Enhancing charge prediction through the collaboration of large and small models

Charge prediction is a fundamental task in AI&Law, where the goal is to predict charges based on fact descriptions. Although various methods have been introduced to enhance performance, challenges remain. Specifically, small models (SMs)-based methods such as BERT struggle with hard cases involving low-frequency or confusing charges due to their limited capacity, whereas large language models (LLMs)-based approaches like GPT-4 exhibit difficulties in handling diverse charges owing to insufficient legal knowledge. To overcome these limitations, we propose a hybrid framework that collaborates both large and small models to improve charge prediction performance, based on the idea that combining the strengths of each can overcome their limitations. Initially, SMs provide an initial prediction along with a predicted probability distribution. If the maximum predicted probability falls below a threshold, LLMs step in to reflect and re-predict as needed. Additionally, we construct a confusing charges dictionary and design a two-stage legal inference prompt, which helps LLMs make the secondary prediction for the hard cases. Extensive experiments on two datasets from China and Italy demonstrate the effectiveness of this approach, yielding average F1 improvements of 7.94% and 11.46% respectively. Moreover, a fine-grained analysis demonstrates that our proposed framework is effective in identifying low-frequency and confusing charges.