Knowledge Base Question Generation via Data Augmentation with Dynamic-Prompt

Knowledge Base Question Generation (KBQG) aims to generate natural language questions according to a subgraph and the target answer(s). Existing methods consume significant computational resources for pre-training or fine-tuning multiple models, which are often is impractical in reality. To address it, we propose a KBQG method via data augmentation with dynamic-prompt (DADP), which leverages the semantic similarity of relational paths across different subgraphs. We prompt the large language model (LLM) to generate questions for data augmentation by utilizing our dynamic-prompt approach. Depending on the quality of the generations, we design different level tokens as prefixes for fine-tuning a pre-trained language model (PLM), enabling it to learn the characteristics of questions with different quality. During the inference stage, we incorporate the highest level token as prefix to inspire the PLM to generate high-quality questions. Remarkably, experimental results on two widely used benchmarks for KBQG task demonstrate that our method achieves new state-of-the-art (SOTA) performance. Furthermore, DADP can be served as a data augmentation approach for knowledge base question answering (KBQA) task, enhancing the performance of various methods in this domain.