Report quality inspection method and system combining large model and deep learning

Abstract

The invention discloses a report quality inspection method and system combining a large model and deep learning, and the method comprises the steps: carrying out the preprocessing of an original report text, including text cleaning, Chinese word segmentation and stop word filtering; performing clause classification prediction on the preprocessed text based on a BERT model, and outputting predicted clauses; comparing the predicted terms with the original report terms; if the predicted terms are consistent with the original report terms, outputting a verification result True; if not, performing deep semantic verification through a large language model, and outputting a final verification result; and standardizing and outputting a verification result. According to the method, the large model and the BERT are combined to check and judge the consistency of the result records and terms in the report, and the deep semantic analysis of the large model and the fast reasoning and judgment advantages of the BERT are combined, so that the accuracy and reasoning efficiency of the overall check and judgment of the algorithm are improved.

Description

Technical Field

[0001] This application relates to the field of report inspection technology, specifically a report quality inspection method and system that combines large models and deep learning. Background Technology

[0002] Inconsistencies often arise between the descriptions of clauses and results in evaluation reports, necessitating semantic analysis to determine their semantic consistency and achieve automatic verification. However, due to the complexity of evaluation scenarios, general semantic analysis models are insufficient for accurate semantic analysis. In comparison, large-scale models possess stronger semantic understanding capabilities, but their output stability needs improvement. Furthermore, the large volume of report data makes the overall analysis and judgment by large-scale models time-consuming and inefficient. Therefore, this invention designs a self-verification method for report clauses and results records that combines large-scale models and deep learning, enabling automatic and efficient report quality detection. Summary of the Invention

[0003] This embodiment provides a report quality inspection method, system, electronic device, and storage medium that combine large models and deep learning to solve the problems of insufficient inspection efficiency and accuracy in related technologies.

[0004] In a first aspect, embodiments of the present invention provide a report quality inspection method combining large models and deep learning, the report quality inspection method combining large models and deep learning includes: Preprocess the original report text, including text cleaning, Chinese word segmentation, and stop word filtering; The BERT model is used to classify and predict clauses in the preprocessed text, and the predicted clauses are output. Compare the forecast terms with the original report terms: If they match, the output verification result is True; If there is a discrepancy, a deep semantic verification is performed using a large language model, and the final verification result is output. Standardize the output verification results.

[0005] In an optional embodiment, the preprocessing further includes: Use a custom domain-specific dictionary for word segmentation; Stop words are dynamically added based on the report's theme to filter out irrelevant vocabulary.

[0006] In an optional embodiment, the training process of the BERT model includes: Construct a category template containing 276 types of clause labels; The binary cross-entropy loss function and AdamW optimizer were used, with a learning rate of 2e-5, a batch size of 16, and 10 training epochs.

[0007] In an optional embodiment, the preprocessed text is classified and predicted based on the BERT model, and the predicted terms are output, including: The probability of a text belonging to each category is calculated using the softmax activation function; The clause with the highest probability is selected as the prediction clause.

[0008] In an optional embodiment, the deep semantic verification includes: Design a structured Prompt template to guide the large language model in performing validation tasks; Text similarity is calculated using cosine similarity, with a similarity threshold of ≥0.85. Extract key entities from the original terms; if more than 30% of the key entities are missing, a mismatch is determined.

[0009] In an optional embodiment, the deep semantic verification further includes: For clauses containing logical relationships, verify the consistency of the logical structure between the predicted clause and the original clause; If the logical structures are inconsistent, they are directly judged as mismatches.

[0010] In an optional embodiment, the Prompt template design satisfies the following conditions: Clearly define task instructions and output format specifications; Include both positive and negative examples to improve recognition accuracy.

[0011] In an optional embodiment, the standardized output is in JSON format, containing the clause matching status and verification basis for each report.

[0012] Compared with existing technologies, the beneficial effects of the report quality inspection method combining large models and deep learning in this invention are as follows: By combining a large model with BERT to verify the consistency of results recording and terms in the report, the advantages of deep semantic analysis of the large model and fast reasoning of BERT are combined, which improves the overall accuracy and reasoning efficiency of the algorithm.

[0013] Secondly, embodiments of the present invention propose a report quality inspection system that combines large models and deep learning, including: The text preprocessing module is used to preprocess the original report text, including text cleaning, Chinese word segmentation, and stop word filtering; The BERT classification module is used to perform clause classification prediction on preprocessed text based on the BERT model and output the predicted clauses. The large language model validation module is used to compare the predicted terms with the original report terms: If they match, the output verification result is True; If there is a discrepancy, a deep semantic verification is performed using a large language model, and the final verification result is output. The results output module is used to standardize the output of verification results.

[0014] Thirdly, embodiments of the present invention provide an electronic device, including a processor, a communication interface, a memory, and a bus, wherein the processor, the communication interface, and the memory communicate with each other through the bus, and the processor can call logical instructions in the memory to execute the steps of the method provided in the first aspect.

[0015] Fourthly, embodiments of the present invention provide a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the report quality inspection method combining large models and deep learning as described in the first aspect.

[0016] Compared with the prior art, the beneficial effects of the report quality inspection system, electronic device and storage medium combining large models and deep learning of the present invention are the same as those of the report quality inspection method combining large models and deep learning described in the first aspect, so they will not be repeated here. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the report quality check method combining large models and deep learning in an embodiment of the present invention; Figure 2 This is a schematic diagram of the report quality inspection method combining large models and deep learning in an embodiment of the present invention; Figure 3 This is a sample diagram of one type of Prompt design in an embodiment of the present invention; Figure 4 This is an example diagram of another Prompt design in an embodiment of the present invention; Figure 5 This is a structural block diagram of the report quality inspection system combining large models and deep learning in an embodiment of the present invention; Figure 6 This is a structural block diagram of the electronic device in an embodiment of the present invention. Detailed Implementation

[0019] To better understand the purpose, technical solution, and advantages of this application, the application is described and explained below in conjunction with the accompanying drawings and embodiments.

[0020] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by one of ordinary skill in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.

[0021] This invention provides a report quality check method that combines large models and deep learning. Figure 1 This is a flowchart of the report quality inspection method combining large models and deep learning according to the present invention, such as... Figure 1 and Figure 2 As shown, the process includes the following steps: S100. Preprocess the original report text, including text cleaning, Chinese word segmentation, and stop word filtering; Furthermore, preprocessing also includes: Use a custom domain-specific dictionary for word segmentation; Stop words are dynamically added based on the report's theme to filter out irrelevant vocabulary.

[0022] Specifically, firstly, the incoming text is cleaned to remove special symbols and illegal characters, unify the encoding format, and standardize the text format; then, a Chinese word segmentation tool is used in conjunction with a custom professional domain dictionary to segment the text into words; finally, a general stop word list is loaded, and relevant stop words are manually added according to the report topic to filter words without actual semantic meaning and irrelevant to business, thereby reducing interference from irrelevant characters and improving the accuracy of semantic extraction.

[0023] S200. Based on the BERT model, perform clause classification and prediction on the preprocessed text, and output the predicted clauses. It should be noted that, in this embodiment, the training process of the BERT model includes: Construct a category template containing 276 types of clause labels; The binary cross-entropy loss function and AdamW optimizer were used, with a learning rate of 2e-5, a batch size of 16, and 10 training epochs.

[0024] For example, firstly, the text classification template is prepared: all clauses of the evaluation report are summarized, a standard clause classification template is created, all original report data used for model training are deduplicated, and data is batch cleaned according to the clause classification template; then, the original result records are labeled according to the clause classification template, training and test data are divided according to a specified ratio, and standard training and test sets are constructed for model training.

[0025] The BERT classification network was trained using the binary cross-entropy loss function, classifying data into 276 classes (labels 0-275). During training, AdamW was selected as the optimization method. This optimizer effectively balances training speed and parameter update stability when handling large-scale pre-trained models like BERT, avoiding overfitting. A learning rate of 2e-5 was set, a value widely validated in fine-tuning BERT models and well-suited for parameter updates. The batch size was set to 16, fully utilizing GPU memory to improve training efficiency while ensuring gradient calculation stability. Ten training epochs were set to ensure the model fully learns the feature distribution on the training data while preventing overfitting due to excessive training epochs.

[0026] Furthermore, based on the BERT model, the preprocessed text is classified and predicted to output predicted terms, including: The probability of a text belonging to each category is calculated using the softmax activation function; The clause with the highest probability is selected as the prediction clause.

[0027] Specifically, during the inference process, each piece of text data input to the model is classified, and the activation function uses softmax to calculate the probability that the current input belongs to each category. The prediction value with the highest probability is taken as the prediction term that matches the current result record.

[0028] S300. Compare the forecast terms with the original report terms: If they match, the output verification result is True; If there is a discrepancy, a deep semantic verification is performed using a large language model, and the final verification result is output. If the predicted terms match the original terms corresponding to the result record, the algorithm returns True, meaning the original report result record matches the terms. Conversely, if the predicted terms do not match the original terms corresponding to the result record, the large model performs further in-depth semantic analysis to understand and make a judgment.

[0029] It should be noted that deep semantic verification includes: Design a structured Prompt template to guide the large language model in performing validation tasks; Text similarity is calculated using cosine similarity, with a similarity threshold of ≥0.85. Extract key entities from the original terms; if more than 30% of the key entities are missing, a mismatch is determined.

[0030] Deep semantic validation also includes: For clauses containing logical relationships, verify the consistency of the logical structure between the predicted clause and the original clause; If the logical structures are inconsistent, they are directly judged as mismatches.

[0031] If the predicted clause matches the original clause corresponding to the result record, the algorithm returns True, meaning the original report result record matches the clause; otherwise, the algorithm returns False. During the large model validation process, multiple conditions are used for comprehensive judgment: first, cosine similarity is used to calculate the text similarity between the two texts; only texts with a similarity score higher than the threshold of 0.85 are considered consistent. At the same time, key entities of the original clauses are extracted. If the predicted clause is missing more than 30% of the key information, it is directly judged as a mismatch. For clauses containing logical relationships, it is also verified whether their logical structure is consistent. If the logic is changed, even if the text is similar, it is judged as a mismatch.

[0032] It should be further noted that the Prompt template design must meet the following conditions: Clearly define task instructions and output format specifications; Include both positive and negative examples to improve recognition accuracy.

[0033] The standardized output is in JSON format, containing the clause matching status and verification basis for each report.

[0034] Specifically, the design and optimization of the Prompt template aims to improve the overall accuracy of the large model in identifying positive and negative examples, while also ensuring the uniformity and stability of the output format to facilitate batch extraction of subsequent judgment results. A specific example of a Prompt template design is shown below. Figure 4 and Figure 5 As shown. (Through) Figure 4 and Figure 5 The structured Prompt design not only clearly guides large models to complete specific tasks, but also facilitates efficient parsing and batch processing of model-generated results by standardizing the output format.

[0035] S400, standardized output verification results.

[0036] Integrate the output results of BERT and large models, save them in JSON format as individual items, and output the final verification and judgment results of the input report.

[0037] This invention combines Deepseek with BERT to verify the consistency of results recording and terms in the report. By combining the deep semantic analysis of Deepseek with the fast reasoning and judgment advantages of BERT, the accuracy and reasoning efficiency of the algorithm's overall verification and judgment are improved.

[0038] This invention also provides a report quality inspection system combining large models and deep learning. This system is used to implement the above-described method embodiments, and details already described will not be repeated. The terms "module," "unit," and "subunit," etc., used below refer to combinations of software and / or hardware that perform a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation or a combination of software and hardware is also possible and contemplated.

[0039] like Figure 5 As shown, Figure 5 This is a structural block diagram of the report quality inspection system combining large models and deep learning in this invention. The system includes: The text preprocessing module 101 is used to preprocess the original report text, including text cleaning, Chinese word segmentation and stop word filtering; BERT classification module 102 is used to perform clause classification prediction on preprocessed text based on the BERT model and output predicted clauses; The large language model validation module 103 is used to compare the predicted terms with the original report terms. If they match, the output verification result is True; If there is a discrepancy, a deep semantic verification is performed using a large language model, and the final verification result is output. Result output module 104 is used to standardize the output of verification results.

[0040] Figure 6 A structural block diagram of the electronic device provided in the embodiments of the present invention, such as... Figure 6 As shown, the electronic device may include a processor 610, a communications interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute the following methods: S100. Preprocess the original report text, including text cleaning, Chinese word segmentation, and stop word filtering; S200. Based on the BERT model, perform clause classification and prediction on the preprocessed text, and output the predicted clauses. S300. Compare the forecast terms with the original report terms: If they match, the output verification result is True; If there is a discrepancy, a deep semantic verification is performed using a large language model, and the final verification result is output. S400, standardized output verification results.

[0041] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0042] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0043] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A report quality checking method combining large models and deep learning, characterized in that, The report quality checking method combining large models with deep learning comprises the following steps: The original report text is preprocessed, including text cleaning, Chinese word segmentation, and stop word filtering; The preprocessed text is classified and predicted based on a BERT model, and the predicted clauses are outputted; The predicted clauses are compared with the original report clauses: If they are consistent, the checking result is True; If they are inconsistent, deep semantic checking is performed through a large language model, and the final checking result is outputted; The checking result is standardized. 2.The method of claim 1, wherein, The preprocessing further comprises the following steps: Word segmentation is performed using a self-defined professional field dictionary; Irrelevant words are filtered by dynamically adding stop words according to the report theme. 3.The method of claim 1, wherein, The training process of the BERT model comprises the following steps: A classification template containing 276 clause labels is constructed; A binary cross-entropy loss function and an AdamW optimizer are adopted, the learning rate is set to 2e-5, the batch size is 16, and the training rounds are 10. 4.The method of claim 3, wherein, The preprocessed text is classified and predicted based on the BERT model, and the predicted clauses are outputted, including the following steps: The probability that the text belongs to each type of clause is calculated through a softmax activation function; The clause corresponding to the maximum probability is taken as the predicted clause. 5.The method of claim 1, wherein, The deep semantic checking comprises the following steps: A structured Prompt template is designed to guide the large language model to perform the checking task; The text similarity is calculated through cosine similarity, and the similarity threshold is set to be greater than or equal to 0.85; If more than 30% of the key entities of the original clauses are missing, it is determined that they are not matched. 6.The method of claim 5, wherein, The deep semantic checking further comprises the following steps: For clauses containing logical relationships, the logical structure consistency of the predicted clauses and the original clauses is verified; If the logical structure is inconsistent, it is directly determined that they are not matched. 7.The method of claim 5, wherein, The Prompt template design meets the following conditions: The task instructions and output format specifications are clear; Positive examples and negative examples are included to improve the correct recognition rate. 8.The method of claim 1, wherein, The standardized output is in JSON format, including the clause matching state of each report and the checking basis. 9.A report quality checking system combining large models and deep learning, characterized in that, It comprises the following parts: A text preprocessing module for preprocessing the original report text, including text cleaning, Chinese word segmentation, and stop word filtering; A BERT classification module for classifying and predicting the preprocessed text based on a BERT model, and outputting the predicted clauses; A large language model checking module for comparing the predicted clauses with the original report clauses: If they are consistent, the checking result is True; If they are inconsistent, deep semantic checking is performed through a large language model, and the final checking result is outputted; A result output module for standardizing the checking result.

10. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to realize the report quality checking method combining large models with deep learning according to any one of claims 1 to 8.

Images