Improving answer quality using image-text coherence on social Q&A sites

There has been a significant rise in the use of social Q&A to get answers to a variety of queries. One common problem faced by most social Q&A is how to help unskillful answerers construct well-received answers. Prior studies in answer quality assessment usually focus on ranking candidate answers for the sake of askers but show little value for the answerers. Moreover, existing work employs textural aspects and image quantity to predict answer quality, but semantic information inherent in answer images is rarely considered. To bridge the research gap, we designed an artifact, answer advisor (AA), to help answerers produce well-received answers. Our AA uses an image-text coherence measure that is obtained by integrating topic modeling with a deep learning approach. On a real-world dataset, the proposed measure can reduce the prediction error of answer popularity before the answer is actually posted on the Q&A site by 38.12%.