Semi-Supervised Answer Extraction from Discussion Forums

International Joint Conference on Natural Language Processing, pages 1–9, Nagoya, Japan, 14-18 October 2013. Semi-Supervised Answer Extraction from Discussion Forums Rose Catherine, Rashmi Gangadharaiah, Karthik Visweswariah, Dinesh Raghu IBM Research Bangalore, India {rosecatherinek, rashgang, v-karthik, diraghu1} @in.ibm.com Abstract Mining online discussions to extract an- swers is an important research problem. Methods proposed in the past used super- vised classifiers trained on labeled data. But, collecting training data for each target forum is labor intensive and time consum- ing, thus limiting their deployment. A re- cent approach had proposed to extract an- swers in an unsupervised manner, by tak- ing cues from their repetitions. This as- sumption however, does not hold true in many cases. In this paper, we propose two semi-supervised methods for extract- ing answers from discussions, which uti- lize the large amount of unlabeled data available, alongside a very small training set to obtain improved accuracies. We show that it is possible to boost the perfor- mance by introducing a related, but paral- lel task of identifying acknowledgments to the answers. The accuracy achieved by our approaches surpass the baselines by a wide margin, as shown by our experiments. 1 Introduction Online discussion forums, also known as com- munity question answering (CQA) sites, are inter- net sites that provide a medium for users to dis- cuss and share information on a wide range of top- ics. Due to their vast popularity, gradually, they have aggregated a massive collection of discussion data. Mining such forums have numerous appli- cations such as improving question–answer (QA) retrieval (Cong et al., 2008), learning important insights like features of products that are draw- ing negative reviews (Lakkaraju et al., 2011) or discovering longstanding unresolved severe tech- nical issues (Gangadharaiah and Catherine, 2012) etc. For this reason, substantial research effort has been directed at mining discussions, in recent times. In this paper, we focus on the specific prob- lem of extracting answers from these discussions. In forums, typically a user starts a discussion by posting a question to which multiple mem- bers of the forum suggest answers. The discus- sion evolves into a complex multi-party conversa- tion as the question gets refined, with additional details specified, clarifications sought, multiple answers provided, frequent digressions, and oc- casional follow-up discussions and acknowledg- ments, altogether spanning several pages. An- swers easily get buried deep within this and locat- ing them automatically is far from straightforward. In this paper, we propose two semi-supervised approaches that require only a very small amount of training data (only 3 manually tagged discus- sion threads ) and achieve high accuracy levels by using the available unlabeled data. With this, we eliminate the need to collect vast amounts of train- ing data, thus aiding faster deployment for new do- mains. Specifically, our contributions are: • A semi-supervised answer extraction method for discussions: This paper makes the first attempt at extracting answers from discussions in a semi- supervised manner. We show how existing fea- tures can be engineered into a co-training frame- work to …

Content truncated to 500 words