Word polarity attention in sentiment analysis

Abstract

Neural network approaches are end-to-end learning approaches without well-designed training data and achieve high performance in sentiment analysis. Because of complex architecture of a neural network, it is difficult to analyze how they work and find their bottleneck to improve their performance. To remedy it, we propose neural sentiment analysis with attention mechanism. Using attention mechanism, we can find important words to determine sentiment polarity of a sentence. Moreover, we can understand why the sentiment analysis could not classify sentiment polarity correctly. We compare our method with neural sentiment analysis without attention mechanism over TSUKUBA corpus and Stanford Sentiment Treebank (SST). Experimental results show that our method is interpretable and can achieve higher precision.