Electric power material thinking chain soft label generation method and device

Abstract

The invention discloses an electric power material thinking chain soft label generation method and apparatus. The method comprises the steps of obtaining electric power material historical review data; based on a pre-constructed multi-modal neural network model, carrying out feature extraction and analysis on the historical evaluation data of the electric power materials, and giving a first thinking chain; verifying and correcting the first thinking chain in combination with an electric power material review inference rule base to obtain a second thinking chain; fusing the Bayesian network and the conditional random field, quantifying the dependency relationship among the steps in the second thinking chain of the to-be-reviewed data of the electric power materials to obtain the basic similarity, dependency coefficient and uncertainty coefficient of each step, determining the weight by combining with a preset label, determining the soft label of each step, and determining the soft label of each step; and generating an electric power material thinking chain soft label. According to the method, the thinking chain and the soft label generation are improved, the review opinions are quantified, and the corresponding soft labels are given, so that the review process is traceable, and the review effect is ensured to a certain extent.

Description

Technical Field

[0001] The present invention belongs to the technical field of electric power material data analysis, and specifically relates to a method and device for generating soft labels for an electric power material thinking chain. Background Art

[0002] Power supplies are all technology-intensive products, involving complex technical fields such as high voltage, insulation, thermal stability, and electromagnetic compatibility. Their quality standards are far higher than those of general industrial supplies. The procurement of power supplies requires strict control of the review process under highly unified technical specifications. Currently, the actual bid evaluation process in the power supply sector suffers from the problem of "complex evaluation standards and inconsistent application of veto clauses." Due to differences in document structure and inconsistent content expression in the bidding documents corresponding to different batches and bidding sections, bid evaluation experts often rely on personal experience in their judgments, which can easily lead to situations such as unclear references to veto clauses and misjudgments of applicable conditions. In addition, the evaluation opinions given by the evaluation experts during the review process are freely written, with inconsistent language styles, which increases the difficulty of reviewing the evaluation opinions.

[0003] Patent application CN120219055A discloses an intelligent assisted bid evaluation method and system for electric power bidding, which includes: obtaining uploaded procurement information, and creating procurement batches based on the configured basic parameters and procurement review rules to generate a review task list with a unique batch number. Based on the review task list and the configured basic parameters, the context information of each review task is extracted to construct a context feature vector for the review task. The context feature vector is matched with the preset rejection rule to obtain a rejection suggestion list corresponding to each review task in the review task list. The rejection suggestion list is bound to the corresponding bidding task, and the bidding task bound to the rejection suggestion list is pushed to the user terminal. The user's operation on each rejection suggestion is received through the user terminal to generate a final rejection opinion, and the final rejection opinion is stored and archived, thereby improving the efficiency of generating review opinions, standardizing rejection expressions, and reducing the workload of experts.

[0004] As in the above-mentioned related prior art, the evaluation results obtained by comparing a single bidding document with the corresponding bidding rules do not comprehensively consider the evaluation opinions of experts, making the evaluation results inaccurate.

[0005] How to quantify the experts' review opinions and make the review process of power materials explicit, structured, and traceable to ensure the review effect and improve the review efficiency is a problem that needs to be solved at present. Summary of the Invention

[0006] To address the deficiencies in the above-mentioned prior art, the present invention provides a method and device for generating soft labels for a thought chain for electric power materials. The method comprises: obtaining historical review data for electric power materials; extracting and analyzing features of the historical review data for electric power materials based on a pre-built multimodal neural network model to generate a first thought chain; verifying and correcting the first thought chain in combination with a reasoning rule base for electric power material review to generate a second thought chain; integrating a Bayesian network and a conditional random field to quantify the dependencies between the steps in the second thought chain for the electric power material to be reviewed data, thereby obtaining a basic similarity, a dependency coefficient, and an uncertainty coefficient for each step; determining weights based on preset labels, integrating the basic similarity, the uncertainty coefficient, and the dependency coefficient to determine soft labels for each step in the second thought chain for the electric power material to be reviewed data, and generating soft labels for the thought chain for electric power materials. By improving the dynamic thought chain and soft label generation, the method is adapted to the scenario of electric power material bidding and review, transforming the "implicit thinking" of expert review into "explicit reasoning steps," quantifying the expert review opinions, and providing soft labels for the corresponding steps. This makes the review process explicit, structured, and traceable, thereby ensuring the review effect and improving the review efficiency during the bidding process.

[0007] In a first aspect, the present invention provides a method for generating a soft label for a power material thinking chain, which specifically comprises the following steps: Obtain historical review data of power materials; Based on the pre-built multimodal neural network model, the historical review data of power materials is extracted and analyzed to provide the first chain of thought; Combined with the power material review reasoning rule base, the first thinking chain is verified and revised to obtain the second thinking chain; By integrating Bayesian networks and conditional random fields, the dependencies between the various steps in the second thought chain of the power material review data were quantified, and the basic similarity, dependency coefficient, and uncertainty coefficient of each step were obtained. The weights are determined according to the preset labels, and the basic similarity, uncertainty coefficient and dependency coefficient are integrated to determine the soft labels of each step in the second thinking chain of the power material data to be reviewed, and generate the soft labels of the power material thinking chain.

[0008] Furthermore, the multimodal neural network model includes multimodal input channels, a shared encoder, a fusion layer, and a decoder; Based on the pre-built multimodal neural network model, the historical review data of power materials is extracted and analyzed to give the first chain of thought, which is obtained through the following steps: In the multi-modal input channel, the historical review data of electric power materials of corresponding modes are respectively subjected to tensor conversion processing to form multiple groups of tensors; The shared encoder extracts the semantic vectors of each group of tensors and transmits the semantic vectors to the fusion layer; After the semantic vectors of each group are concatenated, they are passed through multiple layers of cross attention units to form fusion features; Based on the fusion features, the fusion features are decoded in the decoder to obtain the natural language steps and target phrase positioning, and the first thinking chain is given.

[0009] Furthermore, based on the fusion features, the fusion features are decoded in the decoder to obtain the natural language steps and target phrase positioning, and the first thought chain is given, which specifically includes: Perform linear transformation on the fused features to give the initial hidden state of the decoder; Using the initial hidden state as the starting state for decoder phrase sequence generation, generate the initial phrase and update the initial hidden state; Generate the current phrase based on the current hidden state and the previously generated phrase, and update the current hidden state until a phrase sequence containing all phrases is generated; Based on the regular expression formed by the electric power material corpus, the phrase sequence is matched and analyzed, the target phrase is given, the natural language steps and the target phrase positioning are obtained, and the first chain of thought is given.

[0010] Furthermore, the power material review inference rule base includes at least one of a parameter and specification association rule, a step sequence rule, a threshold judgment rule, and a step dependency rule; Combined with the power material review reasoning rule base, the first chain of thought is verified and revised to obtain the second chain of thought, which specifically includes: Verify each step in the first thinking chain according to the rules in the power material review reasoning rule library. If the current step does not comply with the current rules, obtain a step similar to the current step and replace it to complete the correction of the current step. When all steps in the first chain of thought are verified, the second chain of thought is obtained.

[0011] Furthermore, the Bayesian network and conditional random field are integrated to quantify the dependencies between the various steps in the second thought chain of the power material review data, and the basic similarity, dependency coefficient, and uncertainty coefficient of each step are obtained, including: Analyze the similarity between each step in the second thinking chain of the power materials to be reviewed and the corresponding expert result vector, and provide the basic similarity of each step; Based on the Bayesian network model, the dependency relationship between each step in the second thinking chain of the power materials to be reviewed data is analyzed, and the uncertainty coefficient corresponding to each step is given; Each step in the second thinking chain of the data on power materials to be reviewed is input into a pre-constructed conditional random field to obtain the transition probability from the previous step to the current step and determine the dependency coefficient of each step.

[0012] Furthermore, the similarity between each step in the second thought chain of the power materials to be reviewed and the corresponding expert result vector is analyzed, and the basic similarity of each step is given, including: Quantify each step in the second thinking chain of the power material data to be reviewed to obtain the model result vector; Searching for an expert result vector corresponding to the model result vector from an expert intermediate result library to provide at least one target result vector, wherein the expert intermediate result library is obtained by analyzing and quantifying multi-modal historical review data of electric power materials; Analyze the similarity between the model result vector and the target result vector, and select the one with the highest similarity as the basic similarity.

[0013] Furthermore, the expert intermediate result library is obtained through the following steps: Obtain historical bidding documents and technical specifications for power materials; Perform entity recognition on historical bidding documents for electric power materials to obtain structured document entity parameters; The document entity parameters are associated with the specification clauses in the electric power material technical specification to obtain the electric power material specification knowledge graph, where the nodes of the electric power material specification knowledge graph are specification clauses or document entity parameters, and the edges of the electric power material specification knowledge graph are the associations between specification clauses and document entity parameters; According to the modal type of the historical review data of electric power materials, feature extraction is performed on the historical review data of electric power materials of each modality to obtain modal features corresponding to multiple modalities. The historical review data of electric power materials includes document entity parameters, specification clauses in technical specifications, and review texts. The modal features corresponding to multiple modes are weighted and spliced, and the knowledge graph of power material specifications is integrated to obtain the expert result vector and form an expert intermediate result library.

[0014] Furthermore, based on the Bayesian network model, the dependency relationships of each step in the second thinking chain of the power material review data are analyzed, and the uncertainty coefficients corresponding to each step are given, including: Build a Bayesian network model, initialize model parameters and define prior distribution; Input each step into the Bayesian network model, combine the dependency relationship of each step and the prior distribution, obtain the model prediction results and give the posterior distribution of each step; Sampling from the posterior distribution and calculating the variance of the model predictions in the sample set; The variance of the model prediction results in the sample set is normalized to obtain the uncertainty coefficient.

[0015] Furthermore, each step in the second thought chain of the power material review data is input into the pre-constructed conditional random field to obtain the transition probability from the previous step to the current step and determine the dependency coefficient of each step, including: Input each step in the second thought chain of the power material review data into the pre-built conditional random field, obtain the transition probability from the previous step to the current step and determine it as the dependency coefficient of the current step; The construction of the conditional random field specifically includes: Define observation sequence and implicit state sequence, where the observation sequence includes each step and the implicit state sequence includes the dependency relationship between each step; Based on the conditional probability of the conditional random field, set the characteristic function; Combined with multimodal historical review data of electric power materials, the conditional random field is trained and the value of the characteristic function is extracted until convergence.

[0016] Further, the tag determination weight includes a first weight, a second weight, and a third weight; Determine the weights based on the preset labels, integrate the basic similarity, uncertainty coefficient, and dependency coefficient, determine the soft labels for each step in the second thought chain of the power material data to be reviewed, and generate the soft labels for the power material thought chain, specifically including: Calculate the product of the basic similarity of the current step and the corresponding first weight to obtain the first parameter; Calculate the product of the uncertainty coefficient and the corresponding second weight to obtain the second parameter; Calculate the product of the dependency coefficient of the current step and the corresponding third weight to obtain the third parameter; The first parameter, the second parameter and the third parameter of the current step are summed to obtain the soft label of the current step in the second thinking chain of the electric power material to be reviewed data, and generate the soft label of the electric power material thinking chain.

[0017] Furthermore, the multimodal neural network model is obtained through the following steps: Perturb the data in the historical review data of electric power materials to generate adversarial samples; The adversarial samples and the historical review data of electric power materials are combined to obtain the adversarial training set; The initial network model is trained adversarially using the adversarial training set, and the comprehensive loss function is analyzed. The comprehensive loss function includes the adversarial loss function and the original loss function. The adversarial loss function is the loss generated by the adversarial sample during the training process, and the original loss function is the loss generated by the historical review data of electric power materials during the training process. The initial network model is trained using the historical review data of electric power materials. If the comprehensive loss function converges or reaches the maximum number of iterations, a multimodal neural network model is given.

[0018] In a second aspect, the present invention further provides a device for generating a soft label for a power material thinking chain, which adopts any of the above-mentioned methods for generating a soft label for a power material thinking chain, comprising: Data acquisition module, used to obtain historical review data of power materials; The thought chain generation module is used to extract and analyze the features of the historical review data of power materials based on the pre-built multimodal neural network model to generate the first thought chain; The thinking chain correction module is used to verify and correct the first thinking chain in combination with the power material review reasoning rule library to obtain the second thinking chain; The step analysis module is used to integrate Bayesian networks and conditional random fields to quantify the dependencies between the various steps in the second thought chain of the power material review data, and obtain the basic similarity, dependency coefficient, and uncertainty coefficient of each step; The label generation module is used to determine the weight according to the preset label, integrate the basic similarity, uncertainty coefficient and dependency coefficient, determine the soft label of each step in the second thinking chain of the power material to be reviewed data, and generate the soft label of the power material thinking chain.

[0019] The present invention provides a method and device for generating soft labels for electric power material thinking chain, which has at least the following beneficial effects: (1) By using a multimodal neural network model to extract and analyze the features of the historical review data of power materials, the first thought chain is given, the expert review opinions are quantified, and the "implicit thinking" of the expert review is transformed into "explicit reasoning steps"; then the Bayesian network and conditional random field are integrated to quantify the dependency between the various steps in the second thought chain of the power material data to be reviewed, evaluate the accuracy of each step from different aspects, determine the weight based on the preset label, and determine the soft label of each step in the second thought chain of the power material data to be reviewed. By improving the dynamic thought chain and soft label generation, it is adapted to the power material bidding review scenario, making the review process traceable and ensuring the review effect in the bidding process.

[0020] (2) By using a multimodal neural network model to process the historical review data of power materials in various modes and extract features in different modes, multimodal data provides more comprehensive review information, which helps to improve the accuracy of the review results; the features of different modes are fused, and the first thinking chain is constructed using the fused multimodal features. The integration of features of different modes can reduce the deviation that may be caused by a single mode and enhance the robustness of the model. At the same time, the first thinking chain can explicitly display the review logic, making the review logic traceable while improving the accuracy and reliability of the review.

[0021] (3) By using conditional random fields, the dependencies between steps are explicitly simulated and the rationality and coherence of these steps are evaluated, thereby improving the coherence between the steps. The Bayesian network is used to evaluate the uncertainty and causal relationship in each step, and probabilistic reasoning results are obtained, which helps to understand the causal logic of the reasoning steps. At the same time, the structure and parameters of the Bayesian network provide the interpretability of the model, which helps to understand the rationality between the steps. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A flow chart of a method for generating a soft label for a power material thinking chain according to an embodiment of the present invention; Figure 2 A flowchart for determining a first chain of thought provided by an embodiment of the present invention; Figure 3 An architectural diagram of a multimodal neural network model provided by an embodiment of the present invention; Figure 4 A flowchart of the dependency relationship between the quantified steps provided in an embodiment of the present invention; Figure 5 A flowchart of determining soft labels for each step provided by an embodiment of the present invention; Figure 6 This is a structural block diagram of the power material thinking chain soft label generation device provided by an embodiment of the present invention.

[0023] Among them, 201 is a data acquisition module; 202 is a thinking chain generation module; 203 is a thinking chain correction module; 204 is a step analysis module; 205 is a label generation module. DETAILED DESCRIPTION

[0024] To better understand the above technical solution, the following will be described in detail with reference to the accompanying drawings and specific implementation methods. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] The terms used in the embodiments of the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The singular forms "a," "an," "the," and "the" used in the embodiments of the present invention and the appended claims are also intended to include plural forms, and unless the context clearly indicates otherwise, "a plurality" generally includes at least two.

[0026] It should also be noted that the terms "include," "comprises," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a product or device comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such product or device. In the absence of further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the product or device comprising the element.

[0027] Tender documents are issued by the tenderee to potential bidders. They provide them with the necessary information for preparing their bids and outline the tender requirements. They are the core document of the entire tender process, stipulating key details such as the project's technical requirements, commercial terms, bid evaluation criteria, and contract terms.

[0028] From a technical perspective, the technical specifications and requirements in the bidding documents must be reasonable. For example, when procuring a complex set of industrial equipment, the technical parameters should accurately reflect the bidder's requirements for equipment performance, compatibility, and safety. Excessively high or low technical requirements can lead to a failed bid. Too high may mean no suitable supplier can meet the requirements, while too low may result in the bidder purchasing a product that does not meet actual usage needs.

[0029] The review panel is composed of the tenderee's internal professionals (such as technical, business, and legal experts). Drawing on their expertise and experience, they review each section of the tender documents. For example, technical experts will focus on whether the technical specifications comply with industry standards, business experts will examine whether the commercial terms are reasonable, and legal experts will review whether the tender documents comply with laws and regulations.

[0030] With the development of information technology, some bidding management software has incorporated rudimentary bidding document review capabilities. These software uses pre-set rules to check whether bidding documents contain key clauses (such as evaluation criteria and bid validity periods) and whether there are obvious typos or formatting issues. However, software review primarily serves a supporting role; complex technical and legal issues require further human judgment.

[0031] Different reviewers may have different understandings of the tender documents. For example, when judging the rationality of technical requirements, one technical expert may deem a certain parameter necessary, while another may consider it overly restrictive. This subjective difference can lead to inconsistent review results and affect the quality of the tender documents. Furthermore, tender documents involve knowledge from multiple fields, including technology, business, and law. If reviewers lack knowledge in a particular area, this can lead to review errors. For example, for tender projects in emerging technology fields, reviewers may have limited knowledge of the technical indicators and market conditions of the new technologies, making it impossible to accurately determine whether the technical requirements in the tender documents are reasonable.

[0032] Currently, quantitative standards for evaluating bidding documents are inadequate. For example, when assessing the reasonableness of commercial terms, it's difficult to use a fixed value to determine whether the bid security amount is appropriate. Factors such as varying project sizes and industry characteristics make it difficult to establish a standardized quantitative standard, forcing reviewers to rely solely on experience and subjective judgment.

[0033] The present invention provides a method and device for generating soft labels for a thought chain of electric power materials, the method comprising: obtaining historical review data of electric power materials; extracting and analyzing features of the historical review data of electric power materials based on a pre-built multimodal neural network model to provide a first thought chain; verifying and correcting the first thought chain in combination with a power material review inference rule base to obtain a second thought chain; fusing a Bayesian network and a conditional random field to quantify the dependency between each step in the second thought chain of the data to be reviewed for electric power materials, to obtain the basic similarity, dependency coefficient, and uncertainty coefficient of each step; determining weights based on preset labels, fusing the basic similarity, uncertainty coefficient, and dependency coefficient to determine the soft labels for each step in the second thought chain of the data to be reviewed for electric power materials, and generating soft labels for the thought chain of electric power materials. The present invention addresses the core pain points of "the implicitness of expert review logic and the difficulty in quantifying and tracing conclusions" in digital and intelligent review scenarios for bidding and procurement, and adapts to the bidding and review scenarios for electric power materials by improving dynamic thought chains and soft label generation. By converting the "implicit thinking" of expert reviewers into "explicit reasoning steps", quantifying the expert review opinions, and mapping the key nodes in the reasoning process into explainable soft labels, the review process is traceable, achieving the goal of "traceable review logic, quantifiable results, and assisted decision-making", and to a certain extent ensuring the review effect.

[0034] like Figure 1 As shown, the embodiment of the present invention provides a method for generating a soft label of a power material thinking chain, and the specific steps are as follows: S101: Obtain historical review data of power materials.

[0035] Among them, the historical review data of power materials refers to various multimodal data related to the review opinions during the bidding process, such as the voice review records of the review experts, expert discussion clips in video conferences, charts in the bidding documents (such as equipment parameter tables, engineering drawings, etc.), and pictures (such as equipment appearance pictures, installation site pictures, etc.).

[0036] S102: Based on the pre-built multimodal neural network model, feature extraction and analysis are performed on the historical review data of electric power materials to provide the first chain of thought.

[0037] Furthermore, the multimodal neural network model includes multimodal input channels, a shared encoder, a fusion layer, and a decoder; Based on the pre-built multimodal neural network model, the feature extraction and analysis of the historical review data of power materials are carried out to give the first chain of thought, referring to Figure 2 , specifically including: In the multi-modal input channel, the historical review data of electric power materials of corresponding modes are respectively subjected to tensor conversion processing to form multiple groups of tensors; The shared encoder extracts the semantic vectors of each group of tensors and transmits the semantic vectors to the fusion layer; After the semantic vectors of each group are concatenated, they are passed through multiple layers of cross attention units to form fusion features; Based on the fusion features, the fusion features are decoded in the decoder to obtain the natural language steps and target phrase positioning, and the first thinking chain is given.

[0038] The architecture of the above multimodal neural network model is as follows Figure 3 As shown, the historical review data of electric power materials of multiple modalities are input into the multimodal input channel, and the historical review data of electric power materials of each modality are respectively transformed into tensors. The transformed results enter the shared encoder, and the corresponding semantic vectors are extracted. The multiple semantic vectors are spliced ​​and fused in the fusion layer to obtain the fusion features. Finally, the fusion features are decoded by the decoder to obtain the corresponding first thought chain.

[0039] In a specific embodiment, historical review data for electric power materials, including modalities such as text, text table, image, and metadata, is input into a multimodal input channel. In one specific example, the historical review data for electric power materials in the text mode includes data in portable document format (PDF) / optical character recognition (OCR), such as technical specifications, bid letters, and commercial terms response forms. The historical review data for the text table mode includes data in comma-separated values ​​(CSV) / Excel format, such as structured quotation forms, warranty periods, and delivery dates. The historical review data for electric power materials in the image mode includes image data such as scanned copies of qualification certificates, photos of product nameplates, and factory production site maps. The historical review data for the metadata mode includes data in one-hot encoding (One-Hot Encoding) / embedded vector format, such as timestamps, voltage levels, material categories, and supplier numbers.

[0040] The shared encoder includes a text encoder, a table encoder, an image encoder, and a metadata encoder. The text encoder extracts sentence vectors or local n-tuples. In one specific example, the text encoder is a bidirectional encoder representation from transformers + bidirectional long short-term memory-convolutional neural network (BERT-base + Bi-LSTM-CNN). The table encoder converts tabular data into a format suitable for model input. In one specific example, the table encoder is a tabular neural network (TabNet) or a Transformer-Encoder. The image encoder removes fully connected layers (FC) and retains 2048-dimensional feature maps. In one specific example, the image encoder is a residual network with 50 layers (ResNet50). The metadata encoder processes and converts metadata. In one specific example, the metadata encoder is a two-layer multilayer perceptron (MLP).

[0041] The fusion layer includes Early-Fusion and Cross-Attention. Early-Fusion is used to concatenate vectors from multiple modalities obtained from the shared encoder. The Cross-Attention unit enables Transformer-Cross to query each modality and output a 512-dimensional fused feature.

[0042] The multimodal neural network model's task head generates multiple outputs, including a completeness classification, with results including complete, partially missing, and inconsistent binding. If the classification result is partially missing, the exact number of missing pages is provided. The multimodal neural network model's task head also generates a thought chain, which is decoded using a two-layer gated recurrent unit (GRU) to generate a natural language thought chain token sequence, the first thought chain.

[0043] In a specific example, the entire process of feature extraction, analysis, and generation is presented, taking the "first step in the first thought chain is file integrity check." First, multimodal data is input, including the text modal data of the "10kV Ring Main Unit Technical Specifications" obtained by scanning and OCR, the text table modal data of the 27 rows and 9 columns of the bid quotation form (including delivery date, warranty period, and unit price), the image modal data of three scanned copies of the manufacturer's qualification certificate, and the metadata modal data of "Material Category = Ring Main Unit, Voltage Level = 10kV, Bid Date = 2024-05-20."

[0044] Feature extraction is then performed on the data from each modality. For data in the text modality, BERT is used, outputting z_T∈R^768 [classification token (CLS)]. For data in the text table modality, TabNet is used, outputting z_Tab∈R^128, for example, highlighting the "Warranty Period" column with column attention weights. For data in the image modality, ResNet50 is used, outputting z_I∈R^2048. The final feature map is then subjected to global average pooling (GAP). For data in the metadata modality, MLP is used, outputting z_M∈R^64.

[0045] Next, the features obtained from each modality are fused and analyzed and input into the fusion layer, resulting in z_f = CrossAttn([z_T,z_Tab,z_I,z_M])∈R^512, where z_T is the semantic vector corresponding to the text modality, z_Tab is the semantic vector corresponding to the text table modality, z_I is the semantic vector corresponding to the image modality, and z_M is the semantic vector corresponding to the metadata modality. The softmax output of the integrity classifier is [0.87, 0.11, 0.02]. The probability of the classification result being complete is 0.87, the probability of the classification result being partially missing is 0.11, and the probability of the classification result being inconsistent with the binding is 0.02, indicating that the classification result is complete.

[0046] Finally, the first thought chain is generated and decoded using the Chain-Head mechanism. Input z_f and the start symbol <|start|>, and gradually decode the input fusion features to obtain: Step 1: "First"; Step 2: "Check"; Step 3: "Cover"; ...; the final sentence is: First, check the cover, table of contents, bid letter, technical specifications, and commercial terms response form, a total of five key items; then, check that pages 1-46 of the table of contents are consecutive and complete; then, check that the resolution of the three scanned copies of the qualification certificate is greater than 300 DPI; and comprehensively determine that the documents are complete and without any missing items. This final sentence is the first chain of thought automatically generated by the multimodal neural network model.

[0047] Furthermore, based on the fusion features, the fusion features are decoded in the decoder to obtain the natural language steps and target phrase positioning, and the first thought chain is given, which specifically includes: Perform linear transformation on the fused features to give the initial hidden state of the decoder; Using the initial hidden state as the starting state for decoder phrase sequence generation, generate the initial phrase and update the initial hidden state; Generate the current phrase based on the current hidden state and the previously generated phrase, and update the current hidden state until a phrase sequence containing all phrases is generated; Based on the regular expression formed by the electric power material corpus, the phrase sequence is matched and analyzed, the target phrase is given, the natural language steps and the target phrase positioning are obtained, and the first chain of thought is given.

[0048] The process of decoding the fusion features specifically includes: First, the fusion feature is linearly transformed to obtain the initial hidden state, that is, the fusion feature z_f is linearly mapped to obtain the initial hidden state , where φ() is the activation function, W_h is the weight matrix used to learn the relationship between the fusion features and the initial hidden state, and b_h is the bias vector. Then the joint vocabulary is constructed, the vocabulary V={common words}∪{positioning words}∪{<page_i> ,<col_j> ,<img_k> Next, GRU is used for step-by-step decoding, which is specifically expressed as follows:

[0049] Among them, h_t is the hidden state at the current time step t, x_t is the input feature at the current time step t, h_{t-1} is the hidden state at time step t-1, y_t is the output distribution at the current time step t, which is a probability distribution that represents the probability of each possible output. The softmax function is used to convert a vector into a probability distribution. W_out is the weight matrix of the output layer, which is used to map the hidden state to the output space. b_out is the bias vector of the output layer. The argmax function is used to find the index of the element with the highest probability in the probability distribution. x_{t+1} is the input feature at time step t+1. Regular extraction and positioning are then performed, using multiple regular expressions for extraction and positioning, including regular expressions for determining page numbers, column indexes, and image indexes. For example, regular expression 1:<page_(\d+)(-(\d+))?> →Page number; Regular 2:<col_(\d+)> → Column index; Regular 3:<img_(\d+)> →Image index. In other implementations, other regular expressions can be configured based on actual conditions, and this is not limited to this. Finally, natural language interpretation is used for splicing. The words are generated into corresponding word sequences, and different words are extracted and located. Finally, the extracted and located words are spliced ​​into "step + reason + location" triples. For example, the output: S = "Step description, because [reason], see [location] for details." In a specific example, first, the data of each modality needs to be fused to obtain the fused features, that is, the four types of information, namely text, table, picture, and metadata, are all compressed into a 512-dimensional vector z_f. Then, GRU is used to process the fused feature z_f to obtain the corresponding initial hidden state h0. In order to accurately insert the corresponding content in the chain of thought, multiple regular expressions are used to locate the page number <page_i>, column index <col_j>, and image index <img_k>. For example, <page_23> is "page 23", <col_3> is "column 3 of Excel", and <img_1> is "the first scanned image". Finally, characters are output sequentially using h0, and h0 is updated once for each character output. This loop continues until the end symbol is encountered, and the output result is obtained. After obtaining the output result, the regular expressions in the output result need to be translated completely. For example, when seeing <page_number>, it is translated into "page number", when seeing <col_number>, it is translated into "column number", and when seeing <img_number>, it is translated into "image number". The content corresponding to the regular expressions is highlighted on the displayed PDF or table. Finally, the translated output result is adjusted to make it a complete review opinion. For example, "Checked pages 1 - 22 of the technical specification, found that page 23 of the tender letter is missing, and the clarity of the first scanned certificate meets the requirements. Therefore, it is determined that the document is partially missing."

[0050] In other embodiments, further, the multi-modal neural network model is obtained through the following steps: Perturb the data part in the historical review data of power materials to generate adversarial samples; Fuse the adversarial samples and the historical review data of power materials to obtain an adversarial training set; Use the adversarial training set to perform adversarial training on the initial network model, and analyze the comprehensive loss function. Among them, the comprehensive loss function includes an adversarial loss function and an original loss function. The adversarial loss function is the loss generated by the adversarial samples during the training process, and the original loss function is the loss generated by the historical review data of power materials during the training process. The initial network model is trained through the historical review data of power materials; If the comprehensive loss function converges or reaches the maximum number of iterations, the multi-modal neural network model is given.

[0051] S103: Combine the power material review inference rule base to verify and correct the first chain of thought to obtain the second chain of thought.

[0052] Further, the power material review inference rule base includes at least one of a parameter and specification association rule, a step sequence rule, a threshold judgment rule, and a step dependency rule; Combined with the power material review reasoning rule library, the first chain of thought is verified and revised to obtain the second chain of thought, which specifically includes: Verify each step in the first thinking chain according to the rules in the power material review reasoning rule library. If the current step does not comply with the current rules, obtain a step similar to the current step and replace it to complete the correction of the current step. When all steps in the first chain of thought are verified, the second chain of thought is obtained.

[0053] The parameter and specification association rule indicates that a certain parameter or parameter type is associated with a certain specification clause. That is, when the parameter or parameter type appears, there must be a corresponding specification clause. It is expressed as follows using a logical expression:

[0054] The step sequence rule indicates the order relationship between one step and another. For example, the parameter extraction step A must precede the specification matching step B. It is expressed as a logical expression:

[0055] The threshold judgment rule determines whether a parameter value exceeds or falls below the corresponding threshold. If so, a conflict analysis is triggered. For example, if the parameter value V exceeds the corresponding threshold T, a conflict analysis is triggered. The logical expression is:

[0056] A step dependency rule indicates that a certain step must depend on one or more other steps. For example, the compliance determination step must be after the parameter extraction step and the specification matching step. It is expressed as a logical expression:

[0057] Traverse all the rules in the power material review reasoning rule library, judge the first thinking chain obtained by the multimodal neural network model, check whether the first thinking chain meets the rules, if so, determine the first thinking chain as the second thinking chain, if not, correct the first thinking chain and give the corrected second thinking chain, which meets all the rules in the power material review reasoning rule library.

[0058] In a specific example, there are a parameter extraction step and a compliance judgment step. The relationship between the two steps complies with the specification, but there is a missing step between the two steps. At this time, according to the step dependency rule, the specification matching step is inserted to obtain the first thought chain that complies with the specification, that is, parameter extraction step → specification matching step → compliance judgment step. If there is an error in the order of steps, the order between the steps is adjusted according to the step order rule. If there is a mismatch between the parameters and the specification clauses, the specification clauses corresponding to the parameters are queried from the knowledge graph of the power material specification and replaced. If the parameter value in a certain step exceeds or falls below the corresponding threshold, the parameter value is reread and re-judged according to the threshold judgment rule. If the judgment result still exceeds or falls below the corresponding threshold, a contradiction analysis is triggered, and the parameter value is marked or the parameter value or the corresponding threshold is manually adjusted.

[0059] Logic verification and correction are performed on each step of the generated first chain of thought according to the preset rules in the power material review inference rule library to ensure that the first chain of thought meets the logical requirements of the power material review. For example, the logic verification includes checking whether the step sequence is reasonable and whether the parameters match the specifications accurately. If any logical errors are found in the steps (such as matching parameters with irrelevant specifications), the chain of thought steps are corrected to ensure that they meet the logical requirements of the power material review.

[0060] S104: By integrating the Bayesian network and the conditional random field, the dependency relationship between each step in the second thinking chain of the power material review data is quantified to obtain the basic similarity, dependency coefficient, and uncertainty coefficient of each step.

[0061] Furthermore, the Bayesian network and conditional random field are integrated to quantify the dependency between each step in the second thinking chain of the power material review data, and the basic similarity, dependency coefficient and uncertainty coefficient of each step are obtained. Figure 4 , specifically including: Analyze the similarity between each step in the second thinking chain of the power materials to be reviewed and the corresponding expert result vector, and provide the basic similarity of each step; Based on the Bayesian network model, the dependency relationship between each step in the second thinking chain of the power materials to be reviewed data is analyzed, and the uncertainty coefficient corresponding to each step is given; Each step in the second thinking chain of the data on power materials to be reviewed is input into a pre-constructed conditional random field to obtain the transition probability from the previous step to the current step and determine the dependency coefficient of each step.

[0062] It can be understood that the data of electric power materials to be reviewed are relevant data that need to be reviewed, including bidding documents, technical specifications and other related content. According to the second thinking chain obtained from the above content, the data of electric power materials to be reviewed are brought into the corresponding multimodal neural network model and the output results are corrected. The various steps in the second thinking chain of the data of electric power materials to be reviewed can be obtained. Next, the data of electric power materials to be reviewed need to be reviewed and corresponding soft labels need to be given.

[0063] Furthermore, the similarity between each step in the second thought chain of the power materials to be reviewed and the corresponding expert result vector is analyzed, and the basic similarity of each step is given, including: Quantify each step in the second thinking chain of the power material data to be reviewed to obtain the model result vector; Searching for an expert result vector corresponding to the model result vector from an expert intermediate result library to provide at least one target result vector, wherein the expert intermediate result library is obtained by analyzing and quantifying multi-modal historical review data of electric power materials; Analyze the similarity between the model result vector and the target result vector, and select the one with the highest similarity as the basic similarity.

[0064] In a specific embodiment, the data on electric power materials to be reviewed are substituted into the obtained second thinking chain, and the thinking review data corresponding to the second thinking chain can be obtained. Each step in the thinking review data is quantified to obtain a model result vector. The above quantization process can be implemented through the multimodal input channel and shared encoder of the multimodal neural network model, or through a bidirectional encoder, so that the format of the model result vector and the target result vector are the same. For example, the model result vector contains the material type, and the target result vector with the same material type as the model result vector is searched from the expert intermediate result library. The target result vector can be one or more, and can be searched in the expert intermediate result library by material type, or can be searched from other aspects, without limitation. Then calculate the similarity between the model result vector and the target result vector. In this example, the cosine similarity is used to measure the similarity between the model result vector and the target result vector, which is specifically expressed as:

[0065] in, is the cosine similarity, is the model result vector, is the expert result vector, and n is the dimension of the model result vector and the target result vector.

[0066] For example, in the specification matching step, the parameter "rated voltage deviation (value +5.1%)" is matched with "Specification DL / T 5432-202153.2.1 (threshold ≤ ±5%)". The model result vector uses BERT to encode the above text and obtain a 300-dimensional vector P T300 The vector of "standard matching step" and material type "transformer" is retrieved from the expert intermediate result database, which is the expert result vector P Z300 (The expert result vector represents the expert's record of similar steps in the historical review data of power materials.) Calculate P T300 and P Z300 The basic similarity.

[0067] Furthermore, the expert intermediate result library is obtained through the following steps: Obtain historical bidding documents and technical specifications for power materials; Perform entity recognition on historical bidding documents for electric power materials to obtain structured document entity parameters; The document entity parameters are associated with the specification clauses in the electric power material technical specification to obtain the electric power material specification knowledge graph, where the nodes of the electric power material specification knowledge graph are specification clauses or document entity parameters, and the edges of the electric power material specification knowledge graph are the associations between specification clauses and document entity parameters; According to the modal type of the historical review data of electric power materials, feature extraction is performed on the historical review data of electric power materials of each modality to obtain modal features corresponding to multiple modalities. The historical review data of electric power materials includes document entity parameters, specification clauses in technical specifications, and review texts. The modal features corresponding to multiple modes are weighted and spliced, and the knowledge graph of power material specifications is integrated to obtain the expert result vector and form an expert intermediate result library.

[0068] In one specific implementation, the BERT-NER model is used to perform entity recognition on historical bidding documents, extracting entity parameters from the bid documents (such as "rated voltage deviation," "temperature rise test value," "delivery period," and "warranty period"). This results in structured document entity parameters, including the bid document number, material type, parameter type, parameter name, and parameter value. The BERT-NER model is pre-trained on a power-related corpus, including technical documentation for power equipment and power engineering standards and specifications, to learn the specific vocabulary, semantics, and named entity features of the power sector.

[0069] The extracted document entity parameters are associated with the specification clauses in the technical specifications to generate a knowledge graph for power material specifications. For example, "rated voltage deviation" is associated with Section 3.2.1 of DL / T 5432-2021 (requirement: ≤±5%), and "delivery period" is associated with Section 4.1.2 of the State Grid Distribution Engineering Procurement Specifications (requirement: ≤60 days). The nodes of the knowledge graph for power material specifications are specification clauses (attributes are clause numbers / thresholds) and entity parameters (attributes are parameter names). The edges of the knowledge graph for power material specifications represent the associations between specification clauses and document entity parameters. By analyzing the relationships between specification clauses and document entity parameters in the technical specifications, we establish association edges for the corresponding nodes (for example, "rated voltage deviation" is associated with "DL / T 5432-2021 §3.2.1").

[0070] The multimodal data of the historical review data of electric power materials are subjected to feature extraction, and finally feature fusion is performed to obtain the expert result vector. The specific feature extraction and feature fusion process is realized through the multimodal input channel and shared encoder of the multimodal neural network model, and can also be realized through a bidirectional encoder, which will not be elaborated here.

[0071] It's important to understand that historical review data for power supplies includes used regulatory clauses and various document entity parameters. Using modal features corresponding to multiple modalities, relevant parameters and / or regulatory clauses can be searched within the power supply regulatory knowledge graph. By constructing a knowledge graph for power supply regulatory knowledge, relationships between historical review data for power supplies can be established.

[0072] For example, modal features include file number, material type, parameter name, parameter type, and parameter value. Each feature dimension in the modal features is associated with a node in the power material specification knowledge graph. For example, when the parameter name is "rated voltage deviation," the power material specification knowledge graph can quickly retrieve the specification clause associated with "rated voltage deviation," DL / T 5432-2021 Section 3.2.1 (requirement: ≤±5%). Another example is when the parameter name is "delivery period," the power material specification knowledge graph can quickly retrieve the specification clause associated with "delivery period," State Grid Distribution Network Project Procurement Specification Section 4.1.2 (requirement: ≤60 days). For another example, when the material type is a transformer, the power material specification knowledge graph can retrieve the specification clauses or parameter names related to transformers.

[0073] Furthermore, based on the Bayesian network model, the dependency relationships of each step in the second thinking chain of the power material review data are analyzed, and the uncertainty coefficients corresponding to each step are given, including: Build a Bayesian network model, initialize model parameters and define prior distribution; Input each step into the Bayesian network model, combine the dependency relationship of each step and the prior distribution, obtain the model prediction results and give the posterior distribution of each step; Sampling from the posterior distribution and calculating the variance of the model predictions in the sample set; The variance of the model prediction results in the sample set is normalized to obtain the uncertainty coefficient.

[0074] In a specific embodiment, a Bayesian network is used to model the uncertainty of the thought chain steps (such as input ambiguity, model parameter uncertainty).

[0075] To build a Bayesian network, we first define each node in the model and initialize the model parameters. These parameters include the parameter value, confidence level, correlation level, weight matrix, rule threshold, and step correctness probability for each node. We then define a prior distribution and initialize the node distribution based on historical power material review data. For example, the model parameters follow a Gaussian distribution.

[0076] By inputting each step in the second thought chain of the power material review data into the constructed Bayesian network model, we can obtain the model prediction results and posterior distribution of each step. Monte Carlo sampling is performed from the posterior distribution, and the variance of the model prediction results is taken as the uncertainty coefficient of the current step, which is specifically expressed as:

[0077] Among them, U i is the uncertainty coefficient of the i-th step in the second thinking chain of the power material data to be reviewed, y k is the value sampled from the posterior distribution of the i-th step, K is the number of Monte Carlo sampling, and μ is the mean of the K sampled values.

[0078] A Bayesian network model is used to model the uncertainty of each step. The step inputs, model parameters, and outputs are used as nodes in the Bayesian network. The prior distribution of each step is determined based on prior knowledge (such as the model's performance on historical power material review data and expert reliability assessments of the step). After each thought chain step is executed, the posterior distribution of each node is updated using the Bayesian theorem based on the actual output results. The variance of each step's output is calculated using the belief propagation algorithm to quantify the uncertainty of the current step and obtain the uncertainty coefficient.

[0079] Furthermore, each step in the second thought chain of the power material review data is input into the pre-constructed conditional random field to obtain the transition probability from the previous step to the current step and determine the dependency coefficient of each step, including: Input each step in the second thought chain of the power material review data into the pre-built conditional random field, obtain the transition probability from the previous step to the current step and determine it as the dependency coefficient of the current step; The construction of the conditional random field specifically includes: Define observation sequence and implicit state sequence, where the observation sequence includes each step and the implicit state sequence includes the dependency relationship between each step; Based on the conditional probability of the conditional random field, set the characteristic function; Combined with multimodal historical review data of electric power materials, the conditional random field is trained and the value of the characteristic function is extracted until convergence.

[0080] In a specific embodiment, a conditional random field (CRF) is used to model the dependency between steps (eg, the previous step affects the next step).

[0081] A CRF is used to model the dependencies between steps in the second chain of thought for power material review data. The sequence of steps in the second chain of thought for power material review data is used as the observation sequence, and the dependencies between steps (such as the impact of the output of step i on the input of step i+1) is used as the implicit state sequence. The conditional random field is then trained using historical power material review data to learn the transition and state features between steps. This is combined with the characteristic function to extract the value of the characteristic function until convergence.

[0082] By training a CRF (using historical thought chain step sequences and the corresponding expert-annotated step dependencies as training samples), we learn the transition probabilities between steps (such as the transition probability from the parameter extraction step to the specification matching step), and then analyze the dependency patterns between steps. For example, we found that the compliance determination step often depends on the results of the parameter extraction step and the specification matching step.

[0083] The probability formula of CRF is specifically expressed as:

[0084] Among them, x is the observation sequence, y is the hidden state sequence, and f k is a characteristic function used to capture the state characteristics and transition characteristics between each step, λ k is the feature weight, and Z(x) is the normalization factor.

[0085] In a specific example, the step sequence is parameter extraction step → specification matching step → compliance determination step. The dependency relationship is that the parameter extraction step is the input of the specification matching step, and the output of the specification matching step is the input of the compliance determination step. Using the conditional random field to calculate the transition probability from parameter extraction step to specification matching step is 0.9, and the transition probability from specification matching step to compliance determination step is 0.8. Then, the dependency coefficient of the current step, i.e., the specification matching step, is the average of the forward transition probability (0.9) and the backward transition probability (0.8), i.e., the dependency coefficient is 0.85. In other examples, the product of the forward transition probability and the backward transition probability can also be taken, and there is no limitation on this.

[0086] S105: Determine the weight according to the preset label, integrate the basic similarity, uncertainty coefficient and dependency coefficient, determine the soft label of each step in the second thinking chain of the power material to be reviewed data, and generate the soft label of the power material thinking chain.

[0087] Further, the tag determination weight includes a first weight, a second weight, and a third weight; Determine the weight according to the preset label, integrate the basic similarity, uncertainty coefficient and dependency coefficient, determine the soft label of each step in the second thinking chain of the power material to be reviewed data, generate the soft label of the power material thinking chain, and refer to Figure 5 , specifically including: Calculate the product of the basic similarity of the current step and the corresponding first weight to obtain the first parameter; Calculate the product of the uncertainty coefficient and the corresponding second weight to obtain the second parameter; Calculate the product of the dependency coefficient of the current step and the corresponding third weight to obtain the third parameter; The first parameter, the second parameter and the third parameter of the current step are summed to obtain the soft label of the current step in the second thinking chain of the electric power material to be reviewed data, and generate the soft label of the electric power material thinking chain.

[0088] By comprehensively considering basic similarity, uncertainty coefficient, and dependency coefficient, and combining labels to determine weights, soft labels are calculated for each step in the second thought chain of the power material data to be reviewed. The soft label value reflects the degree of confidence in the correctness of the step. The soft label for each step is used to identify issues with the review comments. For example, a low soft label for the "Compliance Determination Step" indicates a problem with the results of that step, requiring review, or that the corresponding review comments were not considered.

[0089] softLabel of the current step i i , specifically expressed as:

[0090] Among them, α is the first weight, Si is the basic similarity of the current step i, β is the second weight, U i is the uncertainty coefficient of the current step i, γ is the third weight, T i is the dependency coefficient of the current step i.

[0091] The first weight, second weight and third weight can be set according to the actual application scenario and the degree of importance. For example, the first weight corresponding to the basic similarity is 0.4, the second weight corresponding to the uncertainty coefficient is 0.3, and the third weight corresponding to the dependency coefficient is 0.3.

[0092] Reference Figure 6 The embodiment of the present invention provides a device for generating soft labels for a power material thinking chain, including: Data acquisition module 201, used to obtain historical review data of power materials; The thought chain generation module 202 is used to extract and analyze the features of the historical review data of electric power materials based on the pre-built multimodal neural network model to generate a first thought chain; The thinking chain correction module 203 is used to verify and correct the first thinking chain in combination with the power material review reasoning rule library to obtain a second thinking chain; The step analysis module 204 is used to integrate the Bayesian network and the conditional random field to quantify the dependency between each step in the second thought chain of the power material to be reviewed data, and obtain the basic similarity, dependency coefficient, and uncertainty coefficient of each step; The label generation module 205 is used to determine the weight according to the preset label, integrate the basic similarity, uncertainty coefficient and dependency coefficient, determine the soft label of each step in the second thinking chain of the power material to be reviewed data, and generate the soft label of the power material thinking chain.

[0093] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0094] Although preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they are aware of the basic inventive concepts. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the invention. Thus, the present invention is intended to include such changes and modifications as fall within the scope of the claims and their equivalents.

Claims

1. A method for generating soft labels for electric power material thinking chain, characterized in that: include: Obtain historical review data of power materials; Based on the pre-built multimodal neural network model, the historical review data of power materials is extracted and analyzed to provide the first chain of thought; Combined with the power material review reasoning rule base, the first thinking chain is verified and revised to obtain the second thinking chain; By integrating Bayesian networks and conditional random fields, the dependencies between the various steps in the second thought chain of the power material review data were quantified, and the basic similarity, dependency coefficient, and uncertainty coefficient of each step were obtained. The weights are determined according to the preset labels, and the basic similarity, uncertainty coefficient and dependency coefficient are integrated to determine the soft labels of each step in the second thinking chain of the power material data to be reviewed, and generate the soft labels of the power material thinking chain.

2. The method for generating soft labels for electric power material thinking chain according to claim 1, characterized in that: The multimodal neural network model includes multimodal input channels, a shared encoder, a fusion layer, and a decoder; Based on the pre-built multimodal neural network model, the historical review data of power materials is extracted and analyzed to provide the first chain of thought, which includes: In the multi-modal input channel, the historical review data of electric power materials of corresponding modes are respectively subjected to tensor conversion processing to form multiple groups of tensors; The shared encoder extracts the semantic vectors of each group of tensors and transmits the semantic vectors to the fusion layer; After the semantic vectors of each group are concatenated, they are passed through multiple layers of cross attention units to form fusion features; Based on the fusion features, the fusion features are decoded in the decoder to obtain the natural language steps and target phrase positioning, and the first thinking chain is given.

3. The method for generating soft labels for electric power material thinking chain according to claim 2, characterized in that: Based on the fusion features, the fusion features are decoded in the decoder to obtain the natural language steps and target phrase positioning, and the first thought chain is given, which specifically includes: Perform linear transformation on the fused features to give the initial hidden state of the decoder; Using the initial hidden state as the starting state for decoder phrase sequence generation, generate the initial phrase and update the initial hidden state; Generate the current phrase based on the current hidden state and the previously generated phrase, and update the current hidden state until a phrase sequence containing all phrases is generated; Based on the regular expression formed by the electric power material corpus, the phrase sequence is matched and analyzed, the target phrase is given, the natural language steps and the target phrase positioning are obtained, and the first chain of thought is given.

4. The method for generating soft labels for electric power material thinking chain according to claim 1, characterized in that: The power material review reasoning rule base includes at least one of parameter and specification association rules, step sequence rules, threshold judgment rules and step dependency rules; Combined with the power material review reasoning rule library, the first chain of thought is verified and revised to obtain the second chain of thought, which specifically includes: Verify each step in the first thinking chain according to the rules in the power material review reasoning rule library. If the current step does not comply with the current rules, obtain a step similar to the current step and replace it to complete the correction of the current step. When all steps in the first chain of thought are verified, the second chain of thought is obtained.

5. The method for generating soft labels for electric power material thinking chain according to claim 1, characterized in that: By integrating Bayesian networks and conditional random fields, the dependencies between the various steps in the second thought chain of the power material review data are quantified, and the basic similarity, dependency coefficient, and uncertainty coefficient of each step are obtained, including: Analyze the similarity between each step in the second thinking chain of the power materials to be reviewed and the corresponding expert result vector, and provide the basic similarity of each step; Based on the Bayesian network model, the dependency relationship between each step in the second thinking chain of the power materials to be reviewed data is analyzed, and the uncertainty coefficient corresponding to each step is given; Each step in the second thinking chain of the data on power materials to be reviewed is input into a pre-constructed conditional random field to obtain the transition probability from the previous step to the current step and determine the dependency coefficient of each step.

6. The method for generating soft labels for electric power material thinking chain according to claim 5, characterized in that: Analyze the similarity between each step in the second thinking chain of the power materials to be reviewed and the corresponding expert result vector, and provide the basic similarity of each step, including: Quantify each step in the second thinking chain of the power material data to be reviewed to obtain the model result vector; Searching for an expert result vector corresponding to the model result vector from an expert intermediate result library to provide at least one target result vector, wherein the expert intermediate result library is obtained by analyzing and quantifying multi-modal historical review data of electric power materials; Analyze the similarity between the model result vector and the target result vector, and select the one with the highest similarity as the basic similarity.

7. The method for generating soft labels for electric power material thinking chain according to claim 6, characterized in that: The expert intermediate result library is obtained through the following steps: Obtain historical bidding documents and technical specifications for power materials; Perform entity recognition on historical bidding documents for electric power materials to obtain structured document entity parameters; The document entity parameters are associated with the specification clauses in the electric power material technical specification to obtain the electric power material specification knowledge graph, where the nodes of the electric power material specification knowledge graph are specification clauses or document entity parameters, and the edges of the electric power material specification knowledge graph are the associations between specification clauses and document entity parameters; According to the modal type of the historical review data of electric power materials, feature extraction is performed on the historical review data of electric power materials of each modality to obtain modal features corresponding to multiple modalities. The historical review data of electric power materials includes document entity parameters, specification clauses in technical specifications, and review texts. The modal features corresponding to multiple modes are weighted and spliced, and the knowledge graph of power material specifications is integrated to obtain the expert result vector and form an expert intermediate result library.

8. The method for generating soft labels for electric power material thinking chain according to claim 5, characterized in that: Based on the Bayesian network model, the dependency relationships of each step in the second thinking chain of the data on power materials to be reviewed are analyzed, and the uncertainty coefficients corresponding to each step are given, including: Build a Bayesian network model, initialize model parameters and define prior distribution; Input each step into the Bayesian network model, combine the dependency relationship of each step and the prior distribution, obtain the model prediction results and give the posterior distribution of each step; Sampling from the posterior distribution and calculating the variance of the model predictions in the sample set; The variance of the model prediction results in the sample set is normalized to obtain the uncertainty coefficient.

9. The method for generating soft labels for electric power material thinking chain according to claim 1, characterized in that: The tag determination weight includes a first weight, a second weight, and a third weight; Determine the weights based on the preset labels, integrate the basic similarity, uncertainty coefficient, and dependency coefficient, determine the soft labels for each step in the second thought chain of the power material data to be reviewed, and generate the soft labels for the power material thought chain, specifically including: Calculate the product of the basic similarity of the current step and the corresponding first weight to obtain the first parameter; Calculate the product of the uncertainty coefficient and the corresponding second weight to obtain the second parameter; Calculate the product of the dependency coefficient of the current step and the corresponding third weight to obtain the third parameter; The first parameter, the second parameter and the third parameter of the current step are summed to obtain the soft label of the current step in the second thinking chain of the electric power material to be reviewed data, and generate the soft label of the electric power material thinking chain.

10. A device for generating soft labels for electric power material thinking chain, characterized in that: The method for generating a soft label of a power material thinking chain according to any one of claims 1 to 9 comprises: Data acquisition module, used to obtain historical review data of power materials; The thought chain generation module is used to extract and analyze the features of the historical review data of power materials based on the pre-built multimodal neural network model to generate the first thought chain; The thinking chain correction module is used to verify and correct the first thinking chain in combination with the power material review reasoning rule library to obtain the second thinking chain; The step analysis module is used to integrate Bayesian networks and conditional random fields to quantify the dependencies between the various steps in the second thought chain of the power material review data, and obtain the basic similarity, dependency coefficient, and uncertainty coefficient of each step; The label generation module is used to determine the weight according to the preset label, integrate the basic similarity, uncertainty coefficient and dependency coefficient, determine the soft label of each step in the second thinking chain of the power material to be reviewed data, and generate the soft label of the power material thinking chain.

Images