MixKMeans: Clustering Question-Answer Archives

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1576–1585, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics MixKMeans: Clustering Question-Answer Archives Deepak P Centre for Data Sciences and Scalable Computing Queen’s University Belfast, UK deepaksp@acm.org Abstract Community-driven Question Answering (CQA) systems that crowdsource experiential information in the form of questions and answers and have accumulated valuable reusable knowledge. Clustering of QA datasets from CQA systems provides a means of organizing the content to ease tasks such as manual curation and tagging. In this paper, we present a clustering method that exploits the two-part question-answer structure in QA datasets to improve clustering quality. Our method, MixKMeans, composes question and answer space similarities in a way that the space on which the match is higher is allowed to dominate. This construction is motivated by our observation that semantic similarity between question-answer data (QAs) could get localized in either space. We empirically evaluate our method on a variety of real-world labeled datasets. Our results indicate that our method significantly outperforms state- of-the-art clustering methods for the task of clustering question-answer archives. 1 Introduction Community-based Question Answering (CQA) sys- tems such as Yahoo! Answers1, StackOverflow2 and Baidu Zhidao3 have become dependable sources of knowledge to solve common user problems. Unlike factoid question answering4, CQA systems focus on 1http://answers.yahoo.com 2http://www.stackoverflow.com 3http://en.wikipedia.org/en/Baidu Knows 4e.g., http://trec.nist.gov/data/qa.html crowdsourcing how and why questions and their an- swers. As is the case with any system where con- tent is generated by web users, the generated con- tent would be of varying quality, reliability, readabil- ity and abstraction. Thus, manual curation of such datasets is inevitable to weed out low quality and duplicate content to ensure user satisfaction. A nat- ural way to aid manual curation of such broad-based CQA archives is to employ clustering so that seman- tically related QAs are grouped together; this would help organize the corpus in a way that experts en- gaged in manual curation be assigned specific clus- ters relating to areas of their expertise. Cluster- ing also provides a platform to enable tagging the QA dataset; cluster topics could be used as tags, or other QAs in the same cluster could be tagged as being related to a QA. The fundamental difference between CQA archives and general text document collections is the existence of a two-part structure in QAs and the difference in lexical “character” be- tween the question and answer parts. This lexical chasm (i.e., gap) (Berger et al., 2000) between ques- tion and answer parts has been a subject of much study, especially, in the context of improving QA re- trieval. In this paper, we consider using the two-part structure in QAs for clustering CQA datasets. Motivating Example: Table 1 lists four example QAs from the context of a CQA system focused on addressing myriad technical issues. These QAs have been tagged in the table with a manually identified root-cause to aid understanding; the root-cause is not part of the CQA data per se. QA1 …

Content truncated to 500 words