Intelligent document review and traceability positioning method based on LLM natural language processing

Abstract

The invention discloses an intelligent document reviewing and traceability positioning method based on LLM natural language processing. The method comprises the steps that 1, a reviewing document enters an OCR analyzer to be split into text blocks and positioning information; 2, vectorizing the text blocks and storing positioning information corresponding to each text block into a vector database; step 3, vectorizing the review rule; step 4, performing enhanced hybrid retrieval based on the vectorized review rule; step 5, rearranging the most relevant K text blocks with the positioning information retrieved in the step 4; step 6, the large language model extracts key information from the rearranged text blocks according to an examination rule based on a cue word project; 7, the big language model returns a judgment result and positioning information based on the cue word project according to the review rule and the key information; and 8, displaying a judgment result and positioning traceability information. According to the invention, the efficiency and accuracy of examination can be improved, and accurate traceability and positioning of error contents can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing in computer science, and in particular to an intelligent document review and source tracing method based on LLM natural language processing. Background Art

[0002] In today's information age, document review is a vital task, especially in the fields of law, finance, and medicine. Traditional document review usually relies on manual inspection, which is intuitive but inefficient when faced with a large number of documents and is prone to omissions or misjudgments. With the increase in business volume, manual review is not only time-consuming and labor-intensive, but also difficult to ensure the consistency and accuracy of the review.

[0003] In recent years, with the development of natural language processing (NLP) technology, especially the progress of deep learning and large-scale language models (LLM), automated document review has become possible. Existing technologies can already assist humans in completing the initial screening of documents to a certain extent, but these technologies often have the following problems:

[0004] Slow review speed: Despite the introduction of automation, existing review systems are still unable to achieve real-time or fast review for large document sets.

[0005] Inaccurate positioning: Existing automated review tools often have difficulty accurately locating specific erroneous paragraphs or sentences, making subsequent manual review work complicated.

[0006] Limited review accuracy: Existing automated tools still need to be improved in identifying complex logical errors or semantic errors.

[0007] Therefore, in view of the shortcomings of the existing technology, it is necessary to provide an intelligent document review and traceability positioning method based on LLM natural language processing to overcome the shortcomings of the existing technology. Summary of the invention

[0008] The purpose of the present invention is to avoid the shortcomings of the prior art and provide an intelligent document review and source tracing method based on LLM natural language processing, which can improve the efficiency and accuracy of the review and realize the accurate source tracing of erroneous content. It aims to solve the problems existing in the prior art by combining the technical means of document block vector storage and large-scale language model, and provide an efficient and accurate document review solution.

[0009] The purpose of the present invention is achieved through the following technical measures.

[0010] An intelligent document review and traceability positioning method based on LLM natural language processing is provided, which is carried out through the following steps:

[0011] Step 1. Review the document and enter the OCR parser to split the text blocks and positioning information;

[0012] Step 2. Vectorize the text blocks and store the positioning information corresponding to each text block into a vector database;

[0013] Step 3. Vectorize the review rules;

[0014] Step 4. Perform enhanced hybrid retrieval on the document information stored in the vector database based on the vectorized review rules;

[0015] Step 5. Rearrange the most relevant K text blocks with positioning information retrieved in step 4, where K is a natural number and means the number of matches that are expected to be obtained;

[0016] Step 6. The large language model extracts key information from the rearranged text block based on the prompt word engineering and the review rules;

[0017] Step 7. The large language model returns the judgment result and positioning information based on the prompt word engineering according to the review rules and key information;

[0018] Step 8. Display the judgment results and location tracing information.

[0019] Preferably, in the above-mentioned intelligent document review and traceability positioning method based on LLM natural language processing, step 1 specifically involves performing OCR optical recognition on the document to be reviewed, and after entering the parser for parsing, the document is split into text blocks (chunks) according to chapters, and each text block has corresponding positioning information (position).

[0020] Preferably, in the above-mentioned intelligent document review and traceability positioning method based on LLM natural language processing, step 2 specifically vectorizes each text block through a vector model.

[0021] Preferably, in the above-mentioned intelligent document review and traceability positioning method based on LLM natural language processing, in step 2, the text is converted into a vector by using an embedding model, using the open source model bge-large-zh-v1.5.

[0022] Preferably, in the above-mentioned intelligent document review and traceability positioning method based on LLM natural language processing, step 3 specifically vectorizes the review rules through a vector model.

[0023] Preferably, in the above-mentioned intelligent document review and traceability positioning method based on LLM natural language processing, step 3 specifically converts the text of the review rule into a vector through an embedding model, using the open source model bge-large-zh-v1.5.

[0024] Preferably, in the above-mentioned intelligent document review and traceability positioning method based on LLM natural language processing, the enhanced hybrid search in step 4 is a vector search combined with a full-text search.

[0025] Preferably, in the above-mentioned intelligent document review and traceability positioning method based on LLM natural language processing, in step 5, rearrangement is performed through a rearrangement model, and the documents are arranged from front to back in a descending order of relevance.

[0026] The intelligent document review and source tracing method based on LLM natural language processing is carried out through eight steps. By combining the technical means of document block vector storage and large-scale language model, the efficiency and accuracy of review can be improved, and accurate source tracing of erroneous content can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The present invention is further described with reference to the accompanying drawings, but the contents in the accompanying drawings do not constitute any limitation to the present invention.

[0028] Figure 1 It is a flow chart of an intelligent document review and traceability positioning method based on LLM natural language processing of the present invention.

[0029] Figure 2 This is a partial screenshot of the directory of the document in Example 2 of the present invention.

[0030] Figure 3 It is a schematic diagram of text block information obtained after parsing and splitting in Example 2 of the present invention.

[0031] Figure 4 This is a screenshot of an example of positioning one of the text blocks obtained after parsing and splitting in Example 2 of the present invention.

[0032] Figure 5 It is a schematic diagram of the result after rearrangement in Example 2 of the present invention.

[0033] Figure 6 It is a schematic diagram of the result obtained by the method of Example 2 of the present invention.

[0034] Figure 7 It is another schematic diagram of the result obtained by the method of Example 2 of the present invention. DETAILED DESCRIPTION

[0035] The present invention will be further described with reference to the following examples.

[0036] Example 1.

[0037] Large Language Model (LLM), also known as Large Language Model, is an artificial intelligence model designed to understand and generate human language. They are trained on large amounts of text data and can perform a wide range of tasks, including text summarization, translation, sentiment analysis, and more. LLMs are characterized by their large size, which helps them learn complex patterns in language data. These models are usually based on deep learning architectures, which helps them achieve impressive performance on various NLP tasks.

[0038] The present invention aims to solve the problems existing in the prior art by combining the technical means of document block vectorized storage with large-scale language models, and to provide an efficient and accurate document review solution, which can improve the efficiency and accuracy of the review and can also achieve accurate tracing and positioning of erroneous content.

[0039] This embodiment provides an intelligent document review and traceability positioning method based on LLM natural language processing, such as Figure 1 As shown, proceed as follows:

[0040] Step 1. Review the document and enter the OCR parser to split the text blocks and positioning information;

[0041] Step 2. Vectorize the text blocks and store the positioning information corresponding to each text block into a vector database;

[0042] Step 3. Vectorize the review rules;

[0043] Step 4. Perform enhanced hybrid retrieval based on the vectorized review rules;

[0044] Step 5. Rearrange the most relevant K text blocks with positioning information retrieved in step 4, where K is a natural number and means the number of matches that are expected to be obtained;

[0045] Step 6. The large language model extracts key information from the rearranged text block based on the prompt word engineering and the review rules;

[0046] Step 7. The large language model returns the judgment result and positioning information based on the prompt word engineering according to the review rules and key information;

[0047] Step 8. Display the judgment results and location tracing information.

[0048] The main problems of traditional methods are slow document review, inaccurate positioning and limited review accuracy. In order to improve the efficiency and accuracy of the review, the document must first be analyzed by optical character recognition, split into text blocks and obtain positioning information. Then, these text blocks are vectorized and stored in a vector database. The review rules are also vectorized for enhanced hybrid retrieval. The most relevant text blocks obtained through retrieval are rearranged, and the large language model extracts key information based on the prompt word engineering, and finally returns the judgment result and positioning information. The judgment result and positioning traceability information are displayed, which effectively solves the above problems.

[0049] In this technical solution, the document is split into text blocks by the OCR parser, and the positioning information is obtained. Each text block is vectorized by the vector model and stored in the vector database. The review rules are also vectorized, and the embedding model is used to convert the text into vectors. Enhanced hybrid retrieval is performed based on the vectorized review rules, and the most relevant text blocks retrieved are rearranged. The large language model extracts key information from the rearranged text blocks based on the prompt word engineering, and returns the judgment result and positioning information. Finally, the judgment result and positioning traceability information are displayed.

[0050] Specifically, the document is first split into text blocks by the OCR parser, and the corresponding positioning information is obtained. The text blocks are vectorized through the vector model and stored in the vector database. The review rules are also vectorized through the vector model, and the embedding model is used to convert the text into vectors. Enhanced hybrid retrieval is performed based on the vectorized review rules, combining the advantages of vector retrieval and full-text retrieval, and the most relevant text blocks retrieved are rearranged. The large language model extracts key information from the rearranged text blocks based on the prompt word engineering, and returns the judgment results and positioning information. Finally, the judgment results and positioning traceability information are displayed.

[0051] The intelligent document review and traceability positioning method based on LLM natural language processing, step 1 specifically is to perform OCR optical recognition on the document to be reviewed, and after entering the parser for parsing, the document is split into text blocks (chunks) according to chapters, and each text block has corresponding positioning information (position).

[0052] OCR optical recognition is used to digitize document content, parser analysis is used to process digitized document content, document chapter splitting is used to divide document content into smaller units, namely text blocks, and text block positioning information is used to record the position of each text block in the original document. These technical features work together to digitize document content and split it into text blocks with positioning information through OCR optical recognition and parser analysis, thus solving the technical problem of how to perform OCR optical recognition on documents that need to be reviewed and split them into text blocks with positioning information.

[0053] Specifically, OCR optical recognition technology can use existing open source OCR tools, such as Tesseract OCR, to convert paper documents or scans into editable digital text. The parser is used to further process these digital texts, parsing and organizing them according to predefined rules and formats. Document chapter splitting can break down the document content into smaller text blocks based on features such as titles and paragraph marks in the document. The positioning information of each text block can be achieved by recording its page number, line number, etc. in the original document.

[0054] Through the above-mentioned technical means, the documents that need to be reviewed can be effectively digitized and the document content can be split into smaller units, thereby improving the efficiency and accuracy of document review. Compared with the prior art, the technical solution of the present application can complete the OCR recognition and parsing of the document more quickly, and can accurately locate the position of each text block in the original document, facilitating subsequent review and tracing. As a result, the present application overcomes the problems of slow review speed, inaccurate positioning, and limited review accuracy in the prior art.

[0055] In step 2, each text block is vectorized through a vector model and the location information corresponding to each text block is stored in a vector database. For example, the text can be converted into a vector through an embedding model, using the open source model bge-large-zh-v1.5.

[0056] The purpose of this step is to convert the text block into a vector representation for subsequent processing and retrieval. Vectorized text blocks can be more easily calculated and matched for similarity, improving the efficiency and accuracy of document review. This vectorization method can effectively solve the efficiency and accuracy problems of text blocks in subsequent steps of retrieval and matching. Vectorized text blocks can better match vectorized review rules, thereby improving the accuracy of review and positioning.

[0057] Vector models can be implemented in a variety of ways, such as using an embedding model to convert text blocks into vectors. Specifically, you can use an open source model such as bge-large-zh-v1.5, input the text block into the model, and obtain the corresponding vector representation through the model's calculation. These vector representations can be stored in a vector database for subsequent retrieval and processing. In addition, the choice of vector model can be adjusted according to specific application scenarios and requirements to achieve the best results.

[0058] By vectorizing text blocks, the technical solution of the present application can significantly improve the efficiency and accuracy of document review. Compared with the prior art, the method of the present application shows higher review speed and positioning accuracy when processing large-scale document sets. At the same time, through vectorization processing, the review system has also significantly improved in identifying complex logical errors or semantic errors, thereby effectively solving the problems of slow review speed, inaccurate positioning and limited review accuracy existing in traditional review tools. Therefore, the technical solution of the present application shows significant technical advantages in document review and traceability positioning.

[0059] The role of the embedding model is to vectorize text blocks, while the role of the open source model bge-large-zh-v1.5 is to serve as a specific implementation tool to convert text blocks into vectors. These two technical features work together to implement the process of text block vectorization through the embedding model and the specific open source model, solving the technical problem of how to convert text blocks into vectors. By using the embedding model and the open source model bge-large-zh-v1.5, the solution of this application can effectively convert text blocks into vectors, providing basic data support for subsequent review rule vectorization and enhanced hybrid retrieval steps, thereby realizing intelligent document review and traceability positioning.

[0060] Specifically, the embedding model is a technology that converts text data into numerical vectors, which captures the semantic information of the text through a pre-trained neural network model. The open source model bge-large-zh-v1.5 is an embedding model that has been pre-trained with a large amount of Chinese corpus and can efficiently convert Chinese text into vector representations. By inputting a block of text into the model, the model will output a high-dimensional vector that can represent the semantic information of the input text. This process can be implemented by calling relevant open source libraries and APIs. The specific implementation method may include but is not limited to using the Python programming language and related deep learning frameworks such as TensorFlow or PyTorch.

[0061] Therefore, this step provides an efficient and accurate text block vectorization method by combining the embedding model and the open source model bge-large-zh-v1.5. Compared with the prior art, the solution of this application can process large-scale text data more quickly while ensuring the accuracy and consistency of the vectorization results. This provides a solid foundation for subsequent intelligent document review and traceability positioning, and significantly improves the efficiency and accuracy of document review.

[0062] Step 3 specifically vectorizes the examination rules through a vector model. The technical feature of vectorizing the examination rules through a vector model plays a key role in solving the problem of vectorizing the examination rules. Vectorization can convert the examination rules into a vector form that can be processed by computers, making subsequent retrieval and matching operations more efficient and accurate.

[0063] The vectorization process can be implemented in the following ways: First, an open source embedding model, such as bge-large-zh-v1.5, can be used to convert the text of the review rules into vectors. Second, different review rules can be trained based on the vector model to generate specific vector representations. Finally, the accuracy and efficiency of vectorization can be improved by optimizing the parameters of the vector model.

[0064] Compared with the existing technology, the review rules are vectorized by the vector model, which has higher efficiency and accuracy. The vectorized review rules can better match the text blocks in the document, improving the speed and accuracy of document review. In addition, this technical means can also reduce the workload of manual participation and reduce the possibility of human error. Therefore, this application provides a more intelligent and automated solution in the field of document review.

[0065] Using the embedding model, the text of the review rules is converted into vectors. This feature plays a key role in solving the problem of how to convert the text of the review rules into vectors. Specifically, through the embedding model, especially the open source model bge-large-zh-v1.5, the text data of the review rules can be converted into a vector representation. This vectorized representation facilitates subsequent calculations and processing, allowing the system to apply and match review rules more efficiently, thereby improving the accuracy and efficiency of document review.

[0066] In this application, the use of embedding models is one of the key steps. The embedding model converts text data into vectors, allowing complex text information to be represented in mathematical form, making it easier for computers to process and analyze. The open source model bge-large-zh-v1.5 is a trained deep learning model specifically for vectorization of Chinese text. Specifically, the model is trained with a large amount of Chinese corpus, which can capture semantic information in the text and convert it into high-dimensional vectors. These vectors can not only represent the content of the text, but also reflect the similarities and differences between texts.

[0067] For example, in practical applications, each rule of the review rule can be input into the bge-large-zh-v1.5 model, and the model will output a high-dimensional vector to represent the semantic information of the rule. These vectors can be stored in a vector database for subsequent retrieval and matching. When a document needs to be reviewed, the text block in the document can also be input into the model to obtain the corresponding vector representation. Then, by calculating the similarity between the text block vector and the review rule vector, the part of the document that is most relevant to the review rule can be quickly located, achieving efficient and accurate document review.

[0068] Therefore, this application converts the text of the review rules into vectors by using an embedding model, which has significant advantages over the prior art. First, this application can significantly improve the review speed, and through vectorization processing, fast text matching and retrieval can be achieved. Secondly, this application can improve positioning accuracy, and through high-dimensional vectors to represent the semantic information of the text, it can more accurately locate the relevant paragraphs or sentences in the document. Finally, this application can improve the review accuracy, and through the deep learning model to capture the complex semantic information in the text, it can better identify complex logical errors or semantic errors.

[0069] Step 4. Perform enhanced hybrid search based on the vectorized review rules. Enhanced hybrid search is a combination of vector search and full-text search. The previous step has vectorized the entire document and stored it in the vector database, and then the review rules are vectorized. Step 4 performs enhanced hybrid search on the document information stored in the vector database based on the vectorized review rules. Enhanced hybrid search is an implementation of ES's full-text search + vector search, which can more accurately retrieve document text blocks that are closer to the review rules.

[0070] This application proposes an enhanced hybrid retrieval method, which aims to improve the accuracy and efficiency of retrieval by combining two technical means: vector retrieval and full-text retrieval. Vector retrieval can process semantic similarity, while full-text retrieval can quickly locate specific text content. By combining these two methods, the technical problem of how to combine vector retrieval and full-text retrieval for enhanced hybrid retrieval can be better solved, thereby improving the accuracy and efficiency of document review.

[0071] The implementation of enhanced hybrid retrieval can include the following: First, vector retrieval can use pre-trained embedding models to vectorize text blocks and review rules. Specifically, the open source model bge-large-zh-v1.5 can be used. Then, full-text retrieval can quickly retrieve text blocks containing specific keywords through technologies such as inverted indexing. Finally, combining the two retrieval results, the results are rearranged through the rearrangement model, and arranged in descending order of relevance, so as to obtain the most relevant K text blocks.

[0072] This application can significantly improve the accuracy and efficiency of document review by combining two technical means: vector retrieval and full-text retrieval. Compared with the prior art, the enhanced hybrid retrieval method of this application can better handle the problems of semantic similarity and specific text positioning, thereby improving the accuracy and reliability of the review results. Therefore, this application improves the efficiency of document review while also improving the accuracy of the review and reducing the workload of manual review.

[0073] Step 5. Rearrange the most relevant K text blocks with positioning information retrieved in step 4, where K is a natural number and a pre-set parameter, which means the number of matches you want to obtain. Rearrangement can be performed using a rearrangement model, and the text blocks are arranged from front to back in a descending order of relevance. The more relevant the text blocks, the higher they are placed, which is beneficial for subsequent large model extraction and judgment.

[0074] The rearrangement model in step 5 rearranges the retrieved text blocks so that the text blocks with high relevance are arranged in front and the text blocks with low relevance are arranged in the back. This can improve the accuracy of the arrangement of relevant text blocks in the review document, facilitate the processing and judgment of text blocks by the large language model in the subsequent steps, and improve the overall review efficiency and accuracy.

[0075] The core of the rearrangement model is to optimize the arrangement order of text blocks. Specifically, the rearrangement model can be implemented based on a variety of algorithms, such as a machine learning-based sorting algorithm or a rule-based sorting algorithm. The machine learning sorting algorithm can learn the correlation between text blocks through training data and sort them according to the learned model. The rule-based sorting algorithm can sort based on predefined rules, such as keyword matching, text block importance score, etc. Furthermore, a hybrid sorting method can be used in combination with multiple algorithms to improve the accuracy and robustness of the sorting results.

[0076] The technical solution of rearranging by rearrangement model can significantly improve the accuracy of arrangement of relevant text blocks in the review document. Compared with the prior art, the technical solution of the present application can more efficiently identify and sort relevant text blocks, thereby reducing manual intervention and improving the automation and accuracy of the review process. As a result, it can significantly improve the review efficiency, reduce the error rate and ensure the reliability of the review results in large-scale document review tasks.

[0077] The technical solution of the present application involves rearranging the retrieved text blocks through a rearrangement model during the intelligent document review and traceability positioning process. The function of the rearrangement model is to arrange the text blocks in a descending order of relevance, thereby improving the accuracy and efficiency of the document review. In this way, the text blocks that match the review rules can be more effectively located, ensuring that the returned results are more relevant and accurate.

[0078] The reordering model can be implemented in many ways, such as using a deep learning-based sorting model, which can accurately determine the correlation between text blocks and review rules after learning from a large amount of training data. Specifically, pre-trained language models such as BERT and RoBERTa can be used for fine-tuning to enable them to sort text blocks according to specific review rules. In addition, other features, such as the length and position of text blocks, can be combined to further improve the accuracy of sorting.

[0079] Through the above technical solution, this application can significantly improve the efficiency and accuracy of intelligent document review. Compared with the prior art, this application rearranges the text blocks through the rearrangement model to ensure that the returned results are more relevant and accurate, thereby improving the effect and efficiency of document review. Therefore, this application has obvious advantages in solving the shortcomings of the prior art.

[0080] Step 6. The large language model extracts key information from the rearranged text block based on the review rules based on the prompt word engineering. The key information is what the review rules need to review. For example, to determine whether a land area meets the specifications, the key information is "land area". The prompt word engineering can use some prompt word techniques, such as using Cot, Tot, few-shot and other techniques to write some prompt word sentences to improve the accuracy of the review of the large model.

[0081] Step 7. The large language model returns the judgment result and positioning information based on the review rules and key information based on the prompt word engineering. The prompt word engineering can use some prompt word techniques, such as using Cot, Tot, few-shot and other techniques to write some prompt word sentences to improve the accuracy of the review of the large model.

[0082] Finally, step 8 is used to display the judgment results and location tracing information.

[0083] The intelligent document review and source tracing method based on LLM natural language processing is carried out through eight steps. By combining the technical means of document block vector storage and large-scale language model, the efficiency and accuracy of review can be improved, and accurate source tracing of erroneous content can be achieved.

[0084] The invention discloses an intelligent document review and source tracing method based on LLM natural language processing. By dividing the document content into blocks and vectorizing the storage, and utilizing efficient vector database retrieval technology, it is possible to quickly locate potential problem paragraphs, thereby improving the review efficiency. Combining a large-scale language model (LLM) to analyze the document content can improve the ability to identify complex logical errors and semantic errors, thereby enhancing the accuracy of the review. By rearranging the blocks and matching them with the rules, accurate source tracing of the erroneous content can be achieved, which facilitates subsequent manual review and realizes source tracing. Through automated processing procedures, human intervention is reduced, the automation level of the overall review process is improved, and the workload of reviewers is reduced, with a high degree of automation. This technical solution is applicable to a variety of document types and industry fields, such as legal contracts, financial reports, medical records, etc., and is adaptable to a variety of application scenarios.

[0085] By dividing the document content into blocks and storing them vectorized, the document blocks most relevant to the rules can be quickly retrieved, greatly improving the efficiency of review and having the characteristics of efficient retrieval. By using vectorization technology combined with large model rule matching, it is possible to accurately locate the specific erroneous paragraphs or sentences in the document, improving the accuracy of review and having the characteristics of precise positioning. Combined with the Large Scale Language Model (LLM), the system can understand and judge the content of the document, identify complex logical errors and semantic errors, and has the function of intelligent judgment.

[0086] Example 2.

[0087] The intelligent document review and traceability positioning method based on LLM natural language processing of the present invention is described in detail with a specific example.

[0088] First, prepare a document that needs to be reviewed. A partial screenshot of the document directory is as follows: Figure 2 shown.

[0089] Step 1. Enter the review document into the OCR parser to split the text blocks and positioning information. Enter the document OCR parsing, and split the document into text blocks according to each chapter and title. After parsing, the split text blocks are as follows: Figure 3 As shown, each line represents a text block.

[0090] Each text block also has its positioning information, such as Figure 4 Here is a screenshot of an example of positioning a text block. Figure 4Here, each number corresponds to the array representing ["Page Number","Left","Top","Right","Top","Left","Bottom","Right","Bottom"].

[0091] Step 3: Vectorize the review rules. Step 4: Perform enhanced hybrid retrieval on the document information stored in the vector database based on the vectorized review rules.

[0092] In this embodiment, some pre-built-in review rules are used to perform vectorization retrieval to find the text blocks with the highest similarity, and then rearrange them. The result after rearrangement is as follows: Figure 5 shown.

[0093] The LLM large language model is used to extract key information from the rearranged text block based on the prompt word engineering according to the review rules. The large language model returns the judgment result and positioning information based on the prompt word engineering according to the review rules and key information. Figure 6 This is a result obtained by the method of the present invention. It can be seen that the review document mentions education and development background and the jump file location can be highlighted. Figure 7 It is another schematic diagram of a result obtained by the method of the present invention, and the result checks that the total building area in the document does not meet the specifications.

[0094] The experimental results verify that the intelligent document review and source tracing method based on LLM natural language processing is carried out in eight steps. By combining the technical means of document block vector storage and large-scale language model, the efficiency and accuracy of review can be improved, and accurate source tracing of erroneous content can be achieved.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit the scope of protection of the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the essence and scope of the technical solution of the present invention.

Claims

1. An intelligent document review and traceability method based on LLM natural language processing, characterized in that: Proceed as follows: Step 1. Review the document and enter the OCR parser to split the text blocks and positioning information; Step 2. Vectorize the text blocks and store the positioning information corresponding to each text block into a vector database; Step 3. Vectorize the review rules; Step 4. Perform enhanced hybrid retrieval based on the vectorized review rules; Step 5. Rearrange the most relevant K text blocks with positioning information retrieved in step 4, where K is a natural number and means the number of matches that are expected to be obtained; Step 6. The large language model extracts key information from the rearranged text block based on the prompt word engineering and the review rules; Step 7. The large language model returns the judgment result and positioning information based on the prompt word engineering according to the review rules and key information; Step 8. Display the judgment results and location tracing information.

2. The intelligent document review and traceability positioning method based on LLM natural language processing according to claim 1 is characterized in that: Step 1 specifically involves performing OCR optical recognition on the document to be reviewed. After entering the parser for parsing, the document is split into text blocks according to chapters, and each text block has corresponding positioning information.

3. The intelligent document review and traceability positioning method based on LLM natural language processing according to claim 1 or 2 is characterized in that: In step 2, each text block is specifically vectorized through a vector model.

4. The intelligent document review and traceability positioning method based on LLM natural language processing according to claim 3 is characterized in that: In step 2, the text is converted into a vector through the embedding model, using the open source model bge-large-zh-v1.

5.

5. The intelligent document review and traceability positioning method based on LLM natural language processing according to claim 3 is characterized in that: Step 3 specifically vectorizes the review rules through a vector model.

6. The intelligent document review and traceability positioning method based on LLM natural language processing according to claim 5 is characterized in that: In step 3, the text of the review rules is converted into a vector through the embedding model, using the open source model bge-large-zh-v1.

5.

7. The intelligent document review and traceability positioning method based on LLM natural language processing according to claim 5 is characterized in that: The enhanced hybrid search in step 4 is a combination of vector search and full-text search.

8. The intelligent document review and traceability positioning method based on LLM natural language processing according to claim 7 is characterized in that: In step 5, the re-ranking model is used to rank the data from front to back in descending order of relevance.

Images