Extracting Problem and Resolution Information from Online Discussion Forums.

The ability to obtain quick solutions to problems is an important requirement in many practical applications such as help desks and technical support. In this paper, we describe an approach that will enable easy identification of potential solutions to a given problem. The proposed approach involves extraction of useful problem- and resolution-related information from online discussion forums. We specifically focus on identifying important parts of a discussion thread by classifying message posts in the thread as problem- or resolution-related. We cast this problem as a sequence labeling task and train a discriminative Conditional Random Field for supervised learning. Results are presented from classification experiments done at a sentence as well as phrase-level. The structure and information flow pattern in a typical discussion thread helps generate effective features for classification. We also discuss the effect of inducing different features on the precision and recall of the sequence labeling task. Sentence level classification with feature induction yielded F1 scores of 66% for problem related information and 58.8% for resolution-related information, which is a significant improvement over our baseline scores.