Learning to detect text-code inconsistencies with weak and manual supervision

• Ano de defesa: 2023
• Autor(a) principal: SOUZA, Beatriz Bezerra de
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Pernambuco; UFPE; Brasil; Programa de Pos Graduacao em Ciencia da Computacao
• Palavras-chave em Português: Engenharia de software; Detecção de inconsistência
• Link de acesso: https://repositorio.ufpe.br/handle/123456789/49318

Source code often is associated with a natural language summary, enabling developers to understand the behavior and intent of the code. For example, method-level comments summarize the behavior of a method and test descriptions summarize the intent of a test case. Unfortunately, the text and its corresponding code sometimes are inconsistent, which may hinder code understanding, code reuse, and code maintenance. We propose TCID, an approach for Text-Code Inconsistency Detection, which trains a neural model to distinguish consistent from inconsistent text-code pairs. Our key contribution is to combine two ways of training such a model. First, TCID performs weakly supervised pre-training based on large amounts of consistent examples extracted from code as-is and inconsistent examples created by randomly recombining text-code pairs. Then, TCID fine-tunes the model based on a small and curated set of manually labeled examples. This combination is motivated by the observation that weak supervision alone leads to models that generalize poorly to real-world inconsistencies. Our evaluation applies the two-step training procedure to four state-of-the-art models and evaluates it on two text-vs-code problems: 40.7K method-level comments checked against the corresponding Java method body, and—as a problem not considered in prior work— 338.8K test case descriptions checked against corresponding JavaScript implementations. Our results show that a small amount of manual labeling enables the approach to significantly improve effectiveness, outperforming the current state of the art and improving the F1 score by 5% in Java and by 17% in JavaScript. We validate the usefulness of TCID’s predictions by submitting pull requests, of which 10 have been accepted so far.