Word-level Contextual Sentiment Analysis with Interpretability

Word-level contextual sentiment analysis (WCSA) is an important task for mining reviews or opinions. When analyzing this type of sentiment in the industry, both the interpretability and practicality are often required. However, such a WCSA method has not been established. This study aims to develop a WCSA method with interpretability and practicality. To achieve this aim, we propose a novel neural network architecture called Sentiment Interpretable Neural Network (SINN). To realize this SINN practically, we propose a novel learning strategy called Lexical Initialization Learning (LEXIL). SINN is interpretable because it can extract word-level contextual sentiment through extracting word-level original sentiment and its local and global word-level contexts. Moreover, LEXIL can develop the SINN without any specific knowledge for context: therefore, this strategy is practical. Using real textual datasets, we experimentally demonstrate that the proposed LEXIL is effective for improving the interpretability of SINN and that the SINN features both the high WCSA ability and high interpretability.