Word Sense Disambiguation and Human Intuition for Semantic Classification on Homonyms

Word Sense Disambiguation and Human Intuition for Semantic Classification on Homonyms* Dong-Sung Kim1 and Jae-Woong Choe1 1 Department of Linguistics, Korea University, Anam-Dong 5 Ga, Seongbuk-Gu, Seoul, Korea {dsk202, jchoe}@korea.ac.kr Abstract. This paper reports a psycholinguistic research for the human intuition on the sense classification. The goal of this research is to find a computational model that fits best with our experiments on human intuition. In this regard, we compare three different computational models; the Boolean model, the probabilistic model, and the probabilistic inference model. We first measured the values of each models found in the semantically annotated Sejong corpus. Then the experimental result was compared with the values in the initial measurements. Kappa statistics supports that this agreement experiment is homogeneously coincidental. The Pearson correlation coefficient test shows that the Boolean model is strongly correlated with the human intuition. Keywords: Human Intuition, Homonyms, Computational Model, Psycholinguistics, Word Sense Disambiguation, Boolean Model, Probabilistic Model, Probabilistic Inference Model 1 Introduction * This work was supported by the Second Brain Korea 21 Project. Word Sense Disambiguation (WSD) is the basis of the natural language processing and other AI-related problems (Ide and Vernois 1998). Much of WSD mainly appeals to efficient computational techniques or some interesting discord in human annotators’ classifications. In previous studies, human intuitive computational model on the sense classification is explained on the basis of estimation or probabilistic model (Lapata and Lascarides 2003; Lapata and Brew 2004). However, few of WSD work, if any, has addressed the issue of both linguistic and computational mechanism of the human intuition. This paper is a psycholinguistic research on the computational model of human sense classifications. The well-known Distributional hypothesis (Harris 1964) claims that a sense of an ambiguous word is dependent on collocations around it. In other words, an analysis on the collocations elucidates the sense of their keyword(s). This also includes either statistical or logical behaviors of collocations in the context. Computational model is to predict such behaviors of collocating words. We, in this paper, compare three different models, one from each of the three bases, namely, logic, estimations, and probability. The particular models that are chosen are the Boolean model, the probabilistic model, and the probabilistic inference model. Each model also consists of several different versions of WSD methods. Among them, we choose the Boolean search, the Maximum Likelihood Estimation (MLE), and the naïve Bayesian classifier. The reason we choose those kinds of WSD methods is that they have previously been used for WSD technique and sense classification model (Ide and Vernois 1998, Lapata and Lascarides 2003, Lapata and Brew 2004), and thus have been proven to be relevant. We conducted an experiment to observe the human intuition which provides us with clues to choice of the computational model. We investigate the correlation between each model’s prediction and the experimental result. Also, we compare the decision made by each model with the sense classification made by humans. The rest of the paper is composed of three parts: Section 2 is about …

Content truncated to 500 words