Abstract

Community Question Answering (cQA) sites have emerged as platforms designed specifically for the exchange of questions and answers among communities of users. Although users tend to find good quality answers in cQA sites, there is evidence that they also engage in a significant volume of QA in other types of social sites, such as microblog platforms. Research indicates that users opt for these non-specific QA social networks because they contain up-to-date information on current events, also due to their rapid information propagation, and social trust. In this sense, we propose that microblog platforms can emerge as a novel, valuable source of information for QA information retrieval tasks. However, we have found that it is not straightforward to transfer existing approaches for automatically retrieving relevant answers in traditional cQA platforms for use in microblogs. This occurs because there are unique characteristics that differentiate microblog data from that of traditional cQA, such as noise and very short text length. In this work, we study (1) if microblog data can be used to automatically provide relevant answers for the QA task, and, in addition, (2) which features contribute the most for finding relevant answers for a particular query. In particular, we introduce a conversation (thread)-level document model, as well as a machine learning ranking framework for microblog QA. We validate our proposal by using factoid-QA as a proxy task, showing that Twitter conversations can indeed be used to automatically provide relevant results for QA. We are able to identify the importance of different features that contribute the most for QA ranking. In addition, we provide evidence that our method allows us to retrieve complex answers in the domain of non-factoid questions. (C) 2019 Elsevier B.V. All rights reserved.