Aspect Sentiment Classification Based On Sequence To Sequence Reinforced Learning

The task of aspect sentiment classification (ASC) is a fundamental task in sentiment analysis. Given an aspect and a sentence, the task classifies the sentiment polarity expressed on the target in the sentence. Previous work usually distinguish the sentiment based on one-way LSTM, which are often complicated and need more training time. In this paper, motivated by the BERT from Google AI Language, we propose a novel two-way encoder-decoder framework that automatically extracts appropriate sentiment information according to sequence to sequence reinforced learning. We use reinforcement learning to explore the space of possible extractive targets, where useful information provided by earlier predicted antecedents could be utilized for making later coreference decisions. The experiments on SemEval datasets demonstrate the efficiency and effectiveness of our models.