Seeing Several Stars: A Rating Inference Taskfor a Document Containing Several Evaluation Criteria

摘要

In this paper we address a novel sentiment analysis task of rating inference. Previous rating inference tasks, which are sometimes referred to as "seeing stars", estimate only one rating in a document. However reviewers judge not only the overall polarity for a product but also details for it. A document in this new task contains several ratings for a product. Furthermore the range of the ratings is zero to six points (i.e., stars). In other words this task denotes "seeing several stars in a document". If significant words or phrases for evaluation criteria and their strength as positive or negative opinions are extracted, a system with the knowledge can recommend products for users appropriately. For example, the system can output a detailed summary from review documents. In this paper we compare several methods to infer the ratings in a document and discuss a feature selection approach for the methods. The experimental results are useful for new researchers who try this new task.