Extraction de résumé personnalisé, basé sur les aspects des critiques de films

Online reviewing has become a key resource to help users decide what to buy and watch or places to go. A number of platforms allow users to express their opinions about various products or services using numerical ratings as well as textual comments. The numerical ratings give a coarse idea of whether the service or product is good or bad, without further explanation behind the rating. On the other hand, textual comments give full details, but reading all comments is time-consuming and tedious for users. Unfortunately, even summarizing the reviews is unlikely to serve all users in a large population as prior studies have showed that users tend to have different interests regarding the aspects of products or services. Aspects, also called features, describe the characteristics of a product that users give opinions about, i.e., “Story” and “Visual effects” for movies. Most previous studies on aspect-based summary generation have focused on extracting information that summarizes the opinions about aspects in general, without considering the personal preferences of the user. In this dissertation, we develop novel methods and algorithms to generate personalized, aspect-based summaries of the reviews of a given movie for a given user. We rely on a large dataset with movie reviews from the IMDB website to address three important problems of personalized aspect-based summarization. The first problem is to extract the “signature” of an aspect, or the set of words that are related to the aspect, from movie reviews. We develop an algorithm that automatically extracts the signature of each aspect. Our evaluation shows that our method is able to extract even unpopular terms that represent an aspect, such as compound terms or abbreviations, as opposed to the methods proposed in the related work, which focus on popular terms alone. We then study the problem of annotating sentences with aspects, and propose a new method that annotates sentences based on a similarity between the aspect signature and the terms in the sentence. Our large-scale study on the entire dataset shows that our method is able to annotate 61% more sentences than the baseline method with a precision of 0.8. The third problem we tackle is the generation of personalized, aspect-based summaries. We propose an optimization algorithm to maximize the coverage of the aspects the user is interested in and the representativeness of sentences in the summary subject to a length and similarity constraints. We propose a novel method for measuring the representativeness of sentences based on the popularity of opinions that describe the aspects. Finally, we perform three user studies to assess the quality of the aspect-based personalized summaries we generate with real users. The results of the user study show that the approach we propose extracts less redundant and more representative summaries compared to the state of art method for generating summaries.Les sites web de critiques en ligne aident les utilisateurs à décider quoi acheter ou quels hotel choisir. Ces plateformes permettent aux utilisateurs d’exprimer leurs opinions à l’aide d’évaluations numériques et de commentaires textuels. Les notes numériques donnent une idée approximative du service. D'autre part, les commentaires textuels donnent des détails complets, ce qui est fastidieux à lire. Dans cette thèse, nous développons de nouvelles méthodes et algorithmes pour générer des résumés personnalisés de critiques de films, basés sur les aspects, pour un utilisateur donné.Le premier problème que nous abordons consiste à extraire un ensemble de mots liés à un aspect des critiques de films. Notre évaluation montre que notre méthode est capable d'extraire même des termes impopulaires qui représentent un aspect, tels que des termes composés ou des abréviations. Nous étudions ensuite le problème de l'annotation des phrases avec des aspects et proposons une nouvelle méthode qui annote les phrases en se basant sur une similitude entre la signature d'aspect et les termes de la phrase. Le troisième problème que nous abordons est la génération de résumés personnalisés, basés sur les aspects. Nous proposons un algorithme d'optimisation pour maximiser la couverture des aspects qui intéressent l'utilisateur et la représentativité des phrases dans le résumé sous réserve de contraintes de longueur et de similarité. Enfin, nous réalisons trois études d’utilisateur qui montrent que l’approche que nous proposons est plus performante que la méthode de pointe en matière de génération de résumés