Extraction of Hyponymic Relations in French with Knowledge-Pattern-Based Word Sketches

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 574–580 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 574 A French Medical Conversations Corpus Annotated for a Virtual Patient Dialogue System Fr´ejus A. A. Laleye1, Ga¨el de Chalendar1, Antonia Blani´e2, Antoine Brouquet2, Dan Benhamou2 1 CEA, LIST, Laboratoire d’Analyse S´emantique Texte et Image, Gif-sur-Yvette, F-91191 2 Laboratoire de Formation par la Simulation et l’Image en M´edecine et en Sant´e (LabForSIMS), Facult´e de M´edecine Paris-Sud, 94275 Le Kremlin Bicˆetre {frejus.laleye,gael.de-chalendar}@cea.fr, {antonia.blanie,antoine.brouquet,dan.behnamou}@aphp.fr Abstract Data-driven approaches for creating virtual patient dialogue systems require the availability of large data specific to the language, domain and clinical cases studied. Based on the lack of dialogue corpora in French for medical education, we propose an annotated corpus of dialogues including medical consultation interactions between doctor and patient. In this work, we detail the building process of the proposed dialogue corpus, describe the annotation guidelines and also present the statistics of its contents. We then conducted a question categorization task to evaluate the benefits of the proposed corpus that is made publicly available. Keywords: Virtual Patient, Medical Corpus, Question Answering System 1. Introduction Virtual patients allow medical students to perform clinical simulations to acquire experience through the practice of critical procedures without misbehaving with or endanger- ing real patients. For the students to be able to interact nat- urally with their virtual patient, the dialogue system must be quite versatile. It must have the capacity to understand a wide variety of questions and to answer with natural sound- ing sentences. Purely pattern-based systems do not have this capacity. That’s why they must be completed or re- placed by statistical or, more recently, neural-based ones. To train the latter, large corpora of medical interviews are necessary. Several recent works such as (Jin et al., 2017; Maicher et al., 2019; Sp¨anig et al., 2019) have explored the potential advantages of statistical and neural approaches us- ing a large amounts data to enhance the mechanism for in- terpreting questions in dialogue systems. In this paper, we present the corpus of medical conversa- tions in French that we have built in the context of the de- velopment of such a virtual patient dialogue system. We de- scribe the methodology we used, the annotations we made and several data about its content. We also describe one of the subsystems that were built using the corpus, namely a question categorization tool. The remaining of the dialog system is described and evaluated in (Laleye et al., 2020). We make our corpus freely available under a Free/Libre Open Source licence1. 2. Related Work There are different approaches to achieving a dialogue sys- tem for a virtual patient system. Depending on the clini- cal case and the educational objectives, the dialogue man- agement approaches are divided into handcrafted (Laroche et al., 2010) and probabilistic (Celikyilmaz et al., 2018). Instead of defining rules for the dialog strategy by hand, 1Temporary location during review: https://github. com/kleag/labforsims2-corpus. probabilistic dialogue management …

Content truncated to 500 words