Adaptive semi supervised opinion classifier with forgetting mechanism

Opinion stream classification methods face the challenge of learning with a limited amount of labeled data: inspecting and labeling opinions is a tedious task, so systems analyzing opinions must devise mechanisms that label the arriving stream of opinionated documents with minimal human intervention. We propose an opinion stream classifier that only uses a seed of labeled documents as input and thereafter adapts itself, as it reads documents with unknown labels. Since the stream of opinions is subject to concept drift, we use two adaptation mechanisms: forward adaptation, where the classifier incorporates to the training set only those un-labeled documents that it considers informative enough in comparison to those seen thus far; and backward adaptation, where the classifier gradually forgets old documents by eliminating them from the model. We evaluate our method on opinionated tweets and show that it performs comparably or even better than a fully supervised baseline.