Enhancing Mention Detection Using Projection via Aligned Corpora

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 993–1001, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Enhancing Mention Detection using Projection via Aligned Corpora Yassine Benajiba Center for Computational Learning Systems Columbia University, NY ybenajiba@ccls.columbia.edu Imed Zitouni IBM T.J. Watson Research Center Yorktown Heights, NY izitouni@us.ibm.com Abstract The research question treated in this paper is centered on the idea of exploiting rich re- sources of one language to enhance the per- formance of a mention detection system of an- other one. We successfully achieve this goal by projecting information from one language to another via a parallel corpus. We exam- ine the potential improvement using various degrees of linguistic information in a statisti- cal framework and we show that the proposed technique is effective even when the target language model has access to a significantly rich feature set. Experimental results show up to 2.4F improvement in performance when the system has access to information obtained by projecting mentions from a resource-rich- language mention detection system via a par- allel corpus. 1 Introduction The task of identifying and classifying entity textual references in open-domain texts, i.e. the Mention Detection (MD) task, has become one of the most important subtasks of Information Extraction (IE). It might intervene both as one step to structure nat- ural language texts or as a text enrichment prepro- cessing step to help other Natural Language Process- ing (NLP) applications reach higher accuracy. Simi- larly to the Automatic Content Extraction (ACE) 1 nomenclature, we consider that a mention can be either named (e.g., John, Chicago), nominal (e.g., president, activist) or pronominal (e.g., he, she). It has also a specific class which describes the type of the entity it refers to. For instance, in the sentence: 1http://www.itl.nist.gov/iad/mig/tests/ace/2007/doc/ace07- evalplan.v1.3a.pdf Michael Bloomberg, the Mayor of NYC, declared his war on tobacco and sugary drinks in the city. we find the mentions ‘Michael Bloomberg’, ‘Mayor’ and ‘his’ of the same person entity. Their types are named, nominal and pronominal, respectively. ‘NYC’ and ‘city’, on the other hand, are mentions of the same geopolitical (GPE) entity of type named and nominal, respectively. Consequently, MD is a more general and complex task than the well known Named Entity Recognition (NER) task which aims solely at the identification and classification of the named mentions. The difficulty of the MD task is directly related to the nature of the language and the linguistic re- sources available, i.e. it is easier to build accu- rate MD systems for languages with a simple mor- phology and a high amount of linguistic resources. For this reason, we explore the idea of using an MD system, which has been designed and built for a resource-rich language (RRL), to help enhance the performance of an MD system in a target lan- guage (TL). More specifically, the goal of the re- search work we present in this paper is to employ the richness of English, in terms of natural lan- guage resources, to …

Content truncated to 500 words