Opinion Mining with Semantic Analysis – Københavns Universitet

Opinion Mining with Semantic Analysis


Specialeforsvar ved Søren Houen


Abstract Opinion mining is a subtopic of information retrieval with considerable research done. Several methods exist to determine an author's view on a topic from natural language textual information. These generally employ some form of machine learning approach, and have varying degrees of effectiveness. This thesis explores the use of the semantic frame-based analyzer FrameNet in order to improve upon existing methods of sentiment analysis by providing automatically tagged semantic information for use by a sentiment classifying algorithm. This is achieved through the application of semantic information as a feature for a machine learning-based classifier, as well as using semantic analysis in a bootstrapping process to create a sentiment lexicon for use by a classifier.

Several different approaches of using FrameNet meta-data features were attempted, with pure FrameNet data achieving an F1-Score of 0.651, as opposed to a simple bag-of-tokens approach which yielded an F1-Score of 0.819. Combination approaches of bag-of-tokens and FrameNet improved upon the pure FrameNet score, but did not achieve results better than the bag-of-tokens. The sentiment lexicon approach was found to achieve results of 0.630 F1-Score. From the findings in this study, it would seem that using FrameNet as a tool for opinion mining is not a rewarding avenue of research.

Vejleder: Jakob Grue Simonsen, DIKU

Censor: Troels Andreasen, Roskilde Universitet