Co-REX: A Collaborative Transformer for Try Location Detection and Catch Exception Prediction

Try Location Detection (TLD) and Catch Exception Prediction (CEP) are two interrelated tasks in runtime exception handling which is a crucial aspect in Java programming.The performance of each task can significantly benefit from the information provided by the other.However, previous methods primarily treated these tasks in isolation or addressed them separately using multi-task frameworks, which overlooked their interconnected information flow.In this work, we propose a Collaborative transformer model for Runtime EXception detector (Co-REX), which establishes a bidirectional information flow between TLD and CEP and enables mutual enhancement through joint learning.Additionally, to evaluate the performance of Co-REX, we collect a large number of Java methods from Gitee and GitHub to create a benchmark dataset.Our experimental results demonstrate that Co-REX achieves 78.1% accuracy outperforming other state-of-the-art methods by a margin of at least 2.1% for TLD.It also performs better by 3.1% improvement in terms of Top-1 accuracy than others for CEP.