TLDBERT: Leveraging Further Pre-Trained Model for Issue Typed Links Detection.

Issue links are crucial for promoting software information flowing and development efficiency, but manually conducting typed links detection (TLD) is time-consuming and error-prone due to the numerous candidate issues. The general pre-trained NLP models, such as BERT, provide promising automated approaches for TLD task after fine-tuning. In addition, the further pre-training method, which is an intermediate pre-training process on the in-domain corpora, has been proven to further improve the performance of pre-trained models on many downstream tasks. In this paper, we apply the further pre-training method on TLD task and assess the improved performance. We further pre-train BERT by utilizing the in-domain corpora constructed by Jira issues dataset, and then finetune it on the links dataset to obtain the more applicable model for TLD task - TLDBERT. The experimental results indicate the statistically significant improvement in the performance of TLDBERT, with a range of 0.4%-6.6% for accuracy and 0.1%-9.7% for macro F1-score. Finally, based on correlation analysis, we conclude that the coverage and the ratio of issues to users are significantly correlated with the improvement, although the two directions are opposite.