Semantically enhanced tag recommendation for software CQAs via deep learning

Abstract

Most software CQAs (e.g. Stack Overflow) mainly rely on users to assign tags for posted questions. This leads to many redundant, inconsistent and inaccurate tags that are detrimental to the communities. Therefore tag quality becomes a critical challenge to deal with. In this work, we propose STR, a deep learning based approach that automatically recommends tags through learning the semantics of both tags and questions in such software CQAs. First, word embedding is employed to convert text information to high-dimension vectors for better representing questions and tags. Second, a Multi-tasking-like Convolutional Neural Network, the core modules of STR, is designed to capture short and long semantics. Third, the learned semantic vectors are fed into a gradient descent based algorithm for classification. Finally, we evaluate STR on three datasets collected from popular software CQAs, and experimental results show that STR outperforms state-of-the-art approaches in terms of Precision@k, Recall@k and F1 - Measure@k.