Deep Learning Based vs. Markov Chain Based Text Generation for Cross Domain Adaptation for Sentiment Classification

Cross-domain adaptation for sentiment classification is the process of adapting a classifier that uses knowledge from one or multiple source domains and little data from the target domain to function with acceptable accuracy, precision, and recall on the target domain. One of the challenges facing cross-domain adaptation for sentiment classification is the limited availability of labeled samples in the target domain. In this paper, we introduce text generation in the target domain as a solution to provide a set of labeled data in the target domain then compare deep learning based text generators such as LSTM RNN and GRU RNN against Markov chain based text generators. We first use a rule-based classifier that utilizes knowledge from different source domains in labeling the unlabeled samples in the target domain (Kitchen Product reviews) then we have selected high confidence labeled samples for training LSTM RNN, GRU RNN and Markov chain based text generators. We have evaluated the deep learning based and Markov chain based text generators by measuring the fscores and accuracies of the end classifier when trained on the data generated from each of these models when tested on the kitchen benchmark test set.