Système de recommandations vidéo fondé sur l’analyse des sous-titres dans le but de soutenir le développement linguistique d’enfants autistes

Abstract

It has been observed that through exposure to videos containing letters, numbers, and shapes, preschool children with autism, who are usually non-verbal or minimally verbal, quickly acquire a self-taught knowledge of the written code and a vocabulary that will facilitate their schooling for several years later. The research program of which this project is part aims to develop a personalized multimodal video recommendation system that supports their linguistic progress and to design a database on language development in these children. Specifically, this project involves developing a recommendation system based on video subtitles to integrate later features extracted from other signals, such as colors, objects, and sounds, to understand the child's interests better and personalize recommendations further. To achieve this, we developed an optimization framework with a thorough evaluation of the stability of the LDA model parameterization. The framework was then used to model a corpus of automatic transcriptions from children's videos available on YouTube. Using this subtitle modeling, we developed a recommendation system and an interface illustrating the evolution of the estimation of interests through the themes describing the video subtitles (semantic map). To validate the quality of the optimized model, recommendation calculations, and the descriptive potential of the semantic map, we designed a web application to conduct an initial evaluation of this system with participants from the academic community. Following the different experiments, the analysis of the results concerning the optimization framework demonstrates the importance of stability analysis to select a near-optimal model instead of blindly relying on metaheuristic-driven procedures. Regarding evaluating the recommender system, the results demonstrate the potential of using LDA as a computational basis for designing a recommender system. The participants were interested in adopting the semantic map in a video search and selection context. Finally, the participants appreciated visualizing the evolution of their interests and their relevance in describing the content of the videos they liked.

Date	17 May 2023
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Sylvie Ratté (Supervisor) & Laurent Mottron (Co-supervisor)