Process parameter extraction and semantic verification method and system for gold ore dressing and smelting literature

Abstract

The application discloses a kind of process parameters extraction and semantic verification methods and systems for gold mine beneficiation literature.The method is associated with original text evidence by process parameter extraction agent to establish parameter name, value, unit, test group, process stage and the like;By semantic verification agent, evidence existence, process stage attribution, value and unit, test group association and result level check are executed;When field is rejected or doubtful, error correction feedback module generates field-level error correction instructions, and retry control module directs re-extraction to failed field according to error type;Finally, structured process parameter results with original text evidence positioning, verification status and version information are output;Effectively solve the problems of parameter attribution error, recovery rate level confusion, value fabrication and result untraceability in gold mine beneficiation multi-test group and multi-process stage literature, and realize the structured process parameter extraction that can be reviewed and traced.

Description

Technical Field

[0001] This invention relates to the field of information processing technology for gold ore beneficiation and metallurgy literature, specifically to a method and system for extracting process parameters and performing semantic verification for gold ore beneficiation and metallurgy literature, and more particularly to a method, system, and device for extracting process parameters and performing semantic verification for multi-agent collaborative processes in gold ore beneficiation and metallurgy literature. Background Technology

[0002] Research papers, technical reports, experimental records, and scanned texts on gold ore beneficiation typically record key process parameters such as ore type, gold grade of the raw ore, sulfur content, arsenic content, main mineral composition, grinding fineness, flotation reagent system, roasting conditions, leaching conditions, stage recovery rate, and total recovery rate. This information is crucial for conducting process comparisons, flow analysis, indicator evaluation, and knowledge base construction. Therefore, it is essential to accurately, uniformly, and traceably extract these parameters from the literature into structured data. Currently, these process parameters are mostly scattered throughout the text, tables, experimental results descriptions, and comparative experiments of unstructured literature. Furthermore, parameter names, values, units, experimental groups, and process stages may be separated on the page. Existing extraction methods are prone to problems such as numerical mismatches, experimental group mismatches, and process stage mismatches. Manual extraction is labor-intensive, inefficient, and unsuitable for large-scale literature processing. Therefore, how to automatically extract structured process parameters from gold ore beneficiation literature has become one of the technical problems that needs to be solved in related fields.

[0003] Currently, the mainstream technologies for extracting process parameters mainly fall into three categories: First, knowledge graph construction technologies for the gold mining field. These solutions focus on gold mining beneficiation and metallurgy knowledge modeling, knowledge fusion, and graph storage, primarily addressing the problems inherent in graph construction itself. However, they lack specialized processing for the front-end extraction of process parameters from unstructured gold mining beneficiation and metallurgy literature, making it difficult to ensure the accuracy of parameter sources, attributions, and hierarchies before graph formation. Second, document content extraction technologies based on large language models. These solutions can generate structured content from documents using semantic understanding, but they are mostly geared towards general text and lack specific applications. Domain-specific process parameter models in gold ore beneficiation literature struggle to reliably handle multi-stage process flows, the attribution of process conditions and experimental results, hierarchical relationships such as total recovery rate / stage recovery rate / leaching rate, and OCR anomalies such as numerical breaks, percentage sign separation, and misalignment of numerical values ​​and units. Thirdly, text processing technologies based on multi-agent collaboration, while capable of task division and workflow arrangement, are mostly used in general text generation or processing scenarios and do not include field-level semantic verification, error type identification, and targeted error correction and retry mechanisms for gold ore beneficiation process parameter extraction tasks. However, the aforementioned technologies have at least the following shortcomings: existing gold mine knowledge graph construction technologies focus on graph construction itself and lack specialized technical solutions for the front-end extraction process of literature process parameters; existing literature extraction technologies based on large language models lack parameter structure models for the gold ore beneficiation and metallurgy field, making it difficult to stably handle multi-stage process parameters and mining-specific numerical expressions; existing multi-agent collaborative technologies lack schemes for extracting beneficiation and metallurgy process parameters and independent semantic verification and error correction and retry mechanisms for extraction results; existing technologies cannot effectively solve the problem of confusion in the attribution of total recovery rate, stage recovery rate, process conditions, and process results; existing technologies cannot simultaneously establish the correlation between process parameters and original evidence, experimental groups, process stages, and result levels, and also cannot perform field-level error correction for different types of extraction errors.

[0004] In view of this, there is an urgent need to develop a multi-agent collaborative and evidence-constrained process parameter extraction and verification method and system for gold ore beneficiation and metallurgy literature. While preserving the original text evidence and page layout, the method should verify the existence of process parameter execution evidence, process stage attribution, numerical values ​​and units, experimental group association, and result hierarchy. Furthermore, it should perform targeted error correction and re-extraction for fields that fail verification or are questionable, thereby improving the accuracy, consistency, and traceability of process parameter extraction in complex gold ore beneficiation and metallurgy literature. Summary of the Invention

[0005] In view of the technical problems existing in the background art, where process parameters in gold ore beneficiation and metallurgy literature are scattered in the main text, tables, and figure captions, and where numerical values, units, test groups, process stages, and recovery rate levels are easily mismatched, this invention provides a method and system for extracting process parameters and performing semantic verification on gold ore beneficiation and metallurgy literature. This method performs text parsing, layout parsing, table extraction, and OCR numerical preprocessing on the input gold ore beneficiation and metallurgy literature to generate document content units with original text location information; a process parameter extraction agent extracts parameters such as ore properties, process conditions, reagent regimes, process results, and recovery rates according to a domain-specific process parameter model, and associates these parameters with test groups, process stages, and original text evidence; a semantic verification agent performs verification on the associated results for evidence existence, process stage attribution, numerical values ​​and units, test group association, and result level; when the verification result fails or is questionable, an error correction feedback module generates field-level error correction instructions, and a retry control module performs targeted re-extraction of the corresponding field according to the error type; when the verification passes, a structured process parameter result with original text evidence location information is output.

[0006] In a first aspect, embodiments of the present invention provide a method for extracting process parameters and performing semantic verification for gold ore beneficiation literature, comprising the following steps: The obtained gold ore beneficiation and metallurgy literature is subjected to text parsing, page layout parsing and table extraction. Then, the data is subjected to anomaly identification and candidate correction to obtain a set of literature evidence units. Each literature evidence unit is associated with a unique evidence identifier and at least one of the following: page number, page layout, chapter level and table row and column information. Based on the domain-specific process parameter model, a process parameter extraction agent extracts candidate process parameters from the literature evidence unit. The domain-specific process parameter model stipulates that each process parameter must be associated with at least the parameter name, parameter value, parameter unit, experimental group identifier, process stage identifier, result type, and evidence identifier. For values ​​not recorded in the original text and not clearly obtained from the calculation relationship in the original text, the process parameter extraction agent does not perform automatic completion and marks the corresponding field as pending verification. A semantic verification agent, independent of the process parameter extraction agent, performs field-level verification on each field of each candidate process parameter. The field-level verification includes at least evidence existence verification, process stage attribution verification, numerical value and unit verification, experimental group association verification, and result-level verification. Each field outputs a verification status of pass, rejection, or doubt. For fields whose verification status is rejected or questionable, a field-level error correction instruction is generated; the field-level error correction instruction includes at least the path of the field to be corrected, the current field value, the error type, and the evidence location information. Based on the error type in the field-level error correction instruction, select the corresponding targeted re-extraction strategy, and call the process parameter extraction agent to perform targeted re-extraction only on fields with a verification status of rejection or doubt, while fields with a verification status of pass remain unchanged; re-input the targeted re-extracted process parameter candidates into the semantic verification agent, and repeat the field-level verification, field-level error correction instruction generation, and targeted re-extraction until the corresponding field passes the verification or reaches the preset number of retries; The verified process parameters, along with the corresponding evidence identifiers and location information, will be output in a structured format.

[0007] As a further improvement of the present invention, the field-specific process parameter model includes: ore property parameters, including at least ore type, raw ore grade, sulfur content, arsenic content and main mineral composition; process condition parameters, including at least grinding fineness, flotation time, reagent dosage, roasting temperature, roasting time, leaching agent concentration, leaching temperature and leaching time; and experimental result parameters, including at least concentrate grade, tailings grade, leaching residue grade, stage recovery rate, total recovery rate and control improvement value.

[0008] As a further improvement of the present invention, the error types include at least the following: lack of evidence, conflict of process stage attribution, conflict of numerical value or unit, conflict of experimental group association, conflict of result hierarchy, and conflict of multiple candidate values. The targeted re-extraction strategy corresponds one-to-one with error types, including: for errors of missing evidence, narrowing the extraction scope to the original sentence, table cells, and adjacent sentences, and prohibiting the generation of numerical values ​​that do not exist in the original text; for conflicts in process stage attribution, adding chapter titles, table titles, and adjacent process descriptions as extraction inputs; for numerical or unit conflicts, rereading candidate correction values ​​and performing unit normalization; for conflicts in experimental group association, searching for other parameters with the same experimental number, the same sample number, or the same table row and column position; for conflicts in result hierarchy, searching all result statements within the same experimental group and redetermining the result type.

[0009] As a further improvement of the present invention, the result hierarchy verification includes: when the original text qualifier indicates a percentage change, the corresponding value is determined to be an absolute recovery rate result; when the original text qualifier indicates a percentage point change, the corresponding value is determined to be an improvement value; when the original text description is limited to the output material of a specific process stage, the corresponding value is determined to be a stage result rather than an overall result. The test group association verification includes: determining whether the process condition parameters and experimental result parameters belong to the same test group, the same sample group, or a clearly associated control group based on at least one of the test number, sample number, table row and column relationship, and chapter level; and rejecting process condition parameters and experimental result parameters that do not belong to the same test group.

[0010] As a further improvement of the present invention, the result types include stage recovery rate, total recovery rate, control recovery rate, and improvement value; wherein, when the original text describes the change in recovery rate as an increase or decrease of a specific order of magnitude relative to the control baseline, its result type is determined as improvement value, and the change is recorded as an improvement dimension that is different from the dimension of absolute recovery rate, rather than being converted to absolute recovery rate.

[0011] As a further improvement of the present invention, the anomaly identification and candidate correction includes: identifying at least one of the following anomalies: broken numbers, separation of decimal point and value, separation of percentage sign and value, loss of negative sign, separation of value and unit, and misalignment of values ​​in adjacent cells of a table; generating one or more candidate correction values ​​for the values ​​with anomalies, and retaining the original text, correction rules, corrected text and identification confidence, without overwriting the original text with the candidate correction values.

[0012] As a further improvement of the present invention, the structured output includes outputting process parameters as JSON format files, relational database records, tabular data, or knowledge graph nodes and relational data; each output process parameter also includes at least one of the following: original unit and normalized unit, verification status, version number, and number of retries.

[0013] As a further improvement of the present invention, the method further includes: saving the original document, document evidence unit, process parameter candidates, field-level verification results, field-level error correction instructions, field values ​​before revision, field values ​​after revision, and number of retries; establishing a connection through at least one of the document number, evidence identifier, test group identifier, and process stage identifier, so that any output process parameter can be traced back to the original page number, page location, chapter, table, test group, and complete verification and retry process record.

[0014] Secondly, embodiments of the present invention provide a system for extracting process parameters and verifying semantics in gold ore beneficiation literature, which is used to execute the aforementioned method for extracting process parameters and verifying semantics in gold ore beneficiation literature, including: The document preprocessing module is used to perform text parsing, page layout parsing, and table extraction on gold ore beneficiation and metallurgy documents, identify anomalies and correct candidates in the data, and obtain a set of document evidence units with associated unique evidence identifiers and location information. The process parameter extraction agent module is used to extract candidate process parameters from literature evidence units based on a domain-specific process parameter model. The domain-specific process parameter model specifies that each process parameter is associated with at least the parameter name, parameter value, parameter unit, experimental group identifier, process stage identifier, result type, and evidence identifier. The result type at least distinguishes between stage results, overall results, control results, and improvement results. The semantic verification agent module, independent of the process parameter extraction agent module, is used to perform field-level verification on each field of each process parameter candidate. The field-level verification includes at least evidence existence verification, process stage attribution verification, numerical value and unit verification, test group association verification, and result level verification, and outputs the verification status of each field as passed, rejected, or questionable. The error correction feedback module is used to generate field-level error correction instructions for fields whose verification status is rejected or questionable, including at least the path of the field to be corrected, the current field value, the error type, and evidence location information. The retry control module is used to select the corresponding targeted re-extraction strategy according to the error type, control the process parameter extraction intelligent agent module to only perform targeted re-extraction on fields whose verification status is rejected or questionable, and keep the fields that have passed the verification unchanged, and mark the fields that still fail the verification as manual review status when the preset number of retries is reached. The structured output module is used to output the verified process parameters along with the corresponding evidence identifiers and location information in a structured manner.

[0015] Thirdly, embodiments of the present invention provide an electronic device including a processor and a memory, wherein the memory stores a computer program that includes the above-described system and, when executed by the processor, implements the above-described method.

[0016] Beneficial effects: This invention addresses the problems of insufficient accuracy in extracting process parameters, incorrect parameter attribution, and difficulty in verifying extraction results in existing gold ore beneficiation literature through the following technical means: (1) By performing text analysis, page layout analysis and table structure analysis on the documents, the main text, title, table title, table rows and columns, figure captions and OCR recognition results are divided into evidence units with document number, page number, page layout and chapter level, so that each extracted process parameter can be associated with the corresponding original text evidence, thereby solving the problem that the extracted results lack original text basis and cannot be located for verification in the prior art; (2) By establishing a domain-specific process parameter model that includes ore properties, process stages, process conditions, reagent system, experimental results and literature metadata, and linking parameter names, values, units, test groups, process stages, result types and evidence identifiers, the process conditions and experimental results scattered in the text and tables can be represented according to a unified data structure, thereby reducing the problem of inconsistent parameter structures between different documents; (3) The process parameter extraction agent performs term recognition, parameter recognition, context acquisition, structure extraction and unit normalization, and the semantic verification agent independently performs evidence existence, process stage attribution, numerical value and unit, experimental group association and result level verification, so that parameter extraction and result verification are completed by different processing subjects, thereby reducing the risk of fictitious parameters, numerical errors or parameter attribution errors in a single model extraction. (4) By associating process parameters with test group identifiers, sample identifiers, table row and column information and chapter levels, it is possible to bind process conditions and experimental results in the same test group and identify parameters with the same name in different test groups, thereby reducing parameter mismatch problems in multi-test group literature. (5) By setting different result types for stage recovery rate, total recovery rate, control recovery rate and improvement value, and determining the recovery rate level based on the limiting words and context in the original text, it is possible to avoid misidentifying improvement values ​​such as "increase by 1.17 percentage points" as absolute recovery rate, thereby reducing the problem of result level confusion in multi-stage process literature; (6) The error correction feedback module outputs field-level error correction instructions containing the error field path, current field value, error type, original evidence, suggested correction value and re-extraction strategy. The retry control module performs targeted re-extraction only on fields that fail the verification or are questionable, so that fields that have passed the verification remain unchanged, thereby reducing result drift and duplicate errors caused by full re-extraction. (7) By saving the original documents, evidence units, parameter candidates, verification results, error correction instructions, field values ​​before and after revision and number of retries, the final structured parameters can be traced back to the original page number, page position, test group and processing process, thereby improving the traceability, verifiability and anomaly handling capabilities of the process parameter extraction results.

[0017] In summary, the technical improvement of this invention does not lie in using multi-agent or large language models alone, but in combining literature evidence localization, domain-specific parameter representation, experimental group and process stage association, field-level semantic verification, and error type-driven targeted re-extraction to form a process parameter extraction and verification process for gold ore beneficiation literature.

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0019] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the present invention will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0020] Figure 1 This is a schematic diagram of the structure of the process parameter extraction and semantic verification system for gold ore beneficiation literature provided in this embodiment of the invention.

[0021] Figure 2 This is a flowchart illustrating the method for extracting process parameters and performing semantic verification on gold ore beneficiation literature provided in this embodiment of the invention. Detailed Implementation

[0022] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the invention, are intended to cover non-exclusive inclusion.

[0024] In the description of the embodiments of this invention, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this invention, "multiple" means two or more, unless otherwise explicitly defined.

[0025] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0026] To address the challenges of existing methods in accurately extracting and assigning process parameters and distinguishing result levels such as recovery rate / leaching rate in gold mining literature with multiple test groups, multiple process stages, multiple material flows, and OCR noise scenarios, and the lack of field-level evidence verification, targeted error correction, and traceability mechanisms, leading to easily mismatched, confusing, and difficult-to-verify extraction results, this invention provides a method and system for extracting process parameters and performing semantic verification in gold mining literature. While retaining the document's page numbers and layout information, the method associates the numerical values, units, test groups, process stages, and original text evidence of the process parameters. It automatically judges the extraction results through evidence existence, process stage attribution, numerical and unit association, test group association, and result level verification. For fields that fail verification or are questionable, field-level targeted error correction and re-extraction are performed based on the error type. This method solidifies knowledge in the gold mining field into a domain-specific process parameter model with result type hierarchies, and drives field-level targeted retries with semantic verification independent of the extraction process, forming a constrained technical route of extraction-verification-correction-retry. The semantic verification agent and the extraction agent have separate functions. The agent performs five types of field-level verification: evidence existence, process stage attribution, numerical unit, experimental group association, and result level. It intercepts numerical values ​​that do not exist in the original text, mismatches across experimental groups, and errors in attribution across stages before output. Error correction feedback generates field-level error correction instructions containing field path, current value, error type, and evidence location according to the error type. Retry control selects the corresponding retry strategy according to the error type. Only fields that failed or are questionable are re-extracted. Fields that have passed the verification remain unchanged, avoiding the result drift introduced by full re-extraction. Fields that exceed the limit are marked for manual review instead of automatically generating speculative values.

[0027] Example 1 Please see Figure 1 As shown, Embodiment 1 of the present invention provides a system for extracting process parameters and performing semantic verification on gold ore beneficiation literature. The system is used to extract process parameters such as ore properties, process conditions, reagent regimes, process results, and recovery rates from gold ore beneficiation papers, experimental reports, experimental records, and scanned literature. The system then performs original text evidence verification, process stage attribution verification, numerical and unit verification, experimental group association verification, and result hierarchy verification on the extracted process parameters.

[0028] Specifically, the system includes a document input module, a document preprocessing module, a relevant content filtering module, a process parameter extraction agent module, a semantic verification agent module, an error correction feedback module, a retry control module, a structured output module, and a storage module. The connections between these modules are as follows: the output of the document input module is connected to the input of the document preprocessing module; the output of the document preprocessing module is connected to the input of the relevant content filtering module; the output of the relevant content filtering module is connected to the input of the process parameter extraction agent module; the output of the process parameter extraction agent module is connected to the input of the semantic verification agent module; the output of the semantic verification agent module is connected to the input of the error correction feedback module; the output of the error correction feedback module is connected to the retry input of the process parameter extraction agent module via the retry control module, forming an extraction-verification-error correction-retry closed loop; the pass result output of the semantic verification agent module is connected to the input of the structured output module; and the storage module is connected to each of the above modules for full-scale trace storage.

[0029] In this system, the process parameter extraction agent module and the semantic verification agent module are two independent processing entities with separate functions: the former is only responsible for generating parameter candidates based on the domain-specific process parameter model, while the latter is only responsible for making field-level decisions on the candidates based on the original evidence and domain constraints. The two communicate data through structured messages, do not replace each other, and check and balance each other. The error correction feedback module and the retry control module form a closed-loop scheduling hub, which transforms the verification decision into an executable correction action, realizing targeted re-extraction only for failed fields.

[0030] The specific configurations and functions of each module in the system are as follows: (1) Document Input Module The document input module is used to receive gold ore beneficiation and metallurgy papers, experimental reports, experimental records, scanned documents, PDF files, or text files converted by document parsing tools.

[0031] The document input module generates a unique document identifier (document_id) for each document to be processed and records document metadata such as document title, document source, file type, number of pages, and receipt time.

[0032] For electronic documents containing a text layer, the document input module sends the text layer content, page information, and layout information to the document preprocessing module; for scanned documents, the document input module sends the page image to the document preprocessing module for OCR recognition.

[0033] The document input module outputs an original document object, which includes at least: document_id, original document file, document type, page image, text layer content, document page number, and document metadata.

[0034] (2) Document preprocessing module The document preprocessing module is connected to the document input module and is used to perform text parsing, page layout parsing, formula cleaning, paragraph parsing, table extraction, and OCR numerical preprocessing on the original documents.

[0035] The document preprocessing module specifically performs the following processing: ① Identify content areas in a document, such as titles, chapters, main text paragraphs, tables, table titles, table rows, table columns, table cells, figure captions, and footnotes; ② Sort the content areas according to the reading order of the document pages, and generate a unique evidence identifier (evidence_id) for each content area; ③ Save the document number, page number, page coordinates, chapter level, table number, row number, column number, original text and content type corresponding to each evidence unit; ④ Clean up formula symbols, subscripts, superscripts, special symbols, hyphens, and OCR recognition residues; ⑤ Identify abnormal formats such as broken numbers, separated decimal points, missing minus signs, separated percent signs, and separation of numerical values ​​and units; ⑥ For numerical values ​​with OCR anomalies, generate candidate correction results and save the text before correction, correction rules, text after correction, and OCR recognition confidence. ⑦ Reconstruct the table structure based on the table headers, row headers, column headers, and cell position relationships.

[0036] Each evidence unit includes at least the following: evidence_id, document_id, page, bbox, block_type, section_path, table_id, row_id, column_id, original_text, normalized_text, and ocr_confidence.

[0037] Here, bbox represents the page coordinates of the evidence unit, block_type represents the content type of the evidence unit, and section_path represents the chapter level to which the evidence unit belongs.

[0038] The document preprocessing module outputs a set of evidence units, paragraph structure information, table structure information, OCR numerical candidate set, and document metadata, and sends the above data to the relevant content filtering module.

[0039] (3) Related content filtering module The relevant content filtering module is connected to the document preprocessing module and is used to filter document content related to the extraction of process parameters from the evidence unit set.

[0040] The relevant content filtering module matches evidence units based on a preset gold ore beneficiation and metallurgy terminology dictionary, chapter title dictionary, and process parameter keywords, filtering abstracts, experimental sections, process condition sections, result sections, comparative test sections, table sections, and contextual content related to the above content.

[0041] The terminology dictionary for gold ore beneficiation and metallurgy shall include at least: ore type, mineral name, process stage, equipment name, reagent name, parameter name, process index, and result index.

[0042] The process stages include at least crushing, grinding, classification, gravity separation, flotation, roasting, leaching, adsorption, desorption, and tailings treatment.

[0043] The relevant content filtering module also establishes a candidate context for parameter extraction based on chapter path, table number, test number, sample number, table row and column relationship, and page proximity relationship between parameter name and value.

[0044] The relevant content filtering module outputs a candidate context set, which includes at least: evidence unit, chapter title, table title, test number, sample number, parameter keywords, and adjacent text content.

[0045] (4) Process parameter extraction agent module (hereinafter referred to as process parameter extraction agent) The process parameter extraction agent module is connected to the relevant content filtering module and is used to extract structured process parameters from the candidate context set according to the domain-specific process parameter model.

[0046] The process parameter extraction intelligent agent module includes a terminology recognition unit, a context construction unit, a parameter extraction unit, and a unit normalization unit.

[0047] The process parameter extraction agent module specifically performs the following processing: ① Use a gold ore beneficiation terminology dictionary and entity recognition model to identify ore type, mineral name, process stage, equipment name, reagent name, parameter name, value, unit, test number, and sample number; ② Using the identified parameter name or value as the center, obtain the corresponding evidence unit, adjacent sentences, chapter titles, table titles, table row titles, and table column titles to form the parameter extraction context; ③ Based on the table number, row number, column number, test number, sample number, and chapter path, establish candidate associations between parameter names, parameter values, units, process stages, and experimental results; ④ Output structured parameter candidates according to the preset domain-specific process parameter model; ⑤ Standardize the parameter values ​​and units, and save the original values, original units, standardized values, and standardized units at the same time; ⑥ Associate each parameter field with the corresponding evidence_id, page number, and page coordinates; ⑦ For values ​​that cannot be found in the original evidence and cannot be explicitly calculated from the original text, do not perform automatic completion, and mark the field as pending verification.

[0048] The process parameter extraction agent module outputs a set of parameter candidates. Each parameter candidate includes at least: field path, parameter name, original text, normalized value, original unit, normalized unit, trial_id, stage_id, result_scope, evidence_id, page number, page coordinates, extraction confidence score, and inference tag.

[0049] (5) Semantic Validation Agent Module (hereinafter referred to as Semantic Validation Agent) The semantic verification agent module is connected to the process parameter extraction agent module and is used to receive the parameter candidate set, the evidence unit set, and the domain-specific process parameter model, and to perform field-level semantic verification on the extraction results.

[0050] The semantic verification agent module performs at least the following verifications: ① Evidence existence verification is used to determine whether the parameter name, value, unit and qualifier exist in the corresponding original text evidence, or whether they can be obtained through the explicit calculation relationship in the original text; ② Process stage attribution verification, used to determine whether process conditions and experimental results belong to the correct process stage; ③ Numerical and unit verification, used to check numerical format, unit type, unit conversion relationship, percentage sign binding relationship, decimal point integrity and parameter value range; ④ Test group correlation verification is used to determine whether the process conditions and experimental results belong to the same test group, sample group or clearly related control group; ⑤ Recovery rate level (result level) verification, used to distinguish between stage recovery rate, total recovery rate, control recovery rate and improvement value.

[0051] The semantic verification intelligent agent module outputs three verification statuses for each field: pass, reject, or questionable, and outputs the field path, current field value, verification type, error reason, original evidence, and evidence location information.

[0052] (6) Error correction feedback module The error correction feedback module is connected to the semantic verification intelligent agent module and is used to receive the rejected or questionable fields output by the semantic verification intelligent agent module and generate field-level error correction suggestions.

[0053] The field-level error correction suggestions should include at least: field path, current field value, error type, suggested correction value, error reason, original evidence, evidence location, and targeted re-extraction strategy.

[0054] The error types include at least: evidence_missing, stage_conflict, value_or_unit_conflict, trial_conflict, recovery_scope_conflict, and multiple_candidate_conflict.

[0055] (7) Retry control module The retry control module is connected to the process parameter extraction agent module and the error correction feedback module, respectively, and is used to control the field range, context range and number of retries for targeted re-extraction based on field-level error correction suggestions.

[0056] The retry control module selects the corresponding targeted re-extraction strategy based on the error type: ① For errors related to missing evidence, the extraction scope will be narrowed down to the original sentence, table cells, and adjacent sentences, and automatic inference will be prohibited; ② For incorrect process stage attribution, add chapter titles, table titles, material descriptions, and adjacent process stages as inputs; ③ For numerical or unit errors, reread the OCR candidate values ​​and perform unit normalization. ④ For incorrect association of test groups, retrieve relevant parameters under the same table, test number, or sample number; ⑤ For recovery rate hierarchical errors, retrieve all recovery rate expressions in the same experimental group and redetermine the result type.

[0057] The retry control module only re-extracts fields that fail or are questionable in the verification process, while fields that have already passed the verification remain unchanged.

[0058] Each retry should save at least the following: previous round extraction results, failed fields, reasons for failure, candidate evidence, values ​​before revision, values ​​after revision, and number of retries.

[0059] If the number of retries reaches the preset limit and the field value still cannot be determined, the corresponding field will be marked as manually reviewed and the automatic generation of new speculative values ​​will be stopped.

[0060] (8) Structured output module The structured output module is connected to the semantic verification intelligent agent module and is used to receive the process parameter results that have passed semantic verification and output them as structured data.

[0061] Each process parameter output by the structured output module includes at least the following: document_id, trial_id, stage_id, parameter_type, value, unit, result_scope, evidence_id, page, bbox, validation_state, and version.

[0062] The structured output module supports outputting results as: JSON format structured result files; relational database records; tabular data; knowledge graph nodes and relational data.

[0063] For fields that are judged as rejected, questionable, or subject to manual review, the structured output module synchronously outputs the error type, error reason, evidence location, candidate value, and review status.

[0064] (9) Storage module The storage module is connected to the structured output module and is used to save the original document, document metadata, evidence units, parameter candidate set, final structured result, semantic verification result, error correction suggestions and retry record.

[0065] The storage module associates documents with document_id, original evidence with evidence_id, experimental groups with trial_id, and process stages with stage_id.

[0066] Through the above relationships, the system can query the corresponding original page number, page coordinates, chapter, table, test group, process stage and verification process based on any structured process parameter.

[0067] This embodiment further provides an electronic device, including a processor and a memory, the memory storing a computer program that includes the system described in the above embodiment, and the processor executing the computer program to implement the steps in the following method embodiments.

[0068] Example 2 Please see Figure 2 As shown, Embodiment 2 of the present invention provides a method for extracting and semantically verifying process parameters (multi-agent collaboration) in gold ore beneficiation literature, comprising the following steps: Step S101: Obtain the document to be processed. Acquire gold ore beneficiation and metallurgy papers, test reports, experimental records, scanned documents, or PDF documents to be processed, and generate a unique document identifier (document_id) for each document.

[0069] Simultaneously, it records metadata such as document title, source, file type, number of pages, and receipt time, and sends the original document to the document preprocessing module.

[0070] Step S102: Perform literature preprocessing. The document preprocessing module performs text parsing, page layout parsing, heading level recognition, paragraph parsing, table extraction, formula cleaning, and OCR numerical preprocessing on the input documents.

[0071] The document page is divided into content units such as title, body paragraphs, tables, table titles, table rows, table columns, table cells, figure captions, and footnotes, and a unique evidence identifier (evidence_id) is generated for each content unit according to the page reading order.

[0072] For abnormal formats such as broken numbers, separated decimal points, separated percent signs, missing minus signs, and separation of numerical values ​​and units generated by OCR recognition, candidate correction results are generated, and the text before correction, correction rules, text after correction, and OCR confidence scores are saved.

[0073] This step outputs a set of evidence units that includes the original text, page numbers, page coordinates, chapter level, table number, row and column information, and OCR numerical candidate data.

[0074] Step S103: Filter relevant content The relevant content filtering module filters abstracts, experimental sections, process condition sections, results sections, comparative test sections, tables, and their related contexts from the evidence unit set based on the preset gold ore beneficiation and metallurgy terminology dictionary, chapter title dictionary, and process parameter keywords.

[0075] The relevant content filtering module establishes a set of candidate contexts for parameter extraction based on chapter path, table number, test number, sample number, table row and column relationship, and page proximity relationship between parameter name and value.

[0076] This step outputs a candidate context set, which includes at least evidence units, chapter titles, table titles, test numbers, sample numbers, and adjacent text content.

[0077] Step S104: Perform the first parameter extraction. The candidate context set is input into the process parameters to extract the intelligent agent.

[0078] The process parameter extraction agent uses a gold ore beneficiation terminology dictionary and entity recognition model to identify ore type, mineral name, process stage, equipment name, reagent name, parameter name, value, unit, test number, and sample number, and generates structured process parameter candidates based on chapter titles, table titles, table row and column titles, and adjacent context.

[0079] The structured process parameter candidates include at least: parameter name, original text, numerical value, unit, test group, process stage, result type, evidence identifier, page number, and page coordinates.

[0080] This step outputs the first version of candidate structured process parameters.

[0081] Step S105: Prepare semantic verification data The first version of the structured process parameter candidate results, the corresponding original text evidence units, the domain-specific process parameter model, and the relevant context are all input into the semantic verification agent.

[0082] The original text evidence unit includes at least the original text, page number, page coordinates, chapter level, table number, row and column information, and OCR confidence level.

[0083] Step S106, Perform semantic verification The semantic validation agent performs field-level validation on the first version of the candidate results for structured process parameters. The validation includes at least the following: (1) Evidence existence verification is used to determine whether the parameter name, value, unit and qualifier exist in the corresponding original text evidence, or whether they can be obtained from the explicit calculation relationship in the original text; (2) Process stage attribution verification, used to determine whether the process conditions and experimental results belong to the correct process stage; (3) Numerical and unit verification, used to check numerical format, unit type, unit conversion relationship, percentage sign binding relationship and parameter value range; (4) Test group correlation verification, used to determine whether the process conditions and experimental results belong to the same test group, sample group or clearly related control group; (5) Recovery rate hierarchical verification, used to distinguish between stage recovery rate, total recovery rate, control recovery rate and improvement value.

[0084] The semantic verification agent outputs the verification status of each field as passed, rejected, or questionable, and generates the field path, current field value, verification type, error reason, original evidence, and evidence location information.

[0085] Step S107: Output the verification result. When all key fields in the candidate results of structured process parameters pass semantic validation, the structured output module outputs the final structured process parameter results.

[0086] The final structured process parameter results include at least the following: document_id, trial_id, stage_id, parameter_type, value, unit, result_scope, evidence_id, page, bbox, validation_state, and version.

[0087] Where trial_id represents the experimental group or sample group, stage_id represents the process stage, result_scope represents the stage result, overall result, control result or improvement result, and evidence_id represents the corresponding original evidence unit.

[0088] Step S108: Generate field-level error correction instructions When any key field is determined to be rejected or questionable, the error correction feedback module generates a field-level error correction instruction based on the semantic verification result.

[0089] The error correction instructions include at least: field path, current field value, error type, suggested correction value, error reason, original evidence, evidence location, suggested context scope, and re-extraction strategy.

[0090] Among them, error types include at least missing evidence, conflicting stage attribution, conflicting numerical values ​​or units, conflicting trial groups, conflicting recovery rates at different levels, and conflicting multiple candidate values.

[0091] Step S109, perform targeted re-sampling The retry control module receives field-level error correction instructions and determines the field to be re-extracted, the context range, and the processing strategy based on the error type. It then calls the process parameter extraction agent again to perform field-level targeted re-extraction.

[0092] For errors related to missing evidence, narrow the extraction scope to the original sentence, table cell, and adjacent sentences; for errors related to stage attribution, add chapter titles, table titles, material descriptions, and adjacent process stages; for errors related to numerical values ​​or units, reread the OCR candidate values ​​and perform unit normalization; for conflicts in test groups, retrieve relevant parameters under the same test number, sample number, or table row and column.

[0093] The retry control module only updates fields that fail or are questionable in the verification process, while fields that have passed the verification remain unchanged. It also saves the original value, the revised value, the reason for failure, and the number of retries.

[0094] Step S110: Determine the processing result The parameter results after targeted re-extraction are re-input into the semantic verification agent, and steps S105 to S109 are repeated.

[0095] When a re-extracted field passes semantic validation, the corresponding field is updated and the final structured process parameter result is output. When the field value still cannot be determined after the number of retries reaches the preset maximum number, the corresponding field is marked as manual_review and the candidate value, error reason and original evidence location of the field are output. No new speculative value is automatically generated.

[0096] In this embodiment, the structured process parameter results include at least: (1) Document number; (2) Test group or sample group; (3) Ore type and ore properties; (4) A sequence of process stages or one or more process stages; (5) Process conditions corresponding to each process stage; (6) Experimental results corresponding to each process stage; (7) Stage recovery rate, total recovery rate, control recovery rate, and improvement value; (8) The original evidence identifiers, page numbers, and page layouts corresponding to each parameter; (9) Verification status and version information of each parameter.

[0097] Example 3 Based on Example 1, Example 3 further employs a rule verification module and a model semantic verification module to jointly implement the semantic verification function.

[0098] The rule verification module receives the structured parameter candidates and corresponding evidence units output by the process parameter extraction agent, and performs the following verification: (1) Parameter field integrity verification; (2) Numerical format verification; (3) Verification of parameter values ​​and units; (4) Unit conversion verification; (5) Verification of percentage and percentage point expressions; (6) Parameter value range verification; (7) Verification of the existence of original evidence markers.

[0099] The model semantic verification module receives structured parameter candidates, original text evidence units, chapter context, and table context, and performs the following verifications: (1) Verification of the semantic correspondence between parameters and original evidence; (2) Verification of the attribution relationship between parameters and process stages; (3) Verification of the correlation between parameters and experimental groups or sample groups; (4) Verification of the correspondence between process conditions and experimental results; (5) Validation of the hierarchical relationship between stage recovery rate, total recovery rate, control recovery rate and improvement value.

[0100] The rule verification module first verifies the format, value, and unit of the candidate parameters. After the rule verification passes, the model semantic verification module then verifies the original evidence, process stage, experimental group, and result level of the candidate parameters. If any verification module determines that a field has an error or uncertainty, it sends the field and the corresponding verification result to the error correction feedback module.

[0101] The rule validation module and model semantic validation module described above can run as independent modules or be integrated into the semantic validation agent module. Regardless of the implementation method, they both output field-level validation results in a uniform format.

[0102] Example 4 Based on the method in Example 2, Example 4 further adopts a structured output module, which can output the process parameter results that have passed semantic verification as JSON files, relational database records, tabular data, or knowledge graph nodes and relational data according to subsequent application requirements.

[0103] When the output is a JSON file, each process parameter is organized according to the preset parameter fields, and the document number, test group, process stage, parameter name, value, unit, evidence identifier, page number, page position and verification status are saved.

[0104] When the output is a relational database record, the literature, evidence unit, test group, process stage, parameter results, verification results and retry records are stored in the corresponding data tables, and the association is established through the literature number, evidence identifier, test group identifier and parameter number.

[0105] When the output is tabular data, the parameters of the same test group or the same process stage are organized in a columnar manner according to parameter name, value, unit and result type.

[0106] When the output is knowledge graph data, literature, experimental group, process stage, process parameters and experimental results are used as nodes, and the relationships such as "originating from", "belonging to the experimental group", "belonging to the process stage" and "generating results" are used as edges.

[0107] The different output methods mentioned above only change the storage medium of the structured results, without changing the relationship between process parameters and original evidence, experimental groups, process stages, and result types, nor do they change the semantic verification and field-level retry process.

[0108] Although different combinations of verification modules or different structured output methods are used in embodiments 3 and 4 above, the following technical features are retained: (1) Perform text parsing, page layout parsing, table parsing and OCR numerical preprocessing on gold ore beneficiation literature, and generate evidence units with original text location information; (2) Based on the domain-specific process parameter model, associate the parameter name, value, unit, test group, process stage, result type and original evidence; (3) The extraction results are verified by a verification process independent of the parameter extraction process, including evidence, attribution, numerical values, test group and result level verification. (4) When the verification result is rejection or doubt, generate field-level error correction instructions based on the error field and error type; (5) The retry control module performs targeted re-extraction only on fields that fail or are questionable, and retains fields that have passed the verification and historical processing records.

[0109] Example 5 Based on Example 1, Example 5 further includes an OCR numerical correction submodule.

[0110] The OCR numerical correction submodule receives the OCR text, page image, evidence unit, and OCR confidence score output by the document preprocessing module, and handles the following anomalies: (1) There are spaces or breaks between numeric characters; (2) The decimal point is separated from the digits; (3) The percentage sign is separated from the numerical value; (4) The negative sign was misidentified or lost; (5) There is an error gap between the numerical value and the unit; (6) The numbers in adjacent cells of the table are misaligned.

[0111] The OCR numerical correction submodule does not directly overwrite the original OCR text, but generates one or more candidate values ​​and saves them respectively: the original OCR text, the candidate correction value, the correction rule, the evidence identifier, the page number, the page coordinates, and the OCR confidence level.

[0112] Candidate correction values ​​are confirmed by a semantic verification agent by combining parameter type, numerical format of the same table column, contextual qualifiers, and parameter value range.

[0113] For example, when the OCR recognition result is 91.25%, the OCR numerical correction submodule generates a candidate value of 91.25% and records "removing spaces between numbers, merging decimal points, and binding percent signs" as numerical correction rules.

[0114] Example 6 Based on Example 1, Example 6 further includes a recovery rate hierarchical identification submodule.

[0115] The recovery rate hierarchical identification submodule receives evidence units containing recovery rate keywords and identifies the result type of recovery rate based on recovery rate qualifiers, process stage, test group, and contextual relationship.

[0116] The result types include at least: stage_recovery: stage recovery rate; overall_recovery: overall recovery rate; control_recovery: control group recovery rate; improvement: absolute improvement value; improvement_percentage_point: percentage point improvement value.

[0117] The original text states: flotation recovery rate was 78.4%; total recovery rate after roasting-leaching was 91.25%; an increase of 1.17 percentage points compared to the control group. The recovery rate hierarchical identification submodule outputs: flotation stage recovery rate: 78.4%; total recovery rate after roasting-leaching: 91.25%; increase compared to the control group: 1.17 percentage points. The "increase of 1.17 percentage points" is recorded as a numerical value of 1.17, expressed in percent_points, and is not converted to an absolute recovery rate.

[0118] Example 7 Based on Example 1, Example 7 further includes a multi-stage process attribution refinement submodule.

[0119] The multi-stage process attribution refinement submodule receives parameter candidates, chapter titles, table titles, process stage keywords, material names, test numbers, and adjacent context, and determines the process stage to which the parameter belongs based on the following factors: (1) Matching relationship between parameter names and process stage terms; (2) The correspondence between the chapter titles containing the parameters and the process stages; (3) The correspondence between the table title, row title, and column title of the parameter and the process stage; (4) The correspondence between the input and output materials corresponding to the parameters and the process stages; (5) The correspondence between the order of appearance of parameters in the text and the order of process stages.

[0120] When the above factors correspond to multiple candidate process stages, the candidate process stages, matching results, and corresponding scores are written into the stage attribution record; when the difference between the highest candidate stage and the second candidate stage is less than the preset judgment difference, the stage attribution of the parameter is marked as questionable and submitted for manual review or targeted re-extraction.

[0121] Example 8 Based on Example 1, Example 8 further includes a field-level error correction and retry submodule.

[0122] The field-level error correction and retry submodule selects the retry range and retry method based on the error type output by the semantic verification agent: (1) When the error type is missing evidence, the context is narrowed to the original sentence, table cell and adjacent sentences, and the generation of numerical values ​​that do not exist in the original text is prohibited; (2) When the error type is stage affiliation conflict, add the chapter title, table title, material description and adjacent process stage to the retry input; (3) When the error type is a numerical or unit conflict, reread the OCR candidate values ​​and perform unit conversion; (4) When the error type is test group conflict, retrieve the relevant parameters under the same test number, sample number or table row and column; (5) When the error type is recovery rate level conflict, retrieve all recovery rate expressions within the same experimental group and redetermine the stage recovery rate, total recovery rate, control recovery rate or boost value.

[0123] Each retry only updates fields that failed or are questionable in the validation process; fields that passed validation remain unchanged. Fields that cannot be determined after reaching the maximum number of retries are marked as requiring manual review, and candidate values, original evidence, and the reason for the error are output.

[0124] The OCR numerical correction submodule in Example 5 can reduce numerical errors caused by broken numbers, decimal point separation, percent sign separation, and unit misalignment.

[0125] The recovery rate hierarchical identification submodule in Example 6 can avoid misidentifying "increase by 1.17 percentage points" as "increase by 71.17%" or other absolute recovery rates.

[0126] By using the multi-stage process classification and refinement sub-modules in Example 7, parameter mismatches between multiple process stages such as grinding, flotation, roasting, and leaching in the same literature can be reduced.

[0127] The field-level error correction and retry submodule in Example 8 enables targeted correction of only erroneous fields while retaining those that have passed, thereby reducing result drift caused by full re-extraction.

[0128] In summary, this invention discloses a method and system for extracting process parameters and performing semantic verification on gold ore beneficiation and metallurgy literature. This method performs text parsing, layout parsing, table extraction, and numerical preprocessing on gold ore beneficiation and metallurgy literature, generating document evidence units with document number, page number, layout location, chapter level, table rows and columns, and evidence identifiers. Based on a domain-specific process parameter model, it uniformly represents ore properties, process stages, process conditions, reagent regimes, experimental results, and document metadata. A process parameter extraction agent establishes associations between parameter names, values, units, experimental groups, process stages, result types, and original text evidence. A semantic verification agent performs verification of evidence existence, process stage attribution, values ​​and units, experimental group associations, and result level. When a field is rejected or questionable, an error correction feedback module generates field-level error correction instructions, and a retry control module re-extracts failed fields according to error type, leaving passed fields unchanged and marking out-of-limit fields for manual review. Finally, it outputs structured process parameter results with original text evidence location, verification status, and version information. This invention solves the problems of incorrect parameter attribution, confusion of recovery rate levels, fictitious numerical values, and untraceable results in literature on multiple experimental groups and multiple process stages of gold ore beneficiation and metallurgy, and realizes the extraction of structured process parameters that are verifiable and traceable.

[0129] It should be noted that the present invention is not limited to the above-described embodiments. The above embodiments are merely examples, and any embodiments that have the same structure and perform the same effects as the technical concept within the scope of the present invention are included within the scope of the present invention. Furthermore, various modifications that can be conceived by those skilled in the art to the embodiments, and other ways of constructing by combining some of the constituent elements of the embodiments, without departing from the spirit of the present invention, are also included within the scope of the present invention.

Claims

1. A method for extracting process parameters and performing semantic verification in gold ore beneficiation literature, characterized in that, Includes the following steps: The obtained gold ore beneficiation and metallurgy literature is subjected to text parsing, page layout parsing and table extraction. Then, the data is subjected to anomaly identification and candidate correction to obtain a set of literature evidence units. Each literature evidence unit is associated with a unique evidence identifier and at least one of the following: page number, page layout, chapter level and table row and column information. Based on the domain-specific process parameter model, a process parameter extraction agent extracts candidate process parameters from the literature evidence unit. The domain-specific process parameter model stipulates that each process parameter must be associated with at least the parameter name, parameter value, parameter unit, experimental group identifier, process stage identifier, result type, and evidence identifier. For values ​​not recorded in the original text and not clearly obtained from the calculation relationship in the original text, the process parameter extraction agent does not perform automatic completion and marks the corresponding field as pending verification. A semantic verification agent, independent of the process parameter extraction agent, performs field-level verification on each field of each candidate process parameter. The field-level verification includes at least evidence existence verification, process stage attribution verification, numerical value and unit verification, experimental group association verification, and result-level verification. Each field outputs a verification status of pass, rejection, or doubt. For fields whose verification status is rejected or questionable, a field-level error correction instruction is generated; the field-level error correction instruction includes at least the path of the field to be corrected, the current field value, the error type, and the evidence location information. Based on the error type in the field-level error correction instruction, select the corresponding targeted re-extraction strategy, and call the process parameter extraction agent to perform targeted re-extraction only on fields with a verification status of rejection or doubt, while fields with a verification status of pass remain unchanged; re-input the targeted re-extracted process parameter candidates into the semantic verification agent, and repeat the field-level verification, field-level error correction instruction generation, and targeted re-extraction until the corresponding field passes the verification or reaches the preset number of retries; The verified process parameters, along with the corresponding evidence identifiers and location information, will be output in a structured format.

2. The method for extracting process parameters and performing semantic verification for gold ore beneficiation literature according to claim 1, characterized in that, The domain-specific process parameter model includes: ore property parameters, including at least ore type, raw ore grade, sulfur content, arsenic content, and main mineral composition; process condition parameters, including at least grinding fineness, flotation time, reagent dosage, roasting temperature, roasting time, leaching agent concentration, leaching temperature, and leaching time; and experimental result parameters, including at least concentrate grade, tailings grade, leaching residue grade, stage recovery rate, total recovery rate, and control enhancement value.

3. The method for extracting process parameters and performing semantic verification for gold ore beneficiation literature according to claim 1, characterized in that, The error types include at least the following: missing evidence, conflict of process stage attribution, conflict of numerical values ​​or units, conflict of experimental group association, and conflict of result hierarchy. The targeted re-extraction strategy corresponds one-to-one with error types, including: for errors of missing evidence, narrowing the extraction scope to the original sentence, table cells, and adjacent sentences, and prohibiting the generation of numerical values ​​that do not exist in the original text; for conflicts in process stage attribution, adding chapter titles, table titles, and adjacent process descriptions as extraction inputs; for numerical or unit conflicts, rereading candidate correction values ​​and performing unit normalization; for conflicts in experimental group association, searching for other parameters with the same experimental number, the same sample number, or the same table row and column position; for conflicts in result hierarchy, searching all result statements within the same experimental group and redetermining the result type.

4. The method for extracting process parameters and performing semantic verification for gold ore beneficiation literature according to claim 1, characterized in that, The result hierarchy verification includes: when the original text qualifier indicates a percentage change, the corresponding value is determined to be an absolute recovery rate result; when the original text qualifier indicates a percentage point change, the corresponding value is determined to be an improvement value; when the original text description is limited to the output material of a specific process stage, the corresponding value is determined to be a stage result rather than an overall result. The test group association verification includes: determining whether the process condition parameters and experimental result parameters belong to the same test group, the same sample group, or a clearly associated control group based on at least one of the test number, sample number, table row and column relationship, and chapter level; and rejecting process condition parameters and experimental result parameters that do not belong to the same test group.

5. The method for extracting process parameters and performing semantic verification for gold ore beneficiation literature according to claim 1, characterized in that, The result types include stage recovery rate, total recovery rate, control recovery rate, and improvement value; wherein, when the original text describes the change in recovery rate as an increase or decrease of a specific order of magnitude relative to the control baseline, its result type is determined as improvement value, and the change is recorded as an improvement dimension that is different from the dimension of absolute recovery rate, and is not converted to absolute recovery rate.

6. The method for extracting process parameters and performing semantic verification for gold ore beneficiation literature according to claim 1, characterized in that, The anomaly identification and candidate correction include: identifying at least one of the following anomalies: broken numbers, separation of decimal point and value, separation of percentage sign and value, loss of negative sign, separation of value and unit, and misalignment of values ​​in adjacent cells of a table; generating one or more candidate correction values ​​for the values ​​with anomalies, and retaining the original text, correction rules, corrected text, and identification confidence, without overwriting the original text with the candidate correction values.

7. The method for extracting process parameters and performing semantic verification for gold ore beneficiation literature according to claim 1, characterized in that, The structured output includes outputting process parameters as JSON format files, relational database records, tabular data, or knowledge graph nodes and relational data; each output process parameter also includes at least one of the following: original unit and normalized unit, verification status, version number, and number of retries.

8. The method for extracting process parameters and performing semantic verification for gold ore beneficiation literature according to claim 1, characterized in that, The method further includes: saving the original document, document evidence unit, process parameter candidates, field-level verification results, field-level error correction instructions, field values ​​before revision, field values ​​after revision, and number of retries; establishing a connection through at least one of the document number, evidence identifier, test group identifier, and process stage identifier, so that any output process parameter can be traced back to the original page number, page location, chapter, table, test group, and complete verification and retry process record.

9. A system for extracting process parameters and performing semantic verification for gold ore beneficiation literature, characterized in that, The method for extracting process parameters and performing semantic verification of gold ore beneficiation literature as described in any one of claims 1 to 8 includes: The document preprocessing module is used to perform text parsing, page layout parsing, and table extraction on gold ore beneficiation and metallurgy documents, identify anomalies and correct candidates in the data, and obtain a set of document evidence units with associated unique evidence identifiers and location information. The process parameter extraction agent module is used to extract candidate process parameters from literature evidence units based on a domain-specific process parameter model. The domain-specific process parameter model specifies that each process parameter is associated with at least the parameter name, parameter value, parameter unit, experimental group identifier, process stage identifier, result type, and evidence identifier. The result type at least distinguishes between stage results, overall results, control results, and improvement results. The semantic verification agent module, independent of the process parameter extraction agent module, is used to perform field-level verification on each field of each process parameter candidate. The field-level verification includes at least evidence existence verification, process stage attribution verification, numerical value and unit verification, test group association verification, and result level verification, and outputs the verification status of each field as passed, rejected, or questionable. The error correction feedback module is used to generate field-level error correction instructions for fields whose verification status is rejected or questionable, including at least the path of the field to be corrected, the current field value, the error type, and evidence location information. The retry control module is used to select the corresponding targeted re-extraction strategy according to the error type, control the process parameter extraction intelligent agent module to only perform targeted re-extraction on fields whose verification status is rejected or questionable, and keep the fields that have passed the verification unchanged, and mark the fields that still fail the verification as manual review status when the preset number of retries is reached. The structured output module is used to output the verified process parameters along with the corresponding evidence identifiers and location information in a structured manner.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program that includes the system of claim 9 and, when executed by the processor, implements the method of any one of claims 1 to 8.

Images