Evaluating and predicting answer quality in community QA

Abstract

Question answering (QA) helps one go beyond traditional keywords-based querying and retrieve information in more precise form than given by a document or a list of documents. Several community-based QA (CQA) services have emerged allowing information seekers pose their information need as questions and receive answers from their fellow users. A question may receive multiple answers from multiple users and the asker or the community can choose the best answer. While the asker can thus indicate if he was satisfied with the information he received, there is no clear way of evaluating the quality of that information. We present a study to evaluate and predict the quality of an answer in a CQA setting. We chose Yahoo! Answers as such CQA service and selected a small set of questions, each with at least five answers. We asked Amazon Mechanical Turk workers to rate the quality of each answer for a given question based on 13 different criteria. Each answer was rated by five different workers. We then matched their assessments with the actual asker’s rating of a given answer. We show that the quality criteria we used faithfully match with asker’s perception of a quality answer. We furthered our investigation by extracting various features from questions, answers, and the users who posted them, and training a number of classifiers to select the best answer using those features. We demonstrate a high predictability of our trained models along with the relative merits of each of the features for such prediction. These models support our argument that in case of CQA, contextual information such as a user’s profile, can be critical in evaluating and predicting content quality.