A Generic Anaphora Resolution Engine for Indian Languages

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1824–1833, Dublin, Ireland, August 23-29 2014. A Generic Anaphora Resolution Engine for Indian Languages Sobha Lalitha Devi AU-KBC Research Centre MIT Campus of Anna University, Chennai, India sobha@au-kbc.org Vijay Sundar Ram AU-KBC Research Centre MIT Campus of Anna University, Chennai, India sundar@au-kbc.org Pattabhi RK Rao AU-KBC Research Centre MIT Campus of Anna University, Chennai, India pattabhi@au-kbc.org Abstract In this paper, we present a generic anaphora engine for Indian languages, which are mostly resource- poor languages. We have analysed the similarities and variations between pronouns and their agreement with antecedents in Indian languages. The generic algorithm developed uses the morphological richness of Indian languages. The machine learning approach uses the features which can handle major Indian languages. We have tested the system with Indo-Aryan and Dravidian languages namely Bengali, Hindi and Tamil. The results are encouraging. 1 Introduction Natural language has different types of anaphoric expressions and these expressions bring elegance and make the natural language text interesting to read. Anaphoric expression in a discourse refers to another item in a discourse. The task of resolving anaphors with its referent, antecedent is called as anaphora resolution. Anaphora resolution is required in most of the NLP applications to achieve re- quired performance. The importance of anaphora resolution in various tasks is demonstrated by re- searchers by integrating anaphora resolution with answer extraction system, automatic summarization system, relation extraction system, document similarity identifier etc. Most of the anaphora resolution systems are developed for particular languages. The researchers have analyzed anaphors across languages at various levels such as syntactic, semantic, discourse, structured, and unstructured features. But there are very few attempts for language independent ap- proaches. In this paper, we present a generic anaphora resolution engine for Indian languages. We have come up with a language independent engine, which takes shallow parsed text as input. The mor- phological richness of Indian languages is tapped to come up with a language independent anaphora resolution engine. Early works in anaphora resolution by Hobbs (1978), Carbonell and Brown (1988), Rich and Lu- perFoy (1988) etc. were mentioned as knowledge intensive approach, where syntactic, semantic in- formation, world knowledge and case frames were used. Centering theory, a discourse based approach for anaphora resolution was presented by Grosz (1977), Joshi and Kuhn (1979), Joshi and Weinstein (1981), Strube and Hahn (1999). Salience feature based approaches were presented by Lappin and Leass (1994), Kennedy Boguraev (1996) and Sobha et al., (2000). Indicator based resolution methods were presented by Mitkov (1997, 1998). One of the early works using machine learning technique was Dagan Itai‟s (1990) unsupervised approach based on co-occurrence words. With the use of machine learning techniques researchers work on anaphora resolution and noun phrase anaphora resolution si- multaneously. The other machine learning approaches for anaphora resolution were the following. Aone and Bennett (1995), McCarty and Lahnert (1995), Soon et al., (2001), Ng and Cardia (2002) had used decision tree based classifier. Daelman and Van de …

Content truncated to 500 words