ElektroCHAT: A Knowledge Base-Driven Dialogue System for Electrical Engineering Students : A Proposal for Interactive Tutoring

Abstract: Universities worldwide face challenges both with students dropping out of educational programmes and repetitive questions directed toward teaching staff which both consume resources and result in delays. Recent progress in natural language processing (NLP) introduces the possibility of more sophisticated dialogue systems that could help alleviate the situation. Dialogue systems in education are complex to construct for multiple reasons. Two such reasons are that domain-specific data is often not readily available and extending an existing system often requires configuring the system again and re-training models. In this thesis, a graph-based knowledge base (KB) which is the foundation of a heavily rule-based dialogue system is proposed. The core of the natural language understanding (NLU) in the pipeline-based dialogue system includes the transformer-based DIET classifier for intent classification and entity extraction. The custom logic of the dialogue system relies on contextual and distributional embeddings. While the proposed solution is used in electrical engineering specifically, the KB and the architecture of the dialogue system are designed with generalization in mind. An emphasis is to maintain a low level of system maintenance after deployment allowing teaching staff without expertise in computer science and machine learning to operate the system. The utilization of transfer learning with pre-trained language models helps achieve this objective. The findings suggest that the system is sufficiently sophisticated to improve learning environments for students while potentially alleviating the workload of teaching staff. They further indicate that computer science and machine learning expertise are not required to operate the system over time.