Group Non-negative Matrix Factorization with Natural Categories for Question Retrieval in Community Question Answer Archives

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 89–98, Dublin, Ireland, August 23-29 2014. Group Non-negative Matrix Factorization with Natural Categories for Question Retrieval in Community Question Answer Archives Guangyou Zhou, Yubo Chen, Daojian Zeng, and Jun Zhao National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences 95 Zhongguancun East Road, Beijing 100190, China {gyzhou,yubo.chen,djzeng,jzhao}@nlpr.ia.ac.cn Abstract Community question answering (CQA) has become an important service due to the popularity of CQA archives on the web. A distinctive feature is that CQA services usually organize questions into a hierarchy of natural categories. In this paper, we focus on the problem of question re- trieval and propose a novel approach, called group non-negative matrix factorization with natural categories (GNMFNC). This is achieved by learning the category-specific topics for each cate- gory as well as shared topics across all categories via a group non-negative matrix factorization framework. We derive an efficient algorithm for learning the factorization, analyze its complex- ity, and provide proof of convergence. Experiments are carried out on a real world CQA data set from Yahoo! Answers. The results show that our proposed approach significantly outperforms various baseline methods and achieves the state-of-the-art performance for question retrieval. 1 Introduction Community question answering (CQA) such as Yahoo! Answers1 and Quora2, has become an important service due to the popularity of CQA archives on the web. To make use of the large-scale questions and their answers, it is critical to have functionality of helping users to retrieve previous answers (Duan et al., 2008). Typically, such functionality is achieved by first retrieving the historical questions that best match a user’s queried question, and then using answers of these returned questions to answer the queried question. This is what we called question retrieval in this paper. The major challenge for question retrieval, as for most information retrieval tasks, is the lexical gap between the queried questions and the historical questions in the archives. For example, if a queried ques- tion contains the word “company” but a relevant historical question instead contains the word “firm”, then there is a mismatch and the historical question may not be easily distinguished from an irrelevant one. To solve the lexical gap problem, most researchers focused on translation-based approaches since the relationships between words (or phrases) can be explicitly modeled through word-to-word (or phrases) translation probabilities (Jeon et al., 2005; Riezler et al., 2007; Xue et al., 2008; Lee et al., 2008; Bern- hard and Gurevych, 2009; Zhou et al., 2011; Singh, 2012). However, these existing methods model the relevance ranking without considering the category-specific and shared topics with natural categories, it is not clear whether this information is useful for question retrieval. A distinctive feature of question-answer pairs in CQA is that CQA services usually organize questions into a hierarchy of natural categories. For example, Yahoo! Answers contains a hierarchy of 26 categories at the first level and more than 1262 subcategories at the leaf level. When a user asks …

Content truncated to 500 words