Knowledge Base Enhanced & User-centric Dialogue Design for OTF Computing

This thesis aims to provide a bidirectional chatbot solution for the requirement engineering process. The Sonderforschungsbereich (SFB) 901 intends to provide the composition of software service On-the-Fly (OTF). The sub-project (B1) of the SFB 901 project deals with the parameters of service configuration. OTF Computing aims to eradicate the dependency on the requirement engineers for the software development process. However, there is no existing bidirectional chatbot solution that analyses user software requirements and provides viable suggestions to the user regarding their service. Previously, CORDULA chatbot was developed to analyze the software requirements but cannot keep the conversation’s context. The Rasa framework is integrated with the knowledge base to solve the issue, the knowledge base provides domain-specific knowledge to the chatbot. The software description is passed through the natural language understanding process to give consciousness to the chatbot. This process involves various machine learning models, including app family classification, to correctly identify the domain for user OTF service. The statistical models like naïve Bayes, kNN and SVM are compared with transformer models for this classification task. Furthermore, the entities (functional requirements) are also separated from the user description. The chatbot provides the suggestion of requirements from the preliminary service template with the support of the knowledge base. Furthermore, the generated response is compared with the state-of-the-art DialoGPT transformer model and ChatterBot conversational library. These models are trained over the software development related conversational dataset. All the responses are ranked using the DialoRPT model, and the BLEU score to evaluates the models’ responses. Moreover, the chatbot mod- els are tested with human participants, they used and scored the chatbot responses based on effectiveness, efficiency and satisfaction. The overall response accuracy is also measured by averaging the user approval over the generated responses.