Nutrient salt reduction measure identification and hierarchical classification method based on large language model

Abstract

The application discloses a method for identifying and classifying nutrient salt reduction measures based on a large language model, and belongs to the technical field of environmental pollution control. The method comprises the following steps: S1, converting RIS literature data into a structured data set; S2, determining the relevance of the structured data, extracting the measure text, and identifying the corresponding emission source type; S3, constructing an emission source field dictionary and performing consistency verification on the emission source identification result by using the emission source field dictionary; S4, obtaining normalized measure text by normalizing the measure text, and filtering, cleaning and merging the normalized measure text by using a Python script; S5, mapping the measures to the corresponding classification level by using a large language model according to a hierarchical measure standard library; S6, introducing a rollback mechanism to optimize the processing of unmatched items and classification conflict items; and S7, performing dimension aggregation statistics on the structured classification result by using a Python script and generating a word frequency report, so that reliable data support can be provided.

Description

Technical Field

[0001] This invention belongs to the field of environmental pollution control technology, specifically involving a method for identifying and hierarchically classifying nutrient reduction measures based on a large language model. Background Technology

[0002] Because nutrient pollution in the Gulf is characterized by diverse emission sources, strong land-sea coupling, and complex pollution migration pathways, the reduction measures corresponding to different emission sources vary significantly in implementation paths, applicable targets, and mechanisms of action. Therefore, the completeness of reduction measure information and the accuracy of classification results directly affect the reliability and relevance of Gulf pollution control decision-making schemes.

[0003] Currently, the extraction and organization of nutrient reduction measures in the Gulf mainly relies on manual, in-depth reading, screening, and merging of massive amounts of literature. When faced with large-scale literature data, complex combinations of measures, and cross-domain governance scenarios, this manual processing model suffers from bottlenecks such as long processing times, heavy workloads, and susceptibility to subjective experience. Traditional methods struggle to avoid problems such as inconsistent classification standards, duplicate statistics of synonymous measures, and omissions of low-frequency, long-tail measures, making it difficult to efficiently obtain a structured database of measures. With the development of generative artificial intelligence, large language models, with their strong semantic understanding and information extraction capabilities, provide a new technical path for knowledge mining in the field of environmental governance.

[0004] However, nutrient remediation in the Gulf involves multiple highly specialized remediation units, and the technological approaches under different emission sources have strong domain characteristics and logical connections. When general-purpose large language models are directly applied in this vertical domain, they often face technical limitations such as machine illusion, inconsistent classification logic, and unstable output formats due to a lack of professional knowledge constraints. These limitations make it difficult to meet the practical application needs of accurate identification of measures, standardized expression, and hierarchical classification. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a method for identifying and hierarchically classifying nutrient reduction measures based on a large language model. This method can automatically identify nutrient reduction measures in the Gulf from massive amounts of literature data, and standardize and hierarchically classify the information on measures under different emission sources. This improves the efficiency, accuracy, and consistency of measure extraction and organization, facilitating the subsequent construction of a structured database of Gulf nutrient reduction measures. It can also provide reliable data support for precise pollution control and nutrient reduction decision-making in the Gulf.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for identifying and classifying nutrient reduction measures based on a large language model includes the following steps: S1. Use the rispy literature parsing toolkit to convert RIS literature data into structured datasets; S2. Using a large language model guided by prompt words, perform summary relevance determination, measure text extraction, and corresponding emission source type identification on structured data; S3. Construct an emission source domain dictionary and use the emission source domain dictionary constraints to perform consistency verification on the emission source identification results; S4. Using a large language model, the measures text is standardized to obtain standardized measures text, and then filtered, cleaned and merged using Python scripts; S5. Based on the hierarchical measure standard library, use a large language model to map the measures to the corresponding classification levels; S6. Introduce a rollback mechanism to optimize the handling of unmatched entries and entries with category conflicts; S7. Use a Python script to perform dimensional aggregation statistics on the structured classification results and generate a word frequency report.

[0007] Preferably, the specific process of step S1 is as follows: S11. Use the rispy literature parsing toolkit to read the literature storage path information in the configuration file, automatically identify the RIS literature to be processed, and use Python scripts to complete the literature parsing in batches. After parsing S12 and RIS documents, they are uniformly converted into structured datasets. Each document record contains at least the title, abstract, and year fields. For RIS documents that do not exist, are corrupted, or fail to be parsed, the system automatically skips them and outputs log information.

[0008] Preferably, in step S2, the prompt words used to guide the large language model include: Task Description: This task instructs a large language model to act as an expert in identifying Gulf nutrient reduction measures, and based on the structured dataset obtained in step S1, determines whether the literature involves Gulf nutrient reduction, and extracts the text of the measures and the corresponding emission source types. Relevance determination: This is used to instruct the large language model to determine whether the document abstract simultaneously contains information on nutrient emission sources, nutrient pollution issues, and mitigation measures; if it contains all of these, it is output as selected; otherwise, it is output as rejected. Measure text extraction: This is used to extract the specific reduction measures text from the selected abstracts while preserving the original expression form; when there are multiple measures in the abstract, they are split into multiple independent measure items according to semantic relationships; for combined measures, their combination relationship is preserved. Emission source identification: This is used to indicate that the large language model adopts an emission source identification method based on summary context constraints. It combines the study area, study object and pollution medium to dynamically determine the emission source type of the measure, and outputs an unclassified label when information is missing. Field definition and output format: This instructs the large language model to encapsulate the results into a list of structured objects containing fields for filtering results, measure text, and emission source type; Machine illusion suppression: This is used to instruct large language models not to output explanatory content, reasoning processes, or additional explanations that are irrelevant to the task; and not to fabricate measure information for documents for which measure texts have not been extracted.

[0009] Preferably, the specific process of step S3 is as follows: S31. Dictionary Construction: Construct corresponding sets of measure keywords for different emission source types. Each emission source type corresponds to several measure keywords, resulting in an emission source domain dictionary. The emission source domain dictionary is represented as follows: ,in, A dictionary for the field of emission sources; For the first Types of emission sources; To correspond with emission source type The corresponding set of keywords for measures; Number the emission source type; Total number of emission source types; Emission source type The first in the corresponding keyword set One keyword; Number the keywords; Emission source type The number of keywords in the corresponding keyword set; S32. Score Calculation: For each measure text output in step S2, match it with the keyword set corresponding to each emission source type, count the number of matched keywords, and use the number of keywords as the matching score between the measure text and the corresponding emission source type. The calculation formula is as follows: ,in, For the first Text of Measures With the Types of emission sources Match score between them; Number the measure text; For the first Text of the measures; For indicator functions, when the emission source type The first in the corresponding keyword set Key words Appeared in the Text of Measures The value is 1 if the condition is met, and 0 otherwise. The matching scores of all candidate emission source types are sorted, and the emission source type with the highest matching score is selected as the dictionary-suggested emission source type, denoted as . ; S33. Result Validation: Based on the relative difference between the highest matching score and the second highest matching score, calculate the confidence level of the dictionary suggestion result. The calculation formula is as follows: ,in, For the first Text of Measures The confidence level of the corresponding dictionary suggestion results; The highest matching score among all candidate emission source types; The second-highest matching score among all candidate emission source types; To obtain Or the maximum value in 1, used to avoid a denominator of 0; When the highest matching score is less than or equal to 0, the dictionary is deemed to have no valid evidence, and the original recognition result of the large language model output in step S2 is retained; when the dictionary suggests an emission source type that is inconsistent with the original recognition result of the large language model output in step S2, and the confidence level reaches the threshold for manual review. When the automatic correction mode is enabled and the confidence level reaches the automatic correction threshold, it is marked as an object for manual review; When the dictionary-suggested emission source type is used, the final emission source type is used; when automatic correction mode is not enabled, or the confidence level is below the automatic correction threshold. At that time, the original recognition result of the large language model output in step S2 is retained as the final emission source type; The final emission source type obtained after the consistency verification in step S3 corresponding to the measures is denoted as follows: .

[0010] Preferably, the specific process of step S4 is as follows: S41. Standardized Expression: Input the original measure text, document title and corresponding document abstract into the large language model to generate a standardized measure text under contextual constraints; the standardized measure text preferentially adopts the form of action verbs and operation objects to reduce repetition and ambiguity caused by differences in expression between different documents; S42. Result Validation: Use Python scripts to automatically validate the standardized measure text; for results with expanded, missing key objects, or generated generalized expressions, call the large language model for correction; if the requirements are still not met after correction, retain the original measure text as the final output; S43. Filtering, Cleaning and Merging: Use Python scripts to perform null value filtering, blacklist filtering, duplicate item identification and near-duplicate merging on the normalized measure text, remove category words, meaningless generalized expressions and pseudo measure text, and merge near-duplicate measures that only differ in spaces, punctuation or conjunctions. S44. Cache reuse: Establish a cache mapping relationship between the original measure text and the standardized measure text; when the same original measure text appears again in subsequent processing, the cached result is called first to improve processing efficiency and maintain the consistency of the standardized result.

[0011] Preferably, the specific process of step S5 is as follows: S51. Standard Library Construction: For different emission source types, establish corresponding primary measure candidate sets and secondary measure candidate sets. Then, based on the primary and secondary measure candidate sets, establish mapping relationships between hierarchical measure categories to form a hierarchical measure standard library; the hierarchical measure standard library is represented as follows: ,in, A hierarchical standard library for measures; For the first Types of emission sources The corresponding set of primary measures candidates; For the first Types of emission sources The corresponding set of candidate secondary measures; S52. Measure Mapping: Input the standardized measure text obtained in step S4 and the corresponding emission source type into the large language model, and perform restricted classification only within the secondary measure candidate set of the corresponding emission source type to obtain the secondary measure classification result. The calculation formula is as follows: ,in, For the first The classification results of the secondary measures corresponding to the measures; For classification functions of large language models based on cue word constraints; For the first The standardized text of the corresponding measures; Then, based on the preset hierarchical mapping relationship, the corresponding first-level measure category is determined, as expressed in the following expression: ,in, For the first The classification results of the primary measures corresponding to each measure; This is a hierarchical mapping function between primary and secondary measures; If the preset mapping relationship cannot complete the classification, the large language model is called to perform first-level measure classification; if it still cannot match, the output is unclassified.

[0012] Preferably, the specific process of step S6 is as follows: S61, Candidate Expansion and Rollback: For the measures that are unclassified in step S5, reconstruct the candidate category set and relax the original candidate constraint range; input the standardized measure text, corresponding emission source type and literature abstract back into the large language model, and require the model to select only the closest category from the expanded candidate category set; if the rollback result still does not belong to the expanded candidate category set, retain the unclassified result. S62. Hierarchical Consistency Verification: Perform hierarchical mapping verification on the measure classification results output in step S5 to determine whether the secondary measures belong to the corresponding emission source type and the candidate set of secondary measures under the primary measures. The consistency determination result is expressed as follows: ,in, For the first Consistency determination results of the classification results of the measures; For the first The corresponding primary measures; For the first Types of emission sources The following and the first The first-level measures corresponding to the measures The corresponding set of candidate secondary measures; S63. Optimization of conflict results: For measure items that still have classification conflicts after rollback, priority is given to retaining the classification results that meet the hierarchical consistency constraint; if multiple candidate results meet the hierarchical consistency constraint, priority is given to retaining the result with the highest degree of matching with the original candidate set; if none of the candidate results meet the hierarchical consistency constraint, the original classification result is retained and marked as unclassified or pending subsequent manual processing.

[0013] Preferably, the specific process of step S7 is as follows: S71. Field Alignment and Long Table Construction: The fields of emission source type, primary measure, secondary measure, standardized measure and original description are split and aligned in order, and expanded into long table details for each record; where the original description corresponds to the measure text extracted in step S2, and the standardized measure corresponds to the standardized measure text obtained in step S4. S72. Dimensional Aggregation Statistics: Aggregation statistics are performed on the long table details using emission source type, primary measures, secondary measures, and standardized measures as joint grouping keys; the frequency of standardized measures under the corresponding emission source type and level of measure category is expressed as: ,in, For standardization measures In terms of emission source type Level 1 measures and secondary measures Frequency of occurrence of the following; For standardization measures; This is a Level 1 measure; This is a level-two measure; Number the records in the long detail; This indicates the total number of records in the long display details; For the normalization measures corresponding to the r-th long table record; For the r-th long table record, specify the corresponding first-level measure; For the first The long table records the corresponding secondary measures; For the first The long table records the corresponding emission source types; It is a conjunction of two variables; For indicator functions; Through aggregation and statistical processes, the main nutrient reduction measures and their frequency of occurrence under different emission source types were obtained; S73. Output Results: Output the aggregated statistical results as an Excel word frequency report file; the Excel word frequency report includes at least categorized statistical results and detailed records, wherein the categorized statistical results include at least the emission source type, primary measure, secondary measure, standardized measure, frequency, and original description summary fields, and the detailed records include at least the emission source type, primary measure, secondary measure, standardized measure, and original description fields; when abnormal records occur during field splitting or alignment, the abnormal records are marked and output.

[0014] After adopting the above technical solution, the present invention has the following beneficial effects: 1. This invention constructs a fully automated process for processing Gulf nutrient reduction measures. From literature analysis, relevance screening, measure extraction, emission source identification to standardized expression, hierarchical classification and statistical report generation, the entire process requires no manual intervention, which greatly improves the processing efficiency of massive amounts of literature and reduces the workload of traditional manual work from several months to several hours.

[0015] 2. By constructing an emission source domain dictionary and a hierarchical measure standard library, and combining prompt word engineering and multi-round verification mechanism, this invention effectively suppresses the machine illusion problem of large language models, significantly improves the accuracy and consistency of emission source identification and measure classification, and solves the problems of inconsistent classification standards and large subjective bias in traditional manual processing.

[0016] 3. This invention introduces standardized expression and near-duplication merging mechanism, which unifies the expression of synonymous measures in different documents, avoids duplicate statistics, and further improves processing efficiency and result consistency through caching and reuse technology.

[0017] 4. This invention designs a candidate expansion rollback and hierarchical consistency verification mechanism to optimize the processing of unmatched and classification conflict entries, thereby minimizing the omission of measures and classification errors and ensuring the completeness of the measure database.

[0018] 5. The multi-dimensional aggregated statistical report generated by this invention can intuitively display the main nutrient reduction measures and their application frequency under different emission sources and levels, providing a comprehensive and reliable data foundation for the formulation of Gulf pollution control plans, the selection of measures, and decision support. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] like Figure 1 As shown, the method for identifying and classifying nutrient reduction measures based on a large language model includes the following steps: S1. Use the rispy literature parsing toolkit to convert RIS literature data into structured datasets; The specific process of step S1 is as follows: S11. Use the rispy literature parsing toolkit to read the literature storage path information in the configuration file, automatically identify the RIS literature to be processed, and use Python scripts to complete the literature parsing in batches. After parsing S12 and RIS documents, they are uniformly converted into structured datasets. Each document record contains at least the title, abstract, and year fields. For RIS documents that are missing, corrupted, or fail to be parsed, the system automatically skips them and outputs log information. S2. Using a large language model guided by prompt words, perform summary relevance determination, measure text extraction, and corresponding emission source type identification on structured data; In step S2, the prompt words used to guide the large language model include: Task Description: This task instructs a large language model to act as an expert in identifying Gulf nutrient reduction measures, and based on the structured dataset obtained in step S1, determines whether the literature involves Gulf nutrient reduction, and extracts the text of the measures and the corresponding emission source types. Relevance determination: This is used to instruct the large language model to determine whether the document abstract simultaneously contains information on nutrient emission sources, nutrient pollution issues, and mitigation measures; if it contains all of these, it is output as selected; otherwise, it is output as rejected. Measure text extraction: This is used to extract the specific reduction measures text from the selected abstracts while preserving the original expression form; when there are multiple measures in the abstract, they are split into multiple independent measure items according to semantic relationships; for combined measures, their combination relationship is preserved. Emission source identification: This is used to indicate that the large language model adopts an emission source identification method based on summary context constraints. It combines the study area, study object and pollution medium to dynamically determine the emission source type of the measure, and outputs an unclassified label when information is missing. Field definition and output format: This instructs the large language model to encapsulate the results into a list of structured objects containing fields for filtering results, measure text, and emission source type; Machine illusion suppression: This is used to instruct large language models not to output explanatory content, reasoning processes, or additional explanations that are irrelevant to the task; and not to fabricate measure information for documents where the measure text has not been extracted. S3. Construct an emission source domain dictionary and use the emission source domain dictionary constraints to perform consistency verification on the emission source identification results; The specific process of step S3 is as follows: S31. Dictionary Construction: Construct corresponding sets of measure keywords for different emission source types. Each emission source type corresponds to several measure keywords, resulting in an emission source domain dictionary. The emission source domain dictionary is represented as follows: ,in, A dictionary for the field of emission sources; For the first Types of emission sources; To correspond with emission source type The corresponding set of keywords for measures; Number the emission source type; Total number of emission source types; Emission source type The first in the corresponding keyword set One keyword; Number the keywords; Emission source type The number of keywords in the corresponding keyword set; S32. Score Calculation: For each measure text output in step S2, match it with the keyword set corresponding to each emission source type, count the number of matched keywords, and use the number of keywords as the matching score between the measure text and the corresponding emission source type. The calculation formula is as follows: ,in, For the first Text of Measures With the Types of emission sources Match score between them; Number the measure text; For the first Text of the measures; For indicator functions, when the emission source type The first in the corresponding keyword set Key words Appeared in the Text of Measures The value is 1 if the condition is met, and 0 otherwise. The matching scores of all candidate emission source types are sorted, and the emission source type with the highest matching score is selected as the dictionary-suggested emission source type, denoted as . ; S33. Result Validation: Based on the relative difference between the highest matching score and the second highest matching score, calculate the confidence level of the dictionary suggestion result. The calculation formula is as follows: ,in, For the first Text of Measures The confidence level of the corresponding dictionary suggestion results; The highest matching score among all candidate emission source types; The second-highest matching score among all candidate emission source types; To obtain Or the maximum value in 1, used to avoid a denominator of 0; When the highest matching score is less than or equal to 0, the dictionary is deemed to have no valid evidence, and the original recognition result of the large language model output in step S2 is retained; when the dictionary suggests an emission source type that is inconsistent with the original recognition result of the large language model output in step S2, and the confidence level reaches the threshold for manual review. When the automatic correction mode is enabled and the confidence level reaches the automatic correction threshold, it is marked as an object for manual review; When the dictionary-suggested emission source type is used, the final emission source type is used; when automatic correction mode is not enabled, or the confidence level is below the automatic correction threshold. At that time, the original recognition result of the large language model output in step S2 is retained as the final emission source type; The final emission source type obtained after the consistency verification in step S3 corresponding to the measures is denoted as follows: ; S4. Using a large language model, the measures text is standardized to obtain standardized measures text, and then filtered, cleaned and merged using Python scripts; The specific process of step S4 is as follows: S41. Standardized Expression: Input the original measure text, document title and corresponding document abstract into the large language model to generate a standardized measure text under contextual constraints; the standardized measure text preferentially adopts the form of action verbs and operation objects to reduce repetition and ambiguity caused by differences in expression between different documents; S42. Result Validation: Use Python scripts to automatically validate the standardized measure text; for results with expanded, missing key objects, or generated generalized expressions, call the large language model for correction; if the requirements are still not met after correction, retain the original measure text as the final output; S43. Filtering, Cleaning and Merging: Use Python scripts to perform null value filtering, blacklist filtering, duplicate item identification and near-duplicate merging on the normalized measure text, remove category words, meaningless generalized expressions and pseudo measure text, and merge near-duplicate measures that only differ in spaces, punctuation or conjunctions. S44. Cache reuse: Establish a cache mapping relationship between the original measure text and the standardized measure text; when the same original measure text appears again in subsequent processing, the cached result is called first to improve processing efficiency and maintain the consistency of the standardized result; S5. Based on the hierarchical measure standard library, use a large language model to map the measures to the corresponding classification levels; The specific process of step S5 is as follows: S51. Standard Library Construction: For different emission source types, establish corresponding primary measure candidate sets and secondary measure candidate sets. Then, based on the primary and secondary measure candidate sets, establish mapping relationships between hierarchical measure categories to form a hierarchical measure standard library; the hierarchical measure standard library is represented as follows: ,in, A hierarchical standard library for measures; For the first Types of emission sources The corresponding set of primary measures candidates; For the first Types of emission sources The corresponding set of candidate secondary measures; S52. Measure Mapping: Input the standardized measure text obtained in step S4 and the corresponding emission source type into the large language model, and perform restricted classification only within the secondary measure candidate set of the corresponding emission source type to obtain the secondary measure classification result. The calculation formula is as follows: ,in, For the first The classification results of the secondary measures corresponding to the measures; For classification functions of large language models based on cue word constraints; For the first The standardized text of the corresponding measures; Then, based on the preset hierarchical mapping relationship, the corresponding first-level measure category is determined, as expressed in the following expression: ,in, For the first The classification results of the primary measures corresponding to each measure; This is a hierarchical mapping function between primary and secondary measures; If the preset mapping relationship cannot complete the classification, the large language model is called to perform first-level measure classification; if it still cannot match, the output is unclassified. S6. Introduce a rollback mechanism to optimize the handling of unmatched entries and entries with category conflicts; The specific process of step S6 is as follows: S61, Candidate Expansion and Rollback: For the measures that are unclassified in step S5, reconstruct the candidate category set and relax the original candidate constraint range; input the standardized measure text, corresponding emission source type and literature abstract back into the large language model, and require the model to select only the closest category from the expanded candidate category set; if the rollback result still does not belong to the expanded candidate category set, retain the unclassified result. S62. Hierarchical Consistency Verification: Perform hierarchical mapping verification on the measure classification results output in step S5 to determine whether the secondary measures belong to the corresponding emission source type and the candidate set of secondary measures under the primary measures. The consistency determination result is expressed as follows: ,in, For the first Consistency determination results of the classification results of the measures; For the first The corresponding primary measures; For the first Types of emission sources The following and the first The first-level measures corresponding to the measures The corresponding set of candidate secondary measures; S63. Optimization of conflict results: For measure items that still have classification conflicts after rollback, priority is given to retaining the classification results that meet the hierarchical consistency constraint; if multiple candidate results meet the hierarchical consistency constraint, priority is given to retaining the result with the highest degree of matching with the original candidate set; if none of the candidate results meet the hierarchical consistency constraint, the original classification result is retained and marked as unclassified or to be processed manually later. S7. Use a Python script to perform dimensional aggregation statistics on the structured classification results and generate a word frequency report; The specific process of step S7 is as follows: S71. Field Alignment and Long Table Construction: The fields of emission source type, primary measure, secondary measure, standardized measure and original description are split and aligned in order, and expanded into long table details for each record; where the original description corresponds to the measure text extracted in step S2, and the standardized measure corresponds to the standardized measure text obtained in step S4. S72. Dimensional Aggregation Statistics: Aggregation statistics are performed on the long table details using emission source type, primary measures, secondary measures, and standardized measures as joint grouping keys; the frequency of standardized measures under the corresponding emission source type and level of measure category is expressed as: ,in, For standardization measures In terms of emission source type Level 1 measures and secondary measures Frequency of occurrence of the following; For standardization measures; This is a Level 1 measure; This is a level-two measure; Number the records in the long detail; This indicates the total number of records in the long display details; For the normalization measures corresponding to the r-th long table record; For the r-th long table record, specify the corresponding first-level measure; For the first The long table records the corresponding secondary measures; For the first The long table records the corresponding emission source types; It is a conjunction of two variables; For indicator functions; Through aggregation and statistical processes, the main nutrient reduction measures and their frequency of occurrence under different emission source types were obtained; S73. Output Results: Output the aggregated statistical results as an Excel word frequency report file; the Excel word frequency report includes at least categorized statistical results and detailed records, wherein the categorized statistical results include at least the emission source type, primary measure, secondary measure, standardized measure, frequency, and original description summary fields, and the detailed records include at least the emission source type, primary measure, secondary measure, standardized measure, and original description fields; when abnormal records occur during field splitting or alignment, the abnormal records are marked and output.

[0022] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying and hierarchically classifying nutrient reduction measures based on a large language model, characterized in that, Includes the following steps: S1. Use the rispy literature parsing toolkit to convert RIS literature data into structured datasets; S2. Using a large language model guided by prompt words, perform summary relevance determination, measure text extraction, and corresponding emission source type identification on structured data; S3. Construct an emission source domain dictionary and use the emission source domain dictionary constraints to perform consistency verification on the emission source identification results; S4. Using a large language model, the measures text is standardized to obtain standardized measures text, and then filtered, cleaned and merged using Python scripts; S5. Based on the hierarchical measure standard library, use a large language model to map the measures to the corresponding classification levels; S6. Introduce a rollback mechanism to optimize the handling of unmatched entries and entries with category conflicts; S7. Use a Python script to perform dimensional aggregation statistics on the structured classification results and generate a word frequency report.

2. The method for identifying and hierarchically classifying nutrient reduction measures based on a large language model as described in claim 1, characterized in that, The specific process of step S1 is as follows: S11. Use the rispy literature parsing toolkit to read the literature storage path information in the configuration file, automatically identify the RIS literature to be processed, and use Python scripts to complete the literature parsing in batches. After parsing S12 and RIS documents, they are uniformly converted into structured datasets. Each document record contains at least the title, abstract, and year fields. For RIS documents that do not exist, are corrupted, or fail to be parsed, the system automatically skips them and outputs log information.

3. The method for identifying and hierarchically classifying nutrient reduction measures based on a large language model as described in claim 1, characterized in that, In step S2, the prompt words used to guide the large language model include: Task Description: This task instructs a large language model to act as an expert in identifying Gulf nutrient reduction measures, and based on the structured dataset obtained in step S1, determines whether the literature involves Gulf nutrient reduction, and extracts the text of the measures and the corresponding emission source types. Relevance determination: This is used to instruct the large language model to determine whether the document abstract simultaneously contains information on nutrient emission sources, nutrient pollution issues, and mitigation measures; if it contains all of these, it is output as selected; otherwise, it is output as rejected. Measure text extraction: This is used to extract the specific reduction measures text from the selected abstracts while preserving the original expression form; when there are multiple measures in the abstract, they are split into multiple independent measure items according to semantic relationships; for combined measures, their combination relationship is preserved. Emission source identification: This is used to indicate that the large language model adopts an emission source identification method based on summary context constraints. It combines the study area, study object and pollution medium to dynamically determine the emission source type of the measure, and outputs an unclassified label when information is missing. Field definition and output format: This instructs the large language model to encapsulate the results into a list of structured objects containing fields for filtering results, measure text, and emission source type; Machine illusion suppression: This is used to instruct large language models not to output explanatory content, reasoning processes, or additional explanations that are irrelevant to the task; and not to fabricate measure information for documents for which measure texts have not been extracted.

4. The method for identifying and hierarchically classifying nutrient reduction measures based on a large language model as described in claim 1, characterized in that, The specific process of step S3 is as follows: S31. Dictionary Construction: Construct corresponding sets of measure keywords for different emission source types. Each emission source type corresponds to several measure keywords, resulting in an emission source domain dictionary. The emission source domain dictionary is represented as follows: ,in, A dictionary for the field of emission sources; For the first Types of emission sources; To correspond with emission source type The corresponding set of keywords for measures; Number the emission source type; This represents the total number of emission source types. Emission source type The first in the corresponding keyword set One keyword; Number the keywords; Emission source type The number of keywords in the corresponding keyword set; S32. Score Calculation: For each measure text output in step S2, match it with the keyword set corresponding to each emission source type, count the number of matched keywords, and use the number of keywords as the matching score between the measure text and the corresponding emission source type. The calculation formula is as follows: ,in, For the first Text of Measures With the Types of emission sources Match score between them; Number the measure text; For the first Text of the measures; For indicator functions, when the emission source type The first in the corresponding keyword set Key words Appeared in the Text of Measures The value is 1 if the condition is met, and 0 otherwise. The matching scores of all candidate emission source types are sorted, and the emission source type with the highest matching score is selected as the dictionary-suggested emission source type, denoted as . ; S33. Result Validation: Based on the relative difference between the highest matching score and the second highest matching score, calculate the confidence level of the dictionary suggestion result. The calculation formula is as follows: ,in, For the first Text of Measures The confidence level of the corresponding dictionary suggestion results; The highest matching score among all candidate emission source types; The second-highest matching score among all candidate emission source types; To obtain Or the maximum value in 1, used to avoid a denominator of 0; When the highest matching score is less than or equal to 0, the dictionary is deemed to have no valid evidence, and the original recognition result of the large language model output in step S2 is retained; when the dictionary suggests an emission source type that is inconsistent with the original recognition result of the large language model output in step S2, and the confidence level reaches the threshold for manual review. When the automatic correction mode is enabled and the confidence level reaches the automatic correction threshold, it is marked as a manually reviewed object; When the dictionary suggests the emission source type, it is used as the final emission source type; when automatic correction mode is not enabled, or the confidence level is below the automatic correction threshold. At that time, the original recognition result of the large language model output in step S2 is retained as the final emission source type; The final emission source type obtained after the consistency verification in step S3 corresponding to the measures is denoted as .

5. The method for identifying and hierarchically classifying nutrient reduction measures based on a large language model as described in claim 1, characterized in that, The specific process of step S4 is as follows: S41. Standardized Expression: Input the original measure text, document title and corresponding document abstract into the large language model to generate a standardized measure text under contextual constraints; the standardized measure text preferentially adopts the form of action verbs and operation objects to reduce repetition and ambiguity caused by differences in expression between different documents; S42. Result Validation: Use Python scripts to automatically validate the standardized measure text; for results with expanded, missing key objects, or generated generalized expressions, call the large language model for correction; if the requirements are still not met after correction, retain the original measure text as the final output; S43. Filtering, Cleaning and Merging: Use Python scripts to perform null value filtering, blacklist filtering, duplicate item identification and near-duplicate merging on the normalized measure text, remove category words, meaningless generalized expressions and pseudo measure text, and merge near-duplicate measures that only differ in spaces, punctuation or conjunctions. S44. Cache reuse: Establish a cache mapping relationship between the original measure text and the standardized measure text; when the same original measure text appears again in subsequent processing, the cached result is called first to improve processing efficiency and maintain the consistency of the standardized result.

6. The method for identifying and hierarchically classifying nutrient reduction measures based on a large language model as described in claim 4, characterized in that, The specific process of step S5 is as follows: S51. Standard Library Construction: Establish corresponding primary measure candidate sets and secondary measure candidate sets for different emission source types, and then establish a mapping relationship between hierarchical measure categories based on the primary measure candidate sets and secondary measure candidate sets to form a hierarchical measure standard library; The hierarchical standard library is represented as follows: ,in, A hierarchical standard library for measures; For the first Types of emission sources The corresponding set of primary measures candidates; For the first Types of emission sources The corresponding set of candidate secondary measures; S52. Measure Mapping: Input the standardized measure text obtained in step S4 and the corresponding emission source type into the large language model, and perform restricted classification only within the secondary measure candidate set of the corresponding emission source type to obtain the secondary measure classification result. The calculation formula is as follows: ,in, For the first The classification results of the secondary measures corresponding to the measures; For classification functions of large language models based on cue word constraints; For the first The standardized text of the corresponding measures; Then, based on the preset hierarchical mapping relationship, the corresponding first-level measure category is determined, as expressed in the following expression: ,in, For the first The classification results of the primary measures corresponding to each measure; This is a hierarchical mapping function between primary and secondary measures; If the preset mapping relationship cannot complete the classification, the large language model is called to perform first-level measure classification; if it still cannot match, the output is unclassified.

7. The method for identifying and hierarchically classifying nutrient reduction measures based on a large language model as described in claim 6, characterized in that, The specific process of step S6 is as follows: S61, Candidate Expansion and Rollback: For the measures that are unclassified in step S5, reconstruct the candidate category set and relax the original candidate constraint range; input the standardized measure text, corresponding emission source type and literature abstract back into the large language model, and require the model to select only the closest category from the expanded candidate category set; if the rollback result still does not belong to the expanded candidate category set, retain the unclassified result. S62. Hierarchical Consistency Verification: Perform hierarchical mapping verification on the measure classification results output in step S5 to determine whether the secondary measures belong to the corresponding emission source type and the candidate set of secondary measures under the primary measures. The consistency determination result is expressed as follows: ,in, For the first Consistency determination results of the classification results of the measures; For the first The corresponding primary measures; For the first Types of emission sources The following and the first The first-level measures corresponding to the measures The corresponding set of candidate secondary measures; S63. Optimization of conflict results: For measure items that still have classification conflicts after rollback, priority is given to retaining the classification results that meet the hierarchical consistency constraint; if multiple candidate results meet the hierarchical consistency constraint, priority is given to retaining the result with the highest degree of matching with the original candidate set; if none of the candidate results meet the hierarchical consistency constraint, the original classification result is retained and marked as unclassified or pending subsequent manual processing.

8. The method for identifying and hierarchically classifying nutrient reduction measures based on a large language model as described in claim 7, characterized in that, The specific process of step S7 is as follows: S71. Field Alignment and Long Table Construction: The fields of emission source type, primary measure, secondary measure, standardized measure and original description are split and aligned in order, and expanded into long table details for each record; where the original description corresponds to the measure text extracted in step S2, and the standardized measure corresponds to the standardized measure text obtained in step S4. S72. Dimensional Aggregation Statistics: Aggregation statistics are performed on the long table details using emission source type, primary measures, secondary measures, and standardized measures as joint grouping keys; the frequency of standardized measures under the corresponding emission source type and level of measure category is expressed as: ,in, For standardization measures In terms of emission source type Level 1 measures and secondary measures Frequency of occurrence of the following; For standardization measures; This is a Level 1 measure; This is a level-two measure; Number the records in the long detail; This indicates the total number of records in the long display details; For the normalization measures corresponding to the r-th long table record; For the r-th long table record, specify the corresponding first-level measure; For the first The long table records the corresponding secondary measures; For the first The long table records the corresponding emission source types; It is a conjunction of two variables; For indicator functions; Through aggregation and statistical processes, the main nutrient reduction measures and their frequency of occurrence under different emission source types were obtained; S73. Output Results: Output the aggregated statistical results as an Excel word frequency report file; the Excel word frequency report includes at least categorized statistical results and detailed records, wherein the categorized statistical results include at least the emission source type, primary measure, secondary measure, standardized measure, frequency, and original description summary fields, and the detailed records include at least the emission source type, primary measure, secondary measure, standardized measure, and original description fields; when abnormal records occur during field splitting or alignment, the abnormal records are marked and output.

Images