Unsupervised Domain Adaptation for Word Sense Disambiguation using Stacked Denoising Autoencoder

In this paper, we propose an unsupervised domain adaptation for Word Sense Disambiguation (WSD) using Stacked Denoising Autoencoder (SdA). SdA is an unsupervised learning method of obtaining the abstract feature set of input data using Neural Network. The abstract feature set absorbs the difference of domains, and thus SdA can solve a problem of domain adaptation. However, SdA does not always cope with any problems of domain adaptation. Especially, difficulty of domain adaptation for WSD depends on the combination of a source domain, a target domain and a target word. As a result, any method of domain adaptation for WSD has adverse effect for a part of the problem, Therefore, we defined the similarity between two domains, and judge whether we use SdA or not through this similarity. This approach avoids an adverse effect of SdA. In the experiments, we have used three domains from the Balanced Corpus of Contemporary Written Japanese and 16 target words. In comparison with baseline, our method has got higher average accuracies for all combinations of two domains. Furthermore, we have obtained better results against conventional domain adaptation methods.