Enriching Topic Modelling with Users’ histories for Improving Tag Prediction in Q&A Systems

The automatic attribution of tags in Question & Answering (Q&A) systems like Stack Exchange can significantly reduce the human effort in tagging as well as improve the consistency among users. Existing approaches typically either rely on Natural Language Processing solely or employ collaborative filtering techniques. In this paper, we attempt to combine the best of both worlds by investigating whether incorporating a personal profile, consisting of a user's history or its social network can significantly improve the predictions of state-of-the-art text-based methods. Our research has found that enriching content-based text features with this personal profile allows to trade-off the precision of predictions for recall and as such improve the "exact match" (predicting the number of tags and the tags themselves correctly) in a multi-label setting from a baseline of 18.2% text-only to 54.3%.