A method and device for dynamically constructing an electric power material evaluation sample

Abstract

The application discloses a kind of power material evaluation sample dynamic construction method and device, method includes: obtaining historical original evaluation file and carrying out format conversion, obtain structured initial evaluation sample pool;The entropy of each initial evaluation sample in initial evaluation sample pool is analyzed and is filtered, and evaluation sample set is obtained;Initial evaluation sample in evaluation sample set is standardly marked, and is combined with pre-constructed thought chain model to carry out thought chain analysis, and obtain standard evaluation triplets;According to preset evaluation reward model, standard evaluation triplets are evaluated and screened, and power material evaluation sample library is generated;In response to real-time original evaluation file, the update of power material evaluation sample library is completed, and the dynamic construction of power material evaluation sample library is completed.Based on the construction of power material evaluation sample library, it is updated to ensure the quality of evaluation sample in power material evaluation sample library, and provide good data basis for model training.

Description

Technical Field

[0001] This invention belongs to the technical field of sample construction, specifically relating to a method and apparatus for dynamically constructing bidding samples for power materials. Background Technology

[0002] The evaluation of bids for power materials is a crucial step in the power materials procurement bidding process. It involves professional evaluators comprehensively reviewing, quantitatively assessing, and comparing the bids submitted by various bidders based on the technical standards, pricing specifications, and performance requirements set forth in the bidding documents. The core objective is to select suppliers with advanced technology, reasonable prices, and reliable services from numerous bids through a scientific and standardized evaluation mechanism. The evaluation of bids for power materials not only reduces procurement costs and ensures the quality of materials but also plays a vital role in maintaining fair competition in the power market and promoting technological upgrading within the power industry.

[0003] With the deep penetration of artificial intelligence technology into the power industry, intelligent bidding evaluation systems based on large models have become an important direction for improving the efficiency and accuracy of bidding evaluation. The performance of such large bidding evaluation models is highly dependent on high-quality training samples. The "tender document-bid document-evaluation conclusion" triplet is the core form of training samples, and its quality and timeliness directly determine the bidding evaluation accuracy and generalization ability of the large model.

[0004] Currently, the construction of bidding samples for power materials relies excessively on manual annotation. Sample generation and verification require the participation of numerous bidding experts, which not only consumes significant manpower and time costs but also fails to meet the needs of large-scale sample construction. The quality of manually annotated samples is unstable and easily influenced by subjective factors; different experts may have different annotation standards, leading to large fluctuations in sample quality and affecting the training effect of large-scale bidding models. Furthermore, the distribution of bidding data dynamically shifts with factors such as technological development in the power industry and changes in the market environment. Current sample construction methods lack effective dynamic update mechanisms, failing to capture data changes in a timely manner, resulting in insufficient sample timeliness and affecting the generalization ability of large-scale models.

[0005] How to construct a sample set of bidding data for power materials, and how to achieve iterative updates of the bid evaluation samples while ensuring their quality, is a problem that needs to be solved. Summary of the Invention

[0006] To address the shortcomings of the existing technology, this invention provides a method and apparatus for dynamically constructing bidding samples for power materials. The method includes: acquiring historical original bidding documents; converting the format of the historical original bidding documents to obtain a structured initial bidding sample pool; analyzing the entropy value of each initial bidding sample in the initial bidding sample pool, and filtering the initial bidding samples according to the entropy value to obtain a bidding sample set; standardizing the initial bidding samples in the bidding sample set, and performing mind chain analysis on the standardized initial bidding samples using a pre-constructed mind chain model to obtain standard bidding triples; evaluating and filtering the standard bidding triples according to a preset evaluation reward model, and generating a power materials bidding sample library; and updating the power materials bidding sample library in response to real-time original bidding documents, thus completing the dynamic construction of the power materials bidding sample library.

[0007] By analyzing and filtering the entropy values ​​of each initial bid evaluation sample in the initial bid evaluation sample pool, a bid evaluation sample set is obtained. Combining the thinking chain model, the initial bid evaluation samples with standard annotations are analyzed to obtain standard bid evaluation triples, which are then evaluated and filtered to generate a power materials bid evaluation sample library. Then, in response to the real-time original bid evaluation files, the power materials bid evaluation sample library is updated, completing the dynamic construction of the power materials bid evaluation sample library. Based on the construction of the power materials bid evaluation sample library, the update of the power materials bid evaluation sample library is triggered to ensure the quality of the bid evaluation samples in the power materials bid evaluation sample library and provide a good data foundation for model training.

[0008] In a first aspect, the present invention provides a method for dynamically constructing bidding samples for power materials, specifically including the following steps:

[0009] Obtain the original historical bid evaluation documents;

[0010] The original historical bid evaluation documents were converted into a structured initial bid evaluation sample pool.

[0011] Analyze the entropy value of each initial bid evaluation sample in the initial bid evaluation sample pool, and filter the initial bid evaluation samples according to the entropy value to obtain the bid evaluation sample set;

[0012] The initial bid evaluation samples in the bid evaluation sample set are labeled with standards, and the standard-labeled initial bid evaluation samples are analyzed by combining the pre-constructed mind chain model to obtain standard bid evaluation triples.

[0013] Based on the preset evaluation and reward model, the standard bid evaluation triplet is evaluated and screened, and a sample library of power material bid evaluation is generated.

[0014] It responds to the original bid evaluation documents in real time, updates the bid evaluation sample library for power materials, and completes the dynamic construction of the bid evaluation sample library for power materials.

[0015] Furthermore, the original historical bid evaluation documents are formatted to obtain a structured initial bid evaluation sample pool, specifically including:

[0016] Extract the text content of the historical original bid evaluation documents to obtain the historical original bid evaluation text;

[0017] Data cleaning was performed on the original historical bid evaluation texts to obtain the first bid evaluation text;

[0018] Extract the field content from the first bid evaluation text and perform format conversion to obtain standardized bid evaluation fields and corresponding field labels;

[0019] Based on each evaluation field, the standardization of the first evaluation text is analyzed to determine the structured initial evaluation sample.

[0020] Furthermore, the entropy value of each initial bid evaluation sample in the initial bid evaluation sample pool is analyzed, and the initial bid evaluation samples are filtered according to the entropy value to obtain the bid evaluation sample set, specifically including:

[0021] Using a pre-set sample evaluation model, the initial evaluation samples are analyzed and evaluated to obtain the evaluation probability of different evaluation categories for each initial evaluation sample.

[0022] Calculate the entropy value of each initial bid evaluation sample based on the evaluation probability of different evaluation categories of the initial bid evaluation samples;

[0023] Based on the entropy value of the initial bid evaluation samples, the initial bid evaluation samples are sorted, and a preset number of initial bid evaluation samples are selected as the bid evaluation sample set.

[0024] Furthermore, based on the evaluation probabilities of different evaluation categories in the initial evaluation samples, the entropy value of each initial evaluation sample is calculated, specifically including:

[0025] Substituting the evaluation probabilities of different evaluation categories into a preset logarithmic function yields the first evaluation parameter;

[0026] The first evaluation parameter and the evaluation probability of the same evaluation category are combined to obtain the second evaluation parameter;

[0027] Based on each initial evaluation sample, and combined with the evaluation coefficients of each evaluation category, the second evaluation parameter is weighted and fused to obtain the entropy value of the initial evaluation sample.

[0028] Furthermore, by combining the pre-constructed thought chain model with the initial evaluation samples of the standard annotations, thought chain analysis is performed to obtain the standard evaluation triplet, which specifically includes:

[0029] Extract the evaluation fields and corresponding annotation information from the initial evaluation samples with standard annotations;

[0030] Based on a pre-built evaluation rule library, combined with evaluation fields and corresponding annotation information, corresponding standard evaluation guidance instructions are generated.

[0031] Based on the standard evaluation guidance instructions, analyze the relationship between the evaluation fields and the corresponding annotation information, and in conjunction with the preset evaluation criteria, judge the evaluation fields and the corresponding annotation information, and provide the analysis logic and analysis results;

[0032] Based on the output constraints of the inference chain, the analysis logic and analysis results are organized to obtain the inference chain, where the output constraints of the inference chain contain analysis logic of multiple steps;

[0033] By integrating the evaluation fields, annotation information, and inference chain, a standard evaluation triplet is obtained.

[0034] Furthermore, the evaluation rules in the evaluation rule library include evaluation scenarios, rule application time, and evaluation standard priority, and the annotation information includes annotation scenarios, annotation time, and annotation standards;

[0035] Based on a pre-built evaluation rule library, and combining evaluation fields and corresponding annotation information, corresponding standard evaluation guidance instructions are generated, specifically including:

[0036] The similarity between the evaluation scenarios and the annotation scenarios in the evaluation rules, the matching degree between the rule application time and the annotation time are analyzed respectively, and the scenario similarity and time matching degree are given.

[0037] Based on scene similarity and time matching degree, the priority of the evaluation criteria corresponding to the annotation criteria is integrated to obtain the rule matching degree of each evaluation rule in the evaluation rule base;

[0038] Based on the rule matching degree of each bid evaluation rule in the bid evaluation rule library, the bid evaluation rules are filtered to obtain the target bid evaluation rule, and the corresponding standard bid evaluation guidance instruction is generated by combining it with the instruction generation model.

[0039] Furthermore, scene similarity is obtained through the following steps:

[0040] Obtain the first keyword vector of the evaluation scenario and the second keyword vector of the labeled scenario in each evaluation rule;

[0041] By analyzing the keyword matching degree of the first keyword vector and the second keyword vector, the scene similarity is obtained.

[0042] Furthermore, the time matching degree is obtained through the following steps:

[0043] Based on the applicable time of the rule, the applicable start time and applicable end time shall be given;

[0044] Based on the first time difference between the labeled time and the applicable start time, a time smoothing coefficient is fused, and an activation function is used to process the data to obtain the first time parameters.

[0045] Based on the second time difference between the labeled time and the applicable end time, a time smoothing coefficient is fused, and an activation function is used to process the result to obtain the second time parameter.

[0046] The time matching degree is obtained by multiplying the first time parameter and the second time parameter.

[0047] Furthermore, the instruction generation model includes a compliance layer, a technology layer, and a benefit layer;

[0048] The corresponding standard evaluation guidance instructions are generated by combining the instruction generation model, specifically including:

[0049] The evaluation rules for the target bids are analyzed from the perspectives of compliance, technology, and efficiency, and the corresponding feature vectors for the evaluation rules are given.

[0050] The evaluation scenarios of the target evaluation rules are vectorized to obtain the evaluation scenario features;

[0051] By combining the scene feature weights corresponding to the evaluation scene features, the feature vectors corresponding to the evaluation scene features and the target evaluation rules are concatenated to give the rule-scene fusion vector;

[0052] Based on the rule-based scenario fusion vector, an instruction logic chain is constructed to provide standard evaluation guidance instructions.

[0053] Furthermore, the construction of the thought chain model is determined through the following steps:

[0054] Based on the preset evaluation criteria and combined with historical evaluation data, model evaluation guidance instructions are obtained;

[0055] Based on the pre-trained language model, the model evaluation guidance instructions are logically analyzed to obtain the evaluation logic corresponding to the historical evaluation data.

[0056] Based on the evaluation logic, historical evaluation data is analyzed, and the evaluation criteria results are generated by combining the output constraints in the model's evaluation guidance instructions.

[0057] Based on the accuracy of the evaluation criteria results, the pre-trained language model is iteratively trained until convergence, resulting in the thought chain model.

[0058] Furthermore, based on a pre-defined evaluation and reward model, the standard bid evaluation triplets are evaluated and screened, and a sample library of power material bid evaluations is generated, specifically including:

[0059] The completeness, accuracy, and standardization of the standard evaluation tripartite are analyzed and evaluated respectively, and the completeness score, accuracy score, and standardization score are obtained.

[0060] The completeness score, accuracy score, standardization score, and their corresponding weights are combined to obtain a comprehensive evaluation.

[0061] Based on comprehensive evaluation, the standard bid evaluation triplets are ranked and screened according to preset ratios to obtain a sample library of power material bid evaluations.

[0062] Furthermore, in response to the real-time original bid evaluation documents, the bid evaluation sample library for power materials is updated, and the dynamic construction of the bid evaluation sample library for power materials is completed, specifically including:

[0063] Obtain real-time raw bid evaluation samples and extract real-time bid evaluation fields;

[0064] The distribution of evaluation fields in real-time bidding and in the sample database of power material bidding is analyzed separately to obtain the real-time bidding distribution and the sample bidding distribution.

[0065] The difference between the real-time bid evaluation distribution and the sample bid evaluation distribution triggers an update to the power materials bid evaluation sample database.

[0066] Analyze and annotate the real-time bidding fields and add them to the power material bidding sample library.

[0067] Furthermore, the difference between the real-time bid evaluation distribution and the sample bid evaluation distribution is obtained through KL divergence.

[0068] Secondly, the present invention also provides a dynamic construction device for power material bidding samples, employing any of the above-mentioned methods for dynamic construction of power material bidding samples, including:

[0069] The data acquisition module is used to acquire historical original bid evaluation documents;

[0070] The file processing module is used to convert the original historical bid evaluation documents into a structured initial bid evaluation sample pool.

[0071] The sample analysis module is used to analyze the entropy value of each initial bid evaluation sample in the initial bid evaluation sample pool, and to filter the initial bid evaluation samples based on the entropy value to obtain the bid evaluation sample set;

[0072] The standard annotation module is used to standardize the initial evaluation samples in the evaluation sample set, and combine it with the pre-built thinking chain model to perform thinking chain analysis on the standard-annotated initial evaluation samples to obtain standard evaluation triples.

[0073] The sample library generation module is used to evaluate and screen the standard bid evaluation triplet according to the preset evaluation and reward model, and generate a sample library for power material bid evaluation.

[0074] The sample library update module is used to update the power materials bid evaluation sample library in response to real-time original bid evaluation documents, and to complete the dynamic construction of the power materials bid evaluation sample library.

[0075] The present invention provides a method and apparatus for dynamically constructing bidding samples for power materials, which has at least the following beneficial effects:

[0076] (1) By analyzing and filtering the entropy values ​​of each initial bid evaluation sample in the initial bid evaluation sample pool, a bid evaluation sample set is obtained. The initial bid evaluation sample with standard annotation is analyzed by combining the thinking chain model to obtain the standard bid evaluation triplet and evaluate and filter it to generate the power material bid evaluation sample library. Then, in response to the real-time original bid evaluation file, the power material bid evaluation sample library is updated to complete the dynamic construction of the power material bid evaluation sample library. On the basis of constructing the power material bid evaluation sample library, the update of the power material bid evaluation sample library is triggered to ensure the quality of the bid evaluation samples in the power material bid evaluation sample library and provide a good data foundation for model training.

[0077] (2) By calculating the entropy value and analyzing the entropy value of the initial bid evaluation sample, the most uncertain initial bid evaluation sample can be effectively identified and submitted to experts for annotation, thereby improving the accuracy of the initial bid evaluation sample annotation.

[0078] (3) By analyzing and evaluating the completeness, accuracy and standardization of the standard evaluation triplet, a comprehensive evaluation is obtained, and the power material evaluation sample library is finally selected, which ensures the quality and consistency of the samples in the power material evaluation sample library and improves the accuracy and reliability of the evaluation decision.

[0079] (4) By analyzing the difference between the real-time bidding distribution and the sample bidding distribution, the deviation of the data distribution is detected, and the power material sample library is updated and supplemented to ensure the timeliness and representativeness of the sample library.

[0080] (5) By using the double sigmoid function to process the matching of the rule application time and the annotation time, smooth matching of the boundary ambiguous scene is achieved. By integrating factors such as scene similarity, time matching degree (smoothing processing), and evaluation standard priority, the adaptability of the evaluation rules and annotation information is fully considered, and the scene and time of the evaluation rules and annotation information are accurately matched, thereby improving the quality of the evaluation sample construction. Attached Figure Description

[0081] Figure 1 A flowchart of a method for dynamically constructing bidding samples for power materials provided in an embodiment of the present invention;

[0082] Figure 2 This is a flowchart of obtaining the initial evaluation sample pool provided in an embodiment of the present invention;

[0083] Figure 3 A flowchart for obtaining the evaluation sample set provided in an embodiment of the present invention;

[0084] Figure 4 A flowchart for constructing a thought chain model provided in an embodiment of the present invention;

[0085] Figure 5 Example diagram of the evaluation criteria results provided in the embodiments of the present invention;

[0086] Figure 6 A flowchart for generating a sample library for power material bidding provided in an embodiment of the present invention;

[0087] Figure 7 A flowchart for updating the power material bidding sample library provided in this embodiment of the invention;

[0088] Figure 8 The structural block diagram of the dynamic construction device for power material bidding samples provided in the embodiments of the present invention.

[0089] The modules are as follows: 201. Data acquisition module; 202. File processing module; 203. Sample analysis module; 204. Standard annotation module; 205. Sample library generation module; 206. Sample library update module. Detailed Implementation

[0090] To better understand the above technical solutions, a detailed description of the solutions will be provided below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0091] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms, and “multiple” generally includes at least two unless the context clearly indicates otherwise.

[0092] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0093] Power materials are the core material support for the entire chain of power generation, transmission, transformation, distribution and consumption. They cover various equipment and materials such as power generation equipment, transmission line materials, transformers, switchgear, cables, insulators, and metering instruments. The quality of these materials is directly related to the safe and stable operation of the power system, the reliability of power supply and the construction quality of power projects. They are the basic elements to ensure the sustainable development of the power industry.

[0094] Bidding evaluation for power equipment is a crucial step in the power equipment procurement bidding process. It not only reduces procurement costs and ensures equipment quality, but also plays a vital role in maintaining fair competition in the power market and promoting technological upgrades in the power industry. Intelligent bidding evaluation systems based on large-scale models have become an important direction for improving the efficiency and accuracy of bidding evaluation. The performance of such large-scale models highly depends on high-quality training samples; the quality and timeliness of the training samples directly determine the evaluation accuracy and generalization ability of the large-scale model.

[0095] Currently, the construction of bidding evaluation samples for power materials relies excessively on manual annotation, leading to unstable sample quality and affecting the training effect of large-scale bidding evaluation models. How to construct a sample set of power material bidding data that ensures the quality of the bidding evaluation samples while achieving iterative updates is a problem that needs to be solved.

[0096] Related technologies provide a method and apparatus for constructing tuple samples. The method includes: obtaining a first feature set composed of first sample features; obtaining a preset first tuple sample required for the current training of a model to be trained based on the first feature set; obtaining second sample features by performing forward computation on the model to be trained based on the first tuple sample; updating the first sample features corresponding to the first tuple sample in the first feature set with the second sample features to obtain a second feature set; and obtaining a second tuple sample required for the next training of the model to be trained based on the second feature set. This improves the quality of tuple sample construction, and the tuple samples constructed using this method can adapt to different stages of learning of the model to be trained, improving the convergence speed and performance of the model. However, since the types of evaluation samples involve multiple modalities, general sample construction methods are not suitable for evaluation samples.

[0097] To address the issues of unstable quality and difficulty in adapting to data changes in power material bidding samples, this invention provides a dynamic construction method for power material bidding samples. The method includes: acquiring historical original bidding documents; converting the format of the historical original bidding documents to obtain a structured initial bidding sample pool; analyzing the entropy value of each initial bidding sample in the initial bidding sample pool and filtering the initial bidding samples based on the entropy value to obtain a bidding sample set; standardizing the initial bidding samples in the bidding sample set and performing mind chain analysis on the standardized initial bidding samples using a pre-constructed mind chain model to obtain standard bidding triples; evaluating and filtering the standard bidding triples according to a preset evaluation reward model to generate a power material bidding sample library; and updating the power material bidding sample library in response to real-time original bidding documents, thus completing the dynamic construction of the power material bidding sample library. This end-to-end design achieves continuous output of high-quality "bid document-bid proposal-bidding conclusion" triples, providing stable training support for the bidding model.

[0098] like Figure 1 As shown in the figure, this embodiment of the invention provides a method for dynamically constructing bidding samples for power materials, and the specific steps are as follows:

[0099] S101: Obtain the original historical bid evaluation documents.

[0100] Historical original bid evaluation documents include the technical specifications in the bidding documents, the technical response documents submitted by the bidders, the index comparison reports issued by third-party testing institutions, and the review conclusions formed by the historical bid evaluation committees.

[0101] S102: Convert the format of the historical original bid evaluation documents to obtain a structured initial bid evaluation sample pool.

[0102] Furthermore, the original historical bid evaluation documents are converted into a structured initial bid evaluation sample pool, which is then referenced. Figure 2 Specifically, it includes:

[0103] Extract the text content of the historical original bid evaluation documents to obtain the historical original bid evaluation text;

[0104] Data cleaning was performed on the original historical bid evaluation texts to obtain the first bid evaluation text;

[0105] Extract the field content from the first bid evaluation text and perform format conversion to obtain standardized bid evaluation fields and corresponding field labels;

[0106] Based on each evaluation field, the standardization of the first evaluation text is analyzed to determine the structured initial evaluation sample.

[0107] In one specific implementation, text recognition technologies such as OCR are first used to extract the content of historical original bid evaluation documents, obtaining historical original bid evaluation text. A structured parsing tool is then used to extract fields from the historical original bid evaluation text, constructing a unified initial bid evaluation sample pool. This involves converting all data into a unified structured format, such as a table, containing fields such as material type, model, technical parameters, and price. Data cleaning is then performed to remove duplicate data, correct erroneous data, and fill in missing values. Standardized bid evaluation fields include key fields such as core bidding parameters (e.g., equipment operating environment requirements, performance index thresholds), bid technical responses (e.g., parameter matching descriptions, explanations of technical deviations), indicator comparison data (e.g., analysis of differences between the bid product and industry standards), and historical bid evaluation conclusions (e.g., technical compliance judgments, explanations of scoring criteria).

[0108] Then, relevant fields are extracted from the standardized evaluation fields for subsequent entropy analysis. These include: technical parameters such as transformer rated capacity and insulation class, and wire and cable material and conductivity; price information such as unit price and total price; historical evaluation results including past winning bids and expert evaluations; and supplier information such as supplier reputation and past cooperation records.

[0109] Field labels are used to distinguish the content of different evaluation fields, ensuring the readability and processability of the evaluation field content. Each evaluation field has a unique label, which can be a simple text identifier or a numerical code.

[0110] Field standardization measures the normalization and consistency of field content, specifically expressed as:

[0111]

[0112] Where FSD represents the field standardization degree of the first evaluation text, and L... i M represents the valid length of the i-th evaluation field in the first evaluation text. i M is the standardized label for the i-th evaluation field in the first evaluation text. i =1 indicates that the content of the i-th evaluation field in the first evaluation text is standardized, M i =0 indicates that the content of the i-th evaluation field in the first evaluation text is not standardized, L sum is the total length of all evaluation fields in the first evaluation text, and n is the number of evaluation fields in the first evaluation text.

[0113] In a specific example, if the first evaluation text is for evaluating the transformer, then the corresponding evaluation fields include:

[0114] From the evaluation field 1 (tender parameter item), field label (tender parameter), and the corresponding field content (rated capacity: 300kVA; insulation class: C; price limit: 12000 yuan), we can obtain the effective length L1=40 (number of characters) of evaluation field 1, and the standardization mark: M1=1 (content standardization).

[0115] From the evaluation field 2 (bid response content), field label (bid response), and the corresponding field content (rated capacity: 300kVA; insulation class: C; price: 11500 yuan), we can obtain the effective length L2=42 (number of characters) of evaluation field 2, and the standardization mark: M2=1 (content standardization).

[0116] From the evaluation field 3 (technical indicator comparison data), field label (technical comparison), and the corresponding field content (rated capacity matching: yes; insulation level matching: yes; price within the upper limit: yes), we can obtain L3=50 (number of characters) and standardization mark M3=1 (content standardization) for evaluation field 3.

[0117] By analyzing the evaluation field 4 (historical evaluation conclusion), the field label (evaluation conclusion), and the corresponding field content (the transformer meets the bidding requirements, quality level: high; cost-effectiveness: high), we can obtain the evaluation field L4=45 (number of characters) and the standardization mark M4=1 (content standardization).

[0118] The degree of field standardization is calculated as follows:

[0119] Calculate the contribution of standardization for each field: Field 1: L1×M1=40×1=40; Field 2: L2×M2=42×1=42; Field 3: L3×M3=50×1=50; Field 4: L4×M4=45×1=45.

[0120] The total length L of all evaluation fields in the first evaluation text sum =L1+L2+L3+L4=40+42+50+45=177.

[0121] The field standardization degree (FSD) of the first evaluation text is calculated as follows: FSD = (40 + 42 + 50 + 45) / 177 = 1

[0122] By converting the original historical bid evaluation documents into a new format, a unified initial bid evaluation sample pool is obtained. The use of field labels allows each bid evaluation field to be accurately distinguished and processed, providing a high-quality data foundation for the subsequent initial bid evaluation sample pool.

[0123] S103: Analyze the entropy value of each initial bid evaluation sample in the initial bid evaluation sample pool, and filter the initial bid evaluation samples according to the entropy value to obtain the bid evaluation sample set.

[0124] Furthermore, the entropy value of each initial bid evaluation sample in the initial bid evaluation sample pool is analyzed, and the initial bid evaluation samples are filtered according to the entropy value to obtain the bid evaluation sample set, referring to... Figure 3 Specifically, it includes:

[0125] Using a pre-set sample evaluation model, the initial evaluation samples are analyzed and evaluated to obtain the evaluation probability of different evaluation categories for each initial evaluation sample.

[0126] Calculate the entropy value of each initial bid evaluation sample based on the evaluation probability of different evaluation categories of the initial bid evaluation samples;

[0127] Based on the entropy value of the initial bid evaluation samples, the initial bid evaluation samples are sorted, and a preset number of initial bid evaluation samples are selected as the bid evaluation sample set.

[0128] In one specific implementation, a sample evaluation model is used to evaluate the input initial evaluation samples and obtain the evaluation probability of each evaluation category corresponding to the initial evaluation samples. The preset sample evaluation model can be a machine learning model, or it can use other methods such as classifiers or random forests; there is no limitation on this. It analyzes and evaluates the initial evaluation samples to obtain the evaluation probability of different evaluation categories for each initial evaluation sample. For example, if the initial evaluation samples are set to have three levels: high, medium, and low, the evaluation probability of each level is obtained through the sample evaluation model. Then, the entropy value of each initial evaluation sample is calculated. Entropy can measure the uncertainty of the model's classification of the sample. The higher the entropy, the more uncertain the model's classification of the sample.

[0129] Furthermore, based on the evaluation probabilities of different evaluation categories in the initial evaluation samples, the entropy value of each initial evaluation sample is calculated, specifically including:

[0130] Substituting the evaluation probabilities of different evaluation categories into a preset logarithmic function yields the first evaluation parameter;

[0131] The first evaluation parameter and the evaluation probability of the same evaluation category are combined to obtain the second evaluation parameter;

[0132] Based on each initial evaluation sample, and combined with the evaluation coefficients of each evaluation category, the second evaluation parameter is weighted and fused to obtain the entropy value of the initial evaluation sample.

[0133] Entropy, specifically represented as:

[0134]

[0135] Where H(x) is the entropy value of the initial evaluation sample x, and Pi Let x be the evaluation probability that the initial evaluation sample x belongs to evaluation category i.

[0136] In a specific example, the probability of the initial evaluation sample x belonging to the evaluation category "excellent" is 0.5, and the probability of it belonging to the evaluation category "good" is also 0.5. Then the entropy value is: H(x)=−(0.5log0.5+0.5log0.5)=1, indicating that the classification of the initial evaluation sample x is very uncertain.

[0137] Calculate the entropy value of each initial bid evaluation sample, sort the initial bid evaluation samples in descending order, and select a predetermined number of initial bid evaluation samples as the bid evaluation sample set. In this example, the top 1% of the initial bid evaluation samples are selected as the bid evaluation sample set.

[0138] In a specific example, there are 1000 initial evaluation samples. The entropy value of each initial evaluation sample is calculated. For example, the entropy value of sample 1 is 1.0, the entropy value of sample 2 is 0.8, the entropy value of sample 3 is 0.9, and so on. All samples are sorted from highest to lowest entropy value, resulting in the following ranking: Sample 1 (entropy 1.0), Sample 3 (entropy 0.9), Sample 2 (entropy 0.8), and so on. The top 10 samples (1%) with the highest entropy values ​​are selected to form the evaluation sample set.

[0139] By calculating entropy values ​​to analyze the uncertainty of the initial bid evaluation samples, the most uncertain initial bid evaluation samples can be effectively identified and submitted to experts for annotation, thereby improving the accuracy of the initial bid evaluation sample annotation.

[0140] S104: Standardize the initial bid evaluation samples in the bid evaluation sample set, and combine them with the pre-constructed mind chain model to perform mind chain analysis on the standardized initial bid evaluation samples to obtain the standard bid evaluation triplet.

[0141] Furthermore, by combining the pre-constructed thought chain model with the initial evaluation samples of the standard annotations, thought chain analysis is performed to obtain the standard evaluation triplet, which specifically includes:

[0142] Extract the evaluation fields and corresponding annotation information from the initial evaluation samples with standard annotations;

[0143] Substitute the evaluation fields and corresponding annotation information into the thinking chain model, analyze the relationship between the evaluation fields and annotation information, and generate a reasoning chain;

[0144] By integrating the evaluation fields, annotation information, and inference chain, a standard evaluation triplet is obtained.

[0145] Furthermore, the evaluation fields and corresponding annotation information are substituted into the thinking chain model to analyze the relationship between the evaluation fields and annotation information, generating a reasoning chain, specifically including:

[0146] Based on a pre-built evaluation rule library, combined with evaluation fields and corresponding annotation information, corresponding standard evaluation guidance instructions are generated.

[0147] Based on the standard evaluation guidance instructions, analyze the relationship between the evaluation fields and the corresponding annotation information, and in conjunction with the preset evaluation criteria, judge the evaluation fields and the corresponding annotation information, and provide the analysis logic and analysis results;

[0148] Based on the output constraints of the inference chain, the analysis logic and analysis results are organized to obtain the inference chain, where the output constraints of the inference chain contain analysis logic for multiple steps.

[0149] It is understandable that the standard evaluation guidance instructions and the model evaluation guidance instructions correspond to each other. Different instructions are generated based on different evaluation fields and corresponding annotation information to achieve the effect of using a pre-trained language model for logical analysis. Among them, the analysis logic corresponds to the evaluation logic, and the analysis results correspond to the judgment results of historical evaluation data, that is, the conclusions in the examples described later.

[0150] Furthermore, the evaluation rules in the evaluation rule library include evaluation scenarios, rule application time, and evaluation standard priority, and the annotation information includes annotation scenarios, annotation time, and annotation standards;

[0151] Based on a pre-built evaluation rule library, and combining evaluation fields and corresponding annotation information, corresponding standard evaluation guidance instructions are generated, specifically including:

[0152] The similarity between the evaluation scenarios and the annotation scenarios in the evaluation rules, the matching degree between the rule application time and the annotation time are analyzed respectively, and the scenario similarity and time matching degree are given.

[0153] Based on scene similarity and time matching degree, the priority of the evaluation criteria corresponding to the annotation criteria is integrated to obtain the rule matching degree of each evaluation rule in the evaluation rule base;

[0154] Based on the rule matching degree of each bid evaluation rule in the bid evaluation rule library, the bid evaluation rules are filtered to obtain the target bid evaluation rule, and the corresponding standard bid evaluation guidance instruction is generated by combining it with the instruction generation model.

[0155] Furthermore, scene similarity is obtained through the following steps:

[0156] Obtain the first keyword vector of the evaluation scenario and the second keyword vector of the labeled scenario in each evaluation rule;

[0157] By analyzing the keyword matching degree of the first keyword vector and the second keyword vector, the scene similarity is obtained.

[0158] Scene similarity S(rs, ss) is specifically represented as:

[0159]

[0160] in, This is the first keyword vector for the evaluation scenario (rs). Let s be the second keyword vector of the labeled scene s, and ||·|| be the first norm.

[0161] Scene similarity is used to measure the degree of matching between the evaluation scene and the labeled scene for each evaluation rule. In a specific example, the first keyword vector can be "110kV", "transformer", "open tender", etc. The second keyword vector can be "110kV", "transformer", "invitation to tender", etc.

[0162] Furthermore, the time matching degree is obtained through the following steps:

[0163] Based on the applicable time of the rule, the applicable start time and applicable end time shall be given;

[0164] Based on the first time difference between the labeled time and the applicable start time, a time smoothing coefficient is fused, and an activation function is used to process the data to obtain the first time parameters.

[0165] Based on the second time difference between the labeled time and the applicable end time, a time smoothing coefficient is fused, and an activation function is used to process the result to obtain the second time parameter.

[0166] The time matching degree is obtained by multiplying the first time parameter and the second time parameter.

[0167] The time matching degree V(rt,st) is specifically represented as:

[0168]

[0169] Where V(rt,st) is the temporal matching degree between the rule application time rt and the annotation time st, σ() is the activation function, k is the temporal smoothing coefficient, and t sample To indicate the time, t start To apply the start time, t end To apply the end time, σ(k·(t) sampl et start )) is the first time parameter, σ(-k·(t) sample -t end )) is the second time parameter.

[0170] The time matching score measures the smoothness of the boundary between the annotation time and the rule's applicable time. It is calculated using a double sigmoid function, where the sigmoid function is the activation function used to convert the first and second time differences into a validity value between 0 and 1. The variable k controls the boundary smoothness; a larger k results in a steeper boundary. In this example, k = 0.1. In a specific example, the rule's applicable time is [2023-01-01, 2025-12-31]. If the annotation time is 2025-08-01, it means the annotation time is within the rule's applicable time range, and V(rt,st)≈0.6. If the annotation time is 2026-01-01, it means the annotation time is outside the rule's applicable time range, and V(rt,st)≈0.1. A higher time matching score indicates higher validity, and a lower time matching score indicates lower validity.

[0171] The rule matching degree RMS(r,s) of each evaluation rule in the evaluation rule base is specifically expressed as follows:

[0172]

[0173] Wherein, RMS(r,s) represents the degree of rule matching between evaluation rule r and annotation information s in the evaluation rule base, α is the first weight coefficient, β is the second weight coefficient, P(rp) is the priority of the evaluation standard corresponding to the annotation standard, and the first and second weight coefficients are adjusted according to the importance of the evaluation rule.

[0174] The priority of the evaluation criteria corresponding to the labeling criteria is set according to the authority of the evaluation criteria. For example, if the evaluation criteria is a national mandatory standard, the priority of the evaluation criteria P(rp) = 1; if the evaluation criteria is an industry standard, the priority of the evaluation criteria P(rp) = 0.8; if the evaluation criteria is an enterprise standard, the priority of the evaluation criteria P(rp) = 0.5; and if the evaluation criteria is a custom rule, the priority of the evaluation criteria P(rp) = 0.3.

[0175] By using a double sigmoid function to match the applicable time of the rules with the annotation time, a smooth matching of scenarios with ambiguous boundaries is achieved. By integrating factors such as scene similarity, time matching degree (smoothing processing), and evaluation criteria priority, the adaptability of evaluation rules and annotation information is fully considered, and the scene and time of evaluation rules and annotation information are accurately matched, thereby improving the quality of evaluation sample construction.

[0176] The evaluation rules in the evaluation rule library are analyzed, and each rule corresponds to a rule matching degree. The evaluation rules are sorted in descending order according to the rule matching degree, and the evaluation rule with the highest rule matching degree is selected to obtain the target evaluation rule. After obtaining the target evaluation rule, the target evaluation rule, evaluation fields, and corresponding annotation information are input into the instruction generation model to generate the corresponding standard evaluation guidance instruction.

[0177] By incorporating scenario similarity and time matching into the selection process of the bidding evaluation rules, we avoid hard judgments on scenario validity and the inability to handle scenarios with ambiguous boundaries. We also incorporate the priority of evaluation criteria, considering the priority of evaluation rules (e.g., national mandatory standards are higher than industry standards), to improve the accuracy of the selection results.

[0178] Furthermore, the instruction generation model includes a compliance layer, a technology layer, and a benefit layer;

[0179] The corresponding standard evaluation guidance instructions are generated by combining the instruction generation model, specifically including:

[0180] The evaluation rules for the target bids are analyzed from the perspectives of compliance, technology, and efficiency, and the corresponding feature vectors for the evaluation rules are given.

[0181] The evaluation scenarios of the target evaluation rules are vectorized to obtain the evaluation scenario features;

[0182] By combining the scene feature weights corresponding to the evaluation scene features, the feature vectors corresponding to the evaluation scene features and the target evaluation rules are concatenated to give the rule-scene fusion vector;

[0183] Based on the rule-based scenario fusion vector, an instruction logic chain is constructed to provide standard evaluation guidance instructions.

[0184] In one specific implementation, the compliance layer is used to reason about the evaluation fields and corresponding annotation information based on the target evaluation rules, and to provide the feature vector corresponding to the compliance layer. In a specific example, the target evaluation rule is "the bidder possesses a Class A qualification for general contracting of power engineering construction". The bidder is found in the evaluation fields, and the existence of "Class A qualification for general contracting of power engineering construction" is determined from the annotation information. If it exists, it is determined whether the qualification is valid. If it is, the feature vector corresponding to the compliance layer is "compliant and passed"; otherwise, the feature vector corresponding to the compliance layer is "compliant and rejected". The technical layer is used to determine the data deviation of the technical parameters in the evaluation fields. In a specific example, the target evaluation rule is "the rated capacity of the transformer is 200kV and the deviation range is ±5%". The evaluation field shows "Transformer rated capacity is 300kV". The deviation rate corresponding to this field is calculated as follows: Deviation rate = (Bid value in the evaluation field - Value in the target evaluation rule) × 100% / Value in the target evaluation rule = (300kV - 200kV) × 100% / 200kV = 50% > 5%, which does not meet the target evaluation rule. Therefore, the feature vector corresponding to the technical layer is "Technical rejection". In other examples, if the deviation rate ≤ 0%, the feature vector corresponding to the technical layer is "Technical pass"; if the deviation rate ≤ the deviation range, the feature vector corresponding to the technical layer is "Technical pass and can trigger expert review"; if the deviation rate > the deviation range, the feature vector corresponding to the technical layer is "Technical rejection". The benefit layer is used to judge the bid price in the evaluation field. The deviation rate of the bid price is calculated as follows: Deviation rate of bid price = (Bid price in the evaluation field - Bid price in the target evaluation rule) × 100% / Bid price in the target evaluation rule. Depending on the different bid price deviation rates, the feature vector corresponding to the benefit layer is either "reasonable price" or "price rejection".

[0185] Based on the analysis of the compliance layer, technology layer, and benefit layer in the instruction generation model, different feature vectors are obtained from three aspects. The compliance layer corresponds to laws, regulations, and qualification requirements, such as verifying the bidder's qualifications. The technology layer corresponds to parameter standards and performance indicators, such as verifying equipment parameters. The benefit layer corresponds to price deviations and cost requirements, such as considering the bid price.

[0186] The characteristics of the bidding evaluation scenario can be at the material level (material type, voltage level requirements), bidding method (open bidding / competitive negotiation), and usage environment (region, altitude, climate), etc., which can be obtained from the bidding documents. The scenario feature weights are calculated using the entropy weight method. The product of the bidding evaluation scenario features and the scenario feature weights is concatenated with the feature vectors of the compliance layer, technology layer, and benefit layer to generate a rule-based scenario fusion vector. Based on the rule-based scenario fusion vector, an instruction logic chain is constructed, and standard bidding evaluation guidance instructions are given.

[0187] By incorporating the characteristics of the bidding evaluation scenario (such as "material type = transformer", "bidding method = open bidding", "voltage level = 110kV") into the standard bidding evaluation guidance instructions, the scenario adaptability of the instructions is ensured.

[0188] In a specific example, dependency parsing is used to parse the logical structure of the target evaluation rule (such as the conditional clause "If A, then B"). For example, the dependency parsing tree for "The bidder must possess a Class A qualification for general contracting of power engineering construction" is: "Bidder (subject) - Possess (predicate) - Qualification (object) - Class A qualification for general contracting of power engineering construction (modifier)". Then, the evaluation scenario features (such as "public bidding for transformers") are inserted into the logical structure of the target evaluation rule as scenario qualifiers. For example, inserting "Applicable to the public bidding scenario for transformers" into the object of the above target evaluation rule forms "Possess a Class A qualification for general contracting of power engineering construction (applicable to the public bidding scenario for transformers)". Based on the logical structure of the target evaluation rule (such as "Verify whether [subject] is [predicate] [object] (scenario qualifier)"), an instruction template is generated. For example, the logical structure of the rule is "Verify whether [the bidder] has a Class A qualification for general contracting of power engineering construction (applicable to the [transformer public bidding] scenario)", and the corresponding template is "Verify whether {subject} is {predicate} {object} (applicable to the {scenario} scenario)".

[0189] Enter the content of the target evaluation rule and the characteristics of the evaluation scenario into the template to generate a standard evaluation guidance instruction. For example, if the target evaluation rule is "the bidder must have a Class A qualification for general contracting of power engineering construction", and the evaluation scenario characteristic is "public bidding for transformers", the generated standard evaluation guidance instruction would be: "Verify whether the bidder has a Class A qualification for general contracting of power engineering construction".

[0190] By fusing vectors with rules and scenarios, the target evaluation rules and evaluation scenarios are adapted to ensure that the generated standard evaluation guidance instructions are highly relevant to the current evaluation scenario.

[0191] Furthermore, referring to Figure 4 The construction of the mind chain model is determined through the following steps:

[0192] Based on the preset evaluation criteria and combined with historical evaluation data, model evaluation guidance instructions are obtained;

[0193] Based on the pre-trained language model, the model evaluation guidance instructions are logically analyzed to obtain the evaluation logic corresponding to the historical evaluation data.

[0194] Based on the evaluation logic, historical evaluation data is analyzed, and the evaluation criteria results are generated by combining the output constraints in the model's evaluation guidance instructions.

[0195] Based on the accuracy of the evaluation criteria results, the pre-trained language model is iteratively trained until convergence, resulting in the thought chain model.

[0196] In one specific implementation, a pre-trained language model (PLM) is a deep learning model that performs unsupervised or semi-supervised learning on large-scale text data. A pre-trained language model can capture grammatical, semantic, and contextual information of a language to generate text, understand text, and perform various natural language processing tasks, such as text classification, sentiment analysis, machine translation, and question-answering systems. Pre-trained language models can be BERT (bidirectional encoder representations from transformers), GPT (generative pre-trained transformer), RoBERTa (a robustly optimized BERT approach), T5 (text-to-text transfer transformer), etc., and are not limited thereto.

[0197] How to obtain guidance instructions for model evaluation so that it can guide the pre-trained language model to perform logical reasoning and generate appropriate thought chains? In the context of power material bidding evaluation, the specific steps are as follows:

[0198] First, clearly define the evaluation criteria and output limitations for the evaluation task. For example, specify the evaluation criteria, such as requirements for technical parameters, price, and supplier qualifications. Output a "problem-reasoning chain-conclusion" triple as the evaluation criterion result, and refer to... Figure 5After clarifying the output constraints of the pre-trained language model, historical evaluation data is analyzed to extract key information and logical relationships, resulting in model evaluation guidance instructions. These instructions include the evaluation task, evaluation logic, historical evaluation data, and output constraints. In a specific example, the evaluation task is to determine the cost-effectiveness level of a transformer based on its technical parameters and price. The cost-effectiveness level is divided into three categories: high, medium, and low. The evaluation logic is as follows: 1. Analyze the transformer's rated capacity and assess its applicable scenarios; 2. Analyze the transformer's insulation level and assess its safety and reliability; 3. Analyze the transformer's price and assess its cost-effectiveness; 4. Considering both technical parameters and price, determine the cost-effectiveness level. Historical evaluation data shows that: 1. Transformer model S15-M-300, rated capacity 300kVA, insulation class C, price 12,000 yuan, was rated by experts as high cost-effectiveness; 2. Transformer model S13-M-200, rated capacity 200kVA, insulation class B, price 8,000 yuan, was rated by experts as medium cost-effectiveness; 3. Transformer model S11-M-100, rated capacity 100kVA, insulation class A, price 5,000 yuan, was rated by experts as low cost-effectiveness. The output constraints include three parts: the problem, the reasoning chain, and the conclusion. The problem is a specific description of the evaluation task; the reasoning chain includes a step-by-step logical reasoning process; and the conclusion provides a quantitative score and cost-effectiveness rating.

[0199] The aforementioned model evaluation guidance instructions are integrated and input into the pre-trained language model. The pre-trained language model performs logical analysis based on the model evaluation guidance instructions to obtain the evaluation logic corresponding to the historical evaluation data. Combining the evaluation logic, the historical evaluation data is evaluated, and the evaluation criteria result is generated based on the output constraints in the model evaluation guidance instructions. It can be understood that the evaluation criteria result and the output constraints correspond; that is, the evaluation criteria result meets the requirements of the output constraints.

[0200] In a specific example, the historical evaluation data consists of relevant parameters of a transformer of model S15-M-300. Based on the evaluation criteria, a model evaluation guidance instruction is generated. Combining the evaluation logic, the historical evaluation data is evaluated, and combined with the output constraints in the model evaluation guidance instruction, the evaluation criteria result is generated, i.e., a "problem-reasoning chain-conclusion" triple is generated.

[0201] In this example, the evaluation criteria result is: "Question: Determine the cost-effectiveness level of the transformer based on its technical parameters and price."

[0202] Reasoning Chain: 1. Rated capacity is 300kVA, classifying it as a high-capacity transformer, typically used in large power systems; 2. Insulation class is C, indicating good insulation performance, suitable for high-voltage environments; 3. Price is 12,000 yuan, which is relatively high, but considering its high capacity and good insulation performance, the price is within a reasonable range; 4. Combined with the expert's "high" cost-performance rating, it can be inferred that this transformer has a high balance between performance and price, making it suitable for use in large power systems.

[0203] Conclusion: This model of transformer offers a high cost-performance ratio.

[0204] Based on the accuracy of the evaluation criteria results, the pre-trained language model is iteratively trained until convergence, resulting in the thought chain model.

[0205] The accuracy of the bid evaluation criteria results includes the accuracy of technical parameters, the rationality of logical reasoning, and the correctness of the conclusions. Specifically, the accuracy of technical parameters refers to whether the technical parameters mentioned in the bid evaluation criteria are consistent with actual data. The rationality of logical reasoning refers to whether the reasoning process conforms to the bid evaluation logic and industry standards. The correctness of the conclusions refers to whether they are consistent with expert annotations or the actual bid evaluation results.

[0206] Based on the above three aspects, accuracy indicators are constructed. The accuracy of technical parameters is represented by the parameter matching rate, which is the proportion of technical parameters mentioned in the evaluation criteria that match the actual data. Specifically, the parameter matching rate = number of matched parameters / total number of parameters. The rationality of logical reasoning is represented by the logical consistency score, which is scored by experts on the logic of the reasoning chain. Specifically, the logical consistency score is the sum of expert scores divided by the number of experts. The correctness of the conclusion is represented by the conclusion accuracy rate, which is the proportion of the conclusion in the evaluation criteria that is consistent with the expert annotations or the actual evaluation results. Conclusion accuracy rate = number of samples with correct conclusions / total number of samples.

[0207] Based on the specific examples above, the rated capacity, insulation class, and price mentioned in the evaluation criteria are consistent with the actual data, therefore the parameter matching rate is 1. Three experts scored the logical consistency of the reasoning chain, giving scores of 8, 9, and 7 respectively. Therefore, the logical consistency score is (8+9+7) / 3=8. Based on 100 historical evaluation data points, the conclusions of 90 of the evaluation criteria results are consistent with the expert annotations, therefore the conclusion accuracy rate is 90 / 100=0.9.

[0208] The above accuracy indicators are combined to form a comprehensive evaluation index. A weighted average method can be used to assign different weights to each accuracy indicator according to its importance, thus obtaining the comprehensive evaluation index.

[0209] The above steps quantify the accuracy of the evaluation criteria results. Based on this, the pre-trained language model is trained to obtain the thought chain model.

[0210] Understandably, step S103, which uses entropy values ​​to filter uncertain initial bid evaluation samples, could directly add the filtered results to the power materials bid evaluation sample library. While this increases the number of sample types, it also affects the stability of the sample library. When using the sample library for model training, these uncertain initial bid evaluation samples would influence the model's training accuracy. If only step S104 is used without entropy value filtering, all initial bid evaluation samples would need to be labeled, increasing the workload of standard labeling and reducing its accuracy, thus affecting the accuracy of the samples in the sample library. Therefore, combining steps S103 and S104 retains the uncertain initial bid evaluation samples while simultaneously standardizing them, ensuring the accuracy of the bid evaluation sample set. This approach effectively balances sample diversity and accuracy, guaranteeing the sample quality in the power materials bid evaluation sample library.

[0211] S105: Based on the preset evaluation and reward model, evaluate and screen the standard bid evaluation triplet and generate a sample library of power material bid evaluations.

[0212] Furthermore, based on the pre-set evaluation and reward model, the standard bid evaluation triplets are evaluated and screened, and a sample library of power material bid evaluations is generated, referring to... Figure 6 Specifically, it includes:

[0213] The completeness, accuracy, and standardization of the standard evaluation tripartite are analyzed and evaluated respectively, and the completeness score, accuracy score, and standardization score are obtained.

[0214] The completeness score, accuracy score, standardization score, and their corresponding weights are combined to obtain a comprehensive evaluation.

[0215] Based on comprehensive evaluation, the standard bid evaluation triplets are ranked and screened according to preset ratios to obtain a sample library of power material bid evaluations.

[0216] In one specific implementation, completeness primarily checks whether the standard evaluation triple contains all necessary information. For example, a complete standard evaluation triple should include the following three parts: question, reasoning chain, and conclusion. If the standard evaluation triple contains all parts, the completeness score is 100; if any part is missing, the completeness score is reduced accordingly. For example, if the conclusion is missing, the completeness score is 60. A corresponding completeness score is defined based on the importance of the missing part. Accuracy evaluation primarily checks whether the information in the standard evaluation triple is consistent with the actual situation. Whether the content mentioned in the standard evaluation triple is consistent with the actual data. If all information is accurate, the accuracy score is 100. If errors exist, the accuracy score is reduced accordingly. For example, if multiple technical parameters are incorrect, the accuracy score is 75. Normative evaluation primarily checks whether the format and content of the standard evaluation triple conform to predefined standards. For example, whether the format of the standard evaluation triple conforms to predefined output constraints, whether the logic of the reasoning chain is clear, and whether the conclusion is explicit. If all content conforms to the norms, the normative score is 100. If there are non-norms, the normative score is reduced accordingly. For example, the reasoning chain is not logically clear, and the normativity score is 80.

[0217] The completeness score, accuracy score, and standardization score, along with their corresponding weights, are combined to obtain a comprehensive evaluation. For example, the weight for the completeness score is 0.3, the weight for the accuracy score is 0.4, and the weight for the standardization score is 0.3. Based on these weights, a weighted sum is performed to obtain the comprehensive evaluation. All standard evaluation triplets are then sorted in descending order according to the comprehensive evaluation. A preset ratio is then used for screening to obtain the power materials evaluation sample library. Assuming the preset ratio is 80%, the top 80% of standard evaluation triplets with the highest comprehensive evaluations are retained to form the power materials evaluation sample library.

[0218] By analyzing and evaluating the completeness, accuracy, and standardization of the standard bid evaluation triplet, a comprehensive evaluation is obtained, and a power materials bid evaluation sample library is finally selected. This ensures the quality and consistency of the samples in the power materials bid evaluation sample library and improves the accuracy and reliability of bid evaluation decisions.

[0219] S106: Respond to the original bid evaluation documents in real time, update the bid evaluation sample library for power materials, and complete the dynamic construction of the bid evaluation sample library for power materials.

[0220] Furthermore, in response to the real-time original bid evaluation documents, the bid evaluation sample library for power materials is updated, completing the dynamic construction of the bid evaluation sample library for power materials, with reference to... Figure 7 Specifically, it includes:

[0221] Obtain real-time raw bid evaluation samples and extract real-time bid evaluation fields;

[0222] The distribution of evaluation fields in real-time bidding and in the sample database of power material bidding is analyzed separately to obtain the real-time bidding distribution and the sample bidding distribution.

[0223] The difference between the real-time bid evaluation distribution and the sample bid evaluation distribution triggers an update to the power materials bid evaluation sample database.

[0224] Analyze and annotate the real-time bidding fields and add them to the power material bidding sample library.

[0225] Furthermore, the difference between the real-time bid evaluation distribution and the sample bid evaluation distribution is obtained through KL divergence.

[0226] Understandably, after analyzing historical original bid evaluation documents and initially constructing a power materials bid evaluation sample library, real-time original bid evaluation documents will be generated over time. If the power materials bid evaluation sample library is updated in real-time, it will affect its stability and accuracy. If updates are performed on a timed basis, the update efficiency will be affected when the number of original bid evaluation documents in a certain period is too large or too small. Therefore, to ensure the effectiveness of updating the power materials bid evaluation sample library, real-time original bid evaluation samples are analyzed and compared with samples in the power materials bid evaluation sample library. When certain conditions are met, an update of the power materials sample library is triggered, thus improving the update efficiency.

[0227] The process involves acquiring real-time raw bidding samples and extracting real-time bidding fields. These fields are then combined to form a real-time sample library, which consists of the real-time bidding fields of all raw bidding samples obtained after the most recent update of the power materials bidding sample library. The distribution of these real-time bidding fields in the sample library is then calculated, including average, maximum, and minimum values, to obtain the real-time bidding distribution. Simultaneously, the distribution of bidding fields within the power materials bidding sample library is also calculated, yielding the sample bidding distribution. Finally, the difference between the real-time and sample bidding distributions, known as the KL divergence, is analyzed. When this difference reaches a preset update threshold, an update to the power materials bidding sample library is triggered. The real-time bidding fields are then analyzed, labeled, and added to the power materials bidding sample library. When the distribution difference does not reach the preset update threshold, the power materials bidding sample library is not updated. Instead, it continues to absorb real-time original bidding samples, adding data to the real-time sample library, judging the new distribution difference, and repeating the above process to continuously update the power materials bidding sample library. In other words, updating the power materials bidding samples involves adding the content from the real-time sample library, clearing the real-time sample library after addition, and then continuing to receive real-time original bidding samples, repeating the above process.

[0228] KL divergence is an asymmetric measure of the difference between two probability distributions, exhibiting both non-negativity and asymmetry. By calculating the KL divergence values ​​of the real-time evaluation distribution and the sample evaluation distribution, the degree of shift in the evaluation distribution can be determined. The KL divergence value is specifically expressed as:

[0229]

[0230] Among them, D KL (P||Q) represents the KL divergence value between the real-time evaluation distribution P and the sample evaluation distribution Q, where P(x) is the probability distribution of the real-time evaluation distribution in the real-time evaluation field x, and Q(x) is the probability distribution of the sample evaluation distribution in the evaluation field x.

[0231] The above steps enable the updating of the power material sample database. By detecting the offset of data distribution, the power material sample database is triggered to update and supplement the samples, ensuring the timeliness and representativeness of the sample database.

[0232] Reference Figure 8 This invention provides a device for dynamically constructing bidding samples for power materials, comprising:

[0233] Data acquisition module 201 is used to acquire historical original bid evaluation documents;

[0234] The file processing module 202 is used to convert the format of historical original bid evaluation documents to obtain a structured initial bid evaluation sample pool.

[0235] The sample analysis module 203 is used to analyze the entropy value of each initial bid evaluation sample in the initial bid evaluation sample pool, and to filter the initial bid evaluation samples according to the entropy value to obtain the bid evaluation sample set;

[0236] The standard annotation module 204 is used to standardize the initial evaluation samples in the evaluation sample set, and to perform mind chain analysis on the standard-annotated initial evaluation samples in combination with the pre-built mind chain model to obtain standard evaluation triples.

[0237] The sample library generation module 205 is used to evaluate and screen the standard bid evaluation triplet according to the preset evaluation reward model, and generate a sample library for power material bid evaluation.

[0238] The sample library update module 206 is used to update the power material bid evaluation sample library in response to real-time original bid evaluation documents, and to complete the dynamic construction of the power material bid evaluation sample library.

[0239] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the described module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0240] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and variations of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and variations.

Claims

1. A method for dynamically constructing an electric power material bid evaluation sample, characterized in that, The method comprises the following steps: acquire historical original bid evaluation files; convert the historical original bid evaluation files into structured initial bid evaluation sample pools; analyze the entropy values of each initial bid evaluation sample in the initial bid evaluation sample pool, and screen the initial bid evaluation samples according to the entropy values to obtain a bid evaluation sample set; standardly label the initial bid evaluation samples in the bid evaluation sample set, and analyze the standardly labeled initial bid evaluation samples by combining a pre-constructed thought chain model to obtain standard bid evaluation triples, specifically including: extracting bid evaluation fields and corresponding label information in the standardly labeled initial bid evaluation samples; inputting the bid evaluation fields and corresponding label information into the thought chain model to analyze the relationship between the bid evaluation fields and the label information, generating a reasoning chain, specifically including: based on a pre-constructed bid evaluation rule library, combining the bid evaluation fields and the corresponding label information to generate corresponding standard bid evaluation guide instructions; analyzing the relationship between the bid evaluation fields and the corresponding label information according to the standard bid evaluation guide instructions, and combining the pre-set bid evaluation standard to judge the bid evaluation fields and the corresponding label information, and giving analysis logic and analysis results; according to the output limit of the reasoning chain, the analysis logic and the analysis results are arranged to obtain the reasoning chain, wherein the output limit of the reasoning chain contains the analysis logic of multiple steps; fusing the bid evaluation fields, the label information and the reasoning chain to obtain the standard bid evaluation triples; evaluating and screening the standard bid evaluation triples according to a pre-set evaluation reward model, and generating a power material bid evaluation sample library; acquiring real-time original bid evaluation samples and extracting real-time bid evaluation fields; analyzing the distribution of the real-time bid evaluation fields and the bid evaluation fields in the power material bid evaluation sample library respectively to obtain real-time bid evaluation distribution and sample bid evaluation distribution; based on the distribution difference between the real-time bid evaluation distribution and the sample bid evaluation distribution, triggering the update of the power material bid evaluation sample library; analyzing and labeling the real-time bid evaluation fields and adding them to the power material bid evaluation sample library.

2. The method of claim 1, wherein the method further comprises: The historical original bid evaluation files are converted into structured initial bid evaluation sample pools, specifically including: extracting the text content of the historical original bid evaluation files to obtain historical original bid evaluation texts; performing data cleaning on the historical original bid evaluation texts to obtain first bid evaluation texts; extracting field contents in the first bid evaluation texts and performing format conversion to obtain standardized bid evaluation fields and corresponding field labels; analyzing the standardization degree of the first bid evaluation texts based on each bid evaluation field to determine structured initial bid evaluation samples.

3. The method of claim 1, wherein the method further comprises: The entropy values of each initial bid evaluation sample in the initial bid evaluation sample pool are analyzed, and the initial bid evaluation samples are screened according to the entropy values to obtain a bid evaluation sample set, specifically including: using a pre-set sample evaluation model to analyze and evaluate the initial bid evaluation samples to obtain evaluation probabilities of different evaluation categories of each initial bid evaluation sample; calculating the entropy values of each initial bid evaluation sample according to the evaluation probabilities of different evaluation categories of the initial bid evaluation samples; sorting the initial bid evaluation samples according to the entropy values of the initial bid evaluation samples, and screening out a pre-set number of initial bid evaluation samples as a bid evaluation sample set.

4. The method for dynamically constructing bidding samples for power materials as described in claim 3, characterized in that, The entropy values of each initial bid evaluation sample are calculated according to the evaluation probabilities of different evaluation categories of the initial bid evaluation samples, specifically including: The evaluation probability of different evaluation categories is substituted into a preset logarithmic function to obtain a first evaluation parameter; The first evaluation parameter and the evaluation probability of the same evaluation category are fused to obtain a second evaluation parameter; Based on each initial evaluation sample, the second evaluation parameter is weighted and fused based on the evaluation coefficients of each evaluation category to obtain an entropy value of the initial evaluation sample.

5. The method of claim 1, wherein the method further comprises: Wherein, The evaluation rules in the evaluation rule library include evaluation scenes, rule application times, and evaluation standard priorities, and the labeling information includes labeling scenes, labeling times, and labeling standards; Based on the pre-constructed evaluation rule library, the corresponding standard evaluation guide instructions are generated in combination with the evaluation fields and the corresponding labeling information, specifically including: The similarity of the evaluation scenes and the labeling scenes in the evaluation rules and the matching degree of the rule application times and the labeling times are analyzed respectively, and the scene similarity and the time matching degree are given; Based on the scene similarity and the time matching degree, the evaluation standard priority corresponding to the labeling standard is fused to obtain the rule matching degree of each evaluation rule in the evaluation rule library; According to the rule matching degree of each evaluation rule in the evaluation rule library, the evaluation rules are screened to obtain the target evaluation rule and generate the corresponding standard evaluation guide instructions in combination with the instruction generation model, specifically including: the target evaluation rule is analyzed from the aspects of compliance, technology, and benefit to give the feature vector corresponding to the target evaluation rule; the evaluation scene of the target evaluation rule is vectorized to obtain the evaluation scene feature; the evaluation scene feature and the feature vector corresponding to the target evaluation rule are spliced to give the rule scene fusion vector in combination with the scene feature weight corresponding to the evaluation scene feature; the rule scene fusion vector is used to construct an instruction logic chain to give the standard evaluation guide instruction, wherein the instruction generation model includes a compliance layer, a technology layer, and a benefit layer.

6. The method of claim 1, wherein the method further comprises: The construction of the thought chain model is determined by the following steps: According to the preset evaluation standard, the model evaluation guide instruction is obtained in combination with the historical evaluation data; Based on the pre-trained language model, the evaluation logic corresponding to the historical evaluation data is obtained by logically analyzing the model evaluation guide instruction; The historical evaluation data is judged in combination with the evaluation logic, and the evaluation standard result is generated in combination with the output restriction in the model evaluation guide instruction; Based on the accuracy of the evaluation standard result, the pre-trained language model is iteratively trained until convergence to obtain the thought chain model.

7. The method of claim 1, wherein the method further comprises: receiving a request for a new sample of the power asset; and generating the new sample of the power asset based on the received request. According to the preset evaluation reward model, the standard evaluation triplets are evaluated and screened to generate the power material evaluation sample library, specifically including: The integrity, accuracy, and standardization of the standard evaluation triplets are analyzed and evaluated respectively to obtain integrity scores, accuracy scores, and standardization scores; The integrity scores, accuracy scores, standardization scores, and corresponding score weights are fused to obtain a comprehensive evaluation; Based on the comprehensive evaluation, the standard evaluation triplets are sorted and screened in combination with the preset proportion to obtain the power material evaluation sample library.

8. A power material evaluation sample dynamic construction device, characterized in that, The power material evaluation sample dynamic construction method according to any one of claims 1-7, comprising: A data acquisition module for acquiring historical original evaluation files; The file processing module is configured to convert a historical original bid evaluation file into a structured initial bid evaluation sample pool. The sample analysis module is configured to analyze an entropy value of each initial bid evaluation sample in the initial bid evaluation sample pool, and screen the initial bid evaluation sample according to the entropy value to obtain a bid evaluation sample set. The standard annotation module is configured to perform standard annotation on the initial bid evaluation sample in the bid evaluation sample set, and perform thought chain analysis on the standard-annotated initial bid evaluation sample in combination with a pre-constructed thought chain model to obtain a standard bid evaluation triple, specifically including: extracting a bid evaluation field and corresponding annotation information in the standard-annotated initial bid evaluation sample; inputting the bid evaluation field and the corresponding annotation information into the thought chain model to analyze a relationship between the bid evaluation field and the annotation information, and generating a reasoning chain, specifically including: generating a corresponding standard bid evaluation guide instruction based on a pre-constructed bid evaluation rule library in combination with the bid evaluation field and the corresponding annotation information; analyzing the relationship between the bid evaluation field and the annotation information according to the standard bid evaluation guide instruction, and judging the bid evaluation field and the annotation information in combination with a pre-set bid evaluation standard to give an analysis logic and an analysis result; arranging the analysis logic and the analysis result according to an output limit of the reasoning chain to obtain the reasoning chain, wherein the output limit of the reasoning chain contains analysis logics of multiple steps; and fusing the bid evaluation field, the annotation information and the reasoning chain to obtain the standard bid evaluation triple. The sample library generation module is configured to evaluate and screen the standard bid evaluation triple according to a pre-set evaluation reward model, and generate a power material bid evaluation sample library. The sample library update module is configured to obtain a real-time original bid evaluation sample, and extract a real-time bid evaluation field; analyze a distribution of the real-time bid evaluation field and a distribution of the bid evaluation field in the power material bid evaluation sample library to obtain a real-time bid evaluation distribution and a sample bid evaluation distribution; trigger an update of the power material bid evaluation sample library based on a distribution difference between the real-time bid evaluation distribution and the sample bid evaluation distribution; and analyze and annotate the real-time bid evaluation field to add the real-time bid evaluation field to the power material bid evaluation sample library.

Images