Investigation of sentence structure in domain adaptation for sentiment classification

Abstract

A popular use case of computational linguistics is the identification of sentiment in text. Many current methods for sentiment classification focus on word features within sentences of a text. These methods employ different mathematical and computational techniques to achieve increasing accuracies. Additionally, these techniques are being applied to domain adaptation for sentiment classification which allow sentiment classifiers to be even more flexible. This thesis intends to show the relevance of sentence structure in combination with word features for determining sentiment and the benefits to be seen in domain adaptation contexts. By using part-of-speech (POS) representations for sentences in the Amazon product reviews dataset we find that there is useful sentiment information to be gleaned from sentence structures. This information can be subsequently used by classifiers to improve sentiment classification accuracies.