Research And System Implementation Of Software Bug-Oriented Question Answering Technology

During the bug fixing process,developers often need to exploit historical related bug information form the software bug repository or other bug resources to support various bug analysis activities,i.e.,bug prediction,bug localization,bug root cause prediction,in order to better to complete the task of the bug fixing.Bug text contains rich semantic information,and effectively acquiring and harnessing the bug information is useful for bug comprehension.However,traditional keyword matching technology is often used to obtain relevant bug information in the current bug research field.The feedback information is often not related to the bug information required by users,and even contains invalid information that just matches a certain word or character.The native search engines of these software resource platforms have gradually been unable to return the answers that satisfy the developers'retrieval requirements.With the emerging natural language question answering techniques,they provide us an effective and direct way to attain the required bug information.In addition,the form of bug text information itself is not uniform,and unstructured and semi-structured data are mixed,which brings great obstacles to the task of bug understanding.In view of the above problems,this thesis conducts studies on the software bug question answering technology from the perspective of bug comprehension.First,we propose a bug question answering approach based on the structured templates.Entity and entity relations are extracted from bug data to form the SPARQL template,and then we process natural language bug question answering into bug question answering based on templates matching.Then,the deep learning model is used to transform the bug question answering problem into a task of bug natural language reading comprehension,and bug question answering with pretrained encoders is proposed.The main work is as follows:(1)A bug question answering approach with structured templates is proposed for software bug domain.First,we define the basic concepts and introduce the data preparation in this work.Then,we introduce the construction process of structured templates and how to use these structured templates for bug question answering.Finally,we conduct an empirical study on the bug data of two projects in the Bugzilla project management repository,i.e.,Mozilla and Eclipse.We perform an experimental comparison with the existing question answering approaches,and the comparison results indicate that the bug question answering is better than the state-of-the-art Q&A approaches over the Mozilla and Eclipse projects.(2)A bug question answering approach with pretrained encoders is proposed.As there is no public bug dataset available for bug question answering in the field of bug studies,we exploit the public accessible open source datasets to train the BERT model,and then construct the related bug reading comprehension dataset to fine-tune this model.In this work,we transform the bug question answering problem into a task of bug natural language reading comprehension,and design a common paradigm to construct the bug reading comprehension dataset for this approach.The empirical study results indicate that the construction of the bug reading comprehension dataset can indeed improve the performance of bug question answering,and the bug question answering is better than the state-of-the-art Q&A approaches over the Mozilla and Eclipse projects.(3)We design and implement the bug-oriented question answering platform.The platform is designed for developers and other relevant researchers in the bug domain.The platform includes four modules of bug factual triple extraction,bug structured template generation,bug query reformation and bug question answering.Users can effectively obtain and understand the bug data in the form of natural question answering,so as to achieve the purpose of accelerating the bug fixing process.