This is a two-part talk. In the first part, I will report our work on TREC 2006 Blog track opinion retrieval task which which involves locating English blog posts that express an opinion about a given target. We test three main ideas: effect of natural paragraphs vs fixed sized passages on topic retrieval and opinion retrieval; effect of demotion of non-opinionated documents in the ranked list on opinion retrieval; query formulation using the title fields only vs using both the title and the description fields. We found that, for both topic retrieval and opinion retrieval, natural paragraphs work better than fixed sized passages, and query formulation using both the title and description fields works better than using the title fields only, but demotion does not help or hurt much. Our opinion detection approach was simple - we collected Wilson and Wiebe's senitment lexicon, and computed the semantic orientations of words using Turney's Pointwise Mutual Information approach, then aggregated the semantic orientations of words to get the sentiment of documents.
In the second part, I will talk about my dissertation proposal which focuses on Chinese attitude classification and aspect classification for English and Chinese documents in the context of bilingual sentiment aggregation. Previous work on sentiment analysis has been focusing on the analyzing the sentiment of English targets overally, whereas Chinese sentiment anslysis has been sparse. Moreover, we assume people not only evaluate things overally, but also evaluate the aspects of things, so a component of aspect classification is needed in the framework of attitude classification. I will address the approaches for dealing with these two research problems and would like to hear your advice and comments.
Yejun Wu is a doctoral candidate at College of Information Studies, working with Dr. Oard on information retrieval and evaluation problems.
This talk is part of the CLIP Colloquium Series, organized by Jimmy Lin (jimmylin -at- umd .dot. edu). For the complete schedule, please visit http://www.umiacs.umd.edu/research/CLIP/colloq/.