteristics, and/or iii) multilingual experience, and associated differences in language processing and use. We consider potential consequences for usability, communication, and LLM development.

Tatiana Anikina, Arne Binder, David Harbecke et al.

In this work, we reimagine classical probing to evaluate knowledge transfer from simple source to more complex target tasks. Instead of probing frozen representations from a complex source task on diverse simple target probing tasks (as usually done in probing), we explore the effectiveness of embeddings from multiple simple source tasks on a single target task. We select coreference resolution, a linguistically complex problem requiring contextual understanding, as focus target task, and test the usefulness of embeddings from comparably simpler tasks tasks such as paraphrase detection, named entity recognition, and relation extraction. Through systematic experiments, we evaluate the impact of individual and combined task embeddings. Our findings reveal that task embeddings vary significantly in utility for coreference resolution, with semantic similarity tasks (e.g., paraphrase detection) proving most beneficial. Additionally, representations from intermediate layers of fine-tuned models often outperform those from final layers. Combining embeddings from multiple tasks consistently improves performance, with attention-based aggregation yielding substantial gains. These insights shed light on relationships between task-specific representations and their adaptability to complex downstream tasks, encouraging further exploration of embedding-level task transfer.

Ben Hauptvogel, Malte Ostendorff, Georg Rehm et al.

Recent advancements in large language models (LLMs) have led to their increased application across various tasks, with reinforcement learning from human feedback (RLHF) being a crucial part of their training to align responses with user intentions. In the RLHF process, a reward model is trained using responses preferences determined by human labelers or AI systems, which then refines the LLM through reinforcement learning. This work introduces weak supervision as a strategy to extend RLHF datasets and enhance reward model performance. Weak supervision employs noisy or imprecise data labeling, reducing reliance on expensive manually labeled data. By analyzing RLHF datasets to identify heuristics that correlate with response preference, we wrote simple labeling functions and then calibrated a label model to weakly annotate unlabeled data. Our evaluation show that while weak supervision significantly benefits smaller datasets by improving reward model performance, its effectiveness decreases with larger, originally labeled datasets. Additionally, using an LLM to generate and then weakly label responses offers a promising method for extending preference data.

Qianli Wang, Tatiana Anikina, Nils Feldhus et al.

Conversational explainable artificial intelligence (ConvXAI) systems based on large language models (LLMs) have garnered significant interest from the research community in natural language processing (NLP) and human-computer interaction (HCI). Such systems can provide answers to user questions about explanations in dialogues, have the potential to enhance users' comprehension and offer more information about the decision-making and generation processes of LLMs. Currently available ConvXAI systems are based on intent recognition rather than free chat, as this has been found to be more precise and reliable in identifying users' intentions. However, the recognition of intents still presents a challenge in the case of ConvXAI, since little training data exist and the domain is highly specific, as there is a broad range of XAI methods to map requests onto. In order to bridge this gap, we present CoXQL, the first dataset in the NLP domain for user intent recognition in ConvXAI, covering 31 intents, seven of which require filling multiple slots. Subsequently, we enhance an existing parsing approach by incorporating template validations, and conduct an evaluation of several LLMs on CoXQL using different parsing strategies. We conclude that the improved parsing approach (MP+) surpasses the performance of previous approaches. We also discover that intents with multiple slots remain highly challenging for LLMs.

Leonhard Hennig, Philippe Thomas, Sebastian Möller

Relation extraction (RE) is a fundamental task in information extraction, whose extension to multilingual settings has been hindered by the lack of supervised resources comparable in size to large English datasets such as TACRED (Zhang et al., 2017). To address this gap, we introduce the MultiTACRED dataset, covering 12 typologically diverse languages from 9 language families, which is created by machine-translating TACRED instances and automatically projecting their entity annotations. We analyze translation and annotation projection quality, identify error categories, and experimentally evaluate fine-tuned pretrained mono- and multilingual language models in common transfer learning scenarios. Our analyses show that machine translation is a viable strategy to transfer RE instances, with native speakers judging more than 83% of the translated instances to be linguistically and semantically acceptable. We find monolingual RE model performance to be comparable to the English original for many of the target languages, and that multilingual models trained on a combination of English and target language data can outperform their monolingual counterparts. However, we also observe a variety of translation and annotation projection errors, both due to the MT systems and linguistic features of the target languages, such as pronoun-dropping, compounding and inflection, that degrade dataset quality and RE model performance.

Anik Jacobsen, Salar Mohtaj, Sebastian Möller

Vocabulary learning is vital to foreign language learning. Correct and adequate feedback is essential to successful and satisfying vocabulary training. However, many vocabulary and language evaluation systems perform on simple rules and do not account for real-life user learning data. This work introduces Multi-Language Vocabulary Evaluation Data Set (MuLVE), a data set consisting of vocabulary cards and real-life user answers, labeled indicating whether the user answer is correct or incorrect. The data source is user learning data from the Phase6 vocabulary trainer. The data set contains vocabulary questions in German and English, Spanish, and French as target language and is available in four different variations regarding pre-processing and deduplication. We experiment to fine-tune pre-trained BERT language models on the downstream task of vocabulary evaluation with the proposed MuLVE data set. The results provide outstanding results of > 95.5 accuracy and F2-score. The data set is available on the European Language Grid.

Salar Mohtaj, Vera Schmitt, Sebastian Möller

The detection of hate speech online has become an important task, as offensive language such as hurtful, obscene and insulting content can harm marginalized people or groups. This paper presents TU Berlin team experiments and results on the task 1A and 1B of the shared task on hate speech and offensive content identification in Indo-European languages 2021. The success of different Natural Language Processing models is evaluated for the respective subtasks throughout the competition. We tested different models based on recurrent neural networks in word and character levels and transfer learning approaches based on Bert on the provided dataset by the competition. Among the tested models that have been used for the experiments, the transfer learning-based models achieved the best results in both subtasks.

Neslihan Iskender, Oleg Vasilyev, Tim Polzehl et al.

Evaluating large summarization corpora using humans has proven to be expensive from both the organizational and the financial perspective. Therefore, many automatic evaluation metrics have been developed to measure the summarization quality in a fast and reproducible way. However, most of the metrics still rely on humans and need gold standard summaries generated by linguistic experts. Since BLANC does not require golden summaries and supposedly can use any underlying language model, we consider its application to the evaluation of summarization in German. This work demonstrates how to adjust the BLANC metric to a language other than English. We compare BLANC scores with the crowd and expert ratings, as well as with commonly used automatic metrics on a German summarization data set. Our results show that BLANC in German is especially good in evaluating informativeness.

Gabriel Mittags, Sebastian Möller

In this paper, we present a full-reference speech quality prediction model with a deep learning approach. The model determines a feature representation of the reference and the degraded signal through a siamese recurrent convolutional network that shares the weights for both signals as input. The resulting features are then used to align the signals with an attention mechanism and are finally combined to estimate the overall speech quality. The proposed network architecture represents a simple solution for the time-alignment problem that occurs for speech signals transmitted through Voice-Over-IP networks and shows how the clean reference signal can be incorporated into speech quality models that are based on end-to-end trained neural networks.