Document consistency auditing method and system based on large model and storage medium

Abstract

The invention provides a document consistency auditing method and system based on a large model and a storage medium, and the method comprises the steps: obtaining a to-be-audited document, inputting the to-be-audited document into a pre-trained large model, and carrying out the fragment division to obtain segmented documents; information extraction is carried out on the segmented documents to obtain document extraction information, information filtering is carried out on the document extraction information to obtain extraction filtering information, and the document extraction information comprises attribute names, attribute values, extraction original texts and reference documents; and carrying out attribute alignment on the extraction filtering information and attribute standard information in the reference literature, and generating a document consistency auditing result according to an attribute alignment result. According to the embodiment of the invention, the extraction filtering information and the attribute standard information in the reference document are subjected to attribute alignment, so that the to-be-extracted document and the reference document can be effectively subjected to semantic consistency auditing, and the accuracy of document consistency auditing is improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a document consistency review method, system and storage medium based on a large model. Background Art

[0002] With the deepening of the Internet and digital transformation, the speed of document information generation and dissemination is accelerating, and the demand for document review is becoming more urgent. For example, the expansion of urban transportation, the introduction of new electronic products, the construction of new buildings, etc. Most of these things require technical solutions, feasibility plans, etc., and follow certain standards and construction plans. Therefore, the consistency review of documents is increasingly valued by people.

[0003] In the existing document consistency review process, the review is generally conducted based on keyword matching, resulting in low accuracy of document consistency review. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a document consistency review method, system and storage medium based on a large model to solve the problem of low accuracy of document consistency review in the prior art.

[0005] The embodiment of the present invention is implemented as follows: a document consistency review method based on a large model, the method comprising:

[0006] Obtaining a document to be reviewed, and inputting the document to be reviewed into a pre-trained large model for segmentation to obtain segmented documents;

[0007] Extracting information from the segmented documents to obtain document extraction information, and filtering the document extraction information to obtain extraction filtering information, wherein the document extraction information includes attribute names, attribute values, extracted original texts, and references;

[0008] The extracted filtering information is attribute-aligned with the attribute standard information in the cited document, and a document consistency review result is generated according to the attribute alignment result.

[0009] Preferably, extracting information from the segmented documents to obtain document extraction information includes:

[0010] Acquire a preset specified attribute, and perform attribute specified matching between the preset specified attribute and the segmented document;

[0011] According to the attribute-specified matching result, the matched characters in the segmented document are determined as the attribute name, and the value corresponding to the attribute name is determined as the attribute value;

[0012] Determine the sentence corresponding to the attribute name as the extracted original text, and determine the cited document according to the preset specified attribute;

[0013] The attribute name, the attribute value, the extracted original text and the cited document determined corresponding to the same preset designated attribute are combined to obtain the document extraction information.

[0014] Preferably, extracting information from the segmented documents to obtain document extraction information includes:

[0015] Obtaining the field name of the segmented document, and determining the attribute name according to the field name;

[0016] Determine the value corresponding to the attribute name as the attribute value, and determine the sentence corresponding to the attribute name as the extracted original text;

[0017] The cited document is determined according to the field name, and the attribute name, the attribute value, the extracted original text and the cited document determined corresponding to the same segmented document are combined to obtain the document extraction information.

[0018] Preferably, filtering the document extraction information to obtain extraction and filtering information includes:

[0019] Acquire the attribute semantics of the attribute name, and semantically match the attribute semantics with the extracted original text;

[0020] If the semantic match between the attribute semantics and the extracted original text is unqualified, deleting the document extraction information corresponding to the attribute name in the document extraction information;

[0021] Combining the attribute name with the attribute value to obtain an attribute combination, and performing description matching between the attribute combination and the extracted original text;

[0022] If the description match between the attribute combination and the extracted original text is unqualified, deleting the document extraction information corresponding to the attribute combination in the document extraction information;

[0023] The remaining document extraction information is determined as the extraction filtering information.

[0024] Preferably, the extraction and filtering information is attribute-aligned with the attribute standard information in the reference document, including:

[0025] Performing name alignment on the attribute names in the extracted filtering information to obtain a name alignment list;

[0026] Determine whether the reference corresponding to the attribute name in the name alignment list has an attribute standard value;

[0027] If the attribute standard value exists in the cited document, semantically compare the attribute value corresponding to the attribute name with the attribute standard value to obtain a first semantic comparison result;

[0028] If the attribute standard value does not exist in the cited document, the first attribute value corresponding to the attribute name is determined as the attribute standard value, and the remaining attribute values ​​corresponding to the attribute name are semantically compared with the attribute standard value to obtain a second semantic comparison result;

[0029] The first semantic comparison result and the second semantic comparison result are combined to obtain the attribute alignment result.

[0030] Preferably, performing name alignment on the attribute names in the extracted filtering information to obtain a name alignment list includes:

[0031] In the extracted filtering information, similarities between different attribute names are calculated to obtain name similarities, and the attribute names whose name similarities are greater than a first similarity threshold are combined to obtain the name alignment list.

[0032] Preferably, the document to be reviewed is input into the pre-trained large model for segmentation to obtain segmented documents, including:

[0033] According to the pre-trained large model, sequentially determine the character lengths of the document paragraphs in the document to be reviewed;

[0034] If the character length of any of the document paragraphs is less than a length threshold, combining the document paragraph with an adjacent document paragraph in a preset direction to obtain a combined document;

[0035] If the character length of the combined document is greater than or equal to the length threshold, dividing the document paragraphs into the segmented documents;

[0036] If the character length of the combined document is less than the length threshold, the combined document is combined with adjacent document paragraphs in a preset direction until the character length of the combined document is greater than or equal to the length threshold, and the combined document before being combined is divided into the segmented documents.

[0037] Another object of an embodiment of the present invention is to provide a document consistency review system based on a large model, the system comprising:

[0038] A segmentation module is used to obtain a document to be reviewed, and input the document to be reviewed into a pre-trained large model for segmentation to obtain segmented documents;

[0039] An information extraction module, used to extract information from the segmented documents to obtain document extraction information, and to filter the document extraction information to obtain extraction filtering information, wherein the document extraction information includes attribute names, attribute values, extracted original texts and references;

[0040] The consistency review module is used to perform attribute alignment on the extracted filtering information and the attribute standard information in the cited document, and generate a document consistency review result according to the attribute alignment result.

[0041] Preferably, the information extraction module is also used for:

[0042] Acquire a preset specified attribute, and perform attribute specified matching between the preset specified attribute and the segmented document;

[0043] According to the attribute-specified matching result, the matched characters in the segmented document are determined as the attribute name, and the value corresponding to the attribute name is determined as the attribute value;

[0044] Determine the sentence corresponding to the attribute name as the extracted original text, and determine the cited document according to the preset specified attribute;

[0045] The attribute name, the attribute value, the extracted original text and the cited document determined corresponding to the same preset designated attribute are combined to obtain the document extraction information.

[0046] The embodiments of the present invention can effectively divide the document to be reviewed into segmented documents by inputting the document to be reviewed into a pre-trained large model for segmentation. By extracting information from the segmented documents, the attribute names, attribute values, extracted original texts and cited documents corresponding to the segmented documents can be effectively extracted. By aligning the extracted filter information with the attribute standard information in the cited documents, the semantic consistency review can be effectively performed on the extracted document and the cited file, thereby improving the accuracy of the document consistency review. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of a document consistency review method based on a large model provided in a first embodiment of the present invention;

[0048] Figure 2 The document to be reviewed is in a table format provided by the first embodiment of the present invention;

[0049] Figure 3 is a flow chart of knowledge base retrieval provided by the first embodiment of the present invention;

[0050] Figure 4 It is a structural diagram of a document consistency review system based on a large model provided by a second embodiment of the present invention;

[0051] Figure 5 It is a schematic diagram of the structure of a terminal device provided in the third embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0053] In order to illustrate the technical solution of the present invention, a specific embodiment is provided below for illustration.

[0054] Embodiment 1

[0055] See also Figure 1 , is a flow chart of a document consistency review method based on a large model provided by a first embodiment of the present invention. The document consistency review method based on a large model can be applied to any device or system. The document consistency review method based on a large model includes the following steps:

[0056] Step S10, obtaining a document to be reviewed, and inputting the document to be reviewed into a pre-trained large model for segmentation to obtain segmented documents;

[0057] The document to be reviewed is divided into chapters and paragraphs. At this time, the chapter information of the source fragment can be traced. For the content in the table, it can be converted into text description according to the row and column names, for example Figure 2 The table part can be converted into "the maximum throughput capacity of the manual ticket gate (passengers / h) is 1200".

[0058] Optionally, the document to be reviewed is input into a pre-trained large model for segmentation to obtain segmented documents, including:

[0059] According to the pre-trained large model, sequentially determine the character lengths of the document paragraphs in the document to be reviewed;

[0060] If the character length of any of the document paragraphs is less than a length threshold, combining the document paragraph with an adjacent document paragraph in a preset direction to obtain a combined document;

[0061] If the character length of the combined document is greater than or equal to the length threshold, dividing the document paragraphs into the segmented documents;

[0062] If the character length of the combined document is less than the length threshold, the combined document is combined with adjacent document paragraphs in a preset direction until the character length of the combined document is greater than or equal to the length threshold, and the combined document before being combined is divided into the segmented documents;

[0063] In this step, the document paragraphs are merged in sequence, that is, if the total length of paragraph 1 and paragraph 2 is greater than or equal to the length threshold, paragraph 1 is regarded as a segmented document. If the total length of paragraph 1 and paragraph 2 is less than the length threshold, the total length of paragraph 1, paragraph 2 and paragraph 3 is calculated. If it is greater than or equal to the length threshold, paragraph 1 and paragraph 2 are regarded as a segmented document. The above steps are repeated in sequence to complete the segment division. The length threshold can be set according to needs.

[0064] Step S20, extracting information from the segmented documents to obtain document extraction information, and filtering the document extraction information to obtain extraction filtering information;

[0065] Among them, the document extraction information includes attribute name, attribute value, extracted text and cited literature, that is, the big model outputs four results of "attribute name, attribute value, original text (extracted text), and cited literature". The output of "original text" is because there is a situation where one attribute corresponds to multiple values ​​in the text fragment input to the big model. By outputting the original text, the position and order of different values ​​can be distinguished. On the other hand, it is beneficial to narrow the inspection scope in the subsequent filtering and alignment stages. The output of "cited literature" is to accurately locate the comparison document in the alignment stage, narrow the alignment scope, and improve the accuracy and alignment speed.

[0066] Optionally, extracting information from the segmented documents to obtain document extraction information includes:

[0067] Obtaining a preset specified attribute, and performing attribute-specified matching between the preset specified attribute and the segmented document; wherein the preset specified attribute can be set according to requirements, and by performing attribute-specified matching between the preset specified attribute and the segmented document, a document position corresponding to the same character as the preset specified attribute in the segmented document is obtained;

[0068] According to the attribute specified matching result, the matched characters in the segmented document are determined as the attribute name, and the value corresponding to the attribute name is determined as the attribute value; wherein, by determining the value corresponding to the attribute name as the attribute value, the attribute value corresponding to the attribute name in each segmented document can be effectively obtained;

[0069] Determine the sentence corresponding to the attribute name as the extracted original text, and determine the cited document according to the preset specified attribute;

[0070] The attribute name, the attribute value, the extracted original text and the referenced document determined corresponding to the same preset designated attribute are combined to obtain the document extraction information;

[0071] This embodiment is provided with an instruction template PROMPT1 for extracting a specified attribute value. The content of the instruction template PROMPT1 is as follows:

[0072] "You are a structured data extractor and are required to extract the following target attributes from the input text: {key_lists}; Requirements:

[0073] 1. Please output the extraction results in a markdown table. The columns of the table are: attribute name, attribute value, extracted original text, and cited literature name;

[0074] 2. If there is no target attribute content in the original text, the target attribute value is "no corresponding value";

[0075] 3. The content of the original text column requires the extraction of the corresponding original text fragments, and no modification is allowed;

[0076] 4. The cited document name refers to the officially released standard and specification document. If the current target attribute value is a cited document, the document name is returned; otherwise, "None" is returned.

[0077] Input: {segment}

[0078] Output: "

[0079] Among them, key_lists is all the specified attribute names concatenated with the preset concatenation character, and segment is the segmented document.

[0080] Further, information extraction is performed on the segmented documents to obtain document extraction information, including:

[0081] Obtaining the field name of the segmented document, and determining the attribute name according to the field name;

[0082] Determine the value corresponding to the attribute name as the attribute value, and determine the sentence corresponding to the attribute name as the extracted original text;

[0083] The cited document is determined according to the field name, and the attribute name, the attribute value, the extracted original text and the cited document determined corresponding to the same segmented document are combined to obtain the document extraction information.

[0084] This embodiment also provides an instruction template PROMPT2 for extracting common attribute values. The content of instruction template PROMPT2 is as follows:

[0085] “You are a professional who is proficient in data structuring. Please read the text and extract the key information: Requirements:

[0086] 1. Only return the extraction results;

[0087] 2. Please use "field name / / field value / / original text / / referenced document name" as one line to represent the KV value of a key information. If there are multiple extraction results, return multiple lines as required;

[0088] 3. When there are multiple attributes in the "Field Name", please use "_" to connect multiple attributes;

[0089] 4. If there is no extraction result, return "None";

[0090] Example:

[0091] Input: The total planned land area of ​​Chengjiang Road Depot is 60ha, of which Line 5 occupies about 8ha;

[0092] Output: Depot name / / Chengjiang Road Depot / / Chengjiang Road Depot has a total planned land area of ​​60ha / / None;

[0093] Chengjiang Road Depot_Land Area / / 60ha / / Chengjiang Road Depot Total Planned Land Area 60ha / / None;

[0094] Chengjiang Road Depot_Line 5_Land Area / / 8ha / / The total planned land area of ​​Chengjiang Road Depot is 60ha, of which Line 5 occupies about 8ha / / None;

[0095] Please combine the requirements and examples for accurate extraction.

[0096] Input: {segment}

[0097] Output:"

[0098] Among them, segment is a segmented document. Since the values ​​of common attributes are extracted here, in order to improve the extraction effect and the standardization of model output results, PROMPT2 adopts the ICL (In-Context Learning) method to help the model understand the task by adding examples.

[0099] Since users can obtain part / all of the attribute information of the document to be reviewed in actual application scenarios, this embodiment designs two extraction instruction templates, uses a large model to obtain corresponding results, and improves the accuracy of information extraction.

[0100] Furthermore, the document extraction information is filtered to obtain extraction and filtering information, including:

[0101] Acquire the attribute semantics of the attribute name, and semantically match the attribute semantics with the extracted text; wherein, the attribute name is matched with the semantic query table to obtain the attribute semantics, and the attribute semantics is semantically matched with the extracted text to determine whether the attribute name appears in the extracted text. If the attribute name appears in the extracted text, it is determined that the semantic match between the attribute semantics and the extracted text is qualified; if the attribute name does not appear in the extracted text, it is determined that the semantic match between the attribute semantics and the extracted text is unqualified;

[0102] If the semantic match between the attribute semantics and the extracted original text is unqualified, the document extraction information corresponding to the attribute name is deleted from the document extraction information; wherein, by deleting the document extraction information corresponding to the attribute name that does not appear in the extracted original text, the document extraction information is effectively filtered, thereby improving the accuracy of the document extraction information;

[0103] Combining the attribute name with the attribute value to obtain an attribute combination, and performing description matching between the attribute combination and the extracted original text;

[0104] If the description match between the attribute combination and the extracted original text is unqualified, the document extraction information corresponding to the attribute combination is deleted from the document extraction information; wherein, by deleting the document extraction information corresponding to the attribute combination that does not appear in the extracted original text, the accuracy of the document extraction information is further improved;

[0105] The remaining document extraction information is determined as the extraction filtering information.

[0106] In this embodiment, a filtering instruction template PROMPT3 is provided, and the content of the filtering instruction template PROMPT3 is as follows:

[0107] "You are a professional text reviewer, and your task is to determine whether the entry "{key}:{value}" comes from the given text.

[0108] Enter text: {content}; Please think as follows:

[0109] 1. Is the concept of {key} mentioned in the original text? Answer "yes" or "no"

[0110] 2. Can you find any description in the text that matches {key} and {value}? Answer "yes" or "no" and give specific explanations.

[0111] 3. Please give your conclusion in the format of "Conclusion: ***" and follow the following judgment rules:

[0112] *If the answer to either 1 or 2 is negative, then it should be determined that the term does not come from the input text, and the answer is "no"; if the answers to 1 and 2 are both positive, it indicates that the term comes from the input text, and the answer is "yes";

[0113] Requirements: Please answer "yes" or "no" to the conclusion given, no explanation is required;

[0114] Please start giving step-by-step thinking as required:”

[0115] Among them, content is the original text. If the output is "yes", the extraction result is retained, and if the output is "no", the extraction result is aligned.

[0116] After information extraction and information filtering, all extraction results of the document to be reviewed can be obtained and stored in the following list format: "attribute name": key, "attribute value": value, "original text segment": content, "reference document".

[0117] Step S30, performing attribute alignment on the extracted filtering information and the attribute standard information in the cited document, and generating a document consistency review result according to the attribute alignment result;

[0118] Among them, according to the attribute alignment results, the contents with the same attributes but different attribute values ​​are output in the order of their appearance in the document to obtain the document consistency review results. The output format is as follows: "error text", "chapter where the error text is located", "fragment where the error text is located", "correct text", "document name where the correct text is located", "fragment where the correct text is located", "chapter where the correct text is located".

[0119] Optionally, performing attribute alignment on the extracted filtering information and attribute standard information in the referenced document includes:

[0120] Performing name alignment on the attribute names in the extracted filtering information to obtain a name alignment list;

[0121] Determine whether the reference corresponding to the attribute name in the name alignment list has an attribute standard value;

[0122] If the attribute standard value exists in the cited document, semantically compare the attribute value corresponding to the attribute name with the attribute standard value to obtain a first semantic comparison result;

[0123] If the attribute standard value does not exist in the cited document, the first attribute value corresponding to the attribute name is determined as the attribute standard value, and the remaining attribute values ​​corresponding to the attribute name are semantically compared with the attribute standard value to obtain a second semantic comparison result;

[0124] Combining the first semantic comparison result and the second semantic comparison result to obtain the attribute alignment result;

[0125] Among them, for all attribute values ​​of the same attribute name, if the "referenced document" has no valid value (attribute standard value), the first value is taken as the correct value by default, and the remaining values ​​are compared with the first value. When the two are semantically exactly the same, they are consistent, otherwise they are inconsistent and stored in the final output result; if its "referenced document" is a valid value, all the extraction results of the referenced document are obtained, and the values ​​corresponding to the aligned attribute names are found, and then the semantics are compared to see if they are the same.

[0126] Further, the attribute names in the extracted filtering information are aligned to obtain a name alignment list, including:

[0127] In the extracted filtering information, similarities between different attribute names are calculated to obtain name similarities, and attribute names whose name similarities are greater than a first similarity threshold are combined to obtain the name alignment list;

[0128] Among them, for different attribute names, it is determined whether their strings are completely equal. If so, the two attribute names are aligned. If not, the large model is used to score the similarity and obtain the name similarity. If the name similarity is lower than the second similarity threshold, the two attribute names are not aligned; if the name similarity is greater than the first similarity threshold, the two attribute names are aligned; if the name similarity is between the second similarity threshold and the first similarity threshold, the RAG method is used to use an external knowledge base (professional library / web search, etc.) to provide the large model with richer semantic information to determine whether the attribute names are aligned, and then the aligned attribute values ​​are stored in the same list in order.

[0129] For example, see Figure 3 , design user questions and retrieve information related to the questions through the knowledge base;

[0130] The large model combines the retrieved information to give an accurate answer, as shown below:

[0131] The question "Please find the professional explanation of the following two terms: 1. Maximum continuous operating speed; 2. Line grade speed" is input into the external knowledge base for knowledge base question and answer, and the question and answer results are obtained. The question and answer results include:

[0132] Maximum sustained operating speed refers to the maximum speed that a train can continuously reach under normal operating conditions to ensure safety and performance.

[0133] The line grade speed refers to the maximum safe speed for train operation designed according to line and train conditions.

[0134] The question-and-answer results are input into the big model for semantic judgment to determine whether the two entries have the same meaning. After the big model makes the judgment, the semantic judgment result is output.

[0135] In this embodiment, the large model is pre-trained according to the model training samples, so that the trained large model can effectively perform functions such as segmentation, information extraction, information filtering and information filtering. Specifically, the model training samples are input into the large model for segmentation to obtain segmented samples, information is extracted from the segmented samples to obtain sample extraction information, information is filtered on the sample extraction information to obtain sample filtering information, the sample filtering information is attribute-aligned with the attribute standard information in the referenced document to obtain the sample alignment result, the model loss is determined according to the sample alignment result, and the parameters of the large model are updated according to the model loss until the large model converges to obtain the pre-trained large model.

[0136] In this embodiment, by inputting the document to be reviewed into the pre-trained large model for segmentation, the document to be reviewed can be effectively divided into segmented documents. By extracting information from the segmented documents, the attribute names, attribute values, extracted original texts and cited documents corresponding to the segmented documents can be effectively extracted. By aligning the extracted filter information with the attribute standard information in the cited documents, the semantic consistency of the extracted document and the cited file can be effectively reviewed, thereby improving the accuracy of the document consistency review.

[0137] Embodiment 2

[0138] See also Figure 4 , is a schematic diagram of the structure of a document consistency review system 100 based on a large model provided in a second embodiment of the present invention, including:

[0139] The segmentation module 10 is used to obtain the document to be reviewed, and input the document to be reviewed into the pre-trained large model for segmentation to obtain segmented documents.

[0140] Optionally, the segment division module 10 is further used to: determine the character lengths of the document paragraphs in the document to be reviewed in sequence according to the pre-trained large model;

[0141] If the character length of any of the document paragraphs is less than a length threshold, combining the document paragraph with an adjacent document paragraph in a preset direction to obtain a combined document;

[0142] If the character length of the combined document is greater than or equal to the length threshold, dividing the document paragraphs into the segmented documents;

[0143] If the character length of the combined document is less than the length threshold, the combined document is combined with adjacent document paragraphs in a preset direction until the character length of the combined document is greater than or equal to the length threshold, and the combined document before being combined is divided into the segmented documents.

[0144] The information extraction module 11 is used to extract information from the segmented documents to obtain document extraction information, and to filter the document extraction information to obtain extraction filtering information. The document extraction information includes attribute names, attribute values, extracted original texts and cited documents.

[0145] Optionally, the information extraction module 11 is further used to: obtain a preset specified attribute, and perform attribute specified matching between the preset specified attribute and the segmented document;

[0146] According to the attribute-specified matching result, the matched characters in the segmented document are determined as the attribute name, and the value corresponding to the attribute name is determined as the attribute value;

[0147] Determine the sentence corresponding to the attribute name as the extracted original text, and determine the cited document according to the preset specified attribute;

[0148] The attribute name, the attribute value, the extracted original text and the cited document determined corresponding to the same preset designated attribute are combined to obtain the document extraction information.

[0149] Furthermore, the information extraction module 11 is also used to: obtain the field name of the segmented document, and determine the attribute name according to the field name;

[0150] Determine the value corresponding to the attribute name as the attribute value, and determine the sentence corresponding to the attribute name as the extracted original text;

[0151] The cited document is determined according to the field name, and the attribute name, the attribute value, the extracted original text and the cited document determined corresponding to the same segmented document are combined to obtain the document extraction information.

[0152] Furthermore, the information extraction module 11 is also used to: obtain the attribute semantics of the attribute name, and semantically match the attribute semantics with the extracted original text;

[0153] If the semantic match between the attribute semantics and the extracted original text is unqualified, deleting the document extraction information corresponding to the attribute name in the document extraction information;

[0154] Combining the attribute name with the attribute value to obtain an attribute combination, and performing description matching between the attribute combination and the extracted original text;

[0155] If the description match between the attribute combination and the extracted original text is unqualified, deleting the document extraction information corresponding to the attribute combination in the document extraction information;

[0156] The remaining document extraction information is determined as the extraction filtering information.

[0157] The consistency review module 12 is used to perform attribute alignment between the extracted filtering information and the attribute standard information in the referenced document, and generate a document consistency review result according to the attribute alignment result.

[0158] The optional consistency audit module 12 is further used to: perform name alignment on the attribute names in the extracted filtering information to obtain a name alignment list;

[0159] Determine whether the reference corresponding to the attribute name in the name alignment list has an attribute standard value;

[0160] If the attribute standard value exists in the cited document, semantically compare the attribute value corresponding to the attribute name with the attribute standard value to obtain a first semantic comparison result;

[0161] If the attribute standard value does not exist in the cited document, the first attribute value corresponding to the attribute name is determined as the attribute standard value, and the remaining attribute values ​​corresponding to the attribute name are semantically compared with the attribute standard value to obtain a second semantic comparison result;

[0162] The first semantic comparison result and the second semantic comparison result are combined to obtain the attribute alignment result.

[0163] Furthermore, the consistency audit module 12 is also used to: calculate the similarity between different attribute names in the extracted filtering information to obtain name similarity, and combine the attribute names whose name similarity is greater than a first similarity threshold to obtain the name alignment list.

[0164] In this embodiment, by inputting the document to be reviewed into the pre-trained large model for segmentation, the document to be reviewed can be effectively divided into segmented documents. By extracting information from the segmented documents, the attribute names, attribute values, extracted original texts and cited documents corresponding to the segmented documents can be effectively extracted. By aligning the extracted filter information with the attribute standard information in the cited documents, the semantic consistency of the extracted document and the cited file can be effectively reviewed, thereby improving the accuracy of the document consistency review.

[0165] Embodiment 3

[0166] Figure 5 2 is a block diagram of a terminal device 2 provided in the third embodiment of the present application. Figure 5 As shown, the terminal device 2 of this embodiment includes: a processor 20, a memory 21, and a computer program 22 stored in the memory 21 and executable on the processor 20, such as a program of a document consistency review method based on a large model. When the processor 20 executes the computer program 22, the steps in each embodiment of the document consistency review method based on a large model are implemented.

[0167] Exemplarily, the computer program 22 may be divided into one or more modules, which are stored in the memory 21 and executed by the processor 20 to complete the present application. The one or more modules may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program 22 in the terminal device 2. The terminal device may include, but is not limited to, a processor 20 and a memory 21.

[0168] The processor 20 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0169] The memory 21 may be an internal storage unit of the terminal device 2, such as a hard disk or memory of the terminal device 2. The memory 21 may also be an external storage device of the terminal device 2, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the terminal device 2. Further, the memory 21 may also include both an internal storage unit and an external storage device of the terminal device 2. The memory 21 is used to store the computer program and other programs and data required by the terminal device. The memory 21 may also be used to temporarily store data that has been output or is to be output.

[0170] In addition, each functional module in each embodiment of the present application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of software functional unit.

[0171] If the integrated module is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium can be non-volatile or volatile. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable storage medium may include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in computer-readable storage media can be appropriately increased or decreased according to the requirements of legislation and patent practices in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practices, computer-readable storage media do not include electrical carrier signals and telecommunication signals.

[0172] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A document consistency review method based on a large model, characterized in that: The method comprises: Obtaining a document to be reviewed, and inputting the document to be reviewed into a pre-trained large model for segmentation to obtain segmented documents; Extracting information from the segmented documents to obtain document extraction information, and filtering the document extraction information to obtain extraction filtering information, wherein the document extraction information includes attribute names, attribute values, extracted original texts, and references; The extracted filtering information is attribute-aligned with the attribute standard information in the cited document, and a document consistency review result is generated according to the attribute alignment result.

2. The document consistency review method based on a large model as described in claim 1 is characterized in that: Extracting information from the segmented documents to obtain document extraction information includes: Acquire a preset specified attribute, and perform attribute specified matching between the preset specified attribute and the segmented document; According to the attribute-specified matching result, the matched characters in the segmented document are determined as the attribute name, and the value corresponding to the attribute name is determined as the attribute value; Determine the sentence corresponding to the attribute name as the extracted original text, and determine the cited document according to the preset specified attribute; The attribute name, the attribute value, the extracted original text and the cited document determined corresponding to the same preset designated attribute are combined to obtain the document extraction information.

3. The document consistency review method based on a large model as claimed in claim 1, characterized in that: Extracting information from the segmented documents to obtain document extraction information includes: Obtaining the field name of the segmented document, and determining the attribute name according to the field name; Determine the value corresponding to the attribute name as the attribute value, and determine the sentence corresponding to the attribute name as the extracted original text; The cited document is determined according to the field name, and the attribute name, the attribute value, the extracted original text and the cited document determined corresponding to the same segmented document are combined to obtain the document extraction information.

4. The document consistency review method based on a large model as claimed in claim 1, characterized in that: Filtering the document extraction information to obtain extraction and filtering information includes: Acquire the attribute semantics of the attribute name, and semantically match the attribute semantics with the extracted original text; If the semantic match between the attribute semantics and the extracted original text is unqualified, deleting the document extraction information corresponding to the attribute name in the document extraction information; Combining the attribute name with the attribute value to obtain an attribute combination, and performing description matching between the attribute combination and the extracted original text; If the description match between the attribute combination and the extracted original text is unqualified, deleting the document extraction information corresponding to the attribute combination in the document extraction information; The remaining document extraction information is determined as the extraction filtering information.

5. The document consistency review method based on a large model as claimed in claim 1, characterized in that: Attribute alignment of the extracted filtering information with attribute standard information in the referenced document includes: Performing name alignment on the attribute names in the extracted filtering information to obtain a name alignment list; Determine whether the reference corresponding to the attribute name in the name alignment list has an attribute standard value; If the attribute standard value exists in the cited document, semantically compare the attribute value corresponding to the attribute name with the attribute standard value to obtain a first semantic comparison result; If the attribute standard value does not exist in the cited document, the first attribute value corresponding to the attribute name is determined as the attribute standard value, and the remaining attribute values ​​corresponding to the attribute name are semantically compared with the attribute standard value to obtain a second semantic comparison result; The first semantic comparison result and the second semantic comparison result are combined to obtain the attribute alignment result.

6. The document consistency review method based on a large model as claimed in claim 5, characterized in that: Perform name alignment on the attribute names in the extracted filtering information to obtain a name alignment list, including: In the extracted filtering information, similarities between different attribute names are calculated to obtain name similarities, and the attribute names whose name similarities are greater than a first similarity threshold are combined to obtain the name alignment list.

7. The document consistency review method based on a large model as claimed in claim 1, characterized in that: The document to be reviewed is input into the pre-trained large model for segmentation to obtain segmented documents, including: According to the pre-trained large model, sequentially determine the character lengths of the document paragraphs in the document to be reviewed; If the character length of any of the document paragraphs is less than a length threshold, combining the document paragraph with an adjacent document paragraph in a preset direction to obtain a combined document; If the character length of the combined document is greater than or equal to the length threshold, dividing the document paragraphs into the segmented documents; If the character length of the combined document is less than the length threshold, the combined document is combined with adjacent document paragraphs in a preset direction until the character length of the combined document is greater than or equal to the length threshold, and the combined document before being combined is divided into the segmented documents.

8. A document consistency review system based on a large model, characterized in that: The system comprises: A segmentation module is used to obtain a document to be reviewed, and input the document to be reviewed into a pre-trained large model for segmentation to obtain segmented documents; An information extraction module, used to extract information from the segmented documents to obtain document extraction information, and to filter the document extraction information to obtain extraction filtering information, wherein the document extraction information includes attribute names, attribute values, extracted original texts and references; The consistency review module is used to perform attribute alignment on the extracted filtering information and the attribute standard information in the cited document, and generate a document consistency review result according to the attribute alignment result.

9. The document consistency review system based on a large model as claimed in claim 8, characterized in that: The information extraction module is also used for: Acquire a preset specified attribute, and perform attribute specified matching between the preset specified attribute and the segmented document; According to the attribute-specified matching result, the matched characters in the segmented document are determined as the attribute name, and the value corresponding to the attribute name is determined as the attribute value; Determine the sentence corresponding to the attribute name as the extracted original text, and determine the cited document according to the preset specified attribute; The attribute name, the attribute value, the extracted original text and the cited document determined corresponding to the same preset designated attribute are combined to obtain the document extraction information.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images