2025Activity report​​​‌Project-TeamMARIANNE

Axis​ A - Argument mining:​‌

The first research axis​​ is the development of​​​‌ models and algorithms designed​ for mining natural language​‌ arguments from text. The​​ Argument Mining (AM) scientific​​​‌ community has focused till​ now on the two​‌ main tasks constituting the​​ AM pipeline, i.e., the​​​‌ detection of argumentative components​ (namely, evidence and claims),​‌ and the prediction of​​ the relations of support​​​‌ and attack holding among​ them. They represent an​‌ obligatory starting step, but​​ the resulting argumentation frameworks​​​‌ (i.e., the graph structure​ composed by the arguments​‌ as the nodes of​​ the graph and the​​​‌ relations representing the links)​ are still quite simple​‌ with respect to the​​ needs raised in the​​​‌ team application scenarios. Our​ goal is to enhance​‌ the extraction of machine-processable​​ natural language argument structures​​​‌ to allow reasoning over​ complex real world natural​‌ arguments, with a focus​​ on the English, French,​​​‌ Italian, Spanish and German​ languages.

Axis B -​‌ Natural language argument quality​​ assessment:

The second research​​​‌ axis is focused on​ the definition of computational​‌ methods to automatically assess​​ the quality of natural​​​‌ language arguments. Despite a​ few existing approaches, the​‌ issue of automatically assessing​​ the quality of an​​​‌ argumentation remains largely unexplored.​ On the one side,​‌ it consists in assessing​​ the quality of the​​ mined arguments to decide,​​​‌ for instance, whether a‌ certain argument has to‌​‌ be selected for synthesizing​​ a debate, or whether​​​‌ the overall debate is‌ of good quality. On‌​‌ the other side, it​​ consists in ensuring that​​​‌ the newly generated arguments‌ satisfy the defined quality‌​‌ criteria in order to​​ assess them from the​​​‌ qualitative point of view,‌ i.e., a counter-argument to‌​‌ attack a fake news​​ needs to be concise​​​‌ and without repetitions. The‌ quality of the arguments‌​‌ is also characterized by​​ formal properties of the​​​‌ argumentation graph, e.g., the‌ argument strength, argument preferences,‌​‌ and argument acceptability.

Axis​​ C - Generation of​​​‌ natural language arguments:

In‌ addition to the definition‌​‌ of more effective methods​​ to mine (fine-grained) argumentative​​​‌ structures from text, and‌ to automatically assess their‌​‌ quality, the third research​​ axis of the team​​​‌ consists in the definition‌ of new (generative and‌​‌ not generative) methods to​​ generate natural language arguments,​​​‌ with a focus on‌ English and French initially.‌​‌ This process is incremental​​ and starts with the​​​‌ generation of single argumentative‌ components towards the generation‌​‌ of arguments in the​​ context of interactive dialogues​​​‌ with users. These dialogues‌ are then employed in‌​‌ different use cases with​​ different goals, i.e., explanation,​​​‌ counter-argumentation. These arguments are‌ firstly generated starting from‌​‌ the mined arguments, and​​ they rely on abductive​​​‌ reasoning schema based on‌ the set of critical‌​‌ questions and reasoned responses​​ necessary to reach the​​​‌ user's understanding.

Ethics.

Like‌ all AI algorithms, the‌​‌ results of the technological​​ solutions developed in MARIANNE​​​‌ may potentially be used‌ in a misleading way.‌​‌ This potential danger concerns​​ mainly the argumentation generation​​​‌ part, e.g., the generation‌ of hateful and fallacious‌​‌ arguments. This is why​​ a significant part of​​​‌ the research activity of‌ MARIANNE is dedicated to‌​‌ the definition of computational​​ solutions to “ensure” that​​​‌ the argumentation generated by‌ our algorithms in human-machine‌​‌ conversations will not contain​​ harmful or unfair content.​​​‌ This means to go‌ beyond the pure inclusion‌​‌ of language guardrails, as​​ in current Large Language​​​‌ Model (LLM) solutions, as‌ the ethical foundation of‌​‌ an argumentation must be​​ “controlled” on the discourse​​​‌ level too. The team‌ members are already working‌​‌ on this topic, as​​ witnessed by the recent​​​‌ publication at EMNLP 2024.‌ In addition, the awareness‌​‌ of the ethical concerns​​ around AI is high,​​​‌ thanks also to the‌ participation of team members‌​‌ (Serena Villata) to national​​ instances like the National​​​‌ Pilot Committee of Digital‌ Ethics (CNPEN).

Long​‌ term objectives

Current models​​ are shown to learn​​​‌ linguistic patterns or cues​ rather than a real​‌ semantic understanding of the​​ arguments. For instance, a​​​‌ model could predict a​ relation correctly, but it​‌ cannot explain the underlying​​ reasoning of why the​​​‌ two linked components are​ in a relationship, since​‌ the required warrants are​​ not explicitly mentioned in​​​‌ the text. Current classification​ algorithms can therefore be​‌ effective until a certain​​ point, where the relation​​​‌ can be inferred from​ explicit mentions in the​‌ text. The team will​​ therefore investigate methods to​​​‌ inject both common sense​ knowledge and expert domain​‌ knowledge that humans exploit​​ to carry out such​​​‌ kind of reasoning, in​ the form of knowledge​‌ bases or ontologies, to​​ boost the performance of​​​‌ current neural methods applied​ to Argument Mining (AM)​‌ tasks. This long-term objective​​ takes different forms depending​​​‌ on the application scenario:​ for the healthcare and​‌ the legal ones it​​ consists in precise domain​​​‌ knowledge pointed out by​ domain experts, whilst in​‌ the disinformation, hate speech​​ and the politics ones,​​​‌ it consists in a​ subtle mixture of precise​‌ domain knowledge (e.g., if​​ we talk about vaccines​​​‌ or a new tax​ law) and common sense​‌ knowledge. This step is​​ mandatory to boost the​​​‌ performances of argument mining​ models.

Thus far, we​‌ concentrated on the task​​ of predicting intra-argument relations,​​​‌ i.e., those relations holding​ between the evidence and​‌ claims of a single​​ argument. However, a necessary​​​‌ further step consists in​ moving from intra-argument relation​‌ prediction to inter-argument relation​​ prediction. Establishing an argumentative​​​‌ link between two (or​ more) different arguments is​‌ mandatory to obtain a​​ full argumentation framework (e.g.,​​​‌ a graph representation of​ all the evidence and​‌ claims extracted from the​​ clinical essays about Covid-19,​​​‌ where arguments are linked​ through argumentative relations) able​‌ to capture possible inconsistencies​​ (e.g., self-attacking arguments, rebuttal​​​‌ on points supporting the​ candidate's viewpoint in political​‌ debates) and fallacious arguments​​ (e.g., ad hominem attacks),​​​‌ that are at the​ basis of the research​‌ challenge on argument quality​​ discussed below. We plan​​​‌ to tackle this issue​ through neural models like​‌ Graph Neural Networks (GNNs),​​ which represent a natural​​​‌ choice given the graph-based​ structure of the argumentation,​‌ with a particular focus​​ on the use of​​​‌ GNNs as a model​ of neural-symbolic computing, ensuring​‌ explainability.

Whilst a high-level​​ classification of the argument​​​‌ components using the classes​ evidence and claim and​‌ of the relations with​​ attack and support is​​​‌ required to identify the​ main argumentative elements present​‌ in the text, as​​ soon as we move​​​‌ to precise application scenarios​ this general classification becomes​‌ insufficient and needs to​​ be refined with more​​​‌ precise classes. Both for​ argumentative components and relations,​‌ the finer-grained classes strongly​​ depend on the application​​ domain (e.g., in the​​​‌ political domain we may‌ have critical discussion and‌​‌ strategic maneuvering as argument​​ classes, and rebut and​​​‌ undercut as relation classes).‌ For argumentative relations, the‌​‌ idea is to start​​ considering the two classical​​​‌ inter-argument relations in formal‌ computational models of argument,‌​‌ namely rebut (i.e., the​​ claim of an argument​​​‌ attacks the claim of‌ another argument) and undercut‌​‌ (i.e., the claim of​​ an argument attacks the​​​‌ premises of another argument).‌

Overall, in this first‌​‌ axis, the team will​​ extend and empower neural​​​‌ AM models and algorithms‌ to detect even more‌​‌ precisely both argument components​​ and relations in the​​​‌ application scenarios of the‌ team (i.e., medicine, politics,‌​‌ legal cases, hate speech,​​ disinformation), with a focus​​​‌ on the English, French,‌ Italian, Spanish and German‌​‌ languages. All the above​​ mentioned objectives require to​​​‌ pass through (i) the‌ definition of detailed annotation‌​‌ guidelines of high quality​​ linguistic resources, (ii) the​​​‌ annotation of inter-argument relations‌ and finer-grained classes of‌​‌ components/relations, and (iii) the​​ definition of neural AM​​​‌ algorithms empowered with knowledge‌ bases and domain ontologies.‌​‌ The team will also​​ release a number of​​​‌ annotated linguistic resources for‌ these tasks that will‌​‌ be made available to​​ contribute to the advancements​​​‌ of the research community.‌

Medium-term projects

In order‌​‌ to reach the long​​ term objectives, we work​​​‌ on the following projects:‌

• Mining argument misalignments and‌​‌ conflicts: Another mid-term​​ objective of MARIANNE will​​​‌ be the identification and‌ classification of misalignments and‌​‌ overt conflicts that can​​ hinder argumentative discussion and​​​‌ block democratic deliberation. We‌ will design technology-enhanced dialogue‌​‌ spaces based on dispute​​ mediation, that can be​​​‌ applied also in the‌ context of social media‌​‌ discussions to detect and​​ counter the use of​​​‌ abusive language. This innovative‌ method, which scales up‌​‌ mediation techniques for large-scale​​ dialogues, foresees two steps:​​​‌ (i) Conflict analysis: analysis‌ of the main sources‌​‌ and the pattern of​​ escalation of conflicting episodes​​​‌ in the use cases,‌ based on argument mining‌​‌ methods (e.g., frequent debated​​ issues, conflicting arguments) and​​​‌ argument-based qualitative analysis of‌ discourses (which allows to‌​‌ see the dynamics of​​ conflict escalation). The argument​​​‌ conflict analysis will uncover‌ the real issues, recurring‌​‌ types of conflicting arguments​​ that are based on​​​‌ different starting points, but‌ also uncover real issues‌​‌ on which the discussion​​ can proceed and points​​​‌ of commonalities. To achieve‌ this goal, our objective‌​‌ will be to leverage​​ the structural information present​​​‌ in the debates to‌ enhance the performance of‌​‌ relation prediction between argument​​ components. We will tackle​​​‌ this task by integrating‌ Knowledge Graph Embedding models‌​‌ with Transformer-based models based​​ on Pre-trained Large Language​​​‌ Models.
• Mining implicit arguments‌: In argumentation, it‌​‌ is crucial to comprehend​​ the main arguments being​​​‌ put forward, the underlying‌ reasoning, and how these‌​‌ arguments interrelate within a​​ specific context. In this​​​‌ mid-term objective, we aim‌ to develop systems capable‌​‌ of automatically extracting arguments​​ and discerning their relationships.​​​‌ However, human argumentation is‌ more intricate than mere‌​‌ argument extraction and lexical​​​‌ or semantic analysis. Argumentative​ units may not always​‌ be explicitly stated in​​ discourse; instead, they are​​​‌ often connected through implicit​ inferences. These inferences depend​‌ on various factors, including​​ background general knowledge, domain-specific​​​‌ expertise (as medical or​ juridical ones), subjective reasoning,​‌ or other patterns, which​​ complicate argument comprehension and​​​‌ comprehensive analysis in the​ different domains. Therefore, our​‌ objectives are to investigate​​ the role of external​​​‌ knowledge and other extra-linguistic​ features in elucidating implicit​‌ reasoning chains in argumentation,​​ and to utilize these​​​‌ implicit elements to enhance​ existing methods for extracting​‌ arguments and predicting their​​ relationships.
• Mining arguments over​​​‌ time: The way​ argumentation is carried out​‌ in public discourse, political​​ debates and scientific communication​​​‌ changes over time. Another​ mid-term goal of the​‌ team will be to​​ automatically explore the dynamics​​​‌ of inter and intra-argument​ structures over time. On​‌ the one side, the​​ task will be to​​​‌ investigate through semi-supervised and​ unsupervised learning methods how​‌ the argumentation evolved over​​ time. For instance, for​​​‌ political debates, we plan​ to start with the​‌ USElecDeb60To20 dataset we built,​​ containing all the US​​​‌ presidential debates since 1960​ to 2020. The objective​‌ is to study the​​ temporal evolution trends of​​​‌ the argumentation, to see​ how the structure of​‌ the arguments evolved (e.g.,​​ number and fine-grade degree​​​‌ of the premises, presence​ of major claims, employment​‌ of rhetorical elements, choice​​ of news events, change​​​‌ points). On the other​ side, the task will​‌ consist in the investigation​​ of the dynamics of​​​‌ the argumentation in terms​ of attacks and supports​‌ among the candidates' arguments​​ (i.e., graph level analysis​​​‌ of the argumentation). The​ final goal will be​‌ to assess if and​​ how the dynamics of​​​‌ the argumentation impacted the​ outcome of the decision​‌ making process. For instance,​​ this would allow us​​​‌ to learn from past​ argumentation dynamics to predict​‌ the results of the​​ future elections in a​​​‌ country.

Long term objectives

To​​ automatically assess the quality​​​‌ of argument structures, we​ will consider first the​‌ following three main high-level​​ dimensions to characterize arguments'​​​‌ quality: (i) Cogency: an​ argument should be seen​‌ as cogent if it​​ has individually acceptable premises​​​‌ that are relevant to​ the argument's conclusion and​‌ that are sufficient to​​ draw the conclusion; (ii)​​​‌ Rhetorical effectiveness: an argumentation​ should be seen as​‌ effective if it persuades​​ the audience of the​​​‌ author's stance on the​ discussed issue. Besides the​‌ logical grounding of the​​ actual arguments (logos​​​‌), it also takes​ into consideration the credibility​‌ of the argument source​​ (ethos) and​​​‌ the emotional force of​ the argumentation (pathos​‌); (iii) Reasonableness: an​​ argumentation should be seen​​​‌ as reasonable if it​ contributes to the resolution​‌ of the given issue​​ in a sufficient way​​​‌ that is acceptable to​ everyone from the expected​‌ target audience. These three​​ dimensions of argument quality​​​‌ will then be specified​ in further sub-dimensions, like​‌ the fact that a​​ premise of an argument​​ should be seen as​​​‌ relevant if it contributes‌ to the acceptance or‌​‌ rejection of the argument's​​ conclusion (i.e., if it​​​‌ is worthy considering it‌ as a reason, evidence,‌​‌ or similar regarding the​​ conclusion), or that an​​​‌ argumentation is successful in‌ creating credibility if it‌​‌ conveys arguments in a​​ way that makes the​​​‌ author trustworthy (e.g., by‌ indicating the honesty of‌​‌ the author or by​​ revealing the author's expertise​​​‌ regarding the discussed issue),‌ and it should be‌​‌ seen as not successful​​ if the opposite holds.​​​‌

In addition to these‌ three main argument quality‌​‌ dimensions, we will consider​​ also two other elements,​​​‌ namely conceptual notions, and‌ influence factors. Conceptual notions‌​‌ are the notions of​​ maximal quality (based on​​​‌ arguing goals such as‌ agreement and deliberation or‌​‌ on preferences between different​​ arguments), and the notions​​​‌ of minimal quality, which‌ represent what makes an‌​‌ argument valuable or appropriate​​ to be stated as​​​‌ well as how to‌ avoid fallacies. Influence factors‌​‌ are factors that influence​​ the perception of quality​​​‌ beyond the content, structure,‌ and style of the‌​‌ argument itself like argument-related​​ factors (such as the​​​‌ argument's length, its structure‌ in terms of relations‌​‌ between units, and revisions​​ applied to it), and​​​‌ context-related factors (such as‌ the domain of the‌​‌ discussion, the audience addressed,​​ and the debaters involved).​​​‌

These three dimensions takes‌ into account both the‌​‌ content of the argument​​ itself (cogency and rhetorical​​​‌ effectiveness) and the argumentation‌ structure (reasonableness, considering both‌​‌ counter-arguments and their rebuttals),​​ whilst maximal/minimal quality and​​​‌ the influence factors concerns‌ more the context of‌​‌ the arguments to be​​ assessed. To address the​​​‌ automatic quality assessment of‌ the arguments, we will‌​‌ need to jointly consider​​ the content of the​​​‌ textual arguments (i.e., by‌ fine-tuning generalist or domain-specific‌​‌ Large Language Models) and​​ the graph structure of​​​‌ the whole argumentation (i.e.,‌ through the generation of‌​‌ graph embeddings or the​​ usage of Graph Neural​​​‌ Network models). These node-level‌ features will be used‌​‌ to create graph level​​ statistics useful for the​​​‌ reasonableness dimension in particular.‌

Finally, the same criteria‌​‌ proposed for assessing the​​ quality of the mined​​​‌ arguments will be used‌ to assess, as a‌​‌ long-term objective, the quality​​ of the generated arguments.​​​‌ In addition, the argument‌ generation will also be‌​‌ evaluated with diversity, which​​ is of paramount importance,​​​‌ since verbatim repetition of‌ arguments can become detrimental‌​‌ in explanation and counter-argumentation​​ scenarios. Both lexical diversity​​​‌ and semantic diversity can‌ be considered at this‌​‌ stage, where lexical diversity​​ focuses on the variability​​​‌ of the generated arguments‌ and can be captured‌​‌ by word overlapping metrics,​​ and semantic diversity focuses​​​‌ on meaning and is‌ harder to be captured,‌​‌ as in the case​​ of generated arguments with​​​‌ similar meaning but different‌ wordings. The model to‌​‌ generate arguments will be​​ also evaluated along with​​​‌ standard metrics, i.e., BLEU‌ and BertScore (computing a‌​‌ similarity score, using contextual​​ embeddings, for each token​​​‌ in the candidate sentence‌ with each token in‌​‌ the reference sentence) concerning​​​‌ the lexical and semantic​ generation performances.

Medium-term projects​‌

In order to reach​​ the long term objectives,​​​‌ we work on the​ following projects:

• Counter-narrative Quality​‌ Assessment: This first​​ mid-term objective aims to​​​‌ perform a quality assessment​ of Counter Narratives (CN).​‌ In particular, it focuses​​ on the evaluation of​​​‌ the effectiveness of counter​ narratives against hate speech.​‌ Furthermore, it delves into​​ fairness evaluation, proposing the​​​‌ development of metrics to​ assess representation, and inclusivity,​‌ along with employing bias​​ detection techniques to mitigate​​​‌ biases in counter narratives.​ The research will be​‌ conducted on publicly available​​ CN datasets, created both​​​‌ by experts from Non-Governmental​ organisations (NGOs) or collected​‌ directly from user-generated messages​​ on social media. To​​​‌ assess effectiveness, various dimensions​ have to be considered.​‌ These include: the importance​​ of emotional appeal, the​​​‌ analysis of argumentative structures​ and identification of argumentative​‌ language. Additionally, we will​​ consider the need for​​​‌ soundness, target specificity, and​ complexity analysis in counter​‌ narratives. To assess fairness,​​ we will define metrics​​​‌ for evaluating counter narratives​ against hate speech, focusing​‌ on avoidance of harmful​​ stereotypes or biases, and​​​‌ appropriate language. The automatic​ quality assessment of CN​‌ will be addressed through​​ a novel neural model​​​‌ including word embeddings, sentiment​ and emotion features, and​‌ pre-trained LLMs. Furthermore, this​​ task addresses the differences​​​‌ in style between expert-generated​ and user-generated counter narratives,​‌ emphasizing the need for​​ research to understand their​​​‌ impact on effectiveness and​ fairness.
• Fallacious and Fake​‌ Argument Assessment: Fallacies​​ are arguments that employ​​​‌ faulty reasoning. Given their​ persuasive and seemingly valid​‌ nature, fallacious arguments are​​ often used in political​​​‌ debates. Employing these misleading​ arguments in politics can​‌ have detrimental consequences for​​ society, since they can​​​‌ lead to inaccurate conclusions​ and invalid inferences from​‌ the public opinion and​​ policymakers. Automatically detecting and​​​‌ classifying fallacious arguments represents​ therefore a crucial challenge​‌ to limit the spread​​ of misleading or manipulative​​​‌ claims, and promote a​ more informed and healthier​‌ political discourse. In this​​ mid-term objective, we aim​​​‌ to go beyond the​ six categories of fallacious​‌ arguments we considered thus​​ far (i.e., Ad Hominem,​​​‌ Appeal to Authority, Appeal​ to Emotion, False Cause,​‌ Slippery Slope, and Slogans),​​ in order to address​​​‌ the two-fold task of​ fallacious argument detection and​‌ classification through defining neural​​ network architectures based on​​​‌ Transformers models, combining text,​ argumentative features, and engineered​‌ features. More precisely, we​​ will tackle challenging fallacy​​​‌ categories like causal ones​ (e.g., Common Cause, False​‌ Cause, Reversing Causation, Post-Hoc),​​ where reasoning and knowledge-based​​​‌ features are jointly required​ to identify the fallacy.​‌ Finally, we also target​​ the usage of AM​​​‌ methods to provide an​ argumentative representation of fact-checkers'​‌ justifications to improve the​​ precision and explainability of​​​‌ fake news classification systems.​

Long​​ term objectives

The two​​​‌ last long term objectives​ are related to the​‌ generation of natural language​​ arguments.

More precisely, we​​​‌ plan to start with​ the extractive argument generation,​‌ following the example of​​ extractive summarization approaches. The​​ idea is to select​​​‌ the main arguments identified‌ in the text during‌​‌ the mining phase (Axis​​ A), and generating these​​​‌ arguments verbatim producing a‌ kind of argument-based summary‌​‌ of the original text.​​ To select the main​​​‌ arguments to generate, we‌ will exploit the quality‌​‌ criteria defined in Axis​​ B. We will incrementally​​​‌ work starting from the‌ generation of claims to‌​‌ the generation of whole​​ argumentative structures composed of​​​‌ evidence and claims linked‌ by attack and support‌​‌ relations. Secondly, we plan​​ to move to the​​​‌ abstractive argument generation. The‌ argument generation models we‌​‌ target will employ a​​ text planning module to​​​‌ conduct content selection over‌ the mined arguments and‌​‌ specify the suitable language​​ style at sentence-level in​​​‌ order to select the‌ talking points to cover‌​‌ for each sentence to​​ be generated, combined with​​​‌ a content generation module‌ to produce the final‌​‌ fluent argument on a​​ certain topic. For this​​​‌ task, we will fine-tune‌ Large Language Models for‌​‌ general application scenarios and​​ for our use cases​​​‌ (i.e., counter-argument generation to‌ fight online hate speech‌​‌ and disinformation, argument-based natural​​ language explanations). The generated​​​‌ arguments need to meet‌ two criteria, namely quality‌​‌ (e.g., variability of the​​ arguments, no repetitiveness) and​​​‌ quantity. The evaluation of‌ the generated arguments will‌​‌ be addressed by employing​​ standard metrics for natural​​​‌ language generation, i.e., ROUGE‌ (a recall-oriented metric), BLEU‌​‌ (based on n-gram precision)​​ and METEOR (measuring unigram​​​‌ precision and recall by‌ considering synonyms, paraphrases, and‌​‌ stemming).

The second long-term​​ objective is to propose​​​‌ and test a more‌ concrete argument-based dialogue, based‌​‌ on the results of​​ the previous step. First,​​​‌ the argument-based dialogue neural‌ model selects the mined‌​‌ arguments (i.e., components and​​ relations) supporting or attacking​​​‌ a given (stance about‌ a) topic and generates‌​‌ new arguments using the​​ mined ones. Then, a​​​‌ task-oriented dialogue model provides‌ the generated arguments about‌​‌ a certain topic and​​ allows the user to​​​‌ ask clarification questions. It‌ is in charge of‌​‌ generating appropriate queries to​​ the argumentation model, and​​​‌ of generating an answer.‌ The argumentation can be‌​‌ evaluated along with three​​ axes: 1) validity of​​​‌ the used inferences; 2)‌ quality of the generated‌​‌ interactions; 3) epistemic quality​​ of conclusions in a​​​‌ given domain. Measurable factors‌ such as the time‌​‌ spent by each user​​ interacting with the system​​​‌ or the number of‌ repeated questions are taken‌​‌ as indicators of this​​ model. A quantitative analysis​​​‌ of the dialogues will‌ be also conducted, based‌​‌ on different configurations of​​ the dialogue simulator.

This​​​‌ research axis can be‌ questioned with respect to‌​‌ the hype on generative​​ AI the scientific community​​​‌ is facing these days.‌ In the team, in‌​‌ the three scientific axes​​ we propose, we will​​​‌ make use of Large‌ Language Models which we‌​‌ will fine tune for​​ our precise tasks, and​​​‌ we will employ generative‌ AI models like GPT,‌​‌ Gemini, Claude or Mistral​​ for zero-shot or few-shot​​​‌ evaluations. As this issue‌ concerns more an engineering‌​‌ (i.e., prompt engineering) contribution​​​‌ than a scientific one,​ in this research program,​‌ we did not develop​​ further this part, even​​​‌ though we are obliged​ to address it to​‌ compare with state-of-the-art systems.​​

Medium-term projects

In order​​​‌ to reach the long​ term objectives, we work​‌ on the following projects:​​

• Argument-based XAI: Explainable​​​‌ Artificial Intelligence (XAI) has​ the goal to explain​‌ the decisions made by​​ an intelligent system. This​​​‌ topic has recently raised​ a lot of interest​‌ in the AI research​​ community due to the​​​‌ need to explain machine​ decisions, in particular in​‌ sensitive application domains like​​ medicine. AI models must​​​‌ be able to provide​ interpretable and robust explanations​‌ for their decisions, as​​ well as to learn​​​‌ the best way to​ explain their predictions to​‌ humans. This popular issue​​ has already been investigated​​​‌ by different communities in​ the field, and several​‌ approaches have been proposed​​ to get a better​​​‌ understanding of AI models​ from a mathematical point​‌ of view. Current results​​ emphasize that there is​​​‌ the need to generate​ explanations coping with human​‌ reasoning methods, such as​​ argumentation, and that natural​​​‌ language explanations is one​ of the most challenging​‌ ways to materialize them.​​ Our objective is to​​​‌ connect argumentation to XAI​ through the interactive generation​‌ of arguments to help​​ humans to get a​​​‌ better understanding of machine​ predictions. Arguments need not​‌ only to be rational,​​ but “manifestly” rational, so​​​‌ that arguers can see​ for themselves the rationale​‌ behind the inferential steps​​ taken. We will propose​​​‌ novel argument generation methods​ to generate natural language​‌ explanations targeting different domains​​ (i.e., medicine, fake news,​​​‌ hate speech, law, policy​ making). The important feature​‌ we target for generating​​ these natural language explanations​​​‌ is two-fold: on the​ one side, they must​‌ unveil the reasons behind​​ a decision through manifestly​​​‌ rational arguments, and they​ need to consider the​‌ human feedback to improve​​ the firstly generated explanation​​​‌ (i.e., adding more details,​ broaden the discussion, provide​‌ reference to institutional sources).​​ The generated argument-based explanations​​​‌ will benefit the user​ in terms of justification​‌ (exposing the reasoning behind​​ a decision, thus aiding​​​‌ the user to decide​ how much credence to​‌ give in it), user​​ involvement (allowing the user​​​‌ to add her knowledge​ and inference skills to​‌ the overall decision process),​​ and system acceptance (in​​​‌ that the system’s functionality​ is fully transparent and​‌ its suggestions are adequately​​ justified).
• Counter-Argument Dialogue Generation​​​‌: Existing resources highlight​ the proficiency of large​‌ language models in generating​​ counter-arguments, for instance, against​​​‌ hate speech or fake​ news. However, they primarily​‌ focus on simple two-turn​​ exchanges, whereas real-life interactions​​​‌ often involve multiple, complex​ dialogues. Furthermore, among those​‌ approaches, it still remains​​ unclear which arguments are​​​‌ crucial for steering conversations​ toward positive shifts. To​‌ address these issues, our​​ research aims to delve​​​‌ deeper into the structure​ and dynamics of counter-argumentation​‌ in human and machine-generated​​ resources extended dialogues. Our​​​‌ objectives are four-fold: i)​ explore the use and​‌ impact of argumentative structures​​ (premise, conclusion, major claims,​​ and attack/support relations) and​​​‌ rhetorical strategies (logos, pathos,‌ and ethos) in longer‌​‌ conversations; (ii) examine how​​ the generated argument structures​​​‌ may evolve across multiple‌ turns; (iii) assess how‌​‌ well existing counter-arguments taxonomies​​ align with effective argument​​​‌ generation methods; and (iv)‌ develop novel methods for‌​‌ generating counter-arguments that guarantee​​ these conversation shifts, through​​​‌ reinforcement learning with human‌ feedback (RLHF).

6.1.1 HERACLES

• Name:​​
  Online Topic Change Point​​​‌ Detection
• Keywords:
  Incremental clustering,​ Continual Learning
• Functional Description:​‌
  This script performs topic​​ modeling on chunks of​​​‌ documents using BERTopic and​ detects change points in​‌ topic distributions over time​​ using Online Change Point​​​‌ Detection (OCPD). It processes​ document chunks iteratively, updating​‌ the topic model with​​ each new chunk, and​​​‌ identifies significant changes in​ topics over time.
• Contact:​‌
  Serena Villata

6.1.2 StreamETM​​

• Keywords:
  Machine learning, Artificial​​​‌ intelligence, Natural language processing​
• Scientific Description:

  StreamETM is​‌ an application designed for​​ dynamic topic modeling using​​​‌ the Embedded Topic Model​ (ETM). It processes streaming​‌ text data, merges topic​​ models over time, and​​​‌ detects change points in​ topic distributions.

  Features: -​‌ Dynamic Topic Modeling: Continuously​​ update topic models with​​​‌ new data chunks. -​ Topic Merging: Merge new​‌ topic models with existing​​ ones to maintain a​​​‌ coherent topic structure. -​ Change Point Detection: Detect​‌ significant changes in topic​​ distributions over time using​​​‌ Online Change Point Detection​ (OCPD). - Preprocessing: Preprocess​‌ text data, including lemmatization,​​ stopword removal, and frequency-based​​​‌ filtering.

• Functional Description:
  StreamETM​ is an application designed​‌ for online topic modeling​​ using the Embedded Topic​​​‌ Model (ETM). It processes​ streaming text data, merges​‌ topic models over time,​​ and detects change points​​​‌ in topic distributions.
• News​ of the Year:
  The​‌ software has been developed​​ as part of the​​​‌ paper, Merging Embedded Topics​ with Optimal Transport for​‌ Online Topic Modeling on​​ Data Streams (https://arxiv.org/pdf/2504.07711)
• Publication:​​​‌
  hal-05468591
• Contact:
  Federica Granese​
• Participant:
  4 anonymous participants​‌

6.1.3 ACTA

• Name:
  A​​ Tool for Argumentative Clinical​​​‌ Trial Analysis
• Keywords:
  Artificial​ intelligence, Natural language processing,​‌ Argument mining
• Functional Description:​​
  Argumentative analysis of textual​​​‌ documents of various nature​ (e.g., persuasive essays, online​‌ discussion blogs, scientific articles)​​ allows to detect the​​ main argumentative components (i.e.,​​​‌ premises and claims) present‌ in the text and‌​‌ to predict whether these​​ components are connected to​​​‌ each other by argumentative‌ relations (e.g., support and‌​‌ attack), leading to the​​ identification of (possibly complex)​​​‌ argumentative structures. Given the‌ importance of argument-based decision‌​‌ making in medicine, ACTA​​ is a tool for​​​‌ automating the argumentative analysis‌ of clinical trials. The‌​‌ tool is designed to​​ support doctors and clinicians​​​‌ in identifying the document(s)‌ of interest about a‌​‌ certain disease, and in​​ analyzing the main argumentative​​​‌ content and PICO elements.‌
• Release Contributions:
  In 2024,‌​‌ ACTA has been integrated​​ in a suite called​​​‌ ANTIDOTE, which collects the‌ software results from the‌​‌ ANTIDOTE project (all partners).​​ In 2025, ACTA has​​​‌ been optimised to improve‌ the mining and analysis‌​‌ of natural language arguments​​ from clinical texts.
• URL:​​​‌
  http://­antidote.­i3s.­unice.­fr/­acta/
• Contact:
  Serena Villata‌

6.1.4 PEACE

• Name:
  Providing‌​‌ Explanations and Analysis for​​ Combating Hate Expressions
• Keywords:​​​‌
  Hate Speech Detection, Generating‌ Explanations, Implicit Hate Speech,‌​‌ Subtle Hate Speech
• Functional​​ Description:

  PEACE is a​​​‌ web tool conceived to‌ support content moderators in‌​‌ exploring and evaluating implicit​​ and subtle hate speech​​​‌ on social media. It‌ comprises three main functionalities:‌​‌ i) the exploratory analysis​​ of hate speech messages​​​‌ characteristics (exploration), ii) the‌ prediction of hatefulness (detection),‌​‌ and iii) the explanation​​ of system predictions (explanation).​​​‌

  These functionalities incorporate not‌ only a binary classification‌​‌ of whether a message​​ is hateful (including explicit,​​​‌ implicit, and subtle messages),‌ with a detailed explanation‌​‌ in natural language that​​ clarifies why a message​​​‌ is considered hateful and‌ an exploratory analysis of‌​‌ the message characteristics.

• URL:​​
  https://­3ia-demos.­inria.­fr/­peace/
• Contact:
  Elena Cabrio​​​‌

6.1.5 DispuTool 2.0

• Keywords:‌
  Argument mining, NLP, Web‌​‌ API, Relation Extraction, Component​​ Detection
• Functional Description:
  DISPUTool​​​‌ 2.0 is an automated‌ tool which relies on‌​‌ Argument Mining methods to​​ analyze the political debates​​​‌ from the US presidential‌ campaigns to extract argument‌​‌ components (i.e., premise and​​ claim) and relations (i.e.,​​​‌ support and attack), and‌ highlight fallacious arguments. DISPUTool‌​‌ 2.0 allows also for​​ the automatic analysis of​​​‌ a piece of a‌ debate proposed by the‌​‌ user to identify and​​ classify the arguments contained​​​‌ in the text. A‌ REST API is provided‌​‌ to exploit the tool's​​ functionalities.
• Release Contributions:
  The​​​‌ new version of DispuTool‌ of 2025 allows to‌​‌ "repair" a fallacious argument​​ with its non-fallacious version.​​​‌
• URL:
  https://­3ia-demos.­inria.­fr/­disputool/
• Contact:
  Serena‌ Villata

6.1.6 MARIANNE-SAFE

• Name:‌​‌
  Structured Argumentation for Fact-checking​​ with Explanations
• Keywords:
  Artificial​​​‌ intelligence, Disinformation detection
• Functional‌ Description:
  - Generate argument-structured‌​‌ summaries based on a​​ fact-checking article or a​​​‌ list of evidence. -‌ Retrieve evidence and generate‌​‌ argument-structured summaries when the​​ fact-checking article is not​​​‌ available. - Assign a‌ truthfulness label to a‌​‌ news claim which is​​ enriched with a summary.​​​‌
• Release Contributions:
  First version‌ published in 2025.
• URL:‌​‌
  https://­3ia-demos.­inria.­fr/­safe/
• Contact:
  Serena Villata​​

7.1.1 FALCON:​ A multi-label graph-based dataset​‌ for fallacy classification in​​ the COVID-19 infodemic

Participants:​​​‌ Mariana Chaves, Elena​ Cabrio, Serena Villata​‌.

Keywords: Argument mining,​​ fallacy detection, disinformation

Fallacies​​​‌ are arguments that seem​ valid but contain logical​‌ flaws. During the COVID-19​​ pandemic, they played a​​​‌ role in spreading misinformation,​ causing confusion and eroding​‌ public trust in health​​ measures. Therefore, there is​​​‌ a critical need for​ automated tools to identify​‌ fallacies in media, which​​ can help mitigate harmful​​​‌ narratives in future health​ crises. We present two​‌ key contributions to address​​ this task. First, we​​​‌ introduce FALCON 15,​ a multi-label, graph-based dataset​‌ containing COVID-19-related tweets. This​​ dataset includes expert annotations​​​‌ for six fallacy types—loaded​ language, appeal to fear,​‌ appeal to ridicule, hasty​​ generalization, ad hominem, and​​​‌ false dilemma—and allows for​ the detection of multiple​‌ fallacies in a single​​ tweet. The dataset's graph​​​‌ structure enables analysis of​ the relationships between fallacies​‌ and their progression in​​ conversations. Second, we evaluate​​​‌ the performance of language​ models on this dataset​‌ and propose a dual-transformer​​ architecture that integrates engineered​​​‌ features. Beyond model ranking,​ we conduct statistical analyses​‌ to assess the impact​​ of individual features on​​​‌ model performance.

7.1.2 AM4DSP:​ Argumentation Mining in Structured​‌ Decentralized Discussion Platforms for​​ Deliberative Democracy

Participants: Sofiane​​​‌ Elguendouze, Erwan Hain​, Elena Cabrio,​‌ Serena Villata.

Keywords:​​ Argument mining, e-Democracy, Deliberation​​​‌ Platforms

Collaborations: Lucas Anastasiou​ (The Open University (UK)),​‌ Anna de Liddo (The​​ Open University (UK))

Argument(ation)​​​‌ mining (AM) is the​ automated process of identification​‌ and extraction of argumentative​​ structures in natural language.​​​‌ This field has seen​ rapid advancements, offering powerful​‌ tools to analyze and​​ interpret complex and large​​ discourse in diverse domains​​​‌ (political debates, medical reports,‌ etc.). In this paper‌​‌ 20, we introduce​​ an AM-boosted version of​​​‌ BCause, a large-scale deliberation‌ platform. Figure 1 visualizes‌​‌ the overall achitecture of​​ the platform. The system​​​‌ enables the extraction and‌ analysis of arguments from‌​‌ online discussions in the​​ context of deliberative democracy,​​​‌ which aims to enhance‌ the understanding and accessibility‌​‌ of structured argumentation in​​ large-scale deliberation processes.

The​​​‌ AM4DSP system performs argumentation‌ analysis on structured discussions‌​‌ in BCause, a large​​ scale deliberation platform.

7.1.3 RooseBERT:​​ A New Deal For​​​‌ Political Language Modeling

Participants:‌ Deborah Dore, Elena‌​‌ Cabrio, Serena Villata​​.

Keywords: Argument mining,​​​‌ language modeling, political debates,‌ sentiment analysis, stance detection‌​‌

The increasing amount of​​ political debates and politics-related​​​‌ discussions calls for the‌ definition of novel computational‌​‌ methods to automatically analyze​​ such content with the​​​‌ final goal of lightening‌ up political deliberation to‌​‌ citizens. However, the specificity​​ of the political language​​​‌ and the argumentative form‌ of these debates (employing‌​‌ hidden communication strategies and​​ leveraging implicit arguments) make​​​‌ this task very challenging,‌ even for current general-purpose‌​‌ pre-trained Language Models. To​​ address this issue, we​​​‌ introduce a novel pre-trained‌ Language Model for political‌​‌ discourse language called Roose-BERT.​​ Pre-training a language model​​​‌ on a specialized domain‌ presents different technical and‌​‌ linguistic challenges, requiring extensive​​ computational resources and large-scale​​​‌ data. RooseBERT has been‌ trained on large political‌​‌ debate and speech corpora​​ (8K debates, each composed​​​‌ of several sub-debates on‌ different topics) in English.‌​‌ To evaluate its performances,​​ we fine-tuned it on​​​‌ four downstream tasks related‌ to political debate analysis,‌​‌ i.e., stance detection, sentiment​​ analysis, argument component detection​​​‌ and classification, and argument‌ relation prediction and classification.‌​‌ Our results demonstrate significant​​ improvements over general-purpose Language​​​‌ Models on these four‌ tasks, highlighting how domain-specific‌​‌ pre-training enhances performance in​​ political debate analysis. We​​​‌ release RooseBERT 34 for‌ the research community.

7.1.4‌​‌ Leveraging Graph Structural Knowledge​​ to Improve Argument Relation​​​‌ Prediction in Political Debates‌

Participants: Deborah Dore,‌​‌ Serena Villata.

Keywords:​​ Argument mining, political debates,​​​‌ knowledge graphs

Collaborations: Stefano‌ Faralli (Sapienza University of‌​‌ Rome)

Argument Mining (AM)​​ aims at detecting argumentation​​​‌ structures (i.e., premises and‌ claims linked by attack‌​‌ and support relations) in​​ text. A natural application​​​‌ domain is political debates,‌ where uncovering the hidden‌​‌ dynamics of a politician's​​ argumentation strategies can help​​​‌ the public to identify‌ fallacious and propagandist arguments.‌​‌ Despite the few approaches​​ proposed in the literature​​​‌ to apply AM to‌ political debates, this application‌​‌ scenario is still challenging,​​ and, more precisely, concerning​​​‌ the task of predicting‌ the relation holding between‌​‌ two argument components. Most​​ of AM relation prediction​​​‌ approaches only consider the‌ textual content of the‌​‌ argument component to identify​​ and classify the argumentative​​​‌ relation holding among them‌ (i.e., support, attack), and‌​‌ they mostly ignore the​​​‌ structural knowledge that arises​ from the overall argumentation​‌ graph. In this paper​​ 19, we propose​​​‌ to address the relation​ prediction task in AM​‌ by combining the structural​​ knowledge provided by a​​​‌ Knowledge Graph Embedding Model​ with the contextual knowledge​‌ provided by a fine-tuned​​ Language Model (see Figure​​​‌ 2). Our experimental​ setting is grounded on​‌ a standard AM benchmark​​ of televised political debates​​​‌ of the US presidential​ campaigns from 1960 to​‌ 2020. Our extensive experimental​​ setting demonstrates that integrating​​​‌ these two distinct forms​ of knowledge (i.e., the​‌ textual content of the​​ argument component and the​​​‌ structural knowledge of the​ argumentation graph) leads to​‌ novel pathways that outperform​​ existing approaches in the​​​‌ literature on this benchmark​ and enhance the accuracy​‌ of the predictions.

Architecture​​ integrating Knowledge Graphs and​​​‌ Language Models for Argument​ Component Relation Prediction and​‌ Classification.

7.1.5 Merging Embedded​ Topics with Optimal Transport​‌ for Online Topic Modeling​​ on Data Streams

Participants:​​​‌ Federica Granese, Serena​ Villata.

Keywords: Online​‌ Topic Modeling, Optimal Transport,​​ NLP

Collaborations: Benjamin Navet​​​‌ (Université Côte d'Azur), Charles​ Bouveyron (Université Côte d'Azur)​‌

Online topic models are​​ unsupervised algorithms to identify​​​‌ latent topics in data​ streams that continuously evolve​‌ over time. Although these​​ methods naturally align with​​​‌ real-world scenarios, they have​ received considerably less attention​‌ from the community compared​​ to their offline counterparts.​​​‌ In this work, we​ introduce a novel approach​‌ to online topic modeling​​ named StreamETM. This approach​​​‌ builds on the Embedded​ Topic Model (ETM) to​‌ handle data streams by​​ merging models learned on​​​‌ consecutive partial document batches​ using unbalanced optimal transport.​‌ Additionally, an online change​​ point detection algorithm is​​​‌ employed to identify shifts​ in topics over time,​‌ allowing for the detection​​ of significant changes in​​​‌ the dynamics of text​ streams 24.

7.1.6​‌ Stick-Breaking Embedded Topic Model​​ with Continuous Optimal Transport​​​‌ for Online Analysis of​ Document Streams

Participants: Federica​‌ Granese, Serena Villata​​.

Keywords: Online Topic​​​‌ Modeling, Optimal Transport, NLP​

Collaborations: Charles Bouveyron (Université​‌ Côte d'Azur)

In this​​ work, we propose the​​​‌ Stick-Breaking Stream Embedded Topic​ Model (SB-SETM), an improved​‌ version of the StreamETM​​ model presented in 24​​​‌, for the online​ topic modeling setting. SB-SETM​‌ (i) leverages a truncated​​ stick-breaking construction for the​​​‌ topic–per-document distribution, enabling the​ model to automatically infer​‌ from the data the​​ appropriate number of active​​​‌ topics at each timestep;​ and (ii) introduces a​‌ merging strategy for topic​​ embeddings based on a​​​‌ continuous formulation of optimal​ transport adapted to the​‌ high dimensionality of the​​ latent topic space. Numerical​​​‌ experiments show that SB-SETM​ outperforms baselines on simulated​‌ scenarios. We extensively test​​ it on a real-world​​​‌ corpus of news articles​ covering the Russian–Ukrainian war​‌ from 2022 to 2023​​ 30.

7.1.7 A​​​‌ Topicality-Driven QUD Model for​ Discourse Processing

Participants: Yingxue​‌ Fu, Anaïs Ollagnier​​.

Keywords: QUD, discourse​​ modeling, discourse parsing, Rhetorical​​​‌ Structure Theory, Penn Discourse‌ Treebank

Collaborations: Mark-Jan Nederhof‌​‌ (University of St Andrews)​​

Question Under Discussion (QUD)​​​‌ is a discourse framework‌ that has attracted growing‌​‌ interest in NLP in​​ recent years. Among existing​​​‌ QUD models, the QUD‌ tree approach focuses on‌​‌ reconstructing QUDs and their​​ hierarchical relationships, using a​​​‌ single tree to represent‌ discourse structure. Prior implementation‌​‌ shows moderate inter-annotator agreement,​​ highlighting the challenging nature​​​‌ of this task. In‌ this paper, we propose‌​‌ a new QUD model​​ for annotating hierarchical discourse​​​‌ structure 21. Our‌ annotation achieves high inter-annotator‌​‌ agreement: 81.45% for short​​ files and 79.53% for​​​‌ long files of Wall‌ Street Journal articles. We‌​‌ show preliminary results on​​ using GPT-4 for automatic​​​‌ annotation, which suggests that‌ one of the best-performing‌​‌ LLMs still struggles with​​ capturing hierarchical discourse structure.​​​‌ Moreover, we compare the‌ annotations with RST annotations.‌​‌ Lastly, we present an​​ approach for integrating hierarchical​​​‌ and local discourse relation‌ annotations with the proposed‌​‌ model.

7.2.1 Fast Computing of‌ Dung Semantics in Acyclic‌​‌ Probabilistic Argumentation Frameworks

Participants:​​ Victor David.

Keywords:​​​‌ Probabilistic Argumentation, Algorithm, Complexity‌

Collaborations: Stefano Bistarelli (University‌​‌ of Perugia), Pierre Monnin​​ (Inria), Francesco Santini (University​​​‌ of Perugia), Carlo Taticchi‌ (University of Perugia)

This‌​‌ paper published in 14​​ presents fast and exact​​​‌ methods for computing the‌ probability of an argument’s‌​‌ acceptance using Dung’s semantics​​ in the Constellation paradigm​​​‌ of Abstract Argumentation. For‌ (directed) Singly-Connected Graphs (SCGs),‌​‌ the problem can now​​ be solved in linearithmic​​​‌ time instead of being‌ exponential in the number‌​‌ of attacks, as reported​​ in the literature. Moreover,​​​‌ in the more general‌ case of Directed Acyclic‌​‌ Graphs (DAGs), we provide​​ an algorithm whose time​​​‌ complexity is linearithmic in‌ the product of the‌​‌ out-degree of dependent arguments,​​ i.e., arguments reaching the​​​‌ argument considered for acceptance‌ through multiple paths in‌​‌ the graph. We theoretically​​ show that this complexity​​​‌ is lower than the‌ lower bound of the‌​‌ (exact) Constellation method, which​​ is also supported by​​​‌ empirical results. Our approach‌ to DAGs is also‌​‌ compared with the (approximate)​​ Monte-Carlo method, which is​​​‌ stopped when exact results‌ are obtained. Within this‌​‌ time constraint, Monte-Carlo still​​ outputs significant errors, underlying​​​‌ the fast computation of‌ our approach.

7.2.2 A‌​‌ Logic-based Framework for Decoding​​ Enthymemes in Argument Maps​​​‌ involving Implicitness in Premises‌ and Claims

Participants: Victor‌​‌ David.

Keywords: Enthymeme,​​ Logical Argumentation, Argument mining​​​‌

Collaborations: Anthony Hunter (UCL)‌

Argument mining is a‌​‌ natural language processing technology​​ aimed at identifying the​​​‌ explicit premises and claims‌ of arguments in text,‌​‌ and the support and​​ attack relationships between them.​​​‌ To better understand and‌ automatically analyze the argument‌​‌ maps that are output​​ from argument mining, it​​​‌ would be desirable to‌ instantiate the arguments in‌​‌ the argument map with​​ logical arguments. However, most​​​‌ real-world arguments are enthymemes‌ (i.e. some of the‌​‌ premises and/or claims are​​​‌ implicit), which need to​ be decoded (i.e. the​‌ implicit aspects need to​​ be identified). A key​​​‌ challenge is to decode​ enthymemes so as to​‌ respect the support and​​ attack relationships in the​​​‌ argument map. To address​ this, we present a​‌ novel framework 18,​​ based on default logic,​​​‌ for representing arguments including​ enthymemes. We show how​‌ decoding an enthymeme means​​ identifying the default rules​​​‌ that are implicit in​ the premises and claims.​‌ We then show how​​ choosing a decoding of​​​‌ the enthymemes in an​ argument map can be​‌ formalized as an optimization​​ problem, and that a​​​‌ solution can be obtained​ using MaxSAT solvers.

7.2.3​‌ An Axiomatic Study of​​ a Modular Evaluation of​​​‌ Enthymeme Decoding in Weighted​ Structured Argumentation

Participants: Victor​‌ David.

Keywords: Enthymeme,​​ Logical Argumentation, Axiomatization

Collaborations:​​​‌ Jonathan Ben-Naim (CNRS), Anthony​ Hunter (UCL)

An argument​‌ can be seen as​​ a pair of premises​​​‌ and a claim they​ support. Human arguments are​‌ often approximate, with some​​ premises left implicit, leading​​​‌ to an implicit inference​ of the claim, i.e.,​‌ forming enthymemes. To better​​ understand and use them,​​​‌ we must decode these​ approximate enthymemes, typically by​‌ identifying missing premises to​​ make the inference explicit,​​​‌ and, as we propose,​ by also removing irrelevant​‌ content to improve argument​​ quality in specific contexts.​​​‌ Often, multiple decodings of​ an enthymeme are possible.​‌ However, no formal method​​ has yet been proposed​​​‌ for identifying higher-quality decodings.​ To pave the way,​‌ we introduce 13 six​​ types of criteria for​​​‌ evaluating aspects of decodings.​ Then, we introduce the​‌ concept of a criterion​​ measure, designed to evaluate​​​‌ decodings based on a​ specific criterion. In parallel,​‌ we define desirable properties​​ for criterion measures, referred​​​‌ to as axioms, and​ we systematically evaluate our​‌ criterion measures with respect​​ to them. Finally, we​​​‌ introduce the notion of​ quality measure that combine​‌ specific criterion measures to​​ give an overall evaluation​​​‌ of the quality of​ decodings.

7.2.4 Similarity Measures​‌ for First-Order Logical Arguments​​

Participants: Victor David.​​​‌

Keywords: Similarity Measure, First​ Order Logic, Argumentation

Collaborations:​‌ Jérôme Delobelle (Université Paris​​ Cité), Jean-Guy Mailly (Université​​​‌ de Toulouse)

Similarity in​ formal argumentation has recently​‌ gained attention due to​​ its significance in problems​​​‌ such as argument aggregation​ in semantics and enthymeme​‌ decoding. While prior work​​ has focused on propositional​​​‌ logic arguments, we extend​ these approaches to First-Order​‌ Logic (FOL) arguments, enabling​​ reasoning based on the​​​‌ similarity of arguments in​ more complex and realistic​‌ contexts. We present a​​ comprehensive framework 17 for​​​‌ FOL argument similarity, including:​ 1. An extended axiomatic​‌ foundation for similarity measures,​​ 2. A parametric model​​​‌ decomposed into four levels​ to efficiently evaluate structured​‌ knowledge, 3. A Tversky-based​​ family of measures to​​​‌ instantiate these concepts, 4.​ A set of constraints​‌ ensuring well-behaved models that​​ satisfy axioms, and 5.​​​‌ We introduce and analyze​ non-symmetric similarity measures in​‌ formal argumentation for the​​ first time.

7.2.5 DataLens:​​​‌ Enhancing Dataset Discovery via​ Network Topologies

Participants: Anaïs​‌ Ollagnier.

Keywords: Dataset​​ search, Network visualization, Faceted​​ search

Collaborations: Aline Menin​​​‌ (Université Côte d'Azur)

The‌ rapid growth of publicly‌​‌ available textual resources, such​​ as lexicons and domain-specific​​​‌ corpora, presents challenges in‌ efficiently identifying relevant resources.‌​‌ While repositories are emerging,​​ they often lack advanced​​​‌ search and exploration features.‌ Most search methods rely‌​‌ on keyword queries and​​ metadata filtering, which require​​​‌ prior knowledge and fail‌ to reveal connections between‌​‌ resources. To address this,​​ we present DataLens, a​​​‌ web-based platform that combines‌ faceted search with advanced‌​‌ visualization techniques to enhance​​ resource discovery. DataLens offers​​​‌ network-based visualizations, where the‌ network structure can be‌​‌ adapted to suit the​​ specific analysis task. It​​​‌ also supports a chained‌ views approach, enabling users‌​‌ to explore data from​​ multiple perspectives. A formative​​​‌ user study involving six‌ data practitioners revealed that‌​‌ users highly value visualization​​ tools—especially network-based exploration—and offered​​​‌ insights to help refine‌ our approach to better‌​‌ support dataset search.

7.2.6​​ Mining Implicit Arguments for​​​‌ Reasoning : A Survey‌

Participants: Ekaterina Sviridova,‌​‌ Elena Cabrio, Serena​​ Villata.

Keywords: Argument​​​‌ mining, Implicitness

Argumentation is‌ the process of creating‌​‌ arguments for and against​​ competing claims. Computational argumentation​​​‌ involves different ways of‌ analyzing and reasoning upon‌​‌ arguments and their relations.​​ More precisely, Argument Mining​​​‌ is the research field‌ aiming at automatically identifying‌​‌ and classifying argument structures​​ in text. The research​​​‌ field is mainly focused‌ on the extraction of‌​‌ explicit argument structures (i.e.,​​ claims and premises connected​​​‌ by support and attack‌ relations). However, an even‌​‌ more challenging task consists​​ in extracting implicit argument​​​‌ structures in text (e.g.,‌ enthymemes). These structures are‌​‌ particularly valuable to then​​ address argument reasoning, e.g.,​​​‌ on incomplete and uncertain‌ information, to finally compute‌​‌ the set of acceptable​​ arguments, i.e., argument justification​​​‌ and skepticism. In this‌ paper 11, we‌​‌ present and compare current​​ approaches and available datasets​​​‌ for the novel task‌ of Implicit Argument Mining.‌​‌ Future work perspectives are​​ discussed to pave the​​​‌ way to further studies‌ in this direction.

7.2.7‌​‌ Before the Outrage: Challenges​​ and Advances in Predicting​​​‌ Online Antisocial Behavior

Participants:‌ Anaïs Ollagnier.

Keywords:‌​‌ Antisocial behavior prediction, Systematic​​ literature review, Abusive behavior​​​‌

Antisocial behavior (ASB) on‌ social media—including hate speech,‌​‌ harassment, and trolling—poses escalating​​ challenges for platform safety​​​‌ and societal well-being. While‌ prior research has largely‌​‌ focused on detecting harmful​​ content after its appearance,​​​‌ predictive approaches seek to‌ anticipate harmful behaviors—such as‌​‌ hate speech propagation, conversation​​ derailment, or user recidivism—before​​​‌ they fully unfold. Despite‌ growing scholarly attention, the‌​‌ field of ASB prediction​​ remains structurally uncoordinated and​​​‌ conceptually diffuse, lacking a‌ unified taxonomy or integrative‌​‌ synthesis. This paper 36​​ addresses this gap through​​​‌ a systematic review of‌ 49 machine learning studies‌​‌ on ASB prediction published​​ between 2010 and 2025.​​​‌ It introduces a coherent‌ framework that organizes existing‌​‌ work into five core​​ task types: early harm​​​‌ detection, harm emergence prediction,‌ harm propagation prediction, behavioral‌​‌ risk prediction, and proactive​​ moderation support. The analysis​​​‌ compares tasks by temporal‌ framing, prediction granularity, and‌​‌ operational goals, while tracing​​​‌ methodological trends from classical​ machine learning to pre-trained​‌ language models. By consolidating​​ fragmented research and clarifying​​​‌ conceptual boundaries, this review​ establishes a structured foundation​‌ for future work. It​​ identifies key challenges—including dataset​​​‌ scarcity, limited benchmarks, and​ temporal drift—and calls for​‌ ASB-specific evaluation protocols, multimodal​​ and pragmatically informed representations,​​​‌ and explainable human-in-the-loop systems​ to advance predictive moderation.​‌

7.2.8 Addressing Antisocial Behavior​​ in Multi-Party Dialogs Through​​​‌ Multimodal Representation Learning

Participants:​ Hajar Bakarou, Mohamed​‌ Sinane El Messoussi,​​ Anaïs Ollagnier.

Keywords:​​​‌ Social networks, antisocial behavior,​ multi-party dialogs, multimodal representation​‌ learning

Antisocial behavior (ASB)​​ on social media, including​​​‌ hate speech, harassment, and​ cyberbullying, poses growing risks​‌ to platform safety and​​ societal well-being. Prior research​​​‌ has focused largely on​ networks such as X​‌ and Reddit, while multi-party​​ conversational settings remain underexplored​​​‌ due to limited data.​ To address this gap,​‌ we use CyberAgressionAdo-Large 32​​, a French open-access​​​‌ dataset simulating ASB in​ multi-party conversations, and evaluate​‌ three tasks: abuse detection,​​ bullying behavior analysis, and​​​‌ bullying peer-group identification. We​ benchmark six text-based and​‌ eight graph-based representation-learning methods,​​ analyzing lexical cues, interactional​​​‌ dynamics, and their multimodal​ fusion. Results show that​‌ multimodal models outperform unimodal​​ baselines. The late fusion​​​‌ model mBERT + WD-SGCN​ achieves the best overall​‌ results, with top performance​​ on abuse detection (0.718)​​​‌ and competitive scores on​ peer-group identification (0.286) and​‌ bullying analysis (0.606). Error​​ analysis highlights its effectiveness​​​‌ in handling nuanced ASB​ phenomena such as implicit​‌ aggression, role transitions, and​​ context-dependent hostility.

7.2.9 Effectiveness​​​‌ of Counter-Speech against Abusive​ Content: A Multidimensional Annotation​‌ and Classification Study

Participants:​​ Greta Damo, Elena​​​‌ Cabrio, Serena Villata​.

Keywords: Hate speech,​‌ counter speech, social media​​

The research presents a​​​‌ novel computational framework for​ evaluating the effectiveness of​‌ counter-speech (CS) in mitigating​​ online hate speech  16​​​‌. Grounded in linguistics,​ communication, and argumentation theories,​‌ the framework defines six​​ core dimensions – Clarity,​​​‌ Evidence, Emotional Appeal, Rebuttal,​ Audience Adaptation, and Fairness​‌ – used to annotate​​ 4,214 counter-speech instances from​​​‌ two benchmark datasets. A​ human-annotated resource was created​‌ and released to the​​ community, enabling further research​​​‌ in this area. The​ study also proposes two​‌ classification strategies, multi-task and​​ dependency-based models, which outperform​​​‌ standard baselines, achieving average​ F1 scores of 0.94​‌ and 0.96, respectively, across​​ expert- and user-written CS.​​​‌ Results highlight the interdependence​ of effectiveness dimensions and​‌ provide a structured approach​​ for assessing counter-speech, with​​​‌ potential societal impact in​ reducing online hostility and​‌ supporting safer digital discourse.​​

7.2.10 `Detectors Lead, LLMs​​​‌ Follow`: Integrating LLMs and​ traditional models on implicit​‌ hate speech detection to​​ generate faithful and plausible​​​‌ explanations

Participants: Nicolás Benjamín​ Ocampo, Greta Damo​‌, Elena Cabrio,​​ Serena Villata.

Keywords:​​​‌ Hate speech, explainability, Large​ Language Models

The journal​‌ paper addresses the challenge​​ of detecting and explaining​​​‌ implicit hate speech (HS)​ on social media 9​‌. It proposes a​​ novel framework that enhances​​​‌ Large Language Models (LLMs)​ for both classification and​‌ natural language explanation of​​ hateful content, particularly targeting​​ nuanced and implicit instances​​​‌ often missed by traditional‌ methods. The approach combines‌​‌ binary classification (HS vs.​​ Non-HS) with generated explanations,​​​‌ and further investigates the‌ impact of incorporating information‌​‌ from BERT-based models on​​ detection performance. Comprehensive evaluation​​​‌ on widely used datasets,‌ the HateCheck and Implicit‌​‌ Hate Corpus datasets, demonstrates​​ that LLMs coupled with​​​‌ explanations outperform classical detectors,‌ improving both accuracy and‌​‌ interpretability. The work provides​​ insights into how predictive​​​‌ and explanatory systems can‌ support automated moderation and‌​‌ enhance understanding of implicit​​ hateful messages.

7.2.11 DISPUTool​​​‌ 3.0: Fallacy Detection and‌ Repairing in Argumentative Political‌​‌ Debates

Participants: Deborah Dore​​, Elena Cabrio,​​​‌ Serena Villata.

Keywords:‌ Argument mining, political debates,‌​‌ fallacy detection, fallacy repairing​​

Collaborations: Pierpaolo Goffredo (ALTEN)​​​‌

This paper introduces and‌ evaluates a novel web-based‌​‌ application designed to identify​​ and repair fallacious arguments​​​‌ in political debates. DISPUTool‌ 3.0 22 offers a‌​‌ comprehensive tool for argumentation​​ analysis of political debate,​​​‌ integrating state-of-the-art natural language‌ processing techniques to mine‌​‌ and classify argument components​​ and relations (see Figure​​​‌ 3). DISPUTool 3.0‌ builds on the ElecDeb60to20‌​‌ dataset, covering US presidential​​ debates from 1960 to​​​‌ 2020. In this paper,‌ we introduce a novel‌​‌ task which is integrated​​ as a new module​​​‌ in DISPUTool, i.e., the‌ automatic detection and classification‌​‌ of fallacious arguments, and​​ the automatic repairing of​​​‌ such misleading arguments. The‌ goal is to show‌​‌ to the user a​​ tool which not only​​​‌ identifies fallacies in political‌ debates, but also shows‌​‌ how the argument looks​​ like once the veil​​​‌ of fallacy falls down.‌ An extensive evaluation of‌​‌ the module is addressed​​ employing both automated metrics​​​‌ and human assessments. With‌ the inclusion of this‌​‌ module, DISPUTool 3.0 advances​​ even more user critical​​​‌ thinking in front of‌ the augmenting spread of‌​‌ such nefarious kind of​​ content in political debates​​​‌ and beyond.

DispuTOOL 3.0's‌ Architecture.

7.2.12 Repairing‌ Fallacious Argumentation in Political‌​‌ Debates

Participants: Deborah Dore​​, Elena Cabrio,​​​‌ Serena Villata.

Keywords:‌ Argument mining, political debates,‌​‌ fallacy detection, fallacy repairing​​

Collaborations: Pierpaolo Goffredo (ALTEN)​​​‌

Fallacious arguments are defined‌ as “invalid” arguments (e.g.,‌​‌ the conclusion does not​​ follow from the premises)​​​‌ or wrong moves in‌ argumentative discourse. This kind‌​‌ of argumentation is therefore​​ misleading or deceptive, in​​​‌ particular when employed in‌ political debates. As the‌​‌ spreading of this nefarious​​ content severely impacts the​​​‌ society and the decision-making‌ of both citizens and‌​‌ policymakers, it is vital​​ to prevent fallacious and​​​‌ propagandist arguments to circulate.‌ To address this challenging‌​‌ task, several approaches proposing​​ to identify fallacious argumentation​​​‌ in text have been‌ presented in the literature.‌​‌ However, merely identifying this​​ content is insufficient to​​​‌ ensure the audience realizes‌ the impact of the‌​‌ fallacious argument on its​​ deliberation process and to​​​‌ support the development of‌ critical thinking skills. To‌​‌ tackle this challenging goal,​​ it is necessary to​​​‌ unveil why a particular‌ argument is fallacious and‌​‌ to demonstrate how it​​​‌ could be repaired as​ a valid, non-fallacious argument.​‌ In this paper 23​​, we address this​​​‌ key challenge by proposing​ a new task called​‌ repairing fallacious argumentation. The​​ goal of this task​​​‌ is to modify statements​ that contain fallacious arguments​‌ into versions that are​​ clearer, fairer, and free​​​‌ from any technique that​ could negatively persuade listeners.​‌ We carry out this​​ task on political debates,​​​‌ where the need for​ this kind of solution​‌ is urgent. Our contribution​​ in addressing this task​​​‌ is manifold: i) a​ novel dataset, FallacyFix, comprising​‌ repaired examples across various​​ fallacy categories (Appeal to​​​‌ Fear, Appeal to Pity,​ Appeal to Popular Opinion,​‌ Flag Waving, and Loaded​​ Language) based on the​​​‌ ElecDeb60to20-fallacy dataset; ii) modular​ prompt techniques for generating​‌ non-fallacious arguments, both dependent​​ and independent of the​​​‌ specific fallacy label being​ addressed. Through an extensive​‌ evaluation, we assess these​​ techniques using the most​​​‌ widely used Large Language​ Models (in Zero-Shot, Few-Shot,​‌ and Fine-Tuning settings) and​​ a standard baseline model​​​‌ (BART); iii) a rigorous​ evaluation framework to assess​‌ the accuracy of the​​ generated non-fallacious argument repairing​​​‌ the fallacy in the​ original argument, with respect​‌ to the manually annotated​​ benchmark of non-fallacious arguments​​​‌ we built from the​ ElecDeb60to20 dataset; iv) a​‌ human evaluation of the​​ generated non-fallacious arguments to​​​‌ assess the acceptability of​ these arguments across three​‌ dimensions, i.e., Relevance, Suitability,​​ and Cogency. Future research​​​‌ will focus on integrating​ domain-specific knowledge to address​‌ complex fallacy categories, further​​ analyzing language models’ behavior​​​‌ in countering fallacies, and​ exploring real-time fallacy repair​‌ methodologies. These efforts aim​​ to enhance our ability​​​‌ to address fallacies dynamically​ in various argumentation contexts,​‌ potentially improving the quality​​ of public discourse and​​​‌ decision-making.

7.2.13 Contextualizing Toxicity:​ An Annotation Framework for​‌ Unveiling Pragmatics in Conversations​​ of Online Discussion Forums​​​‌

Participants: Yingxue Fu,​ Anaïs Ollagnier.

Keywords:​‌ pragmatics ; toxicity ;​​ annotation ; Reddit conversations​​​‌

The role of context​ has attracted increasing attention​‌ in research on toxicity​​ detection. Interpreting toxic language​​​‌ remains a complex and​ multifaceted challenge, shaped by​‌ numerous linguistic, contextual, and​​ social factors. However, current​​​‌ approaches often define “context”​ narrowly, focusing primarily on​‌ surface lexical cues such​​ as hate lexicons, profanity​​​‌ markers, or sentiment polarity.​ These features, while useful,​‌ are insufficient to capture​​ the interactional dynamics, user​​​‌ behaviors, and intentionality that​ shape such phenomena. To​‌ address this gap, this​​ paper introduces a novel​​​‌ and systematic annotation framework​ 35, grounded in​‌ Speech Act Theory, aimed​​ at deciphering the illocutionary​​​‌ and perlocutionary dimensions of​ conversation, which are unexplored​‌ in existing studies. We​​ apply this framework to​​​‌ a new dataset of​ complete Reddit conversation threads,​‌ sampled to include discussions​​ that turn toxic (124​​​‌ conversations, 1990 messages). We​ evaluate the performance of​‌ GPT models (GPT-3, GPT-4,​​ and GPT-5) on this​​​‌ challenging annotation task, providing​ insights into how large​‌ language models capture pragmatic​​ and contextual dimensions of​​​‌ online toxicity.

7.3.1 Beating Harmful Stereotypes‌​‌ Through Facts: RAG-based Counter-speech​​ Generation

Participants: Greta Damo​​​‌, Elena Cabrio,‌ Serena Villata.

Keywords:‌​‌ Hate speech, counter speech,​​ Retrieval-Augmented Generation

The study​​​‌ introduces a novel knowledge-grounded‌ framework for automatic counter-speech‌​‌ (CS) generation 33.​​ The framework leverages advanced​​​‌ Retrieval-Augmented Generation (RAG) pipelines‌ to produce trustworthy and‌​‌ factual counter-speech for eight​​ target groups commonly affected​​​‌ by hate speech, including‌ women, people of color,‌​‌ persons with disabilities, migrants,​​ Muslims, Jews, LGBT persons,​​​‌ and others. A knowledge‌ base of 32,792 texts‌​‌ from the United Nations​​ Digital Library, EUR-Lex, and​​​‌ the EU Agency for‌ Fundamental Rights supports the‌​‌ generation process. Evaluation on​​ the MultiTarget-CONAN dataset, using​​​‌ both automated and LLM-based‌ metrics, and human assessments,‌​‌ demonstrates that the framework​​ outperforms standard large language​​​‌ model baselines and competitive‌ approaches. The resulting system‌​‌ and knowledge base provide​​ a reusable resource for​​​‌ research on effective and‌ reliable counter-speech generation.

7.3.2‌​‌ Overview of the Critical​​ Questions Generation Shared Task​​​‌

Participants: Ekaterina Sviridova,‌ Elena Cabrio, Serena‌​‌ Villata.

Collaborations: Blanca​​ Calvo Figueras (University of​​​‌ the Basque Country UPV/EHU),‌ Jaione Bengoetxea (University of‌​‌ the Basque Country UPV/EHU),​​ Maite Heredia (University of​​​‌ the Basque Country UPV/EHU),‌ Rodrigo Agerri (University of‌​‌ the Basque Country UPV/EHU)​​

The proliferation of AI​​​‌ technologies has reinforced the‌ importance of developing critical‌​‌ thinking skills. One promising​​ direction involves leveraging Large​​​‌ Language Models (LLMs) to‌ support the generation of‌​‌ critical questions: inquiries aimed​​ at identifying weaknesses or​​​‌ fallacies in argumentative texts.‌ The Critical Questions Generation‌​‌ (CQs-Gen) shared task 29​​ offers the first benchmark​​​‌ for this task. Thirteen‌ participating teams explored diverse‌​‌ strategies for generating such​​ questions, with the highest-performing​​​‌ system achieving 67.6 %‌ accuracy, highlighting the challenge‌​‌ and complexity of the​​ task. Notably, three of​​​‌ the four top-ranked submissions‌ used argumentation scheme annotations‌​‌ to enhance performance. While​​ most systems were based​​​‌ on open-weight LLMs, the‌ two leading teams employed‌​‌ proprietary models.

7.3.3 Argument​​ generation for fact-checking

Participants:​​​‌ Xiaoou Wang, Elena‌ Cabrio, Serena Villata‌​‌.

Keywords: Argument generation,​​ RAG, fact-checking, disinformation detection.​​​‌

The widespread availability of‌ the internet and social‌​‌ media platforms has fundamentally​​ transformed how people consume​​​‌ and share information. This‌ transformation has brought with‌​‌ it a significant challenge:​​ online disinformation, i.e., namely,​​​‌ the intentional spread of‌ false or misleading information.‌​‌ The speed and scale​​ at which disinformation circulates​​​‌ present two major hurdles:‌ (1) the development of‌​‌ automated fact-checking algorithms, and​​ (2) the generation of​​​‌ explanations to accompany these‌ systems. Explanations are crucial‌​‌ for several reasons: they​​ help convince readers of​​​‌ the interpretation of evidence;‌ they enable a feedback‌​‌ loop to correct judgment​​ errors; and they reduce​​​‌ the risk of the‌ “backfire effect,” where opaque‌​‌ predictions from black-box models​​ actually strengthen belief in​​​‌ false claims. Currently, explanations‌ are typically derived either‌​‌ from human-written fact-checking articles​​ or from lists of​​​‌ evidence extracted from textual‌ corpora. Meanwhile, the field‌​‌ of Argument Mining offers​​​‌ promising tools to enhance​ fact-checking. In the context​‌ of the ANR ATTENTION​​ project, we enriched a​​​‌ subset of the LIAR-PLUS​ dataset with argumentative annotations,​‌ creating LIARArg 12 the​​ first fake news classification​​​‌ dataset enhanced with argumentative​ layers. It includes 2,832​‌ news titles with justifications,​​ comprising 3,956 claims, 7,130​​​‌ premises, and 8,205 argument​ relations. Moreover, we introduced​‌ a novel summarization method​​ 27 that generates argument-structured​​​‌ explanations, significantly boosting the​ performance of top fact-checking​‌ algorithms across multiple datasets—including​​ those with human-written fact-checks​​​‌ (LIAR-PLUS, FNC-1, Check-Covid) and​ those requiring automatic evidence​‌ retrieval (ExClaim). This retrieval-summarization​​ framework, called SAFE 28​​​‌, is also made​ available via a user-facing​‌ web interface. Our work​​ bridges Argument Mining and​​​‌ automated fact-checking to create​ systems that are not​‌ only more effective, but​​ also more transparent and​​​‌ persuasive. We showed that​ argumentative structures in evidence​‌ can improve both classification​​ and explanation.

7.4.1 On​​ Estimating the Strength of​​​‌ Differentially Private Mechanisms in​ a Black-Box Setting

Participants:​‌ Federica Granese.

Keywords:​​ Differential Privacy , Histogram-based​​​‌ Sampling, Impossibility of Provable​ Guarantees

Collaborations: Daniele Gorla​‌ (Sapienza University of Rome),​​ Louis Jalouzot (CEA), Catuscia​​​‌ Palamidessi (Inria), Pablo Piantanida​ (CNRS)

We analyze to​‌ what extent final users​​ can infer information about​​​‌ the level of protection​ of their data when​‌ the data obfuscation mechanism​​ is a priori unknown​​​‌ to them (the so-called​ “black-box” scenario). In particular,​‌ we explore four notions​​ of differential privacy. On​​​‌ the one hand, we​ prove that, without any​‌ assumptions on the underlying​​ distributions, it is not​​​‌ possible to have an​ algorithm that can infer​‌ the level of data​​ protection with provable guarantees.​​​‌ On the other hand,​ we demonstrate that, under​‌ reasonable assumptions (namely Lipschitzness​​ of the involved densities​​​‌ on a closed interval),​ such guarantees exist for​‌ the local versions and​​ can be achieved by​​​‌ a simple histogram-based estimator.​ We validate our results​‌ experimentally and note that,​​ in two particularly well-behaved​​​‌ distributions (namely the Laplace​ and the Gaussian noise),​‌ our method performs better​​ than expected, in the​​​‌ sense that in practice​ the number of samples​‌ needed to achieve the​​ desired confidence is smaller​​​‌ than the theoretical bound,​ and the estimate of​‌ $ϵ$ is more precise​​ than predicted 10.​​​‌

8.1.1 Associate Teams in​ the framework of an​‌ Inria International Lab or​​ in the framework of​​​‌ an Inria International Program​

Victor David coordinated the​‌ activity of the Inria–UCL​​ associated team EXPLAINER, which​​​‌ focuses on argumentation, natural​ language processing, logic, and​‌ neuro-symbolic approaches involving large​​ language models. The team​​​‌ addresses the challenge of​ implicit arguments (enthymemes), which​‌ are pervasive in human​​ communication but poorly handled​​ by current argument mining​​​‌ and learning-based methods. By‌ combining NLP with logical‌​‌ representations and automated reasoning,​​ the goal is to​​​‌ enable principled analysis of‌ argumentative structures, including consistency,‌​‌ validity, and relations between​​ arguments. A central objective​​​‌ is the decoding and‌ evaluation of enthymemes. Other‌​‌ team members working within​​ the EXPLAINER team are​​​‌ Elena Cabrio and Serena‌ Villata .

In 2025,‌​‌ the collaboration has already​​ yielded strong outcomes, including​​​‌ two papers accepted at‌ top-tier A* conferences. Through‌​‌ the EXPLAINER team, funding​​ was secured for a​​​‌ two-year postdoctoral position starting‌ in January 2026, dedicated‌​‌ to enthymeme decoding using​​ LLMs and argumentation schemes.​​​‌ In parallel, additional projects‌ enabled the recruitment of‌​‌ a six-month intern followed​​ by a six-month research​​​‌ engineer position, with a‌ view toward a PhD‌​‌ on the persuasiveness of​​ enthymemes. Further support was​​​‌ obtained through a one-year‌ postdoctoral position funded by‌​‌ the IDEX AMI-Idées call​​ of Université Côte d’Azur,​​​‌ focused on evaluating enthymeme‌ decoding quality.

8.1.2 Participation‌​‌ in other International Programs​​

Serena Villata co-led with​​​‌ Jaime Sichman (USP, Brasil)‌ the UNBIAS Team (argUmentation‌​‌ and Norm Based Intelligent​​ AgentS) within the IRC​​​‌ CNRS-USP (UNBIAS website‌). The scientific goal‌​‌ of the UNBIAS team​​ is to incorporate AI​​​‌ argumentation and normative agents‌ techniques to provide better‌​‌ interactions in social-technical systems.​​

8.2.1‌ Visits of international scientists‌​‌

The collaboration with UCL​​ within the EXPLAINER associated​​​‌ team was also strengthened‌ by a one-month visit‌​‌ from Anthony Hunter in​​ June-July 2025 (funded by​​​‌ the I3S lab), and‌ by the submission of‌​‌ a JCJC ANR proposal​​ directly linked to the​​​‌ EXPLAINER project.

8.2.2 Visits‌ to international teams

Federica‌​‌ Granese participated in the​​ first edition of the​​​‌ IVADO–Inria Initiative: Franco–Quebec Exchange‌ Semester on Artificial Intelligence,‌​‌ a joint program designed​​ to strengthen collaboration between​​​‌ AI research communities in‌ France and Québec. The‌​‌ program provided financial support​​ for a five-week research​​​‌ stay (mid-November to mid-December‌ 2025) at the CNRS‌​‌ International Laboratory on Learning​​ Systems (ILLS) and Mila​​​‌ – Quebec AI Institute‌ (Montréal, Québec, Canada). The‌​‌ research project addressed emerging​​ safety risks in multi-agent​​​‌ AI systems, with a‌ particular emphasis on modeling‌​‌ manipulative interactions between AI​​ agents. The project involved​​​‌ Federica Granese and Serena‌ Villata (MARIANNE, Inria-I3S) and‌​‌ Pablo Piantanida (ILLS, Mila).​​

8.3.1 Horizon Europe

The​​​‌ team participated to one‌ European project in 2025.‌​‌

• ORBIS project (2022-2025) ORBIS​​ addresses the disconnects between​​​‌ ambitious ideas and collective‌ actions at a large‌​‌ socio-technical scale. It responds​​ to the profound lack​​​‌ of dialogue between citizenship‌ and policy making institutions‌​‌ by providing a theoretically​​ sound and highly pragmatic​​​‌ socio-technical solution to enable‌ the transition to a‌​‌ more inclusive, transparent and​​ trustful Deliberative Democracy in​​​‌ Europe. The project shapes‌ and supports new democratic‌​‌ models that are developed​​ through deliberative democracy processes;​​​‌ it follows a socio-constructive‌ approach in which deliberative‌​‌ democracy is not a​​​‌ theory which prescribes new​ democratic practices and models,​‌ but rather the process​​ through which we can​​​‌ collectively imagine and realize​ them. ORBIS provides new​‌ ways to understand and​​ facilitate the emergence of​​​‌ new participatory democracy models,​ together with the mechanisms​‌ to scale them up​​ and consolidate them at​​​‌ institutional level. It delivers:​ (i) a sound methodology​‌ for deliberative participation and​​ co-creation at scale; (ii)​​​‌ novel AI-enhanced tools for​ deliberative participation across diverse​‌ settings; (iii) a novel​​ socio-technical approach that augments​​​‌ the articulation between deliberative​ processes and representative institutions​‌ in liberal democracies; (iv)​​ new evidence-based democratic models​​​‌ that emerge from the​ application of citizen deliberation​‌ processes; (v) demonstrated measurable​​ impact of such innovations​​​‌ in real-world settings. The​ project builds on cutting-edge​‌ AI tools and technologies​​ to develop a sustainable​​​‌ digital solution to e-deliberation.​ In this project, the​‌ MARIANNE team contributes with​​ the development of argumentation-based​​​‌ solutions to analyze in​ a transparent manner the​‌ discussions and enhance decision​​ making. The MARIANNE members​​​‌ involved in the project​ are: Sofiane Elguendouze ,​‌ Erwan Hain , Elena​​ Cabrio and Serena Villata​​​‌ . Link to the​ project webpage

9.1.1 Scientific​​​‌ events: organization

9.1.2 Scientific events:‌ selection

9.1.3​ Journal

9.1.4​​​‌ Invited talks

Federica Granese​ delivered the following invited​‌ talks:

• When Pragmatic AI​​ Agents Misalign: Safety Risks​​​‌ in Multi-Agent AI Systems​ – A Focus on​‌ Deceptive Behavior. Inria–IVADO Closing​​ Event, December 2–3, 2025,​​​‌ Montréal, Québec, Canada. Franco–Canadian​ Dialogue on AI. Round​‌ table on AI and​​ Safety, November 28, 2025,​​​‌ Mila – Quebec AI​ Institute, Montréal, Québec, Canada;​‌ English.
• When Pragmatic AI​​ Agents Misalign: Safety Risks​​​‌ in Multi-Agent AI Systems​ – A Focus on​‌ Deceptive Behavior. Invited seminar,​​ International Laboratory on Learning​​​‌ Systems (ILLS), November 27,​ 2025, Montréal, Québec, Canada.​‌

Serena Villata delivered the​​ following invited talks:

• Generative​​​‌ AI for fighting online​ abusive content: results and​‌ open challenges, Séminaire​​ IA Génératives: Promesses et​​​‌ Défis, CNRS, March 12th,​ 2025.
• Argumentation for medicine:​‌ extracting arguments and generating​​ explanations for a better​​​‌ human understanding, Inria​ Meetup NLP Santé, April​‌ 8th, 2025.
• Intelligence Artificielle​​ : questions techniques et​​ d’éthique, Journée d’étude​​​‌ sur « l’impact de‌ l’IA sur la pratique‌​‌ de l’enseignement universitaire »Académie​​ Royale de Belgique, Bruxelles,​​​‌ September 29th, 2025.
• L‘IA‌ contre les fake news‌​‌ et le cyberharcèlement,​​ Allianz Conference, Paris, October​​​‌ 16th, 2025.
• Generative AI:‌ technical and ethical challenges‌​‌ in game applications,​​ First Workshop on Artificial​​​‌ Intelligence for Human-Game Interaction,‌ co-located with the 28th‌​‌ European Conference of Artificial​​ Intelligence (ECAI ?25) Bologna,​​​‌ October 25th-30th, 2025.

9.1.5‌ Leadership within the scientific‌​‌ community

• Elena Cabrio is​​ a chairholder of the​​​‌ 3IA Institute, project title:‌ “Advanced Natural Language Understanding”‌​‌ (Chair renewed in 2025).​​
• Anais Ollagnier is member​​​‌ of the Academy 1‌ Board: Networks, Information, Digital‌​‌ Society (term 2025–2026).
• Anais​​ Ollagnier is contributor to​​​‌ the European EMAI4EU project,‌ as part of the‌​‌ EIT Digital Master School,​​ a European master's-level training​​​‌ program in deep-tech technologies,‌ combining academic excellence, entrepreneurship,‌​‌ and innovation, and involving​​ a network of university​​​‌ partners across Europe.
• Serena‌ Villata is a chairholder‌​‌ of the 3IA Institute,​​ project title: “Artificial Argumentation​​​‌ for Humans”.
• Since June‌ 2025, Serena Villata is‌​‌ nominated member of the​​ new Conseil de l'intelligence​​​‌ artificielle et du numerique‌ CIANum. The Conseil‌​‌ is an independent instance​​ reporting to the Minister​​​‌ responsible for artificial intelligence‌ and digital technology. Its‌​‌ mission is to study​​ all issues relating to​​​‌ the development of digital‌ technology and artificial intelligence,‌​‌ as well as their​​ impact on society, the​​​‌ economy, and local areas.‌
• Till June 2025, Serena‌​‌ Villata has been member​​ of the scientific committee​​​‌ of the Direction Generale‌ des Finances Publiques. The‌​‌ committee had the goal​​ to discuss with the​​​‌ Direction Generale des Finances‌ Publiques about their strategy‌​‌ and open issues about​​ digital transformation and data​​​‌ science.

9.1.6 Scientific expertise‌

Elena Cabrio has been:‌​‌

• Reviewer for the ERC​​ Advanced Grant 2024 Call.​​​‌
• Reviewer for Research Foundation‌ Flanders (FWO).
• Reviewer for‌​‌ the scientific project call​​ from COMUE Université de​​​‌ Toulouse.
• Expert tutor for‌ the AI Projects of‌​‌ the Fondation de recherche​​ pour l'aeronautique et l'espace​​​‌ and IRT ANTOINE SAINT‌ EXUPERY.
• Member of the‌​‌ advisory committee of the​​ MultiPOD projet (Multilingual and​​​‌ Multicultural Spaces for Political‌ Deliberation), funded by the‌​‌ EU.
• Member of the​​ AI Advisory Council of​​​‌ the company ENGIE Energy.‌
• Member of the Scientific‌​‌ Committee of ISTEX.fr (CNRS),​​ preparing the platform's application​​​‌ as a national research‌ infrastructure.

Anais Ollagnier has‌​‌ been member of the​​ selection committee for:

• the​​​‌ scientific projects AAP 2025‌ – DIM AI4IDF
• the‌​‌ ATALA thesis prize, 2025​​

Serena Villata has been​​​‌ member of:

• the recruitment‌ committee for the Inria‌​‌ Paris CRCN-ISFP positions, 2025​​ (7 applications to review,​​​‌ participation to the jury‌ meetings, oral interviews of‌​‌ the pre-selected candidates, deliberation).​​
• the HCRES committee for​​​‌ the LTCI Laboratory, 2025‌ (3 teams to evaluate,‌​‌ participation to the committee​​ meetings, oral interviews of​​​‌ the teams, report writing).‌

9.1.7 Research administration

• Elena‌​‌ Cabrio is Appointed Member​​ of the Academic Board​​​‌ of Université Côte d’Azur.‌
• Victor David managed the‌​‌ Inria–UCL associated team.
• Anais​​​‌ Ollagnier obtained and managed​ research and training funding,​‌ more precisely, the Call​​ for Expressions of Interest​​​‌ in "Digital Transformation" for​ the Introduction to Deep​‌ Learning course, and a​​ six-month internship grant (AAP​​​‌ MIDI 02) on the​ DataLens project (a web-based​‌ platform that integrates faceted​​ search with advanced information​​​‌ visualization techniques to facilitate​ the search and exploration​‌ of machine learning (ML)​​ datasets).
• Serena Villata has​​​‌ been Deputy Scientific Director​ of the Interdisciplinary Institute​‌ of Artificial Intelligence (3IA​​ Côte d'Azur) till October​​​‌ 2025. The 3IA Côte​ d'Azur Institute was one​‌ of the four Interdisciplinary​​ Institutes of Artificial Intelligence​​​‌ labeled in April 2019​ by the Ministry of​‌ Higher Education, Research and​​ Innovation, and it is​​​‌ now one of the​ nine French National Centers​‌ of Excellence in AI​​ labeled "IA Cluster" (since​​​‌ May 2024). Since November​ 2025, Serena Villata took​‌ the responsibility of Scientific​​ Director of the Institute.​​​‌
• Federica Granese has been​ the organizer and coordinator​‌ of MARIANNE Journal Club​​ (Exploring LLMs from the​​​‌ ground up, covering seminal​ papers, key advancements, and​‌ their limitations)1.​​

9.2.1 Teaching

• Elena Cabrio​‌ delivered the following teaching​​ hours: Natural Language Processing​​​‌, Master 1 Computer​ Science. 30 hours (eq.​‌ TD); Natural Language Processing​​, Master 2 MIAGE​​​‌ - IA2: Applied Artificial​ Intelligence. 8 hours; Introduction​‌ to Computational Linguistics,​​ Master 1 EUR CREATES,​​​‌ Parcours Linguistique, traitements informatiques​ du texte et processus​‌ cognitifs. 30 hours; Introduction​​ to Python, Master​​​‌ 1 EUR CREATES, Parcours​ Linguistique, traitements informatiques du​‌ texte et processus cognitifs.​​ 15 hours; Natural Language​​​‌ Processing, DUT IA​ et santé. 2 hours;​‌ Natural Language Processing,​​ Lic. 3 Sciences et​​​‌ Technologies: Intelligence Artificielle. 11​ hours; Introduction to AI​‌. Licence 2. IUT.​​ 22 hours; Introduction to​​​‌ DataBases. Licence 1.​ IUT. 90 hours; Web​‌ interfaces. Licence 1.​​ IUT. 23.5 hours.
• Serena​​​‌ Villata delivered the following​ teaching hours: Master II​‌ Droit de la Création​​ et du Numérique -​​​‌ Sorbonne University: Approche de​ l'Elaboration et du Fonctionnement​‌ des Logiciels, 15 hours​​ (CM), 30 students; Master​​​‌ 2 MIAGE IA -​ University Côte d'Azur: I.A.​‌ et Langage : Traitement​​ automatique du langage naturel,​​​‌ 28 hours (CM+TP), 30​ students; Master 2 Algorithmic​‌ law and data governance​​ - University Côte d'Azur:​​​‌ Introduction to AI, 12​ hours (CM), 20 students;​‌ Master NeuroMod - University​​ Côte d'Azur: Text Analysis,​​​‌ Deep Learning and Statistics,​ 15 hours (CM), 10​‌ students.
• Anais Ollagnier delivered​​ the following teaching hours:​​​‌ EUR DS4H, Enseignement mutualisé​ (S2 et S1), M1​‌ Droit des affaires, UE​​ Artificial Intelligence: Introduction to​​​‌ Deep Learning (KMUIAIUI) (CM​ : 12,5 h ;​‌ TD : 16 h)​​ x2; EFELIA, M1 Anthropologie​​​‌ des arts vivants (S1),​ UE Artificial Intelligence and​‌ Societal Transformation (BMUIST1), CM​​ : 12 h; EUR​​ DS4H – Enseignement mutualisé,​​​‌ M1 Informatique – parcours‌ informatique (S2), UE Individual‌​‌ Project S2 (EN) (KMUMP22U),​​ TD : 2 h;​​​‌ Portail Sciences et Technologies‌ – Enseignement mutualisé, PO1‌​‌ Sciences, ingénierie, technologie, environnement​​ (S2), UE INFO :​​​‌ Système 1 (SPUF201), TP‌ : 60 h; EIT‌​‌ Digital Master School /​​ EMAI4EU, Conception et dispensation​​​‌ de cours en Traitement‌ Automatique des Langues. Cours‌​‌ : Introduction to Natural​​ Language Processing (enseignement en​​​‌ ligne, public international, niveau‌ master).
• Victor David delivered‌​‌ the following teaching hours:​​ University Tutor for Internships​​​‌ and Apprenticeships, Université Côte‌ d’Azur & Polytech Nice‌​‌ Sophia; Fundamentals of AI,​​ Practical Sessions (18 hours),​​​‌ DUT2, IUT Nice Côte‌ d’Azur (x 3 groups).‌​‌
• Sofiane Elguendouze delivered the​​ following teaching hours: Advanced​​​‌ NLP course - M2‌ Computer Science (27h): NLP‌​‌ basics, deep neural networks​​ & backpropagation, recurrent neural​​​‌ networks, sequence to sequence‌ models and attention, transformers,‌​‌ LLMs and prompt engineering;​​ Introduction to AI -​​​‌ M1 applied foreign languages‌ (15h): definition, demystification, introduction‌​‌ to machine learning, introduction​​ to rule-based/statistical and neural​​​‌ NLP, introduction to language‌ modeling; Applied AI -‌​‌ M2 applied foreign languages​​ (15h): reminder about AI​​​‌ and machine learning, reminder‌ about NLP, introduction to‌​‌ language models and prompting;​​ Introduction to AI -​​​‌ M2 humanities and social‌ sciences (15h): introduction to‌​‌ AI, introduction to machine​​ learning, introduction to deep​​​‌ learning, introduction to NLP,‌ language modeling and language‌​‌ models, ethical and societal​​ limits and challenges of​​​‌ AI methods; Introduction to‌ AI - M2 archaeology‌​‌ (12h): introduction to deep​​ learning and computer vision,​​​‌ applied computer vision techniques‌ to archaeological learning problems.‌​‌
• Greta Damo : In​​ 2025, as part of​​​‌ her teaching duty, Greta‌ helped evaluating the group‌​‌ project presentations for the​​ course I.A. et langage​​​‌ at the M2 Intelligence‌ Artificielle Appliquée program (Semester‌​‌ 2 of the 2024/25​​ academic year), which focused​​​‌ on natural language processing‌ topics and techniques, dedicating‌​‌ 5 hours to this​​ activity. Additionally, she taught​​​‌ both theoretical and practical‌ sessions for the Data‌​‌ Analysis and Visualization course​​ at CentraleDigitalLab, an AI​​​‌ postgraduate program at Centrale‌ Méditerranée, during the first‌​‌ semester of the 2025/26​​ academic year, totaling 22​​​‌ hours.
• Ekaterina Sviridova :‌ Ekaterina delivered 12 hours‌​‌ of TD (Travaux dirigés)​​ for Master 2 students​​​‌ in Traitement automatique des‌ textes (EUR CREATES), in‌​‌ the course HMEDAC3 –​​ ECUE Annotation de corpus.​​​‌ The course introduced the‌ role of annotated data‌​‌ in NLP, presented different​​ types of annotation tasks​​​‌ and tools, and included‌ hands-on annotation sessions. It‌​‌ also covered post-processing and​​ preprocessing of annotated data​​​‌ for language models, followed‌ by a practical project‌​‌ in which students implemented​​ the simple pipeline from​​​‌ raw data annotation to‌ model training to address‌​‌ a specific application task.​​
• Deborah Dore delivered the​​​‌ following teaching hours: Data‌ Analysis and Visualization (22H),‌​‌ Centrale Méditerranée, Theory and​​ laboratory sessions for students​​​‌ from different backgrounds. The‌ course focused on how‌​‌ data visualization can provide​​ new insights and support​​​‌ data driven analysis; Traitement‌ du Langage Naturel (24H),‌​‌ Université Côte d’Azur, MIAGE.​​​‌ Laboratory courses focused on​ exploring the field of​‌ natural language processing, from​​ text preprocessing to the​​​‌ creation of AI models;​ Traitement du Langage Naturel​‌ (14H), Université Côte d’Azur,​​ L3 Informatique. Laboratory courses​​​‌ focused on exploring the​ field of natural language​‌ processing, from text preprocessing​​ to the creation of​​​‌ AI models.

9.2.2 Supervision​

• Elena Cabrio and Anais​‌ Ollagnier co-supervised the doctoral​​ thesis of Thi Thao​​​‌ Ha , “Toward Culturally​ and Contextually Grounded Language​‌ Models”.
• Elena Cabrio and​​ Serena Villata co-supervised the​​​‌ PhD thesis of Greta​ Damo (counter-narratives against hate​‌ speech), Deborah Dore (political​​ argumentation evolving in time),​​​‌ Ekaterina Sviridova (implicitness in​ argumentation), Cyprien Michel-Deletie (bridging​‌ natural language argumentation with​​ formal reasoning), Benjamin Ocampo​​​‌ (implicit and subtle hate​ speech detection), Xiaoou Wang​‌ (argumentation-based fact checking).
• Victor​​ David supervised the two​​​‌ Master’s level internships (5​ and 6 months) of​‌ Loup Doinel and Nino​​ Pireaud . Victor also​​​‌ supervised the activity of​ the research engineer Theo​‌ Alkibiades Collias in the​​ context of the AGGREY​​​‌ ANR project.
• Anais Ollagnier​ supervised the scientific activity​‌ of post-doctoral research Yingxue​​ Fu on “Improving argument​​​‌ mining through contextual information​ synthesis”.
• Anais Ollagnier also​‌ supervised the two Master's​​ level internships of Hajar​​​‌ Bakarou and Mohamed Sinane​ El Messoussi .
• Serena​‌ Villata supervised Sofiane Elguedouze​​ as post-doc on the​​​‌ CIGAIA and then the​ ORBIS project, on argument​‌ mining for online deliberation​​ and dispute resolution, Mariana​​​‌ Chavez as engineer on​ the AI4MEDIA and then​‌ CIGAIA project, working on​​ fallacy identification in online​​​‌ social media.

9.2.3 Juries​

Elena Cabrio took part​‌ to the following juries:​​

• President of the jury​​​‌ for a tenure track​ position at Pompeu Fabra​‌ Unversity, Spain.
• Jury member​​ for a tenure track​​​‌ position at Fondazione Bruno​ Kessler, Italy.
• Reviewer of​‌ the PhD thesis of​​ Arij Riabi, Sorbonne Université.​​​‌
• Reviewer of the PhD​ thesis of Amir Reza​‌ Jafari, Institut-Mines Télécom, Télécom​​ SudParis.
• Reviewer of the​​​‌ PhD thesis of Andrea​ Zaninello, Bolzen University, Italy.​‌
• Examiner of the PhD​​ thesis of Virgile Rennard,​​​‌ LIX, École Polytechnique.

Serena​ Villata took part to​‌ the following juries:

• Marta​​ Marchiori Manerba, University of​​​‌ Pisa. Title of the​ thesis: "Fairness Auditing, Explanation​‌ and Debiasing in Linguistic​​ Data and Language Models".​​​‌ Role: reviewer. PhD defense:​ 2025.
• Mustapha Bounoua, EURECOM.​‌ Title of the thesis:​​ "Harnessing Multimodality: Diffusion based​​​‌ Generative Modeling and Information​ Estimation". Role: reviewer. PhD​‌ defense: 2025.
• Youri Peskine,​​ EURECOM. Title of the​​​‌ thesis: "Using Knowledge Graph​ To Detect And Explain​‌ Misinformation Spread On The​​ Web". Role: reviewer. PhD​​​‌ defense: 2025.

9.3.1 Productions (articles, videos,​ podcasts, serious games, ...)​‌

Serena Villata has been​​ interviewed by the newspaper​​​‌ Epsiloon, in the article​ "La fin d’Internet" by​‌ Pierre-Yves Bocquet, March 26,​​ 2025. Link to the​​ article.

9.3.2 Participation​​​‌ in Live events

• Elena‌ Cabrio took part in‌​‌ the following live events:​​
  • Panelist in the roundtable​​​‌ “The metamorphosis of democracy”‌ at the World AI‌​‌ Cannes Festival (WAICF), 2025.​​
  • Festival des Sciences de​​​‌ Nice, 2025.
• Federica Granese‌ took part to the‌​‌ live event "Attractivité de​​ la France dans le​​​‌ domaine de la recherche‌ en IA" (3 minutes‌​‌ interview, link to the​​ online event).
• Serena​​​‌ Villata participated to the‌ following popularization events:
  • Round‌​‌ table at the Conférence​​ Intelligence Artificielle de l'Association​​​‌ OBJECTIF SCIENCE: "L’humanité à‌ l’heure de l’IA :‌​‌ Entre Maîtrise & Dépendance​​ (social, démographie, santé, éthique,​​​‌ démocratie, liberté, informations et‌ communications...)", Isle sur la‌​‌ Sorgue, April 4th, 2025.​​
  • Round table at the​​​‌ Printemps des technologies 2025,‌ March 2025, Centre de‌​‌ Congres Ville Saint Raphael,​​ "Faut-il tout accepter de​​​‌ la techno ?".

International​ journals

• 9 articleG.​‌Greta Damo, N.​​ B.Nicolás Benjamín Ocampo​​​‌, E.Elena Cabrio​ and S.Serena Villata​‌. “Detectors Lead, LLMs​​ Follow”: Integrating LLMs and​​​‌ traditional models on implicit​ hate speech detection to​‌ generate faithful and plausible​​ explanations.Data and​​​‌ Knowledge Engineering162November​ 2025HALDOIback​‌ to text
• 10 article​​D.Daniele Gorla,​​​‌ L.Louis Jalouzot,​ F.Federica Granese,​‌ C.Catuscia Palamidessi and​​ P.Pablo Piantanida.​​​‌ On Estimating the Strength​ of Differentially Private Mechanisms​‌ in a Black-Box Setting​​.IEEE Transactions on​​​‌ Dependable and Secure Computing​225September 2025​‌, 5494-5507HALDOI​​back to text
• 11​​​‌ articleE.Ekaterina Sviridova​, E.Elena Cabrio​‌ and S.Serena Villata​​. Mining implicit arguments​​​‌ for reasoning: A survey​.Argument and Computation​‌June 2025HALDOI​​back to text
• 12​​​‌ articleX.Xiaoou Wang​, E.Elena Cabrio​‌ and S.Serena Villata​​. When automated fact-checking​​​‌ meets argumentation: unveiling fake​ news through argumentative evidence​‌.Argument and Computation​​16April 2025HAL​​​‌DOIback to text​

International peer-reviewed conferences

• 13​‌ inproceedingsJ.Jonathan Ben-Naim​​, V.Victor David​​​‌ and A.Anthony Hunter​. An Axiomatic Study​‌ of a Modular Evaluation​​ of Enthymeme Decodingin Weighted​​​‌ Structured Argumentation.Proceedings​ of the 22nd International​‌ Conference on Principles of​​ Knowledge Representation and Reasoning​​​‌22nd International Conference on​ Principles of Knowledge Representation​‌ and ReasoningMelbourne, Australia​​2025HALback to​​​‌ text
• 14 inproceedingsS.​Stefano Bistarelli, V.​‌Victor David, P.​​Pierre Monnin, F.​​​‌Francesco Santini and C.​Carlo Taticchi. Fast​‌ Computing of Dung Semantics​​ in Acyclic Probabilistic Argumentation​​​‌ Frameworks.The 39th​ Annual AAAI Conference on​‌ Artificial Intelligence (AAAI 2025)​​Thirty-Ninth AAAI Conference on​​​‌ Artificial Intelligence, AAAI 2025​Philadelphia, United StatesFebruary​‌ 2025HALback to​​ text
• 15 inproceedingsM.​​​‌Mariana Chaves, E.​Elena Cabrio and S.​‌Serena Villata. FALCON:​​ A multi-label graph-based dataset​​​‌ for fallacy classification in​ the COVID-19 infodemic.​‌SAC 2025 - ACM/SIGAPP​​ Symposium on Applied Computing​​Catania, Italy2025HAL​​​‌DOIback to text‌
• 16 inproceedingsG.Greta‌​‌ Damo, E.Elena​​ Cabrio and S.Serena​​​‌ Villata. Effectiveness of‌ Counter-Speech against Abusive Content:‌​‌ A Multidimensional Annotation and​​ Classification Study.IEEE​​​‌ XploreWI-IAT 2025 -‌ 24th IEEE/WIC International Conference‌​‌ on Web Intelligence and​​ Intelligent Agent TechnologyLondon,​​​‌ United KingdomNovember 2025‌HALback to text‌​‌
• 17 inproceedingsV.Victor​​ David, J.Jérôme​​​‌ Delobelle and J.-G.Jean-Guy‌ Mailly. Similarity Measures‌​‌ for First-Order Logical Arguments​​.CEUR workshopNMR​​​‌ 2025 - 23rd International‌ Workshop on Nonmonotonic Reasoning‌​‌CEUR-4071Melbourne, AustraliaNovember​​ 2025HALback to​​​‌ text
• 18 inproceedingsV.‌Victor David and A.‌​‌Anthony Hunter. A​​ Logic-based Framework for Decoding​​​‌ Enthymemes in Argument Maps‌ Involving Implicitness in Premises‌​‌ and Claims.IJCAI​​ 2025 - Thirty-Fourth International​​​‌ Joint Conference on Artificial‌ IntelligenceMontreal, CanadaInternational‌​‌ Joint Conferences on Artificial​​ Intelligence OrganizationAugust 2025​​​‌, 4445-4453HALDOI‌back to text
• 19‌​‌ inproceedingsD.Deborah Dore​​, S.Stefano Faralli​​​‌ and S.Serena Villata‌. Leveraging Graph Structural‌​‌ Knowledge to Improve Argument​​ Relation Prediction in Political​​​‌ Debates.12th Argument‌ Mining WorkshopProceedings of‌​‌ the 12th Argument Mining​​ WorkshopVienne, AustriaAssociation​​​‌ for Computational LinguisticsJuly‌ 2025, 74-86HAL‌​‌DOIback to text​​
• 20 inproceedingsS.Sofiane​​​‌ Elguendouze, L.Lucas‌ Anastasiou, E.Erwan‌​‌ Hain, E.Elena​​ Cabrio, A.Anna​​​‌ de Liddo and S.‌Serena Villata. AM4DSP:‌​‌ Argumentation Mining in Structured​​ Decentralized Discussion Platforms for​​​‌ Deliberative Democracy.Proceedings‌ of the 2025 Conference‌​‌ on Empirical Methods in​​ Natural Language Processing: System​​​‌ DemonstrationsSuzhou, ChinaAssociation‌ for Computational LinguisticsNovember‌​‌ 2025, 796-805HAL​​DOIback to text​​​‌
• 21 inproceedingsY.Yingxue‌ Fu, M.-J.Mark-Jan‌​‌ Nederhof and A.Anaïs​​ Ollagnier. A Topicality-Driven​​​‌ QUD Model for Discourse‌ Processing.ACL Anthology‌​‌SIGDIAL 2025 - The​​ 26th Annual Meeting of​​​‌ the Special Interest Group‌ on Discourse and Dialogue‌​‌2025.sigdial-1.0Proceedings of the​​ 26th Annual Meeting of​​​‌ the Special Interest Group‌ on Discourse and Dialogue‌​‌Avignon, FranceAugust 2025​​, 214–230HALback​​​‌ to text
• 22 inproceedings‌P.Pierpaolo Goffredo,‌​‌ D.Deborah Dore,​​ E.Elena Cabrio and​​​‌ S.Serena Villata.‌ DISPUTool 3.0: Fallacy Detection‌​‌ and Repairing in Argumentative​​ Political Debates.ACL​​​‌ Anthology63rd Annual Meeting‌ of the Association for‌​‌ Computational Linguistics (ACL 2025)​​3Proceedings of the​​​‌ 63rd Annual Meeting of‌ the Association for Computational‌​‌ Linguistics (Volume 3: System​​ Demonstrations)Vienne, AustriaAssociation​​​‌ for Computational LinguisticsJuly‌ 2025, 472-480HAL‌​‌DOIback to text​​
• 23 inproceedingsP.Pierpaolo​​​‌ Goffredo, D.Deborah‌ Dore, E.Elena‌​‌ Cabrio and S.Serena​​ Villata. Repairing Fallacious​​​‌ Argumentation in Political Debates‌.Argumentation in the‌​‌ Digital Society: Proceedings of​​ the 5th European Conference​​​‌ on ArgumentationWarsaw, Poland‌September 2025HALback‌​‌ to text
• 24 inproceedings​​F.Federica Granese,​​​‌ B.Benjamin Navet,‌ S.Serena Villata and‌​‌ C.Charles Bouveyron.​​​‌ Merging Embedded Topics with​ Optimal Transport for Online​‌ Topic Modeling on Data​​ Streams.ECML PKDD​​​‌ 2025 - Joint European​ Conference on Machine Learning​‌ and Knowledge Discovery in​​ Databases16019Lecture Notes​​​‌ in Computer SciencePorto,​ PortugalSpringer Nature Switzerland​‌September 2025, 290-307​​HALDOIback to​​​‌ textback to text​
• 25 inproceedingsB.Benjamin​‌ Molinet, E.Elena​​ Cabrio and S.Serena​​​‌ Villata. Assessing Argument-based​ Natural Language Explanations in​‌ Medical Text.The​​ 40th ACM/SIGAPP Symposium On​​​‌ Applied Computing (SAC 2025)​Catania, Sicily, ItalyMarch​‌ 2025HAL
• 26 inproceedings​​N. B.Nicolás Benjamín​​​‌ Ocampo, E.Elena​ Cabrio and S.Serena​‌ Villata. From Hidden​​ to Harmful: Connecting Implicit​​​‌ and Explicit Hate Through​ Implied Statements.WI-IAT​‌ 2025 - 24th IEEE/WIC​​ International Conference on Web​​​‌ Intelligence and Intelligent Agent​ TechnologyLondon, United Kingdom​‌November 2025HAL
• 27​​ inproceedingsX.Xiaoou Wang​​​‌, E.Elena Cabrio​ and S.Serena Villata​‌. Leveraging Argumentation Schemes​​ in Justification Generation for​​​‌ Automated Fact-checking.The​ 24th IEEE/WIC International Conference​‌ on Web Intelligence and​​ Intelligent Agent Technology (WI-IAT​​​‌ 2025)Londres, United Kingdom​November 2025HALback​‌ to text
• 28 inproceedings​​X.Xiaoou Wang,​​​‌ E.Elena Cabrio and​ S.Serena Villata.​‌ SAFE: Structured Argumentation for​​ Fact-checking with Explanations.​​​‌Proceedings of the Thirty-Fourth​ International Joint Conference on​‌ Artificial Intelligence (IJCAI-25)34th​​ International Joint Conference on​​​‌ Artificial Intelligence (IJCAI-25)Montreal,​ CanadaInternational Joint Conferences​‌ on Artificial IntelligenceAugust​​ 2025, 11114-11118HAL​​​‌DOIback to text​

Conferences without proceedings

• 29​‌ inproceedingsB.Blanca Calvo​​ Figueras, J.Jaione​​​‌ Bengoetxea, M.Maite​ Heredia, E.Ekaterina​‌ Sviridova, E.Elena​​ Cabrio, S.Serena​​​‌ Villata and R.Rodrigo​ Agerri. Overview of​‌ the Critical Questions Generation​​ Shared Task.Proceedings​​​‌ of the 12th Argument​ mining WorkshopVienna, Austria​‌Association for Computational Linguistics​​July 2025, 243-257​​​‌HALDOIback to​ text
• 30 inproceedingsF.​‌Federica Granese, S.​​Serena Villata and C.​​​‌Charles Bouveyron. Stick-Breaking​ Embedded Topic Model with​‌ Continuous Optimal Transport for​​ Online Analysis of Document​​​‌ Streams.29th International​ Conference on Artificial Intelligence​‌ and Statistics (AISTATS)Tangier,,​​ Morocco2026HALback​​​‌ to text

Doctoral dissertations​ and habilitation theses

• 31​‌ thesisN. B.Nicolás​​ Benjamín Ocampo. Unmasking​​​‌ implicit and subtle hate​ speech : NLP approaches​‌ for detecting and countering​​ online harm.Université​​​‌ Côte d'AzurJune 2025​HAL

Reports & preprints​‌

• 32 miscH.Hajar​​ Bakarou, M.Mohamed​​​‌ Sinane El Messoussi and​ A.Anaïs Ollagnier.​‌ Addressing Antisocial Behavior in​​ Multi-Party Dialogs Through Multimodal​​​‌ Representation Learning.October​ 2025HALDOIback​‌ to text
• 33 misc​​G.Greta Damo,​​​‌ E.Elena Cabrio and​ S.Serena Villata.​‌ Beating Harmful Stereotypes Through​​ Facts: RAG-based Counter-speech Generation​​​‌.October 2025HAL​back to text
• 34​‌ miscD.Deborah Dore​​, E.Elena Cabrio​​​‌ and S.Serena Villata​. RooseBERT: A New​‌ Deal For Political Language​​ Modelling.October 2025​​HALback to text​​​‌
• 35 miscY.Yingxue‌ Fu and A.Anaïs‌​‌ Ollagnier. Contextualizing Toxicity:​​ An Annotation Framework for​​​‌ Unveiling Pragmatics in Conversations‌ of Online Discussion Forums‌​‌.December 2025HAL​​back to text
• 36​​​‌ miscA.Anaïs Ollagnier‌. Before the Outrage:‌​‌ Challenges and Advances in​​ Predicting Online Antisocial Behavior​​​‌.July 2025HAL‌back to text
• 37‌​‌ miscA.Anaïs Ollagnier​​ and A.Aline Menin​​​‌. DataLens: Enhancing Dataset‌ Discovery via Network Topologies‌​‌.July 2025HAL​​