On prediction of the answer affordance with the vocabulary power in CQA systems

Abstract

Research into routing questions to appropriate answerers has recently attracted a lot of attention. The answerers are required to have expertise in the question area as well as availability to answer the question. However, when choosing proper answerers through a data set, there is a trade-off relationship between those two features. An expert is usually identified by past abundant answer data while the expert is possibly not available at the present moment. Thus, finding experts in an early phase when they are still active is essential to improve the chances of getting answers back. In this thesis, we propose a framework to predict the expertise level and availability of new answerers who only possess a small amount of data, which is a cold start problem. Co-occurrence of vocabulary in a similar category between answerers, irrespective of time, is a key to solving this problem. Experts can disclose their expertise using their vocabulary in a tiny set of content. Their expertise along with their recent behavior is reflected in the Answer Affordance defined in this thesis. Extensive experiments are conducted in two categories with a large data set of Naver Knowledge-In, the top CQA service in Korea. First, we verify the expertise of the answerers ranked at the top by our algorithm. Then, we show how such users will play important roles in the future by evaluating their expertise and availability, with the scores of the Answer Affordance. As a result, the new answerers with high Answer Affordance scores clearly show not only their expertise but also their real activity.