Pattern Learning for Event Extraction using Monolingual Statistical Machine Translation

Proceedings of Recent Advances in Natural Language Processing, pages 371–377, Hissar, Bulgaria, 12-14 September 2011. Pattern Learning for Event Extraction using Monolingual Statistical Machine Translation Marco Turchi, Vanni Zavarella, Hristo Tanev European Commission - Joint Research Centre (JRC), IPSC - GlobeSec Via Fermi 2749, 21020 Ispra (VA) - Italy marco.turchi@jrc.ec.europa.eu, {vanni.zavarella, hristo.tanev}@ext.jrc.ec.europa.eu Abstract Event extraction systems typically take ad- vantage of language and domain-specific knowledge bases, including patterns that are used to identify specific facts in text; techniques to acquire these patterns can be considered one of the most challenging is- sues. In this work, we propose a language- independent and weakly-supervised algo- rithm to automatically discover linear pat- terns from texts. Our approach is based on a phrase-based statistical machine transla- tion system trained on monolingual data. A bootstrapping version of the algorithm is proposed. Our method was tested on patterns with different domain-specific se- mantic roles in three languages: English, Spanish and Russian. Performance shows the feasibility of our approach and its ca- pability of working with texts in various languages. 1 Introduction Multilingual event extraction task consists of re- trieving information about particular facts from text documents in different languages and produc- ing event-description templates, which typically contain slots about event participants, location, time and means. In this work we use an event extraction sys- tem which aims at identifying violent events, man made and natural disasters and humanitar- ian crises, in title and first sentence of news re- ports. An event is represented as a template, whose main slots correspond to event-specific se- mantic roles, such as: event-type, killed-victims, injured-victims, perpetrators, and others. Slot fillers are typically extracted by matching linear patterns in text. For example, killed <PERSON- GROUP> represents a sample pattern for the se- mantic role DEAD-VICTIM. It will match text snippets like killed five people, where five people fills the pattern slot <PERSON-GROUP>1. this paper, we are concerned with surface-level,In one-slot patterns which accept as slot fillers person names/descriptions such as two Italian women. Building a lexicon of linear patterns is a crucial step in the development and customization of an event extraction system, particularly in news texts which are characterized by an open domain and a large vocabulary. Different approaches have been proposed but most of them require a large man- ual effort and linguistic expertise. Moreover, due to lexical and syntactic variability and to Zipf’s law-based word distribution in language, acquired patterns can only partially cover the range of lin- guistic constructions. These are serious obstacles faced by every effort to adapt an event extraction system across domains or languages. In order to address these problems, we put for- ward a novel language-independent and weakly- supervised algorithm to automatically learn linear event extraction patterns from an unannotated cor- pus of texts. The method allows knowledge-poor pattern acquisition without any data annotation. It is based on the noisy-channel model developed for Phrase-Based Statistical Machine Translation (PBSMT). (e.g.For a particular event-specific semantic role seed.DEAD-VICTIM) a pattern is proposed as The most frequent person …

Content truncated to 500 words