KUNLP system using Classification Information Model at SENSEVAL-2

KUNLP system using Classification Information Model at SENSEVAL-2 Hee-Cheol Seo, Sang-Zoo Lee, Rae-Chang Rim Dept. of Computer Science and Engineering, Ho Lee Astronest Inc. Korea University 1, 5-ka, Anarn-dong Seongbuk-Gu, Seoul, 136-701, Korea {hcseo,zoo,rirn }@nlp.korea.ac.kr Abstract The classification information model or CIM classi- fies instances by considering the discrimination abil- ity of their features, which was proven to be useful for word sense disambiguation at SENSEVAL-1. But the CIM has a problem of information loss. KUNLP system at SENSEVAL-2 uses a modified version of the CIM for word sense disambiguation. We used three types of features for word sense disambiguation: local, topical, and bigram context. Local and topical context are similar to Chodorow's context and refer to only unigram information. The window of a bigram context is similar to that of a local context but a bigram context refers to only bigram information. We participated in the English lexical sample task and the Korean lexical sample task, where our sys- tems ranked high. 1 Introduction The classification information model(Ho, 1997) is the model that classifies instances by considering the discrimination ability of their features. In the CIM, a feature with high discrimination ability con- tributes to the classification more than one with low discrimination ability. Hence, we can omit the fea- ture selection procedure. · The CIM has a kind of information loss problem due to the assumption that a feature contributes to only one class. We devised a modified version of the CIM where a feature can contribute to all classes. \Vord sense disambiguation task can be treated as a kind of classification process(Ho, 2000). When a classification technique is applied to word sense dis- ambiguation, an instance corresponds to a context containing a polysemous word and its class to the proper sense of the word, and one of its features to a piece of context information. As a classifica- tion problem, word sense disambiguation task can be solved by the CIM. \Ve used three types of features for word sense disambiguation: local, topical, and bigram context. Local and topical context are similar to Chodorow's context(Chodorow, 2000) and consist of only uni- 135-090 3rd floor, Hanarn BD 157-18 Sarnsung-Dong Kangnarn-Gu, Seoul, Korea leeho@astronest.corn gram information. A bigram context has a similar window to a local context but consists of only bigram information. 2 KUNLP system To disambiguate senses, we did two phases: corpus preprocessing and sense disambiguation. Figure 1 shows the flow chart of our system. Corpus ~ Tokenizer ~ POS-Tagger Corpus Preprocessing ~ Phrase Filter ~ Sense Tagger using Modified CIM Sense Disambiguation ~ Sense-Tagged Corpus Figure 1: Flow chart of KUNLP system 2.1 Corpus preprocessing At the corpus preprocessing phase, we tokenized a corpus and then tagged it with parts-of-speech using Brill's Tagger(Brill, 1994). The tokenizer just sepa- rates symbols from a word. For example, a sentence "I'm straight, white, no longer middle class, anti- IRA, have ... " is tokenized to "I 'm stright , white , no longer middle class , anti - IRA , …

Content truncated to 500 words