Research And Realization Of Explanatory Relationship Extraction In Commentary Text

Sentiment representation methods and automatic analyzing approaches have received great attentions in recent years. Most existing researches have focused focuses on classifying or identifying opinions, emotions, and sentiments at various levels of granularity from unstructured text. Fine-grained methods try to extract opinion holder, opinion expression, opinion target, and other opinion elements, in order to answer the basic question "who expresses what opinion on which target".This paper studies the problem of extracting sentences which are explanations of corresponding opinion fragments from product reviews. From analyzing product reviews with high human assigned helpfulness ratings, we observe that those reviews include not only the positive and negative aspects of the product, but also explanations of them. Since most of the consumers are usually not experts, given the more easily comprehensible explanations with the corresponding sentiments would be helpful to the average consumer. In this work, we investigate to use a learning method based on Integer Linear Programming to extract explanatory sentences. Experimental evaluations on a manually labeled corpus demonstrate the effectiveness of the proposed approach.The main contributions of our work are as follows:1. We formally define the explanatory sentence. To the best of our knowledge, this is the first focused study on this relationship.2. Since there is no public corpus available, we collected a large number of product reviews and deployed the corpus labeling task on Amazon’s Mechanical Turk.3. We propose a supervised learning method which takes an article as input and sentiments with their corresponding explanatory sentences as output. The algorithm is two phrases and based on Integer Linear Programming, which helps to concisely handle various global constraints.