View article

In this paper, we present our methods for the 2016 CLPPsych shared task. We extracted and selected eight features from the corpus consisting of posts from ReachOut. com including the information of the post’s source board, numbers of kudos and views, post time, ranks of the authors, unigram of the body and subject, frequency of the used emotion icons, and the topic model features. Two support vector machine models were trained with the extracted features. A baseline system was also developed, which uses the calculated log likelihood ratio (LLR) for each token to rank a post. Finally, the prediction results of the above three systems were integrated by using a global ranking algorithm with the weighted Borda-fuse (WBF) model and the linear combination model. The best F-score achieved by our systems is 0.3 which is based on the global ranking method with WBF.