Design Extractor: A ML-based Tool for CapturingSoftware Design Decisions

Abstract: Context: A software project’s success; involvinga larger group of individuals, relies on efficient teamcommunication. Part of efficient communication is avoidingmiscommunication, misunderstandings, and losingknowledge. These consequences of poor communication canlead to negative repercussions such as loss of time, money,and customer approval. Much effort has been put intocreating tools and systems to aid software engineers inretaining knowledge and decisions made during meetings,but many existing solutions require additional manualintervention on the part of software meeting participants.The objective of this thesis is to explore and develop a toolcalled Design Extractor (DE) which creates concisesummaries of design meetings from recorded voiceconversations. These summaries include both the designdecisions made during a meeting as well as the rationalebehind them. This thesis used readily available Pythonframeworks for machine learning to train two transformermodels based on DistilBert and Google’s BERT. Fine-tuningthese models with data sourcedfrom six different softwaredesign meetings found that the best base model wasDistilBert, which resulted in a fine-tuned model reporting anF1 score of 82.63%. This study created a simple Python tool,built upon many publicly available Python frameworks andthe fine-tuned transformer model, that takes in voicerecordings and outputs labeled sentence-label pairs that canbe used to quickly notate a design meeting. Short summariesare also provided by the tool through the use of pre-existingtext summarisation machine learning models such as BART.Design extractor therefore provides a simple quick way toreview longer meeting recordings in the context of softwareengineering decisions.