Detecting Salient Events in Large Corpora by a Combination of NLP and Data Mining Techniques

In this paper, we present a framework and a system that extracts “salient” events relevant to a query from a large collection of documents, and which also enables events to be placed along a timeline. Each event is represented by a sentence extracted from the collection. We have conducted some experiments showing the interest of the method for this issue. Our method is based on a combination of linguistic modeling (concerning temporal adverbial meanings), symbolic natural language processing techniques (using cascades of morpho-lexical transducers) and data mining techniques (namely, sequential pattern mining under constraints). The system was applied to a corpus of newswires in French provided by the Agence France Presse (AFP). Evaluation was performed in partnership with French newswire agency journalists.