A Semi-Supervised Self-Adaptive Classifier over Opinionated Streams

We investigate the problem of polarity learning over a stream of opinionated documents. We deal with two challenges. First, if the opinions are not labeled, then we cannot assume that a human expert will be regularly and frequently available to assess the sentiment of arriving documents for learning and model adaption. Further, the vocabulary of the stream, and thus the feature space used for learning, changesover time: people use an abundancy of words, and sometime seven invent new ones to express their feelings. We propose a semi-supervised opinion stream classification algorithm that uses only an initial training set of labeled documents for polarity learning and gradually adapts to changes in the vocabulary. In particular, our algorithm S*3 Learner starts with the vocabulary of opinionated words that are in the documents of the initial training set, and then expands it with new words, as soon as there is enough evidence for estimating their polarity. We study the performance of S*3 Learneron opinionated streams under the natural order of document arrival and under a modified ordering that allows us to simulate vocabulary evolution.