Domain Adaptation Using Domain Similarity- and Domain Complexity-Based Instance Selection for Cross-Domain Sentiment Analysis

摘要

We propose an approach to domain adaptation that selects instances from a source domain training set, which are most similar to a target domain. The factor by which the original source domain training set size is reduced is determined automatically by measuring domain similarity between source and target domain as well as their domain complexity variance. Domain similarity is measured as divergence between term unigram distributions. Domain complexity is measured as homogeneity, i.e. self-similarity. We evaluate our approach in a semi-supervised cross-domain document-level polarity classification experiment. Thereby we show, that it yields small but statistically significant improvements over several natural baselines and achieves results competitive to other state-of-the-art domain adaptation schemes.