Distilled Deep Prediction for Community QA

Community Question Answer (cQA) platforms such as Reddit, Stack Overflow and Quora are helpful to find relevant and timely information about several issues. They provide users the flexibility of asking open-ended questions to a large audience, and replies to such questions may be useful both to the user and the community. However, sometimes users must scour through countless repetitive questions and unrelated answers on cQA to find accurate and appropriate responses to their inquiries. In this paper, a community question and answer relevancy ranking system has been developed via advanced deep machine-learning technology motivated by the feature learning capacity of deep neural networks. With its ability to accurately retrieve the most relevant answers to a user given a question, our model greatly assists users with much more ease. It shows superior performance over baseline based on extensive offline experiments on two different datasets. One is the Yahoo Answers [1] and another is the COVID-19 Open Research Dataset [2]. Our model significantly improves the accuracy and effectiveness of community-based question and answer platforms. Furthermore, in order to meet much more stringent latency and computation requirements for real-time usage, a much more light-weighted model is needed, for which we applied knowledge distillation and developed a novel pipeline for model compression. Results indicate that our distilled deep learning model can perform as well as the original one since the knowledge from the big model has been successfully transferred. The distilled deep learning model has a significant lift on AUC compared with baseline, however it is only around 16 % of the size of the original heavy one.