Interpretable Sentiment Analysis Using the Attention-Based Multiple Instance Classification Model: An Application to Wine Reviews

Sentiment analysis (SA) is widely used for analyzing text data to identify the underlying opinion or emotion expressed in a document. Neural network methods in natural language processing have produced state-of-the-art results in SA. However, many of those methods are ‘black-box’ algorithms that do not provide interpretable results. In this article, we develop an interpretable word-level context-based SA tool, which is an attention-based multiple instance classification model (AMIC). In text analysis, the attention mechanism helps models understand the context of words and sentences. AMIC uses word embedding to transform text to data, employs a multiple instance classification structure, and incorporates the attention mechanism to include context information from documents. AMIC has a transparent model structure with interpretable word-level context-based sentiment metrics. In addition, AMIC can be used to construct domain-specific sentiment dictionary without requiring prior information on seed words or a pre-trained list of sentiment words. We demonstrate the performance of AMIC using a large online wine review data set, decoding the sentiment embedded in the texts and gaining insight on the specific usage of wine review languages.