Zero-Shot Relation Triplet Extraction via Knowledge-Driven LLM Synthetic Data Generation

Large language models have demonstrated astonishing empirical success in the field of zero-shot relational triple extraction. However, using LLMs interface may cause ongoing overhead and privacy leakage issues, which is unacceptable in certain scenarios. On the other hand, a large language model may not always be necessary for relatively simple tasks. In this paper, we propose a method to leverage a LLM to supervise the training process of another language model, rather than using it directly as a extractor. Specifically, it synthesizes high-quality training data through knowledge-driven external large-scale models and utilizes ranking-based filtering methods to further augment the data. The proposed method tackle with the problem of low quality of synthetic data by reducing the noise that may affect the accuracy of triple extraction in certain relation categories. During the data generation process, no local data is uploaded to LLM. Privacy is thus guaranteed.