Detecção de opiniões e análise de polaridade em documentos financeiros com múltiplas entidades.

Abstract

Polarity analysis aims at classifying the author’s opinion into positive, negative, or neutral. However, given the sheer volume of information available on the web, manually carrying out such task is unfeasible. In particular, in the financial domain this type of analysis is useful for companies in making decisions related to the financial market which is particularly prone to changes according to shifting of opinions. Most studies in literature deal with this problem by considering that documents have a global polarity. However, in general, documents cite several entities with possibly different polarities. This suggests that the classification should be performed in an entity level. Besides this problem, we also noted that many financial documents do not always emit opinion. Thus, a first task of interest in this research field is to identify documents on which opinions are expressed, that is, the subjective ones. Therefore, in this paper we propose a supervised polarity classification method based on multiple models to deal with financial documents with multiple entities. In particular, we study text segmentation strategies that use heuristics such as string matching and anaphora resolution and we propose a hierarchical classification method based on subjectivity detection. Our results showed that the multiple-models approach significantly outperformed the global-model baseline. The segmentation of the documents restricted to sentences that mention entities and the adoption of a hierarchical strategy also achieved gains, although modest.