Classification and Retrieval from Mailing Lists and Forums

In this paper, we address the tasks of classifying the posts of mailing lists and forums, and retrieving relevant information from the same (describing our participation in the FIRE 2010 evaluation sub-tasks for the same). We approach the classification problem in two ways. In the first method, we pose it as a sequence labeling task using conditional random fields (CRF), where the entire thread information is used to find the labels for each of the posts in that thread. In the second approach, we model it as a document classification task considering each post as an independent document and using a support vector machine (SVM) like classifier. We discuss the performances of both of these approaches, and also, the effect of different features that we used for training the CRFs. For the retrieval task, the approach we propose is to combine the text match score with a goodness score of the thread. This score is a measure of the quality of the resolution text contained in the thread. A classifier is employed to identify posts of the thread that could be a solution, and the goodness score is computed based on its contents.