Gwénolé Lecorvé

Quentin Brabant, Gwenole Lecorve, Lina M. Rojas-Barahona

Questions asked by humans during a conversation often contain contextual dependencies, i.e., explicit or implicit references to previous dialogue turns. These dependencies take the form of coreferences (e.g., via pronoun use) or ellipses, and can make the understanding difficult for automated systems. One way to facilitate the understanding and subsequent treatments of a question is to rewrite it into an out-of-context form, i.e., a form that can be understood without the conversational context. We propose CoQAR, a corpus containing $4.5$K conversations from the Conversational Question-Answering dataset CoQA, for a total of $53$K follow-up question-answer pairs. Each original question was manually annotated with at least 2 at most 3 out-of-context rewritings. CoQAR can be used in the supervised learning of three tasks: question paraphrasing, question rewriting and conversational question answering. In order to assess the quality of CoQAR's rewritings, we conduct several experiments consisting in training and evaluating models for these three tasks. Our results support the idea that question rewriting can be used as a preprocessing step for question answering models, thereby increasing their performances.

Quentin Brabant, Gwenole Lecorve, Lina M. Rojas-Barahona et al.

We present KGConv, a large, conversational corpus of 71k conversations where each question-answer pair is grounded in a Wikidata fact. Conversations contain on average 8.6 questions and for each Wikidata fact, we provide multiple variants (12 on average) of the corresponding question using templates, human annotations, hand-crafted rules and a question rewriting neural model. We provide baselines for the task of Knowledge-Based, Conversational Question Generation. KGConv can further be used for other generation and analysis tasks such as single-turn question generation from Wikidata triples, question rewriting, question answering from conversation or from knowledge graphs and quiz generation.

Rashedur Rahman, Gwénolé Lecorvé, Nicolas Béchet

Children have less text understanding capability than adults. Moreover, this capability differs among the children of different ages. Hence, automatically predicting a recommended age based on texts or sentences would be a great benefit to propose adequate texts to children and to help authors writing in the most appropriate way. This paper presents our recent advances on the age recommendation task. We consider age recommendation as a regression task, and discuss the need for appropriate evaluation metrics, study the use of state-of-the-art machine learning model, namely Transformers, and compare it to different models coming from the literature. Our results are also compared with recommendations made by experts. Further, this paper deals with preliminary explainability of the age prediction model by analyzing various linguistic features. We conduct the experiments on a dataset of 3, 673 French texts (132K sentences, 2.5M words). To recommend age at the text level and sentence level, our best models achieve MAE scores of 0.98 and 1.83 respectively on the test set. Also, compared to the recommendations made by experts, our sentence-level recommendation model gets a similar score to the experts, while the text-level recommendation model outperforms the experts by an MAE score of 1.48.

Hichem Ammar Khodja, Frédéric Béchet, Quentin Brabant et al.

The factuality of large language model (LLMs) tends to decay over time since events posterior to their training are "unknown" to them. One way to keep models up-to-date could be factual update: the task of inserting, replacing, or removing certain simple (atomic) facts within the model. To study this task, we present WikiFactDiff, a dataset that describes the evolution of factual knowledge between two dates as a collection of simple facts divided into three categories: new, obsolete, and static. We describe several update scenarios arising from various combinations of these three types of basic update. The facts are represented by subject-relation-object triples; indeed, WikiFactDiff was constructed by comparing the state of the Wikidata knowledge base at 4 January 2021 and 27 February 2023. Those fact are accompanied by verbalization templates and cloze tests that enable running update algorithms and their evaluation metrics. Contrary to other datasets, such as zsRE and CounterFact, WikiFactDiff constitutes a realistic update setting that involves various update scenarios, including replacements, archival, and new entity insertions. We also present an evaluation of existing update algorithms on WikiFactDiff.

Michelle Elizabeth, Alicja Kasicka, Natalia Krawczyk et al.

The growing number of generative AI-based dialogue systems has made their evaluation a crucial challenge. This paper presents our contribution to this important problem through the Dialogue System Technology Challenge (DSTC-12, Track 1), where we developed models to predict dialogue-level, dimension-specific scores. Given the constraint of using relatively small models (i.e. fewer than 13 billion parameters) our work follows two main strategies: employing Language Models (LMs) as evaluators through prompting, and training encoder-based classification and regression models. Our results show that while LM prompting achieves only modest correlations with human judgments, it still ranks second on the test set, outperformed only by the baseline. The regression and classification models, with significantly fewer parameters, demonstrate high correlation for some dimensions on the validation set. Although their performance decreases on the test set, it is important to note that the test set contains annotations with significantly different score ranges for some of the dimensions with respect to the train and validation sets.

Aline Étienne, Delphine Battistelli, Gwénolé Lecorvé

The objective of this paper is to predict (A) whether a sentence in a written text expresses an emotion, (B) the mode(s) in which it is expressed, (C) whether it is basic or complex, and (D) its emotional category. One of our major contributions, through a dataset and a model, is to integrate the fact that an emotion can be expressed in different modes: from a direct mode, essentially lexicalized, to a more indirect mode, where emotions will only be suggested, a mode that NLP approaches generally don't take into account. Another originality is that the scope is on written texts, as opposed usual work focusing on conversational (often multi-modal) data. In this context, modes of expression are seen as a factor towards the automatic analysis of complexity in texts. Experiments on French texts show acceptable results compared to the human annotators' agreement, and outperforming results compared to using a large language model with in-context learning (i.e. no fine-tuning).

Hichem Ammar Khodja, Frédéric Béchet, Quentin Brabant et al.

This paper explores the robustness of language models (LMs) to variations in the temporal context within factual knowledge. It examines whether LMs can correctly associate a temporal context with a past fact valid over a defined period, by asking them to differentiate correct from incorrect contexts. The LMs' ability to distinguish is analyzed along two dimensions: the distance of the incorrect context from the validity period and the granularity of the context. To this end, a dataset called TimeStress is introduced, enabling the evaluation of 18 diverse LMs. Results reveal that the best LM achieves a perfect distinction for only 11% of the studied facts, with errors, certainly rare, but critical that humans would not make. This work highlights the limitations of current LMs in temporal representation.

Camille Barboule, Viet-Phi Huynh, Adrien Bufort et al.

Despite outstanding processes in many tasks, Large Language Models (LLMs) still lack accuracy when dealing with highly technical domains. Especially, telecommunications (telco) is a particularly challenging domain due the large amount of lexical, semantic and conceptual peculiarities. Yet, this domain holds many valuable use cases, directly linked to industrial needs. Hence, this paper studies how LLMs can be adapted to the telco domain. It reports our effort to (i) collect a massive corpus of domain-specific data (800M tokens, 80K instructions), (ii) perform adaptation using various methodologies, and (iii) benchmark them against larger generalist models in downstream tasks that require extensive knowledge of telecommunications. Our experiments on Llama-2-7b show that domain-adapted models can challenge the large generalist models. They also suggest that adaptation can be restricted to a unique instruction-tuning step, dicarding the need for any fine-tuning on raw texts beforehand.

Ismaël Rousseau, Claire Perroux, Pierre Adam et al.

This paper presents the work carried out by the O_FT team, joint with Orange and Ouest-France, on adapting language models to the defense domain as part of the EvalLLM2025 challenge. This work focused on adapting the \texttt{Mistral-7B-Instruct-v0.3} model using classical techniques of continued pre-training and instruction-tuning. The core of our efforts is based on collecting, generating, and selecting data for these two stages as well as for model evaluation. Experiments show that our adapted models have better domain-specific knowledge and improved domain-specific task processing skills, along with comparable (or even superior) performance on general knowledge and skills. Considering the carbon footprint of our adaptations, this work demonstrates the feasibility of domain adaptation for relatively small models. -- Ce document présente les travaux réalisés par l'équipe O_FT conjointe à Orange et Ouest-France sur l'adaptation de modèles de langue au domaine de la défense dans le cadre du challenge EvalLLM2025. Ces travaux se sont concentrés sur l'adaptation du modèle \texttt{Mistral-7B-Instruct-v0.3} avec des techniques classiques de poursuite du pré-entraînement et d'affinage sur instructions. L'essentiel de nos travaux a porté sur la constitution, génération et sélection de données pour ces deux étapes ainsi que pour l'évaluation des modèles. Les expériences montrent que nos modèles adaptés ont de meilleures de connaissances de fond et une meilleure capacité de traitement de tâches sur le domaine de la défense, ainsi que des performances comparables (voire supérieures) sur des connaissances ou capacités généralistes. Mis au regard des empreintes carbones de nos adaptations, ces travaux démontrent ainsi la viabilité de l'adaptation à un domaine de modèles relativement petits.

Juliette Faille, Quentin Brabant, Gwenole Lecorve et al.

We explore question generation in the context of knowledge-grounded dialogs focusing on explainability and evaluation. Inspired by previous work on planning-based summarisation, we present a model which instead of directly generating a question, sequentially predicts first a fact then a question. We evaluate our approach on 37k test dialogs adapted from the KGConv dataset and we show that, although more demanding in terms of inference, our approach performs on par with a standard model which solely generates a question while allowing for a detailed referenceless evaluation of the model behaviour in terms of relevance, factuality and pronominalisation.

Morgan Veyret, Jean-Baptiste Duchene, Kekeli Afonouvi et al.

Typically available dialogue frameworks have adopted a semantic representation based on dialogue-acts and slot-value pairs. Despite its simplicity, this representation has disadvantages such as the lack of expressivity, scalability and explainability. We present WEBDial: a dialogue framework that relies on a graph formalism by using RDF triples instead of slot-value pairs. We describe its overall architecture and the graph-based semantic representation. We show its applicability from simple to complex applications, by varying the complexity of domains and tasks: from single domain and tasks to multiple domains and complex tasks.