Science and technology document intelligent error correction and logic consistency verification method based on large model

Abstract

The invention discloses a science and technology document intelligent error correction and logic consistency verification method based on a large model, and belongs to the technical field of artificial intelligence and document intelligent processing. Comprising the steps of science and technology document access and structured analysis; carrying out semantic unit division and feature extraction on the science and technology document; performing multi-level intelligent error correction based on a large model; checking the logic consistency of the science and technology documents; and fusing and outputting error correction and verification results. According to the method, automatic identification and prompt of language errors, professional errors and cross-paragraph logic inconsistency problems in the document are realized through comprehensive analysis of the science and technology document structure, professional semantics and logic relationship, so that the normalization and technical reliability of the science and technology document are improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and intelligent document processing technology, specifically relating to a method for intelligent error correction and logical consistency verification of scientific and technological documents based on a large model. Background Technology

[0002] In research institutes and engineering and technical units, scientific and technological documents are an important carrier for technology transfer and scientific research management. These documents typically have the following characteristics: dense technical terminology and rigorous technical descriptions; numerous numbered clauses, formulas, parameters, and derivations; and cross-chapter references and dependencies.

[0003] Existing document correction technologies primarily focus on general text spelling or grammar correction, which have significant shortcomings in the context of scientific and technical documents: they cannot identify errors at the professional semantic level, such as incorrect parameter units, mixed use of technical concepts, and non-standard equipment names. They lack logical consistency verification capabilities, failing to determine whether conclusions in different chapters are contradictory or whether conditions match conclusions. Furthermore, rule systems are costly to maintain, and traditional rules struggle to cover the complex and ever-changing writing scenarios of scientific and technical documents.

[0004] Currently, the following technical solutions are mainly used for document content correction and verification: Rule-based text correction technology This type of technology is usually based on predefined rules or templates to detect spelling errors, formatting errors, and non-standard numbering in documents, for example, through regular expressions, keyword matching, or fixed format comparison.

[0005] However, this technology relies heavily on manual maintenance of the rule base, which has a large number of rules and high update costs. It is difficult to cover the complex and diverse professional expressions in scientific and technological documents, and its ability to identify errors at the professional semantic level is limited.

[0006] Error correction methods based on statistical models or traditional natural language processing techniques Some existing systems employ language models based on statistical features or traditional natural language processing algorithms to perform spelling and grammar correction on text. These methods mainly rely on word frequency and contextual statistical features for judgment, and have some effectiveness in general text scenarios.

[0007] However, in scientific and technical documents, due to the diverse forms of technical terminology and expression, this type of method is difficult to accurately understand the professional semantics, and is prone to misjudging correct terms as incorrect ones. Furthermore, it cannot effectively handle non-natural language content such as parameters, units, and formulas.

[0008] Scientific document verification methods based on templates or manual review In practical scientific research and engineering applications, scientific and technological documents still rely heavily on manual review or template-based filling methods for quality control.

[0009] This method is not only inefficient and costly in terms of manpower, but also relies on the experience of the reviewers, making it highly subjective and difficult to detect hidden problems such as logical conflicts across chapters, missing conditions, or invalid conclusions in a timely manner.

[0010] Existing technologies generally have the following shortcomings in the correction and verification of scientific and technological documents: It focuses on detecting surface-level text errors, making it difficult to identify errors at the professional semantic level; it lacks the ability to comprehensively judge the rationality of parameters, units, and technical indicators; it cannot effectively reason about the logical relationships between conditions, data, and conclusions; it is difficult to achieve content consistency verification across chapters and paragraphs; the system has high maintenance costs and is difficult to adapt to the complex and ever-changing writing scenarios of scientific and technological documents.

[0011] Therefore, existing technologies are insufficient to meet the high requirements of scientific and technological documents in terms of professional accuracy and logical consistency. There is an urgent need for an intelligent method that can combine the semantic understanding and reasoning capabilities of large models, and is tailored to the structural characteristics of scientific and technological documents, to achieve professional error correction and logical consistency verification. Summary of the Invention

[0012] This invention aims to provide a method for intelligent error correction and logical consistency verification of scientific and technological documents based on a large model. By comprehensively analyzing the structure, professional semantics, and logical relationships of scientific and technological documents, it can automatically identify and prompt language errors, professional errors, and logical inconsistencies across paragraphs, thereby improving the standardization and technical reliability of scientific and technological documents.

[0013] The method of the present invention includes the following steps: S1: Access to and structured parsing of scientific and technological documents; S2: Semantic unit segmentation and feature extraction of scientific and technological documents; S3: Multi-level intelligent error correction based on a large model; S4: Logical consistency check of scientific and technological documents; S5: Output the combined results of error correction and verification.

[0014] Step S1: Accessing and Structured Parsing of Scientific and Technical Documents The technical documents received include formats such as PDF, Word, PPT, and scanned copies. The received documents undergo the following processing: format parsing and unified conversion; OCR recognition of scanned text; extraction of titles, chapter numbers, paragraphs, tables, formulas, and figure captions; and the creation of a document structure tree to record chapter levels and citation relationships. The parsing results form a structured intermediate representation of the technical documents.

[0015] Step S2: Semantic unit segmentation and feature extraction. Based on structural analysis, the scientific and technological document is segmented into semantic units to form the smallest semantic processing units, including: technical description sentences; conditional statements; conclusion or judgment sentences; parameter definitions and indicator description sentences.

[0016] For each semantic unit, the following features are extracted: technical terminology features; numerical and unit features; logical connective features; chapter and contextual dependencies.

[0017] Step S3: Multi-level intelligent error correction based on a large model Based on the above semantic units and features, a large model is invoked to perform multi-level error correction: Language and format layer error correction Identify and correct the following issues: spelling errors, punctuation errors; inconsistent numbering formats; and issues related to the standardization of scientific writing.

[0018] Correction of technical terminology and expressions By combining a domain terminology database with the semantic understanding capabilities of large models, we can identify: the use of non-standard terms; the mixed use of professional concepts; and ambiguity in technical expressions.

[0019] Parameter and numerical rationality correction Verify the parameter range and unit usage to identify problems such as missing units and inconsistent dimensions.

[0020] Step S4: Logical consistency verification of scientific and technological documents. Based on the completion of single-point error correction, logical consistency verification of the scientific and technological documents is performed, including: Cross-chapter semantic consistency verification compares descriptions of the same object in different chapters to determine whether there are contradictory conclusions.

[0021] Condition-conclusion logic verification analyzes whether there are logical jumps or invalid conclusions between the condition description and the conclusion statement.

[0022] Verify the consistency of citations and numbering, checking whether the citation relationships of chapter numbers, figure numbers, and formula numbers are correct.

[0023] During the logic verification process, the large model combines contextual information to output consistency judgment results and reasons.

[0024] Step S5: Output the fused error correction and verification results The error correction results and logical consistency verification results are merged to generate a unified output, including: error location; error type classification; correction suggestions or risk warnings; and explanation of logical conflicts.

[0025] The results can be output in the form of document comments, lists, or interfaces.

[0026] The present invention has the following beneficial effects: it enables intelligent error correction for scientific and technological documents, surpassing the capabilities of general text error correction; it provides cross-chapter logical consistency verification, reducing the risk of hidden errors in technical documents; it reduces the cost of manual proofreading and improves document quality; and it can directly serve scientific research management systems, knowledge platforms, and large model application platforms. Detailed Implementation

[0027] This invention is applicable to the automatic error correction, content standardization, and logical consistency verification of scientific and technological documents such as research reports, technical standards, test reports, operating procedures, and design specifications. Dielectric loss and capacitance measurement

[0028] Evaluation criteria: At 20℃, the value shall not exceed the following values: 750kV 0.5% 330kV~500kV 0.6% 110kV~220kV 0.8% 35kV 1.5% Experimental conclusion: Qualified The system identified the following during the analysis: There is a logical conflict between the experimental conclusions of the dielectric loss angle test and the description of the evaluation criteria. The system automatically provides the analysis results: According to the standard, the analysis results are as follows: The dielectric loss angle of HV-LV and G is 1.25%, which exceeds the standard of 0.8%. The dielectric loss angle of HV and LV-G is 1.12%, which also exceeds the standard of 0.8%. The dielectric loss angle of HV-LV is 1.2%, which also exceeds the standard of 0.8%; The dielectric loss angle of LV--HV and G is 0.272%, which is lower than the standard and is qualified.

[0029] Since at least three measurement results exceeded the specified standard value (0.8%), the test conclusion should have been unqualified. However, the report concluded that it was "qualified," which is inconsistent with the data analysis results. Therefore, the test conclusion is incorrect.

[0030] In conclusion, the review found that the test results were incorrect, and the correct conclusion should be that the test was unqualified.

[0031] This embodiment uses a transformer dielectric loss and capacitance measurement test report as an example to explain in detail the specific working process and working principle of the technical solution of the present invention in experimental scientific and technological documents.

[0032] The test report includes the measurement results of the dielectric loss angle, the evaluation criteria, and the final test conclusion. The document structure is typical and has a clear logical relationship of "measurement data - standard constraints - conclusion judgment".

[0033] The process of accessing and parsing scientific and technological documents The system first receives the test report document to be analyzed. After receiving the document, the system performs format parsing and text recognition processing on the document content, and automatically identifies different functional content areas in the document. During the parsing process, the system automatically distinguishes and extracts the following contents by combining rules and models: numerical descriptions of the dielectric loss measurement results; standard statements describing the dielectric loss evaluation criteria and threshold conditions; and summary statements expressing the test conclusions.

[0034] The system further establishes a structured representation within the document, logically linking measurement results, evaluation criteria, and conclusions to provide a unified data foundation for subsequent analysis.

[0035] Working principle of semantic unit segmentation and feature extraction After completing the structured parsing, the system divides the document text into semantic units, breaking the entire document down into multiple minimal analyzable semantic units. Specifically, the description of the dielectric loss angle measurement is identified as a parametric semantic unit, the evaluation criteria as a constraint semantic unit, and the experimental conclusions as judgment semantic units.

[0036] For various semantic units, the system automatically extracts key semantic features, including: the numerical value and unit corresponding to the dielectric loss angle; the judgment conditions implicit in the standard description (such as "not greater than a certain value"); and the qualification judgment keywords contained in the conclusion statement. Through the above feature extraction, the system constructs a correspondence model of "parameter-condition-conclusion" at the semantic level.

[0037] Parameter and Standard Inference Analysis Process Based on Large Models The system inputs the extracted parametric semantic units and standard constraint semantic units into the large model for inference analysis.

[0038] In this process, the large model does not simply compare strings, but rather, based on semantic understanding, automatically identifies the implicit judgment logic in the standard, namely: under specified temperature conditions, the dielectric loss angle must not exceed the threshold of the corresponding voltage level. The large model independently judges each dielectric loss angle measurement result, analyzes whether its value meets the constraints described in the standard, and outputs the compliance judgment result for each measurement result.

[0039] Through this process, the system identified several instances where measurement results exceeded the standard allowable range.

[0040] Working principle of logical consistency verification of experimental conclusions After completing the compliance analysis of individual parameters, the system further calls the large model to perform logical consistency verification of the test conclusions.

[0041] This verification process is based on the following working principle: If the document contains measurement results that do not meet the evaluation criteria, the overall test conclusion should not be judged as qualified.

[0042] The system compares the parameter analysis results with the semantic units of the conclusion judgment, and the large model automatically determines whether the "qualified" judgment expressed in the conclusion statement can be logically derived from the aforementioned measurement results and evaluation criteria.

[0043] In this embodiment, the large model inference results show that there is a logical inconsistency between the conclusion statement and the aforementioned data analysis results, and the conclusion judgment lacks data support.

[0044] The process of error correction and verification result output After confirming the logical inconsistency issue, the system organizes and integrates the analysis results to form unified error correction and verification output information.

[0045] The output information includes: It was clearly pointed out that there was a logical error in the experimental conclusions; Explain the cause of the error, namely, that the measurement results do not meet the evaluation criteria; Suggestions for revising the conclusion are provided, suggesting that the test conclusion should be adjusted to unqualified.

[0046] The results are output in the form of text annotations and structured review comments to help technical personnel quickly complete document revisions.

Claims

1. A method for intelligent error correction and logical consistency verification of scientific and technological documents based on a large model, characterized in that, S1: Access to and structured parsing of scientific and technological documents; S2: Semantic unit segmentation and feature extraction of scientific and technological documents; S3: Multi-level intelligent error correction based on a large model; S4: Logical consistency check of scientific and technological documents; S5: Output the combined results of error correction and verification.

2. The method for intelligent error correction and logical consistency verification of scientific and technological documents based on a large model according to claim 1, characterized in that, Step S1: Accessing and structured parsing technical documents includes: The technical documents that can be accessed include PDF, Word, PPT, and scanned formats; The following processing is performed on the accessed documents: format parsing and unified conversion; OCR recognition of scanned text; extraction of titles, chapter numbers, paragraphs, tables, formulas and figure captions; establishment of a document structure tree to record chapter levels and citation relationships; and the parsing results to form a structured intermediate representation of the scientific and technological documents.

3. The method for intelligent error correction and logical consistency verification of scientific and technological documents based on a large model according to claim 1, characterized in that, Step S2: Semantic unit segmentation and feature extraction includes: Based on structural analysis, the scientific and technological documents are divided into semantic units to form the smallest semantic processing units, and features are extracted from each semantic unit.

4. The method for intelligent error correction and logical consistency verification of scientific and technological documents based on a large model according to claim 3, characterized in that, Semantic unit division includes: technical description sentences; conditional description sentences; conclusion or judgment sentences; parameter definition and indicator description sentences.

5. The method for intelligent error correction and logical consistency verification of scientific and technological documents based on a large model according to claim 3, characterized in that, Extract the following features for each semantic unit: Characteristics of technical terms; Numerical and unit characteristics; Logical connective features; Chapter and context dependencies.

6. The method for intelligent error correction and logical consistency verification of scientific and technological documents based on a large model according to claim 1, characterized in that, Step S3: Multi-level intelligent error correction based on a large model includes: Based on the above semantic units and features, the large model is invoked to perform multi-level error correction: including language and format layer error correction, professional terminology and technical expression error correction, and parameter and numerical rationality error correction.

7. The method for intelligent error correction and logical consistency verification of scientific and technological documents based on a large model according to claim 6, characterized in that, Language and formatting layer error correction identifies and corrects the following issues: spelling errors, punctuation errors; inconsistent numbering formats; and issues related to scientific and technical writing standards. The system corrects errors in technical terminology and expressions by combining a domain terminology database with the semantic understanding capabilities of a large model to identify: the use of non-standard terms; and the mixed use of professional concepts. Ambiguity in technical terminology; The system corrects errors related to the rationality of parameters and values, verifies parameter ranges and unit usage, and identifies issues such as unit omissions and inconsistencies in dimensions.

8. The method for intelligent error correction and logical consistency verification of scientific and technological documents based on a large model according to claim 1, characterized in that, Step S4: Logical consistency verification of scientific and technological documents includes: Cross-chapter semantic consistency verification compares descriptions of the same object in different chapters to determine whether there are contradictory conclusions. Condition-conclusion logic verification analyzes whether there are logical jumps or invalid conclusions between the condition description and the conclusion statement. Verify the consistency between citations and numbering, checking whether the citation relationships of chapter numbers, figure numbers, and formula numbers are correct; During the logic verification process, the large model combines contextual information to output consistency judgment results and reasons.

9. The method for intelligent error correction and logical consistency verification of scientific and technological documents based on a large model according to claim 1, characterized in that, Step S5: Merge and output the error correction and verification results. Merge the error correction results and the logical consistency verification results to generate a unified output result, including: error location; error type classification; correction suggestions or risk warnings; logical conflict explanations; and output the results in the form of document comments, lists, or interfaces.

Images