Research On Leaveraging Knowledge For Sentiment Analysis

Sentiment analysis is a hot research topic in natural language processing, it mainly focuses on identifying and extracting subjective information from online reviews. Accurate understanding sentiment in reviews plays an important role in guiding decision making and business intelligence. For many research problems in sentiment analysis, we more or less have some common sense. For instance, we know that “price” is a synonym of “cost”. Leveraging such prior knowledge as weak supervision might be the key to promote the performance in sentiment analysis.In this thesis, we study the problems of sentiment extraction, sentiment & aspect classification, review summarization from the perspective of leveraging knowledge. The main contributions are listed as follows:? Sentiment extraction, which targets at extracting feature and opinion word from reviews, is a fundamental work in sentiment analysis. For this task, we build a simple yet robust unsupervised model with the grammar knowledge incorporated.The grammar knowledge is obtained through statistics on large data, which does not rely on manually collected patterns or labeled data. Our approach leverages rich language structures on large data, which is an advantage over existing methods.? Sentiment & aspect classification refers to document-level sentiment classification and sentence-level aspect assignment. In this task, we formalize the use of aspect signature terms as weak supervision in a structural learning framework, which remarkably promotes aspect-level analysis. Also, through joint modeling, the performance of aspect analysis and document-level sentiment classification are mutually enhanced. Our approach learns a sentence-level aspect classifier only with several aspect signature terms, and does not require training data with aspect labeled.? To summarize a single review, we study the problem of sentence informativeness ranking, because for one review, not every part is equally informative. We rank each sentence considering sentence text and sentence sentiment. For sentence text ranking, our approach is very flexible to incorporate various heuristic ranking rules as prior knowledge. With the help of such ranking knowledge, our approach does not need manual annotation.? For the task of summarizing a collection of reviews, we propose to extend topic model with semantic knowledge incorporated, and then generate personalized review summary. By introducing semantic knowledge, our model is capable of modeling sentiment, aspects and ratings of online reviews without increasing model complexity.