Caractérisation de la couverture d'information : une approche computationnelle fondée sur les asymétries

Abstract

Nowadays, accelerated production of information requires people to adopt strategies to select information, to exclude repeated information and even to merge information, to build a complete panorama of a topic. These strategies fit well with the process of coverage of information, which is becoming an everyday task, but also a complex exercise. Natural Language Processing (NLP) techniques attempt to achieve automatically the coverage of information. In this thesis, we address the coverage of information with a computational approach based on asymmetries. We applied our analysis in two different scenarios : In the first scenario, we analyzed the coverage of information in students’ dissertations by verifying the presence of terminology from the official bibliographic references as suggested in the syllabus of the course. We performed this analysis using a hybrid asymmetric coverage coefficient that uses lexical and semantic information. This hybrid characteristic allows us to capture the different forms of lexical surface that a student can use to express the same concept. To determine if the concepts of a book are covered in the content of a dissertation, we implemented a text-alignment strategy. Our approach can detect a dissertation containing low degree of coverage of information among a group of dissertations that have a better coverage. To corroborate the interpretations of our results, we conducted a qualitative evaluation with the course’s teachers. This evaluation revealed that the results of our analyzes coincided with the grades given to the dissertations. Consequently, the coverage of concepts in student dissertations helps to explain the grades that teachers attributed to the dissertations. In the second scenario, we analyzed the coverage of information in narrative journalistic texts. In this type of texts, events, which occur in the world, are told and discussed by journalists. Events become our interest in this case. An event presents a structure, which can find its form in the answers to the questions : who did what ? To whom? Where? And when? In order to capture the most information about an event, we designed an information coverage coefficient based on linear linguistic patterns. These patterns, although simple, try to capture the structure of an event. We also used a strategy of weighting patterns to highlight a particular pattern. We addressed the coverage of information, in this case, with a strategy of novelty detection, which corresponds to information not covered by other sources. In the quantitative evaluation, our asymmetric approach is able to perform as well as the symmetric measures of the state of the art. In addition, our approach offers the advantage of explaining the origin of the new information because of the strategy of weighting of the patterns.

Date	26 Jun 2017
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Sylvie Ratté (Supervisor)