BCL is the first framework that uses particle filtering and Bayesian updates to systematically refine label representations in information extraction. It achieves consistent performance across model scales and generalizes to both sequence labeling and relation classification through four key steps: initialization, observation, weight update, and resampling.