Unsupervised Solution Post Identification from Discussion Forums

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 155–164, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Unsupervised Solution Post Identification from Discussion Forums Deepak P IBM Research - India Bangalore, India deepak.s.p@in.ibm.com Karthik Visweswariah IBM Research - India Bangalore, India v-karthik@in.ibm.com Abstract Discussion forums have evolved into a de- pendable source of knowledge to solve common problems. However, only a mi- nority of the posts in discussion forums are solution posts. Identifying solution posts from discussion forums, hence, is an important research problem. In this pa- per, we present a technique for unsuper- vised solution post identification leverag- ing a so far unexplored textual feature, that of lexical correlations between problems and solutions. We use translation mod- els and language models to exploit lex- ical correlations and solution post char- acter respectively. Our technique is de- signed to not rely much on structural fea- tures such as post metadata since such features are often not uniformly available across forums. Our clustering-based itera- tive solution identification approach based on the EM-formulation performs favor- ably in an empirical evaluation, beating the only unsupervised solution identifica- tion technique from literature by a very large margin. We also show that our unsu- pervised technique is competitive against methods that require supervision, outper- forming one such technique comfortably. 1 Introduction Discussion forums have become a popular knowl- edge source for finding solutions to common prob- lems. StackOverflow1, a popular discussion forum for programmers is among the top-100 most vis- ited sites globally2. Now, there are discussion fo- rums for almost every major product ranging from 1http://www.stackoverflow.com 2http://www.alexa.com/siteinfo/stackoverflow.com automobiles3 to gadgets such as those of Mac4 or Samsung5. These typically start with a registered user posting a question/problem6 to which other users respond. Typical response posts include so- lutions or clarification requests, whereas feedback posts form another major category of forum posts. As is the case with any community of humans, discussion forums have their share of inflamma- tory remarks too. Mining problem-solution pairs from discussion forums has attracted much atten- tion from the scholarly community in the recent past. Since the first post most usually contains the problem description, identifying its solutions from among the other posts in the thread has been the focus of many recent efforts (e.g., (Gandhe et al., 2012; Hong and Davison, 2009)). Extract- ing problem-solution pairs from forums enables the usage of such knowledge in knowledge reuse frameworks such as case-based reasoning (Kolod- ner, 1992) that use problem-solution pairs as raw material. In this paper, we address the problem of unsupervised solution post identification7 from discussion forums. Among the first papers to address the solution identification problem was the unsupervised ap- proach proposed by (Cong et al., 2008). It em- ploys a graph propagation method that prioritizes posts that are (a) more similar to the problem post, (b) more similar to other posts, and (c) authored by a more authoritative user, to be labeled as so- lution posts. Though seen to …

Content truncated to 500 words