Thomas Demeester

Karel D'Oosterlinck, Winnie Xu, Chris Develder et al.

Large Language Models (LLMs) are often aligned using contrastive alignment objectives and preference pair datasets. The interaction between model, paired data, and objective makes alignment a complicated procedure, sometimes producing subpar results. We study this and find that (i) preference data gives a better learning signal when the underlying responses are contrastive, and (ii) alignment objectives lead to better performance when they specify more control over the model during training. Based on these insights, we introduce Contrastive Learning from AI Revisions (CLAIR), a data-creation method which leads to more contrastive preference pairs, and Anchored Preference Optimization (APO), a controllable and more stable alignment objective. We align Llama-3-8B-Instruct using various comparable datasets and alignment objectives and measure MixEval-Hard scores, which correlate highly with human judgments. The CLAIR preferences lead to the strongest performance out of all datasets, and APO consistently outperforms less controllable objectives. Our best model, trained on 32K CLAIR preferences with APO, improves Llama-3-8B-Instruct by 7.65%, closing the gap with GPT4-turbo by 45%. Our code is available at https://github.com/ContextualAI/CLAIR_and_APO.

Karel D'Oosterlinck, Semere Kiros Bitew, Brandon Papineau et al.

State-of-the-art coreference resolutions systems depend on multiple LLM calls per document and are thus prohibitively expensive for many use cases (e.g., information extraction with large corpora). The leading word-level coreference system (WL-coref) attains 96.6% of these SOTA systems' performance while being much more efficient. In this work, we identify a routine yet important failure case of WL-coref: dealing with conjoined mentions such as 'Tom and Mary'. We offer a simple yet effective solution that improves the performance on the OntoNotes test set by 0.9% F1, shrinking the gap between efficient word-level coreference resolution and expensive SOTA approaches by 34.6%. Our Conjunction-Aware Word-level coreference model (CAW-coref) and code is available at https://github.com/KarelDO/wl-coref.

Matthias De Lange, Warre Veys, Federico Retyk et al.

Today's evolving labor markets rely increasingly on recommender systems for hiring, talent management, and workforce analytics, with natural language processing (NLP) capabilities at the core. Yet, research in this area remains highly fragmented. Studies employ divergent ontologies (ESCO, O*NET, national taxonomies), heterogeneous task formulations, and diverse model families, making cross-study comparison and reproducibility exceedingly difficult. General-purpose benchmarks lack coverage of work-specific tasks, and the inherent sensitivity of employment data further limits open evaluation. We present \textbf{WorkRB} (Work Research Benchmark), the first open-source, community-driven benchmark tailored to work-domain AI. WorkRB organizes 13 diverse tasks from 7 task groups as unified recommendation and NLP tasks, including job/skill recommendation, candidate recommendation, similar item recommendation, and skill extraction and normalization. WorkRB enables both monolingual and cross-lingual evaluation settings through dynamic loading of multilingual ontologies. Developed within a multi-stakeholder ecosystem of academia, industry, and public institutions, WorkRB has a modular design for seamless contributions and enables integration of proprietary tasks without disclosing sensitive data. WorkRB is available under the Apache 2.0 license at https://github.com/techwolf-ai/WorkRB.

Cédric Goemaere, Johannes Deleu, Thomas Demeester

The Hopfield network serves as a fundamental energy-based model in machine learning, capturing memory retrieval dynamics through an ordinary differential equation (ODE). The model's output, the equilibrium point of the ODE, is traditionally computed via synchronous updates using the forward Euler method. This paper aims to overcome some of the disadvantages of this approach. We propose a conceptual shift, viewing Hopfield networks as instances of Deep Equilibrium Models (DEQs). The DEQ framework not only allows for the use of specialized solvers, but also leads to new insights on an empirical inference technique that we will refer to as 'even-odd splitting'. Our theoretical analysis of the method uncovers a parallelizable asynchronous update scheme, which should converge roughly twice as fast as the conventional synchronous updates. Empirical evaluations validate these findings, showcasing the advantages of both the DEQ framework and even-odd splitting in digitally simulating energy minimization in Hopfield networks. The code is available at https://github.com/cgoemaere/hopdeq

Klim Zaporojets, Lucie-Aimee Kaffee, Johannes Deleu et al.

In our continuously evolving world, entities change over time and new, previously non-existing or unknown, entities appear. We study how this evolutionary scenario impacts the performance on a well established entity linking (EL) task. For that study, we introduce TempEL, an entity linking dataset that consists of time-stratified English Wikipedia snapshots from 2013 to 2022, from which we collect both anchor mentions of entities, and these target entities' descriptions. By capturing such temporal aspects, our newly introduced TempEL resource contrasts with currently existing entity linking datasets, which are composed of fixed mentions linked to a single static version of a target Knowledge Base (e.g., Wikipedia 2010 for CoNLL-AIDA). Indeed, for each of our collected temporal snapshots, TempEL contains links to entities that are continual, i.e., occur in all of the years, as well as completely new entities that appear for the first time at some point. Thus, we enable to quantify the performance of current state-of-the-art EL models for: (i) entities that are subject to changes over time in their Knowledge Base descriptions as well as their mentions' contexts, and (ii) newly created entities that were previously non-existing (e.g., at the time the EL model was trained). Our experimental results show that in terms of temporal performance degradation, (i) continual entities suffer a decrease of up to 3.1% EL accuracy, while (ii) for new entities this accuracy drop is up to 17.9%. This highlights the challenge of the introduced TempEL dataset and opens new research prospects in the area of time-evolving entity disambiguation.

Jens-Joris Decorte, Jeroen Van Hautte, Johannes Deleu et al.

The impact of person-job fit on job satisfaction and performance is widely acknowledged, which highlights the importance of providing workers with next steps at the right time in their career. This task of predicting the next step in a career is known as career path prediction, and has diverse applications such as turnover prevention and internal job mobility. Existing methods to career path prediction rely on large amounts of private career history data to model the interactions between job titles and companies. We propose leveraging the unexplored textual descriptions that are part of work experience sections in resumes. We introduce a structured dataset of 2,164 anonymized career histories, annotated with ESCO occupation labels. Based on this dataset, we present a novel representation learning approach, CareerBERT, specifically designed for work history data. We develop a skill-based model and a text-based model for career path prediction, which achieve 35.24% and 39.61% recall@10 respectively on our dataset. Finally, we show that both approaches are complementary as a hybrid approach achieves the strongest result with 43.01% recall@10.

François Remy, Kris Demuynck, Thomas Demeester

This work introduces BioLORD, a new pre-training strategy for producing meaningful representations for clinical sentences and biomedical concepts. State-of-the-art methodologies operate by maximizing the similarity in representation of names referring to the same concept, and preventing collapse through contrastive learning. However, because biomedical names are not always self-explanatory, it sometimes results in non-semantic representations. BioLORD overcomes this issue by grounding its concept representations using definitions, as well as short descriptions derived from a multi-relational knowledge graph consisting of biomedical ontologies. Thanks to this grounding, our model produces more semantic concept representations that match more closely the hierarchical structure of ontologies. BioLORD establishes a new state of the art for text similarity on both clinical sentences (MedSTS) and biomedical concepts (MayoSRS).

François Remy, Pieter Delobelle, Hayastan Avetisyan et al.

The development of monolingual language models for low and mid-resource languages continues to be hindered by the difficulty in sourcing high-quality training data. In this study, we present a novel cross-lingual vocabulary transfer strategy, trans-tokenization, designed to tackle this challenge and enable more efficient language adaptation. Our approach focuses on adapting a high-resource monolingual LLM to an unseen target language by initializing the token embeddings of the target language using a weighted average of semantically similar token embeddings from the source language. For this, we leverage a translation resource covering both the source and target languages. We validate our method with the Tweeties, a series of trans-tokenized LLMs, and demonstrate their competitive performance on various downstream tasks across a small but diverse set of languages. Additionally, we introduce Hydra LLMs, models with multiple swappable language modeling heads and embedding tables, which further extend the capabilities of our trans-tokenization strategy. By designing a Hydra LLM based on the multilingual model TowerInstruct, we developed a state-of-the-art machine translation model for Tatar, in a zero-shot manner, completely bypassing the need for high-quality parallel data. This breakthrough is particularly significant for low-resource languages like Tatar, where high-quality parallel data is hard to come by. By lowering the data and time requirements for training high-quality models, our trans-tokenization strategy allows for the development of LLMs for a wider range of languages, especially those with limited resources. We hope that our work will inspire further research and collaboration in the field of cross-lingual vocabulary transfer and contribute to the empowerment of languages on a global scale.

Semere Kiros Bitew, Amir Hadifar, Lucas Sterckx et al.

Multiple choice questions (MCQs) are widely used in digital learning systems, as they allow for automating the assessment process. However, due to the increased digital literacy of students and the advent of social media platforms, MCQ tests are widely shared online, and teachers are continuously challenged to create new questions, which is an expensive and time-consuming task. A particularly sensitive aspect of MCQ creation is to devise relevant distractors, i.e., wrong answers that are not easily identifiable as being wrong. This paper studies how a large existing set of manually created answers and distractors for questions over a variety of domains, subjects, and languages can be leveraged to help teachers in creating new MCQs, by the smart reuse of existing distractors. We built several data-driven models based on context-aware question and distractor representations, and compared them with static feature-based models. The proposed models are evaluated with automated metrics and in a realistic user test with teachers. Both automatic and human evaluations indicate that context-aware models consistently outperform a static feature-based approach. For our best-performing context-aware model, on average 3 distractors out of the 10 shown to teachers were rated as high-quality distractors. We create a performance benchmark, and make it public, to enable comparison between different approaches and to introduce a more standardized evaluation of the task. The benchmark contains a test of 298 educational questions covering multiple subjects & languages and a 77k multilingual pool of distractor vocabulary for future research.

Semere Kiros Bitew, Johannes Deleu, Chris Develder et al.

Large Language Models (LLMs) such as ChatGPT have demonstrated remarkable performance across various tasks and have garnered significant attention from both researchers and practitioners. However, in an educational context, we still observe a performance gap in generating distractors -- i.e., plausible yet incorrect answers -- with LLMs for multiple-choice questions (MCQs). In this study, we propose a strategy for guiding LLMs such as ChatGPT, in generating relevant distractors by prompting them with question items automatically retrieved from a question bank as well-chosen in-context examples. We evaluate our LLM-based solutions using a quantitative assessment on an existing test set, as well as through quality annotations by human experts, i.e., teachers. We found that on average 53% of the generated distractors presented to the teachers were rated as high-quality, i.e., suitable for immediate use as is, outperforming the state-of-the-art model. We also show the gains of our approach 1 in generating high-quality distractors by comparing it with a zero-shot ChatGPT and a few-shot ChatGPT prompted with static examples.

Amir Hadifar, Semere Kiros Bitew, Johannes Deleu et al.

We introduce a high-quality dataset that contains 3,397 samples comprising (i) multiple choice questions, (ii) answers (including distractors), and (iii) their source documents, from the educational domain. Each question is phrased in two forms, normal and close. Correct answers are linked to source documents with sentence-level annotations. Thus, our versatile dataset can be used for both question and distractor generation, as well as to explore new challenges such as question format conversion. Furthermore, 903 questions are accompanied by their cognitive complexity level as per Bloom's taxonomy. All questions have been generated by educational experts rather than crowd workers to ensure they are maintaining educational and learning standards. Our analysis and experiments suggest distinguishable differences between our dataset and commonly used ones for question generation for educational purposes. We believe this new dataset can serve as a valuable resource for research and evaluation in the educational domain. The dataset and baselines will be released to support further research in question generation.

Jens-Joris Decorte, Jeroen Van Hautte, Johannes Deleu et al.

Skills play a central role in the job market and many human resources (HR) processes. In the wake of other digital experiences, today's online job market has candidates expecting to see the right opportunities based on their skill set. Similarly, enterprises increasingly need to use data to guarantee that the skills within their workforce remain future-proof. However, structured information about skills is often missing, and processes building on self- or manager-assessment have shown to struggle with issues around adoption, completeness, and freshness of the resulting data. Extracting skills is a highly challenging task, given the many thousands of possible skill labels mentioned either explicitly or merely described implicitly and the lack of finely annotated training corpora. Previous work on skill extraction overly simplifies the task to an explicit entity detection task or builds on manually annotated training data that would be infeasible if applied to a complete vocabulary of skills. We propose an end-to-end system for skill extraction, based on distant supervision through literal matching. We propose and evaluate several negative sampling strategies, tuned on a small validation dataset, to improve the generalization of skill extraction towards implicitly mentioned skills, despite the lack of such implicit skills in the distantly supervised data. We observe that using the ESCO taxonomy to select negative examples from related skills yields the biggest improvements, and combining three different strategies in one model further increases the performance, up to 8 percentage points in RP@5. We introduce a manually annotated evaluation benchmark for skill extraction based on the ESCO taxonomy, on which we validate our models. We release the benchmark dataset for research purposes to stimulate further research on the task.

Jens-Joris Decorte, Severine Verlinden, Jeroen Van Hautte et al.

Online job ads serve as a valuable source of information for skill requirements, playing a crucial role in labor market analysis and e-recruitment processes. Since such ads are typically formatted in free text, natural language processing (NLP) technologies are required to automatically process them. We specifically focus on the task of detecting skills (mentioned literally, or implicitly described) and linking them to a large skill ontology, making it a challenging case of extreme multi-label classification (XMLC). Given that there is no sizable labeled (training) dataset are available for this specific XMLC task, we propose techniques to leverage general Large Language Models (LLMs). We describe a cost-effective approach to generate an accurate, fully synthetic labeled dataset for skill extraction, and present a contrastive learning strategy that proves effective in the task. Our results across three skill extraction benchmarks show a consistent increase of between 15 to 25 percentage points in \textit{R-Precision@5} compared to previously published results that relied solely on distant supervision through literal matches.

Paloma Rabaey, Stefan Heytens, Thomas Demeester

Clinical information extraction, which involves structuring clinical concepts from unstructured medical text, remains a challenging problem that could benefit from the inclusion of tabular background information available in electronic health records. Existing open-source datasets lack explicit links between structured features and clinical concepts in the text, motivating the need for a new research dataset. We introduce SimSUM, a benchmark dataset of 10,000 simulated patient records that link unstructured clinical notes with structured background variables. Each record simulates a patient encounter in the domain of respiratory diseases and includes tabular data (e.g., symptoms, diagnoses, underlying conditions) generated from a Bayesian network whose structure and parameters are defined by domain experts. A large language model (GPT-4o) is prompted to generate a clinical note describing the encounter, including symptoms and relevant context. These notes are annotated with span-level symptom mentions. We conduct an expert evaluation to assess note quality and run baseline predictive models on both the tabular and textual data. The SimSUM dataset is primarily designed to support research on clinical information extraction in the presence of tabular background variables, which can be linked through domain knowledge to concepts of interest to be extracted from the text -- namely, symptoms in the case of SimSUM. Secondary uses include research on the automation of clinical reasoning over both tabular data and text, causal effect estimation in the presence of tabular and/or textual confounders, and multi-modal synthetic data generation. SimSUM is not intended for training clinical decision support systems or production-grade models, but rather to facilitate reproducible research in a simplified and controlled setting.

Yiwei Jiang, Klim Zaporojets, Johannes Deleu et al.

This work presents a new dialog dataset, CookDial, that facilitates research on task-oriented dialog systems with procedural knowledge understanding. The corpus contains 260 human-to-human task-oriented dialogs in which an agent, given a recipe document, guides the user to cook a dish. Dialogs in CookDial exhibit two unique features: (i) procedural alignment between the dialog flow and supporting document; (ii) complex agent decision-making that involves segmenting long sentences, paraphrasing hard instructions and resolving coreference in the dialog context. In addition, we identify three challenging (sub)tasks in the assumed task-oriented dialog system: (1) User Question Understanding, (2) Agent Action Frame Prediction, and (3) Agent Response Generation. For each of these tasks, we develop a neural baseline model, which we evaluate on the CookDial dataset. We publicly release the CookDial dataset, comprising rich annotations of both dialogs and recipe documents, to stimulate further research on domain-specific document-grounded dialog systems.

François Remy, Kris Demuynck, Thomas Demeester

In this study, we investigate the potential of Large Language Models to complement biomedical knowledge graphs in the training of semantic models for the biomedical and clinical domains. Drawing on the wealth of the UMLS knowledge graph and harnessing cutting-edge Large Language Models, we propose a new state-of-the-art approach for obtaining high-fidelity representations of biomedical concepts and sentences, consisting of three steps: an improved contrastive learning phase, a novel self-distillation phase, and a weight averaging phase. Through rigorous evaluations via the extensive BioLORD testing suite and diverse downstream tasks, we demonstrate consistent and substantial performance improvements over the previous state of the art (e.g. +2pts on MedSTS, +2.5pts on MedNLI-S, +6.1pts on EHR-Rel-B). Besides our new state-of-the-art biomedical model for English, we also distill and release a multilingual model compatible with 50+ languages and finetuned on 7 European languages. Many clinical pipelines can benefit from our latest models. Our new multilingual model enables a range of languages to benefit from our advancements in biomedical semantic representation learning, opening a new avenue for bioinformatics researchers around the world. As a result, we hope to see BioLORD-2023 becoming a precious tool for future biomedical applications.

Sofie Labat, Thomas Demeester, Véronique Hoste

Due to the rise of user-generated content, social media is increasingly adopted as a channel to deliver customer service. Given the public character of these online platforms, the automatic detection of emotions forms an important application in monitoring customer satisfaction and preventing negative word-of-mouth. This paper introduces EmoTwiCS, a corpus of 9,489 Dutch customer service dialogues on Twitter that are annotated for emotion trajectories. In our business-oriented corpus, we view emotions as dynamic attributes of the customer that can change at each utterance of the conversation. The term `emotion trajectory' refers therefore not only to the fine-grained emotions experienced by customers (annotated with 28 labels and valence-arousal-dominance scores), but also to the event happening prior to the conversation and the responses made by the human operator (both annotated with 8 categories). Inter-annotator agreement (IAA) scores on the resulting dataset are substantial and comparable with related research, underscoring its high quality. Given the interplay between the different layers of annotated information, we perform several in-depth analyses to investigate (i) static emotions in isolated tweets, (ii) dynamic emotions and their shifts in trajectory, and (iii) the role of causes and response strategies in emotion trajectories. We conclude by listing the advantages and limitations of our dataset, after which we give some suggestions on the different types of predictive modelling tasks and open research questions to which EmoTwiCS can be applied. The dataset is available upon request and will be made publicly available upon acceptance of the paper.

François Remy, Pieter Delobelle, Bettina Berendt et al.

Training monolingual language models for low and mid-resource languages is made challenging by limited and often inadequate pretraining data. In this study, we propose a novel model conversion strategy to address this issue, adapting high-resources monolingual language models to a new target language. By generalizing over a word translation dictionary encompassing both the source and target languages, we map tokens from the target tokenizer to semantically similar tokens from the source language tokenizer. This one-to-many token mapping improves tremendously the initialization of the embedding table for the target language. We conduct experiments to convert high-resource models to mid- and low-resource languages, namely Dutch and Frisian. These converted models achieve a new state-of-the-art performance on these languages across all sorts of downstream tasks. By reducing significantly the amount of data and time required for training state-of-the-art models, our novel model conversion strategy has the potential to benefit many languages worldwide.

Henri Arno, Klaas Mulier, Joke Baeck et al.

Models for bankruptcy prediction are useful in several real-world scenarios, and multiple research contributions have been devoted to the task, based on structured (numerical) as well as unstructured (textual) data. However, the lack of a common benchmark dataset and evaluation strategy impedes the objective comparison between models. This paper introduces such a benchmark for the unstructured data scenario, based on novel and established datasets, in order to stimulate further research into the task. We describe and evaluate several classical and neural baseline models, and discuss benefits and flaws of different strategies. In particular, we find that a lightweight bag-of-words model based on static in-domain word representations obtains surprisingly good results, especially when taking textual data from several years into account. These results are critically assessed, and discussed in light of particular aspects of the data and the task. All code to replicate the data and experimental results will be released.

Maarten De Raedt, Semere Kiros Bitew, Fréderic Godin et al.

The brittleness of finetuned language model performance on out-of-distribution (OOD) test samples in unseen domains has been well-studied for English, yet is unexplored for multi-lingual models. Therefore, we study generalization to OOD test data specifically in zero-shot cross-lingual transfer settings, analyzing performance impacts of both language and domain shifts between train and test data. We further assess the effectiveness of counterfactually augmented data (CAD) in improving OOD generalization for the cross-lingual setting, since CAD has been shown to benefit in a monolingual English setting. Finally, we propose two new approaches for OOD generalization that avoid the costly annotation process associated with CAD, by exploiting the power of recent large language models (LLMs). We experiment with 3 multilingual models, LaBSE, mBERT, and XLM-R trained on English IMDb movie reviews, and evaluate on OOD test sets in 13 languages: Amazon product reviews, Tweets, and Restaurant reviews. Results echo the OOD performance decline observed in the monolingual English setting. Further, (i) counterfactuals from the original high-resource language do improve OOD generalization in the low-resource language, and (ii) our newly proposed cost-effective approaches reach similar or up to +3.1% better accuracy than CAD for Amazon and Restaurant reviews.

François Remy, Thomas Demeester

Background: More than 400,000 biomedical concepts and some of their relationships are contained in SnomedCT, a comprehensive biomedical ontology. However, their concept names are not always readily interpretable by non-experts, or patients looking at their own electronic health records (EHR). Clear definitions or descriptions in understandable language are often not available. Therefore, generating human-readable definitions for biomedical concepts might help make the information they encode more accessible and understandable to a wider public. Objective: In this article, we introduce the Automatic Glossary of Clinical Terminology (AGCT), a large-scale biomedical dictionary of clinical concepts generated using high-quality information extracted from the biomedical knowledge contained in SnomedCT. Methods: We generate a novel definition for every SnomedCT concept, after prompting the OpenAI Turbo model, a variant of GPT 3.5, using a high-quality verbalization of the SnomedCT relationships of the to-be-defined concept. A significant subset of the generated definitions was subsequently judged by NLP researchers with biomedical expertise on 5-point scales along the following three axes: factuality, insight, and fluency. Results: AGCT contains 422,070 computer-generated definitions for SnomedCT concepts, covering various domains such as diseases, procedures, drugs, and anatomy. The average length of the definitions is 49 words. The definitions were assigned average scores of over 4.5 out of 5 on all three axes, indicating a majority of factual, insightful, and fluent definitions. Conclusion: AGCT is a novel and valuable resource for biomedical tasks that require human-readable definitions for SnomedCT concepts. It can also serve as a base for developing robust biomedical retrieval models or other applications that leverage natural language understanding of biomedical knowledge.

Semere Kiros Bitew, Johannes Deleu, A. Seza Doğruöz et al.

Since performing exercises (including, e.g., practice tests) forms a crucial component of learning, and creating such exercises requires non-trivial effort from the teacher, there is a great value in automatic exercise generation in digital tools in education. In this paper, we particularly focus on automatic creation of gapfilling exercises for language learning, specifically grammar exercises. Since providing any annotation in this domain requires human expert effort, we aim to avoid it entirely and explore the task of converting existing texts into new gap-filling exercises, purely based on an example exercise, without explicit instruction or detailed annotation of the intended grammar topics. We contribute (i) a novel neural network architecture specifically designed for aforementioned gap-filling exercise generation task, and (ii) a real-world benchmark dataset for French grammar. We show that our model for this French grammar gap-filling exercise generation outperforms a competitive baseline classifier by 8% in F1 percentage points, achieving an average F1 score of 82%. Our model implementation and the dataset are made publicly available to foster future research, thus offering a standardized evaluation and baseline solution of the proposed partially annotated data prediction task in grammar exercise creation.

Alexandre Le Mercier, Thomas Demeester, Chris Develder

State space models (SSMs) like Mamba have gained significant traction as efficient alternatives to Transformers, achieving linear complexity while maintaining competitive performance. However, Hidden State Poisoning Attacks (HiSPAs), a recently discovered vulnerability that corrupts SSM memory through adversarial strings, pose a critical threat to these architectures and their hybrid variants. Framing the HiSPA mitigation task as a binary classification problem at the token level, we introduce the CLASP model to defend against this threat. CLASP exploits distinct patterns in Mamba's block output embeddings (BOEs) and uses an XGBoost classifier to identify malicious tokens with minimal computational overhead. We consider a realistic scenario in which both SSMs and HiSPAs are likely to be used: an LLM screening résumés to identify the best candidates for a role. Evaluated on a corpus of 2,483 résumés totaling 9.5M tokens with controlled injections, CLASP achieves 95.9% token-level F1 score and 99.3% document-level F1 score on malicious tokens detection. Crucially, the model generalizes to unseen attack patterns: under leave-one-out cross-validation, performance remains high (96.9% document-level F1), while under clustered cross-validation with structurally novel triggers, it maintains useful detection capability (91.6% average document-level F1). Operating independently of any downstream model, CLASP processes 1,032 tokens per second with under 4GB VRAM consumption, potentially making it suitable for real-world deployment as a lightweight front-line defense for SSM-based and hybrid architectures. All code and detailed results are available at https://anonymous.4open.science/r/hispikes-91C0.

Alexandre Le Mercier, Chris Develder, Thomas Demeester

State space models (SSMs) like Mamba offer efficient alternatives to Transformer-based language models, with linear time complexity. Yet, their adversarial robustness remains critically unexplored. This paper studies the phenomenon whereby specific short input phrases induce a partial amnesia effect in such models, by irreversibly overwriting information in their hidden states, referred to as a Hidden State Poisoning Attack (HiSPA). Our benchmark RoBench25 allows evaluating a model's information retrieval capabilities when subject to HiSPAs, and confirms the vulnerability of SSMs against such attacks. Even a recent 52B hybrid SSM-Transformer model from the Jamba family collapses on RoBench25 under optimized HiSPA triggers, whereas pure Transformers do not. We also observe that HiSPA triggers significantly weaken the Jamba model on the popular Open-Prompt-Injections benchmark, unlike pure Transformers. Finally, our interpretability study reveals patterns in Mamba's hidden layers during HiSPAs that could be used to build a HiSPA mitigation system. The full code and data to reproduce the experiments can be found at https://anonymous.4open.science/r/hispa_anonymous-5DB0.

Florian Handke, Dejan Stančević, Felix Koulischer et al.

Diffusion models do not recover semantic structure uniformly over time. Instead, samples transition from semantic ambiguity to class commitment within a narrow regime. Recent theoretical work attributes this transition to dynamical instabilities along class-separating directions, but practical methods to detect and exploit these windows in trained models are still limited. We show that tracking the class-conditional entropy of a latent semantic variable given the noisy state provides a reliable signature of these transition regimes. By restricting the entropy to semantic partitions, the entropy can furthermore resolve semantic decisions at different levels of abstraction. We analyze this behavior in high-dimensional Gaussian mixture models and show that the entropy rate concentrates on the same logarithmic time scale as the speciation symmetry-breaking instability previously identified in variance-preserving diffusion. We validate our method on EDM2-XS and Stable Diffusion 1.5, where class-conditional entropy consistently isolates the noise regimes critical for semantic structure formation. Finally, we use our framework to quantify how guidance redistributes semantic information over time. Together, these results connect information-theoretic and statistical physics perspectives on diffusion and provide a principled basis for time-localized control.

Tom Van Der Meersch, Johannes Deleu, Thomas Demeester

On dedicated analog hardware, equilibrium propagation is an energy-efficient alternative to backpropagation. In spite of its theoretical guarantees, its application in the AI domain remains limited to the discriminative setting. Meanwhile, despite its high computational demands, generative AI is on the rise. In this paper, we demonstrate the application of Equilibrium Propagation in training a variational autoencoder (VAE) for generative modeling. Leveraging the symmetric nature of Hopfield networks, we propose using a single model to serve as both the encoder and decoder which could effectively halve the required chip size for VAE implementations, paving the way for more efficient analog hardware configurations.

Maarten De Raedt, Fréderic Godin, Chris Develder et al.

For text classification tasks, finetuned language models perform remarkably well. Yet, they tend to rely on spurious patterns in training data, thus limiting their performance on out-of-distribution (OOD) test data. Among recent models aiming to avoid this spurious pattern problem, adding extra counterfactual samples to the training data has proven to be very effective. Yet, counterfactual data generation is costly since it relies on human annotation. Thus, we propose a novel solution that only requires annotation of a small fraction (e.g., 1%) of the original training data, and uses automatic generation of extra counterfactuals in an encoding vector space. We demonstrate the effectiveness of our approach in sentiment classification, using IMDb data for training and other sets for OOD tests (i.e., Amazon, SemEval and Yelp). We achieve noticeable accuracy improvements by adding only 1% manual counterfactuals: +3% compared to adding +100% in-distribution training samples, +1.3% compared to alternate counterfactual approaches.

Felix Koulischer, Cédric Goemaere, Tom van der Meersch et al.

The recent discovery of a connection between Transformers and Modern Hopfield Networks (MHNs) has reignited the study of neural networks from a physical energy-based perspective. This paper focuses on the pivotal effect of the inverse temperature hyperparameter $β$ on the distribution of energy minima of the MHN. To achieve this, the distribution of energy minima is tracked in a simplified MHN in which equidistant normalised patterns are stored. This network demonstrates a phase transition at a critical temperature $β_{\text{c}}$, from a single global attractor towards highly pattern specific minima as $β$ is increased. Importantly, the dynamics are not solely governed by the hyperparameter $β$ but are instead determined by an effective inverse temperature $β_{\text{eff}}$ which also depends on the distribution and size of the stored patterns. Recognizing the role of hyperparameters in the MHN could, in the future, aid researchers in the domain of Transformers to optimise their initial choices, potentially reducing the necessity for time and energy expensive hyperparameter fine-tuning.

Karel D'Oosterlinck, Thomas Demeester, Chris Develder et al.

Model interpretability and model editing are crucial goals in the age of large language models. Interestingly, there exists a link between these two goals: if a method is able to systematically edit model behavior with regard to a human concept of interest, this editor method can help make internal representations more interpretable by pointing towards relevant representations and systematically manipulating them.

Paloma Rabaey, Cedric De Boom, Thomas Demeester

Bayesian Networks may be appealing for clinical decision-making due to their inclusion of causal knowledge, but their practical adoption remains limited as a result of their inability to deal with unstructured data. While neural networks do not have this limitation, they are not interpretable and are inherently unable to deal with causal structure in the input space. Our goal is to build neural networks that combine the advantages of both approaches. Motivated by the perspective to inject causal knowledge while training such neural networks, this work presents initial steps in that direction. We demonstrate how a neural network can be trained to output conditional probabilities, providing approximately the same functionality as a Bayesian Network. Additionally, we propose two training strategies that allow encoding the independence relations inferred from a given causal structure into the neural network. We present initial results in a proof-of-concept setting, showing that the neural model acts as an understudy to its Bayesian Network counterpart, approximating its probabilistic and causal properties.

Henri Arno, Dennis Frauen, Emil Javurek et al.

Many decision-making problems require ranking individuals by their treatment effects rather than estimating the exact effect magnitudes. Examples include prioritizing patients for preventive care interventions, or ranking customers by the expected incremental impact of an advertisement. Surprisingly, while causal effect estimation has received substantial attention in the literature, the problem of directly learning rankings of treatment effects has largely remained unexplored. In this paper, we introduce Rank-Learner, a novel two-stage learner that directly learns the ranking of treatment effects from observational data. We first show that naive approaches based on precise treatment effect estimation solve a harder problem than necessary for ranking, while our Rank-Learner optimizes a pairwise learning objective that recovers the true treatment effect ordering, without explicit CATE estimation. We further show that our Rank-Learner is Neyman-orthogonal and thus comes with strong theoretical guarantees, including robustness to estimation errors in the nuisance functions. In addition, our Rank-Learner is model-agnostic, and can be instantiated with arbitrary machine learning models (e.g., neural networks). We demonstrate the effectiveness of our method through extensive experiments where Rank-Learner consistently outperforms standard CATE estimators and non-orthogonal ranking methods. Overall, we provide practitioners with a new, orthogonal two-stage learner for ranking individuals by their treatment effects.

Karel D'Oosterlinck, Omar Khattab, François Remy et al.

Multi-label classification problems with thousands of classes are hard to solve with in-context learning alone, as language models (LMs) might lack prior knowledge about the precise classes or how to assign them, and it is generally infeasible to demonstrate every class in a prompt. We propose a general program, $\texttt{Infer--Retrieve--Rank}$, that defines multi-step interactions between LMs and retrievers to efficiently tackle such problems. We implement this program using the $\texttt{DSPy}$ programming model, which specifies in-context systems in a declarative manner, and use $\texttt{DSPy}$ optimizers to tune it towards specific datasets by bootstrapping only tens of few-shot examples. Our primary extreme classification program, optimized separately for each task, attains state-of-the-art results across three benchmarks (HOUSE, TECH, TECHWOLF). We apply the same program to a benchmark with vastly different characteristics and attain competitive performance as well (BioDEX). Unlike prior work, our proposed solution requires no finetuning, is easily applicable to new tasks, alleviates prompt engineering, and requires only tens of labeled examples. Our code is public at https://github.com/KarelDO/xmc.dspy.

Alexandre Le Mercier, Chris Develder, Thomas Demeester

In multi-agent systems (MAS), a single deceptive agent can nullify all gains of an agentic AI collective and evade deployed defenses. However, existing adversarial studies on MAS target only shallow tasks and do not consider adaptive adversaries, which evolve their strategies to evade the very detectors trained to catch them. To address that gap, we introduce GAMBIT, a benchmark with three evaluation modes and two independent scores for evaluating imposter detectors: the first two modes measure zero-shot detection under increasing distribution shift, and a third recalibration mode measures how quickly a detector adapts to novel attacks from just 20 labeled examples. The benchmark comes with a dataset of 27,804 labeled instances spanning 240 co-evolved imposter strategies. Our contributions are threefold: (1) Using chess as a substrate deep reasoning problem and Gemini 3.1 Pro for agents, we release GAMBIT and its dataset to evaluate imposter detectors under realistic constraints against a stealthy adaptive imposter; (2) We introduce an adaptive imposter agent based on an efficient evolutionary framework, generalizable beyond chess, that collapses collective task performance while remaining essentially undetectable (50.5% F1-score with a Gemini-based detector); (3) We show that zero-shot evaluation can be highly misleading for adaptive adversaries: two detectors with near-identical zero-shot scores differ by 8x on few-shot adaptation, while the meta-learned variant converges 20x faster, a gap only visible in the recalibration mode. Altogether, GAMBIT provides the first multi-agent benchmark where adversarial attacks and defenses co-evolve, with an imposter framework generalizable beyond our use case, and promising techniques for fast recalibration in a rapidly evolving adversarial system. Code and data: https://anonymous.4open.science/r/gambit.

Henri Arno, Paloma Rabaey, Thomas Demeester

One of the central goals of causal machine learning is the accurate estimation of heterogeneous treatment effects from observational data. In recent years, meta-learning has emerged as a flexible, model-agnostic paradigm for estimating conditional average treatment effects (CATE) using any supervised model. This paper examines the performance of meta-learners when the confounding variables are expressed in text. Through synthetic data experiments, we show that learners using pre-trained text representations of confounders, in addition to tabular background variables, achieve improved CATE estimates compared to those relying solely on the tabular variables, particularly when sufficient data is available. However, due to the entangled nature of the text embeddings, these models do not fully match the performance of meta-learners with perfect confounder knowledge. These findings highlight both the potential and the limitations of pre-trained text representations for causal inference and open up interesting avenues for future research.

Karel D'Oosterlinck, François Remy, Johannes Deleu et al.

Timely and accurate extraction of Adverse Drug Events (ADE) from biomedical literature is paramount for public safety, but involves slow and costly manual labor. We set out to improve drug safety monitoring (pharmacovigilance, PV) through the use of Natural Language Processing (NLP). We introduce BioDEX, a large-scale resource for Biomedical adverse Drug Event Extraction, rooted in the historical output of drug safety reporting in the U.S. BioDEX consists of 65k abstracts and 19k full-text biomedical papers with 256k associated document-level safety reports created by medical experts. The core features of these reports include the reported weight, age, and biological sex of a patient, a set of drugs taken by the patient, the drug dosages, the reactions experienced, and whether the reaction was life threatening. In this work, we consider the task of predicting the core information of the report given its originating paper. We estimate human performance to be 72.0% F1, whereas our best model achieves 62.3% F1, indicating significant headroom on this task. We also begin to explore ways in which these models could help professional PV reviewers. Our code and data are available: https://github.com/KarelDO/BioDEX.

Klim Zaporojets, Johannes Deleu, Chris Develder et al.

This paper presents DWIE, the 'Deutsche Welle corpus for Information Extraction', a newly created multi-task dataset that combines four main Information Extraction (IE) annotation subtasks: (i) Named Entity Recognition (NER), (ii) Coreference Resolution, (iii) Relation Extraction (RE), and (iv) Entity Linking. DWIE is conceived as an entity-centric dataset that describes interactions and properties of conceptual entities on the level of the complete document. This contrasts with currently dominant mention-driven approaches that start from the detection and classification of named entity mentions in individual sentences. Further, DWIE presented two main challenges when building and evaluating IE models for it. First, the use of traditional mention-level evaluation metrics for NER and RE tasks on entity-centric DWIE dataset can result in measurements dominated by predictions on more frequently mentioned entities. We tackle this issue by proposing a new entity-driven metric that takes into account the number of mentions that compose each of the predicted and ground truth entities. Second, the document-level multi-task annotations require the models to transfer information between entity mentions located in different parts of the document, as well as between different tasks, in a joint learning setting. To realize this, we propose to use graph-based neural message passing techniques between document-level mention spans. Our experiments show an improvement of up to 5.5 F1 percentage points when incorporating neural graph propagation into our joint model. This demonstrates DWIE's potential to stimulate further research in graph neural networks for representation learning in multi-task IE. We make DWIE publicly available at https://github.com/klimzaporojets/DWIE.

Fréderic Godin, Kris Demuynck, Joni Dambre et al.

Character-level features are currently used in different neural network-based natural language processing algorithms. However, little is known about the character-level patterns those models learn. Moreover, models are often compared only quantitatively while a qualitative analysis is missing. In this paper, we investigate which character-level patterns neural networks learn and if those patterns coincide with manually-defined word segmentations and annotations. To that end, we extend the contextual decomposition technique (Murdoch et al. 2018) to convolutional neural networks which allows us to compare convolutional neural networks and bidirectional long short-term memory networks. We evaluate and compare these models for the task of morphological tagging on three morphologically different languages and show that these models implicitly discover understandable linguistic rules. Our implementation can be found at https://github.com/FredericGodin/ContextualDecomposition-NLP .

Paloma Rabaey, Jong Hak Moon, Jung-Oh Lee et al.

Radiology reports are invaluable for clinical decision-making and hold great potential for automated analysis when structured into machine-readable formats. These reports often contain uncertainty, which we categorize into two distinct types: (i) Explicit uncertainty reflects doubt about the presence or absence of findings, conveyed through hedging phrases. These vary in meaning depending on the context, making rule-based systems insufficient to quantify the level of uncertainty for specific findings; (ii) Implicit uncertainty arises when radiologists omit parts of their reasoning, recording only key findings or diagnoses. Here, it is often unclear whether omitted findings are truly absent or simply unmentioned for brevity. We address these challenges with a two-part framework. We quantify explicit uncertainty by creating an expert-validated, LLM-based reference ranking of common hedging phrases, and mapping each finding to a probability value based on this reference. In addition, we model implicit uncertainty through an expansion framework that systematically adds characteristic sub-findings derived from expert-defined diagnostic pathways for 14 common diagnoses. Using these methods, we release Lunguage++, an expanded, uncertainty-aware version of the Lunguage benchmark of fine-grained structured radiology reports. This enriched resource enables uncertainty-aware image classification, faithful diagnostic reasoning, and new investigations into the clinical impact of diagnostic uncertainty.

Adrick Tench, Thomas Demeester

Understanding causal relationships is critical for healthcare. Accurate causal models provide a means to enhance the interpretability of predictive models, and furthermore a basis for counterfactual and interventional reasoning and the estimation of treatment effects. However, would-be practitioners of causal discovery face a dizzying array of algorithms without a clear best choice. This abundance of competitive algorithms makes ensembling a natural choice for practical applications. At the same time, real-world use cases frequently face challenges that violate the assumptions of common causal discovery algorithms, forcing heavy reliance on expert knowledge. Inspired by recent work on dynamically requested expert knowledge and LLMs as experts, we present a flexible model averaging method leveraging dynamically requested expert knowledge to ensemble a diverse array of causal discovery algorithms. Experiments demonstrate the efficacy of our method with imperfect experts such as LLMs on both clean and noisy data. We also analyze the impact of different degrees of expert correctness and assess the capabilities of LLMs for clinical causal discovery, providing valuable insights for practitioners.

Felix Koulischer, Johannes Deleu, Gabriel Raya et al.

Negative Prompting (NP) is widely utilized in diffusion models, particularly in text-to-image applications, to prevent the generation of undesired features. In this paper, we show that conventional NP is limited by the assumption of a constant guidance scale, which may lead to highly suboptimal results, or even complete failure, due to the non-stationarity and state-dependence of the reverse process. Based on this analysis, we derive a principled technique called Dynamic Negative Guidance, which relies on a near-optimal time and state dependent modulation of the guidance without requiring additional training. Unlike NP, negative guidance requires estimating the posterior class probability during the denoising process, which is achieved with limited additional computational overhead by tracking the discrete Markov Chain during the generative process. We evaluate the performance of DNG class-removal on MNIST and CIFAR10, where we show that DNG leads to higher safety, preservation of class balance and image quality when compared with baseline methods. Furthermore, we show that it is possible to use DNG with Stable Diffusion to obtain more accurate and less invasive guidance than NP.

Alexander Decruyenaere, Heidelinde Dehaene, Paloma Rabaey et al.

Recent advances in generative models facilitate the creation of synthetic data to be made available for research in privacy-sensitive contexts. However, the analysis of synthetic data raises a unique set of methodological challenges. In this work, we highlight the importance of inferential utility and provide empirical evidence against naive inference from synthetic data, whereby synthetic data are treated as if they were actually observed. Before publishing synthetic data, it is essential to develop statistical inference tools for such data. By means of a simulation study, we show that the rate of false-positive findings (type 1 error) will be unacceptably high, even when the estimates are unbiased. Despite the use of a previously proposed correction factor, this problem persists for deep generative models, in part due to slower convergence of estimators and resulting underestimation of the true standard error. We further demonstrate our findings through a case study.

Matthieu Meeus, Anthony Rathé, François Remy et al.

While Large Language Models (LLMs) have shown remarkable capabilities in natural language understanding and generation, their performance often lags in lower-resource, non-English languages due to biases in the training data. In this work, we explore strategies for adapting the primarily English LLMs (Llama-2 and Llama-3) to Dutch, a language spoken by 30 million people worldwide yet often underrepresented in LLM development. We collect 104GB of Dutch text ($32$B tokens) from various sources to first apply continued pretraining using low-rank adaptation (LoRA), complemented with Dutch posttraining strategies provided by prior work. For Llama-2, we consider using (i) the tokenizer of the original model, and (ii) training a new, Dutch-specific tokenizer combined with embedding reinitialization. We evaluate our adapted models, ChocoLlama-2, both on standard benchmarks and a novel Dutch benchmark, ChocoLlama-Bench. Our results demonstrate that LoRA can effectively scale for language adaptation, and that tokenizer modification with careful weight reinitialization can improve performance. Notably, Llama-3 was released during the course of this project and, upon evaluation, demonstrated superior Dutch capabilities compared to our Dutch-adapted versions of Llama-2. We hence apply the same adaptation technique to Llama-3, using its original tokenizer. While our adaptation methods enhanced Llama-2's Dutch capabilities, we found limited gains when applying the same techniques to Llama-3. This suggests that for ever improving, multilingual foundation models, language adaptation techniques may benefit more from focusing on language-specific posttraining rather than on continued pretraining. We hope this work contributes to the broader understanding of adapting LLMs to lower-resource languages, and to the development of Dutch LLMs in particular.

Jens-Joris Decorte, Jeroen Van Hautte, Chris Develder et al.

Labor market analysis relies on extracting insights from job advertisements, which provide valuable yet unstructured information on job titles and corresponding skill requirements. While state-of-the-art methods for skill extraction achieve strong performance, they depend on large language models (LLMs), which are computationally expensive and slow. In this paper, we propose \textbf{ConTeXT-match}, a novel contrastive learning approach with token-level attention that is well-suited for the extreme multi-label classification task of skill classification. \textbf{ConTeXT-match} significantly improves skill extraction efficiency and performance, achieving state-of-the-art results with a lightweight bi-encoder model. To support robust evaluation, we introduce \textbf{Skill-XL}, a new benchmark with exhaustive, sentence-level skill annotations that explicitly address the redundancy in the large label space. Finally, we present \textbf{JobBERT V2}, an improved job title normalization model that leverages extracted skills to produce high-quality job title representations. Experiments demonstrate that our models are efficient, accurate, and scalable, making them ideal for large-scale, real-time labor market analysis.

Joachim De Baer, A. Seza Doğruöz, Thomas Demeester et al.

Optimizing language models for use in conversational agents requires large quantities of example dialogues. Increasingly, these dialogues are synthetically generated by using powerful large language models (LLMs), especially in domains where obtaining authentic human data is challenging. One such domain is human resources (HR). In this context, we compare two LLM-based dialogue generation methods for producing HR job interviews, and assess which method generates higher-quality dialogues, i.e., those more difficult to distinguish from genuine human discourse. The first method uses a single prompt to generate the complete interview dialogue. The second method uses two agents that converse with each other. To evaluate dialogue quality under each method, we ask a judge LLM to determine whether AI was used for interview generation, using pairwise interview comparisons. We empirically find that, at the expense of a sixfold increase in token count, interviews generated with the dual-prompt method achieve a win rate 2 to 10 times higher than those generated with the single-prompt method. This difference remains consistent regardless of whether GPT-4o or Llama 3.3 70B is used for either interview generation or quality judging.

Max Hallemeesch, Marija Pizurica, Paloma Rabaey et al.

Cancer diagnosis and prognosis primarily depend on clinical parameters such as age and tumor grade, and are increasingly complemented by molecular data, such as gene expression, from tumor sequencing. However, sequencing is costly and delays oncology workflows. Recent advances in Deep Learning allow to predict molecular information from morphological features within Whole Slide Images (WSIs), offering a cost-effective proxy of the molecular markers. While promising, current methods lack the robustness to fully replace direct sequencing. Here we aim to improve existing methods by introducing a model-agnostic framework that allows to inject prior knowledge on gene-gene interactions into Deep Learning architectures, thereby increasing accuracy and robustness. We design the framework to be generic and flexibly adaptable to a wide range of architectures. In a case study on breast cancer, our strategy leads to an average increase of 983 significant genes (out of 25,761) across all 18 experiments, with 14 generalizing to an increase on an independent dataset. Our findings reveal a high potential for injection of prior knowledge to increase gene expression prediction performance from WSIs across a wide range of architectures.

Alexander Decruyenaere, Heidelinde Dehaene, Paloma Rabaey et al.

While synthetic data hold great promise for privacy protection, their statistical analysis poses significant challenges that necessitate innovative solutions. The use of deep generative models (DGMs) for synthetic data generation is known to induce considerable bias and imprecision into synthetic data analyses, compromising their inferential utility as opposed to original data analyses. This bias and uncertainty can be substantial enough to impede statistical convergence rates, even in seemingly straightforward analyses like mean calculation. The standard errors of such estimators then exhibit slower shrinkage with sample size than the typical 1 over root-$n$ rate. This complicates fundamental calculations like p-values and confidence intervals, with no straightforward remedy currently available. In response to these challenges, we propose a new strategy that targets synthetic data created by DGMs for specific data analyses. Drawing insights from debiased and targeted machine learning, our approach accounts for biases, enhances convergence rates, and facilitates the calculation of estimators with easily approximated large sample variances. We exemplify our proposal through a simulation study on toy data and two case studies on real-world data, highlighting the importance of tailoring DGMs for targeted data analysis. This debiasing strategy contributes to advancing the reliability and applicability of synthetic data in statistical inference.

Sofie Labat, Thomas Demeester, Véronique Hoste

Emotion-aware customer service needs in-domain conversational data, rich annotations, and predictive capabilities, but existing resources for emotion recognition are often out-of-domain, narrowly labeled, and focused on post-hoc detection. To address this, we conducted a controlled Wizard of Oz (WOZ) experiment to elicit interactions with targeted affective trajectories. The resulting corpus, EmoWOZ-CS, contains 2,148 bilingual (Dutch-English) written dialogues from 179 participants across commercial aviation, e-commerce, online travel agencies, and telecommunication scenarios. Our contributions are threefold: (1) Evaluate WOZ-based operator-steered valence trajectories as a design for emotion research; (2) Quantify human annotation performance and variation, including divergences between self-reports and third-party judgments; (3) Benchmark detection and forward-looking emotion inference in real-time support. Findings show neutral dominates participant messages; desire and gratitude are the most frequent non-neutral emotions. Agreement is moderate for multilabel emotions and valence, lower for arousal and dominance; self-reports diverge notably from third-party labels, aligning most for neutral, gratitude, and anger. Objective strategies often elicit neutrality or gratitude, while suboptimal strategies increase anger, annoyance, disappointment, desire, and confusion. Some affective strategies (cheerfulness, gratitude) foster positive reciprocity, whereas others (apology, empathy) can also leave desire, anger, or annoyance. Temporal analysis confirms successful conversation-level steering toward prescribed trajectories, most distinctly for negative targets; positive and neutral targets yield similar final valence distributions. Benchmarks highlight the difficulty of forward-looking emotion inference from prior turns, underscoring the complexity of proactive emotion-aware support.

Paloma Rabaey, Adrick Tench, Stefan Heytens et al.

Electronic health records (EHRs) form an invaluable resource for training clinical decision support systems. To leverage the potential of such systems in high-risk applications, we need large, structured tabular datasets on which we can build transparent feature-based models. While part of the EHR already contains structured information (e.g. diagnosis codes, medications, and lab results), much of the information is contained within unstructured text (e.g. discharge summaries and nursing notes). In this work, we propose a method for multi-modal patient-level information extraction that leverages both the tabular features available in the patient's EHR (using an expert-informed Bayesian network) as well as clinical notes describing the patient's symptoms (using neural text classifiers). We propose the use of virtual evidence augmented with a consistency node to provide an interpretable, probabilistic fusion of the models' predictions. The consistency node improves the calibration of the final predictions compared to virtual evidence alone, allowing the Bayesian network to better adjust the neural classifier's output to handle missing information and resolve contradictions between the tabular and text data. We show the potential of our method on the SimSUM dataset, a simulated benchmark linking tabular EHRs with clinical notes through expert knowledge.

Henri Arno, Thomas Demeester

Existing methods for estimating personalized treatment effects typically rely on structured covariates, limiting their applicability to unstructured data. Yet, leveraging unstructured data for causal inference has considerable application potential, for instance in healthcare, where clinical notes or medical images are abundant. To this end, we first introduce an approximate 'plug-in' method trained directly on the neural representations of unstructured data. However, when these fail to capture all confounding information, the method may be subject to confounding bias. We therefore introduce two theoretically grounded estimators that leverage structured measurements of the confounders during training, but allow estimating personalized treatment effects purely from unstructured inputs, while avoiding confounding bias. When these structured measurements are only available for a non-representative subset of the data, these estimators may suffer from sampling bias. To address this, we further introduce a regression-based correction that accounts for the non-uniform sampling, assuming the sampling mechanism is known or can be well-estimated. Our experiments on two benchmark datasets show that the plug-in method, directly trainable on large unstructured datasets, achieves strong empirical performance across all settings, despite its simplicity.

Paloma Rabaey, Johannes Deleu, Stefan Heytens et al.

Bayesian networks are well-suited for clinical reasoning on tabular data, but are less compatible with natural language data, for which neural networks provide a successful framework. This paper compares and discusses strategies to augment Bayesian networks with neural text representations, both in a generative and discriminative manner. This is illustrated with simulation results for a primary care use case (diagnosis of pneumonia) and discussed in a broader clinical context.