ChunkUIE: Chunked instruction-based unified information extraction.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Large language models (LLMs) excel at linguistic tasks but struggle with cross-lingual information extraction.
• Inconsistent schema usage during training and evaluation degrades LLM performance in information extraction.
• Existing methods often increase training data without addressing schema mismatches.

Purpose of the Study:

• To develop a unified information extraction model (ChunkUIE) for both Chinese and English.
• To enhance the zero-shot performance of LLMs in information extraction tasks.
• To mitigate performance degradation caused by schema inconsistencies.

Main Methods:

• ChunkUIE employs a chunked instruction construction strategy, dividing schemas into reproducible chunks.
• This strategy ensures all schemas are covered while limiting schemas per instruction to manage inconsistencies.
• Challenging negative schemas are generated using a hard schema dictionary to reduce semantic confusion.

Main Results:

• ChunkUIE demonstrates improved zero-shot performance in information extraction tasks.
• The chunked instruction strategy effectively addresses schema count discrepancies between training and evaluation.
• The model shows enhanced ability to differentiate between semantically similar schemas.

Conclusions:

• ChunkUIE offers a robust solution for unified Chinese and English information extraction.
• The proposed chunking strategy is effective in improving LLM performance despite schema variations.
• This work advances the field of cross-lingual information extraction using LLMs.