Potts Model on the Case Fillers for Word Sense Disambiguation

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1382–1386, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Potts Model on the Case Fillers for Word Sense Disambiguation Hiroya Takamura Tokyo Institute of Technology takamura@pi.titech.ac.jp Manabu Okumura Tokyo Institute of Technology oku@pi.titech.ac.jp Abstract We propose a new method for word sense disambiguation for verbs. In our method, sense-dependent selectional preference of verbs is obtained through the probabilistic model on the lexical network. The mean- field approximation is employed to com- pute the state of the lexical network. The outcome of the computation is used as fea- tures for discriminative classifiers. The method is evaluated on the dataset of the Japanese word sense disambiguation. 1 Introduction Polysemous words can be obstacles to many ap- plications of natural language processing such as information retrieval, question answering, and ma- chine translation. The task of distinguishing word senses given a token of a polysemous word and its context is often referred to as word sense dis- ambiguation and has been studied by many re- searchers (Agirre and Edmonds, 2006). Among a number of word sense disambiguation tasks including noun sense disambiguation, adjective sense disambiguation, and named entity disam- biguation, we focus on the verb sense disambigua- tion with the supervised setting, where we are sup- posed to construct a classifier given labeled train- ing instances. It is learned in the previous work that case fillers are often good clues for the verb sense disam- biguation. Consider, for example, the following two sentences: 1. “He drove a car to the next town.” 2. “He drove the dogs away.” The “drive” in Sentence 1 means “operate (a ve- hicle)”, while “drive” in Sentence 2 means “urge (something to move)”. Although there might be long contexts to these instances, looking at the case fillers “car” and “dog” alone would lead to the correct interpretations of the meanings. This kind of preference on nouns as case fillers is called selectional preference, which in this case de- pends on sense. It is sometimes impractical, how- ever, to expect that the training dataset covers all the nouns as case fillers, if case fillers of the target verb are diverse. The purpose of this article is to propose a method for propagating the information on the case fillers in the training dataset to other nouns, so that the sense of polysemous words is correctly disambiguated. In our method, the infor- mation propagation is implemented as estimation of the state of the probabilisitc model on the lexi- cal network. One advantage of our method is that we can overcome the difficulty caused by the noise contained in the lexical network. 2 Related Work Recent work on general word sense disambigua- tion is summarized in the book edited by Agirre and Edmonds (2006). We focus on the work of the verb sense disambiguation. The idea of using case frames for the verb sense disambiguation is not novel, and dates back to 90’s. Fujii et al. (1998) proposed a …

Content truncated to 500 words