Automatic construction of domain-specific sentiment lexicon for unsupervised domain adaptation and sentiment classification

Abstract Sentiment analysis has long been suffering from inaccuracies using either machine learning methods that mostly benefit from text features or sentiment lexicon-based methods that are prone to domain-dependent problems. Furthermore, since labeling is a time-consuming and an expensive task, supervised machine learning methods suffer from the drawback of insufficient labeled samples. To tackle the mentioned issues, this paper proposes a novel approach with a hybrid of a neural network and a sentiment lexicon. This combination can simultaneously adapt word polarities to the target domain and leverage the polarity of whole document in order to alleviate the need for large labeled corpora in an unsupervised manner. In this respect, a sentiment lexicon is constructed from the source domain in the preprocessing phase using the labeled data. In the Next phase, having a Multilayer Perceptron (MLP), the weights of the first hidden layer are set to the corresponding polarity of each word from the retrieved sentiment lexicon and the network is trained. Finally, a Domain-Independent Lexicon (DIL) is introduced which contains words (mostly adjectives) with static positive or negative scores independent from a specific domain. After feeding the target domain to the pre-trained model, the total accuracy of the framework is enhanced by estimating the sentiment polarity of each sentence using the summation of the scores of the constitutive domain independent words. The experiments on Amazon multi-domain sentiment dataset illustrate that our approach significantly outperforms several alternative previous approaches of unsupervised domain adaptation.