Cross collection aspect based opinion mining using topic models

Official URL: http://risc01.sabanciuniv.edu/record=b1817215 (Table of Contents)

Aspect based opinion mining is the automated science of identifying and extracting sentiments associated to individual aspects in a text document. Over the years this science has emerged to be a cornerstone for analysis of public opinion on consumer products and social-political events. The task is more fruitful and likewise more challenging when comparison of opinion on aspects of multiple entities is of essence. Different methods in literature have attempted to extract aspects in a single collection or collection by collection across multiple collection. These approaches do not appeal when number of collections is large and hence su er significant performance drawbacks. In this work we perform aspect based opinion mining across contrasting multiple collections, simultaneously. We utilize existing cross collection topic models to identify topics that prevail across multiple collections, we propose a topic refinement algorithm that successfully converts these topics into semantically coherent and visually identifiable aspects. We compare the quality of aspects extracted by our algorithm to topics returned by two cross collection topic models. Finally we evaluate the accuracy of sentiment scores when measured over features extracted by the two cross collection topic models. We conclude that with proposed improvements cross collection topic models outperform state of art approaches in aspect based sentiment analysis.