Tagging and Linking Web Forum Posts

Proceedings of the Fourteenth Conference on Computational Natural Language Learning, pages 192–202, Uppsala, Sweden, 15-16 July 2010. c⃝2010 Association for Computational Linguistics Tagging and Linking Web Forum Posts Su Nam Kim, Li Wang and Timothy Baldwin Dept of Computer Science and Software Engineering University of Melbourne, Australia sunamkim@gmail.com, li.wang.d@gmail.com, tb@ldwin.net Abstract We propose a method for annotating post- to-post discourse structure in online user forum data, in the hopes of improving troubleshooting-oriented information ac- cess. We introduce the tasks of: (1) post classification, based on a novel dialogue act tag set; and (2) link classification. We also introduce three feature sets (structural features, post context features and seman- tic features) and experiment with three dis- criminative learners (maximum entropy, SVM-HMM and CRF). We achieve above- baseline results for both dialogue act and link classification, with interesting diver- gences in which feature sets perform well over the two sub-tasks, and go on to per- form preliminary investigation of the inter- action between post tagging and linking. 1 Introduction With the advent of Web 2.0, there has been an ex- plosion of web authorship from individuals of all walks of life. Notably, social networks, blogs and web user forums have entered the mainstream of modern-day society, creating both new opportuni- ties and challenges for organisations seeking to en- gage with clients or users of any description. One area of particular interest is web-based user sup- port, e.g. to aid a user in purchasing a gift for a friend, or advising a customer on how to config- ure a newly-acquired wireless router. While such interactions traditionally took place on an indi- vidual basis, leading to considerable redundancy for frequently-arising requests or problems, user forums support near-real-time user interaction in the form of a targeted thread made up of individ- ual user posts. Additionally, they have the poten- tial for perpetual logging to allow other users to benefit from them. This in turn facilitates “sup- port sharing”—i.e. the ability for users to look over the logs of past support interactions to deter- mine whether there is a documented, immediately- applicable solution to their current problem—on a scale previously unimaginable. This research is targeted at this task of enhanced support sharing, in the form of text mining over troubleshooting- oriented web user forum data (Baldwin et al., to appear). One facet of our proposed strategy for enhanc- ing information access to troubleshooting-oriented web user forum data is to preprocess threads to uncover the “content structure” of the thread, in the form of its post-to-post discourse structure. Specifically, we identify which earlier post(s) a given post responds to (linking) and in what man- ner (tagging), in an amalgam of dialogue act tag- ging (Stolcke et al., 2000) and coherence-based discourse analysis (Carlson et al., 2001; Wolf and Gibson, 2005). The reason we do this is gauge the relative role/import of individual posts, to in- dex and weight component terms accordingly, ul- timately in an attempt to enhance information ac- cess. Evidence to suggest that this structure …

Content truncated to 500 words