Learning under Covariate Shift for Domain Adaptation for Word Sense Disambiguation

要旨

We show that domain adaptation for word sense disambiguation (WSD) satisfies the as-sumption of covariate shift, and then solve it by learning under covariate shift. Learning under covariate shift has two key points: (1) calculation of the weight of an instance and (2) weighted learning. For the first point, we em-ploy unconstrained least squares importance fitting (uLSIF), which models the probability density ratio of the source domain against a target domain directly. Additionally, we pro-pose weight only to the particular instance and using a linear kernel rather than a Gaussian kernel in uLSIF. For the second point, we em-ploy a support vector machine (SVM) rather than the maximum entropy method (ME) that is commonly employed in weighted learning. Three corpora in the Balanced Corpus of Con-temporary Written Japanese (BCCWJ) and 16 target words were used in our experiment. The experimental results show that the proposed method demonstrates the highest average pre-cision. 1

主題