KCDC: Word Sense Induction by Using Grammatical Dependencies and Sentence Phrase Structure

Proceedings of the 5th International Workshop on Semantic Evaluation, ACL 2010, pages 351–354, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics KCDC: Word Sense Induction by Using Grammatical Dependencies and Sentence Phrase Structure Roman Kern Know-Center Graz, Austria rkern@know-center.at Markus Muhr Know-Center Graz, Austria mmuhr@know-center.at Michael Granitzer Graz University of Technology, Know-Center Graz, Austria mgrani@know-center.at Abstract Word sense induction and discrimination (WSID) identifies the senses of an am- biguous word and assigns instances of this word to one of these senses. We have build a WSID system that exploits syntactic and semantic features based on the results of a natural language parser component. To achieve high robustness and good general- ization capabilities, we designed our sys- tem to work on a restricted, but grammat- ically rich set of features. Based on the results of the evaluations our system pro- vides a promising performance and robust- ness. 1 Introduction The goal of the SemEval-2 word sense induc- tion and discrimination task, see Manandhar et al. (2010), is to identify the senses of ambiguous nouns and verbs in an unsupervised manner and to label unseen instances of these words with one of the induced senses. The most common approach towards this task is to apply clustering or graph partitioning algorithms on a representation of the words that surround an ambiguous target word, see for example Niu et al. (2007) and Pedersen (2007). We followed this approach by employing a cluster- ing algorithm to detect the individual senses, but focused on generating feature sets different to the mainstream approach. Our feature sets utilize the output of a linguistic processing pipeline that cap- tures the syntax and semantics of sentence parts closely related with the target word. 2 System Overview The base of our system is to apply a parser on the sentence in which the target word occurs. Contex- tual information, for example the sentences sur- rounding the target sentence, are currently not exploited by our system. To analyze the sen- tences we applied the Stanford Parser (Version 1.6.2), which is based on lexicalized probabilis- tic context free grammars, see Klein and Man- ning (2003). This open-source parser not only ex- tracts the phrase structure of a given sentence, but also provides a list of so called grammatical rela- tions (typed dependencies), see de Marneffe et al. (2006). These relations reflect the dependencies between the words within the sentence, for exam- ple the relationship between the verb and the sub- ject. See Chen et al. (2009) for an application of grammatical dependencies for word sense disam- biguation. 2.1 Feature Extraction The phrase structure and the grammatical depen- dencies are sources for the feature extraction stage. To illustrate the result of the parser and feature ex- traction stages we use an example sentence, where the target word is the verb “file”: Afterward , I watched as a butt-ton of good , but misguided people filed out of the theater , and immediately lit up a smoke . 2.1.1 Grammatical Dependency Features The …

Content truncated to 500 words