Hierarchical Neural Network with Bidirectional Selection Mechanism for Sentiment Analysis

Document-level sentiment classification is a process that aims to predict the sentiment rating of a particular document. The most popular methods take word embeddings as the inputs of a neural network. However, there are many different senses that coexist in a word embedding; such ambiguous word representation will cause models to misunderstand a text, which will result in predicting the sentiment incorrectly. Accordingly, to make the information in the text representation more explicit, we propose a Bidirectional Selection Mechanism (Bi-SM), which filters out redundancy from different perspectives by alternately selecting salient information contained in the text representation and the category representation. Bi-SM mainly contains two parts-Selection and Back Selection. At the word level, the Selection module uses the category of each word to “filter” out redundant senses. Meanwhile, the category will be enriched in this module for further being used in the Back Selection module. Subsequently, the Back Selection module uses the “filtered” word information to “select” the salient “enriched category”, which is further used at the sentence level to “filter” out redundancy in the sentence representation. The same thing happened at the sentence level. Our experimental results demonstrate that our proposed models can achieve consistent improvements compared to state-of-the-art methods.