TagDeepRec: Tag Recommendation for Software Information Sites Using Attention-Based Bi-LSTM

Abstract

Software information sites are widely used to help developers to share and communicate their knowledge. Tags in these sites play an important role in facilitating information classification and organization. However, the insufficient understanding of software objects and the lack of relevant knowledge among developers may lead to incorrect tags. Thus, the automatic tag recommendation technique has been proposed. However, tag explosion and tag synonym are two major factors that affect the quality of tag recommendation. Prior studies have found that deep learning techniques are effective for mining software information sites. Inspired by recent deep learning researches, we propose TagDeepRec, a new tag recommendation approach for software information sites using attention-based Bi-LSTM. The attention-based Bi-LSTM model has the advantage of deep potential semantics mining, which can accurately infer tags for new software objects by learning the relationships between historical software objects and their corresponding tags. Given a new software object, TagDeepRec is able to compute the confidence probability of each tag and then recommend top-k tags by ranking the probabilities. We use the dataset from six software information sites with different scales to evaluate our proposed TagDeepRec. The experimental results show that TagDeepRec has achieved better performance compared with the state-of-the-art approaches TagMulRec and FastTagRec in terms of Recall@k, Precision@k and $F1-score@k$.