Novelty Detection in Community Question Answering Forums

Novelty Detection in Community Question Answering Forums Tirthankar Ghosal ´UFAL, MFF Charles University, Czech Republic ghosal@ufal.mff.cuni.cz Vignesh Edithal IIT Patna Patna, India edithal.cs14@iitp.ac.in Tanik Saikh IIT Patna Patna, India 1821CS08@iitp.ac.in Saprativa Bhattacharjee Government Polytechnic Daman Daman, India saprativa.bhatt@gov.in Asif Ekbal IIT Patna Patna, India asif@iitp.ac.in Pushpak Bhattacharyya IIT Bombay Mumbai, India pb@cse.iitb.ac.in Abstract Community Question Answering (CQA) fo- rums are a popular place for open-ended question-answers and discussions by the pub- lic. Popular question answering sites have become a one-stop solution for the diverse information-seeking needs of the netizens. However, it is quite common that users some- times ask the same question which may have been posted before and appropriately an- swered by the community members. It is not uncommon for users to perform an im- proper search and pose the same question again. Community members start responding to that question; however, the answer may al- ready have been proposed previously on a dif- ferent thread. When two questions are get- ting at the same problem, community mem- bers (moderators) in the forum often flag one as a duplicate of the other to help route traf- fic to high-quality novel questions and their correct answers. However, they usually do this manually based on the novelty or redun- dancy of the question. In this work, we try to mitigate this problem of detecting semanti- cally equivalent non-novel questions automat- ically and flagging those. We also propose an approach to identify novel questions from CQA forums so that only the novel questions and corresponding answers threads stay, and semantic duplicates are removed. We make use of a Dynamic Memory Network (DMN) to assimilate information from multiple source questions to answer whether a new question is novel or a semantically equivalent question already exists. We introduce a new dataset for semantic-level novelty detection on com- munity question answering. Our proposed ap- proach attains performance improvements of +6.64% in terms of accuracy and +9.09% in terms of averaged F1 score over recent textual novelty detection methods. We would make our newly created dataset and the proposed approach available at https://github. com/edithal-14/DMN-Novelty. 1 Introduction Community Question Answering forums are a con- venient source of information for web users. Users post their questions on the forum, and fellow com- munity members help them with pointers or answers to their diverse information needs. However, with the rapid growth of content within those websites, redundancy has become a problem. Unaware of existing solutions, users pose questions on a new thread that may already have been resolved in an ex- isting thread. Such activity is widespread, and un- aware community members start interacting on the new thread with solutions/pointers that may have al- ready been discussed on an older thread. Usually, CQA forums have moderators who flag such non- novel questions and route the users to the existing solutions on the forum. To channel traffic to high quality and novel questions and answers on the fo- rum, it is important to weed out redundancies and non-novelties. Most of the …

Content truncated to 500 words