Neural transition-based model for extraction and sentiment analysis of user opinions

Extraction and analysis of user opinions towards products and services is important task in research and applications of natural language processing methods. We frame this task as structured prediction task where each data instance represented by group of interdependent labels. To solve this task, we describe transition-based model that decomposes it to the prediction of transition sequence which incrementally build final structure. The proposed model uses deep neural networks as feature extractor for classifier that predicts next transition based on previous transitions and the parts of predicted structure. To evaluate quality of the proposed model, we conducted a series of experiments on user reviews texts from two sources: English reviews from Amazon and Russian reviews from AliExpress. The experiments show that our model performs equally or better than an alternative and suffer less accuracy drop from distributional shift.