Abstract
Contract anomaly detection, as one of the fundamental building blocks for contract analysis, can identify potential violations and abnormal clauses efficiently, further help to reduce legal and operational risks and evaluate whether the contract complies with regulatory requirements and internal norms. Traditional contract detection relies on legal professionals and is plagued by problems such as low efficiency, high costs, and unstable detection quality. Consequently, automated contract detection has gained significant research attention. However, existing detection methods encounter three pivotal challenges. First, the named entity recognition model based on annotation suffers from inaccurate entity boundary detection. This leads to entity extraction errors and a decrease in the model's recognition performance. Second, contract data scarcity and expensive manual annotation hinder model training. Third, effectively eliciting knowledge from large language models for contract risk detection remains a challenge. To address these challenges, this paper proposes a multi-source feature fusion approach for contract anomaly detection. Four types of features are extracted from contract texts: semantic features, structural features, keyword features and TF-IDF features. These features are then concatenated and input into the Fuzzy Neighborhood Rough Set model to perform anomaly detection. We conducted experiments on 10 datasets, and 9 unsupervised algorithms are evaluated, including both established and contemporary anomaly detection techniques. The results demonstrate that BERT-FNRS achieves favorable performance. It performs prominently in Accuracy and ROCAUC, with average values of 0.9652 and 0.9252, respectively.

This publication has 41 references indexed in Scilit: