Feature Selection for Polarity Classification
Speaker: Michael Wiegand
Institution: Saarland University
Abstract:
In this talk, I will present some of my results on feature selection for semi-supervised polar classification on document level. Feature selection is vital in classification tasks where the amount of labeled data is sparse. This is the case in semi-supervised learning. I will compare different kinds of feature sets encompassing simple bag-of-words, in-domain adjectives (gained from unlabeled documents), statistically induced feature sets and (semi-)manually constructed polarity lexicons. In order to substantiate that the insights of this work are universal, my experiments have been carried out not only on the benchmark dataset of IMDB movie reviews but also on other domains. I will also give a brief overview of my recent work on feature extraction for supervised classification on sentence level. Feature extraction is important in this scenario because standard bag-of-words features tend to provide too sparse feature vectors despite the presence of many labeled data instances.
