Tags’ Recommender to Classify Architectural Knowledge Applying Language Models

Abstract

Agile global software engineering challenges architectural knowledge (AK) management since face-to-face interactions are preferred over comprehensive documentation, which causes AK loss over time. The AK condensation concept was proposed to reduce AK losing, using the AK shared through unstructured electronic media. A crucial part of this concept is a classification mechanism to ease AK recovery in the future. We developed a Slack complement as a classification mechanism based on social tagging, which recommends tags according to a chat/message topic, using natural language processing (NLP) techniques. We evaluated two tagging modes: NLP-assisted versus alphabetical auto-completion, in terms of correctness and time to select a tag. Fifty-two participants used the complement emulating an agile and global scenario and gave us their complement’s perceptions about usefulness, ease of use, and work integration. Messages tagged through NLP recommendations showed fewer semantic errors, and participants spent less time selecting a tag. They perceived the component as very usable, useful, and easy to be integrated into the daily work. These results indicated that a tag recommendation system is necessary to classify the shared AK accurately and quickly. We will improve the NLP techniques to evaluate AK condensation in a long-term test as future work.