Research On Opinion Target Extraction

With the rapid development of Web 2.0 technologies, the opinion-rich resources in the Internet such as online review sites and personal blogs are growing rapidly. The sudden eruption of activities in the area of sentiment analysis, which deals with the computational treatment of opinion, sentiment, and subjectivity in text, has been drawing more and more attention in recent years in natural language processing (NLP). Sentiment analysis is also known as opinion mining, which aims to obtain, organize and analyze opinion related information.Sentiment analysis has made much progress in recent years, especially in sentiment polarity classification. Nowadays, more and more researchers focus on fine-grained sentiment analysis and its related applications, such as sentiment information extraction, sentiment retrieval and opinion summarization. Among them, opinion target extraction/identification is one of the most important subtasks of sentiment information extraction in sentiment analysis.This paper carries out extensive research on the key techniques of opinion target extraction, with focus on:1. CRFs(Conditional Random Fields)-based opinion target extraction. After employing frequently used features in sentiment information extraction, we summarize all the features into four categories, i.e. lexical, syntactic, relative-positional and semantic features, and detailed comparative studies have been made to evaluate the performance by exploring various features and their combination.2. Considering the ubiquitous phenomenon of domain adaptation in sentiment analysis, we address the multi-domain opinion target extraction and present our algorithm. Given the training data from multiple domains, we propose a new model-level fusion method to train classifiers using all the data simultaneously, so that training instances from different domains may help each other. We apply regular fusion methods to this task, and propose our modified meta-learning algorithm according to the task specific settings.3. With the maturity of semantic role labeling (SRL) techniques, it has been successfully applied to many other NLP tasks. In this paper, we explore SRL as a specific feature to improve opinion target extraction.The major contributions of this paper lie on the following four aspects:1. New features, such as dependency relations and whether the word is an opinion word, are proposed. We sum them up into four categories, and successfully apply to opinion target extraction.2. Various experiments have been made to systematically compare the contributions of different features and their combinations.3. A task-specific improved fusion algorithm is proposed to multi-domain opinion target extraction.4. Exploration of SRL-based semantic roles on opinion target extraction shows that SRL is a good indicator for opinion target.Experiments show that above research not only improves the performance of opinion target extraction, but also exhibits great reference value for the future research in sentiment analysis.