Building Contrastive Summaries of Subjective Text Via Opinion Ranking

Abstract

This article investigates methods to automatically compare entities from opinionated text to help users to obtain important information from a large amount of data, a task known as “contrastive opinion summarization”. The task aims at generating contrastive summaries that highlight differences between entities given opinionated text (written about each entity individually) where opinions have been previously identified. These summaries are made by selecting sentences from the input data. The core of the problem is to find out how to choose these more relevant sentences in an appropriate manner. The proposed method uses a heuristic that makesdecisions according to the opinions found in the input text and to traits that a summary is expected to present. The evaluation is made by measuring three characteristics that contrastive summaries are expected to have: representativity (presence of opinions that are frequent in the input), contrastivity (presence of opinions that highlight differences between entities) and diversity (presence of different opinions to avoid redundancy). The novel method is compared to methods previously published and performs significantly better than them according to the measures used. The main contributions of this work are: a comparative analysis of methods of contrastive opinion summarization, the proposal of a systematic way to evaluate summaries, the development of a new method that performs better than others previously known and the creation of a dataset for the task.