Exploring the Usefulness of Cross-lingual Information Fusion for Refining Real-time News Event Extraction: A Preliminary Study

Proceedings of Recent Advances in Natural Language Processing, pages 210–217, Hissar, Bulgaria, 12-14 September 2011. Exploring the Usefulness of Cross-lingual Information Fusion for Refining Real-time News Event Extraction: A Preliminary Study Jakub Piskorski Institute of Computer Science Polish Academy of Sciences Warsaw, Poland Jakub.Piskorski @ipipan.waw.pl Jenya Belayeva, Martin Atkinson Joint Research Centre of the European Commission Ispra, Italy {Jenya.Belayeva,Martin.Atkinson} @jrc.ec.europa.eu Abstract Nowadays, many influential facts are re- ported multiple times by different sources and in different languages. This paper presents the results of an experiment on deploying cross-lingual information fu- sion techniques for refining the results of a large-scale multilingual news event extrac- tion system. An evaluation on a test cor- pus consisting of 618 event descriptions which refer to 523 real-world events re- vealed that the description of circa 10% of the events extracted by the mono-lingual systems could be refined. In particular, an overall gain of 6,4% and 4,8% in re- call and precision against the best mono- lingual system could be obtained respec- tively. 1 Introduction The goal of event extraction is to automatically identify events in free texts and to derive struc- tured and detailed information about them. In the past, a vast bulk of the research focused on the de- velopment of mono-lingual event extraction sys- tems that operate on single documents without tak- ing any advantage of global evidence, i.e., with- out reusing the knowledge acquired in the process of extracting information from other topically- related documents. The advantages of going beyond the classical single-document extraction and exploiting information redundancy to validate facts have recently been explored by various re- search groups (Downey et al., 2005; Finkel et al., 2005; Ji and Grishman, 2008; Lee et al., 2010; Liao and Grishman, 2010; Mann, 2007; Patward- han et al., 2007; Poibeau et al., 2008; Yangarber and Jokipii, 2005; Yangarber, 2006). Since nowa- days many influential facts are not only reported multiple times by different sources, but also in dif- ferent languages, the importance of the ability to aggregate and fuse information across documents in several languages is becoming paramount (Ji, 2010). Several experiments on cross-lingual infor- mation extraction have been reported (Chen et al., 2009; Sudo et al., 2004; Lee et al., 2010), how- ever, they mainly focused on cross-lingual boot- strapping of ML-based event extraction systems. This paper presents the results of an experiment aiming at exploring the usefulness of cross-lingual information fusion for refining the results of a real-time multilingual news event extraction en- gine that is deployed in a large-scale online news monitoring platform. To be more precise, we ex- plored: (a) what fraction of event descriptions ex- tracted could potentially be merged and refined through cross-lingual information fusion; and, (b) whether gain in precision/recall could be obtained. In principle, there are two ways of approaching cross-lingual information fusion in the context of multilingual news event extraction: (1) translate all news articles into one common language for which a high-performance event extraction system exists (e.g., English), and run that system on the …

Content truncated to 500 words