Domain Adaptation of Sentiment Classifiers

27^th June 2013, 11:00 G12
Dr. Danushka Bollegala
Dept. of Information and Communication Engineering
The University of Tokyo
Japan

Abstract

The ability to adapt to novel environments is an important property of human intelligence. Although supervised learning algorithms have come a long way and can accurately learn in numerous tasks using large training data, such approaches fail when the data used for training become obsolete or even marginally different from the actual task. In this talk, I explain the problem of adapting to novel domains in the context of sentiment classification. In sentiment classification, given a review written by a user on a certain product or service, we must determine its sentiment (i.e. positive or negative). However, humans use different words to express sentiment about a particular product, and a sentiment classifier trained using labeled data for one product (e.g. books) does not work well on a different product (e.g. kitchen appliances) because of the feature mismatch problem. In this talk, I will present a solution to adapt a sentiment classifier to a novel domain using unlabeled data. I will also explain the close relationship between domain adaptation and deep learning methods.