Reproduce “IEPile: Unearthing Large-Scale Schema-Based Information Extraction Corpus” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Honghao Gui, Lin Yuan, Hongbin Ye, Ningyu Zhang, Mengshu Sun, Lei Liang, Huajun Chen
IEPILE, a large-scale bilingual (English and Chinese) schema-based instruction corpus from Zhejiang University and Ant Group, significantly enhances Large Language Models' (LLMs) zero-shot performance in Information Extraction tasks, including NER, RE, and EE, through novel instruction generation strategies. The corpus contains approximately 2 million instruction entries and 0.32 billion tokens, demonstrating improved generalization by mitigating schema query inconsistency and semantic confusion.
Key insights
- Intelligently structuring instruction data with both positive and carefully selected 'hard negative' schemas is critical for training robust LLMs that can effectively distinguish between semantically similar types.
- Dynamically varying the number of schema queries presented during instruction tuning is crucial for enabling LLMs to generalize effectively when faced with an inconsistent number of queries during inference.
- A comprehensive, large-scale, and schema-based instruction corpus can effectively bridge the performance gap of LLMs in complex and diverse Information Extraction tasks across multiple languages.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.