Sorting Forum Responses by Relevance to Original Post Aaron

Online forums provide a medium for discussing niche topics. Forums consist of threads, where each thread consists of an original post followed by response posts. A subset of original posts are questions related to the forum’s topic. The responses to these posts are often expert answers due to the niche nature of the forum. However, not all responses will be answers. Users are free to write anything in their response post. In this project, we train a Näıve Bayes classifier and a linear support vector machine (SVM) classifier to distinguish thread answers from nonanswers. Our features are preprocessed whitespace-delimited terms occurring in each response. We obtained training and testing data by implementing a web crawler and a web parser targeted against the online discussion forum Slickdeals.net/forums (1). We crawled T = 8 threads whose original posts were questions. There was an average of 94.25 responses for each post, giving a total of N = 754 responses for classification. We manually assigned a labeling of 0 or 1 to each response depending upon whether that response attempted to answer the original question. We performed leave-one-out crossvalidation on the eight test questions. The Näıve Bayes classifier correctly classified 70.1% (110/157) of the answers and 74% of the non-answers (442/597). The linear-kernel SVM correctly classified 80.2% of the answers (126/157) and 82.5% (493/597) of the non-answers.