Automatic summarization of legal documents requires a thorough understanding of their specificities, mainly with respect to the vocabulary used by legal experts. Indeed, the latter rely heavily on their external knowledge when writing summaries, in order to contextualize the main entities of the source document. This leads to reference summaries containing many abstractions, that sota models struggle to generate. In this paper, we propose an entity-driven approach aiming at learning the model to generate factual hallucinations, as close as possible to the abstractions of the reference summaries. We evaluated our approach on two different datasets, with legal documents in English and French. Results show that our approach allows to reduce non-factual hallucinations and maximize both summary coverage and factual hallucinations at entity-level. Moreover, the overall quality of summaries is also improved, showing that guiding summarization with entities is a valuable solution for legal documents summarization.

This article presents our participation to Task 6 of SemEval-2024, named SHROOM (a Shared-task on Hallucinations and Related Observable Overgeneration Mistakes), which aims at detecting hallucinations. We propose two types of approaches for the task: the first one is based on sentence embeddings and cosine similarity metric, and the second one uses LLMs (Large Language Model). We found that LLMs fail to improve the performance achieved by embedding generation models. The latter outperform the baseline provided by the organizers, and our best system achieves 78% accuracy.

Résumer automatiquement des textes juridiques permettrait aux chargés de veille d’éviter une surcharge informationnelle et de gagner du temps sur une activité particulièrement chronophage. Dans cet article, nous présentons un corpus de textes juridiques en français associés à des résumés de référence produits par des experts, et cherchons à établir quels modèles génératifs de résumé sont les plus intéressants sur ces documents possédant de fortes spécificités métier. Nous étudions quatre modèles de l’état de l’art, que nous commençons à évaluer avec des métriques traditionnelles. Afin de comprendre en détail la capacité des modèles à transcrire les spécificités métiers, nous effectuons une analyse plus fine sur les entités d’intérêt. Nous évaluons notamment la couverture des résumés en termes d’entités, mais aussi l’apparition d’informations non présentes dans les documents d’origine, dites hallucinations. Les premiers résultats montrent que le contrôle des hallucinations est crucial dans les domaines de spécialité, particulièrement le juridique.

This paper describes our two deep learning systems that competed at SemEval-2022 Task 1 “CODWOE: Comparing Dictionaries and WOrd Embeddings”. We participated in the subtask for the reverse dictionary which consists in generating vectors from glosses. We use sequential models that integrate several neural networks, starting from Embeddings networks until the use of Dense networks, Bidirectional Long Short-Term Memory (BiLSTM) networks and LSTM networks. All glosses have been preprocessed in order to consider the best representation form of the meanings for all words that appears. We achieved very competitive results in reverse dictionary with a second position in English and French languages when using contextualized embeddings, and the same position for English, French and Spanish languages when using char embeddings.