Contract review system based on large language models, method and computer readable medium thereof

Abstract

This invention discloses a contract review system and method incorporating a large-scale language model. Based on user-selected review criteria, it parses documents through a clause segmentation module to reconstruct the hierarchical structure of clauses, items, and terms. It then uses a review question generation algorithm to generate corresponding review questions. Subsequently, a topic-based risk review algorithm semantically compares the review questions with the contract clauses across multiple topics, detecting inconsistencies or omissions between clauses and generating clear and directly applicable revision suggestions. This automatically completes the review of contract clauses. Notably, the review results can be inserted into the document according to the original clause position, presented as annotations or revision tracking, thereby improving the efficiency and accuracy of contract review. This invention also provides a computer-readable medium for performing the method of this invention.

Claims

1. A contract review system incorporating a large-scale language model, comprising: The clause segmentation module extracts multiple structural elements from an editable document to be reviewed. It then reconstructs the hierarchical tree structure of the document by parsing its group identifiers, hierarchical indicators, and heading levels, assigning a unique location code to each node as an anchor point for subsequent annotations and comparisons. The review module, connected to the clause segmentation module, includes a contract comparison inference model for contract review and review questions generated by user-selected review criteria. The contract comparison inference model uses a topic-based risk review method to map each contract clause of the document to the review question and then performs basic checks on the document and semantic-level risk checks according to the review questions corresponding to each contract topic, generating review results. The result annotation and tracking revision module, connected to the review module, receives the review results after the review module completes its review. It inserts the review results into the original document according to hierarchical paragraphs and automatically generates annotations and tracking revisions. The annotations detail the review question summary, review opinions, and supporting evidence.

2. The contract review system combining a large language model as described in claim 1 further includes a fine-tuning module for the inference model linked to the review module, used to train the contract comparison inference model so that the contract comparison inference model has the ability to make contract semantic judgments and suggest clause revisions, wherein, The fine-tuning module of this inference model trains the contract comparison inference model using training sample data consisting of demand conditions, contract terms, inference process, and suggested terms for modification. It also incorporates the training sample data into the inference steps using the inference chain fine-tuning method, so that the contract comparison inference model can output the inference process simultaneously when judging the contract terms.

3. The contract review system combining a large language model as described in claim 2, wherein, The inference chain fine-tuning method adds thinking tags to the training sample data to record the thinking process that generates the correct answer. Then, through the thinking tags, the contract comparison inference model automatically links the blocks contained in the thinking tags with the inference process, so that inference can be carried out when judging the contract terms.

4. The contract review system combining a large language model as described in claim 2, wherein, The fine-tuning module of the inference model includes fine-tuning the contract comparison inference model using rank adaptive reliability optimization technique during training. The rank adaptive reliability optimization technique includes embedding trainable low-rank patches into the specific weight matrix of the contract comparison inference model and automatically adjusting the scaling according to the rank of the patches. During training, the output of the contract comparison inference model is adjusted synchronously so that the contract comparison inference model includes the inference steps and correction suggestions, and establishes the semantic relationship between the main body of the clause and the context.

5. A contract review system incorporating a large language model as described in claim 1, wherein, The clause segmentation module parses the review criteria selected by the user to restore the hierarchical structure of the clauses and obtain the content of the clauses. Then, the contract comparison and reasoning model uses the review question generation method to perform detailed normalization and parsing of each criterion clause to extract verifiable specific information. After that, the review questions that can be directly compared are automatically generated based on the specific information.

6. A contract review system incorporating a large language model as described in claim 5, wherein, If the user inputs new review criteria, the review module will generate new review questions using the review question generation method, and compare them with the existing review questions during contract review.

7. A contract review method incorporating a large-scale language model includes the following steps: A clause segmentation module obtains multiple structural elements from an editable document to be reviewed, and then reconstructs the hierarchical tree structure of the document by parsing its group identifiers, hierarchical indicators, and heading levels, assigning a unique location code to each node as an anchor point for subsequent annotations and comparisons; A review module, equipped with a contract comparison inference model for contract review and review questions generated by user-selected review criteria, uses the contract comparison inference model and a topic-based risk review method to map each contract clause of the document to be reviewed to the review question and then performs a basic check on the document and a semantic-level risk check according to the review questions corresponding to each contract topic, generating a review result; and a result annotation and tracking revision module receives the review result obtained by the review module after review, inserts the review result into the original document to be reviewed according to hierarchical paragraphs, and automatically generates annotations and tracking revisions, wherein... The annotations provide a detailed summary of the review issues, review comments, and supporting evidence.

8. The contract review method combining a large language model as described in claim 7 further includes training the contract comparison reasoning model with a fine-tuning module, enabling the contract comparison reasoning model to have the ability to make contract semantic judgments and suggest clause revisions, wherein, The fine-tuning module of this inference model trains the contract comparison inference model using training sample data consisting of demand conditions, contract terms, inference process, and suggested terms for modification. It also incorporates the training sample data into the inference steps using the inference chain fine-tuning method, so that the contract comparison inference model can output the inference process simultaneously when judging the contract terms.

9. The contract review method combining a large language model as described in claim 8, wherein, The inference chain fine-tuning method adds thinking tags to the training sample data to record the thinking process that generates the correct answer. Then, through the thinking tags, the contract comparison inference model automatically links the blocks contained in the thinking tags with the inference process, so that inference can be carried out when judging the contract terms.

10. The contract review method combining a large language model as described in claim 8, wherein, The fine-tuning module of the inference model includes fine-tuning the contract comparison inference model using rank adaptive reliability optimization technique during training. The rank adaptive reliability optimization technique includes embedding trainable low-rank patches into the specific weight matrix of the contract comparison inference model and automatically adjusting the scaling according to the rank of the patches. During training, the output of the contract comparison inference model is adjusted synchronously so that the contract comparison inference model includes the inference steps and correction suggestions, and establishes the semantic relationship between the main body of the clause and the context.

11. The contract review method combining a large language model as described in claim 7, wherein, The clause segmentation module parses the review criteria selected by the user to restore the hierarchical structure of the clauses and obtain the content of the clauses. Then, the contract comparison and reasoning model uses the review question generation method to perform detailed normalization and parsing of each criterion clause to extract verifiable specific information. After that, the review questions that can be directly compared are automatically generated based on the specific information.

12. The contract review method combining a large language model as described in claim 11, wherein, If the user inputs new review criteria, the review module will generate new review questions using the review question generation method, and compare them with the existing review questions during contract review.

13. A computer-readable medium, applied in a computing device or computer, storing instructions for performing a contract review method incorporating a large language model as described in any one of claims 7 to 12.

Images