关键词:  大语言模型, 中文开放领域, 命名实体识别, 关系抽取, 提示学习

Abstract: Large Language Model (LLM) faces issues with unstable extraction performance in the task of entity-relation extraction within the Chinese open-domain, particularly showing lower accuracy in recognizing texts and annotated categories from certain specific fields. To address this, a Chinese open-domain entity-relation extraction strategy based on LLM, called Multi-level Dialog Strategies for Large Language Model（MLDS-LLM）, was proposed. First, the superior semantic understanding and transfer learning capabilities of LLM were leveraged to achieve entity-relation extraction through multi-turn dialogues across different tasks. Then, text summaries were generated by using the large language model based on the structured logic of open-domain texts and a chain-of-thought mechanism, thereby avoiding issues such as relational and factual hallucinations and the inability to consider subsequent information. Next, the limitations of the context window were reduced through a text simplification strategy and the introduction of a replaceable vocabulary. Finally, multi-level Prompt templates were constructed based on structured summaries and simplified texts, and explores the impact of the parameter 'temperature' on entity-relation extraction using the LLaMA-2-70B model. The Precision (P), Recall (R), F1 Score (F1), and Exact Match (EM) values of entity-relation extraction by LLaMA-2-70B were tested before and after applying the proposed strategy. Experimental results demonstrate that the proposed strategy enhances the performance of LLM in Named Entity Recognition （NER） and Relation Extraction (RE) across five different Chinese datasets: CL-NE-DS, DiaKG, CCKS2021, DulE, and IEPA. Particularly for the DiaKG and IEPA datasets, which are highly specialized and initially showed poor zero-shot test results, the accuracy of named entity recognition improved by 9.3% and 6.7% respectively compared to few-shot Prompt testing, with EM values increasing by 2.7% and 2.2%. The accuracy of relation extraction improved by 12.2% and 16.0%, and F1 scores increased by 10.7% and 10.0%, proving that the proposed strategy effectively enhances the performance of LLM in entity-relation extraction and resolves the issue of unstable model performance.

Key words: Large Language Model (LLM), Chinese open-domain, Named Entity Recognition (NER), Relation Extraction (RE), prompt learning