Accepted submissions

At RepRisk, a leading ESG data provider, we harness the power of Natural Language Processing (NLP) and Machine Learning (ML) to extract critical ESG insights from news articles worldwide. Our advanced ML pipeline processes vast amounts of unstructured text to identify key ESG-related events, assess company involvement, and generate structured, actionable insights. In this live demo, attendees will have the opportunity to select news articles of their choice, which will then be processed in real time through RepRisk’s multi-stage ML pipeline. The system will extract ESG-relevant information, classify incidents, map companies to their identifiers, and generate predictive insights, all displayed dynamically in our interactive UI. Each prediction and extracted entity will be clickable, allowing users to explore related incidents and navigate company profiles directly on RepRisk’s platform. This session will not only showcase the sophistication of RepRisk’s NLP-driven ESG analytics but also allow participants to experience firsthand the accuracy and depth of our AI models in transforming raw news into meaningful ESG intelligence.

With rapid advancements in NLP, practitioners face a pivotal choice between Retrieval-Augmented Generation (RAG) and Long-Context Large Language Models (LLMs). Both approaches promise efficient processing of extensive textual data, yet their applicability varies significantly based on the use case. RAG leverages external knowledge retrieval, excelling in precision and adaptability, whereas Long-Context LLMs offer robust context retention beneficial for summarization and complex dialogues. This presentation provides practical insights into selecting the optimal approach, highlighting performance, cost efficiency, complexity of implementation, and real-world benchmarks. Attendees will leave equipped with a clear framework for evaluating and implementing these cutting-edge NLP strategies.

The project aims at bridging the semantic gap between user queries, the company’s information systems (e.g., databases, customer relationship management systems, and software APIs) and its knowledge management infrastructure (e.g., domain ontologies, structured knowledge in databases and knowledge graphs, and corporate knowledge repositories). LLMs interpret user queries and translate them to the corresponding concepts in the knowledge graph. This enables processing of queries using symbolic AI which ensures very high reliability, since reasoning within symbolic AI components is deterministic.

Symbolic reasoning upon domain-specific knowledge graphs also explains query results and decisions based on (human understandable) concepts within these graphs, ensuring that system decisions are traceable and explainable to non-computer scientists.

Once completed, SYMBOL will support clients in the wealth management industry by
1. navigating and aiding users through complicated regulatory requirements;
2. generating regulatory reports and analyses on demand, helping wealth management firms respond to audits, risk assessments, and evolving compliance mandates; and
3. extracting deep business insights from their data, enabling proactive decision-making based on structured, regulatory-compliant intelligence.

By allowing non-technical users to interact with SYMBOL the project will eliminate barriers to data-driven decision-making, ensuring that compliance officers, portfolio managers, and auditors can extract the necessary information when it matters most such as to support high-stake decision-making processes and on-site regulatory reviews.

Although the wealth management use case is central to the SYMBOL project, we aim at adapting the developed neurosymbolic AI components to other high-stake environments in domains such as finance, medicine, and law.

Despite the growing popularity of using large language models (LLMs) such as ChatGPT for qualitative content analysis, current approaches often rely on overly simplistic prompting strategies. As Mayring (2025) highlights in his field report, the primary issue lies in the inadequacy of many prompt designs.
General requests such as “Do a qualitative content analysis according to Mayring” result in superficial outputs that lack adherence to the step-by-step methodology central to rigorous qualitative analysis. Even with more structured prompts, ChatGPT frequently fails to follow essential procedural elements such as inductive category formation, abstraction level calibration, and coder agreement testing. The outcomes typically resemble rough summaries rather than methodologically grounded categorizations, leading to what Mayring refers to as “rough approximations and gross errors” (Mayring, 2025, p. 5).
These limitations are further exacerbated when applied to larger datasets or when theoretical grounding is required. Prompt-based approaches, even when refined, struggle to maintain the iterative and transparent logic required by Mayring’s qualitative content analysis. As a result, the reliability and reproducibility of the outcomes remain questionable. To overcome these limitations, we propose the integration of multi-agent systems that mirror the structured, procedural logic of Mayring’s methodology (Mayring, 2022). Rather than relying on single, monolithic prompts, a system of specialized LLM agents can be deployed, with each agent responsible for a specific task aligned with Mayring’s distinct techniques (e.g., inductive category formation, summarization, explication).
For instance, individual agents can be designated to handle:
• Category definition
• Calibration of abstraction levels
• Identification and validation of coding units
• Verification of coder agreement
Crucially, these agents would operate under human oversight, ensuring interpretive validity and adherence to ethical and methodological standards.
This agent-based architecture is inspired by recent advances in the design of LLM agents (Guo et al., 2024), where specialized agents collaborate under human guidance to plan, execute, and optimize complex experiments. By adapting this collaborative structure to qualitative content analysis, we can reflect Mayring’s method not only in output, but also in process – step-by-step, transparent, and verifiable.
This hybrid system presents a promising way to elevate current practices from surface-level approximations toward structured, scientifically grounded qualitative content analysis.

In the dynamic landscape of machine learning, optimizing model performance relies on a thorough analysis of feature spaces. This study introduces an innovative framework designed to refine and improve machine learning models through meticulous feature analysis. We explore the correlations between the features and the model predictions to identify areas of improvement and potential feature gaps. By targeting misclassified samples, we uncover patterns that may elude conventional models, enabling us to propose targeted adjustments in model architecture and feature engineering.

We leverage SHAP (SHapley Additive exPlanations) analysis together with unsupervised learning techniques, such as PCA or t-SNE, to reveal nonlinear relationships and natural data groupings based on feature vectors. Furthermore, we employ K-Nearest Neighbors (KNN) and cluster analysis to detect annotation errors by identifying homogeneous feature vector clusters and to enhance data integrity by flagging potential misannotations for review.

We applied the proposed framework to an entity matching project, where text-based features are compared between different documents to identify matching pairs. This approach allowed us to identify the limitations of our models and guide the creation of new features specifically designed to distinguish between samples with very similar feature vectors but different annotations. Clustering analysis also helped identify and correct erroneous annotations in the dataset, resulting in a significant improvement in model performance.

Our framework not only identifies and corrects model weaknesses, but also proposes strategies to build more robust, accurate, and interpretable models, ultimately advancing their applicability in real-world scenarios. Although tailored for NLP challenges, the framework is also applicable beyond NLP for any feature-based ML model. This study serves as a guide for data scientists and machine learning practitioners seeking to optimize model performance through comprehensive feature analysis and enhancement techniques.

Methods: History of present illness (HPI) as well as diagnoses and procedures (DXL) were extracted from patient records for three clinical scenarios: planned or elective chemotherapy (PEC), acute coronary syndrome (ACS), i.e. myocardial infarction, and acute lower back pain (ALBP). Three medium-sized LLMs, i.e. Mixtral 8x7B, Mixtral 8x22B and Llama 3.1 70B were prompted to generate discharge summaries based on HPI and DXL inputs. Three approaches of generating discharge notes were compared: prompting without examples (zero-shot approach), In-Context Learning (ICL) which utilized four examples of triplets consisting of HPI, DXL, and human-written discharge summaries (4-shot approach), and supervised fine-tuning (SFT) on Mixtral 8x7B with specific training sets (N_PEC-train = 1028, N_ACS-train = 1920, N_ALBP-train = 1494). For evaluation, five simple and five complex samples were extracted for each of the three scenarios, resulting in 30 triplets of HPI, DXL and human-written discharge summaries. Using the different LLMs and different prompting approaches, this results in a total of 150 generated discharge summaries, which were assessed via BLEU, ROUGE-L, and BERTScore metrics. In addition, a blinded panel of 6 specialists in internal medicine assessed the 150 summaries with a modified Physician Documentation Quality Instrument (mPDQI-9) consisting of nine items rated on a 5-point Likert scale, with higher scores indicating better performance.

Results: Our findings indicate that both ICL and SFT enhance the quality of the generated discharge summaries compared to the zero-shot approach. The improvements were most notable for SFT in the PEC scenario (median 32 vs 28 out of 45). In general, generated reports for simpler cases received higher human ratings compared to more complex cases, particularly for the PEC scenario, but hallucination was a problem. When benchmarked against their respective ground truth discharge summaries, we achieved a BERTScore of 0.75, a BLEU score of 0.18, and a ROUGE-L score of 0.35 for the simple cases with SFT, which was the best approach. Overall, zero-shot Mixtral 8x7B, 8x22B, and Llama 3.1 70B demonstrated similar performance based on the expert panel’s assessment.

Conclusion: Our findings demonstrate that LLMs create medical discharge summaries for simple clinical scenarios with acceptable quality, but struggle with more complex cases. This highlights the need for accurate prompting, technical solutions to hallucination, and high quality input data in training models. Addressing these challenges would alleviate much of the administrative burden for physicians, especially those in training, which currently spend only 30 % of their workdays directly with patients. This approach has the potential of enhancing workflow efficiency, reducing clinician burnout, and improving patient care transitions, ultimately helping hospitals manage increasing workloads more effectively.

At RepRisk we maintain a large database of news incidents coupled to companies accused of ESG issues violations. This data is used by asset managers, investors or institutions to make informed decisions about entities they are interested in. This data is multilingual, with a long history, enriched by new companies created every day and combines both human analysis and machine learning. Our pipeline addresses the complex challenge of associating news documents with corporate entities, a critical need for clients who rely on accurate, timely data. Faced with the absence of a single source of truth, duplicate records, and disparate naming conventions—where legal names, journalistic aliases, and outdated entries coexist—we developed a robust, multi-faceted solution. We leverage our unique dataset where texts are associated with IDs to identify entities corresponding to both legal names and the commonly-used variants as well as custom transliteration routines to address our varied multilingual data. Our approach integrates advanced entity extraction with candidate generation, recall-based linking for candidate selection, and precision-based verification for optimal results. To enhance multilingual performance, we incorporate all this contextual information into cutting-edge transformer models combined with large language models through tailored prompting. This comprehensive system not only resolves data inconsistencies across heterogeneous sources but also sets a new benchmark for technical rigor and operational efficiency in real-time news content processing.

We developed and deployed Life Guide Scout, a GenAI-powered underwriting assistant to more than 3,000 Life & Health underwriters worldwide. The system uses a Retrieval-Augmented Generation (RAG) setup to integrate Swiss Re’s proprietary underwriting guidance and medical knowledge, thereby speeding up information retrieval. Fully integrated in the underwriter’s workflow, it enables intuitive, efficient and trustworthy interactions with highly specific knowledge.

The real challenge of productively deploying an LLM-based system lies in assessing its performance over time and across versions. We present a comprehensive evaluation methodology based on synthetically generated data. For instance, on the specific task of mentioning the right underwriting rating in Life Guide Scout’s answer, we achieve an end-to-end 80% hierarchical recall, a metric particularly suited to our problem. We also examine the various failure modes and suggest mitigations. In addition to this programmatic approach, human feedback played a crucial role in refining Life Guide Scout through three key approaches: expert evaluations for structured assessments, user feedback within the application for real-time insights, and surveys and interviews to gauge adoption trends. This multi-layered approach ensured continuous iteration, improving accuracy, usability, and overall user satisfaction.

Developing GenAI applications also requires a blend of new and traditional engineering practices. We share insights on prompt management techniques, structured outputs, and strategies for handling frequent LLM updates, including new models and versions. While LLMs introduce novel challenges, traditional software engineering practices remain critical. We detail unit, integration, and regression testing methods, which are essential for iterating on an LLM-centric application in a production environment.

Given the risks of incorrect AI-generated outputs in an insurance context, we implemented pre- and post-processing techniques to reduce inaccuracies by leveraging the specificities of our problem. Enhancing transparency, we introduced source anchoring using IDs, which not only links references but also highlights the exact section or phrase within the source that the LLM used to generate its response. This improves user trust and allows for quick verification of information.
GenAI introduces new risks related to safety and security. We conducted extensive adversarial attacks, or Red Teaming, on Life Guide Scout to uncover vulnerabilities and proactively mitigate risks, ensuring alignment with responsible AI principles. By stress-testing the system against adversarial scenarios, we strengthened safeguards, improving both security and reliability.

Finally, we share our approach to developing conversational memory within a RAG setup while managing token usage effectively. Maintaining context across interactions enhances the user experience but presents engineering trade-offs that we addressed through targeted optimizations.

Automated Essay Scoring (AES) demands models that can evaluate student essays with human-like consistency while maintaining computational efficiency. Although standard transformer models like DeBERTa can achieve strong performance, they are often resource-intensive and constrained by a 512-token input limit, which can lead to truncated context in longer essays. This limitation hinders the model’s ability to capture argument flow, coherence, and global structure, which are crucial for accurate scoring. Additionally, many existing approaches also rely on prompt engineering, further restricting practical application.
To address these challenges, we present a novel prompt-free approach using SetFit for AES that achieves competitive accuracy while significantly reducing computational overhead. Unlike traditional transformer-based models, SetFit enables sentence transformer fine-tuning with contrastive learning, making it suitable for essay scoring even in low-data regimes. We extend Longformer into a sentence transformer, allowing SetFit to process full-length essays within a 4096-token window. This overcomes the 512-token restriction of traditional transformers, ensuring that the model can evaluate entire essays rather than isolated sections.
Our approach integrates SetFit’s lightweight contrastive learning to optimize sentence embeddings, enabling efficient, prompt-free fine-tuning with significantly lower GPU requirements compared to full transformer fine-tuning. By using contrastive learning, our model learns rich representations of essay quality without needing large-scale labeled datasets. We train our model on AES-specific datasets, so it captures the complexity of essay evaluation metrics such as coherence, grammar, and argumentation strength.
Our fine-tuned model has been publicly released on Hugging Face, where it has already gained over 6,000 downloads, reflecting strong community interest in efficient, long-text NLP solutions. Our results show that SetFit with an extended Longformer sentence transformer achieves competitive accuracy and offers a cost-effective, scalable alternative to resource-heavy methods. Beyond essay scoring, our approach is applicable to other long-form NLP tasks, including legal document analysis, research paper assessment, and educational content evaluation, providing a cost-effective alternative to computationally expensive transformer-based models.

Over the past three years, LLMs have impressed the world with their powerful capabilities to understand and generate human language. As with most technological innovations, it takes time and significant efforts to build successful productive solutions (and not just prototypes) around LLMs which generate real-world business revenues and ultimately render the upfront investment profitable. Companies and organizations around the world are still at a fairly early stage in exploring how to best leverage LLMs productively, but some trends and best practices are emerging as to which types of use case are worthwhile to pursue and how LLM-based solutions should be built.

In this talk, we will pick up on some of these trends and best practices and sketch out what we believe most commercial GenAI projects might look like in a few years from today. In doing so, we take the perspective of projects which do not have unlimited budget. We will look more closely at some common challenges and ways to mitigate them. We will give some examples from real-life projects that focus on automating business processes.

The widespread availability of wearable devices and sports monitoring applications has enabled individuals, including those with Type 1 diabetes (T1D), to easier track their physical activity. Given the importance of exercise in managing T1D, personalized feedback can play a critical role in optimizing workout routines while mitigating the risks of hypo- and hyper-glycemia. This study explores the feasibility of leveraging Natural Language Processing (NLP) models to generate tailored messages based on an individual’s activity data and expert inputs.
In particular, we consider two types of workouts: with negative-outcome (i.e., where the individual’s glucose level went out of range, further subdivided into hypo- and hyper-glycemia) and with positive-outcome (i.e., where the individual’s glucose level remained within the range). Negative-outcome workouts require a behavior change, and messages should advise the individual on how to adjust. Conversely, if the outcome is positive, the individual should be encouraged to maintain their current behavior.
Driven by the potential future goal to integrate our approach into an app that prioritizes user privacy and transparency, we focus on evaluating several open-source NLP models to determine their effectiveness in producing high-quality, personalized messages. Furthermore, we consider two types of prompts. First, the simpler one, referred to as the observable prompt type, is based on the combination of a behavioral pattern (i.e., a more precise description of the out-of-range behavior selected from a pre-defined set of possibilities) and its accompanying expert-provided information. Second, the more complex one, referred to as the actionable prompt type, adds to the observable prompt type personalized actionable variables (derived by the underlying ML model1). Additionally, we implemented prompt refinement strategies to enhance message quality and safety, though further research is needed to optimize these approaches. We perform quantitative and qualitative evaluation of prompts. For example, within the qualitative evaluation we focused on prompt adherence, correctness, level of detail, emotional tone, and medical content comprehension.
Contrary to expectations, our results reveal that models fine-tuned on medical data or those excelling in medical benchmarks do not necessarily generate superior messages for this application. Among the tested models, Mistral-7B-Instruct-v0.3 demonstrated the most promising performance, while others, including Starling-LM-7B-beta, gemma-2-2b-it, Llama-3.2-3B-Instruct, and JSL-MedPhi2-2.7B, yielded suboptimal outcomes.
This work serves as a proof of concept for the feasibility of using personalized NLP-driven messages in diabetes management, with the ultimate goal of driving behavior change. However, we acknowledge the limitations of our study, particularly regarding dataset size and the narrow scope of actionable variables considered. Future research should focus on expanding the dataset and refining both model selection and prompt engineering techniques to improve the reliability and effectiveness of NLP-generated guidance in diabetes care.

1 Details about the model are not reported here due to space constraints and the focus on NLP methods.

In this study, we investigate the potential of GZIP-KNN, a recently proposed lightweight method, for detecting AI-generated text, specifically content generated by ChatGPT. We evaluate GZIP-KNN’s predictive performance, training time, inference time, and memory footprint in comparison to logistic regression, eXtreme Gradient Boosting (XGB), and Gated Recurrent Unit (GRU). As our focus is on low-resource approaches, we do not consider pre-trained models.

Using five open datasets from different domains, we conduct two experiments. The first examines the trade-off between predictive performance and computational complexity in an in-domain setting. The second assesses performance under data and inference time constraints in an out-of-domain scenario. Experimental results indicate that GZIP-KNN achieves strong predictive accuracy, outperforming alternative methods even with limited data. However, its higher inference time limits its applicability in scenarios requiring rapid decision-making. Nonetheless, findings suggest that GZIP-KNN can match the performance of other methods when trained on only a small subset of available data in an out-of-domain context.

In the fast-paced environment of news production, ensuring editorial quality while maintaining tight publication schedules remains a significant challenge. We present ErrorCatcher, an LLM-powered editorial quality assurance system developed at Reuters News to help journalists identify and correct both syntactic errors and style guide violations before publication.
ErrorCatcher leverages a suite of specialized prompts designed in collaboration with experienced journalists to analyze news articles across multiple dimensions: grammatical correctness, adherence to in-house style guidelines, consistency in terminology, and in-story factual coherence. The system offers targeted feedback, identifying errors and suggesting corrections which reference relevant style guidelines.
Our system addresses a significant challenge in integrating sizable organizational style guidelines by developing a hierarchical approach that categorizes style elements by priority and relevance, enabling the system to focus on the most pertinent rules, lowering costs and improving response coherence.
We evaluate several leading LLMs as the backbone of our system, revealing that LLMs optimized for complex reasoning demonstrate superior capabilities in identifying subtle style inconsistencies and nuanced grammatical issues across journalistic content. Our preliminary deployment of ErrorCatcher as an internal tool has shown promising results, with journalists reporting improved workflow efficiency and heightened awareness of recurring style issues.
We outline our approach to developing ErrorCatcher, discuss the technical and practical challenges of implementing AI editorial assistance in a global news environment, and share our progress in extending ErrorCatcher with additional capabilities while evaluating its performance.

Pre-trained Large Language Models (LLMs) can be leveraged to answer domain-specific questions using prompt engineering and retrieval-augmented generation. However, ensuring the trustworthiness of such systems remains a critical challenge. In this work, we propose a general methodology to evaluate the reliability of LLM-based modules by constructing large, representative, and unbiased datasets of questions and answers through automated variation generation. We define key metrics to assess correctness, robustness, and explainability. We apply our approach to a real-world use case in which a smart wheelchair provides answers about its functioning, exploiting RAG with ChatGPT as the underlying LLM. Our experimental results, based on a dataset of over 1,000 questions, reveal that while correctness and robustness are generally strong, the model struggles with open-ended questions, negations, and idiomatic expressions, with explainability being the most challenging aspect. Beyond the specific results (which heavily depend also on dataset at hand), we emphasize the generalizability of our methodology, which can be adapted to various domains. We are currently working on automating the evaluation pipeline to reduce reliance on human assessment and extending the methodology for real-time monitoring of LLM responses.

The RaDisli (“Radikale Diskurse Lichten”) project introduces an automated, dynamic monitoring tool that systematically collects and analyzes extremist content from Telegram channels. The prototype leverages advanced NLP techniques to provide real-time analytical insights into radical discourse, specifically supporting monitoring and analytical efforts within social work.

Key features include individual filtering by channels, time range, and search terms. Each message undergoes automated classification into categories: “hate speech,” “toxicity,” “threat,” and “extremism.” The Streamlit-based web application visualizes activity patterns through heat maps, highlighting peak communication times. Word clouds summarize frequently used terms per channel or group, and topic modeling via Latent Dirichlet Allocation (LDA) provides insights into prevalent themes within the discourse. Additionally, a network graph visualizes interconnections between channels based on forwarded messages, highlighting influential hubs and dissemination pathways.

Evaluations of the prototype indicate that the application significantly enhances analytical capabilities. Users report that the streamlined, image-free interface reduces emotional stress and allows for a more objective, neutral assessment of extremist content compared to direct interaction within Telegram. To date, the system has processed and analyzed over 3.1 million messages from more than 180 channels, demonstrating robust scalability and performance.

CV of the Team

• Dirk Baier: Head of the Institute for Delinquency and Crime Prevention at ZHAW, specializing in youth crime and extremism research.
• Judith Bühler: Lecturer at ZHAW School of Social Work, expert in extremism prevention, digital transformation, and social work.
• Pius von Däniken: NLP researcher at ZHAW School of Engineering, experienced in misinformation analysis, NLP evaluation, and social media analytics.
• Lars Schmid: Research assistant in NLP at ZHAW, pursuing an MSc in Data Science, focusing on automated text analysis and NLP techniques.

Intended Audience
This presentation targets developers and researchers working in related fields, particularly those interested in social media analysis, NLP applications in social sciences, and extremism research.

This paper presents a new approach to fine-tuning OpenAI’s Whisper model for low-resource languages by introducing a novel data generation method that converts sentence-level data into a long-form corpus, using Swiss German as a case study. Non-sentence-level data, which could improve the performance of long-form audio, is difficult to obtain and often restricted by copyright laws. Our method bridges this gap by transforming more accessible sentence-level data into a format that preserves the model’s ability to handle long-form audio and perform segmentation without requiring non-sentence-level data. Our data generation process improves performance in several real-world applications and leads to the development of a new state-of-the-art speech-to-text (STT) model for Swiss German. We compare our model with a non-fine-tuned Whisper and previous state-of-the-art Swiss German STT models, where our new model achieves higher BLEU scores. Our results also indicate that the proposed method is adaptable to other low-resource languages, supported by written guidance and code that allows the creation of fine-tuned Whisper models, which keep segmentation capabilities and allow the transcription of longer audio files using only sentence-level data with high quality.

This paper explores the idea of using phonemes as a textual representation within a conventional multilingual simultaneous speech-to-speech translation pipeline, as opposed to the traditional reliance on textbased language representations. To investigate this, we trained an open-source sequence-to-sequence model on the WMT17 dataset in two formats: one using standard textual representation and the other employing phonemic representation. The performance of both approaches was assessed using the BLEU metric. Our findings phonemic approach with comparable quality, offering insights into its potential to enhance speech-to-speech translation quality and adaptability.

This paper investigates the extent to which pretrained German BERT encodes knowledge of noun compound semantics. We comprehensively vary combinations of target tokens, layers, and cased vs. uncased models, and evaluate them by predicting the compositionality of 868 gold standard compounds. While our strongest results lag behind equivalent prior work on English — suggesting a more difficult nature of the task in German — we find comparable representational patterns within the transformer architecture, with compositionality information most recoverable in the early layers.

We explore the capability of four open-weight large language models (LLMs) in argumentation mining (AM). We conduct experiments on three different corpora; persuasive essays (PE), argumentative microtexts (AMT) Part 1 and Part 2, based on two argumentation mining subtasks: (i) argument component type classification (ACTC), and (ii) argumentative relation classification (ARC). This work aims to assess the argumentation capability of open-weight LLMs, including Mistral 7B, Mixtral 8x7B, LLaMA2 7B and LLaMA3 8B in both, zero-shot and few-shot scenarios. Our results demonstrate that open-sourced LLMs can effectively tackle argumentation mining subtasks, with context-aware prompting improving relation classification performance, though the models’ effectiveness varies across different argumentation patterns and corpus types, suggesting potential for specialized adaptation in future argumentation systems. Our analysis advances the assessment of computational argumentation capabilities in open-weight LLMs and provides a foundation for future research.

Modern approaches for table recognition consist of an encoder for feature extraction and one or more decoders for structure recognition and cell box detection. Recent advancements in this field have introduced Transformers, initially in the decoders and more recently in the encoder as well. While these improvements have enhanced performance, they have also increased model complexity, requiring larger datasets for training, a pre-training step, and higher inference time.
In this paper, we explore SLANet, a lightweight Transformer-free model originally trained on PubTabNet. By training it on the SynthTabNet dataset, we improve its S-TEDS score by 0.47%, we named this model SLANet-1M. Additionally, SLANet-1M achieves an S-TEDS score on PubTabNet that is only 0.41% lower than the state-of-the-art UniTable Large, while using nearly 14 times fewer parameters. On SynthTabNet, its S-TEDS score is just 0.03% below UniTable Large. Moreover, SLANet-1M outperforms large vision-language models (VLMs) such as GPT-4o, GPT-4-turbo with vision, and LLaVA in this specific task.
SLANet-1M is also more efficient during inference, offering faster processing and CPU-friendly execution, eliminating the need for a GPU.

This paper introduces GENAIVC, a version control system for content creation using generative AI. As generative AI models become integral to creative workflows, managing iterative changes, branching, and merging of content is challenging. Current version control systems are not designed for these workflows, which involve multiple AI assistants exchanging text, images, or other artifacts. In this paper, we identify the core requirements for such a system and show how GENAIVC meets them. Our system provides full traceability and versioning of both artifacts and conversation states, allowing seamless integration of multiple AI assistants into creative workflows.

We present 20min-XD (20 Minuten cross-lingual document-level), a French-German, document-level comparable corpus of news articles, sourced from the Swiss online news outlet 20 Minuten/20 minutes. Our dataset comprises around 15,000 article pairs spanning 2015 to 2024, automatically aligned based on semantic similarity. We detail the data collection process and alignment methodology. Furthermore, we provide a qualitative and quantitative analysis of the corpus. The resulting dataset exhibits a broad spectrum of cross-lingual similarity, ranging from near-translations to loosely related articles, making it valuable for various NLP applications and broad linguistically motivated studies. We publicly release the dataset and alignment scripts.

Guided dictionary-based content analysis has emerged as an effective way to process large-scale text corpora. However, the reproducibility of these analysis efforts is often not guaranteed. We propose a human-in-the-loop approach to dictionary-based content analysis, where users get control over the training pipeline by chunking the process into four distinct steps. Compared to end-to-end and/or purely LLM-based approaches, where the learning and inference process is difficult to understand and, hence, to steer, we advocate for a human-in-the-loop methodology. We demonstrate how, through minimal labeling and intervention, the user can guide the process and achieve competitive performance.

Rising pressure for corporate sustainability has intensified greenwashing as companies seek competitive advantages. In the absence of a standardized definition and labeled data, detecting greenwashing remains challenging. Recent research has explored various machine learning techniques to address this issue. In this work, we employ selected ML tools to uncover greenwashing cues in ESG reports, focusing on both language and content. We narrow our content analysis to traffic-related emissions—a particularly challenging area for verification. Our framework comprises multiple pipelines: a language analysis module that detects overly positive and vague language, and a claim verification pipeline that extracts claims and assesses whether they are supported by evidence (i.e., internal proof statements or corroborative external data).

Many individuals struggle with informal interactions like small-talk, which are vital in daily and professional settings. We introduce a conversational agent that combines a state-based interaction model with a social behaviour regulation (SBR) layer to provide structured coaching and real-time conversational modulation. The agent dynamically addresses issues such as oversharing or topic divergence and triggers coaching interventions based on user disengagement or inappropriateness. An exploratory study with neurodivergent-focused educators suggests the system’s potential to foster socially appropriate communication. Our work shows how modular prompt orchestration can enhance both adaptability and the pedagogical value of conversational agents.

This paper describes a new version of the Swiss Parliaments Corpus, an effort to transform and improve NLP datasets in the Swiss context by moving fr​om sentence level to long-form data. In contrast to the original SPC, our approach introduces substantial changes inspired by recent findings in speech-to-text research that highlight the advantages of long-form data. We first transcribe all the audio using Whisper Large-v3 at elevated computational settings, and th​en apply a three-step correction process using GPT-4o – first refining the transcription, th​en evaluating the final transcription and th​en filtering the data, based on the predicted BLEU and the evaluation, to produce the final output. Context is provided to GPT-4o with relevant chunks fr​om manually generated summaries of the political discussions. Additionally, we show that the average log probability is a strong predictor of the transcription BLEU score, and incorporate this metric into our dataset. The final corpus consists of 801 hours of audio, with 751 hours recommended for practical use. Our results show significant improvements in transcription quality and reduced word error rates after applying our three-steps approach.

We present SwissGPC v1.0, the first mid-to-large-scale corpus of spontaneous Swiss German speech, developed to support research in ASR, TTS, dialect identification, and related fields. The dataset consists of links to talk shows and podcasts hosted on Schweizer Radio und Fernsehen and YouTube, which contain approximately 5400 hours of raw audio. After segmentation and weak annotation, nearly 5000 hours of speech were retained, covering the seven major Swiss German dialect regions alongside Standard German.

We describe the corpus construction methodology, including an automated annotation pipeline, and provide statistics on dialect distribution, token counts, and segmentation characteristics. Unlike existing Swiss German speech corpora, which primarily feature controlled speech, this corpus captures natural, spontaneous conversations, making it a valuable resource for real-world speech applications.