UKT: A Unified Knowledgeable Tuning Framework for Chinese Information Extraction

Jiyong Zhou, Chengyu Wang, Junbing Yan, Jianing Wang, Yukang Xie, Jun Huang, Ying Gao

Abstract: Large Language Models (LLMs) have significantly improved the performance of various NLP tasks. Yet, for Chinese Information Extraction (IE), LLMs can perform poorly due to the lack of fine-grained linguistic and semantic knowledge. In this paper, we propose Unified Knowledgeable Tuning (UKT), a lightweight yet effective framework that is applicable to several recently proposed Chinese IE models based on Transformer. In UKT, both linguistic and semantic knowledge is incorporated into word representations. We further propose the relational knowledge validation technique in UKT to force model to learn the injected knowledge to increase its generalization ability. We evaluate our UKT on five public datasets related to two major Chinese IE tasks. Experiments confirm the effectiveness and universality of our approach, which achieves consistent improvement over state-of-the-art models.