This paper presents a study on the use of word context features for Word Sense Disambiguation (WSD). State-of-the-art WSD systems achieve high accuracy by using resources such as dictionaries, taggers, lexical analyzers or topic modeling packages. However, these resources are either too heavy or don’t have sufficient coverage for large-scale tasks such as information retrieval. The use of local context for WSD is common, but the rationale behind the formulation of features is often based on trial and error. We therefore investigate the notion of relatedness of context words to the target word (the word to be disambiguated), and propose an unsupervised method for finding the optimal weights for context words based on their distance to the target word. The key idea behind the method is that the optimal weights should maximize the similarity of two context models constructed from different context samples of the same word. Our experimental results show that the strength of the relation between two words follows approximately a power law. The resulting context models are used in Naïve Bayes classifiers for word sense disambiguation. Our evaluation on Semeval WSD tasks in both English and Japanese show that our method can achieve state-of-the-art effectiveness even though it does not use external tools like most existing methods. The high efficiency makes it possible to use our method in large scale applications such as information retrieval.
An overview of the SemEval-2 Japanese WSD task is presented. The new characteristics of our task are (1) the task will use the first balanced Japanese sense-tagged corpus, and (2) the task will take into account not only the instances that have a sense in the given set but also the instances that have a sense that cannot be found in the set. It is a lexical sample task, and word senses are defined according to a Japanese dictionary, the Iwanami Kokugo Jiten. This dictionary and a training corpus were distributed to participants. The number of target words was 50, with 22 nouns, 23 verbs, and 5 adjectives. Fifty instances of each target word were provided, consisting of a total of 2,500 instances for the evaluation. Nine systems from four organizations participated in the task.