Abstract
The criminal amount helps the judges select the most appropriate sentence for the defendant. Criminal amount calculation aims to identify a final criminal amount in a judgment document automatically. We propose to model the task as a money entity classification problem, which can be solved by a BERT-based classification model. However, the money entity classification lacks large-scale high-quality human annotated data for training. Thus, we propose to leverage legal domain knowledge from unlabeled data to help this low-resource task. To be specific, we fine-tune a general BERT on an auxiliary task with legal domain data to embed domain knowledge, inspired by task augmentation. The natural language inference (NLI) task is adopted as the auxiliary task. And we use an NLI data generator to generate large-scale legal domain NLI data from unlabeled legal documents. Experimental results show that task augmentation is effective in domain knowledge transferring for low-resource criminal money entity classification.