Semantic deep proofreading and man-machine review method and system based on large model and knowledge graph tracing

Abstract

The application relates to the technical field of electric digital data processing, and discloses a semantic depth proofreading and man-machine rechecking method and system based on a large model and a knowledge graph traceability, which comprises the following steps: constructing a knowledge field vector index composed of authoritative facts; calculating the vector of a fact assertion to be proofread and the gravitational gradient of the fact assertion in the knowledge field; and performing risk grading according to the gradient, so that the subsequent depth analysis process is started or skipped in a differentiated manner. Through the establishment of a prepositive risk quantification mechanism, the application avoids the traditional post-validation mode, can dynamically allocate computing resources to high-risk assertions, and can identify implicit errors inconsistent with the fact basis at the semantic level in advance, thereby improving the proofreading efficiency and quality.

Description

Technical Field

[0001] This invention relates to a semantic deep proofreading and human-computer verification method and system based on large model and knowledge graph tracing, belonging to the field of electronic digital data processing technology. Background Technology

[0002] In current text processing applications, a common technical solution is to combine large language models for deep semantic understanding and use knowledge graphs for fact tracing to achieve automated proofreading. This solution has significant technical advantages over traditional rule-based or shallow matching methods when dealing with issues such as contextual contradictions, logical errors, and inconsistencies in terminology. However, when this technical solution is applied to large-scale text processing, there is an inherent technical contradiction between computational cost and processing efficiency in its workflow. Existing workflows typically adopt a post-verification workflow, where the system first performs a uniform deep scan of all text content using a large language model to identify potential errors, and then calls external resources such as knowledge graphs to verify the identified potential errors. This fixed workflow results in the high computational load of deep semantic analysis being indiscriminately applied to all text information to be processed, including a large amount of information that is correct in itself.

[0003] Therefore, simply improving the reasoning performance of the large language model itself or optimizing the query speed of the knowledge graph cannot change the workflow itself, which uniformly allocates computing resources to all information, nor can it fundamentally solve the problem of excessive overall computational overhead. Analysis reveals the following shortcomings in the workflow design of existing technologies: 1. The processing flow lacks a computational step that can pre-judge the degree of deviation between textual assertions and factual basis before deep semantic analysis; 2. The allocation of computing resources is not correlated with the probability of errors in the information itself, and is evenly distributed to all information to be processed by preceding steps; 3. The initial screening, primarily based on semantic recognition, struggles to generate effective distinguishing features for errors that conform to grammatical logic but are inconsistent with the facts, potentially leading to their omission.

[0004] Moreover, even some existing technologies aimed at optimizing the reasoning process have failed to break free from this approach in terms of the macro-process of data processing. For example, Chinese invention patent CN118606487B discloses a data processing method for a knowledge graph reasoning engine based on logical reasoning. Although this method attempts to ensure the accuracy of the results by analyzing the characteristics of logical paths and the cross-correlation of paths, combined with subsequent source tracing verification, it is essentially still a serial mode that performs unified and complex logical verification on all data entering the processing flow. It does not set up a pre-emptive risk assessment step and cannot divert data according to the degree of deviation between the data and the existing knowledge system. Therefore, it also faces the problem of wasting a lot of computing resources on verifying correct information that does not require in-depth analysis. This makes it difficult to fundamentally avoid the inherent defect of excessive computational overhead when processing massive amounts of data. Therefore, the technical problem to be solved by this invention is how to establish a new data processing method that can quantify the inherent risk of each factual assertion in the text through an efficient pre-computation step before conducting high-cost semantic deep analysis, and then sort the information according to the quantification results, giving priority to allocating computing resources to assertions with high risk. Summary of the Invention

[0005] This invention provides a semantic deep proofreading and human-computer verification method and system based on large model and knowledge graph tracing. Its main purpose is to solve the problems of rigid computing resource allocation, high proofreading costs, and the risk of hidden factual errors being missed due to the lack of a pre-emptive risk quantification step in existing technologies.

[0006] To achieve the above objectives, this invention provides a semantic deep proofreading and human-computer review method based on large model and knowledge graph tracing. This method pre-executes a risk assessment and processing path decision-making step before a semantic deep analysis step is performed. The method includes:

[0007] Step a, construct a knowledge field vectorized index: call a first sentence encoding model to calculate each triple in a knowledge graph consisting of multiple (entity, relation, attribute) triples to generate a knowledge vector, and store all generated knowledge vectors in a vector database. Build a vector index that supports nearest neighbor search for all knowledge vectors. This index is defined as a knowledge field.

[0008] Step b, perform pre-screening based on assertion vector and knowledge field gravitational gradient: call a sentence encoding model that is the same as the first sentence encoding model, calculate an assertion vector for a factual assertion extracted from the text to be proofread; perform a nearest neighbor search in the knowledge field to find a knowledge vector that is geometrically closest to the assertion vector, and calculate the Euclidean distance between the assertion vector and the found nearest neighbor knowledge vector, and define the distance value as the knowledge field gravitational gradient;

[0009] Step c: Based on the comparison results of the knowledge field gravity gradient, execute differentiated processing paths: when the gradient value is less than a preset first threshold, the factual assertion is judged as a high-confidence assertion, and the subsequent semantic deep analysis step performed by the large model and the source tracing step performed by the knowledge graph are skipped; when the gradient value is greater than a preset second threshold, the factual assertion is judged as a high-risk abnormal assertion, and the semantic deep analysis step and the source tracing step are initiated.

[0010] Preferably, the factual assertion extracted from the text to be proofread is an atomized, structured representation that is decomposed from the unstructured sentence of the text to be proofread using natural language processing technology, and contains a core entity, a core relation, and a core attribute.

[0011] Preferably, the method further includes a closed-loop feedback self-evolution step based on the human-machine review results. The closed-loop feedback self-evolution step includes: structurally archiving the final decision made by a human expert on the output results of the semantic deep analysis step and the source tracing step; when the final decision is to confirm a new fact or correct an existing fact, the confirmed or corrected fact is updated to the knowledge graph in the form of (entity, relation, attribute) triples, and the first sentence encoding model is called to recalculate the knowledge vector for the updated triples to update the knowledge field vectorized index.

[0012] Preferably, the method further includes a step of calculating an implicit risk score for high-risk anomaly assertions to determine the processing priority of high-risk anomaly assertions in the semantic deep analysis step; the calculation rule for the implicit risk score is limited to: ,in, This represents the implicit risk score; Gravitational gradient of knowledge fields that assert facts; Score the contextual credibility of factual assertions within the surrounding paragraphs of the text to be proofread; This is historical adoption rate data that characterizes the authority of the source associated with a factual assertion; , , The preset weighting coefficients are used, and the preset weighting coefficients are used... , , The sum of is 1.

[0013] Preferably, the human-computer review is implemented through a human-computer interaction interface, which is configured as follows: for high-confidence assertions, the high-confidence assertion is displayed side by side with the original triples corresponding to the nearest neighbor knowledge vectors found in the knowledge field, and a confirmation option and a correction option are provided; for high-risk anomaly assertions, the high-risk anomaly assertion is displayed side by side with the correction suggestions generated by the semantic deep analysis step performed by the large model, and the tracing evidence provided by the tracing step performed by the knowledge graph, and an option to accept the correction suggestion, an option to reject the correction suggestion, and an option to edit the correction suggestion are provided.

[0014] Preferably, a preset third threshold is set between the preset first threshold and the preset second threshold, wherein the preset first threshold is less than the preset third threshold and the preset third threshold is less than the preset second threshold; when the knowledge field gravity gradient is greater than the preset first threshold and less than the preset third threshold, the factual assertion is judged as a potential risk assertion, and the semantic deep analysis step and the source tracing step are initiated with a priority lower than that of high-risk abnormal assertions.

[0015] Preferably, the data sources for knowledge graphs include at least one of authoritative documents, legal provisions, industry standards, and historical proofreading records.

[0016] Preferably, the sentence encoding model that is the same as the first sentence encoding model is a distilled BERT model that is pre-trained on a medium-sized general corpus and then fine-tuned using domain corpus from the data source.

[0017] Preferred, contextual credibility score The generation steps include: based on a pre-defined semantic pattern library and context dependency analysis, determining whether the factual assertion is an emotion-dependent assertion; if so, invoking a contextual sentiment analysis model to analyze the contextual paragraph containing the factual assertion to extract its sentiment polarity, subjective intensity, and contextual consistency features, and generating a contextual credibility score based on the extracted features. The method also includes a preset weighting coefficient. , , The dynamic calibration process includes: prioritizing the processing of multiple high-risk anomaly assertions by a human expert and comparing the results with the system's implicit risk scores. The calculated priority ranking results are compared to determine a ranking consistency deviation; and based on this deviation, the preset weighting coefficients are adjusted. , , The value is adjusted.

[0018] A semantic deep proofreading and human-computer review system based on large model and knowledge graph tracing, the system includes:

[0019] A knowledge field construction module is configured to: call a first sentence encoding model to compute each triple in a knowledge graph consisting of multiple (entity, relation, attribute) triples to generate a knowledge vector, store all generated knowledge vectors in a vector database, and build a vector index that supports nearest neighbor search for all knowledge vectors. This index is defined as a knowledge field.

[0020] A pre-review module is configured to: invoke a sentence encoding model identical to the first sentence encoding model to compute a fact assertion extracted from the text to be reviewed, thereby generating an assertion vector; perform a nearest neighbor search in the knowledge field to find a knowledge vector that is geometrically closest to the assertion vector, and compute the Euclidean distance between the assertion vector and the found nearest neighbor knowledge vector, defining the distance value as the knowledge field gravity gradient;

[0021] A processing path decision module is configured to: execute differentiated processing paths based on the comparison results of the knowledge field gravity gradient; when the gradient value is less than a preset first threshold, the factual assertion is judged as a high-confidence assertion, and the system controls the system to skip the subsequent semantic deep analysis steps performed by the large model and the source tracing steps performed by the knowledge graph; when the gradient value is greater than a preset second threshold, the factual assertion is judged as a high-risk abnormal assertion, and the system controls the system to start the semantic deep analysis steps and the source tracing steps.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] 1. Before initiating deep semantic analysis, a knowledge field vector index composed of massive facts was first constructed. By calculating the distance between the assertion vector in the text to be proofread and the nearest neighbor knowledge vector in the knowledge field, a gravity gradient value representing the degree of consistency between the assertion and the existing knowledge system was obtained. This approach establishes a pre-emptive risk assessment mechanism based on the inherent correlation of data. This ensures that subsequent computing resources are no longer indiscriminately invested in all text information, but are guided to different processing paths according to the gradient value. High-risk abnormal assertions are prioritized for deep analysis, while high-confidence assertions are directly sent to the manual verification stage. This changes the previous proofreading system's inherent operation of serial processing and post-verification of all information.

[0024] 2. This method places the construction of the knowledge field vectorized index, the real-time calculation of assertion vectors, and the subsequent deep analysis based on the large model into a unified closed-loop feedback system. The final decision of human experts on the results of the deep analysis is structured and archived and used in reverse to update the facts or relationships in the knowledge graph. This directly supplements or strengthens the vector index of the knowledge field and provides high-quality labeled data for the subsequent fine-tuning of the large model. In this way, a continuous and mutually reinforcing data coupling relationship is formed between the benchmark knowledge field of the pre-examination and the core large model of the subsequent deep analysis. The overall verification capability of the system evolves spontaneously as its processing tasks increase.

[0025] 3. By decomposing the text to be proofread into atomic factual assertions and calculating the gravitational gradient of each assertion in the knowledge field, the system can effectively identify implicit errors that are semantically grammatically correct and logically consistent but deviate from the factual basis. This is because the assertion vectors of such errors will be in a relatively sparse position in the knowledge field vector space composed of objective facts due to the lack of sufficient gravitational pull from relevant knowledge vectors, and their calculated gravitational gradient values ​​will be significantly different from those assertions that are consistent with the facts. This provides a direct identification path based on the data space distribution for dealing with key factual errors that are easily missed by traditional proofreading methods based solely on semantic analysis. Attached Figure Description

[0026] Figure 1 This is a flowchart of the closed-loop proofreading and knowledge graph self-evolution process of human-machine collaboration in this invention.

[0027] Figure 2 This is a graph showing the multi-dimensional performance improvement of the system's self-evolution.

[0028] Figure 3 This is a diagram illustrating the risk pre-screening and closed-loop feedback framework of the present invention based on the knowledge field gravitational gradient. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0030] This invention provides a semantic deep proofreading and human-computer verification method and system based on large model and knowledge graph tracing. The overall data processing flow is configured such that a pre-assessment and processing path decision-making step is performed before a semantic deep analysis step. Physically, the system includes a knowledge field construction module, a pre-review module, a processing path decision-making module, and a human-computer verification and closed-loop feedback module. The knowledge field construction module transforms a knowledge graph composed of authoritative facts into a vectorized index that supports efficient retrieval. This index is defined in this technical solution as... The knowledge field; the pre-screening module receives the text to be proofread, breaks it down into atomic factual assertions, and calculates the gravitational gradient of each factual assertion in the knowledge field; the processing path decision module, based on the calculation results of the gravitational gradient, classifies each factual assertion into risk levels and controls the system to execute differentiated subsequent processing paths, thereby dynamically allocating computing resources to high-risk assertions. The operation of this system aims to solve the technical problems of rigid allocation of computing resources and omission of hidden factual errors caused by the lack of a pre-emptive risk quantification step in the existing electronic digital data processing flow; in a specific application scenario, namely for When automating the proofreading of a draft contract containing numerous technical specifications and legal clauses, an objective obstacle in existing processes is that they must uniformly call upon a high-cost large language model for deep scanning of all clauses in the contract, regardless of their correctness, leading to excessive consumption of computing resources. To address this challenge, the knowledge field construction module of this technical solution is configured to pre-execute the construction steps of the knowledge field vectorized index during the system initialization phase. The initial state of this module includes a knowledge graph composed of entity, relation, and attribute triples, with data sources including authoritative documents, legal provisions, and industry standards. A distilled BERT model, pre-trained on a medium-sized general corpus and then fine-tuned using the aforementioned domain corpus, is used as the first sentence encoding model. The controller calls the first sentence encoding model to calculate each triple in the knowledge graph to generate a 768-dimensional knowledge vector. All generated knowledge vectors are stored in a vector database, and a vector index supporting nearest neighbor search is established for all knowledge vectors. This index is defined as the knowledge field. This process establishes a vector space benchmark composed of authoritative facts that can be efficiently computed, providing a data foundation for subsequent risk prediction.

[0031] Furthermore, during the proofreading task, to avoid indiscriminately applying the computational load of deep semantic analysis to all text information to be processed, the pre-review module of this technical solution is configured to perform a pre-review step based on assertion vectors and knowledge field gravity gradients. Upon receiving the draft contract text to be proofread, the processor first uses natural language processing technology to decompose the unstructured sentences in the text into one or more atomic structured fact assertions containing a core entity, a core relation, and a core attribute. For example, the sentence "The contract stipulates that the warranty period for equipment A is three years" is decomposed into the fact assertion (equipment A, warranty period, three years). Next, the processor calls a sentence encoding model identical to the first sentence encoding model used when constructing the knowledge field, and calculates an assertion vector of the same dimension for each extracted fact assertion. For each generated assertion vector, the processor performs a k-nearest neighbor search operation in the knowledge field, where k equals 1, to find a knowledge vector that is geometrically closest to the current assertion vector in the vector space, and calculates the Euclidean distance between the assertion vector and its found nearest neighbor knowledge vector. This calculated distance value is defined as the knowledge field's gravitational gradient. For example, if the assertion vector generated from the aforementioned fact assertions is... The nearest neighbor knowledge vector found in the knowledge field that corresponds to the known fact (device A, warranty period, three years) is: The calculated knowledge field gravitational gradient Its value may be 0.05. This step generates a quantitative risk indicator for each fact assertion through a vector nearest neighbor search and distance calculation, which can characterize the degree of consistency between the assertion and the existing knowledge system.

[0032] The process of decomposing unstructured sentences in text into atomic fact assertions is executed through one of two configurable technical paths. The first path is a pipelined processing approach, which sequentially calls a named entity recognition model based on BERT-CRF and a relation extraction model based on an attention mechanism. The former identifies the core entities and attributes contained in the sentence, while the latter, based on the former's identification results, determines the predefined relationship types between entities and attributes. The second path is a unified generation approach, which takes the unstructured sentence as input and, through a structured prompt containing specific instructions and output format examples, calls a finely tuned model with at least 70 parameters. The system employs a generative large language model with a capacity of hundreds of millions of bytes, directly generating a JSON-formatted text string containing key-value pairs (entities, relations, attributes). During system deployment, for batch processing tasks with high processing speed requirements, the system is configured to use the first path; for online processing tasks with high accuracy requirements for extracting complex sentence structures and implicit relations, the system is configured to use the second path. After obtaining the knowledge field gravity gradient of each fact assertion, to achieve differentiated allocation of computing resources, the processing path decision module of this technical solution is configured to execute differentiated processing paths based on the comparison results of gradient values. The execution of this module depends on two preset thresholds, a first threshold... With a second threshold ,in Less than These two thresholds can be determined through offline calibration experiments, for example, by... Set to 0.1, Set to 0.7; when the gradient value of a fact assertion... Less than the first threshold When the gradient value is 0.05, as in the previous example, the assertion is determined to be a high-confidence assertion. The processing path decision module will control the system to skip the subsequent semantic deep analysis steps performed by the large model and the source tracing steps performed by the knowledge graph, and directly display the assertion and the original triples matched in the knowledge field side by side on the human-computer interaction interface for the operator to confirm; when the gradient value Greater than the second threshold When the gradient value is between [a certain value], it indicates that the assertion deviates significantly from all known facts in the knowledge field, and is therefore judged as a high-risk anomalous assertion. The module will then control the system to initiate subsequent semantic deep analysis and source tracing steps. and If an error occurs between these points, it is considered a potential risk assertion and is given a lower priority than a high-risk anomaly assertion, triggering in-depth analysis and tracing steps. This differentiated processing path decision-making mechanism ensures that computing resources can be focused on text content with a higher probability of error.

[0033] Furthermore, to prioritize the processing of high-risk anomaly assertions and enable the system's risk assessment model to be adaptive, this technical solution also includes an implicit risk score calculation step and a closed-loop feedback self-evolution step based on human-machine review results; for fact assertions judged as high-risk anomaly assertions, the system will apply the defined calculation rules. To calculate its implicit risk score; where, The calculated gravitational gradient of the knowledge field; The process of generating a contextual credibility score involves calling a contextual sentiment analysis model to analyze the contextual paragraph in which the assertion is located, extracting features such as sentiment polarity and subjective intensity, and generating a score based on these features. This is a historical adoption rate data that characterizes the authority of the source associated with the factual assertion; and , , The weights are preset and sum to 1, and can be adjusted through a dynamic calibration step. In the human-machine review phase, a human expert's final decision on the system's deep analysis results is structured and archived. When the decision confirms a new fact or corrects an existing fact, the updated fact is updated in the knowledge graph as a triple, and the knowledge vector is recalculated to update the knowledge field's vectorized index. Simultaneously, the actual ranking of the human expert's processing priorities for multiple high-risk anomaly assertions will be compared with the system's... The calculated ranking results are compared to determine the ranking consistency deviation, and the weighting coefficients are adjusted accordingly. , , The value is adjusted. This mechanism not only enables refined management of high-risk assertions, but also builds a data-driven closed-loop feedback system, which forms a continuous and mutually reinforcing data coupling relationship between the benchmark of pre-examination, i.e., the knowledge field, and the core of post-deep analysis, i.e., the big model.

[0034] Example 1: This example demonstrates the specific operation of the described technical solution in a particular scenario. In the final review process of a large pharmaceutical company submitting a new drug marketing authorization application to a drug regulatory authority, the application document is massive, containing thousands of pages of clinical trial data, production batch numbers, pharmacological and toxicological research reports, and chemical composition descriptions. It contains a type of hidden factual error that traditional proofreading methods struggle to detect: the text is grammatically and semantically correct, but the stated facts deviate from the project knowledge base or other parts of the document. For example, the content of a key component is described as 0.5% in the main text, but recorded as 0.45% in the appendix's test report. Overlooking such errors could delay the entire approval process. When the system using this technical solution processes the new drug marketing authorization application, it first initiates a pre-review step based on assertion vectors and the gravity gradient of the knowledge field, rather than immediately calling a computationally expensive large model for comprehensive semantic analysis. The system's processor decomposes the entire application document into tens of thousands of atomized factual assertions, such as the molecular formula of compound A. The Phase II clinical trial enrolled 350 subjects, and the environmental temperature for product stability testing was 25°C. The system utilizes the same sentence encoding model as when constructing the knowledge field to calculate each factual assertion as an assertion vector. Simultaneously, the system's knowledge field has been pre-constructed using the knowledge field construction module, which has pre-built a vectorized index containing tens of millions of knowledge vectors from all verified historical project data, standard operating procedures, and relevant pharmacopoeia standards accumulated within the pharmaceutical company. For the vast majority of the extracted factual assertions, such as compound molecular formulas or clinical trial enrollment numbers, because their content is highly consistent with authoritative facts already existing in the knowledge field, the calculated knowledge field gravity gradient values ​​after performing nearest neighbor search on their assertion vectors are all less than a preset first threshold. For example, if the value is less than 0.1, the system will classify these assertions as high-confidence assertions and skip the subsequent deep analysis process. This mechanism allows the system to free up computing resources from a large amount of accurate information, thereby resolving the contradiction in traditional data processing processes where high computing costs must be incurred in pursuit of comprehensiveness in the verification process.

[0035] Therefore, the system focuses its computing resources on a small number of assertions deemed risky; during processing, one fact assertion was that the ambient temperature for product stability testing was 25 degrees Celsius. The calculated gravitational gradient of the knowledge field is 0.82, which is greater than the preset second threshold. The system immediately classified it as a high-risk anomaly assertion because the standard operating procedure triplet stored in the knowledge field explicitly recorded an ambient temperature of 30°C. This vector space distance-based comparison allows the system to identify implicit errors that deviate from the factual basis before semantic understanding. At this point, the screening in the pre-screening step and the precise analysis in the subsequent deep analysis step work synergistically. The processing path decision module immediately initiates the semantic deep analysis step executed by the large model and the source tracing step executed by the knowledge graph. However, the target of the analysis is no longer the entire text, but only the part of the text where the ambient temperature is 25 degrees Celsius. This high-risk assertion and its context; the large model combines context with logical reasoning, while the knowledge graph provides tracing evidence, clearly showing that the assertion conflicts with specific clauses of the standard operating procedure document; in this way, this technical solution no longer attempts to directly answer the question of how to perform semantic analysis on the whole text more quickly, but transforms it into how to pre-determine which parts of the text are worthy of semantic analysis. By changing the initial entry point of data processing, the original technical limitations are circumvented under the new processing framework; finally, the system submits a focused proofreading report to the pharmacist in charge of review, which does not include a lot of text polishing suggestions, but only lists the environmental temperature for product stability testing. A few key factual errors, including those identified, were verified, and each error was presented with its location in the original text, suggested corrections from the large model, and authoritative source evidence from the knowledge graph. Based on this report, the pharmacist confirmed and corrected the key errors within minutes, ensuring the accuracy of the submitted documents. Compared to traditional processes, the entire data processing task achieved improvements in both processing efficiency and the detection rate of hidden factual errors. It should be noted that in the final human-computer review stage of Example 1, the human-computer interface operated by the pharmacist was configured to provide differentiated display and operation methods for assertions at different risk levels. The ambient temperature for product stability testing was 25°C. This high-risk anomaly assertion is compared by the interface with a correction suggestion generated by a semantic deep analysis step performed by a large model, with an ambient temperature of 30. The tracing basis provided by the tracing steps performed by the knowledge graph is Standard Operating Procedure (SOP) B-12, Section 3.4: The ambient temperature for stability testing should be maintained at 30°C. ±2 The suggestions are displayed side-by-side, offering options to accept, reject, and edit them for pharmacists to decide. For assertions deemed high-confidence by the system, such as the molecular formula of compound A... The interface then compares the assertion with the original triplet (compound A, molecular formula, ...) found in the knowledge field corresponding to its nearest neighbor knowledge vector. The two options are displayed side-by-side, with a confirmation option and a correction option provided, allowing pharmacists to quickly compare their consistency, thereby significantly improving the decision-making efficiency and reliability of the basis in the manual review process.

[0036] To further verify the beneficial effects of the technical solution of the present invention in solving the problems of the prior art through objective data comparison, the following comparative example 1 is set up.

[0037] Comparative Example 1: This comparative example aims to simulate a scenario where existing conventional technical approaches in the background section are used to automatically proofread a new drug marketing application document that is exactly the same as that in Example 1. The system used in this comparative example has the same hardware configuration (including processor, memory, and graphics processing unit), the large language model it calls, the knowledge graph used as the factual benchmark, and the natural language processing technology used to decompose the text into factual assertions as the system in Example 1. The only essential difference is that the system used in this comparative example does not have the knowledge field gravity gradient-based technology claimed in this invention. The pre-examination module and processing path decision module follow a fixed, undifferentiated serial processing flow, namely: all approximately 15,000 atomic fact assertions extracted from the application documents are submitted to the large language model for deep semantic analysis without exception, and the knowledge graph is called one by one for fact tracing and verification. In order to quantitatively evaluate the performance difference between the two technical solutions when processing the same large-scale professional document, the key performance indicators of the experimental group using the method of this invention (i.e., the execution process of Example 1) and the control group using the conventional method of this comparative example were recorded and compared. The results are given in Table 1.

[0038] Table 1: Comparison of performance indicators between the experimental group and the control group.

[0039]

[0040] As shown in Table 1, the two methods perform similarly in the core proofreading quality indicator, namely the detection rate of hidden factual errors, both effectively identifying the vast majority of critical errors. However, there are fundamental differences in computational efficiency and resource consumption. Due to its post-verification operation mode, the control group applies a high computational load indiscriminately to all information, resulting in a total processing time 7.5 times that of the experimental group. During the processing, the resource utilization rate of its graphics processing unit remains saturated. In practical engineering applications, this not only means higher hardware and energy costs but also makes it difficult to handle scenarios with multiple proofreading tasks running concurrently. The experimental results show that without sacrificing proofreading accuracy, the conventional technical path cannot avoid its inherent high computational cost and low processing efficiency when processing large-scale, highly serious technical documents, lacking the pre-emptive risk quantification and diversion processing mechanism based on the knowledge field gravity gradient of this invention. This also proves that the core technical features of this invention can improve the overall operational efficiency of the automated proofreading system.

[0041] Example 2: To objectively verify the effectiveness of the pre-screening step based on the knowledge field gravity gradient in identifying hidden factual errors and optimizing computational resource allocation in this technical solution, the following comparative experiment was designed and executed. The purpose of this experiment is to quantitatively compare the differences in key performance indicators between the experimental group using this technical solution and the control group using existing technology by running it on a standardized text corpus containing known error types and distributions. The experimental platform is based on a server configured with a general-purpose processor, 64GB of memory, and a graphics processing unit with 40GB of video memory. Both the experimental group and the control group use the same... The specific implementation method involves distilling the BERT model as the sentence encoding model and calling the same large language model to perform deep semantic analysis. The data source used in the experiment is a publicly available legal document dataset. First, 10,000 documents are randomly selected from this dataset to construct the knowledge field for this experiment, ensuring that the facts in the knowledge field are authoritative and unmodified original facts. Subsequently, another 5,000 documents are selected as the basis for the test corpus, and an automated script is used to inject two types of hidden factual errors into them at a preset ratio: the first type is high-risk anomalous assertions, totaling 500, characterized by the insertion of key figures from the original text. The first category consists of minor modifications to entities, such as changing the date from May 10, 2023 to May 20, 2023; the second category comprises 500 potentially risky assertions, characterized by the introduction of new facts that are true but not present in the knowledge field; the remaining assertions, accounting for over 99% of the total, remain unchanged and serve as the benchmark for high-confidence assertions; the control group's technical solution is set to directly call a large language model to perform a complete deep semantic analysis of the entire test corpus; the experimental group adopts this technical solution, which first performs a pre-screening step, and then performs triage processing on the factual assertions based on the calculation results of the knowledge field gravity gradient; the experiment begins After the operation, the pre-screening module of the experimental group first calculated the knowledge field gravity gradient of all factual assertions in the test corpus, and classified them according to the preset first threshold of 0.1 and the second threshold of 0.7. The statistical results showed that 99.1% of the factual assertions were judged as high-confidence assertions and skipped the deep analysis because their gradient values ​​were less than 0.1. Among the 1,000 injected erroneous or newly added assertions, 988 had gradient values ​​greater than 0.1 and were sent to the subsequent processing flow. Subsequently, the two systems processed in parallel until the proofreading task of the entire test corpus was completed, and its key performance indicators were recorded. The specific data are shown in Table 2.

[0042] Table 2: Comparison of performance indicators between the experimental group and the control group.

[0043]

[0044] As shown in Table 2, the performance of the present invention's sample group and the control group was at the same level in terms of the detection rate of hidden fact errors. This indicates that the pre-screening step did not cause the omission of key errors due to the filtering out of a large number of high-confidence assertions. However, in terms of processing efficiency, the total processing time of the present invention's sample group was only 13.2% of that of the control group. This is because the pre-screening step reduced the amount of data requiring deep semantic analysis from 100% to less than 1%. This data directly confirms the core mechanism of this technical solution, namely, the pre-risk quantification through the knowledge field gravity gradient, which can effectively separate computational resources from a large amount of information that is likely to be correct and allocate them to a small number of high-risk information. The experimental results show that by introducing a pre-screening step before deep semantic analysis, this technical solution can reduce the amount of data processing in subsequent high-cost computational steps without reducing the ability to detect hidden fact errors.

[0045] Example 3: This example combines Figures 1 to 3 This document describes a method and system for semantic deep proofreading and human-computer verification based on large-scale models and knowledge graph tracing, such as... Figure 1 As shown, the process begins with the original text to be proofread. After step a: assertion parsing and intent structure extraction to generate factual assertions, it proceeds to step b: factual assertions are classified, and sentiment and contextual analysis is performed. Assertions are divided into simple factual assertions and sentiment-dependent assertions. For the former, the system uses knowledge graphs for efficient verification in step c1, or uses step c2 to automatically assist decision-making on conflicting verification evidence. For the latter, it proceeds to step d, where the sentiment big model and knowledge graph perform in-depth analysis by generating RAGs through retrieval enhancement. The processing results of both paths are gathered in step e: human-machine collaborative review and closed-loop reconstruction, for expert final adjudication. The expert's adjudication ultimately triggers step z1: the system performs self-evolution of the knowledge graph based on the arbitration result, and uses the adjudication data to update the knowledge graph and vertical category big model.

[0046] like Figure 2 As shown in the figure, the horizontal axis represents the training time from the initial state to week 20, and the vertical axis represents the detection rate (%). The figure displays the evolution curves of three key performance indicators: the detection rate of hidden factual errors, the detection rate of logical errors, and the detection rate of terminological inconsistencies. All three curves show a steady upward trend over the training time, indicating that the system of this invention can continuously enhance its ability to identify and detect different types of errors through continuous learning and iteration. Figure 3As shown, it begins with unstructured text from a source text to be proofread. This text is first broken down into multiple atomic fact assertions by the fact assertion extraction module. Subsequently, each fact assertion is fed into the gravity gradient calculation module. This module uses a B1: knowledge field vectorization index, constructed from knowledge graph triples, as a benchmark to generate an assertion vector for the assertion and compare it with the knowledge field, thereby outputting a quantified knowledge field gravity gradient value. This gradient value is passed to the decision processing path module, which performs risk triage based on the gradient value, selecting high-confidence assertions with lower gradient values ​​along with their... The original triples matched in the knowledge field are directly submitted to human experts for review, while the high-risk / potential-risk assertions with high gradient values ​​are sent to the deep analysis and tracing module. This module calls the large model and knowledge graph for analysis, and submits the generated correction suggestions and tracing evidence to human experts. Finally, the final decision made by the human experts is not only used to generate the final proofreading report to complete this proofreading task, but also serves as a closed-loop feedback signal to update the triples in B2: the knowledge graph, and then update the vectorized index of B1: the knowledge field through the updated knowledge vector.

[0047] Example 4: This example aims to illustrate the calibration procedure for key parameters in the preceding technical solution. In a specific application scenario, when the verification system constructed by this technical solution is deployed for the first time in a new professional field, such as the review of prospectuses in the financial securities field, one of the objective technical challenges it faces is the first threshold used to distinguish factual assertions of different risk levels. With the second threshold And implicit risk score calculation rules for refined ranking of high-risk anomaly assertions. Weighting coefficients in , , All parameters are currently unknown. Setting these parameters solely based on experience will lead to deviations in risk assessment and handling path decisions during subsequent system operation. To address this issue, a standardized offline calibration process must be executed before the system is officially put into use. The initial state of this process is defined as: a corpus containing 10,000 authoritative documents in the field, used to construct an initial knowledge field; and a validation set containing 1,000 documents pre-annotated by human experts in the field, where each factual assertion is labeled as one of three categories: correct assertion, incorrect factual assertion, or newly added factual assertion. First, to determine the first threshold... With the second threshold The system calculates the knowledge field gravitational gradient for each fact asserted in the validation set, taking the value of . Statistical analysis of the calculation results revealed that the gradient values ​​of different categories of assertions exhibited different distribution ranges. The gradient values ​​of correct assertions were concentrated in a lower range, the gradient values ​​of incorrect assertions were concentrated in a higher range, and the gradient values ​​of newly added fact assertions fell between the two. Based on this, the first threshold... The determination procedure is limited to: within a preset gradient value range, such as 0.01 to 0.30, iterating in steps of 0.01, calculating the precision and recall for correctly classifying erroneous assertions at each candidate threshold, and selecting the gradient value that maximizes the harmonic mean of these two values, i.e., the F1 score, as the gradient. The final value; similarly, the second threshold The determination procedure is limited to selecting the gradient value that maximizes the F1 score of the fact-false assertion classification within another preset interval; through this procedure, the following can be obtained: and This is a set of parameters derived from statistical data.

[0048] Furthermore, this relates to the weighting coefficients in the implicit risk score calculation rules. , , The system performs calibration by selecting all samples from the validation set that are marked as factual errors or newly added factual assertions, and then calculating the contextual credibility score for each sample. ;Should The calculation of the value is determined to be an algorithmic process, which involves calling a pre-defined contextual sentiment analysis model to analyze the context of the assertion in order to extract its sentiment polarity. Subjective intensity Consistency with context Three eigenvalues, and based on Calculation Value, where, This represents the emotional polarity value. This represents the subjective intensity value. This is a contextual consistency score; this calculation method ensures that objective, neutral statements that are consistent with the context receive higher credibility scores; the gradient values ​​for each sample are obtained. Contextual credibility score and historical adoption rate data (Initial values ​​are uniformly set to 0.5) Then, the system presents a subset of these samples, i.e., 20 items, in random order to a human expert and records the expert's ranking of processing priorities based on their professional judgment; subsequently, the system executes an optimization algorithm... Under constraints, find a set of conditions that allow the system to... The calculated Spearman rank correlation coefficient between the ranking results and the ranking results of human experts reaches the maximum weighted combination; this process can determine a set of initial weights. , , By executing the aforementioned standardized calibration procedures, several previously uncertain core parameters have obtained initial set values ​​based on objective data and clear optimization objectives. This transforms a parameter setting process that relies on experience-based judgment into a logically closed and reproducible engineering implementation step. This ensures that the core risk assessment and decision-making mechanisms of the calibration system can be adapted to the data characteristics of the current field during the initial deployment of the system in different professional fields, providing an initial benchmark for its subsequent stable operation and self-evolution.

[0049] Example 5: This example aims to illustrate the dynamic maintenance procedures for the knowledge field and related parameters that form the basis of the preceding technical solutions. In a specific application scenario, when the verification system constructed by this technical solution is used to audit technical standard documents in a continuously evolving technical field, such as the communications industry, new versions of the standards are periodically released, introducing new technical terms and parameter indicators. If the system's knowledge field is not updated synchronously, it will classify assertions that have become facts in the new standards as high-risk abnormal assertions, thereby generating unnecessary in-depth analysis overhead. To address this situation, this technology... The solution includes a knowledge field timeliness guarantee and reconstruction mechanism, which is set to execute automatically at preset intervals. During execution, the system first accesses multiple pre-defined external authoritative data sources, such as the official release repositories of industry standard organizations, and automatically retrieves all documents newly added or revised within that period. The system processes these documents, extracting the (entity, relation, attribute) triples they contain, and stores these triples in a temporary storage area. Subsequently, the system calculates the knowledge field gravity gradient for each triple in the temporary storage area within the current formal knowledge field, and sets the gradient value above a preset second threshold. The triples are identified as potential new or changed knowledge, and a discrepancy report is generated and submitted to a human expert for final confirmation. After confirmation by the human expert, these new or changed triples are officially updated into the knowledge graph, and the knowledge field construction module recalculates the knowledge vectors for them to update the knowledge field vectorized index. At the same time, the old version triples that are replaced are timestamped and archived to support subsequent version tracking. Through this procedure, the knowledge field, which serves as the pre-review benchmark, can keep pace with the evolution of external knowledge systems.

[0050] Furthermore, to ensure implicit risk scores In the calculation, historical adoption rate data, which characterizes the authority of the information source, is used. This technical solution is also equipped with a feature that can reflect the reliability of information from different sources. The dynamic update procedure for values; in this procedure, each information source in the system is assigned a unique source identifier and an initial value is set. Values, such as 0.5; In the human-machine review process of the system, when a human expert makes a final decision on an assertion originating from a specific source identifier, whether it is acceptance, modification, or rejection, the decision result is associated with that source identifier and recorded; the system periodically counts the total number of all associated decisions under each source identifier. And the number of times it was ruled acceptable. and based on The calculation rules are used to update the historical acceptance rate data of the information source. In addition, in order to make The value can better reflect recent changes in the credibility of the information source. A time decay factor can be introduced into the update calculation, assigning higher weight to recent rulings. This procedure will... The value has been transformed from a static initial setting into a data-driven reputation score that can be dynamically adjusted based on actual verification results.

[0051] Example 6: This example aims to supplement the explanation of the specific selection, construction, and fine-tuning procedures of the first sentence encoding model and the sentence encoding model identical to the first sentence encoding model in the previous technical solution, so as to eliminate the model black box that may exist in specific engineering implementation; in a specific application scenario, when the proofreading system constructed by this technical solution needs to be applied to a field with a highly specialized terminology system, such as the review of aircraft design specifications in the aerospace field, an unadapted general sentence encoding model may not be able to accurately capture the subtle semantic differences when mapping the professional terms in the field (such as flutter boundaries or static aeroelastic divergence) to the vector space, thus causing deviations in the subsequently calculated knowledge field gravity gradient and affecting the accuracy of the pre-review step; to cope with this boundary condition, the sentence encoding model adopted by this technical solution is initially a The distilled BERT model is pre-trained on a medium-sized general corpus, a choice made to balance computational efficiency and model performance. Before applying it to a specific professional domain, a standardized domain-adaptation fine-tuning process is required. The initial state of this process is defined as: a training dataset containing at least 1 million sentence pairs from authoritative literature in the target professional domain (such as aerospace), where each sentence pair is labeled as semantically equivalent or semantically inequivalent. The fine-tuning process adopts a contrastive learning framework, where two input sentences are passed through the distilled BERT model with shared weights to obtain their corresponding sentence vectors. The model parameters are iteratively updated with the optimization objective of minimizing the vector distance between semantically equivalent sentence pairs and maximizing the vector distance between semantically inequivalent sentence pairs. This process is completed when the model's classification accuracy on the reserved validation set converges to a stable value, for example, above 98%.

[0052] Furthermore, to ensure the stable performance of the sentence encoding model during long-term operation and to adapt to the continuous evolution of domain knowledge, this technical solution also includes an incremental fine-tuning mechanism based on human-machine review feedback. In this mechanism, all factual assertions finalized by human experts during the human-machine review process, along with their corresponding original triples in the knowledge field and their nearest-neighbor knowledge vectors, are automatically constructed into new training sentence pairs. Specifically, if a high-confidence assertion is confirmed as correct by a human expert, the assertion and its corresponding original triple are constructed into a semantically equivalent sentence pair; if a high-confidence assertion is confirmed as correct by a human expert, the assertion and its corresponding original triple are constructed into a semantically equivalent sentence pair; if a high-confidence assertion is confirmed as correct by a human expert, the assertion and its corresponding original triple are constructed into a semantically equivalent sentence pair. After a risky or abnormal assertion is corrected by human experts and updated in the knowledge graph, the original assertion and the corrected fact triple are constructed into a semantically inequivalent sentence pair. The system periodically accumulates this high-quality labeled data generated from actual business operations. When the amount of data reaches a preset threshold, such as 10,000 new sentence pairs, an incremental fine-tuning process is automatically triggered. Based on the model after the previous fine-tuning, the newly added data is used for further training. Through this procedure, it is ensured that the sentence encoding model, which serves as the foundation of the semantic representation of the entire system, can continuously optimize its ability to understand domain-specific terminology.

[0053] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A semantic deep proofreading and human-computer verification method based on large model and knowledge graph source tracing, characterized in that, This method performs a preliminary risk assessment and processing path decision-making step before executing a semantic deep analysis step. The method includes: Step a, construct a knowledge field vectorized index: call a first sentence encoding model to calculate each triple in a knowledge graph consisting of multiple triples to generate a knowledge vector, and store all generated knowledge vectors in a vector database. Build a vector index that supports nearest neighbor search for all knowledge vectors. This vector index is defined as a knowledge field. Step b, perform pre-screening based on assertion vector and knowledge field gravitational gradient: call a sentence encoding model that is the same as the first sentence encoding model, calculate an assertion vector for a factual assertion extracted from the text to be proofread; perform a nearest neighbor search in the knowledge field to find a knowledge vector that is geometrically closest to the assertion vector, and calculate the Euclidean distance between the assertion vector and the found nearest neighbor knowledge vector, and define the Euclidean distance value as the knowledge field gravitational gradient; Step c: Based on the comparison results of the knowledge field gravity gradient, execute differentiated processing paths: when the gradient value is less than a preset first threshold, the factual assertion is judged as a high-confidence assertion, and the subsequent semantic deep analysis step performed by the large model and the source tracing step performed by the knowledge graph are skipped; when the gradient value is greater than a preset second threshold, the factual assertion is judged as a high-risk abnormal assertion, and the semantic deep analysis step and the source tracing step are initiated.

2. The semantic deep proofreading and human-computer verification method based on large model and knowledge graph tracing as described in claim 1, characterized in that, The factual assertion extracted from the text to be proofread is an atomized, structured representation that is decomposed from the unstructured sentence of the text to be proofread using natural language processing technology. It contains a core entity, a core relation, and a core attribute.

3. The semantic deep proofreading and human-computer verification method based on large model and knowledge graph tracing as described in claim 1, characterized in that, The method also includes a closed-loop feedback self-evolution step based on the human-machine review results. The closed-loop feedback self-evolution step includes: archiving the final decision made by a human expert on the output results of the semantic deep analysis step and the source tracing step in a structured manner; when the final decision is to confirm a new fact or to correct an existing fact, the confirmed or corrected fact is updated to the knowledge graph in the form of a triple, and the first sentence encoding model is called to recalculate the knowledge vector for the updated triple.

4. The semantic deep proofreading and human-computer verification method based on large model and knowledge graph tracing as described in claim 1, characterized in that, The method also includes a step of calculating implicit risk scores for high-risk anomaly assertions to determine their processing priority in the semantic deep analysis step; the rules for calculating the implicit risk scores are defined as follows: ,in, This represents the implicit risk score; Gravitational gradient of knowledge fields that assert facts; Score the contextual credibility of factual assertions within the surrounding paragraphs of the text to be proofread; This is historical adoption rate data that characterizes the authority of the source associated with a factual assertion; , , The preset weighting coefficients are used, and the preset weighting coefficients are used... , , The sum of is 1.

5. The semantic deep proofreading and human-computer verification method based on large model and knowledge graph tracing as described in claim 1, characterized in that, Human-computer review is implemented through a human-computer interaction interface, which is configured as follows: for high-confidence assertions, the high-confidence assertion is displayed side by side with the original triples corresponding to the nearest neighbor knowledge vectors found in the knowledge field, and a confirmation option and a correction option are provided; for high-risk anomaly assertions, the high-risk anomaly assertion is displayed side by side with the correction suggestions generated by the semantic deep analysis step performed by the large model, and the source tracing evidence provided by the source tracing step performed by the knowledge graph, and an option to accept the correction suggestion, an option to reject the correction suggestion, and an option to edit the correction suggestion are provided.

6. The semantic deep proofreading and human-computer verification method based on large model and knowledge graph tracing as described in claim 1, characterized in that, Between the preset first threshold and the preset second threshold, a preset third threshold is also set, wherein the preset first threshold is less than the preset third threshold, and the preset third threshold is less than the preset second threshold. When the knowledge field gravity gradient is greater than a preset first threshold but less than a preset third threshold, the factual assertion is judged as a potentially risky assertion, and the semantic deep analysis step and the source tracing step are initiated with a lower priority than high-risk abnormal assertions.

7. The semantic deep proofreading and human-computer verification method based on large model and knowledge graph tracing as described in claim 1, characterized in that, The data sources for knowledge graphs include at least one of the following: authoritative documents, legal provisions, industry standards, and historical proofreading records.

8. The semantic deep proofreading and human-computer verification method based on large model and knowledge graph tracing as described in claim 1, characterized in that, The first sentence encoding model is the same as the first sentence encoding model. It is a distilled BERT model that is pre-trained on a medium-sized general corpus and then fine-tuned using domain corpus from the data source.

9. The semantic deep proofreading and human-computer verification method based on large model and knowledge graph tracing as described in claim 4, characterized in that, Contextual credibility score The generation steps include: based on a pre-defined semantic pattern library and context dependency analysis, determining whether the factual assertion is an emotion-dependent assertion; if so, invoking a contextual sentiment analysis model to analyze the contextual paragraph containing the factual assertion to extract its sentiment polarity, subjective intensity, and contextual consistency features, and generating a contextual credibility score based on the extracted features. The method also includes a preset weighting coefficient. , , The dynamic calibration process includes: prioritizing the processing of multiple high-risk anomaly assertions by a human expert and comparing the results with the system's implicit risk scores. The calculated priority ranking results are compared to determine a ranking consistency deviation; and based on this deviation, the preset weighting coefficients are adjusted. , , The value is adjusted.

10. A semantic deep proofreading and human-computer review system based on large model and knowledge graph tracing, characterized in that, The system is configured to perform the method of claim 1, the system comprising: A knowledge field construction module is configured to: call a first sentence encoding model to compute each triple in a knowledge graph consisting of multiple triples to generate a knowledge vector, store all generated knowledge vectors in a vector database, and build a vector index that supports nearest neighbor search for all knowledge vectors. This vector index is defined as a knowledge field. A pre-review module is configured to: invoke a sentence encoding model identical to the first sentence encoding model to compute a fact assertion extracted from the text to be reviewed, thereby generating an assertion vector; perform a nearest neighbor search in the knowledge field to find a knowledge vector that is geometrically closest to the assertion vector, and compute the Euclidean distance between the assertion vector and the found nearest neighbor knowledge vector, defining the Euclidean distance value as the knowledge field gravity gradient; A processing path decision module is configured to: execute differentiated processing paths based on the comparison results of the knowledge field gravity gradient; when the gradient value is less than a preset first threshold, the factual assertion is judged as a high-confidence assertion, and the system controls the system to skip the subsequent semantic deep analysis steps performed by the large model and the source tracing steps performed by the knowledge graph; when the gradient value is greater than a preset second threshold, the factual assertion is judged as a high-risk abnormal assertion, and the system controls the system to start the semantic deep analysis steps and the source tracing steps.

Images