Document data processing method and device, equipment and medium

Abstract

The invention provides a document data processing method and device, equipment and a medium, and relates to the technical field of artificial intelligence. The method comprises the steps of obtaining a pre-established vector library and a to-be-audited document; performing document splitting processing according to the to-be-audited document and a preset splitting rule; performing search processing in a pre-established vector library according to each to-be-audited fragment; adjusting the first reference document name list in response to a user auditing and adjusting operation; according to the second reference document name list, determining reference source knowledge of each to-be-audited fragment, and performing knowledge point range adjustment processing on the reference source knowledge of each to-be-audited fragment; inputting the target reference source knowledge of each to-be-audited fragment, each to-be-audited fragment and the fixed cue word into a pre-training knowledge retrieval large model; and in response to a result processing operation of the user on the audit result of each to-be-audited fragment, obtaining a target audit result. According to the method, the auditing accuracy and efficiency are improved, and the human input in the document content auditing process is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a document data processing method and device, equipment and medium. BACKGROUND

[0002] Document content review is of great significance in ensuring information accuracy. By identifying and correcting potential problems in documents, it helps to provide a reliable information base to support scientific decision-making and avoid misleading or decision-making errors caused by false information.

[0003] In the prior art, document content review is usually implemented by relying on natural language processing (NLP) technology, machine learning algorithms or deep learning models. Among them, natural language processing technology involves key technologies such as word segmentation, part-of-speech tagging, and syntax analysis; while machine learning algorithms and deep learning models need to build a training set based on a large amount of labeled data, and train corresponding classification models or complex neural network structures to realize the understanding and judgment of document content.

[0004] However, in order to ensure the accuracy of the review, the existing document content review method usually needs to invest a large amount of manpower for data labeling and model training, resulting in high review cost, thus limiting its application efficiency in large-scale document processing scenarios. SUMMARY

[0005] The present application provides a document data processing method, device, equipment and medium, which splits the to-be-reviewed document into multiple to-be-reviewed segments, searches each to-be-reviewed segment in the pre-built vector library, obtains the corresponding reference document name list, and then searches the target reference source knowledge within the range corresponding to the reference document name list, and then obtains the review result of each to-be-reviewed segment by pre-training the knowledge retrieval large model, thereby narrowing the search range, improving the review accuracy and efficiency, and reducing the manpower investment in the document content review process.

[0006] In a first aspect, the present application provides a document data processing method, comprising:

[0007] obtaining a pre-built vector library and a to-be-reviewed document, wherein the pre-built vector library is constructed based on sample documents;

[0008] performing document splitting processing according to the to-be-reviewed document and a preset splitting rule to obtain multiple to-be-reviewed segments;

[0009] performing search processing on each to-be-reviewed segment in the pre-built vector library to obtain a first reference document name list composed of the names of the sample documents corresponding to each to-be-reviewed segment;

[0010] In response to a user review adjustment operation, the first reference document name list is adjusted to obtain a second reference document name list composed of names of sample documents corresponding to each to-be-reviewed segment;

[0011] According to the second reference document name list, the reference source knowledge of each to-be-reviewed segment is determined, and the knowledge point range adjustment processing is performed on the reference source knowledge of each to-be-reviewed segment to obtain target reference source knowledge of each to-be-reviewed segment;

[0012] The target reference source knowledge of each to-be-reviewed segment, each to-be-reviewed segment, and a fixed prompt word are input into a pre-trained knowledge retrieval large model to obtain an audit result of each to-be-reviewed segment;

[0013] In response to a user result processing operation on the audit result of each to-be-reviewed segment, a target audit result is obtained.

[0014] In a second aspect, the present application provides a document data processing device, comprising: an acquisition module, a document splitting module, a vector search module, a document review module, and a result output module;

[0015] The acquisition module is configured to acquire a pre-built vector library and a to-be-reviewed document, wherein the pre-built vector library is built based on sample documents;

[0016] The document splitting module is configured to perform document splitting processing according to the to-be-reviewed document and a preset splitting rule to obtain a plurality of to-be-reviewed segments;

[0017] The vector search module is configured to perform search processing on each to-be-reviewed segment in the pre-built vector library to obtain a first reference document name list composed of names of sample documents corresponding to each to-be-reviewed segment;

[0018] The document review module is configured to adjust the first reference document name list in response to a user review adjustment operation to obtain a second reference document name list composed of names of sample documents corresponding to each to-be-reviewed segment;

[0019] The document review module is further configured to determine the reference source knowledge of each to-be-reviewed segment according to the second reference document name list, and perform knowledge point range adjustment processing on the reference source knowledge of each to-be-reviewed segment to obtain target reference source knowledge of each to-be-reviewed segment;

[0020] The document review module is further configured to input the target reference source knowledge of each to-be-reviewed segment, each to-be-reviewed segment, and a fixed prompt word into a pre-trained knowledge retrieval large model to obtain an audit result of each to-be-reviewed segment;

[0021] The result output module is configured to obtain a target audit result in response to a user result processing operation on the audit result of each to-be-reviewed segment.

[0022] In a third aspect, the present application provides an electronic device, comprising: a processor, and a memory connected with the processor in communication;

[0023] The memory stores computer-executable instructions.

[0024] The processor executes the computer-executable instructions stored in the memory, to implement the document data processing method of the first aspect.

[0025] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer-executable instructions, and the computer-executable instructions, when executed by a processor, implement the document data processing method of the first aspect.

[0026] In a fifth aspect, the present application provides a computer program product, comprising a computer program, and the computer program, when executed by a processor, implements the document data processing method of the first aspect.

[0027] This application provides a document data processing method, apparatus, device, and medium, comprising: first, acquiring a pre-built vector library and documents to be reviewed, wherein the pre-built vector library is constructed based on sample documents; then, performing document segmentation processing according to the documents to be reviewed and preset segmentation rules to obtain multiple segments to be reviewed; next, performing search processing in the pre-built vector library according to each segment to be reviewed to obtain a first reference document name list composed of the names of sample documents corresponding to each segment to be reviewed; subsequently, adjusting the first reference document name list in response to user review adjustment operations to obtain a second reference document name list composed of the names of sample documents corresponding to each segment to be reviewed; further, determining the reference source knowledge of each segment to be reviewed based on the second reference document name list, and adjusting the knowledge point scope of the reference source knowledge of each segment to be reviewed to obtain the target reference source knowledge of each segment to be reviewed; then, inputting the target reference source knowledge of each segment to be reviewed, each segment to be reviewed, and fixed prompt words into a pre-trained knowledge retrieval model to obtain the review result of each segment to be reviewed; finally, responding to user result processing operations on the review results of each segment to be reviewed to obtain the target review result. The following technical effects were achieved: By splitting the document to be reviewed into multiple segments and searching for each segment in a pre-built vector library, a list of corresponding reference document names was obtained. Then, within the scope corresponding to the reference document name list, target reference source knowledge was searched. Finally, a pre-trained knowledge retrieval model was used to obtain the review results for each segment. This narrowed the search scope, improving review accuracy and efficiency while reducing human intervention in the document content review process. By determining a second list of reference document names and then searching for reference source knowledge for each segment within the knowledge scope corresponding to the sample document names in the second list, a chapter-level search scope constraint was achieved. Finally, based on the chapter scope corresponding to the second list of reference document names, the reference source knowledge for each segment was searched in the pre-built vector library, improving the accuracy and efficiency of reference source knowledge search. By using only minimal human intervention in each step, the review process was made controllable, reducing user workload and significantly lowering the cost of human involvement in the document content review process. This allows for the rapid processing of large-scale, multi-type document content review tasks. By allowing users to adjust the search results for chapter and knowledge point searches, the accuracy of the review results was further improved. Attached Figure Description

[0028] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0029] Figure 1 This is a flowchart illustrating a document data processing method provided in an embodiment of this application.

[0030] Figure 2 A structural schematic diagram of a document data processing apparatus provided in an embodiment of the present application.

[0031] Figure 3 A structural schematic diagram of an electronic device provided in an embodiment of the present application.

[0032] The specific embodiments of the present application have been shown through the above-described drawings, and will be described in more detail hereinafter. These drawings and the following description are not intended to limit the scope of the present application in any way, but merely to explain the concept of the present application to those skilled in the art by reference to a specific embodiment. DETAILED DESCRIPTION

[0033] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The same reference numbers in different drawings represent the same or similar elements. The following detailed description is not intended to limit the scope of the present application, but merely to explain the concepts of the present application by reference to some embodiments in accordance with the appended claims.

[0034] In the embodiments of the present application, the terms "first", "second", and the like are used to distinguish between similar or identical items or elements having substantially the same function and role. Those skilled in the art can understand that the terms "first", "second", and the like do not limit the number and execution order, and the terms "first", "second", and the like do not necessarily mean different. It should be noted that in the embodiments of the present application, the words "exemplary" or "for example" are used to indicate an example, illustration, or description. Any embodiment or design solution described as "exemplary" or "for example" in the present application should not be interpreted as more preferred or more advantageous than other embodiments or design solutions. Rather, the words "exemplary" or "for example" are used to present the relevant concept in a specific manner. In the embodiments of the present application, "at least one" means one or more, and "multiple" means two or more.

[0035] It should be noted that "at" in the embodiments of the present application can be at the moment when a certain condition occurs, or can be within a period of time after a certain condition occurs, which is not specifically limited in the embodiments of the present application. In addition, the document data processing method provided in the embodiments of the present application is only an example, and a document data processing method can also include more or less content.

[0036] It should be noted that the document data processing method, device, equipment and medium provided by the present application can be used in the field of artificial intelligence technology, and can also be used in any field other than the field of artificial intelligence technology. The application of the document data processing method, device, equipment and medium of the present application is not limited.

[0037] Document content review is crucial for ensuring the accuracy of document information. By accurately identifying and correcting potential problems, it can provide solid support for effective decision-making and avoid misleading or decision-making errors caused by false information.

[0038] In the prior art field, document content review is mainly realized by means of natural language processing technology, machine learning algorithm and deep learning model for automatic processing. Among them, the natural language processing technology relies on the core capabilities of word segmentation, part-of-speech tagging, and syntax analysis; and the machine learning algorithm and deep learning model need to build a classification model or a complex neural network structure based on a large-scale training set to realize the review function.

[0039] However, the current document content review method faces significant challenges in pursuing high review accuracy. To ensure review quality, a large amount of manpower is usually required for data annotation and complex model training, which results in high review cost.

[0040] Therefore, the embodiments of the present application provide a document data processing method, device, equipment and medium, which can be used in the field of artificial intelligence technology, to solve the above technical problems of the prior art. The method first splits the document to be reviewed into multiple independent segments to be reviewed; then, in the pre-constructed knowledge vector library, each segment to be reviewed is efficiently searched to obtain a corresponding reference document name list, and then the target reference source knowledge corresponding to each segment to be reviewed is accurately obtained within the range corresponding to the reference document name list; then, by means of a pre-trained knowledge retrieval large model, each segment is analyzed in depth to obtain the review result of each segment to be reviewed; finally, according to the review result of each segment to be reviewed, the corresponding target review result is generated. By granulating the document content and combining the capabilities of the knowledge vector library and the large model, the search range is effectively reduced, not only the review accuracy is improved, but also the cost of human participation in the document content review process is significantly reduced, which can quickly process large-scale and multi-type document content review tasks.

[0041] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail in the specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.

[0042] Figure 1 A flowchart of a document data processing method provided in an embodiment of the present application is shown. As shown in the figure, the method comprises: Figure 1

[0043] S101, obtaining a pre-built vector library and a to-be-audited document.

[0044] In an embodiment of the present application, the execution subject of a document data processing method can be a data processing server or other device with data processing capability, such as a smart phone, a notebook computer, a personal computer, a tablet computer, etc. The data processing server can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, or a cloud server, etc., which is not specifically limited here. For the convenience of description, the execution subject of a document data processing method is uniformly described as a server in the embodiments of the present application.

[0045] The pre-built vector library is constructed based on sample documents and is a knowledge vectorization database.

[0046] Specifically, before auditing the to-be-audited document, the server can first construct a knowledge vectorization database according to multiple sample documents to support subsequent knowledge point matching and reference source retrieval, and provide an important reference basis for subsequent document content auditing.

[0047] S102, performing document splitting processing on the to-be-audited document according to a preset splitting rule to obtain multiple to-be-audited segments.

[0048] Specifically, since the to-be-audited document is usually long, the overall auditing will exceed the length limit of the large model input parameter, and will result in a relatively rough auditing effort, which cannot meet the actual use requirements of the auditing. Therefore, the server can first split the to-be-audited document according to a preset splitting rule (such as by paragraphs, chapters, fixed number of words, etc.) to split it into multiple independent to-be-audited segments, thereby facilitating subsequent fine-grained processing.

[0049] S103, searching each to-be-audited segment in the pre-built vector library to obtain a first reference document name list composed of the names of the sample documents corresponding to each to-be-audited segment.

[0050] Specifically, for each to-be-audited segment, the server can perform search processing in the pre-built vector library. After the search, the name of the sample document corresponding to each to-be-audited segment is obtained, and these names are combined into a first reference document name list.

[0051] S104, adjusting the first reference document name list in response to a user auditing adjustment operation to obtain a second reference document name list composed of the names of the sample documents corresponding to each to-be-audited segment. ​

[0052] Specifically, when the user finds that the names of the sample documents in the first reference document name list are redundant or lack, the user can input first adjustment information to the server to delete the redundant names of the sample documents or add the names of the missing sample documents by combining the user's personal experience.

[0053] The server can delete the redundant names of the sample documents in the first reference document name list and add the names of the missing sample documents in the first reference document name list according to the first adjustment information input by the user in response to the user's review adjustment operation, so as to generate a second reference document name list composed of the names of the sample documents corresponding to each to-be-reviewed segment after the adjustment of the first reference document name list is completed.

[0054] S105, determine the reference source knowledge of each to-be-reviewed segment according to the second reference document name list, and perform knowledge point range adjustment processing on the reference source knowledge of each to-be-reviewed segment to obtain target reference source knowledge of each to-be-reviewed segment.

[0055] In the embodiments of the present application, the server can search the reference source knowledge of each to-be-reviewed segment in the corresponding knowledge range in the pre-built vector library according to the knowledge range determined by the second reference document name list.

[0056] Specifically, the server can search the reference source knowledge of each to-be-reviewed segment in the knowledge range corresponding to each sample document name in the second reference document name list in the pre-built vector library. The server can obtain a reference source knowledge list according to all the searched reference source knowledge, and then display the reference source knowledge list to the user through a visual interface. When the user finds that the reference source knowledge in the reference source knowledge list is redundant or lacks, the user can input second adjustment information to the server to delete the redundant reference source knowledge in the reference source knowledge list or add the missing reference source knowledge by combining the user's personal experience.

[0057] Then, the server can delete the redundant reference source knowledge in the reference source knowledge list and add the missing reference source knowledge in the reference source knowledge list according to the second adjustment information input by the user in response to the user's adjustment operation, so as to generate the target reference source knowledge of each to-be-reviewed segment after the knowledge point range adjustment of the reference source knowledge of each to-be-reviewed segment is completed.

[0058] By determining the second reference document name list, and then searching for the reference source knowledge of each to-be-audited segment within the knowledge range corresponding to the sample document name in the second reference document name list, the search range at the chapter level is constrained, and finally the reference source knowledge of each to-be-audited segment is searched in the pre-built vector library based on the chapter range corresponding to the second reference document name list, thereby improving the search accuracy and search efficiency of the reference source knowledge.

[0059] S106, input the target reference source knowledge of each to-be-audited segment, each to-be-audited segment and the fixed prompt word into the pre-trained knowledge retrieval large model to obtain the audit result of each to-be-audited segment.

[0060] Specifically, after determining the target reference source knowledge of each to-be-audited segment, the server can input each to-be-audited segment and the corresponding target reference source knowledge, and the fixed prompt word into the pre-trained knowledge retrieval large model, so as to output the audit result of each segment through the pre-trained knowledge retrieval large model.

[0061] Among them, the fixed prompt word refers to a preset prompt, which refers to a piece of text description set in advance by the user, or a parameter description in a certain format; the pre-trained knowledge retrieval large model refers to a deep neural network model, which has super strong semantic understanding and information generation capability.

[0062] S107, in response to the result processing operation of the user on the audit result of each to-be-audited segment, obtaining a target audit result.

[0063] Specifically, the server can show the user the audit result of each to-be-audited segment through a visual interface. The user can choose to directly adopt the audit result of each to-be-audited segment, or to modify part of the content of the audit result and then adopt it.

[0064] The server can generate a target audit result according to the result processing operation of the user on the audit result of each to-be-audited segment, and finally export the audit result to form a complete audit report.

[0065] The document data processing method provided by the embodiment of the application can effectively improve the accuracy and efficiency of document content auditing, and reduce the cost of manual participation.

[0066] The document data processing method provided by the embodiment comprises the following steps: firstly, a pre-built vector library and a to-be-audited document are acquired, wherein the pre-built vector library is built based on sample documents; secondly, the to-be-audited document is subjected to document splitting processing according to a preset splitting rule, thereby obtaining a plurality of to-be-audited segments; thirdly, the to-be-audited segments are subjected to searching processing in the pre-built vector library, thereby obtaining a first reference document name list composed of the names of the sample documents corresponding to the to-be-audited segments; fourthly, the first reference document name list is adjusted in response to a user audit adjustment operation, thereby obtaining a second reference document name list composed of the names of the sample documents corresponding to the to-be-audited segments; fifthly, the reference source knowledge of the to-be-audited segments is determined according to the second reference document name list, and the reference source knowledge of the to-be-audited segments is subjected to knowledge point range adjustment processing, thereby obtaining target reference source knowledge of the to-be-audited segments; sixthly, the target reference source knowledge of the to-be-audited segments, the to-be-audited segments and a fixed prompt word are input into a pre-trained knowledge retrieval large model, thereby obtaining an audit result of the to-be-audited segments; and finally, the target audit result is obtained in response to a result processing operation of the user on the audit result of the to-be-audited segments.

[0067] The following technical effects are achieved: the to-be-audited document is split into a plurality of to-be-audited segments, each to-be-audited segment is searched in the pre-built vector library to obtain a corresponding reference document name list, the target reference source knowledge is searched within the range corresponding to the reference document name list, and then the pre-trained knowledge retrieval large model is used to obtain the audit result of each to-be-audited segment, thereby narrowing the search range, improving the audit accuracy and efficiency, and reducing the human input in the document content audit process; the second reference document name list is determined, the reference source knowledge of each to-be-audited segment is searched within the knowledge range corresponding to the sample document name in the second reference document name list, the search range at the chapter level is constrained, and finally the reference source knowledge of each to-be-audited segment is searched in the pre-built vector library within the chapter range corresponding to the second reference document name list, thereby improving the search accuracy and efficiency of the reference source knowledge; a small amount of manual intervention is used in each step, the effect of the whole-process audit is controllable, and the workload of the user is reduced, thereby significantly reducing the cost of manual participation in the document content audit process, and large-scale and multi-type document content audit tasks can be quickly processed; the user adjusts the search results of the chapter search and the knowledge point search, thereby further improving the accuracy of the audit result.

[0068] In a possible implementation, the preset splitting rule is splitting by chapter, splitting by directory outline structure, or splitting by specific word number; accordingly, in the step S202, the document splitting processing is performed according to the to-be-audited document and the preset splitting rule to obtain the plurality of to-be-audited segments, including: splitting the to-be-audited document by chapter to obtain a plurality of to-be-audited segments each in the unit of chapter, section, or clause; or splitting the to-be-audited document by directory outline structure to obtain a plurality of to-be-audited segments each in the unit of directory page range; or splitting the to-be-audited document by specific word number to obtain a plurality of to-be-audited segments each containing specific word number.

[0069] Specifically, the splitting by chapter is applicable to a to-be-audited document with clear structure and explicit chapter, section, and clause numbers (such as laws and regulations, technical standards, rules and regulations, etc.). The specific splitting manner includes: using a regular expression to identify the chapter and clause structure of the to-be-audited document. Taking a chapter as an independent segment, further dividing each chapter by section, and refining each clause to facilitate clause-by-clause auditing. For example, splitting a paper into chapter segments of introduction, related work, method, experiment, and conclusion. By splitting the to-be-audited document by chapter, the plurality of to-be-audited segments obtained are clear in order, thereby facilitating clause-by-clause auditing.

[0070] The splitting by directory outline structure is applicable to a to-be-audited document with detailed directory structure but not necessarily organized by chapter and clause (such as research reports, academic papers, project plans, etc.). The specific splitting manner is to extract the bookmark structure in the to-be-audited document or parse the title hierarchy of the to-be-audited document, and split the to-be-audited document into a plurality of to-be-audited segments according to the directory page range. By splitting the to-be-audited document by directory outline structure, the to-be-audited segments obtained maintain the integrity of the logical structure, are suitable for chapter content with strong integrity, and are easy to locate the specific position of each to-be-audited segment.

[0071] The splitting by specific word number is applicable to a to-be-audited document without obvious structure or requiring uniform length processing (such as contract texts, translation materials, artificial intelligence generated content, etc.). The specific splitting manner is to use a string splitting function to divide the to-be-audited document into a plurality of to-be-audited segments by a fixed number of words (such as every 500 words). Optionally, an overlapping interval (such as the last 100 words of a to-be-audited segment and the first 100 words of the next to-be-audited segment) can be further set, and a sliding window algorithm can be used to process the overlapping content to avoid semantic fragmentation. By splitting the to-be-audited document by specific word number, the plurality of to-be-audited segments obtained are uniform in length, avoiding the bias caused by human judgment of the structure of the to-be-audited document.

[0072] The document data processing method provided in the embodiment of the application can flexibly adapt to different types of to-be-audited documents by splitting the to-be-audited document into multiple to-be-audited segments according to chapters, or according to a table of contents outline structure, or according to a specific number of words, can perform standardized splitting processing according to the content structure or the number of words of the to-be-audited document, and thus facilitates subsequent search processing in a pre-built vector library according to each to-be-audited segment, and improves overall search efficiency.

[0073] In a possible implementation, in the step S202, the document splitting processing is performed according to the to-be-audited document and the preset splitting rule to obtain the multiple to-be-audited segments, and the step can further include: sequentially parsing and splitting the to-be-audited document according to a preset priority from high to low in the order of chapter, table of contents outline, fixed-length field based on a preset sliding window, and preset fixed-length word number based on a paragraph to obtain the multiple to-be-audited segments.

[0074] Specifically, when splitting the to-be-audited document, the server can sequentially attempt different splitting rules in a preset priority order from high to low, and once a certain splitting rule is met, the splitting mode corresponding to the splitting rule is executed, thereby avoiding conflicts or repeated splitting between multiple splitting rules.

[0075] Specifically, the highest priority is to split the to-be-audited document according to a chapter section item structure. If the server detects that the to-be-audited document has a clear chapter section item structure, the to-be-audited document is preferentially split into multiple to-be-audited segments according to chapters, sections, and items. If the to-be-audited document does not have a clear chapter section item structure but has a complete directory structure (such as a PDF bookmark or a Word title level), the to-be-audited document is split according to a directory page number or a level range. If neither of the first two structures exists in the to-be-audited document, a fixed number of words (such as 500 words) can be used in combination with a sliding window technique (allowing partial overlap) to split the to-be-audited document, that is, the to-be-audited document is split using string cutting in combination with an overlapping interval algorithm. Finally, the last splitting rule is to split the to-be-audited document according to a natural paragraph. If a single paragraph exceeds a preset maximum number of words (such as 800 words), the to-be-audited document is further split, that is, the to-be-audited document is split according to a line feed to divide a paragraph in combination with word number analysis.

[0076] The document data processing method provided in the embodiment of the application can flexibly adapt to different types of to-be-audited documents by splitting the to-be-audited document into multiple to-be-audited segments according to chapters, or according to a table of contents outline structure, or according to a specific number of words, can perform standardized splitting processing according to the content structure or the number of words of the to-be-audited document, and thus facilitates subsequent search processing in a pre-built vector library according to each to-be-audited segment, and improves overall search efficiency.

[0077] In a possible implementation, in step S203, the search processing is performed in the pre-built vector library according to each to-be-audited segment, to obtain a first reference document name list composed of names of sample documents corresponding to each to-be-audited segment, including: performing iterative search in the pre-built vector library according to all to-be-audited segments, to obtain a first quantity of knowledge data in a similarity ranking of each to-be-audited segment; performing reordering processing according to the first quantity of knowledge data, to obtain a second quantity of knowledge data in a score ranking, where the second quantity is less than the first quantity; and performing analysis and deduplication processing on the second quantity of knowledge data corresponding to each to-be-audited segment, to obtain the first reference document name list.

[0078] In the embodiments of the present application, the first quantity can be 30, 50, or 60, and the second quantity can be 10, 5, or 3, which are not limited herein.

[0079] Specifically, the server can convert each to-be-audited segment into a vector form using the same vectorization model as the pre-built vector library. For the vector of each to-be-audited segment, the similarity between the vector and all vectors in the pre-built vector library is calculated, which can be calculated by a similarity calculation method such as cosine similarity, Euclidean distance, etc. Then, the knowledge data in the vector library is sorted in descending order of similarity, and the first quantity (for example, 50) of knowledge data with the highest similarity is selected from the sorted results as the preliminary matching result of the to-be-audited segment.

[0080] Since the first quantity of knowledge data obtained for each to-be-audited segment may have a local optimal condition, that is, a certain knowledge data has a high similarity in the matching of a single to-be-audited segment, but it may not be the most relevant one from the whole. Through reordering processing, the similarity relationship between all to-be-audited segments and knowledge data can be considered comprehensively, so as to obtain more accurate and more relevant knowledge data. Specifically, the similarity of each knowledge data in all to-be-audited segments is summed, averaged, or weighted averaged, etc. to calculate the comprehensive score of each knowledge data. For example, using the summation method, for a knowledge data, the similarity in the matching result of all to-be-audited segments can be added to obtain the comprehensive score of the knowledge data. Then, according to the calculated comprehensive score, all preliminary matching knowledge data is reordered, and the second quantity (for example, 3) of knowledge data with the highest comprehensive score is selected from the reordered results as the final matching result.

[0081] Since the same knowledge data may be matched with multiple to-be-audited segments in the preliminary matching and reordering process, resulting in duplicate document names in the result list. Therefore, the server can extract the document names from the first second number of knowledge data corresponding to each to-be-audited segment, and use a data structure (such as a set) or an algorithm (such as hash deduplication) to remove duplicate document names. And arrange the deduplicated document names in a certain order (such as alphabetical order, frequency of occurrence, etc.) to generate the final first reference document name list.

[0082] By analyzing and deduplicating the knowledge data, duplicate document names can be removed, thereby obtaining a unique and non-duplicate first reference document name list.

[0083] Optionally, the reordering process according to the first first number of knowledge data can also be implemented by using a fine sorting model. The fine sorting model can be a Bidirectional Encoder Representations from Transformers (BERT) model, a Text-to-Text Transfer Transformer (T5) model, or a Contextualized Late Interaction over BERT (ColBERT) model, etc.

[0084] The document data processing method provided by the embodiments of the present application can obtain more accurate and relevant knowledge data by iteratively searching all to-be-audited segments in the pre-built vector library and considering the similarity relationship between all to-be-audited segments and knowledge data through reordering processing; and by analyzing and deduplicating the knowledge data, duplicate document names can be removed, thereby obtaining a unique and non-duplicate first reference document name list.

[0085] Next, the complete process of obtaining the pre-built vector library is described in detail.

[0086] In one possible implementation, in the step S201, obtaining the pre-built vector library includes: obtaining a sample document; performing document splitting processing according to the sample document and a preset splitting rule to obtain a sample knowledge segment; and forming the pre-built vector library according to the sample knowledge segment and a preset library table structure.

[0087] Specifically, the server can obtain public text data as sample documents from public data platforms such as academic databases, news websites, open source project platforms, etc. For example, when studying the similarity of texts in the medical field, medical papers can be downloaded from a medical literature database (PubMed) as sample documents. Alternatively, the user can input product manuals, research reports, etc. to the server as sample documents.

[0088] Then, the server can organize the collected sample documents in a certain format, such as saving them uniformly in TXT, PDF, or Word format, etc. For documents in different formats, use the corresponding parsing tools for conversion. And can use data cleaning techniques such as regular expressions, natural language processing tools, etc. to preprocess the sample documents to remove noise and error information. At the same time, the sample documents can also be de-duplicated to avoid interference of repeated data on subsequent processing. For example, a hash algorithm can be used to calculate the hash value of each sample document to determine whether the sample documents are repeated by comparing the hash values.

[0089] Next, the server can perform document splitting processing on the sample documents according to the preset splitting rules to obtain sample knowledge fragments. Specifically, the document splitting processing according to the sample documents and the preset splitting rules is similar to the above-mentioned document splitting processing according to the to-be-audited documents and the preset splitting rules, and the splitting method is similar, which will not be repeated here.

[0090] Finally, the server can convert the sample knowledge fragments into vector representations according to the preset vectorization model. The preset vectorization model can be a Word to Vector (Word2Vec) model, a Global Vectors for Word Representation (GloVe) model, or a BERT model, etc. For the Word2Vec model or the GloVe model, the server can average or weighted average the vectors of all words in the sample knowledge fragments to obtain the vector representation of the sample knowledge fragments; for the BERT model, the server can directly use the sentence vector output by the BERT model as the vector representation of the sample knowledge fragments.

[0091] The pre-built vector library can store the vector representation of the sample knowledge fragments, the text content of the fragments, the source document information, etc. Table 1 is a mapping relationship table between vector library information and document information provided by the embodiments of the present application.

[0092] Table 1: Mapping relationship table between vector library information and document information

[0093]

[0094] As shown in Table 1, the vector library index field title corresponds to the knowledge point title; the vector library index field content corresponds to the knowledge point content; and the vector library index field fileName corresponds to the document name. The preset library table structure includes the table header of the table and the content of each table header, and the mapping relationship between the vector library information and the document information in the table.

[0095] The server can import the vectorized sample knowledge segments into the specific indexes of the vector library according to the preset library table structure, to form a pre-built vector library. The index field, the index field meaning, the content of the corresponding document knowledge, and the specific role are shown in Table 1.

[0096] In a possible implementation, the step of obtaining the pre-built vector library can be a continuous action. The user can continuously import sample documents into the pre-built vector library according to a preset period, and timely remove expired and invalid document knowledge from the pre-built vector library, to ensure the correctness and real-time performance of the knowledge in the pre-built vector library. The preset period can be adjusted according to different fields and application scenarios, for example, the pre-built vector library is updated daily, weekly, or monthly, and the like, which is not limited here.

[0097] The document data processing method provided by the embodiment of the application can be used to build a pre-built vector library by importing sample documents into the vector library. Subsequently, the user only needs to continuously import sample documents into the pre-built vector library according to a preset period, and timely remove expired and invalid document knowledge from the pre-built vector library, and pay attention to timely maintenance of the pre-built vector library, to ensure the correctness and real-time performance of the knowledge in the pre-built vector library. Therefore, it is not necessary to invest a large amount of manpower in data labeling and model training, and the use threshold of document content review is reduced.

[0098] In a possible implementation, in the step S207, in response to the result processing operation of the user on the review results of each to-be-reviewed segment, the target review result is obtained, including: in response to the result judgment operation of the user on the review results of each to-be-reviewed segment, a result judgment result is obtained; when it is detected that the result judgment result is direct adoption, the review results of each to-be-reviewed segment are directly combined and processed to obtain the whole review result of the to-be-reviewed document as the target review result; when it is detected that the result judgment result is partial modification and adoption, in response to the result modification operation of the user on the review results of each to-be-reviewed segment, a modified review result of the to-be-reviewed document is obtained as the target review result; and when it is detected that the result judgment result is re-review, in response to the reference source knowledge adjustment operation of the user, an adjusted target reference source knowledge is obtained, and the adjusted target reference source knowledge, each to-be-reviewed segment, and a fixed prompt word are input into the pre-trained knowledge retrieval large model to obtain a secondary review result of each to-be-reviewed segment.

[0099] Specifically, the server can further guide the user to judge the audit results of each to-be-audited segment after completing the preliminary audit of all to-be-audited segments.

[0100] The user first judges the audit result of each to-be-audited segment to obtain a result judgment result. This judgment is based on the user's own knowledge, experience, and reference source knowledge provided by the server, and other factors. The types of result judgment results include: direct adoption, that is, the user thinks that the audit result of the to-be-audited segment is completely correct and does not need to be modified; partial modification and adoption, that is, the user thinks that the audit result of the to-be-audited segment is partially correct, but there are some places that need to be modified, such as incorrect citation format, inaccurate expression of part of the viewpoint, etc.; re-audit, that is, the user thinks that the audit result of the to-be-audited segment has a big problem, or the reference source knowledge is insufficient, and needs to re-acquire the reference source knowledge and perform the audit.

[0101] When it is detected that the result judgment result is direct adoption, the server can perform a merging processing on the audit results of the to-be-audited segments. The merging manner can be to sequentially splice the audit results of each to-be-audited segment according to the order of the to-be-audited segments in the to-be-audited document, to form a complete audit report. The complete audit result of the to-be-audited document after merging is the target audit result. For example, when auditing a contract document, if the audit results of all segments are directly adopted, then the merged audit result is a complete contract audit pass report.

[0102] When it is detected that the result judgment result is partial modification and adoption, the server can guide the user to modify the audit results of each to-be-audited segment. The user can modify the incorrect part in the audit result according to his own judgment, such as modifying the incorrect citation, adjusting the inaccurate viewpoint, etc. After the user completes the modification, the server can take the modified audit result as the modified audit result of the to-be-audited document, that is, the target audit result. For example, when auditing a news release, the user finds that the citation source of part of the segment is incorrect, and the modified audit result obtained after modification is the target audit result.

[0103] When the result judgment result is re-audit, the server can first guide the user to adjust the reference source knowledge. The user can add, delete or modify the reference source knowledge according to the user's own needs to obtain more accurate and comprehensive reference source knowledge. After adjusting the reference source knowledge, the server can input the adjusted target reference source knowledge, each to-be-audited segment and the fixed prompt word into the pre-trained knowledge retrieval large model. The pre-trained knowledge retrieval large model can re-audit the to-be-audited segment according to the input information, and output the secondary audit result of each to-be-audited segment. The secondary audit result can be used as the target audit result. For example, when auditing a scientific research paper, if the user thinks that the reference source knowledge of the initial audit is insufficient, the secondary audit result obtained by adjusting and re-inputting the model for auditing is the target audit result.

[0104] The document data processing method provided in the embodiments of the present application obtains the target audit result according to the result processing operation of the user on the audit result of each to-be-audited segment, and further ensures the accuracy of the audit result.

[0105] In a possible implementation, the document data processing method described above further includes, after step S203, when it is detected that there is no user input adjustment information, determining the reference source knowledge of each to-be-audited segment according to the first reference document name list, and performing knowledge point range adjustment processing on the reference source knowledge of each to-be-audited segment to obtain the target reference source knowledge of each to-be-audited segment.

[0106] Specifically, if the user finds that there is no need to adjust the first reference document name list, the server can directly skip the step of adjusting the search range at the document level, and directly determine the reference source knowledge of each to-be-audited segment according to the first reference document name list.

[0107] That is, adjusting the search range at the document level to obtain the second reference document name list is not a necessary step. If the user thinks that the first reference document name list obtained by the initial automatic search is already the correct document range, the user can skip the adjustment and directly search the knowledge points based on the document range corresponding to the first reference document name list to obtain the reference source knowledge of each to-be-audited segment.

[0108] The document data processing method provided in the embodiments of the present application directly determines the reference source knowledge of each to-be-audited segment according to the first reference document name list when the user thinks that there is no need to adjust the first reference document name list, thereby improving the audit efficiency.

[0109] Figure 2 A structural schematic diagram of a document data processing device provided in the embodiments of the present application is shown in FIG. 1. Figure 2As shown, the document data processing apparatus comprises: an acquisition module 210, a document splitting module 220, a vector search module 230, a document review module 240, and a result output module 250.

[0110] The acquisition module 210 is configured to acquire a pre-built vector library and a document to be reviewed, wherein the pre-built vector library is built based on sample documents.

[0111] The document splitting module 220 is configured to perform document splitting processing on the document to be reviewed according to a preset splitting rule, to obtain a plurality of to-be-reviewed segments.

[0112] The vector search module 230 is configured to perform search processing on each to-be-reviewed segment in the pre-built vector library, to obtain a first reference document name list composed of names of sample documents corresponding to each to-be-reviewed segment.

[0113] The document review module 240 is configured to adjust the first reference document name list in response to a user review adjustment operation, to obtain a second reference document name list composed of names of sample documents corresponding to each to-be-reviewed segment.

[0114] The document review module 240 is further configured to determine reference source knowledge of each to-be-reviewed segment according to the second reference document name list, and perform knowledge point range adjustment processing on the reference source knowledge of each to-be-reviewed segment, to obtain target reference source knowledge of each to-be-reviewed segment.

[0115] The document review module 240 is further configured to input the target reference source knowledge of each to-be-reviewed segment, each to-be-reviewed segment, and a fixed prompt word into a pre-trained knowledge retrieval large model, to obtain a review result of each to-be-reviewed segment.

[0116] The result output module 250 is configured to obtain a target review result in response to a user result processing operation on the review result of each to-be-reviewed segment.

[0117] In a possible implementation, the preset splitting rule is splitting by chapter, splitting by directory outline structure, or splitting by specific number of words; correspondingly, the document splitting module 220 is further configured to split the document to be reviewed into a plurality of to-be-reviewed segments in units of chapters, sections, and items according to the chapter, section, and item; or split the document to be reviewed into a plurality of to-be-reviewed segments in units of directory page ranges according to the directory outline structure; or split the document to be reviewed into a plurality of to-be-reviewed segments containing specific numbers of words.

[0118] In a possible implementation, the document splitting module 220 is further configured to sequentially parse and split the document to be reviewed according to a preset priority from high to low in the order of chapter, directory outline structure, fixed-length field based on a preset sliding window, and preset fixed-length number of words based on a paragraph, to obtain the plurality of to-be-reviewed segments.

[0119] In a possible implementation, the vector search module 230 comprises an iterative search module, a reordering module, and a deduplication module.

[0120] The iterative search module is configured to perform iterative search on all to-be-audited segments in the pre-built vector library to obtain a first quantity of knowledge data similar to each to-be-audited segment.

[0121] The reordering module is configured to perform reordering processing on the first quantity of knowledge data to obtain a second quantity of knowledge data in score order, where the second quantity is less than the first quantity.

[0122] The deduplication module is configured to perform analysis and deduplication processing on the second quantity of knowledge data corresponding to each to-be-audited segment to obtain a first reference document name list.

[0123] In a possible implementation, the obtaining module 210 comprises a sample document obtaining module, a sample document splitting module, and a pre-built vector library module.

[0124] The sample document obtaining module is configured to obtain a sample document.

[0125] The sample document splitting module is configured to perform document splitting processing on the sample document according to a preset splitting rule to obtain a sample knowledge segment.

[0126] The pre-built vector library module is configured to form a pre-built vector library according to the sample knowledge segment and a preset library table structure.

[0127] In a possible implementation, the result output module 250 comprises a result judgment module and a detection module.

[0128] The result judgment module is configured to obtain a result judgment result in response to a result judgment operation of a user on the audit results of the to-be-audited segments.

[0129] The detection module is configured to, when it is detected that the result judgment result is direct adoption, directly perform merging processing on the audit results of the to-be-audited segments to obtain an entire audit result of the to-be-audited document as a target audit result.

[0130] The detection module is further configured to, when it is detected that the result judgment result is partial modification and adoption, obtain a modified audit result of the to-be-audited document as the target audit result in response to a result modification operation of a user on the audit results of the to-be-audited segments.

[0131] The detection module is further configured to, when the result of the detection is that the result of the judgment is re-audit, obtain adjusted target reference source knowledge in response to a reference source knowledge adjustment operation of a user, and input the adjusted target reference source knowledge, the to-be-audited segments, and the fixed prompt word into the pre-trained knowledge retrieval large model to obtain secondary audit results of the to-be-audited segments.

[0132] In a possible implementation, the document audit module 240 is further configured to, when it is detected that there is no input adjustment information of the user, determine reference source knowledge of the to-be-audited segments according to the first reference document name list, and perform knowledge point range adjustment processing on the reference source knowledge of the to-be-audited segments to obtain target reference source knowledge of the to-be-audited segments.

[0133] The document data processing apparatus provided in this embodiment can execute the document data processing method provided in the above-described embodiments, and has similar implementation principles and technical effects. Details are not described herein again.

[0134] In the specific implementation of the above-described document data processing apparatus, each module and unit can be implemented as a processor. The processor can execute computer execution instructions stored in a memory, so that the processor executes the above-described document data processing method.

[0135] Figure 3 A structural schematic diagram of an electronic device provided in an embodiment of the present application is shown in FIG. 3. As shown in FIG. 3, the electronic device includes at least one processor 310 and a memory 320. The electronic device further includes a communication component 330. The processor 310, the memory 320, and the communication component 330 are connected through a bus 340. Figure 3

[0136] In the specific implementation process, the at least one processor 310 executes computer execution instructions stored in the memory 320, so that the at least one processor 310 executes the above-described document data processing method executed by the electronic device.

[0137] The specific implementation process of the processor 310 can refer to the above-described method embodiments, and has similar implementation principles and technical effects. Details are not described herein again.

[0138] ​In the above embodiments, it should be understood that the processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The steps of the method disclosed in combination with the application can be directly embodied as hardware processor execution, or executed by a combination of hardware and software modules in the processor.

[0139] The memory can include a high-speed RAM memory, and can also include a non-volatile storage NVM, such as at least one disk memory.

[0140] The bus can be an industry standard architecture (ISA) bus, a peripheral component (PCI) bus, or an extended industry standard architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the sake of representation, the bus in the drawings of the present application does not limit to only one bus or one type of bus.

[0141] The functions realized by the electronic device and the master device described above are introduced for the scheme provided by the embodiments of the present application. It can be understood that the electronic device or the master device contains the hardware structure and / or software module corresponding to the execution of each function in order to realize the above functions. The units and algorithm steps of each example described in combination with the embodiments disclosed in the embodiments of the present application can be realized in the form of hardware or a combination of hardware and computer software. Whether a certain function is executed by hardware or computer software driven hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the technical solution of the embodiments of the present application.

[0142] The present application also provides a computer readable storage medium, the computer readable storage medium stores computer execution instructions, when the processor executes the computer execution instructions, for realizing the above document data processing method.

[0143] The readable storage medium described above can be realized by any type of volatile or nonvolatile storage devices or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general or special purpose computer.

[0144] An exemplary readable storage medium is coupled to the processor, so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be an integral part of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the readable storage medium can also exist as discrete components in an electronic device or a host device.

[0145] The present application also provides a computer program product, the computer program product comprising a computer program stored in a readable storage medium, at least one processor of an electronic device can read the computer program from the readable storage medium, and the at least one processor executes the computer program to enable the electronic device to perform the scheme provided by the above-mentioned embodiments.

[0146] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware. The foregoing program can be stored in a computer readable storage medium. The program executes to perform the steps of the above-mentioned method embodiments; and the foregoing storage medium includes ROM, RAM, magnetic disk or optical disk and various storage medium that can store program codes.

[0147] So far, the technical scheme of the present application has been described in combination with the preferred embodiments shown in the drawings, but those skilled in the art can easily understand that the protection scope of the present application is obviously not limited to these specific embodiments, and the above embodiments are only used to illustrate the technical scheme of the present application, but not to limit it; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical scheme recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical scheme deviate from the scope of the technical scheme of the embodiments of the present application.

Claims

1. A document data processing method, characterized in that, include: Obtain a pre-built vector library and documents to be reviewed, wherein the pre-built vector library is constructed based on sample documents; The document is split according to the document to be reviewed and the preset splitting rules to obtain multiple segments to be reviewed. Based on the search processing performed on each segment to be reviewed in the pre-built vector library, a first reference document name list composed of the names of the sample documents corresponding to each segment to be reviewed is obtained. In response to the user's review and adjustment operation, the first reference document name list is adjusted to obtain a second reference document name list composed of the names of sample documents corresponding to each segment to be reviewed; Based on the second list of reference document names, the reference source knowledge for each segment to be reviewed is determined, and the knowledge point range of the reference source knowledge for each segment to be reviewed is adjusted to obtain the target reference source knowledge for each segment to be reviewed. The target reference source knowledge of each segment to be reviewed, each segment to be reviewed, and fixed prompt words are input into a pre-trained knowledge retrieval model to obtain the review results of each segment to be reviewed. In response to the user's result processing operation on the review results of each segment to be reviewed, the target review result is obtained.

2. The method according to claim 1, characterized in that, The preset splitting rules are splitting by chapter, splitting by table of contents structure, or splitting by specific word count; The document is split according to the document to be reviewed and the preset splitting rules to obtain multiple segments to be reviewed, including: The document to be reviewed is divided into multiple segments to be reviewed, each consisting of a chapter, section, and paragraph. Alternatively, the document to be reviewed can be divided into multiple segments based on the table of contents page number range according to the outline structure. Alternatively, the document to be reviewed can be split into multiple segments containing a specific number of words.

3. The method according to claim 1, characterized in that, The document is split according to the document to be reviewed and the preset splitting rules to obtain multiple segments to be reviewed, including: The document to be reviewed is parsed and split in the following order of priority from high to low: chapters, table of contents, fixed-length fields based on a preset sliding window, and pre-set word count based on paragraphs, to obtain multiple segments to be reviewed.

4. The method according to claim 1, characterized in that, The process involves searching the pre-built vector library for each segment to be reviewed, resulting in a first list of reference document names composed of the names of the sample documents corresponding to each segment. This list includes: Based on all the segments to be reviewed, an iterative search is performed in the pre-built vector library to obtain the first number of knowledge data points ranked first by similarity for each segment to be reviewed. The first number of knowledge data are re-sorted to obtain the second number of knowledge data in the score ranking, wherein the second number is less than the first number. The second number of knowledge data corresponding to each segment to be reviewed are parsed and deduplicated to obtain a list of first reference document names.

5. The method according to claim 1, characterized in that, The acquisition of the pre-built vector library includes: Obtain sample documents; The document is split according to the sample document and the preset splitting rules to obtain sample knowledge fragments; A pre-built vector library is formed based on the sample knowledge fragments and the preset library table structure.

6. The method according to claim 1, characterized in that, The result processing operation in response to the user's review results for each segment to be reviewed, to obtain the target review result, includes: In response to the user's result judgment operation on the review results of each segment to be reviewed, a result judgment result is obtained; When the result is detected as a direct acceptance, the review results of each segment to be reviewed are directly merged to obtain the entire review result of the document to be reviewed as the target review result. When the result of the judgment is detected to be a partial modification adoption, the result of the user's modification operation on the review result of each segment to be reviewed is obtained as the target review result of the document to be reviewed. When the result is detected as requiring re-review, the system responds to the user's reference source knowledge adjustment operation, obtains the adjusted target reference source knowledge, and inputs the adjusted target reference source knowledge, each segment to be reviewed, and fixed prompt words into the pre-trained knowledge retrieval model to obtain the secondary review results of each segment to be reviewed.

7. The method according to any one of claims 1 to 6, characterized in that, After performing search processing in the pre-built vector library based on each segment to be reviewed to obtain a first list of reference document names consisting of the names of sample documents corresponding to each segment to be reviewed, the method further includes: When no user input adjustment information is detected, the reference source knowledge of each segment to be reviewed is determined according to the first reference document name list, and the knowledge point range of the reference source knowledge of each segment to be reviewed is adjusted to obtain the target reference source knowledge of each segment to be reviewed.

8. A document data processing apparatus, characterized in that, include: The acquisition module is used to acquire a pre-built vector library and documents to be reviewed, wherein the pre-built vector library is built based on sample documents; The document splitting module is used to split the document into multiple segments based on the document to be reviewed and preset splitting rules. The vector search module is used to perform search processing on each segment to be reviewed in the pre-built vector library to obtain a first reference document name list composed of the names of sample documents corresponding to each segment to be reviewed. The document review module is used to adjust the first reference document name list in response to the user's review adjustment operation, and obtain a second reference document name list composed of the names of sample documents corresponding to each segment to be reviewed. The document review module is further configured to determine the reference source knowledge of each segment to be reviewed based on the second list of reference document names, and to perform knowledge point range adjustment processing on the reference source knowledge of each segment to be reviewed to obtain the target reference source knowledge of each segment to be reviewed. The document review module is also used to input the target reference source knowledge of each segment to be reviewed, each segment to be reviewed, and fixed prompt words into a pre-trained knowledge retrieval model to obtain the review results of each segment to be reviewed. The results output module is used to respond to the user's results processing operations on each segment to be reviewed, and to obtain the target review result.

9. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1 to 7.

Images