A Hybrid Approach for Anaphora Resolution in Hindi

International Joint Conference on Natural Language Processing, pages 977–981, Nagoya, Japan, 14-18 October 2013. A Hybrid Approach for Anaphora Resolution in Hindi Praveen Dakwale LTRC, IIIT-Hyderabad India dakwale.praveen@gmail.com Vandan Mujadia CSPIT, Charusat Gujarat, India vmujadia@gmail.com Dipti M Sharma LTRC, IIIT-Hyderabad India diptims@gmail.com Abstract In this paper we present a hybrid approach to resolve Entity-pronoun references in Hindi. While most of the existing ap- proaches, syntactic as well as data-driven, use phrase-structure syntax for anaphora resolution, we explore use of dependency structures as a source of syntactic informa- tion. In our approach, dependency struc- tures are used by a rule-based module to resolve simple anaphoric references, while a decision tree classifier is used to resolve more ambiguous instances, using gram- matical and semantic features. Our results show that, use of dependency structures provides syntactic knowledge which helps to resolve some specific types of refer- ences. Semantic information such as ani- macy and Named Entity categories further helps to improve the resolution accuracy. 1 Introduction In various approaches on anaphora resolution syntax has been used as an important feature. Some well-known syntax based approaches in- clude Hobbs algorithm (Hobbs, 1986) and the Centering approach (Brennan et al., 1987). Var- ious rule based and data driven approaches have been proposed which use syntactic information as an important feature. Most of the earlier works have used phrase- structure parse as a source of syntactic informa- tion. However, dependency structures are more suitable representations for relatively free word or- der languages such as Hindi (Bharati et al., 1995; Melˇcuk, 1988) and hence research in many such languages has focused on development of depen- dency based resources resulting in better availabil- ity of dependency data for such languages. In this paper, we explore the possibility of using depen- dency structure for anaphora resolution in Hindi. However, we do not intend either to propose de- pendency as an alternative to phrase structure or to compare the usability of the two frameworks. (Prasad and Strube, 2000) is one of the most important approach for anaphora resolution in Hindi. They applied a discourse salience rank- ing to two pronoun resolution algorithms, the BFP and the S-List algorithm. (Dakwale and Sharma, 2011) reported the best performance for Hindi in Anaphora Resolution tool contest in Indian languages(ICON-2011). They propose a hybrid approach with limited linguistic knowledge such as NER categories and verb similarity. Two earlier approaches explore the use of de- pendency relations for anaphora resolution. For Hindi, (Uppalapu and Sharma, 2009) extends the S-List algorithm by using two different lists in place of a single list. For English, (Bj¨orkelund and Kuhn, 2012) explores the possibility of using dependency relations as a feature for co-reference resolution in a learning based approach. Both the approaches are limited in their exploration of de- pendency for anaphora resolution as they only use dependency relations either as a salience for rank- ing the candidate referents or as an additional fea- ture in a learning based approach. We discuss how resolution of different types of Entity-pronoun …

Content truncated to 500 words