Skip to Main content Skip to Navigation

Analyse des sentiments dans les conversations humain-agent Vers un modèle des goûts de l'utilisateur

Abstract : This thesis is related to two research areas : the sentiment analysis and the embodied conversational agents. Embodied conversational agents can be defined as virtual characters with the ability to converse with a human user. In order to increase the communicative skills, the agent needs a social-emotional intelligence. The agent must be able to manage socio-emotional signals, both for generation and detection. Many works have produced models optimizing the production of gestures or facial expressions expressing either emotions or social attitudes. On the detection side, a majority of the work focuses on the analysis of nonverbal and social-affective features (facial expressions, acoustic features). The verbal content is more and more integrated but remains partially exploited. The rare studies integrating a detection module of user’s sentiments in the context of conversations human-agent do not integrate the specificities of this interaction context. This thesis focuses on the user’s verbal expressions of sentiment. It is dedicated to the development of a detection model for face-to-face human-agent conversations. First, our goal is to define, in the broad class of sentiment expressions, those appearing as the most adapted to the agent’s communicative goals. Second, we aim to select a method that offers the possibility of a fine-grained analysis and can be suited to the conversational context. We choose to refer to the model of Martin and White (2005) and to use a method integrating linguistic rules. Our contributions comprise three parts. First, we provide an in-depth analysis of sentiment expressions that focus on three conversational units : the speech turn, the adjacency pair and the topic sequence. This analysis highlights a number of features used for the development of linguistic rules. Then, we propose a system integrating rules and grammars. The system is grounded on an bottom-up analysis starting from the lexical level to reach the phrastic level. Three level are successively analysed : the speech turn, the adjacency pair and the topic sequence. Finally, we propose an evaluation of the rules. We created two annotation platforms for building two annotations sets on two corpora : a small-talk corpus and a negociation corpus. The annotation sets were used for evaluating the performance of the system.
Document type :
Complete list of metadata

Cited literature [153 references]  Display  Hide  Download
Contributor : Caroline Langlet Connect in order to contact the contributor
Submitted on : Monday, February 11, 2019 - 2:59:16 PM
Last modification on : Wednesday, October 14, 2020 - 4:20:45 AM
Long-term archiving on: : Sunday, May 12, 2019 - 2:28:53 PM


Files produced by the author(s)


  • HAL Id : tel-02002580, version 2



Caroline Langlet. Analyse des sentiments dans les conversations humain-agent Vers un modèle des goûts de l'utilisateur. Traitement du texte et du document. Télécom ParisTech, 2018. Français. ⟨tel-02002580v2⟩



Les métriques sont temporairement indisponibles