Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EntGPT: Entity Linking with Generative Large Language Models

Yifan Ding, Amrit Poudel, Qingkai Zeng, Tim Weninger, Balaji Veeramani, Sanmitra Bhattacharya

TL;DR

EntGPT addresses entity linking by grounding mentions to knowledge base entries using two large-language-model–driven strategies: a three-step prompt-based pipeline (EntGPT-P) and instruction-tuning (EntGPT-I). EntGPT-P achieves substantial zero-shot gains (up to micro-$F_1$ improvement of 36%), while EntGPT-I delivers consistent supervised gains (average micro-$F_1$ +2.1%) and state-of-the-art zero-shot QA across six benchmarks. The methods demonstrate robust performance across ten ED datasets with Wikipedia as the KB, without supervised fine-tuning, and the authors release data and code for reproducibility. This work highlights the value of explicit knowledge grounding and task-specific instruction tuning in improving ED and related QA tasks, with practical impact for scalable knowledge-grounded NLP systems.

Abstract

Entity Linking in natural language processing seeks to match text entities to their corresponding entries in a dictionary or knowledge base. Traditional approaches rely on contextual models, which can be complex, hard to train, and have limited transferability across different domains. Generative large language models like GPT offer a promising alternative but often underperform with naive prompts. In this study, we introduce EntGPT, employing advanced prompt engineering to enhance EL tasks. Our three-step hard-prompting method (EntGPT-P) significantly boosts the micro-F_1 score by up to 36% over vanilla prompts, achieving competitive performance across 10 datasets without supervised fine-tuning. Additionally, our instruction tuning method (EntGPT-I) improves micro-F_1 scores by 2.1% on average in supervised EL tasks and outperforms several baseline models in six Question Answering tasks. Our methods are compatible with both open-source and proprietary LLMs. All data and code are available on GitHub at https://github.com/yifding/In_Context_EL.

EntGPT: Entity Linking with Generative Large Language Models

TL;DR

EntGPT addresses entity linking by grounding mentions to knowledge base entries using two large-language-model–driven strategies: a three-step prompt-based pipeline (EntGPT-P) and instruction-tuning (EntGPT-I). EntGPT-P achieves substantial zero-shot gains (up to micro- improvement of 36%), while EntGPT-I delivers consistent supervised gains (average micro- +2.1%) and state-of-the-art zero-shot QA across six benchmarks. The methods demonstrate robust performance across ten ED datasets with Wikipedia as the KB, without supervised fine-tuning, and the authors release data and code for reproducibility. This work highlights the value of explicit knowledge grounding and task-specific instruction tuning in improving ED and related QA tasks, with practical impact for scalable knowledge-grounded NLP systems.

Abstract

Entity Linking in natural language processing seeks to match text entities to their corresponding entries in a dictionary or knowledge base. Traditional approaches rely on contextual models, which can be complex, hard to train, and have limited transferability across different domains. Generative large language models like GPT offer a promising alternative but often underperform with naive prompts. In this study, we introduce EntGPT, employing advanced prompt engineering to enhance EL tasks. Our three-step hard-prompting method (EntGPT-P) significantly boosts the micro-F_1 score by up to 36% over vanilla prompts, achieving competitive performance across 10 datasets without supervised fine-tuning. Additionally, our instruction tuning method (EntGPT-I) improves micro-F_1 scores by 2.1% on average in supervised EL tasks and outperforms several baseline models in six Question Answering tasks. Our methods are compatible with both open-source and proprietary LLMs. All data and code are available on GitHub at https://github.com/yifding/In_Context_EL.
Paper Structure (20 sections, 1 figure, 7 tables)

This paper contains 20 sections, 1 figure, 7 tables.

Table of Contents

  1. Introduction
  2. EntGPT
  3. EntGPT-P: Entity Disambiguation with Multi-Step Prompting
  4. Step 1: Entity Candidate Generation
  5. Step 2: Augmentation by Prompting
  6. Step 3: Multiple-choice Selection by Prompting
  7. EntGPT-I: Entity Disambiguation with Instruction Tuning
  8. Instruction Dataset Construction
  9. Instruction Tuning
  10. Methodology
  11. Datasets
  12. Evaluation Metrics
  13. Results
  14. Entity Disambiguation
  15. Question Answering
  16. ...and 5 more sections

Figures (1)

  • Figure 1: Pipeline of EntGPT frameworks: Given input document with the annotated mention, both of the approaches (1) generate relevant entity candidates. EntGPT-Prompt then (2) finds an auxiliary content of the annotated mention and then (3) selects the corresponding entity from the list. EntGPT-Instruct (2) constructs an instruction dataset, and (3) is instruction-tuned on the curated data.