Pipeline Chain-of-Thought: A Prompt Method for Large Language Model Relation Extraction
Hangtian Zhao, Hakiz Yilahun, Askar Hamdulla
Abstract
The development of language models has been influencing approaches to relation extraction (RE) problems. Although large language models (LLMs) have demonstrated breakthrough potential in certain aspects, they are still in the exploratory stage for RE tasks. Currently, the mainstream approach to improving the RE performance of LLMs is through prompt fine-tuning, but most methods require providing entity information in the prompt, which effectively only allows the LLMs to perform relationship classification tasks. We propose the Pipeline Chain-of-Thought (Pipeline-COT), which breaks down the RE task into steps and transforms it into reasoning tasks that have flat scaling curves, thereby enabling the use of Chain-of-Thought (COT) to enhance model inference. In addition, our method utilizes n-shot samples to provide signals for the Bayesian inference of the model by prompting the LLMs to focus on specific concepts to generate answers. We evaluated pipeline-COT on the Chinese dataset DuIE2.0, and compared with baseline methods that require including entity information in the prompt, our method still shows competitive performance.