文献类型： 外文期刊

作者： Hu, Haotian ¹ ; Yang, Alex Jie ² ; Deng, Sanhong ² ; Wang, Dongbo ⁴ ; Song, Min ⁵ ;

作者机构： 1.Jiangsu Acad Agr Sci, Nanjing 210014, Peoples R China

2.Nanjing Univ, Sch Informat Management, Nanjing 210023, Peoples R China

3.Nanjing Univ, Int Joint Informat Lab, Nanjing 210023, Peoples R China

4.Nanjing Agr Univ, Sch Informat Management, Nanjing 210095, Peoples R China

5.Yonsei Univ, Dept Lib & Informat Sci, Seoul 03722, South Korea

关键词： Instruction tuning; Large language model; Chain-of-thought; Drug-drug interaction; Triplet extraction; Biomedical information extraction

期刊名称：EXPERT SYSTEMS WITH APPLICATIONS （ 影响因子：7.5； 五年影响因子：7.8 ）

ISSN： 0957-4174

年卷期： 2025 年 273 卷

页码：

收录情况： SCI

摘要： Current state-of-the-art drug-drug interaction (DDI) triplet extraction methods not only fails to exhaustively capture potential overlapping entity relations but also grapples to extract discontinuous drug entities, leading to suboptimal performance in DDI triplet extraction. To address these challenges, we proposed a Chain-of-Thought Enhanced Large Language Model for DDI Triplet Extraction (CoTEL-D3X). Based on the transformer architecture, we designed joint and pipeline methods that can perform end-to-end DDI triplet extraction in a generative manner. Our proposed approach builds upon the novel LLaMA series model as the foundation model and incorporates instruction tuning and Chain-of-Thought techniques to enhance the model's understanding of task requirements and reasoning capabilities. We validated the effectiveness of our methods on the widely-used DDI dataset, which comprises 1025 documents containing 17,805 entity mentions and 4,999 DDIs. Our joint and pipeline methods not only outperformed mainstream generative models, such as ChatGPT, GPT-3, and OPT, on the DDI Extraction 2013 dataset but also improved the current corresponding best F1-score by 9.75% and 5.86%, respectively. Particularly, compared to the currently most advanced few-shot learning methods, our approach achieved more than a two-fold improvement in F1-score. We further validated the method's transferability and generalization performance on the TAC 2018 DDI Extraction and ADR Extraction datasets, and assessed its applicability on real-world data from DrugBank. Performance analysis of the proposed method revealed that the CoT component significantly enhanced the extraction effect. The introduction of generative LLMs allows us to freely define the content and format of inputs and outputs, offering superior usability and flexibility compared to traditional extraction methods based on sequence labeling. Furthermore, as our proposed approach does not rely on external knowledge or manually defined rules, it may lack domain-specific knowledge to some extent. However, it can easily be adapted to other domains.