Project decision and effect prediction method and device based on expert knowledge extraction and medium

Abstract

The invention provides a project decision and effect prediction method and device based on expert knowledge extraction and a medium, and belongs to the technical field of power grid enterprise achievement transformation management. The method comprises the following steps: receiving a feasibility research report text and expert review data, and extracting a core text through a keyword matching rule; carrying out word segmentation and segmentation on the extracted content by using a built-in word segmentation device of the pre-training language model, obtaining vector representation of each segment, and distributing a weight to each segment based on an attention mechanism; unifying the dimension of the expert review score in the expert review data; constructing a deep learning model based on pre-training BERT, and adopting a staged progressive training strategy; and collecting a feasibility research report text to be decided, inputting the trained deep learning model, and outputting a project approval decision. According to the method, through intelligent processing of the feasibility research report text and the expert review data of the incubation project, the structured review knowledge is extracted, the deep learning model is constructed, and the project optimization efficiency and consistency are remarkably improved.

Description

Technical Field

[0001] This invention relates to a method, device, and medium for project decision-making and performance prediction based on expert knowledge extraction, belonging to the field of power grid enterprise achievement transformation management technology. Background Technology

[0002] With the continuous and rapid development of technological innovation in the power industry, power grid companies, as the backbone of the energy sector, have consistently produced a large number of technological achievements with practical value and market potential during the research and development process. These achievements cover multiple key areas such as smart grid technology, new energy access and consumption, electricity market trading, and equipment monitoring and operation and maintenance, and are of great significance for improving grid operating efficiency, promoting the utilization of renewable energy, and ensuring energy security. my country is actively promoting the adjustment and optimization of its energy structure, vigorously developing clean energy and renewable energy to reduce dependence on fossil fuels and lower carbon emissions. At the same time, digital transformation has become an important direction for the development of power grid companies. By applying advanced technologies such as big data, cloud computing, and artificial intelligence, the power grid is becoming more intelligent, automated, and interactive, improving its flexibility and reliability. However, efficiently transforming these practically valuable technological achievements into actual productivity and widely promoting their market application is a crucial link in enhancing the core competitiveness of power grid companies and promoting high-quality development of the industry.

[0003] Currently, power grid companies primarily rely on the review of feasibility study reports for project selection in the management of technology transfer. Experts play a crucial role in this process, comprehensively evaluating projects from multiple dimensions such as material standardization, technical route rationality, team management capabilities, and market prospect forecasts. They then combine their personal experience and professional knowledge to provide scores, ultimately leading to project approval decisions. While expert review plays an indispensable role in project selection, existing technologies still face numerous shortcomings and challenges in practical application. Expert review is highly subjective and lacks consistency: Due to differences in professional backgrounds, research fields, and review experience among experts, the scoring standards and methods they use during the review process vary, leading to poor consistency in review results. This subjectivity not only affects the fairness and objectivity of the review but also increases the difficulty and uncertainty of project selection. In some cases, biases in review results may even arise due to expert personal preferences or conflicts of interest, thus hindering the successful transformation of high-quality projects.

[0004] Manual scoring is time-consuming and laborious, and it is difficult to meet the needs of large-scale reviews: The existing expert review method mainly relies on manual reading and scoring of materials, which is not only time-consuming but also laborious. With the continuous increase in the number of technical achievements of power grid enterprises and the expanding review requirements, the manual scoring method is obviously unable to meet the needs of quickly and efficiently reviewing a large number of projects. Especially when faced with urgent or large-scale project review tasks, the manual scoring method is even more powerless, seriously affecting the progress and efficiency of成果转化 (it seems there is a specific term here that might be misspelled or not fully clear, perhaps "achievement transformation").

[0005] There is a lack of an effective mechanism for sedimentation and reuse of expert review knowledge: During the expert review process, a large amount of review knowledge and experience have been accumulated, including information on project evaluation criteria, technological development trends, market prospect forecasts, etc. However, this knowledge mostly exists in an unstructured form and is difficult to be systematically sorted, stored, and reused. Each review requires re-accumulating knowledge, which not only increases the cost and uncertainty of the review but also leads to the waste of review experience and the breakage of inheritance. The lack of an effective knowledge management mechanism makes it difficult for power grid enterprises to form the ability of continuous improvement and optimization in the management of成果转化 (again, the specific term might need clarification). Summary of the Invention

[0006] The purpose of the present invention is to provide a project decision-making and effectiveness prediction method, device, and medium based on expert knowledge extraction. Through the intelligent processing of the feasibility study report text of incubation projects and expert review data, structured review knowledge is extracted, and a deep learning model is constructed to significantly improve the efficiency and consistency of project selection.

[0007] To achieve the above object, the present invention is realized through the following technical solutions: A project decision-making and effectiveness prediction method based on expert knowledge extraction includes: Receive the feasibility study report text and expert review data, and extract the core text through keyword matching rules; Use the built-in tokenizer of the pre-trained language model to tokenize and segment the extracted content, and obtain the vector representation of each segment. Based on the attention mechanism, assign weights to each segment; Unify the dimension of the expert review scores in the expert review data; Construct a deep learning model based on the pre-trained BERT, and adopt a phased progressive training strategy, including the scoring regression model training stage and the dynamic threshold classification model training stage; in the scoring regression model training stage, the BERT encoder learns and fits the quantitative criteria of expert reviews based on the regression prediction head; in the dynamic threshold classification model training stage, construct an independent classification prediction head to obtain a binary classifier for dynamic threshold decision-making; Collect the feasibility study report text to be decided, input it into the trained deep learning model, output the project approval decision, and predict the effectiveness of the approved projects.

[0008] Preferably, the step of extracting core text through keyword matching rules The specific method is as follows: , in, Indicates the sequential connection of text. This indicates a successfully matched text segment. This represents a matching function used to determine the meaning of a text segment. Does it contain keywords? , Indicates the first A text segment to be matched, Indicates keywords.

[0009] Preferably, the built-in tokenizer of the pre-trained language model uses WordPiece encoding to convert the core text into a token sequence and segments the long sequence. The specific method for obtaining the vector representation of each segment is as follows: , , in, For the first Vector representation of each segment, Represents the BERT model. This represents the input to the BERT model. This represents a vocabulary mapping function that converts text into its corresponding index in the vocabulary. Indicates classification label, Indicates the first A text fragment of input, Indicates a separator.

[0010] Preferably, the weight calculation method is as follows: , , in, For the first The weight of each segment, For the first Vector representation of each segment, Indicates the first Attention scores for each segment after processing Represents an exponential function. Indicates the first Attention scores for each segment after processing This represents the transpose of the trainable parameters. This represents a trainable weight matrix. Representing vector representation, This represents a trainable bias vector. This represents the hyperbolic tangent activation function. This indicates the total number of segments.

[0011] Preferably, during the training phase of the scoring regression model, a regression prediction head is constructed based on the BERT encoder, and the regression prediction head includes a fully connected layer and a ReLU activation function; Adversarial training is employed to improve the model's robustness by applying adversarial perturbations; its loss function... for: , , , in, This represents the mean squared error loss function. This represents the original input data of the model. Indicates a counter-perturbation. Indicates the true label of the regression task. Indicates the mean squared error loss relative to the input. gradient, Indicates the disturbance amplitude or step size. This represents the gradient operator.

[0012] Preferably, during the training phase of the dynamic threshold classification model, the output vector of the BERT encoder is concatenated with the score vector of the regression prediction to form an enhanced feature vector; the classification prediction head is composed of fully connected layers. The dynamic threshold classification model is trained using a differential learning rate strategy. A normal learning rate is set for the classification prediction head, while a low learning rate, which is 0.01-0.001 times the learning rate of the classification prediction head, is set for the optimized BERT encoder and regression prediction head.

[0013] Preferably, the loss function during the training phase of the dynamic threshold classification model is... as follows: , in, The weight hyperparameters representing the regression loss term. Indicates mean square error. Indicates the first The true regression label of each sample The model represents the first The predicted value for each sample, The weight hyperparameters representing the classification loss term. Represents the cross-entropy loss function. This represents the true classification label of the sample. This represents the classification probability predicted by the model.

[0014] Preferably, the classification prediction head adopts a dynamic threshold decision mechanism, and the final decision threshold is jointly determined by the initial threshold, the expert consensus calibration factor, and the project risk factor. , , , in, The final decision threshold, As the initial threshold, As an expert consensus calibration factor, As a project risk factor, This indicates the risk of budget overruns. The standard deviation of the expert rating.

[0015] The advantages of this invention are: By intelligently processing feasibility study report texts and expert review data, structured review knowledge is extracted, and an automated scoring and decision-making model is constructed. This significantly reduces the time spent on manual scoring and improves review efficiency. Simultaneously, the model employs a unified scoring standard, avoiding biases in review results caused by inconsistencies in scoring standards among different experts, thus enhancing the consistency of review results.

[0016] A phased, progressive training strategy, combined with adversarial training and differential learning rate strategies, ensures the model deeply understands the scoring task and improves its generalization ability. A dynamic threshold decision-making mechanism integrates expert consensus, project risk, and model confidence, making the decision boundary closer to real expert experience and enhancing the scientific rigor and robustness of the decision-making process.

[0017] The automated scoring model can quickly process large amounts of project data, meeting the review needs brought about by the increasing number of technology transfers by power grid companies. The dynamic threshold adaptive learning mechanism optimizes parameters based on historical decision results, further enhancing its applicability in large-scale review scenarios. By quantifying scoring standards and introducing a multi-dimensional calibration mechanism, the influence of expert subjective preferences on review results is reduced. Adjusting decision thresholds in conjunction with project risk factors makes the review of high-risk projects more rigorous, improving the objectivity of overall decision-making. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0019] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0020] Figure 2 This is a schematic diagram of the model structure of the present invention. Detailed Implementation

[0021] To provide a more detailed understanding of the features and technical content of the embodiments of this disclosure, the implementation of the embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for illustrative purposes only and are not intended to limit the embodiments of this disclosure. In the following technical description, for ease of explanation, several details are used to provide a full understanding of the disclosed embodiments. However, one or more embodiments may still be implemented without these details. In other cases, well-known structures and devices may be simplified in their depiction to simplify the drawings.

[0022] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this disclosure described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion.

[0023] Unless otherwise stated, the term "multiple" means two or more.

[0024] In this embodiment of the disclosure, the character " / " indicates that the objects before and after it are in an "or" relationship. For example, A / B means: A or B.

[0025] The term "and / or" describes an association between objects, indicating that three relationships can exist. For example, A and / or B means: A or B, or A and B.

[0026] The term "correspondence" can refer to an association or binding relationship. The correspondence between A and B means that there is an association or binding relationship between A and B.

[0027] Example 1 like Figure 1 As shown, a project decision-making and outcome prediction method based on expert knowledge extraction is proposed. By intelligently processing feasibility study report texts and expert review data, structured review knowledge is extracted and a deep learning model is constructed, thereby significantly improving the efficiency and consistency of project selection and providing strong support for the transformation and management of power grid enterprises' achievements.

[0028] Includes the following steps: S1: Receive the feasibility study report text and expert review data, and extract the core text through keyword matching rules; S2: Use the built-in word segmenter of the pre-trained language model to segment and segment the extracted content, obtain the vector representation of each segment, and assign weights to each segment based on the attention mechanism; S3: Standardize the dimensions of expert review scores in the expert review data; S4: Construct a deep learning model based on pre-trained BERT, and adopt a phased progressive training strategy, including a scoring regression model training phase and a dynamic threshold classification model training phase; the scoring regression model training phase enables the BERT encoder to learn and fit the quantitative criteria of expert review based on the regression prediction head; the dynamic threshold classification model training phase constructs an independent classification prediction head to obtain a binary classifier for dynamic threshold decision. S5: Collect the text of the feasibility study report to be decided, input the trained deep learning model, output the project approval decision, and predict the effectiveness of the approved project.

[0029] As a refinement of the above embodiment, step S1 receives the complete text of the feasibility study report and expert review data. To improve the effectiveness of features and filter redundant information, key content is extracted first. By configuring chapter title keyword (such as "technological innovation points", "market prospects", "team strength") matching rules, core text highly relevant to the expert review is automatically located and extracted from the feasibility study report.

[0030] The core text is extracted using keyword matching rules. The specific method is as follows: , in, Indicates the sequential connection of text. This indicates a successfully matched text segment. This represents a matching function used to determine the meaning of a text segment. Does it contain keywords? , Indicates the first A text segment or section of the original text to be matched. Indicates keywords.

[0031] As a refinement of the above embodiment, in step S2, the extracted core text is processed by calling the built-in word segmenter of the pre-trained language model (BERT) to perform WordPiece encoding and convert it into a token sequence.

[0032] To process extremely long texts and simulate the process of experts recognizing key information, fragments are obtained by truncating or segmenting the extremely long sequences.

[0033] Based on segmentation, a vector representation of each segment is obtained during the encoding stage.

[0034] Specifically, the method for obtaining the vector representation of each segment is as follows: , , in, For the first Vector representation of each segment, This refers to the BERT (Bidirectional Encoder Representations from Transformers) model, a pre-trained deep learning language model. This represents the input to the BERT model, which is the tokenized sequence of TokenIDs. This represents a vocabulary mapping function that converts text (Token) into its corresponding index (ID) in the vocabulary. This represents a classification tag, a special token added at the beginning of the sequence. Its corresponding output vector is usually used as the aggregate representation of the entire sequence. Indicates the first A text fragment of input, This represents a separator, a special token used to separate different sentences or segments.

[0035] As a refinement of the above embodiment, step S2 introduces an attention mechanism to focus on key information by assigning corresponding weights to different information fragments. This allows the model to focus on content that has a decisive impact on score prediction. The weights of the attention mechanism are calculated as follows: , , in, For the first The weight of each segment, For the first Vector representation of each segment, Indicates the first The attention scores of each segment after processing are used to calculate the final weights. Represents an exponential function. Indicates the first Attention scores for each segment after processing This represents the transpose of the trainable parameters. This represents a trainable weight matrix. Representing vector representation, This represents a trainable bias vector. This represents the hyperbolic tangent activation function. This indicates the total number of segments.

[0036] vector As a trainable parameter, it represents the aggregation of information that the model learns is most important for the rating, thus enabling the model to focus on the content that has a decisive impact on rating prediction.

[0037] As a refinement of the above embodiment, in step S3, the expert review scores, which serve as the regression target values, are normalized or standardized to unify the dimensions of different scoring dimensions.

[0038] As a refinement of the above embodiment, step S4 constructs a deep learning model based on pre-trained BERT to achieve intelligent analysis of the feasibility study report. The core of this model is a shared BERT encoder, responsible for extracting contextual features of the report text and uniformly using its [CLS] marker bit in the output vector. As a deep semantic representation of the entire report, to ensure that the model first deeply understands the core task of "scoring", the traditional multi-task parallel processing was abandoned, and instead a progressive training strategy of "regression first, classification later" was adopted.

[0039] Specifically, such as Figure 2 As shown, the entire model is divided into two stages: Phase 1: Training phase of the scoring regression model; Phase 2: Training phase of dynamic threshold classification model.

[0040] (1) The core task of the rating regression model training phase is to enable the BERT encoder to fully learn and fit the quantitative criteria of expert review. A regression prediction head is built on top of the shared BERT encoder. This module consists of fully connected layers and ReLU activation functions. To improve the generalization ability and robustness of the model in the fine-tuning phase, we introduce adversarial training in this phase. This technique applies small, gradient-based adversarial perturbations to the word embedding space.

[0041] Optimization objective during the training phase of the rating regression model for: , , , in, This represents the mean squared error loss function. This represents the original input data of the model. Indicates a counter-perturbation. Indicates the true label of the regression task. This represents the gradient of the mean squared error loss with respect to the input x. This indicates the amplitude or step size of the disturbance, controlling the "size" of the adversarial disturbance. This represents the gradient operator.

[0042] This stage forces the model to maintain stable output even when faced with minor changes in the input data (such as synonym substitutions or subtle differences in expression). The entire module receives... The vector is ultimately passed through an N-neuron output layer without an activation function to generate continuous scores corresponding to N rating dimensions.

[0043] (2) The goal of the dynamic threshold classification model stage is to efficiently train a binary classifier for dynamic threshold decision-making based on the first stage.

[0044] This stage will fully load the BERT encoder weights (and regression head weights) trained in Stage 1. At this point, the encoder possesses a deep understanding of report quality. First, a separate classification prediction head is constructed. The input for this stage will undergo crucial feature fusion: in one forward propagation, the output vector of the BERT encoder is concatenated with the regression prediction score vector to form an enhanced feature vector. The newly constructed classification prediction head (consisting of independent fully connected layers) will process based on this fused vector.

[0045] To prevent the classification training in Phase 2 (with different task objectives) from destroying the valuable scoring knowledge learned in Phase 1, we will adopt a differential learning rate strategy. That is, we will set a normal learning rate for the newly constructed classification head, while setting a very low learning rate for the optimized BERT encoder and regression head.

[0046] As a refinement of the above embodiments, the deep learning model loss function as follows: , in, The weight hyperparameters representing the regression loss term. Indicates mean square error. This represents the true regression label of the i-th sample. This represents the model's predicted value for the i-th sample. The weight hyperparameters representing the classification loss term. This represents the cross-entropy loss function. This represents the true classification label of the sample. This represents the classification probability predicted by the model.

[0047] During training, end-to-end fine-tuning is performed using optimizers such as Adam to minimize the loss across all tasks simultaneously. If certain rating dimensions are more important, the corresponding loss weights can be appropriately increased to improve prediction accuracy. Multi-task training enables the model to learn from each task while sharing text representations, thereby improving rating accuracy and decision robustness.

[0048] As a refinement of the above embodiments, in the binary classification output stage of the classification prediction head, a dynamic threshold decision mechanism is proposed to achieve a more scientific output that is closer to real expert experience. This mechanism aims to dynamically adjust the decision boundary by integrating the initial threshold, expert consensus calibration factor, and project risk factor.

[0049] (1) The mechanism sets a baseline initial threshold, which is defined as follows: The value is 0.8. This corresponds to an 80% expert support rate, meaning that the project is recommended for approval when the model predicts a project approval probability P≥0.8.

[0050] (2) In order to ensure that the threshold reflects the actual distribution of expert decisions, this mechanism introduces an "expert consensus calibration factor" through a multi-dimensional calibration module. This factor is used to quantify the variability of expert ratings (standard deviation of expert ratings). When the level of expert consensus is high ( ), decision-making should be more stringent, therefore the threshold should be tightened ( When experts have significant disagreements ( Decision-making should be more inclusive, therefore the threshold should be relaxed ( ).

[0051] Its formula is expressed as follows: , (3) To address the potential risks of different projects, integrate project risk factors. For high-risk projects (such as those with a budget overrun risk >30%), the threshold will be automatically increased by 0.05.

[0052] , In summary, the final dynamic decision threshold It is composed of the above factors: , in, The final decision threshold, As the initial threshold, As an expert consensus calibration factor, As a project risk factor, This indicates the risk of budget overruns. The standard deviation of the expert rating.

[0053] The threshold strategy employs adaptive learning optimization, establishing a dynamic threshold adjustment module that updates threshold parameters based on historical decision results. It integrates the collective wisdom of experts with model predictions to dynamically adapt to different project characteristics and risk levels.

[0054] As a refinement of the above embodiments, the specific method for predicting the effectiveness of the initiated projects is as follows: A hierarchical indicator system is constructed, evaluation criteria and subdivided indicators are set, and the weight of each indicator is determined using the Analytic Hierarchy Process (AHP). Combined with actual data, the overall benefits are calculated, and the transformation effectiveness is quantified and fed back, providing a basis for resource allocation and decision optimization during the transformation process. Specifically, this includes: (1) Constructing a hierarchical structure model Based on the overall evaluation objectives, the indicator system is divided into three levels.

[0055] Target layer (A): Evaluation of the benefits of results transformation.

[0056] Criteria Layer (B): Transformation Objectives (B1), Capital Investment (B2), Technology Maturation (B3), Transformation Efficiency (B4).

[0057] Indicator Layer (C): The third-level indicators under each criterion layer.

[0058] The hierarchical structure model established based on the benefit evaluation index system is shown in Table 1. The construction of the hierarchical structure follows two principles: first, there is no overlap between indicators at the same level to avoid overlapping evaluation dimensions; second, the lower-level indicators are complete for the upper-level indicators to ensure that the connotation of each criterion level is fully covered by the subordinate indicators.

[0059] Table 1 Evaluation Index System for the Benefits of Technology Transfer (2) Constructing a judgment matrix and expert scoring The judgment matrix is ​​established to compare indicators at the same level in the hierarchical structure that are governed by the level above. Specifically, it covers two levels: Level 1, the judgment matrix of the criterion layer (B) to the target layer (A). Level 2, the judgment matrix of the indicator layer (C) to its respective criterion layer (B1-B4).

[0060] The comparison scale uses the Saaty 1-9 scale, where 1 indicates equal importance, 3 indicates slightly important, 5 indicates significantly important, 7 indicates very important, 9 indicates extremely important, and 2, 4, 6, and 8 are the median values ​​of adjacent scales.

[0061] Experts are invited to independently fill out the judgment matrix. The expert scores are then combined using the geometric mean method to generate a unified judgment matrix. A consistency check is performed, and if the check fails, feedback is given to the experts for correction.

[0062] (3) Weight calculation and hierarchical ranking The weight calculation is divided into hierarchical single sorting and hierarchical total sorting. Hierarchical single sorting is carried out for each level of judgment matrix. The eigenvalue method is used to solve the maximum eigenvalue and the corresponding eigenvector. After normalization, the weight of the index at this level relative to the corresponding index of the upper level is obtained. Hierarchical total sorting then stacks the weights of each level layer by layer to obtain the comprehensive weight of the bottom-level index relative to the overall goal.

[0063] Through the above steps, the weight values of each level of indicators relative to the overall goal are finally obtained. The specific results are shown in Table 2.

[0064] Table 2 Index Weight Values of the Achievement Transformation Benefit Evaluation System (4)Calculation of Benefit Score The calculation of the benefit score needs to combine the comprehensive weight of the index layer (C) and the actual score, and use the weighted summation formula: [[ID=!14]] [[ID=!15]] ,[[ID=!17]] where, [[ID=!19]] is the total sorting weight of the [[ID=!21]] th index layer index, [[ID=!23]] is the actual score of this index, [[ID=!25]] is the comprehensive benefit score. This score quantitatively represents the expected effectiveness of the project in the achievement transformation stage.

[0065] (5)Benefit Evaluation Index System in the Achievement Transformation Stage Based on the hierarchical structure model and the weight distribution results, starting from four criterion layers of transformation goals, capital investment, technology maturation, and transformation efficiency, a number of specific indicators are refined. Each indicator clearly defines the corresponding evaluation dimension, score weight, and quantitative calculation rules (see Table 3), providing a standardized basis for accurately measuring the benefit level in the transformation stage.

[0066] |table]]Table 3 Benefit Evaluation Index System in the Achievement Transformation Stage <{ It should be noted that: through the intelligent processing of the feasibility study report text and expert review data, the present invention extracts structured review knowledge and constructs an automated scoring and decision-making model, significantly reducing the time-consuming of manual scoring and improving the review efficiency. At the same time, the model adopts a unified scoring standard, avoiding the deviation of review results caused by inconsistent scoring standards of different experts and enhancing the consistency of review results.

[0067] A phased, progressive training strategy, combined with adversarial training and differential learning rate strategies, ensures the model deeply understands the scoring task and improves its generalization ability. A dynamic threshold decision-making mechanism integrates expert consensus, project risk, and model confidence, making the decision boundary closer to real expert experience and enhancing the scientific rigor and robustness of the decision-making process.

[0068] The automated scoring model can quickly process large amounts of project data, meeting the review needs brought about by the increasing number of technology transfers by power grid companies. The dynamic threshold adaptive learning mechanism optimizes parameters based on historical decision results, further enhancing its applicability in large-scale review scenarios. By quantifying scoring standards and introducing a multi-dimensional calibration mechanism, the influence of expert subjective preferences on review results is reduced. Adjusting decision thresholds in conjunction with project risk factors makes the review of high-risk projects more rigorous, improving the objectivity of overall decision-making.

[0069] Example 2 This disclosure also provides a project decision-making and outcome prediction device based on expert knowledge extraction, including a processor and a memory. Optionally, the device may further include a communication interface and a bus. The processor, communication interface, and memory can communicate with each other via the bus. The communication interface can be used for information transmission. The processor can call logical instructions in the memory to execute the project decision-making and outcome prediction method based on expert knowledge extraction described in the above embodiments.

[0070] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0071] Memory, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of this disclosure. The processor executes the program instructions / modules stored in the memory to perform functional applications and data processing, thereby realizing the project decision-making and effectiveness prediction methods based on expert knowledge extraction described in the above embodiments.

[0072] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory may include high-speed random access memory and may also include non-volatile memory.

[0073] This disclosure provides a computer-readable storage medium storing computer-executable instructions configured to execute the above-described project decision-making and outcome prediction method based on expert knowledge extraction.

[0074] The aforementioned computer-readable storage medium may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0075] The technical solutions of this disclosure can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this disclosure. The aforementioned storage medium can be a non-transitory storage medium, including: a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, and other media capable of storing program code. It can also be a transient storage medium.

[0076] The foregoing description and accompanying drawings fully illustrate embodiments of this disclosure to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or” as used in this application means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprises a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., disclosed in the embodiments, if they correspond to the method section disclosed in the embodiments, the relevant parts can be referred to the description of the method section.

[0077] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this disclosure. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0078] The methods and products (including but not limited to devices and equipment) disclosed in the embodiments herein can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the coupling or direct coupling or communication connection shown or discussed between each other may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this disclosure may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0079] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description, and sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A project decision-making and outcome prediction method based on expert knowledge extraction, characterized in that, Includes the following steps: Receive the feasibility study report text and expert review data, and extract the core text through keyword matching rules; The extracted content is segmented into words and segments using the built-in word segmenter of the pre-trained language model, and a vector representation of each segment is obtained. Weights are assigned to each segment based on an attention mechanism. Standardize the units of measurement for expert review scores in the expert review data; A deep learning model based on pre-trained BERT is constructed, and a phased progressive training strategy is adopted, including a scoring regression model training stage and a dynamic threshold classification model training stage. The scoring regression model training stage enables the BERT encoder to learn and fit the quantitative criteria of expert review based on the regression prediction head. The dynamic threshold classification model training stage constructs an independent classification prediction head to obtain a binary classifier for dynamic threshold decision. Collect the text of the feasibility study report to be decided, input it into the trained deep learning model, output the project approval decision, and predict the effectiveness of the approved project.

2. The project decision-making and outcome prediction method based on expert knowledge extraction according to claim 1, characterized in that, The core text is extracted using keyword matching rules. The specific method is as follows: ， in, Indicates the sequential connection of text. This indicates a successfully matched text segment. This represents a matching function used to determine the meaning of a text segment. Does it contain keywords? , Indicates the first A text segment to be matched, Indicates keywords.

3. The project decision-making and outcome prediction method based on expert knowledge extraction according to claim 1, characterized in that, The pre-trained language model's built-in tokenizer uses WordPiece encoding to convert the core text into a token sequence and segments long sequences. The specific method for obtaining the vector representation of each segment is as follows: ， ， in, For the first Vector representation of each segment, Represents the BERT model. This represents the input to the BERT model. This represents a vocabulary mapping function that converts text into its corresponding index in the vocabulary. Indicates classification label, Indicates the first A text fragment of input, Indicates a separator.

4. The project decision-making and outcome prediction method based on expert knowledge extraction according to claim 1, characterized in that, The weight calculation method is as follows: ， ， in, For the first The weight of each segment, For the first Vector representation of each segment, Indicates the first Attention scores for each segment after processing Represents an exponential function. Indicates the first Attention scores for each segment after processing This represents the transpose of the trainable parameters. This represents a trainable weight matrix. Representing vector representation, This represents a trainable bias vector. This represents the hyperbolic tangent activation function. This indicates the total number of segments.

5. The project decision-making and outcome prediction method based on expert knowledge extraction according to any one of claims 1-4, characterized in that, During the training phase of the scoring regression model, a regression prediction head is constructed based on the BERT encoder. The regression prediction head includes a fully connected layer and a ReLU activation function. Adversarial training is employed to improve the model's robustness by applying adversarial perturbations; its loss function... for: ， ， ， in, This represents the mean squared error loss function. This represents the original input data of the model. Indicates a counter-perturbation. Indicates the true label of the regression task. Indicates the mean squared error loss relative to the input. gradient, Indicates the disturbance amplitude or step size. This represents the gradient operator.

6. The project decision-making and outcome prediction method based on expert knowledge extraction according to claim 5, characterized in that, During the training phase of the dynamic threshold classification model, the output vector of the BERT encoder is concatenated with the score vector of the regression prediction to form an enhanced feature vector; the classification prediction head is composed of fully connected layers. The dynamic threshold classification model is trained using a differential learning rate strategy. A normal learning rate is set for the classification prediction head, while a low learning rate, which is 0.01-0.001 times the learning rate of the classification prediction head, is set for the optimized BERT encoder and regression prediction head.

7. The project decision-making and outcome prediction method based on expert knowledge extraction according to claim 6, characterized in that, The loss function during the training phase of the dynamic threshold classification model as follows: ， in, The weight hyperparameters representing the regression loss term. This represents the mean square error. Indicates the first The true regression label of each sample The model represents the first The predicted value for each sample, The weight hyperparameters representing the classification loss term. Represents the cross-entropy loss function. This represents the true classification label of the sample. This represents the classification probability predicted by the model.

8. The project decision-making and outcome prediction method based on expert knowledge extraction according to claim 7, characterized in that, The classification prediction head adopts a dynamic threshold decision mechanism, and the final decision threshold is jointly determined by the initial threshold, the expert consensus calibration factor, and the project risk factor. ， ， ， in, The final decision threshold, As the initial threshold, As an expert consensus calibration factor, As a project risk factor, This indicates the risk of budget overruns. The standard deviation of the expert rating.

9. A project decision-making and outcome prediction device based on expert knowledge extraction, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute, when running the program instructions, the project decision-making and outcome prediction method based on expert knowledge extraction as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the project decision-making and outcome prediction method based on expert knowledge extraction as described in any one of claims 1-8.

Images