Unsupervised Learning for Verb Sense Disambiguation Using Both Trigger Words and Parsing Relations

摘要

This paper presents an unsupervised machine learning approach to verb sense disambiguation based on context clustering. The context is represented as a combination of co-occurring trigger words within a window size of the verb and the parsing relations of the verb. The context is retrieved from a large indexed repository that stores a raw corpus of 88,000,000 words and its corresponding parsing results. The clusters are mapped onto the senses defined by the SENEVAL-2 standards. Benchmarking shows that when using trigger words only, this approach produces results reaching state-of-the-art performance in the unsupervised category. When parsing relations are combined with triggers words in integrated training, a modest performance enhancement is observed. Alternative approaches in using the parsing relations for contextual clustering are discussed.