Adapting a State-of-the-art Anaphora Resolution System for Resource-poor Language

International Joint Conference on Natural Language Processing, pages 815–821, Nagoya, Japan, 14-18 October 2013. Adapting a State-of-the-art Anaphora Resolution System for Resource-poor Language Utpal Kumar Sikdar 1, Asif Ekbal 1, Sriparna Saha 1 Olga Uryupina 2, Massimo Poesio 3 1 Department of Computer Science and Engineering, IIT Patna, India, {utpal.sikdar,asif,sriparna}@iitp.ac.in 2 University of Trento, Center for Mind/Brain Sciences, uryupina@unitn.it 3 University of Essex, Language and Computation Group, poesio@essex.ac.uk Abstract In this paper we present our work on adapting a state-of-the-art anaphora res- olution system for a resource poor lan- guage, namely Bengali. Performance of any anaphoric resolver greatly depends on the quality of a high accurate mention de- tector. We develop a number of mod- els for mention detection based on heuris- tics and machine learning. Our exper- iments show that, a language-dependent system can attain reasonably good perfor- mance when re-trained on a new language with a proper subset of features. The system yields the MUC recall, precision and F-measure values of 57.80%, 79.00% and 66.70%, respectively. Our experi- ments with other available scorers show the F-measure values of 59.47%, 49.83%, 31.81% and 70.82% for BCUB, CEAFM, CEAFE and BLANC, respectively. 1 Introduction Anaphora/co-reference resolution is the task of identifying noun phrases that are used to refer to the same entity in a text. More precisely, let us assume that C1 and C2 are occurrences of two noun phrases (NPs) and both have a unique ref- erent in the context in which they occur. Here C2 refers to C1 in the context. C1 is called antecedent and C2 is called anaphor. The noun phrases that may participate in co-reference relation are called mentions/markables. Various practical tasks re- quire language technology; for example, informa- tion extraction and text summarization, can be per- formed more reliably if it is possible to automat- ically find parts of the text containing informa- tion about a given topic. Anaphoric information is also needed to solve several other such kinds of Natural Language Processing (NLP) problems. Most of these works on supervised machine learn- ing co-reference resolution have been developed for English (Soon et al., 2001; Ng and Cardie, 2002; Yang et al., 2003; Luo et al., 2004), due to the availability of large corpora such as ACE (Walker et al., 2006) and OntoNotes (Weischedel et al., 2008). BART, the Beautiful Anaphora Res- olution Toolkit (Versley et al., 2008), (Ponzetto and Strube, 2006), (Poesio and Kabadjov, 2004), is the resultant of the project titled ”Exploiting Lexical and Encyclopedic Resources For Entity Disambiguation” carried out at the Johns Hopkins Summer Workshop 2007. It can handle all the preprocessing tasks to perform automatic coref- erence resolution. A variety of machine learning approaches are used in BART; it mainly uses sev- eral machine learning toolkits, including WEKA, MaxEnt and Support Vector Machine (SVM). Literature shows the significant amount of works in the area of anaphora resolution. But these (Pradhan et al., 2012; Ng, 2010; Poesio et al., 2010) are mainy in non-Indian languages. The works related to anaphora …

Content truncated to 500 words