Predicting response probability by embedding questions in online question recommendation

Abstract

Currently, there are many Q&A sites, including Yahoo! Answers, Quora, and StackOverflow. Although the number of questions posted on these sites is enormous, many remain unanswered. This is detrimental to the user experience, so service operators are motivated to obtain more answers to posted questions. It is also diffcult for users to find specific questions they can answer among the vast number that are asked. Therefore, a system that recommends questions that can be answered by the user is needed. In this study, we first propose a method for predicting response probability. Specifically, we propose a method for learning embedding vectors that takes into account cases in which the required answers are similar, even if the question texts are different, based on a contrastive learning method. We also implemented a recommendation method that increases respondent satisfaction by optimizing and analyzing the theoretical properties of our method. Finally, we conduct an experimental validation of these two methods to demonstrate their effectiveness using data from a Q&A service for child care.