Résumé

En contexte de veille, l’extraction d’information non supervisée a pour but d’extraire

Abstract

The purpose of unsupervised information extraction is to extract information from text without fixing the type of information. Our work concentrates on the task of extracting and characterizing new relations between given entity types. We first propose in this article a filtering procedure to remove false relation candidates by combining heuristics and machine learning models. Best results achieve a score of 77.1% F-measure. Similar relations are then grouped together semantically using Markov Clustering and All Pairs Similarity Search algo- rithm, which can efficiently identify similar candidates in large scale. Finally, evaluations of clustering results, using both internal and external measures, show that the integration of the filtering step allows to double the recall while keeping the same precision.