Aspect and Entity Extraction from Opinion Documents.

摘要

Opinion mining has been an active research area in Web mining and Natural Language Processing (NLP) in recent years. In this thesis, we present a comprehensive study of aspect and entity extraction from opinion documents for opinion mining. We first introduce the aspect-based opinion mining model. Then, we propose a new method for aspect extraction and ranking, which is based on language patterns and dependency grammar. Meanwhile, it is capable of ranking extracted aspects by their importance, i.e. relevancy and frequency. In addition, we discover that there are two kinds of special product aspects in some domains. One is noun aspect implying opinion. The other is the resource term. Novel extraction algorithms are proposed to identify them from opinion documents. In terms of entity extraction task, it is similar to the classic named entity extraction (NER) problem. However, there is a major difference. In a typical opinion mining application, the users often want to find opinions on some competing entities, e.g., competing or relevant products. This implies that the discovered entities must be of the same type/class. Basically, this is a set expansion problem. To deal with this problem, we present two set expansion algorithms for entity extraction in opinion documents. One is based on positive and unlabeled (PU) learning model. The other is based on Bayesian Sets. We also discuss extracting topic documents from a collection. Opinion mining system crawls and indexes opinion documents first and then used for different specific tasks later. Typically, the documents are not well categorized because one does not know what the future tasks will be. Normally, keyword search is used to find relevant opinion documents for analysis. However, the documents that are retrieved in this way can have both low recall and low precision. Another way is to train a document classifier. But the training procedure is time-consuming and labor-intensive. We propose an unsupervised technique to solve this problem based on a new PU learning algorithm.