Logic Error Detection Algorithm Based on RNN with Threshold Selection

Logical errors in source code can be detected by probabilities obtained from a language model trained by the recurrent neural network (RNN). Using the probabilities and determining thresholds, places that are likely to be logic errors can be enumerated. However, when the threshold is set inappropriately, user may miss true logical errors because of passive extraction or unnecessary elements obtained from excessive extraction. Moreover, the probabilities of output from the language model are different for each task, so the threshold should be selected properly. In this paper, we propose a logic error detection algorithm using an RNN and an automatic threshold determination method. The proposed method selects thresholds using incorrect codes and can enhance the detection performance of the trained language model. For evaluating the proposed method, experiments with data from an online judge system, which is one of the educational systems that provide the automated judge for many programming tasks, are conducted. The experimental results show that the selected thresholds can be used to improve the logic error detection performance of the trained language model.