Generative AI-Driven Knowledge Base Question Answering with Differentiable Knowledge Graphs for Sustainable Engineering Education

Knowledge Base Question Answering (KBQA) is a challenging and widely studied research direction. Approaches that rely on embedding use implicit reasoning to obtain answers, but they are not well suited to producing complete, transparent reasoning paths. Models built upon differentiable knowledge graphs, by contrast, can deliver explainable outputs with minimal supervision signals, requiring only question-answer pairs. Building on differentiable knowledge graphs, we propose an encoder-decoder model called EDKBQA. The encoder fine-grainedly and sequentially models questions using a multi-head attention mechanism and an LSTM to generate query vectors that better capture the semantic properties of each reasoning hop. The decoder incorporates a feedforward neural network with an attention mechanism for multi-hop reasoning, effectively representing the relative importance of each hop when determining the final answer. The proposed model directly addresses information loss arising from coarse-grained, non-sequential modelling. Experiments on the MetaQA-1hop, MetaQA-2hop, MetaQA-3hop, WebQSP, and CWQ datasets achieve accuracies of 97.5%, 100%, 100%, 77.8%, and 51.4%, respectively. Ablation studies confirm that every module contributes to overall performance. Furthermore, the paper demonstrates the model's practical utility through a case study in sustainable engineering education, showcasing its ability to provide transparent, multi-hop explanations for complex engineering concepts.