Pest control level evaluation method and system based on large language model

Abstract

The present application relates to the field of agricultural informatization and intelligent information processing technology, in particular to a pest control level evaluation method and system based on a large language model. The method comprises: obtaining crop pest control technical rule texts, inputting the texts into a first large language model after cutting by a sliding window to extract substantive semantic units; clustering the substantive semantic units into multiple topic clusters after encoding them into high-dimensional semantic vectors; inputting the topic clusters into a second large language model after hierarchical sampling, and outputting three-level indicators with semantic confidence; inputting the three-level indicators into a third large language model to obtain two-level indicators, constructing an initial evaluation system and calculating weighted three-level indicators; further constructing a complete pest control level evaluation system; obtaining pest control data of an area to be evaluated, and evaluating based on the system. The present application can accurately extract substantive technical rules from functionally heterogeneous texts, and improve the construction accuracy of the evaluation system and the authenticity of the evaluation results.

Description

Technical Field

[0001] This invention relates to the field of agricultural informatization and intelligent information processing technology, specifically to a method and system for evaluating the level of pest and disease control based on a large language model. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, large language models, with their advantages in semantic understanding and text generation, have been applied to the construction of evaluation systems for crop pest and disease control technology rules and the quantitative assessment of regional pest and disease control levels, providing a new technical path for this field.

[0004] However, in actual documents, technical rules for pest and disease control are often mixed with non-technical content of varying attributes, such as quarantine implementation records, fund usage reports, and emergency prevention and control notices, exhibiting strong functional heterogeneity. General-purpose large language models lack effective semantic filtering mechanisms when processing such mixed texts, making it difficult to accurately separate the control rules with substantial technical guidance. This results in the extraction results containing a large amount of non-technical noise. The inclusion of this non-technical noise directly affects the accurate construction of indicators and the reasonable allocation of weights in the subsequent evaluation system, further causing the evaluated level of pest and disease control to deviate from the actual implementation. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method and system for evaluating the level of pest and disease control based on a large language model.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a method for evaluating pest and disease control levels based on a large language model is provided, including: The text of the technical rules for the prevention and control of crop diseases and pests is obtained, segmented using a sliding window, and then input into the first language model to extract substantive semantic units. Substantive semantic units are input into a pre-trained text encoding model and encoded into high-dimensional semantic vectors; a clustering algorithm is then used to divide the high-dimensional semantic vectors into multiple topic clusters. After hierarchical sampling of each topic cluster, the data is input into the second large language model, and the output is a three-level index with semantic confidence. The tertiary indicators are input into the third language model to obtain the secondary indicators. Based on the correspondence between the secondary and tertiary indicators, an initial evaluation system is constructed. The semantic confidence of each tertiary indicator is calculated to obtain the weighted tertiary indicators. Based on weighted three-level indicators and an initial evaluation system, an evaluation system for the level of pest and disease control is constructed. Obtain data on pest and disease control in the area to be evaluated, and evaluate the level of pest and disease control in the area based on the pest and disease control level evaluation system.

[0007] Secondly, a pest and disease control level evaluation system based on a large language model is provided, including: The text acquisition and semantic extraction module is configured to: acquire the text of technical rules for the prevention and control of crop diseases and pests, segment it using a sliding window, input it into the first language model, and extract substantive semantic units; The semantic encoding and clustering module is configured to: input substantive semantic units into a pre-trained text encoding model and encode them into high-dimensional semantic vectors; and use a clustering algorithm to divide the high-dimensional semantic vectors into multiple topic clusters. The third-level indicator generation module is configured to: perform hierarchical sampling on each issue cluster and input it into the second large language model, and output a third-level indicator with semantic confidence. The initial evaluation system construction module is configured to: input the tertiary indicators into the third language model to obtain the secondary indicators, and construct the initial evaluation system based on the correspondence between the secondary and tertiary indicators; calculate the semantic confidence of each tertiary indicator to obtain the weighted tertiary indicators; The final evaluation system construction module is configured to: construct an evaluation system for pest and disease control level based on weighted three-level indicators and the initial evaluation system; The pest and disease control level assessment module is configured to: acquire pest and disease control data for the area to be assessed, and evaluate the level of pest and disease control work in the area to be assessed based on the pest and disease control level evaluation system.

[0008] Thirdly, an electronic device is also provided, comprising: Memory, used for non-transitory storage of computer-readable instructions; and Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in the first aspect above.

[0009] Fourthly, a computer-readable storage medium is provided having a program stored thereon that, when executed by a processor, implements the method described in the first aspect above.

[0010] Fifthly, a computer program product is provided, employing the following technical solution: A computer program product includes software code, wherein a program in the software code performs the steps of the method described in the first aspect of the present invention.

[0011] The above technical solution has the following advantages or beneficial effects: (1) In view of the problem that general large language models are difficult to accurately extract substantive technical rules and are prone to non-technical noise when processing functionally heterogeneous texts, this invention uses sliding window segmentation and negative constraint filtering, combined with unsupervised clustering and adaptive hierarchical sampling based on cluster density, to efficiently separate substantive semantic units with technical significance from mixed texts, significantly reduce the interference of non-technical text noise on the construction of indicators, and thus improve the accuracy and reliability of the evaluation system construction.

[0012] (2) This invention constructs a hierarchical evaluation system and introduces semantic confidence as a weight correction coefficient, so that the indicator weights reflect both the attention distribution in the original text and the semantic purity of the technical connotation, effectively suppressing the weight bias of ambiguous issues. At the same time, through theoretical saturation quantification test and structural Gini coefficient analysis, the completeness of the evaluation system's coverage of the original data and the balance of dimensional distribution are verified from a computational perspective, ensuring that the constructed indicator system has scientificity and structural self-consistency, thereby making the evaluation results of pest and disease control level more realistically reflect the actual implementation in the region.

[0013] (3) The evaluation system constructed in this invention has dynamic feedback and iterative optimization capabilities. By comparing the evaluation results with the weight benchmark, it can automatically identify the links that need to be strengthened in prevention and control and generate suggestions for strategy adjustment; by re-incorporating the newly added prevention and control data into the semantic mapping process, it can detect new technology signals not covered by the existing system and trigger index updates, providing data-driven quantitative support for the continuous improvement of pest and disease control strategies. Attached Figure Description

[0014] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0015] Figure 1 This is a flowchart of a method for evaluating pest and disease control levels based on a large language model, as described in a specific embodiment of the present invention. Figure 2 This is a line graph showing the profile coefficient and sum of squared errors when calculating issue clusters in a specific embodiment of the present invention. Detailed Implementation

[0016] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, 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.

[0017] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments of the invention. The terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0018] In this embodiment of the invention, "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Furthermore, in the description of this invention, "multiple" refers to two or more.

[0019] Furthermore, to facilitate a clear description of the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and the terms "first" and "second" are not necessarily different.

[0020] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0021] Example 1 like Figure 1 As shown, this embodiment provides a method for evaluating the level of pest and disease control based on a large language model, including: S101: Obtain the text of the technical rules for the prevention and control of crop diseases and pests, segment it using a sliding window, input it into the first language model, and extract the substantive semantic units. S102: Input the substantive semantic units into a pre-trained text encoding model and encode them into high-dimensional semantic vectors; use a clustering algorithm to divide the high-dimensional semantic vectors into multiple topic clusters; S103: After hierarchical sampling of each topic cluster, the data is input into the second language model, and the output is a three-level index with semantic confidence. S104: Input the tertiary indicators into the third language model to obtain the secondary indicators, and construct an initial evaluation system based on the correspondence between the secondary and tertiary indicators; calculate the semantic confidence of each tertiary indicator to obtain the weighted tertiary indicators. S105: Based on weighted three-level indicators and an initial evaluation system, construct an evaluation system for the level of pest and disease control; S106: Obtain data on pest and disease control work in the area to be evaluated, and evaluate the level of pest and disease control work in the area to be evaluated based on the pest and disease control level evaluation system.

[0022] In this embodiment, the first to third large language models all adopt the DeepSeek-V3 model. It should be understood that those skilled in the art may also choose other large language models.

[0023] Step S101 is as follows: S101.1: Obtain the text of the technical rules for the prevention and control of crop diseases and pests.

[0024] In some embodiments, crop pest and disease control technical rules can be obtained from websites that publish such rules in the region to be evaluated. For example, data can be collected from documents included on the official website of the agricultural and rural affairs department of the region to be evaluated. In this embodiment, comprehensive and systematic data collection is conducted from the Peking University Legal Database using "pest and disease control + region name" as the core search term.

[0025] In some embodiments, to ensure text reliability, the collected technical rules texts for crop pest and disease control need to be cleaned: First, duplicate files, metadata index pages, and texts with very short lengths (<200 words) with low information content are removed; second, text formatting and noise reduction processing is performed, completely removing HTML tags and garbled characters; finally, through manual review, routine documents such as fund usage reports and meeting notices that are not technically oriented are accurately screened out. In this embodiment, 266 technical rules texts for crop pest and disease control were finally obtained.

[0026] S101.2: To address the issue of contextual detachment that easily occurs when inputting long texts of crop pest and disease control technical rules into the first language model, this embodiment constructs a sliding window segmentation and semantic noise reduction mechanism in the preprocessing stage, aiming to improve the semantic coherence of the entire text.

[0027] First, a sliding window algorithm with overlap constraints is used to sequentially segment long texts of crop pest and disease control technical rules. By setting the window length and overlap step size, the semantic separation problem across paragraphs is prevented. Let the original set of crop pest and disease control technical rules text be... For any text in the text set Set the fixed window length to L and the overlap step size to O (where O <L)。

[0028] In this embodiment, m represents a single original text. The total number of slices generated after sliding window segmentation. Since different original texts have different lengths, the total number of blocks *m* that each text is segmented into is a variable under the same window length *L* and overlap stride *O*, and can usually be calculated using the total text length and the sliding stride (i.e., *LO*). Text set Segmented into slice sequences S i = { s i,1 ,s i,2 ,...,s i,m}. Define the first j slice s i,j The character range is [ start j end j ], start j This refers to the starting character index of the j-th slice in the original text; end j This refers to the index position of the terminating character of the j-th slice in the original text.

[0029] The overlapping regions of adjacent slices satisfy the following: (1); In the formula, adjacent slices s i,j and s i,j+1 There exists a semantic overlap region of length O. : (2); Then, negative constraint filtering is performed to negatively remove formatted domain noise that offers no informational increment. The serialized slices are then... s i,j Input the first major language model Guided function via prompt words f prompt ( · Output a structured set of substantive semantic units A: (3); in, It represents a single element in set A, namely a specific action rule or semantic unit extracted from the set. Constraint pos Represents a positive constraint condition; Constraint neg This is a pre-defined list of negative constraints.

[0030] In this embodiment, the content including "pest and disease targets, triggering conditions, prevention and control methods, and core parameters" is a substantive semantic unit.

[0031] In this embodiment, to prevent information loss due to excessive cleaning, a manual sampling process is further implemented to check the accuracy and recall rate. Substantive semantic units are filtered and verified, and verified substantive semantic units are output. For example, a stratified random sampling method is used to sample 5% of the generated substantive semantic units, which are then double-blindly verified by two plant protection experts.

[0032] In this embodiment, the sliding window length L is set to 500 words, and the overlap threshold between adjacent windows O is set to 100 words. Then, by injecting specific extraction prompts, the first large language model is driven to accurately capture and reconstruct substantial semantic units containing content such as "pest and disease targets, triggering conditions, control methods, and core parameters" from the sliding window text. After automated cleaning and segmentation by the first large language model, the system ultimately extracts 4,039 discrete substantial semantic units from 266 technical rule texts on crop pest and disease control.

[0033] Step S102 is as follows: S102.1: Input the substantive semantic units into the pre-trained text encoding model and encode them into high-dimensional semantic vectors.

[0034] After extracting the substantive semantic units, this embodiment constructs a high-dimensional semantic space to accurately capture the complex plant protection terminology and its deep semantic relationships within the technical rule text. All substantive semantic units are input into the BAIBGE-M3 pre-trained encoding model. The pre-trained encoding model performs contextual understanding on each unstructured short text and accurately transforms it into a dense vector, outputting a high-dimensional feature matrix. This high-dimensional feature matrix transforms massive amounts of discrete text into a continuous mathematical coordinate space capable of precise distance measurement.

[0035] A dense retrieval vector space is constructed, and a vector space model is used to transform discrete semantic units into computable mathematical objects. Unsupervised learning is then used to discover potential issues. A pre-trained encoding model is introduced. Enc ( · ), to semantic units a k ∈A Mapped to 3D dense vector v k : (4); Thus, a high-dimensional semantic matrix is ​​constructed. This mechanism can effectively capture the deep contextual features of text.

[0036] In this embodiment, the pre-trained encoding model performs contextual understanding on each substantial semantic unit and precisely transforms it into a dense vector containing 1024 floating-point numbers. Ultimately, it outputs a high-dimensional feature matrix of size 4039×1024.

[0037] S102.2: A clustering algorithm is used to divide the high-dimensional semantic vector into multiple issue clusters.

[0038] Perform clustering segmentation based on objective function optimization. To overcome the subjectivity of traditional manual classification, an unsupervised clustering algorithm is used to segment the high-dimensional semantic matrix. The MiniBatchK-Means algorithm is then executed on the high-dimensional semantic matrix X. The objective function is defined. J To minimize the sum of squared errors within the cluster:

[0039] in, For clusters C j The center of mass, The high-dimensional embedding vector representing the i-th substantial semantic unit: To determine the optimal number of clusters K opt The contour coefficient SC is introduced for iterative optimization. a ( i ) represents a sample i The average distance to other samples in the same cluster, i.e., the cluster cohesion; b ( i ) represents a sample i The average distance to the nearest neighboring cluster sample is the inter-cluster separation. The algorithm iteratively searches for a K value that maximizes the silhouette coefficient, then stops the search, thus objectively dividing the massive semantic atoms into... K opt A highly cohesive semantic issue cluster is used to complete unsupervised detection of potential pattern structures.

[0040]

[0041] Where A is the aforementioned set of substantive semantic units, and |A| is the total number of samples in this set. In this embodiment, since the number of real issues in crop pest and disease control technology rules is unknown, the study calculates the silhouette coefficient SC and the sum of squared errors SSE within a broad interval of K∈[10,80]. Figure 2As shown, with the increase of K value, the SSE curve, representing intra-cluster cohesion, shows a significant downward trend, while the SC curve, representing inter-cluster separation, exhibits a clear global peak at K=50, where SC=0.0483, revealing that the relative separation of issue clusters reaches its optimum at this dimension. Therefore, this embodiment locks the initial number of issue clusters at [value missing]. K opt =50.

[0042] Step S103 is as follows: S103.1: Perform stratified sampling for each topic cluster.

[0043] After unsupervised clustering, the discrete text of crop pest and disease control technology rules is divided into multiple semantically cohesive issue clusters. However, in actual vector space partitioning, the sample size of each cluster is often unbalanced: there are giant clusters containing massive amounts of text, as well as small micro-clusters. When subsequently calling the second language model for concept extraction, if indiscriminate global direct truncation or simple random sampling is used, the attention mechanism built into the second language model is easily captured by high-frequency mainstream features, resulting in the systematic folding of weak signals hidden in small clusters. To address this, this embodiment constructs an adaptive hierarchical sampling mechanism based on cluster density, matching differentiated sampling logic to issue clusters of different sizes to ensure the comprehensiveness and unbiasedness of theoretical concept generation.

[0044] Adaptive hierarchical sampling based on cluster density is performed. To address the long-tail distribution problem of semantic clusters, a statistical distribution parameter μ (mean) is introduced. σ (Standard deviation) is used to dynamically set the scale threshold, where Used to identify microclusters For identifying giant clusters. Define an adaptive sampling function. Sample ( C j ), which is based on cluster size N j = | C j |Dynamic matching of the three-zone structured sampling logic: (7); in, Top(C j , k top ) Indicates from cluster C j Extract the frontmost element closest to the semantic centroid. k top One core sample; Rand(C j , krand ) Indicates from cluster C j Randomly selected from the remaining samples k rand Individual samples to ensure semantic diversity; H(C j , k head ) , M(C j , k mid ) and T(C j , k tail ) These respectively represent giant clusters C j The sample was divided into three spatial zones—proximal, mid-range, and distal—based on its distance from the centroid, and samples were extracted independently from each zone. k head , k mid , k tail One sample.

[0045] This embodiment uses adaptive hierarchical sampling to significantly prune high-frequency redundant non-technical text noise at the physical level, while preserving the complete semantic distribution of the cluster in the entire high-dimensional space.

[0046] S103.2: Input the stratified samples into the second language model. Output a three-level index with semantic confidence.

[0047] After adaptive stratified sampling, unbiased and representative samples were obtained. S sample Next, the second major language model is used for concept naming. Chain-like thinking prompts guide the model to generate indicator names following a rigorous logical path. Name j : (8); Introducing semantic confidence C iAs a quantitative tool, macroscopic descriptive noise with logical deviations is eliminated. Combined with four-dimensional human-machine verification, a three-level index list with high reliability and interpretability is output. Specifically, it includes: (1) internal consistency test, which automatically eliminates low-scoring noise clusters with insufficient cohesion through model confidence scoring; (2) domain specificity test, which identifies and corrects generalized names that are out of context; (3) factual verification, which compares the generated index with the underlying original text evidence and performs semantic downgrading on the index that has been artificially inflated by the algorithm; (4) mutual exclusivity test, which calculates the semantic similarity between the indexes and performs a merging operation on overlapping indexes.

[0048] In this embodiment, 50 clustered corpora and their semantic centroid features are input into the second large language model, driving it to perform inference through "semantic clustering—functional identification—term conversion". During the prompting process, Temperature is strictly set to 0.1 to minimize the generation illusion of the large model and ensure the fidelity of the output to the original technical corpus. The second large language model, DeepSeek-V3, exhibits strong generalization ability at this stage, but also shows some generalization bias. For example, some clusters, due to their broad names (such as integrated agricultural management technology), appear too general in the plant protection technology system; some small-scale clusters (such as pest and disease monitoring and early warning forecasts with only 2 samples) are identified as weak forward-looking signals and retained.

[0049] To address the issues with the initially generated level 3 indicators, this embodiment employs a four-dimensional human-machine collaborative verification mechanism. As shown in Table 1, experts deeply analyzed the reading data provided by the second major language model, DeepSeek-V3, manually eliminating non-substantive statements and repairing semantic generalization indicators. Simultaneously, through mutual exclusion testing and expert intervention, logical overlaps between specialized unified prevention and control and integrated management of rice pests and diseases were eliminated. Ultimately, 50 level 3 indicators with semantic confidence scores all above 0.5 were successfully retained, ensuring the empirical validity and technical significance of the system.

[0050] Table 1. Example of the Evolution List of Technical Rule Indicators

[0051] Step S104 is as follows: S104.1: Input the tertiary indicators into the third language model to obtain the secondary indicators. Based on the correspondence between the secondary and tertiary indicators, construct the initial evaluation system.

[0052] Following step S103, the technical rules for crop disease and pest control have been transformed into a discrete set of third-level indicators with semantic confidence. However, these discrete third-level indicators remain fragmented concepts at a low level of abstraction and cannot automatically bridge the gap between empirical phenomena and deep theoretical logic. Therefore, this step aims to construct a macroscopic theoretical system from micro-indicators.

[0053] To avoid potential knowledge contamination from the pre-trained parameters of the second-largest language model, this step positions it as a semantic bridging tool rather than a decision-making entity. When generating the initial dimensions, the abductive reasoning ability of the second-largest language model is utilized to analyze the attribute relationships between indicators, and an evidence constraint mechanism is introduced: the second-largest language model is required to simultaneously output a summary of the three-level indicator features supporting that dimension and an index of the original corpus when proposing the macro-dimensional framework. This constraint guides the model to move beyond prior knowledge and instead provide a structured interpretation of the evolution of micro-indicators, thereby inductively generating an initial dimensional framework supported by data.

[0054] The third language model was used as a full-scale inverse mapping classifier, driving it to perform secondary induction on 50 indicators based on the principle of back-reference reasoning. To prevent interference from the prior knowledge of the large language model, evidence verification was performed: the large language model was required to backtrack and label its supporting tertiary indicators and corresponding original text fragments when summarizing secondary indicators. This induction process followed a three-step logic: first, the substantive technical intervention targets of the indicators were identified through encoding; second, the technical system theory was introduced for logical classification; and finally, experts intervened to refer to relevant pest and disease control technical standards to conduct terminology standardization and mutual exclusivity tests, performing splitting and merging operations on the initial dimensions to establish a final theoretical framework containing M core dimensions. This M-dimensional framework not only has deep underlying textual data support but also incorporates the prior knowledge of experts. Finally, an initial evaluation system for crop pest and disease control technical rules was established, consisting of six secondary indicators, including professional unified prevention and control and integrated management, precise targeted intervention with chemical agents, and standardized management system, as shown in Table 2.

[0055] Table 2. Framework of Initial Evaluation System for Crop Disease and Pest Control Technical Rules

[0056] In some embodiments, to verify the coverage of the original data by the newly generated M core dimensions, the cosine similarity between the high-dimensional semantic vector and the semantic centroid of each secondary indicator is calculated, and the semantic centroid vector of each dimension is calculated in the vector space. μ k Then, a classifier is constructed. Category ( x ), calculate each high-dimensional semantic vector Enc ( xThe cosine similarity with each centroid is used to assign it to the dimension of the best match: (9); Perform a quantitative test of theoretical saturation to verify the completeness of knowledge extraction. Transform the traditional concept of theoretical saturation into a precisely calculable quantitative formula and define the verification coverage. S ,in, N total This represents the total number of substantive semantic units obtained during the preprocessing stage. N mapped To successfully classify in the reverse mapping stage M The number of effective corpora in each core dimension.

[0057] (10); In addition to calculating the classification coverage S of the validation corpus, a marginal saturation curve test is also introduced: when a new corpus is randomly injected into the model, if no new clusters are generated and the semantic centroid shift of the core dimension is already present, the test is performed. Δμ< 0.05 If so, it is determined to be theoretically saturated.

[0058] Finally, a structural equilibrium test is performed. The structural Gini coefficient G is introduced to verify the structural equilibrium in the theoretical dimension. n i and n j Representing the first i The and the first j The corpus distribution included in the mapping of each dimension, The average number of corpora for each dimension. By measuring the absolute difference in corpus distribution across dimensions, the Gini coefficient G can effectively quantify the robustness of the theoretical framework. If the G value is within a reasonable range, it indicates that the generated indicator system possesses scientific validity and structural consistency.

[0059] (11); Specifically, in this embodiment, 4,039 substantive semantic units were reverse-aligned to the indicator framework, and semantic mapping was performed using cosine similarity, resulting in a verification coverage rate of S=83.62%. The remaining 16.38% of the corpus was confirmed as non-substantive descriptions through semantic backtracking using the third language model, and did not possess the characteristics for constructing new indicators. In subsequent tests with additional corpus, no information sufficient to disrupt the existing six-dimensional structure was identified, and the semantic centroids of each dimension tended to stabilize, confirming from a computational perspective that the system had reached theoretical saturation. Structural Gini coefficient analysis of the indicator quantity distribution across the six dimensions yielded G=0.2400. According to the evaluation criteria, this value is within a relatively balanced range of 0.2-0.3, reflecting the good robustness and synergy of the evaluation system structure.

[0060] S104.2: Calculate the semantic confidence of each tertiary indicator to obtain the weighted tertiary indicators.

[0061] In this embodiment, when calculating the indicator weights, not only coverage is considered, but semantic confidence is also used as a metric. Semantic confidence is measured by the cosine similarity between the generated indicator and the centroid of the underlying corpus, identifying and filtering common "macro-descriptive noise" (i.e., non-technical specific measures) in the text. The final weight of the indicator is determined by the original attention percentage. P i With semantic confidence C i The joint decision effectively suppressed the weight bias of issues that were semantically ambiguous and lacked substantial specific targets, thereby truly restoring the distribution of the effectiveness of the enforcement of technical rules.

[0062] First, statistics on each of the three levels of indicators. i Number of associated substantive semantic units S i Let the total sample size after preprocessing be... N Then the original attention percentage of this indicator P i Defined as: (12); semantic confidence C i As a quality correction factor, the preliminary weight score of this indicator is calculated. E i : (13); Full normalization was performed to ensure the systematicity and comparability of each indicator within the evaluation system. The final weights of the three-level indicators were determined. W i The calculation formula is: (14); In this embodiment, after constructing an evaluation system containing 50 tertiary indicators, the importance of each indicator in the overall technical rule space is assigned weights.

[0063] Step S105 is as follows: Based on the correspondence between the tertiary and secondary indicators and the weight of each tertiary indicator, an evaluation system for the level of pest and disease control is constructed, as shown in Table 3.

[0064] Table 3 Evaluation System for Pest and Disease Control Level

[0065] As shown in Table 4, to further demonstrate the practical effect of the automatic rule extraction technology of the evaluation system constructed in this embodiment, taking the professional unified prevention and control and comprehensive management among the secondary indicators as an example, the typical rule texts corresponding to the three tertiary indicators with the highest weights are listed. All rule clauses are directly extracted, and the source is marked to the original document title.

[0066] Table 4 Professionalized Unified Prevention and Control and Comprehensive Management

[0067] Step S106 is as follows: Data on pest and disease control efforts in the region to be evaluated during a specific period will be collected. Data sources may include records of control measures implemented, pesticide usage logs, coverage of integrated pest management services, quarantine implementation records, and information on the use of special funds. Subsequently, based on the constructed three-tiered indicators, the collected control data will be quantitatively scored item by item to obtain the actual implementation score for each technical dimension. i .

[0068] Among them, the actual performance score of each tertiary indicator for different types of prevention and control work data collected is as follows: i The scoring is calculated according to the following rules, with the scoring range uniformly normalized to [0, 100]. For numerical data (including but not limited to quantifiable statistical data such as prevention area, coverage rate, and fund allocation), the actual implementation value V is used. i The benchmark reference value T of the corresponding indicator i The ratio is converted using the following formula: (15); Among them, the benchmark reference value T i The target value for pest and disease control can be the preset value of the work objective, the historical statistical average, or the actual scale of pests and diseases involved in the corresponding indicator; the conversion factor λ is used to map the actual implementation ratio to the scoring range of [0, 100], and its value can be dynamically adjusted according to industry benchmark data or expert preset rules; when there are already official quantitative results such as standardized treatment rate, coverage rate, or control compliance rate publicly released by the competent department for the corresponding indicator, the value is directly used to map to the Score. i ; For Boolean or event-based data (including but not limited to records of execution status such as emergency plan formulation and technical training), a discrete tiered scoring system is applied based on the completeness of execution: if there are no execution records, 0 points are assigned; if there are execution records but key steps are missing, a preset intermediate score is assigned; if the execution log is complete and compliant, 100 points are assigned. For unstructured text data (including but not limited to natural language records such as work summaries), the features of the collected execution text are extracted, and semantic similarity is calculated between them and the semantic constraints of the third-level index standard established in the system construction stage of the third language model. The calculated semantic matching degree is then mapped proportionally to a continuous score in the range [0, 100]. When the corresponding indicator data is missing, it will be marked as missing data and will not be included in the calculation of the comprehensive score for that evaluation, so as to ensure the comparability of the evaluation results.

[0069] Based on this, and combining the weights W of each tertiary indicator i The overall score for the prevention and control work in this region was calculated using a weighted summation method: (16); Due to the index weight W i The comprehensive score, calculated using semantic confidence weighting, objectively reflects the relative importance of each technical rule within the current evaluation system. Therefore, compared to the traditional equal-weighted scoring method, this comprehensive score more accurately reflects the actual differences in the actual contribution of different prevention and control technologies to the overall control effect. Furthermore, by comparing the weighted scores of the evaluated regions on each secondary indicator, the strengths and areas needing strengthening in the region's prevention and control efforts can be further identified, providing a quantitative basis for the precise allocation of subsequent prevention and control resources.

[0070] In this embodiment, all 50 tertiary indicators can be used in scoring and weighted calculations in practical applications. Due to the limited availability of publicly available data, this embodiment selects six representative tertiary indicators for demonstrative verification to illustrate the computability of the scoring formula and weighted calculation method. The scoring methods for the remaining tertiary indicators are consistent and can be supplemented with prevention and control data obtained from specific application scenarios.

[0071] To verify the computability of the above assessment method, this embodiment uses the annual implementation data of crop diseases and pests publicly released by the Shandong Provincial Agricultural Technology Extension Center at a media briefing on December 5, 2022, for empirical analysis. This data comes from a report published in the *Dazhong Daily* on December 8, 2022, and represents annual implementation results data released by a provincial-level agricultural technology extension agency.

[0072] According to the report, in 2022, Shandong Province experienced approximately 519 million mu (26.7 million hectares) of crop damage from various pests, diseases, weeds, and rodents, with a total control area of ​​approximately 601 million mu (40.7 million hectares). By crop type, corn affected 132 million mu (8.8 million hectares) with 139 million mu (8.8 million hectares) treated; soybeans affected 6.1155 million mu (412,000 hectares) with 6.4738 million mu (424,000 hectares) treated; cotton affected 8.1082 million mu (424,000 hectares) with 8.2378 million mu (424,000 hectares) treated (saving 36,100 tons of cotton); the Oriental migratory locust affected 3.2941 million mu (244,000 hectares) with 2.0775 million mu (133,000 hectares) meeting standards and 2.2969 million mu (144,000 hectares) treated; and the fall armyworm control rate was reported as 100%.

[0073] This embodiment introduces the prevention and control coverage rate as a quantitative intermediate variable for the intensity of implementation. The calculation method is: Prevention and control coverage rate = Prevention and control area ÷ Affected area. It is then converted into an index score of 0 to 100 using the conversion formula: Score = MIN(100, Prevention and control coverage rate × 70). The statistical scope for wheat stripe rust prevention and control area is the cumulative area of ​​preventative pesticide application across the province (prevention and control coverage rate reaching 7383.05%), which differs from the statistical scope for other crops. Therefore, this embodiment does not include it in the comprehensive score calculation, but only discloses it as a single data point to ensure the comparability of the scoring results.

[0074] Based on mapping calculations, the overall data for the province corresponds to the green control index for major crop diseases and pests, with a converted score of 81.1 points; the corn data corresponds to the chemical control technology index for major rice diseases and pests (since Shandong Province mainly grows wheat and corn, this embodiment uses corn data as a substitute for verification), with a converted score of 73.7 points; the soybean data corresponds to the chemical spray control technology index, with a converted score of 74.1 points; the cotton data corresponds to the integrated pest management technology index for cotton pests, with a converted score of 71.1 points; the East Asian migratory locust data corresponds to the monitoring, early warning, and emergency response system index for major diseases and pests, with a converted score of 48.8 points; and the fall armyworm data corresponds to the monitoring, early warning, and emergency control system index for diseases and pests, with a score of 100 points directly adopted since the treatment rate has been directly published. The scores of each of the above indicators are then combined with their final weights W established in the evaluation system of this embodiment. i After normalizing and weighting the sum, the overall prevention effectiveness index of this embodiment is 75.13 points.

[0075] After completing the evaluation of prevention and control effectiveness, the indicator system constructed in this embodiment further supports dynamic feedback and iterative adjustment of specific prevention and control methods. Specifically, the scores of each dimension obtained in the evaluation phase are compared with the weight benchmarks established in the construction phase of the indicator system. For dimensions or specific indicators whose scores deviate significantly from the expected weights, they can be identified as corresponding areas requiring reinforcement in the current prevention and control work in the region. Based on this, managers evaluate the prevention and control methods adopted in the early stage and provide feedback to adjust subsequent prevention and control methods and strategies. Taking this embodiment as an example, the converted score (48.8 points) of the corresponding indicator for the Oriental migratory locust is significantly lower than that of other crops and pest indicators, reflecting that the implementation intensity of major pest and disease monitoring, early warning, and emergency response in this region is relatively weak. Based on this evaluation result, feedback is provided to adjust the prevention and control methods for this type of pest and disease, such as strengthening the pre-stocking of emergency prevention and control materials and upgrading routine prevention and control to improve the rapid disposal timeliness of qualified areas. It should be understood that the specific judgment criteria and adjustment strategies can be differentiated according to the actual prevention and control scenarios of different regions and different crops.

[0076] In addition to the implementation intensity feedback based on a single assessment result, the indicator system constructed in this invention also supports a deeper iterative evaluation of the prevention and control methods themselves. Specifically, the system can compare the trend of the scoring results of the same tertiary indicator over multiple consecutive assessment periods. If the score of the indicator is consistently lower than the expected level corresponding to its weight, and the score does not improve significantly after adjustments to the implementation intensity (such as increasing investment or increasing the frequency of treatment), it can be determined that the specific prevention and control method currently used for the indicator may have limitations, rather than simply being due to insufficient implementation. In this case, the system can combine the secondary indicators to which the tertiary indicator belongs to further compare the historical scoring performance of other tertiary indicators belonging to the same dimension but using different technical paths in the indicator system of this invention, identify whether there are better-scoring prevention and control technical paths that can be used as alternatives or supplements, and generate iterative suggestions for prevention and control methods accordingly. For example, when a certain type of pest or disease has long relied on a single chemical agent for control but the score remains low, the system can combine the historical performance of relevant third-level indicators (such as ecological regulation, natural enemy protection and utilization, and integrated pest management services) under the physical and ecological control system or the professional integrated pest management and comprehensive governance dimension in the evaluation index system constructed in this embodiment. It can then suggest that the control method for this type of pest or disease be gradually adjusted from single chemical control to an integrated control path that combines chemical control with biological control and physical control. This provides a basis for optimizing and adjusting the control strategy from a methodological level rather than just an implementation intensity level.

[0077] Furthermore, as the region under evaluation continues to generate new prevention and control data in subsequent cycles, this new data can be reintegrated into the semantic extraction and reverse mapping process described in this embodiment. By calculating the cosine similarity between this data and the semantic centroid of the existing indicator system, it can be determined whether the prevention and control practices in this region have generated new technological signals not yet covered by the existing indicator system. If such new signals are identified, they can be fed back to the iterative update stage of the indicator system to supplement or correct the corresponding tertiary indicators. Through the above mechanism, the evaluation indicator system constructed in this embodiment has the ability to continuously optimize as prevention and control practices evolve.

[0078] Example 2 This embodiment provides a pest and disease control level evaluation system based on a large language model, including: The text acquisition and semantic extraction module is configured to: acquire the text of technical rules for the prevention and control of crop diseases and pests, segment it using a sliding window, input it into the first language model, and extract substantive semantic units; The semantic encoding and clustering module is configured to: input the substantive semantic units into a pre-trained text encoding model and encode them into high-dimensional semantic vectors; and use a clustering algorithm to divide the high-dimensional semantic vectors into multiple topic clusters. The third-level indicator generation module is configured to: perform hierarchical sampling on each issue cluster and input it into the second large language model, and output a third-level indicator with semantic confidence. The initial evaluation system construction module is configured to: input the tertiary indicators into the third language model to obtain secondary indicators, and construct the initial evaluation system based on the correspondence between the secondary and tertiary indicators; calculate the semantic confidence of each tertiary indicator to obtain weighted tertiary indicators; The final evaluation system construction module is configured to: construct an evaluation system for pest and disease control level based on weighted three-level indicators and the initial evaluation system; The pest and disease control level assessment module is configured to: acquire pest and disease control data for the area to be assessed, and evaluate the level of pest and disease control work in the area to be assessed based on the pest and disease control level evaluation system.

[0079] It should be noted that the text acquisition and semantic extraction module, semantic encoding and clustering module, three-level indicator generation module, initial evaluation system construction module, final evaluation system construction module, and prevention and control level assessment module mentioned above correspond to steps S101 to S106 in Embodiment 1. The examples and application scenarios implemented by the above modules and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer-executable instructions.

[0080] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0081] The proposed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and the division of modules described above is only a logical functional division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0082] Example 3 This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein, the processor is connected to the memory, and the one or more computer programs are stored in the memory. When the electronic device is running, the processor executes the one or more computer programs stored in the memory to cause the electronic device to perform the method described in Embodiment 1.

[0083] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0084] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0085] In the implementation process, each step of the above method can be completed by the integrated logic circuits in the processor hardware or by software instructions.

[0086] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0087] Those skilled in the art will recognize that the units and algorithm steps described in connection with the various examples of this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention.

[0088] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium.

[0089] A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of the method as described in Embodiment 1 of the present invention.

[0090] The detailed steps are the same as those provided in Example 1, and will not be repeated here.

[0091] Example 5 Embodiment 5 of the present invention provides a computer program product.

[0092] A computer program product includes software code, wherein the program in the software code performs the steps described in Embodiment 1 of the present invention.

[0093] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for evaluating the level of pest and disease control based on a large language model, characterized in that, include: The text of the technical rules for the prevention and control of crop diseases and pests is obtained, segmented using a sliding window, and then input into the first language model to extract substantive semantic units. The substantive semantic units are input into a pre-trained text encoding model and encoded into high-dimensional semantic vectors; a clustering algorithm is then used to divide the high-dimensional semantic vectors into multiple topic clusters. After hierarchical sampling of each topic cluster, the data is input into the second large language model, and the output is a three-level index with semantic confidence. The tertiary indicators are input into the third language model to obtain secondary indicators, and an initial evaluation system is constructed based on the correspondence between the secondary and tertiary indicators. The proportion of the substantive semantic units associated with each third-level indicator to the total sample size is calculated to obtain the original attention proportion. The initial weight scores of each tertiary indicator are obtained based on the original attention ratio and semantic confidence. The initial weight scores of all tertiary indicators are normalized so that the weight of each tertiary indicator is equal to the proportion of its initial weight score to the sum of the initial weight scores of all tertiary indicators, thus obtaining the weight of each tertiary indicator. Based on weighted three-level indicators and an initial evaluation system, an evaluation system for the level of pest and disease control is constructed. Data on pest and disease control in the area to be evaluated are obtained, and the level of pest and disease control in the area to be evaluated is evaluated based on the pest and disease control level evaluation system.

2. The method for evaluating pest and disease control levels based on a large language model as described in claim 1, characterized in that, The process of segmenting using a sliding window and inputting the data into the first language model to extract substantive semantic units involves: using a sliding window algorithm with overlap constraints for serialization segmentation, setting the window length and overlap step size, with the window length being greater than the overlap step size; inputting the serialized slices into the first language model, and then using a prompt word guidance function combined with a preset positive constraint list and negative constraint list to perform negative constraint filtering and extract substantive semantic units.

3. The method for evaluating pest and disease control levels based on a large language model as described in claim 1, characterized in that, The substantive semantic units are input into a pre-trained text encoding model and encoded into high-dimensional semantic vectors. A clustering algorithm is used to divide the high-dimensional semantic vectors into multiple topic clusters. Specifically, the substantive semantic units are input into a pre-trained text encoding model, which performs contextual semantic understanding on each substantive semantic unit and converts it into a dense vector to construct a high-dimensional semantic vector. The MiniBatchK-Means clustering algorithm is used to perform unsupervised clustering and segmentation of all high-dimensional semantic vectors. The centroid of each cluster is updated by iteratively optimizing the sum of squared errors within the cluster. At the same time, the silhouette coefficient is introduced as an evaluation index. The silhouette coefficient corresponding to each candidate cluster number is calculated within a preset range of candidate cluster numbers. The number of clusters corresponding to the silhouette coefficient reaching the global maximum value is taken as the optimal number of clusters, thus dividing the high-dimensional semantic vectors into multiple issue clusters.

4. The method for evaluating pest and disease control levels based on a large language model as described in claim 1, characterized in that, The stratified sampling of each topic cluster specifically involves: dynamically setting a size threshold based on the mean and standard deviation of the sample size of each topic cluster, including a first threshold for identifying micro-clusters and a second threshold for identifying giant clusters; matching differentiated sampling logic according to cluster size: when the cluster size is less than the first threshold, a full extraction mode is used; when the cluster size is between the first and second thresholds, a mixed sampling mode is used; when the cluster size is greater than or equal to the second threshold, a structured stratified mode is used.

5. The method for evaluating pest and disease control levels based on a large language model as described in claim 1, characterized in that, The preliminary weight scores of each third-level indicator are obtained based on the original attention ratio and semantic confidence. Specifically, the original attention ratio is multiplied by the semantic confidence to obtain the preliminary weight scores of each third-level indicator. The semantic confidence is obtained by measuring the cosine similarity between the generated indicator and the centroid of the underlying corpus.

6. The method for evaluating pest and disease control levels based on a large language model as described in claim 1, characterized in that, The evaluation of pest and disease control level in the region to be assessed is based on the aforementioned pest and disease control level evaluation system. Specifically, the collected control work data is quantitatively scored item by item according to the three-level indicators in the pest and disease control level evaluation system to obtain the actual implementation score of each technical dimension. The comprehensive score of the control work in the region is calculated by weighted summation, taking into account the weights of each three-level indicator.

7. A pest and disease control level evaluation system based on a large language model, characterized in that, include: The text acquisition and semantic extraction module is configured to: acquire the text of technical rules for the prevention and control of crop diseases and pests, segment it using a sliding window, input it into the first language model, and extract substantive semantic units; The semantic encoding and clustering module is configured to: input the substantive semantic units into a pre-trained text encoding model and encode them into high-dimensional semantic vectors; and use a clustering algorithm to divide the high-dimensional semantic vectors into multiple topic clusters. The third-level indicator generation module is configured to: perform hierarchical sampling on each issue cluster and input it into the second large language model, and output a third-level indicator with semantic confidence. The initial evaluation system construction module is configured to: input the tertiary indicators into the third language model to obtain secondary indicators, and construct the initial evaluation system based on the correspondence between the secondary and tertiary indicators; The proportion of the substantive semantic units associated with each third-level indicator to the total sample size is calculated to obtain the original attention proportion. The initial weight scores of each tertiary indicator are obtained based on the original attention ratio and semantic confidence. The initial weight scores of all tertiary indicators are normalized so that the weight of each tertiary indicator is equal to the proportion of its initial weight score to the sum of the initial weight scores of all tertiary indicators, thus obtaining the weight of each tertiary indicator. The final evaluation system construction module is configured to: construct an evaluation system for pest and disease control level based on weighted three-level indicators and the initial evaluation system; The pest and disease control level assessment module is configured to: acquire pest and disease control data for the area to be assessed, and evaluate the level of pest and disease control work in the area to be assessed based on the pest and disease control level evaluation system.

8. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the method for evaluating the level of pest and disease control based on a large language model as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the method for evaluating the level of pest and disease control based on a large language model as described in any one of claims 1-6.

10. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps in the method for evaluating the level of pest and disease control based on a large language model as described in any one of claims 1-6.

Images