Semi-Supervised Learning For Identifying Opinions In Web Content

Loading...
Thumbnail Image
Date
2011-10-19
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
[Bloomington, Ind.] : Indiana University
Abstract
Opinions published on the World Wide Web (Web) offer opportunities for detecting personal attitudes regarding topics, products, and services. The opinion detection literature indicates that both a large body of opinions and a wide variety of opinion features are essential for capturing subtle opinion information. Although a large amount of opinion-labeled data is preferable for opinion detection systems, opinion-labeled data is often limited, especially at sub-document levels, and manual annotation is tedious, expensive and error-prone. This shortage of opinion-labeled data is less challenging in some domains (e.g., movie reviews) than in others (e.g., blog posts). While a simple method for improving accuracy in challenging domains is to borrow opinion-labeled data from a non-target data domain, this approach often fails because of the domain transfer problem: Opinion detection strategies designed for one data domain generally do not perform well in another domain. However, while it is difficult to obtain opinion-labeled data, unlabeled user-generated opinion data are readily available. Semi-supervised learning (SSL) requires only limited labeled data to automatically label unlabeled data and has achieved promising results in various natural language processing (NLP) tasks, including traditional topic classification; but SSL has been applied in only a few opinion detection studies. This study investigates application of four different SSL algorithms in three types of Web content: edited news articles, semi-structured movie reviews, and the informal and unstructured content of the blogosphere. SSL algorithms are also evaluated for their effectiveness in sparse data situations and domain adaptation. Research findings suggest that, when there is limited labeled data, SSL is a promising approach for opinion detection in Web content. Although the contributions of SSL varied across data domains, significant improvement was demonstrated for the most challenging data domain--the blogosphere--when a domain transfer-based SSL strategy was implemented.
Description
Thesis (Ph.D.) - Indiana University, Information Science, 2011
Keywords
Semi-Supervised Learning, Sentiment Analysis, Domain Transfer, Opinion Detection, blog, text mining, Co-Training, Self-Training
Citation
DOI
Link(s) to data and video for this item
Relation
Rights
This work is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported (CC BY-NC-SA 3.0) license.
Type
Doctoral Dissertation