Hierarchical Approach to Sentiment Analysis

Abstract

model to discover and distinguish aspecSentiment analysis aims to extract the customer'sattitude and feeling from his/her unstructured reviews by separatingthe subjective information from the other information.We propose a generative probabilistic topic model that detectsboth an aspect and corresponding sentiment, simultaneously,from review articles. Unlike existing sentiment analysis models,which generally consider rating prediction to be a side task, ourproposal, the hierarchical approach to sentiment analysis, identifiesboth an item and its rating by dividing topics, traditionallytreated as one entity, into aspect and sentiment topics. Since ourmodel is aware of both objective and subjective information, itcan discover fine-grained tightly coherent topics, and describethe generative process of each article in a unified manner. Tohandle the differences involved, HASA extends topic models byintroducing both observed variables and a latent switch variableinto each token, where topics are influenced not only by word cooccurrencebut also item/rating information, and then classifyingthem as either aspect or sentiment topics. Experiments on reviewarticles show that the proposed model is useful as a generativets from sentiments.