Article Preview
Top1. Introduction
All natural languages contain words bearing multiple meanings and Hindi is not an exception. For example, the Hindi noun ‘हार’ (haar) may mean defeat or necklace. Human beings can identify the correct sense of a word from the context it is used in, but for machines it is very difficult task. Computational identification of most appropriate sense of a word in a given context is called as Word Sense Disambiguation (WSD). It is one of the most challenging tasks in Natural Language Processing (NLP). WSD algorithms can be broadly categorized into dictionary-based and corpus-based. Dictionary-based approaches use machine readable dictionary, thesaurus or computational lexicon for performing disambiguation. WordNet is the most widely used computational lexicon in English WSD. Work on dictionary-based approaches includes (Lesk, 1986; Banerjee & Pederson, 2002; Patwardhan et al., 2003). Corpus-based approaches can be further categorized into: supervised and unsupervised approaches. Supervised approaches use sense tagged corpus and apply machine learning classifier for disambiguation. The work reported in (Gale et al., 1992; Ng & Lee, 1996; Pedersen, 2000; Lee et al., 2004) used semantically tagged corpus for disambiguation. Unsupervised approaches use raw corpus for disambiguation. The work reported in (Yarowsky, 1995; Resnik, 1997) used raw corpus for disambiguation.
For Hindi WSD, little published work is available including (Sinha et al., 2004; Khapra et al., 2008; Khapra et al., 2009; Singh & Siddiqui, 2012; Singh et al., 2013). Sinha et al. (2004) used Hindi WordNet to extract words in extended sense definitions (synset, glosses and example sentences) and semantic relations (hypernyms, glosses of hypernyms, example sentences of hypernyms, hyponyms, glosses of hypernyms, example sentences of hypernyms, meronyms, glosses of meronyms, example sentences of meronyms). Disambiguation was performed using contextual overlap of words in context of target polysemous word and words in extended sense definitions and semantic relations. Khapra et al. (2008) studied domain specific WSD in a trilingual setting of English, Hindi and Marathi. The dominant senses of words in specific domains were used for disambiguation. They evaluated in Tourism and Health domains and achieved an F1-Score of 65% for all the three languages. Khapra et al. (2009) projected WordNet and corpus parameters from a more resource fortunate language to a less resource fortunate language. Their WSD approach was centered on a novel synset based multilingual dictionary and the observation that distribution of senses remains more or less invariant across languages within a domain. WSD in their work took place in a multilingual setting involving Hindi, Marathi, Bengali and Tamil. Singh et al. (2013) adapted Leacock-Chodorow measure of semantic relatedness for Hindi WSD. Evaluation was done on 20 polysemous Hindi nouns and they achieved an overall average accuracy of 60.65% using Leacock-Chodorow measure. Singh and Siddiqui (2012) evaluated a Lesk-Like algorithm for Hindi WSD. They used direct overlap for measuring the similarity. They evaluated the effects of stemming, stop word removal and context window size on Hindi WSD and obtained maximum overall precision after applying stop word removal and stemming. They observed drop in performance in some cases due to stemming mainly because stemming reduces different vibhakti signs in the same root form thus loosing the discrimination between karakas. For example, ‘के’ (ke), ‘की’ (ki), ‘का’ (ka) and ‘को’ (ko) reduces to ‘क’ (k) after stemming. Frequency analysis of vibhaktis in their dataset reveals that there exist interesting variations in the use of vibhaktis among different senses of a word to express relationship between different parts of speech in a sentence. Some vibhakti signs were more frequent in one sense than others. For example, frequency of ‘का’ (ka), ‘की’ (ki), ‘के’ (ke) and ‘लिए’ (liye) in 30 instances of sense 1, 79 instances of sense 2 and 36 instances of sense 3 of Hindi noun ‘उत्तर’ (uttar) was 3, 1, 7, 2 and 9, 30, 8, 2 and 9, 2, 12, 2 respectively. In this paper, we attempt to utilize this variation to identify correct sense of a word.