Community and Thread Methods for Identifying Best Answers in Online Question Answering Communities

Much research has recently investigated the measurement of quality answers in Question Answering (Q&A) communities in the form of automatic best answer identification. Previous approaches have focused on manual user annotations and diverse features based on intuition for identifying best answers and proved relatively successful despite considering best answer identification as a general classification problem. Best answer modelling is generally distanced from community studies about what users regard as important for identifying quality content. In particular, previous research tends to only focus on the automatic aspects of best answers identification model by applying generic learning algorithms. This thesis introduces the concepts of qualitative and structural design in order to investigate if features derived from community questionnaires can enrich the understanding of best answer identification in Q&A communities and if the thread-like structure of Q&A communities can be exploited for better results. Two different approaches for exploiting the thread structure of Q&A communities are proposed and two new, previously unstudied, features are introduced. First, a measure of question complexity is introduced as a proxy measure of answerer knowledge. Second, different models of contribution effort are proposed for representing the answering reactivity of contributors. The experiments are systematically conducted on datasets issued from three different communities that vary in size, content and structure. The results show that the newly proposed features allow for better understanding of what constitute best answers. The findings also reveal that the thread-wise algorithms and optimisation techniques created from the structural design methodology correlate with best answers. In general both structural and qualitative design appear to improve best answer identification meaning that structural and qualitative methods may improve unrelated classification tasks