Vincent Bouvier

Large Scale Text Mining Approaches for Information Retrieval and Extraction

The issues for Natural Language Processing and Information Retrieval have been studied for long time but the recent availability of very large resources (Web pages, digital documents…) and the development of statistical machine learning methods exploiting annotated texts (manual encoding by crowdsourcing is a new major way) have transformed these fields. This allows not limiting these approaches to highly specialized domains and reducing the cost of their implementation. For this chapter, our aim is to present some popular text-mining statistical approaches for information retrieval and information extraction and to discuss the practical limits of actual systems that introduce challenges for future.

Use of Web Popularity on Entity Centric Document Filtering

Filtering pages about an entity (person, company, music band...) so that only interesting pages are kept is a real challenge. The interest can be qualified using criteria such as recency, novelty. In the last decade, we have seen classification systems trained to detect the interest for a document regarding an entity. For scalability reasons, it is not possible to consider a manual annotation of a training set for each tracked entity. Some approaches strive to build entity independent systems. These approaches obtain the state of the art performances, but we show that they can be improved. Time features differ from one entity to another, therefore no relevant statistics can be estimated out of these observations by a single classifier. Instead of having one model per entity or one model for all entities, we propose an approach that uses one model per cluster of entities based on the entity web popularity. We also introduce different strategies for automatic classification model selection. We test our approach on the Knowledge Base Acceleration (KBA) framework from TREC and we show that our approach brings significant improvements over a non-cluster-based method.

Modèles de langue adaptatifs et métacritères pour le filtrage de documents et le suivi temporel d’entités

Cet article s’interesse a une problematique de filtrage cible de documents. En plus de detecter et de desambiguiser les entites dans un flux de documents, notre approche ambitionne de selectionner seulement les documents qui presentent des informations nouvelles concernant les entites cibles. Nous proposons une nouvelle approche faiblement supervisee a base de combinaison de modeles de langue dynamiques et temporels (time-aware) qui permettent de suivre l’evolution des entites. Nous mettons en place des metacriteres qui permettent la desambiguisation d’entite dans un document, une estimation de la nouveaute et par dela l’interet de conserver ou non les documents selon une approche de classification par forets aleatoires. Nous montrons sur les donnees de la tâche KBA (Knowledge Base Acceleration) de TREC que nos strategies aboutissent a des performances meilleures que l’etat de l’art.