Abstract:Contract documents are typically characterized by long text spans, complex hierarchical structures, and tightly coupled semantics, which pose significant challenges to contract information extraction due to ambiguous field boundaries and unclear structural attribution. Existing supervised extraction methods rely heavily on large-scale annotated data, resulting in high construction costs and limited domain adaptability. To address these issues, this study proposes a zero-shot contract information extraction method based on large language models. The proposed method incorporates boundary-constrained sliding windows aligned with the natural structural units of contracts to enhance contextual completeness and structural consistency. In the forward stage, candidate fields are generated, while in the reverse stage, a chain-of-thought reasoning mechanism is introduced to verify structural attribution and logical consistency. Furthermore, a confidence scoring system is constructed based on semantic similarity and generative stability to enable automated quality evaluation of extraction results. Experiments are conducted on a private enterprise contract dataset (Dataset A) and the public CCKS insurance contract dataset (Dataset B). The experimental results demonstrate that the proposed method significantly improves the extraction performance of large language models across diverse contract scenarios. Specifically, the GPT-4o model achieves F1 improvements of 7.73 and 8.57 percentage points on Datasets A and B, respectively, verifying the effectiveness and practical value of the proposed method for contract information extraction.