Exploring Semantic Information in Hindi WordNet for Hindi Dependency Parsing

International Joint Conference on Natural Language Processing, pages 189–197, Nagoya, Japan, 14-18 October 2013. Exploring Semantic Information in Hindi WordNet for Hindi Dependency Parsing Sambhav Jain Naman Jain Aniruddha Tammewar Riyaz Ahmad Bhat Dipti Misra Sharma Language Technologies Research Centre IIIT Hyderabad {sambhav.jain, riyaz.bhat}@research.iiit.ac.in, dipti@iiit.ac.in {naman.jain, uttam.tammewar}@students.iiit.ac.in Abstract In this paper, we present our efforts to- wards incorporating external knowledge from Hindi WordNet to aid dependency parsing. We conduct parsing experiments on Hindi, an Indo-Aryan language, utiliz- ing the information from concept ontolo- gies available in Hindi WordNet to com- plement the morpho-syntactic information already available. The work is driven by the insight that concept ontologies cap- ture a specific real world aspect of lexical items, which is quite distinct and unlikely to be deduced from morpho-syntactic in- formation such as morph, POS-tag and chunk. This complementing information is encoded as an additional feature for data driven parsing and experiments are conducted. We perform experiments over datasets of different sizes. We achieve an improvement of 1.1% (LAS) when train- ing on 1,000 sentences and 0.2% (LAS) on 13,371 sentences over the baseline. The improvements are statistically significant at p<0.01. The higher improvements on 1,000 sentences suggest that the semantic information could address the data spar- sity problem. 1 Introduction Last decade has witnessed several efforts towards developing robust data driven dependency pars- ing techniques (K¨ubler et al., 2009). The ef- forts, in turn, initiated a parallel drive for building dependency annotated treebanks (Tsarfaty et al., 2013), which serve as a data source for training data driven dependency parsers. The annotations are often multi-layered and furnish information on part of speech category of word forms, their mor- phological features, related word groups and the syntactic relations. The availability of such rich resources have considerably improved the parsing performance of syntactic parsers (Collins et al., 1999). However, the error analysis studies carried out on these parsers later revealed that certain syn- tactic relations are difficult to deduce and disam- biguate with the syntactic information available in the annotated treebanks. The need for richer information invoked sev- eral efforts in the direction of annotating higher order linguistic information in treebanks. It was felt that semantics can be leveraged for syntactic disambiguation and thus semantic annotation was performed in syntactic treebanks to complement the morpho-syntactic annotations (Kingsbury et al., 2002; Montemagni et al., 2003). Fujita et al. (2007) and MacKinlay et al. (2012) illustrated that semantic annotation delivers a significant im- provement in parsing, confirming the hypothesis that semantics can assist syntactic analysis. Among Indian languages, notable efforts on us- ing semantic information in dependency parsing are on Hindi. Bharati et al. (2008) illustrated that mere animacy (human, non-human and inanimate) of a nominal significantly improves the accuracy of the parser. Later studies on extending such information with finer semantic distinctions like time, place, abstract reconfirmed the substantial role of semantics in syntactic parsing (Ambati et al., 2009). These studies are carried out on a data- set with hand annotated semantics. Although these studies …

Content truncated to 500 words