Abstract: Contract risk detection helped prevent project risks and protect rights, improving the efficiency of contract review. However, existing methods still faced limitations in understanding the complex semantics of contract texts and extracting multi-scale features, which made it challenging to balance global semantic relations with the capture of local key information. This study proposed a novel multi-channel BERT-based pretraining method for contract text detection. The BERT pretraining model was utilized to extract the global features of the text, while a multi-channel convolution module was designed to capture multi-scale local features. An integrated attention mechanism adaptively learned the weight distribution between global and local features, enabling the precise extraction of key information from the text. On a publicly available contract text dataset, the proposed algorithm improved the contract risk classification accuracy by 4.6% compared to existing popular methods, validating the effectiveness of the proposed model. The model integrated the advantages of BERT pretraining, multi-channel convolution modules, and attention mechanisms, offering new research ideas and technical support for the field of text detection, with broad application prospects and significant practical value.

Key words: contract risk detection, text detection, BERT pre-training, attention mechanism, multi-channel