BCL: Bayesian In-Context Learning Framework for Information Extraction

2026-06-18 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

BCL, a Bayesian In-Context Learning Framework for Information Extraction, introduces the first optimization method using particle filtering with Bayesian updates to systematically refine label representations across diverse IE tasks. This framework, operating through initialization, observation, weight update, and resampling, generalizes effectively to both sequence labeling and relation classification. Experiments show BCL consistently outperforms existing ICL approaches, achieving up to 30% F1 score improvements. For instance, on CoNLL03 with Qwen-2.5-7B, BCL reached 72.83 F1, significantly surpassing ChatIE's 25.64 F1. It also enables cost-efficient deployment, with Qwen-2.5-3B using BCL matching Llama-3.1-8B's one-shot performance on CoNLL03 with 62% fewer parameters, and demonstrates data efficiency by converging with only 3-5% of training data.

Key takeaway

For Machine Learning Engineers deploying information extraction solutions, if you face inconsistent performance with current in-context learning methods or operate under computational constraints, you should evaluate BCL. This framework provides a systematically optimized and data-efficient approach that generalizes across NER and RE tasks, even on smaller LLMs. Its ability to transfer learned rules to stronger models further enhances deployment flexibility and performance reliability.

Key insights

BCL optimizes in-context learning for information extraction using Bayesian particle filtering to refine semantic label patterns.

Principles

• Decompose labels into fine-grained semantic units to treat context as controllable discrete variables.
• Systematically refine rule representations using Bayesian updates and particle filtering for consistent performance.

Method

BCL refines label representations via particle filtering: initialize patterns, evaluate via ICL, update weights with Bayesian posterior, and resample with LLM-guided mutation.

In practice

• Apply BCL to both sequence labeling (NER) and relation classification (RE) tasks.
• Optimize ICL rules on smaller LLMs and transfer them to larger models for improved performance.

Topics

• Information Extraction
• In-Context Learning
• Bayesian Optimization
• Particle Filtering
• Large Language Models
• Sequence Labeling
• Relation Classification

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• BCL: Bayesian In-Context Learning Framework for Information Extraction — BCL uses Bayesian particle filtering to systematically optimize in-context learning for information extraction, improving performance and generalizability.
• Distilling Examples into Task Instructions: Enhanced In-Context Learning for Real-World B2B Conversations — Distilling ICL examples into structured instructions significantly improves performance and efficiency for complex B2B conversation classification.
• Uncertainty-Aware Clarification in LLM Agents with Information Gain — Information Gain Reward quantifies clarification utility for LLM agents, improving task success by reducing uncertainty in user intent.
• Nested Named Entity Recognition in Plasma Physics Research Articles — A lightweight BERT-CRF model with entity-specific specialization and Bayesian Optimization effectively extracts nested entities from plasma physics texts.
• Context Payload Optimization for ICL-Based Tabular Foundation Models — Optimizing context payloads is crucial for balancing accuracy, latency, and cost in ICL tabular foundation models.
• In-Context Learning for the Imputation of Public Opinion Data with Large Language Models — In-context learning with LLMs significantly improves missing public opinion data imputation over traditional statistical methods.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.