Determining the Polarity of Postings for Discussion Search

When performing discussion search it might be desirable to consider non-topical measures like the number of positive and negative replies to a posting, for instance as one possible indicator for the trustworthiness of a comment. Systems like POLAR are able to integrate such values into the retrieval function. To automatically detect the polarity of postings, they need to be classified into positive and negative ones w.r.t. the comment or document they are annotating. We present a machine learning approach for polarity detection which is based on Support Vector Machines. We discuss and identify appropriate term and context features. Experiments with ZDNet News show that an accuracy of around 79%80% can be achieved for automatically classifying comments according to their polarity.