Selecting Training Data for Unsupervised Domain Adaptation in Word Sense Disambiguation

This paper describes a method of domain adaptation, which involves adapting a classifier developed from source to target data. We automatically select the training data set that is suitable for the target data from the whole source data of multiple domains. This is unsupervised domain adaptation for Japanese word sense disambiguation (WSD). Experiments revealed that the accuracies of WSD improved when we automatically selected the training data set using two criteria, the degree of confidence and the leave-one-out (LOO)-bound score, compared with when the classifier was trained with all the data.