Semi-supervised domain adaptation for WSD: Using a word-by-word model selection approach

This paper proposes a word-by-word model selection approach to domain adaptation for Word Sense Disambiguation. By this approach, the model for a target word is automatically selected from a candidate model set, which is comprised of improved self-training models and a supervised model. The improved self-training uses sense priors to prevent its iteration from converging into undesirable states. Experimental results on a domain-specific corpus show that: (1) our improved self-training model is effective for the words which have target domain linked senses; (2) the selected models obtain higher accuracies than each single model and effectively improve the performance compared to the state-of-the-art supervised model.