Filling the Gap: Semi-Supervised Learning for Opinion Detection Across Domains

Proceedings of the Fifteenth Conference on Computational Natural Language Learning, pages 200–209, Portland, Oregon, USA, 23–24 June 2011. c⃝2011 Association for Computational Linguistics Filling the Gap: Semi-Supervised Learning for Opinion Detection Across Domains Ning Yu Indiana University nyu@indiana.edu Sandra K¨ubler Indiana University skuebler@indiana.edu Abstract We investigate the use of Semi-Supervised Learning (SSL) in opinion detection both in sparse data situations and for domain adapta- tion. We show that co-training reaches the best results in an in-domain setting with small la- beled data sets, with a maximum absolute gain of 33.5%. For domain transfer, we show that self-training gains an absolute improvement in labeling accuracy for blog data of 16% over the supervised approach with target domain training data. 1 Introduction Rich and free opinions published electronically and, more recently, on the WWW offer ample opportuni- ties to discover individual’s attitudes towards certain topics, products, or services. To capitalize on this enormous body of opinions, researchers have been working in the area of opinion mining since the late 1990s. Opinion detection seeks to automatically de- termine the presence or absence of opinions in a text, and it is therefore a fundamental task for opinion mining. In order to capture subtle and creative opinions, opinion detection systems generally assume that a large body of opinion-labeled data are available. However, collections of opinion-labeled data are of- ten limited, especially at the granularity level of sen- tences; and manual annotation is tedious, expensive and error-prone. The shortage of opinion-labeled data is less challenging in some data domains (e.g., reviews) than in others (e.g., blog posts). A sim- ple method for improving accuracies in challenging domains would be to borrow opinion-labeled data from a non-target data domain; but this approach often fails because opinion detection strategies de- signed for one data domain generally do not perform well in another domain. One reason for failure of the simple transfer approach is that the information used for opinion detection is typically lexical, and lexical means of expressing opinions may vary not only from domain to domain, but also from register to register. For example, while the word ”awesome” is a good indicator of an opinion in blogs, it is less likely to occur in the same role in newspaper texts. While it is difficult to obtain opinion-labeled data, one can easily collect almost infinite unlabeled user- generated data that contain opinions. The use of Semi-Supervised Learning (SSL), motivated by lim- ited labeled data and plentiful unlabeled data in the real world, has achieved promising results in vari- ous NLP studies (e.g., (F¨urstenau and Lapata, 2009; Talukdar and Pereira, 2010)), yet it has not been fully investigated for use in opinion detection. Al- though studies have shown that simple SSL meth- ods are promising for extracting opinion features or patterns using limited opinion-labeled data (e.g., (Wiebe and Riloff, 2005)), few efforts have been made either to apply SSL directly to opinion detec- tion or to examine more sophisticated SSL methods. This research is intended to fill the gap …

Content truncated to 500 words