TSP: Learning Task-Specific Pivots for Unsupervised Domain Adaptation

Abstract

Unsupervised Domain Adaptation (UDA) considers the problem of adapting a classifier trained using labelled training instances from a source domain to a different target domain, without having access to any labelled training instances from the target domain. Projection-based methods, where the source and target domain instances are first projected onto a common feature space on which a classifier can be trained and applied have produced state-of-the-art results for UDA. However, a critical pre-processing step required by these methods is the selection of a set of common features (aka. pivots ), this is typically done using heuristic approaches, applied prior to performing domain adaptation. In contrast to the one of heuristics, we propose a method for learning Task-Specific Pivots (TSPs) in a systematic manner by considering both the labelled and unlabelled data available from both domains. We evaluate TSPs against pivots selected using alternatives in two cross-domain sentiment classification applications. Our experimental results show that the proposed TSPs significantly outperform previously proposed selection strategies in both tasks. Moreover, when applied in a cross-domain sentiment classification task, TSP captures many sentiment-bearing pivots.