Our previous work has introduced the newly generated program using the code transformation model GPT-2,verifying the generated programming codes through simhash(SH)and longest common subsequence(LCS)algo-rithms.Howeve...Our previous work has introduced the newly generated program using the code transformation model GPT-2,verifying the generated programming codes through simhash(SH)and longest common subsequence(LCS)algo-rithms.However,the entire code transformation process has encountered a time-consuming problem.Therefore,the objective of this study is to speed up the code transformation process signicantly.This paper has proposed deep learning approaches for modifying SH using a variational simhash(VSH)algorithm and replacing LCS with a piecewise longest common subsequence(PLCS)algorithm to faster the verication process in the test phase.Besides the code transformation model GPT-2,this study has also introduced MicrosoMASS and Facebook BART for a comparative analysis of their performance.Meanwhile,the explainable AI technique using local interpretable model-agnostic explanations(LIME)can also interpret the decision-making ofAImodels.The experimental results show that VSH can reduce the number of qualied programs by 22.11%,and PLCS can reduce the execution time of selected pocket programs by 32.39%.As a result,the proposed approaches can signicantly speed up the entire code transformation process by 1.38 times on average compared with our previous work.展开更多
基金supported by the Ministry of Science and Technology,Taiwan,under Grant Nos.MOST 111-2221-E-390-012 and MOST 111-2622-E-390-001.
文摘Our previous work has introduced the newly generated program using the code transformation model GPT-2,verifying the generated programming codes through simhash(SH)and longest common subsequence(LCS)algo-rithms.However,the entire code transformation process has encountered a time-consuming problem.Therefore,the objective of this study is to speed up the code transformation process signicantly.This paper has proposed deep learning approaches for modifying SH using a variational simhash(VSH)algorithm and replacing LCS with a piecewise longest common subsequence(PLCS)algorithm to faster the verication process in the test phase.Besides the code transformation model GPT-2,this study has also introduced MicrosoMASS and Facebook BART for a comparative analysis of their performance.Meanwhile,the explainable AI technique using local interpretable model-agnostic explanations(LIME)can also interpret the decision-making ofAImodels.The experimental results show that VSH can reduce the number of qualied programs by 22.11%,and PLCS can reduce the execution time of selected pocket programs by 32.39%.As a result,the proposed approaches can signicantly speed up the entire code transformation process by 1.38 times on average compared with our previous work.
