LCC-WSD: System Description for English Coarse Grained All Words Task at SemEval 2007

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 223–226, Prague, June 2007. c⃝2007 Association for Computational Linguistics LCC-WSD: System Description for English Coarse Grained All Words Task at SemEval 2007 Adrian Novischi, Munirathnam Srikanth and Andrew Bennett Language Computer Corp. Richardson, TX � adrian,srikanth,abennet � @languagecomputer.com Abstract This document describes the Word Sense Disam- biguation system used by Language Computer Cor- poration at English Coarse Grained All Word Task at SemEval 2007. The system is based on two su- pervised machine learning algorithms: Maximum Entropy and Support Vector Machines. These algo- rithms were trained on a corpus created from Sem- Cor, Senseval 2 and 3 all words and lexical sample corpora and Open Mind Word Expert 1.0 corpus. We used topical, syntactic and semantic features. Some semantic features were created using WordNet glosses with semantic relations tagged manually and automatically as part of eXtended WordNet project. We also tried to create more training instances from the disambiguated WordNet glosses found in XWN project (XWN, 2003). For words for which we could not build a sense classifier, we used First Sense in WordNet as a back-off strategy in order to have cov- erage of 100%. The precision and recall of the over- all system is 81.446% placing it in the top 5 systems. 1 Introduction The performance of a Word Sense Disambiguation (WSD) system using a finite set of senses depends greatly on the definition of the word senses. Fine grained senses are hard to distinguish while coarse grained senses tend to be more clear. Word Sense Disambiguation is not a final goal, but it is an in- termediary step used in other Natural Processing ap- plications like detection of Semantic Relations, In- formation Retrieval or Machine Translation. Word Sense Disambiguation is not useful if it is not per- formed with high accuracy (Sanderson, 1994). A coarse grained set of sense gives the opportunity to make more precise sense distinction and to make a Word Sense Disambiguation system more useful to other tasks. Our goal at SemEval 2007 was to measure the per- formance of known supervised machine learning al- gorithm using coarse grained senses. The idea of us- ing supervised machine learning for WSD is not new and was used for example in (Ng and Lee, 1996). We made experiments with two supervised methods: Maximum Entropy (ME) and Support Vector Ma- chines (SVM). These supervised algorithms were used with topical, syntactic and semantic features. We trained a classifier for each word using both su- pervised algorithms. New features were added in 3 incremental steps. After an initial set of experi- ments the algorithm performance was enhanced us- ing a greedy feature selection algorithm similar to one in (Mihalcea, 2002). In order to increase the number of training instances, we tried to use the disambiguated WordNet glosses from XWN project (XWN, 2003). Combining other corpora with dis- ambiguated glosses from XWN did not provide any improvement so we used XWN as a fall back strat- egy for 70 …

Content truncated to 500 words