Martin Potthast

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Aleksandra Piktus, Odunayo Ogundepo, Christopher Akiki et al.

Noticing the urgent need to provide tools for fast and user-friendly qualitative analysis of large-scale textual corpora of the modern NLP, we propose to turn to the mature and well-tested methods from the domain of Information Retrieval (IR) - a research field with a long history of tackling TB-scale document collections. We discuss how Pyserini - a widely used toolkit for reproducible IR research can be integrated with the Hugging Face ecosystem of open-source AI libraries and artifacts. We leverage the existing functionalities of both platforms while proposing novel features further facilitating their integration. Our goal is to give NLP researchers tools that will allow them to develop retrieval-based instrumentation for their data analytics needs with ease and agility. We include a Jupyter Notebook-based walk through the core interoperability features, available on GitHub at https://github.com/huggingface/gaia. We then demonstrate how the ideas we present can be operationalized to create a powerful tool for qualitative data analysis in NLP. We present GAIA Search - a search engine built following previously laid out principles, giving access to four popular large-scale text collections. GAIA serves a dual purpose of illustrating the potential of methodologies we discuss but also as a standalone qualitative analysis tool that can be leveraged by NLP researchers aiming to understand datasets prior to using them in training. GAIA is hosted live on Hugging Face Spaces - https://huggingface.co/spaces/spacerini/gaia.

Christopher Akiki, Odunayo Ogundepo, Aleksandra Piktus et al.

We present Spacerini, a tool that integrates the Pyserini toolkit for reproducible information retrieval research with Hugging Face to enable the seamless construction and deployment of interactive search engines. Spacerini makes state-of-the-art sparse and dense retrieval models more accessible to non-IR practitioners while minimizing deployment effort. This is useful for NLP researchers who want to better understand and validate their research by performing qualitative analyses of training corpora, for IR researchers who want to demonstrate new retrieval models integrated into the growing Pyserini ecosystem, and for third parties reproducing the work of other researchers. Spacerini is open source and includes utilities for loading, preprocessing, indexing, and deploying search engines locally and remotely. We demonstrate a portfolio of 13 search engines created with Spacerini for different use cases.

Matthias Hagen, Maik Fröbe, Artur Jurk et al.

We introduce and study the task of clickbait spoiling: generating a short text that satisfies the curiosity induced by a clickbait post. Clickbait links to a web page and advertises its contents by arousing curiosity instead of providing an informative summary. Our contributions are approaches to classify the type of spoiler needed (i.e., a phrase or a passage), and to generate appropriate spoilers. A large-scale evaluation and error analysis on a new corpus of 5,000 manually spoiled clickbait posts -- the Webis Clickbait Spoiling Corpus 2022 -- shows that our spoiler type classifier achieves an accuracy of 80%, while the question answering model DeBERTa-large outperforms all others in generating spoilers for both types.

Lukas Gienapp, Harrisen Scells, Niklas Deckers et al.

Recent advances in large language models have enabled the development of viable generative retrieval systems. Instead of a traditional document ranking, generative retrieval systems often directly return a grounded generated text as a response to a query. Quantifying the utility of the textual responses is essential for appropriately evaluating such generative ad hoc retrieval. Yet, the established evaluation methodology for ranking-based ad hoc retrieval is not suited for the reliable and reproducible evaluation of generated responses. To lay a foundation for developing new evaluation methods for generative retrieval systems, we survey the relevant literature from the fields of information retrieval and natural language processing, identify search tasks and system architectures in generative retrieval, develop a new user model, and study its operationalization.

Niklas Deckers, Maik Fröbe, Johannes Kiesel et al.

Conditional generative models such as DALL-E and Stable Diffusion generate images based on a user-defined text, the prompt. Finding and refining prompts that produce a desired image has become the art of prompt engineering. Generative models do not provide a built-in retrieval model for a user's information need expressed through prompts. In light of an extensive literature review, we reframe prompt engineering for generative models as interactive text-based retrieval on a novel kind of "infinite index". We apply these insights for the first time in a case study on image generation for game design with an expert. Finally, we envision how active learning may help to guide the retrieval of generated images.

Niklas Deckers, Julia Peters, Martin Potthast

Prompt engineering is still the primary way for users of generative text-to-image models to manipulate generated images in a targeted way. Based on treating the model as a continuous function and by passing gradients between the image space and the prompt embedding space, we propose and analyze a new method to directly manipulate the embedding of a prompt instead of the prompt text. We then derive three practical interaction tools to support users with image generation: (1) Optimization of a metric defined in the image space that measures, for example, the image style. (2) Supporting a user in creative tasks by allowing them to navigate in the image space along a selection of directions of "near" prompt embeddings. (3) Changing the embedding of the prompt to include information that a user has seen in a particular seed but has difficulty describing in the prompt. Compared to prompt engineering, user-driven prompt embedding manipulation enables a more fine-grained, targeted control that integrates a user's intentions. Our user study shows that our methods are considered less tedious and that the resulting images are often preferred.

Shuai Wang, Harrisen Scells, Martin Potthast et al.

Screening prioritisation in medical systematic reviews aims to rank the set of documents retrieved by complex Boolean queries. Prioritising the most important documents ensures that subsequent review steps can be carried out more efficiently and effectively. The current state of the art uses the final title of the review as a query to rank the documents using BERT-based neural rankers. However, the final title is only formulated at the end of the review process, which makes this approach impractical as it relies on ex post facto information. At the time of screening, only a rough working title is available, with which the BERT-based ranker performs significantly worse than with the final title. In this paper, we explore alternative sources of queries for prioritising screening, such as the Boolean query used to retrieve the documents to be screened and queries generated by instruction-based generative large-scale language models such as ChatGPT and Alpaca. Our best approach is not only viable based on the information available at the time of screening, but also has similar effectiveness to the final title.

Magdalena Wolska, Christopher Schröder, Ole Borchardt et al.

We present the first dataset and evaluation results on a newly defined computational task of trigger warning assignment. Labeled corpus data has been compiled from narrative works hosted on Archive of Our Own (AO3), a well-known fanfiction site. In this paper, we focus on the most frequently assigned trigger type--violence--and define a document-level binary classification task of whether or not to assign a violence trigger warning to a fanfiction, exploiting warning labels provided by AO3 authors. SVM and BERT models trained in four evaluation setups on the corpora we compiled yield $F_1$ results ranging from 0.585 to 0.798, proving the violence trigger warning assignment to be a doable, however, non-trivial task.

Shahbaz Syed, Ahmad Dawar Hakimi, Khalid Al-Khatib et al.

Current approaches to automatic summarization of scientific papers generate informative summaries in the form of abstracts. However, abstracts are not intended to show the relationship between a paper and the references cited in it. We propose a new contextualized summarization approach that can generate an informative summary conditioned on a given sentence containing the citation of a reference (a so-called "citance"). This summary outlines the content of the cited paper relevant to the citation location. Thus, our approach extracts and models the citances of a paper, retrieves relevant passages from cited papers, and generates abstractive summaries tailored to each citance. We evaluate our approach using $\textbf{Webis-Context-SciSumm-2023}$, a new dataset containing 540K~computer science papers and 4.6M~citances therein.

Vahid Sadiri Javadi, Martin Potthast, Lucie Flek

When searching for products, the opinions of others play an important role in making informed decisions. Subjective experiences about a product can be a valuable source of information. This is also true in sales conversations, where a customer and a sales assistant exchange facts and opinions about products. However, training an AI for such conversations is complicated by the fact that language models do not possess authentic opinions for their lack of real-world experience. We address this problem by leveraging product reviews as a rich source of product opinions to ground conversational AI in true subjective narratives. With OpinionConv, we develop the first conversational AI for simulating sales conversations. To validate the generated conversations, we conduct several user studies showing that the generated opinions are perceived as realistic. Our assessors also confirm the importance of opinions as an informative basis for decision-making.

Yamen Ajjour, Johannes Kiesel, Benno Stein et al.

Many computational argumentation tasks, like stance classification, are topic-dependent: the effectiveness of approaches to these tasks significantly depends on whether the approaches were trained on arguments from the same topics as those they are tested on. So, which are these topics that researchers train approaches on? This paper contributes the first comprehensive survey of topic coverage, assessing 45 argument corpora. For the assessment, we take the first step towards building an argument topic ontology, consulting three diverse authoritative sources: the World Economic Forum, the Wikipedia list of controversial topics, and Debatepedia. Comparing the topic sets between the authoritative sources and corpora, our analysis shows that the corpora topics-which are mostly those frequently discussed in public online fora - are covered well by the sources. However, other topics from the sources are less extensively covered by the corpora of today, revealing interesting future directions for corpus construction.

Alonso Palomino, Martin Potthast, Khalid Al-Khatib et al.

Most research on natural language processing treats bias as an absolute concept: Based on a (probably complex) algorithmic analysis, a sentence, an article, or a text is classified as biased or not. Given the fact that for humans the question of whether a text is biased can be difficult to answer or is answered contradictory, we ask whether an "absolute bias classification" is a promising goal at all. We see the problem not in the complexity of interpreting language phenomena but in the diversity of sociocultural backgrounds of the readers, which cannot be handled uniformly: To decide whether a text has crossed the proverbial line between non-biased and biased is subjective. By asking "Is text X more [less, equally] biased than text Y?" we propose to analyze a simpler problem, which, by its construction, is rather independent of standpoints, views, or sociocultural aspects. In such a model, bias becomes a preference relation that induces a partial ordering from least biased to most biased texts without requiring a decision on where to draw the line. A prerequisite for this kind of bias model is the ability of humans to perceive relative bias differences in the first place. In our research, we selected a specific type of bias in argumentation, the stance bias, and designed a crowdsourcing study showing that differences in stance bias are perceptible when (light) support is provided through training or visual aid.

Janek Bevendorff, Philipp Sauer, Lukas Gienapp et al.

The rapidly growing volume of scientific publications offers an interesting challenge for research on methods for analyzing the authorship of documents with one or more authors. However, most existing datasets lack scientific documents or the necessary metadata for constructing new experiments and test cases. We introduce SMAuC, a comprehensive, metadata-rich corpus tailored to scientific authorship analysis. Comprising over 3 million publications across various disciplines from over 5 million authors, SMAuC is the largest openly accessible corpus for this purpose. It encompasses scientific texts from humanities and natural sciences, accompanied by extensive, curated metadata, including unambiguous author IDs. SMAuC aims to significantly advance the domain of authorship analysis in scientific texts.

Shahbaz Syed, Dominik Schwabe, Khalid Al-Khatib et al.

Online forums encourage the exchange and discussion of different stances on many topics. Not only do they provide an opportunity to present one's own arguments, but may also gather a broad cross-section of others' arguments. However, the resulting long discussions are difficult to overview. This paper presents a novel unsupervised approach using large language models (LLMs) to generating indicative summaries for long discussions that basically serve as tables of contents. Our approach first clusters argument sentences, generates cluster labels as abstractive summaries, and classifies the generated cluster labels into argumentation frames resulting in a two-level summary. Based on an extensively optimized prompt engineering approach, we evaluate 19~LLMs for generative cluster labeling and frame classification. To evaluate the usefulness of our indicative summaries, we conduct a purpose-driven user study via a new visual interface called Discussion Explorer: It shows that our proposed indicative summaries serve as a convenient navigation tool to explore long discussions.

Marcel Gohsen, Matthias Hagen, Martin Potthast et al.

We propose to use image captions from the Web as a previously underutilized resource for paraphrases (i.e., texts with the same "message") and to create and analyze a corresponding dataset. When an image is reused on the Web, an original caption is often assigned. We hypothesize that different captions for the same image naturally form a set of mutual paraphrases. To demonstrate the suitability of this idea, we analyze captions in the English Wikipedia, where editors frequently relabel the same image for different articles. The paper introduces the underlying mining technology, the resulting Wikipedia-IPC dataset, and compares known paraphrase corpora with respect to their syntactic and semantic paraphrase similarity to our new resource. In this context, we introduce characteristic maps along the two similarity dimensions to identify the style of paraphrases coming from different sources. An annotation study demonstrates the high reliability of the algorithmically determined characteristic maps.

Shahbaz Syed, Dominik Schwabe, Martin Potthast

This paper presents Summary Workbench, a new tool for developing and evaluating text summarization models. New models and evaluation measures can be easily integrated as Docker-based plugins, allowing to examine the quality of their summaries against any input and to evaluate them using various evaluation measures. Visual analyses combining multiple measures provide insights into the models' strengths and weaknesses. The tool is hosted at \url{https://tldr.demo.webis.de} and also supports local deployment for private resources.

Sebastian Heineking, Wilhelm Pertsch, Ines Zelch et al.

Large language models (LLMs) enable a new form of advertising for retrieval-augmented generation (RAG) systems in which organic responses are blended with contextually relevant ads. The prospect of such "generated native ads" has sparked interest in whether they can be detected automatically. Existing datasets, however, do not reflect the diversity of advertising styles discussed in the marketing literature. In this paper, we (1) develop a taxonomy of advertising styles for LLMs, combining the style dimensions of explicitness and type of appeal, (2) simulate that advertisers may attempt to evade detection by changing their advertising style, and (3) evaluate a variety of ad-detection approaches with respect to their robustness under these changes. Expanding previous work on ad detection, we train models that use entity recognition to exactly locate an ad in an LLM response and find them to be both very effective at detecting responses with ads and largely robust to changes in the advertising style. Since ad blocking will be performed on low-resource end-user devices, we include lightweight models like random forests and SVMs in our evaluation. These models, however, are brittle under such changes, highlighting the need for further efficiency-oriented research for a practical approach to blocking of generated ads.

Pierre Achkar, Tim Gollub, Arno Simons et al.

Existing benchmarks for systematic reviewing remain limited either in scale or in disciplinary coverage, with some collections comprising only a modest number of topics and others focusing primarily on biomedical research. We present Webis-SR4ALL-26, a large-scale, cross-disciplinary corpus of 301,871 systematic reviews spanning all scientific fields as covered by OpenAlex. Using a multi-stage pre-processing pipeline, we link reviews to resolved OpenAlex metadata and reference lists and extract, when explicitly reported, structured method artifacts relevant to retrieval and screening. These artifacts include reported search strategies (Boolean queries or keyword lists) that we normalize into executable approximations, as well as reported inclusion and exclusion criteria. Together, these layers support cross-domain benchmarking of retrieval and screening components against review reference lists, training and evaluation of extraction methods for review artifacts, and comparative meta-science analyses of systematic review practices across disciplines and time. To demonstrate one concrete use case, we report large-scale baseline retrieval signals by executing normalized search strategies in OpenAlex and comparing retrieved sets to resolved reference lists. We release the corpus and the pre-processing pipeline, along with code used for extraction validation and the retrieval demonstration.

Pierre Achkar, Tim Gollub, Martin Potthast

The exponential growth of scientific publications has made it increasingly difficult for researchers to stay updated and synthesize knowledge effectively. This paper presents XSum, a modular pipeline for multi-document summarization (MDS) in the scientific domain using Retrieval-Augmented Generation (RAG). The pipeline includes two core components: a question-generation module and an editor module. The question-generation module dynamically generates questions adapted to the input papers, ensuring the retrieval of relevant and accurate information. The editor module synthesizes the retrieved content into coherent and well-structured summaries that adhere to academic standards for proper citation. Evaluated on the SurveySum dataset, XSum demonstrates strong performance, achieving considerable improvements in metrics such as CheckEval, G-Eval and Ref-F1 compared to existing approaches. This work provides a transparent, adaptable framework for scientific summarization with potential applications in a wide range of domains. Code available at https://github.com/webis-de/scolia25-xsum

Janek Bevendorff, Martin Potthast, Benno Stein

Web search and other large-scale web data analytics rely on processing archives of web pages stored in a standardized and efficient format. Since its introduction in 2008, the IIPC's Web ARCive (WARC) format has become the standard format for this purpose. As a list of individually compressed records of HTTP requests and responses, it allows for constant-time random access to all kinds of web data via off-the-shelf open source parsers in many programming languages, such as WARCIO, the de-facto standard for Python. When processing web archives at the terabyte or petabyte scale, however, even small inefficiencies in these tools add up quickly, resulting in hours, days, or even weeks of wasted compute time. Reviewing the basic components of WARCIO and analyzing its bottlenecks, we proceed to build FastWARC, a new high-performance WARC processing library for Python, written in C++/Cython, which yields performance improvements by factors of 1.6-8x.

Christopher Schröder, Lydia Müller, Andreas Niekler et al.

We introduce small-text, an easy-to-use active learning library, which offers pool-based active learning for single- and multi-label text classification in Python. It features numerous pre-implemented state-of-the-art query strategies, including some that leverage the GPU. Standardized interfaces allow the combination of a variety of classifiers, query strategies, and stopping criteria, facilitating a quick mix and match, and enabling a rapid and convenient development of both active learning experiments and applications. With the objective of making various classifiers and query strategies accessible for active learning, small-text integrates several well-known machine learning libraries, namely scikit-learn, PyTorch, and Hugging Face transformers. The latter integrations are optionally installable extensions, so GPUs can be used but are not required. Using this new library, we investigate the performance of the recently published SetFit training paradigm, which we compare to vanilla transformer fine-tuning, finding that it matches the latter in classification accuracy while outperforming it in area under the curve. The library is available under the MIT License at https://github.com/webis-de/small-text, in version 1.3.0 at the time of writing.

Martin Potthast, Tim Gollub, Matthias Hagen et al.

Clickbait has grown to become a nuisance to social media users and social media operators alike. Malicious content publishers misuse social media to manipulate as many users as possible to visit their websites using clickbait messages. Machine learning technology may help to handle this problem, giving rise to automatic clickbait detection. To accelerate progress in this direction, we organized the Clickbait Challenge 2017, a shared task inviting the submission of clickbait detectors for a comparative evaluation. A total of 13 detectors have been submitted, achieving significant improvements over the previous state of the art in terms of detection performance. Also, many of the submitted approaches have been published open source, rendering them reproducible, and a good starting point for newcomers. While the 2017 challenge has passed, we maintain the evaluation system and answer to new registrations in support of the ongoing research on better clickbait detectors.

Martin Potthast, Stefan Heindorf, Hannah Bast

The WSDM Cup 2017 was a data mining challenge held in conjunction with the 10th International Conference on Web Search and Data Mining (WSDM). It addressed key challenges of knowledge bases today: quality assurance and entity search. For quality assurance, we tackle the task of vandalism detection, based on a dataset of more than 82 million user-contributed revisions of the Wikidata knowledge base, all of which annotated with regard to whether or not they are vandalism. For entity search, we tackle the task of triple scoring, using a dataset that comprises relevance scores for triples from type-like relations including occupation and country of citizenship, based on about 10,000 human relevance judgements. For reproducibility sake, participants were asked to submit their software on TIRA, a cloud-based evaluation platform, and they were incentivized to share their approaches open source.

Stefan Heindorf, Martin Potthast, Gregor Engels et al.

We report on the Wikidata vandalism detection task at the WSDM Cup 2017. The task received five submissions for which this paper describes their evaluation and a comparison to state of the art baselines. Unlike previous work, we recast Wikidata vandalism detection as an online learning problem, requiring participant software to predict vandalism in near real-time. The best-performing approach achieves a ROC-AUC of 0.947 at a PR-AUC of 0.458. In particular, this task was organized as a software submission task: to maximize reproducibility as well as to foster future research and development on this task, the participants were asked to submit their working software to the TIRA experimentation platform along with the source code for open source release.

Shuai Wang, Harrisen Scells, Shengyao Zhuang et al.

Systematic reviews are crucial for evidence-based medicine as they comprehensively analyse published research findings on specific questions. Conducting such reviews is often resource- and time-intensive, especially in the screening phase, where abstracts of publications are assessed for inclusion in a review. This study investigates the effectiveness of using zero-shot large language models~(LLMs) for automatic screening. We evaluate the effectiveness of eight different LLMs and investigate a calibration technique that uses a predefined recall threshold to determine whether a publication should be included in a systematic review. Our comprehensive evaluation using five standard test collections shows that instruction fine-tuning plays an important role in screening, that calibration renders LLMs practical for achieving a targeted recall, and that combining both with an ensemble of zero-shot models saves significant screening time compared to state-of-the-art approaches.

Sebastian Schmidt, Ines Zelch, Janek Bevendorff et al.

Conversational search engines such as YouChat and Microsoft Copilot use large language models (LLMs) to generate responses to queries. It is only a small step to also let the same technology insert ads within the generated responses - instead of separately placing ads next to a response. Inserted ads would be reminiscent of native advertising and product placement, both of which are very effective forms of subtle and manipulative advertising. Considering the high computational costs associated with LLMs, for which providers need to develop sustainable business models, users of conversational search engines may very well be confronted with generated native ads in the near future. In this paper, we thus take a first step to investigate whether LLMs can also be used as a countermeasure, i.e., to block generated native ads. We compile the Webis Generated Native Ads 2024 dataset of queries and generated responses with automatically inserted ads, and evaluate whether LLMs or fine-tuned sentence transformers can detect the ads. In our experiments, the investigated LLMs struggle with the task but sentence transformers achieve precision and recall values above 0.9.

Shahbaz Syed, Khalid Al-Khatib, Martin Potthast

This paper presents TL;DR Progress, a new tool for exploring the literature on neural text summarization. It organizes 514~papers based on a comprehensive annotation scheme for text summarization approaches and enables fine-grained, faceted search. Each paper was manually annotated to capture aspects such as evaluation metrics, quality dimensions, learning paradigms, challenges addressed, datasets, and document domains. In addition, a succinct indicative summary is provided for each paper, consisting of automatically extracted contextual factors, issues, and proposed solutions. The tool is available online at https://www.tldr-progress.de, a demo video at https://youtu.be/uCVRGFvXUj8

Matti Wiegmann, Jennifer Rakete, Magdalena Wolska et al.

Trigger warnings are labels that preface documents with sensitive content if this content could be perceived as harmful by certain groups of readers. Since warnings about a document intuitively need to be shown before reading it, authors usually assign trigger warnings at the document level. What parts of their writing prompted them to assign a warning, however, remains unclear. We investigate for the first time the feasibility of identifying the triggering passages of a document, both manually and computationally. We create a dataset of 4,135 English passages, each annotated with one of eight common trigger warnings. In a large-scale evaluation, we then systematically evaluate the effectiveness of fine-tuned and few-shot classifiers, and their generalizability. We find that trigger annotation belongs to the group of subjective annotation tasks in NLP, and that automatic trigger classification remains challenging but feasible.

Janek Bevendorff, Maik Fröbe, André Greiner-Petter et al.

The goal of the PAN workshop is to advance computational stylometry and text forensics via objective and reproducible evaluation. In 2026, we run the following five tasks: (1) Voight-Kampff Generative AI Detection, particularly in mixed and obfuscated authorship scenarios, (2) Text Watermarking, a new task that aims to find new and benchmark the robustness of existing text watermarking schemes, (3) Multi-author Writing Style Analysis, a continued task that aims to find positions of authorship change, (4) Generative Plagiarism Detection, a continued task that targets source retrieval and text alignment between generated text and source documents, and (5) Reasoning Trajectory Detection, a new task that deals with source detection and safety detection of LLM-generated or human-written reasoning trajectories. As in previous years, PAN invites software submissions as easy-to-reproduce Docker containers for most of the tasks. Since PAN 2012, more than 1,100 submissions have been made this way via the TIRA experimentation platform.

Theresa Elstner, Martin Potthast

This paper introduces a new dimension for validating algorithmic decisions about humans by measuring the fidelity of their representations. Representation Fidelity measures if decisions about a person rest on reasonable grounds. We propose to operationalize this notion by measuring the distance between two representations of the same person: (1) an externally prescribed input representation on which the decision is based, and (2) a self-description provided by the human subject of the decision, used solely to validate the input representation. We examine the nature of discrepancies between these representations, how such discrepancies can be quantified, and derive a generic typology of representation mismatches that determine the degree of representation fidelity. We further present the first benchmark for evaluating representation fidelity based on a dataset of loan-granting decisions. Our Loan-Granting Self-Representations Corpus 2025 consists of a large corpus of 30 000 synthetic natural language self-descriptions derived from corresponding representations of applicants in the German Credit Dataset, along with expert annotations of representation mismatches between each pair of representations.

Lukas Gienapp, Christopher Schröder, Stefan Schweter et al.

Large language model development relies on large-scale training corpora, yet most contain data of unclear licensing status, limiting the development of truly open models. This problem is exacerbated for non-English languages, where openly licensed text remains critically scarce. We introduce the German Commons, the largest collection of openly licensed German text to date. It compiles data from 41 sources across seven domains, encompassing legal, scientific, cultural, political, news, economic, and web text. Through systematic sourcing from established data providers with verifiable licensing, it yields 154.56 billion tokens of high-quality text for language model training. Our processing pipeline implements comprehensive quality filtering, deduplication, and text formatting fixes, ensuring consistent quality across heterogeneous text sources. All domain subsets feature licenses of at least CC-BY-SA 4.0 or equivalent, ensuring legal compliance for model training and redistribution. The German Commons therefore addresses the critical gap in openly licensed German pretraining data, and enables the development of truly open German language models. We also release code for corpus construction and data filtering tailored to German language text, rendering the German Commons fully reproducible and extensible.

Tim Hagen, Niklas Deckers, Felix Wolter et al.

Research on causality extraction from text has so far almost entirely neglected counterclaims. Existing causality extraction datasets focus solely on "procausal" claims, i.e., statements that support a relationship. "Concausal" claims, i.e., statements that refute a relationship, are entirely ignored or even accidentally annotated as procausal. We address this shortcoming by developing a new dataset that integrates concausality. Based on an extensive literature review, we first show that concausality is an integral part of causal reasoning on incomplete knowledge. We operationalize this theory in the form of a rigorous guideline for annotation and then augment the Causal News Corpus with concausal statements, obtaining a substantial inter-annotator agreement of Cohen's $κ=0.74$. To demonstrate the importance of integrating concausal statements, we show that models trained without concausal relationships tend to misclassify these as procausal instead. Based on our new dataset, this mistake can be mitigated, enabling transformers to effectively distinguish pro- and concausality.

André Greiner-Petter, Maik Fröbe, Jan Philip Wahle et al.

The generative plagiarism detection task at PAN 2025 aims at identifying automatically generated textual plagiarism in scientific articles and aligning them with their respective sources. We created a novel large-scale dataset of automatically generated plagiarism using three large language models: Llama, DeepSeek-R1, and Mistral. In this task overview paper, we outline the creation of this dataset, summarize and compare the results of all participants and four baselines, and evaluate the results on the last plagiarism detection task from PAN 2015 in order to interpret the robustness of the proposed approaches. We found that the current iteration does not invite a large variety of approaches as naive semantic similarity approaches based on embedding vectors provide promising results of up to 0.8 recall and 0.5 precision. In contrast, most of these approaches underperform significantly on the 2015 dataset, indicating a lack in generalizability.

Marcel Gohsen, Matthias Hagen, Martin Potthast et al.

Since paraphrasing is an ill-defined task, the term "paraphrasing" covers text transformation tasks with different characteristics. Consequently, existing paraphrasing studies have applied quite different (explicit and implicit) criteria as to when a pair of texts is to be considered a paraphrase, all of which amount to postulating a certain level of semantic or lexical similarity. In this paper, we conduct a literature review and propose a taxonomy to organize the 25~identified paraphrasing (sub-)tasks. Using classifiers trained to identify the tasks that a given paraphrasing instance fits, we find that the distributions of task-specific instances in the known paraphrase corpora vary substantially. This means that the use of these corpora, without the respective paraphrase conditions being clearly defined (which is the normal case), must lead to incomparable and misleading results.

Timon Ziegenbein, Shahbaz Syed, Felix Lange et al.

Online discussion moderators must make ad-hoc decisions about whether the contributions of discussion participants are appropriate or should be removed to maintain civility. Existing research on offensive language and the resulting tools cover only one aspect among many involved in such decisions. The question of what is considered appropriate in a controversial discussion has not yet been systematically addressed. In this paper, we operationalize appropriate language in argumentation for the first time. In particular, we model appropriateness through the absence of flaws, grounded in research on argument quality assessment, especially in aspects from rhetoric. From these, we derive a new taxonomy of 14 dimensions that determine inappropriate language in online discussions. Building on three argument quality corpora, we then create a corpus of 2191 arguments annotated for the 14 dimensions. Empirical analyses support that the taxonomy covers the concept of appropriateness comprehensively, showing several plausible correlations with argument quality dimensions. Moreover, results of baseline approaches to assessing appropriateness suggest that all dimensions can be modeled computationally on the corpus.

Fabian Ziegner, Janos Borst, Andreas Niekler et al.

This paper evaluates the viability of using fixed language models for training text classification networks on low-end hardware. We combine language models with a CNN architecture and put together a comprehensive benchmark with 8 datasets covering single-label and multi-label classification of topic, sentiment, and genre. Our observations are distilled into a list of trade-offs, concluding that there are scenarios, where not fine-tuning a language model yields competitive effectiveness at faster training, requiring only a quarter of the memory compared to fine-tuning.

Christopher Akiki, Lukas Gienapp, Martin Potthast

This technical report documents our efforts in addressing the tasks set forth by the 2021 AMoC (Advanced Modelling of Cyber Criminal Careers) Hackathon. Our main contribution is a joint visualisation of semantic and temporal features, generating insight into the supplied data on darknet cybercrime through the aspects of novelty, transience, and resonance, which describe the potential impact a message might have on the overall discourse in darknet communities. All code and data produced by us as part of this hackathon is publicly available.

Lukas Gienapp, Wolfgang Kircheis, Bjarne Sievers et al.

We present the Webis-STEREO-21 dataset, a massive collection of Scientific Text Reuse in Open-access publications. It contains more than 91 million cases of reused text passages found in 4.2 million unique open-access publications. Featuring a high coverage of scientific disciplines and varieties of reuse, as well as comprehensive metadata to contextualize each case, our dataset addresses the most salient shortcomings of previous ones on scientific writing. Webis-STEREO-21 allows for tackling a wide range of research questions from different scientific backgrounds, facilitating both qualitative and quantitative analysis of the phenomenon as well as a first-time grounding on the base rate of text reuse in scientific publications.

Maik Fröbe, Matthias Hagen, Janek Bevendorff et al.

Commercial web search engines employ near-duplicate detection to ensure that users see each relevant result only once, albeit the underlying web crawls typically include (near-)duplicates of many web pages. We revisit the risks and potential of near-duplicates with an information retrieval focus, motivating that current efforts toward an open and independent European web search infrastructure should maintain metadata on duplicate and near-duplicate documents in its index. Near-duplicate detection implemented in an open web search infrastructure should provide a suitable similarity threshold, a difficult choice since identical pages may substantially differ in parts of a page that are irrelevant to searchers (templates, advertisements, etc.). We study this problem by comparing the similarity of pages for five (main) content extraction methods in two studies on the ClueWeb crawls. We find that the full content of pages serves precision-oriented near-duplicate-detection, while main content extraction is more recall-oriented.

Christopher Akiki, Martin Potthast

Masked language models have recently been interpreted as energy-based sequence models that can be generated from using a Metropolis--Hastings sampler. This short paper demonstrates how this can be instrumentalized for constrained composition and explores the poetics implied by such a usage. Our focus on constraints makes it especially apt to understand the generated text through the poetics of the OuLiPo movement.

Kim Breitwieser, Allison Lahnala, Charles Welch et al.

We introduce the problem of proficiency modeling: Given a user's posts on a social media platform, the task is to identify the subset of posts or topics for which the user has some level of proficiency. This enables the filtering and ranking of social media posts on a given topic as per user proficiency. Unlike experts on a given topic, proficient users may not have received formal training and possess years of practical experience, but may be autodidacts, hobbyists, and people with sustained interest, enabling them to make genuine and original contributions to discourse. While predicting whether a user is an expert on a given topic imposes strong constraints on who is a true positive, proficiency modeling implies a graded scoring, relaxing these constraints. Put another way, many active social media users can be assumed to possess, or eventually acquire, some level of proficiency on topics relevant to their community. We tackle proficiency modeling in an unsupervised manner by utilizing user embeddings to model engagement with a given topic, as indicated by a user's preference for authoring related content. We investigate five alternative approaches to model proficiency, ranging from basic ones to an advanced, tailored user modeling approach, applied within two real-world benchmarks for evaluation.

Milad Alshomary, Timon Gurcke, Shahbaz Syed et al.

Key point analysis is the task of extracting a set of concise and high-level statements from a given collection of arguments, representing the gist of these arguments. This paper presents our proposed approach to the Key Point Analysis shared task, collocated with the 8th Workshop on Argument Mining. The approach integrates two complementary components. One component employs contrastive learning via a siamese neural network for matching arguments to key points; the other is a graph-based extractive summarization model for generating key points. In both automatic and manual evaluation, our approach was ranked best among all submissions to the shared task.

Shahbaz Syed, Tariq Yousef, Khalid Al-Khatib et al.

This paper introduces Summary Explorer, a new tool to support the manual inspection of text summarization systems by compiling the outputs of 55~state-of-the-art single document summarization approaches on three benchmark datasets, and visually exploring them during a qualitative assessment. The underlying design of the tool considers three well-known summary quality criteria (coverage, faithfulness, and position bias), encapsulated in a guided assessment based on tailored visualizations. The tool complements existing approaches for locally debugging summarization models and improves upon them. The tool is available at https://tldr.webis.de/

Christopher Schröder, Andreas Niekler, Martin Potthast

Active learning is the iterative construction of a classification model through targeted labeling, enabling significant labeling cost savings. As most research on active learning has been carried out before transformer-based language models ("transformers") became popular, despite its practical importance, comparably few papers have investigated how transformers can be combined with active learning to date. This can be attributed to the fact that using state-of-the-art query strategies for transformers induces a prohibitive runtime overhead, which effectively nullifies, or even outweighs the desired cost savings. For this reason, we revisit uncertainty-based query strategies, which had been largely outperformed before, but are particularly suited in the context of fine-tuning transformers. In an extensive evaluation, we connect transformers to experiments from previous research, assessing their performance on five widely used text classification benchmarks. For active learning with transformers, several other uncertainty-based approaches outperform the well-known prediction entropy query strategy, thereby challenging its status as most popular uncertainty baseline in active learning for text classification.

Shahbaz Syed, Khalid Al-Khatib, Milad Alshomary et al.

The purpose of an argumentative text is to support a certain conclusion. Yet, they are often omitted, expecting readers to infer them rather. While appropriate when reading an individual text, this rhetorical device limits accessibility when browsing many texts (e.g., on a search engine or on social media). In these scenarios, an explicit conclusion makes for a good candidate summary of an argumentative text. This is especially true if the conclusion is informative, emphasizing specific concepts from the text. With this paper we introduce the task of generating informative conclusions: First, Webis-ConcluGen-21 is compiled, a large-scale corpus of 136,996 samples of argumentative texts and their conclusions. Second, two paradigms for conclusion generation are investigated; one extractive, the other abstractive in nature. The latter exploits argumentative knowledge that augment the data via control codes and finetuning the BART model on several subsets of the corpus. Third, insights are provided into the suitability of our corpus for the task, the differences between the two generation paradigms, the trade-off between informativeness and conciseness, and the impact of encoding argumentative knowledge. The corpus, code, and the trained models are publicly available.

Milad Alshomary, Shahbaz Syed, Arkajit Dhar et al.

Text generation has received a lot of attention in computational argumentation research as of recent. A particularly challenging task is the generation of counter-arguments. So far, approaches primarily focus on rebutting a given conclusion, yet other ways to counter an argument exist. In this work, we go beyond previous research by exploring argument undermining, that is, countering an argument by attacking one of its premises. We hypothesize that identifying the argument's weak premises is key to effective countering. Accordingly, we propose a pipeline approach that first assesses the premises' strength and then generates a counter-argument targeting the weak ones. On the one hand, both manual and automatic evaluation proves the importance of identifying weak premises in counter-argument generation. On the other hand, when considering correctness and content richness, human annotators favored our approach over state-of-the-art counter-argument generation.

Sebastian Bischoff, Niklas Deckers, Marcel Schliebs et al.

The prerequisite of many approaches to authorship analysis is a representation of writing style. But despite decades of research, it still remains unclear to what extent commonly used and widely accepted representations like character trigram frequencies actually represent an author's writing style, in contrast to more domain-specific style components or even topic. We address this shortcoming for the first time in a novel experimental setup of fixed authors but swapped domains between training and testing. With this setup, we reveal that approaches using character trigram features are highly susceptible to favor domain information when applied without attention to domains, suffering drops of up to 55.4 percentage points in classification accuracy under domain swapping. We further propose a new remedy based on domain-adversarial learning and compare it to ones from the literature based on heuristic rules. Both can work well, reducing accuracy losses under domain swapping to 3.6% and 3.9%, respectively.

Wei-Fan Chen, Shahbaz Syed, Benno Stein et al.

An abstractive snippet is an originally created piece of text to summarize a web page on a search engine results page. Compared to the conventional extractive snippets, which are generated by extracting phrases and sentences verbatim from a web page, abstractive snippets circumvent copyright issues; even more interesting is the fact that they open the door for personalization. Abstractive snippets have been evaluated as equally powerful in terms of user acceptance and expressiveness---but the key question remains: Can abstractive snippets be automatically generated with sufficient quality? This paper introduces a new approach to abstractive snippet generation: We identify the first two large-scale sources for distant supervision, namely anchor contexts and web directories. By mining the entire ClueWeb09 and ClueWeb12 for anchor contexts and by utilizing the DMOZ Open Directory Project, we compile the Webis Abstractive Snippet Corpus 2020, comprising more than 3.5 million triples of the form $\langle$query, snippet, document$\rangle$ as training examples, where the snippet is either an anchor context or a web directory description in lieu of a genuine query-biased abstractive snippet of the web document. We propose a bidirectional abstractive snippet generation model and assess the quality of both our corpus and the generated abstractive snippets with standard measures, crowdsourcing, and in comparison to the state of the art. The evaluation shows that our novel data sources along with the proposed model allow for producing usable query-biased abstractive snippets while minimizing text reuse.

Krisztian Balog, Lucie Flekova, Matthias Hagen et al.

This paper discusses the potential for creating academic resources (tools, data, and evaluation approaches) to support research in conversational search, by focusing on realistic information needs and conversational interactions. Specifically, we propose to develop and operate a prototype conversational search system for scholarly activities. This Scholarly Conversational Assistant would serve as a useful tool, a means to create datasets, and a platform for running evaluation challenges by groups across the community. This article results from discussions of a working group at Dagstuhl Seminar 19461 on Conversational Search.