A multi-dimensional quantification method and device for evaluating document value

Abstract

The application discloses a multi-dimension quantification literature value evaluation method and device, relates to the related field of literature metrology, and comprises the following steps: extracting features of target literature in a preset dimension through a large model to obtain multi-dimension literature information; constructing a benchmark library of each dimension according to the preset dimension based on the field of expertise; introducing a bidirectional adversarial verification mechanism to dynamically check the mapping and quantification relationship between feature extraction and the benchmark library; after dynamic checking, respectively using the benchmark library of each dimension to perform dimension mapping and quantification on the literature information extracted in each dimension; performing knowledge structure entropy change operation according to the dimension quantification results, quantifying the knowledge structure entropy change value, and performing literature value fusion evaluation according to the dimension quantification results and the knowledge structure entropy change value. The application solves the technical problems of strong subjectivity and incomplete and inaccurate evaluation of the existing literature value evaluation, and achieves the technical effect of objectively, comprehensively and accurately evaluating the literature value.

Description

Technical Field

[0001] This application relates to the field of bibliometrics, and in particular to a multi-dimensional quantitative method and apparatus for evaluating the value of literature. Background Technology

[0002] In the fields of academic research and knowledge innovation, accurately evaluating the value of literature is a crucial step in promoting disciplinary development, selecting high-quality research results, rationally allocating research resources, and facilitating academic exchange and cooperation. Currently, the main method for addressing the issue of literature value evaluation is a combination of manual evaluation and some simple quantitative indicators. Manual evaluation typically relies on the professional knowledge and experience of field experts, who subjectively judge the literature based on its content, considering aspects such as innovativeness, practicality, and research methods. Simple quantitative indicators select easily statistically relevant metrics, such as the number of citations and the impact factor of the journal, to conduct a preliminary assessment of the literature's value. However, expert review is highly subjective, inefficient, and difficult to scale; traditional bibliometric indicators lag behind actual impact and fail to reveal the inherent technical novelty and innovative value of the literature, often resulting in a disconnect between the evaluation results and the true technical value of the literature.

[0003] At present, the evaluation of the value of literature suffers from technical problems such as strong subjectivity and incomplete and inaccurate evaluation. Summary of the Invention

[0004] This application provides a multi-dimensional quantitative method and apparatus for evaluating the value of literature. It employs techniques such as extracting multi-dimensional information from target literature using a large model according to preset dimensions, constructing benchmark databases for each dimension based on the target literature's field, verifying the validity of the mapping and quantification relationship between feature extraction and the benchmark databases through a two-way adversarial verification mechanism, mapping and quantifying the literature information extracted from each dimension using the benchmark databases to obtain quantification results for each dimension, performing knowledge structure entropy change calculations based on the quantification results, and integrating the quantification results with the entropy change values ​​to generate the final literature value evaluation conclusion. These techniques solve the technical problems of strong subjectivity and incomplete and inaccurate evaluation in existing literature value evaluation methods, achieving an objective, comprehensive, and accurate evaluation of literature value.

[0005] This application provides a multi-dimensional quantitative method for evaluating the value of literature, comprising: extracting features of target literature in preset dimensions using a large model to obtain multi-dimensional literature information, wherein the preset dimensions include at least technical concepts, core technologies, public coverage, and differentiated increments; constructing benchmark libraries for each dimension based on the field to which the target literature belongs; introducing a two-way adversarial verification mechanism to dynamically verify the mapping and quantification relationship between feature extraction and benchmark libraries; after dynamic verification, using the benchmark libraries for each dimension, performing dimension mapping and quantification on the literature information extracted in each dimension (technical concepts, core technologies, public coverage, and differentiated increments) to obtain quantification results for each dimension; performing knowledge structure entropy change calculation based on the quantification results for each dimension to quantify the knowledge structure entropy change value; and performing a literature value fusion evaluation based on the quantification results for each dimension and the knowledge structure entropy change value to generate a literature value evaluation result.

[0006] In a possible implementation, the benchmark libraries for each dimension are used to perform dimension mapping quantification on the literature information extracted from the dimensions of technical concept, core technology, public coverage, and differentiated increment. The following processing is performed: a quantification mapping relationship is established with the corresponding benchmark library according to the quantification calculation formula for each dimension; a quantification module for each dimension is constructed based on the quantification mapping relationship for each dimension, wherein the corresponding quantification calculation formula is embedded in the quantification module, and a unique dimension identifier is assigned to each quantification module; based on the dimension identifier, the corresponding literature information is mapped to the quantification module of the corresponding dimension, and combined with the data in the benchmark library, the quantification calculation formula is used for calculation processing to output the quantification results for each dimension.

[0007] In a possible implementation, according to the quantization calculation formulas of each dimension, a quantization mapping relationship is established with the corresponding benchmark library, and the following processing is performed: the calculation parameters are parsed according to the quantization calculation formulas, the calculation input data and calculation methods are identified, and the calculation mapping relationship between each dimension and the benchmark library is established according to the calculation mapping relationship between the calculation parameters and the benchmark library; according to the calculation mapping relationship, the quantization calculation formulas are precisely paired with the corresponding benchmark library data to construct a quantization mapping framework, wherein the quantization calculation formulas are quantized and mapped with the corresponding calculation benchmark library.

[0008] In a possible implementation, before establishing the quantization mapping relationship with the corresponding benchmark library according to the quantization calculation formulas of each dimension, the following processing is performed: A sample dataset is constructed according to the aforementioned technical concept, core technology, public coverage, and differentiated incremental preset dimensions; quantization calculation formulas are fitted based on the sample dataset, wherein a positive sample set and a negative sample set are established, and both the positive and negative sample sets include a training set and a test set; the quantization calculation relationships of each preset dimension are trained using the training sets of the positive and negative sample sets respectively, and the trained quantization parameters are verified using the test sets of the positive and negative sample sets until a bidirectional convergence target is reached, thereby obtaining the quantization calculation formulas of each preset dimension.

[0009] In a possible implementation, the following processing is performed: the quantitative calculation formula for the dimension corresponding to the technical concept includes: quantifying the data of technical concept sentences in the literature, and calculating the proportion of the quantified technical concept sentence data with a preset quantity threshold to obtain the quantitative proportion of the concept sentences; segmenting the benchmark library to obtain a literature concept benchmark library and an unsolved problem benchmark library; quantifying the mean cosine similarity between the technical concept sentences and the literature concept benchmark library to obtain a first similarity; quantifying the mean cosine similarity between the technical concept sentences and the unsolved problem benchmark library to obtain a second similarity; performing a weighted calculation on the first similarity and the second similarity, and using the quantitative proportion of the concept sentences as a coefficient to multiply and quantify the weighted result to obtain a comprehensive evaluation result for the technical concept dimension; and normalizing and adjusting the comprehensive evaluation result to obtain the final quantitative evaluation result for the technical concept dimension.

[0010] In a possible implementation, the following processing is performed: the quantitative calculation formula for the core technology dimension includes: quantifying the technical innovation points including the technical contribution characteristics of the literature, and calculating the proportion using the number of technical innovation points and a preset number threshold to obtain the quantitative proportion of contribution points; quantifying the average cosine similarity between the technical innovation points and the innovation features in the benchmark library of the relevant field to obtain the intra-library similarity; quantifying the average cosine similarity between the technical innovation points to obtain the intra-point similarity; weighting the intra-library similarity and the intra-point similarity, and then multiplying the quantitative proportion of contribution points as a coefficient with the weighted calculation result, and normalizing the calculation result to obtain the quantitative evaluation result of the core technology dimension.

[0011] In a possible implementation, the following processing is performed: the quantitative calculation formula for the dimension corresponding to the public coverage includes: segmenting the relevant field benchmark library according to the public disclosure of technical theories, public disclosure of industry applications, and project goals to obtain a theoretical technical benchmark library, an industry technical application benchmark library, and a project goal benchmark library; quantifying the average cosine similarity between the technical concept sentence combined with the technical innovation sentence and the theoretical technical benchmark library to obtain a third similarity, the higher the third similarity, the higher the public coverage; quantifying the average cosine similarity between the technical concept sentence combined with the technical innovation sentence and the industry technical application benchmark library to obtain a fourth similarity; quantifying the average cosine similarity between the technical concept sentence combined with the technical innovation sentence and the project goal benchmark library to obtain a fifth similarity; weighting the third, fourth, and fifth similarities to obtain a similarity weighted result, and calculating the undisclosed coverage based on the difference between the similarity weighted result and 1; weighting the second similarity and the undisclosed coverage, and normalizing the calculation result to obtain a quantitative evaluation result for the public coverage dimension.

[0012] In a possible implementation, the following processing is performed: the quantitative calculation formula for the dimension corresponding to the differentiated increment includes: quantifying the cosine similarity between the technological innovation point and the mean vector of all innovation features in the benchmark library of the corresponding field to obtain the technological innovation similarity, wherein the lower the technological innovation similarity, the greater the differentiation; using a large model, comparing and identifying the technological innovation point with all innovation features in the benchmark library of the corresponding field to extract breakthrough improvement technology sentences; comparing the number of identified and extracted breakthrough improvement technology sentences with a preset increment threshold to obtain the incremental breakthrough feature coefficient; and calculating the quantitative evaluation result of the differentiated increment dimension using the formula: S4=(1-technical innovation similarity)×(1+incremental breakthrough feature coefficient)×12.5.

[0013] In a possible implementation, knowledge structure entropy change calculation is performed based on the quantification results of each dimension to quantify the knowledge structure entropy change value, and the following processing is performed: constructing a domain knowledge structure state representation for the time period before the publication of the target literature, extracting the core literature set of the target literature within the domain for the time period, and constructing a concept co-occurrence network; simulating the knowledge structure reorganization process after the introduction of the target literature, wherein the concept set of the target literature is injected into the concept co-occurrence network to form a new concept co-occurrence network, the weights of all edges of the concept co-occurrence network are recalculated based on the content similarity of the literature, identifying the new technology path branches generated by the introduction of the concept set, and constructing an extended path set; evaluating the degree of reconstruction of the common knowledge structure by the target literature based on the extended path set; calculating multi-dimensional information entropy change components based on the degree of knowledge structure reconstruction, and performing collaborative weighted fusion of the entropy change components based on the quantification results of each dimension to calculate the final knowledge structure entropy change value.

[0014] This application also provides a multi-dimensional quantitative document value evaluation device, comprising: a feature extraction module, used to extract features of the target document in preset dimensions using a large model to obtain multi-dimensional document information, wherein the preset dimensions include at least technical concept, core technology, public coverage, and differentiated increment; a benchmark library construction module, used to construct benchmark libraries for each dimension according to the field to which the target document belongs; a dynamic verification module, used to introduce a two-way adversarial verification mechanism to dynamically verify the mapping and quantification relationship between feature extraction and benchmark library; a dimension mapping and quantification module, used to perform dimension mapping and quantification on the document information extracted in each dimension (technical concept, core technology, public coverage, and differentiated increment) using the benchmark libraries for each dimension after dynamic verification, to obtain quantification results for each dimension; and a document value fusion evaluation module, used to perform knowledge structure entropy change calculation based on the quantification results for each dimension, quantify the knowledge structure entropy change value, and perform document value fusion evaluation based on the quantification results for each dimension and the knowledge structure entropy change value to generate a document value evaluation result.

[0015] This application proposes a multi-dimensional quantitative method and apparatus for evaluating the value of literature. First, a large model is used to extract features from the target literature across preset dimensions, obtaining multi-dimensional literature information. These preset dimensions include at least technical concepts, core technologies, public coverage, and differentiated increments. Next, based on the target literature's field, benchmark libraries are constructed for each preset dimension. A two-way adversarial verification mechanism is then introduced to dynamically verify the mapping and quantification relationship between feature extraction and the benchmark libraries. After dynamic verification, the literature information extracted for each dimension (technical concepts, core technologies, public coverage, and differentiated increments) is quantified using the benchmark libraries for each dimension, obtaining quantification results for each dimension. Finally, knowledge structure entropy change calculations are performed based on the quantification results for each dimension, quantifying the knowledge structure entropy change value. A fusion evaluation of the literature value is then conducted based on the quantification results for each dimension and the knowledge structure entropy change value, generating a literature value evaluation result. Through this process, the method and apparatus proposed in this application achieve the technical effect of objectively, comprehensively, and accurately evaluating the value of literature. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0017] Figure 1This is a flowchart illustrating a multi-dimensional quantitative method for evaluating the value of documents, as provided in an embodiment of this application.

[0018] Figure 2 This is a schematic diagram of a multi-dimensional quantitative document value evaluation device provided in an embodiment of this application.

[0019] Figure labeling: Feature extraction module 10, benchmark library construction module 20, dynamic verification module 30, dimensional mapping quantification module 40, and literature value fusion evaluation module 50. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] This application provides a multi-dimensional quantitative method for evaluating the value of documents, such as... Figure 1 As shown, the method includes:

[0022] Step S100: Extract features from the target document using a large model with preset dimensions to obtain multi-dimensional document information. The preset dimensions include at least technical concepts, core technologies, public coverage, and differentiated increments.

[0023] Specifically, prompt word engineering or model fine-tuning can be used to automatically identify and extract text fragments related to the evaluation dimensions from the full text of the target document. For example, a pre-trained large language model, such as BERT, GPT series, or ChatGLM, can be used, with specific information extraction prompt templates designed for it. These templates explicitly instruct the model to find sentences describing technical ideas, core technologies, public coverage, and differentiated incremental features from the input text. Alternatively, model fine-tuning can be used, which involves collecting a batch of patent documents that have been annotated by experts with technical idea sentences, technical innovation sentences, breakthrough improvement sentences, innovative sentences from papers, unsolved problems sentences, technical application sentences from industry information, and project goal sentences from public projects, etc., as training data. This data is used to conduct supervised training on the large model, enabling it to learn to accurately extract these specific types of sentences.

[0024] Step S200: Based on the field to which the target document belongs, construct a benchmark library for each dimension according to preset dimensions.

[0025] Specifically, a benchmark library serves as a reference for quantitative comparisons; each benchmark library is a collection of texts from a specific domain and of a specific type. The process involves benchmark library construction and text vectorization. The benchmark libraries are built upon existing literature data resources, including patents, academic papers, industry analysis reports, and open-source project documents. Then, text embedding models, such as Qwen3-Embedding-8B and BAAI / bge-m3, are used to convert key sentences from each document—such as the invention content sentences of patents and the innovative points sentences of papers—into high-dimensional numerical vectors. The collection of these vectors constitutes the benchmark library. In the system, a benchmark library is essentially a vector database.

[0026] Step S300 introduces a two-way adversarial verification mechanism to dynamically verify the mapping and quantization relationship between feature extraction and the benchmark library.

[0027] Specifically, the bidirectional adversarial verification mechanism dynamically verifies both the accuracy of the feature extraction results and the stability of the mapping quantization relationship with the benchmark library. It can be implemented using either the bidirectional adversarial verification mechanism or a Generative Adversarial Network (GAN), as detailed below:

[0028] If a two-way adversarial verification mechanism is adopted, a two-dimensional adversarial verification process needs to be designed simultaneously for the mapping relationship between the feature extraction results and the benchmark library. Specifically, for the feature extraction results, based on the extracted text fragments such as technical concept sentences and core technical sentences, three types of adversarial samples are generated through rule-based perturbation: First, semantic interference samples, which replace core terms using a domain thesaurus (e.g., replacing recurrent neural networks with RNNs, adjusting sentence structure such as changing active sentences to passive sentences) to simulate semantic comprehension bias; second, information missing samples, which randomly delete 20%-30% of key information, such as technical parameters and innovation logic, to simulate information omission scenarios; and third, noise-mixed samples, which insert 5%-10% of domain-irrelevant expressions, such as those with application prospects, to simulate noise interference. The three types of adversarial samples and the original feature results are then input into a preset dimensional quantization model to calculate the deviation between the quantized adversarial sample results and the original results. An absolute error ≤ 0.5% and a relative error ≤ 5% are considered acceptable thresholds. If any adversarial sample deviation exceeds the threshold, the feature extraction robustness is deemed insufficient, requiring re-optimization of the prompt word template of the large model or fine-tuning of the training data. Secondly, regarding the mapping and quantization relationship of the benchmark library, two types of adversarial modifications are applied to the benchmark libraries of each dimension: First, vector perturbation modification, adding random Gaussian noise of strength 0.01-0.05 to all text embedding vectors in the benchmark library to simulate vector deviation during data storage or retrieval; second, sample replacement modification, randomly replacing 5%-8% of the sample data in the benchmark library. The replaced samples belong to the same dimension as the original samples but differ in technical features, simulating data quality fluctuations during the benchmark library update process. The original feature extraction results are input into the modified adversarial benchmark library and the original benchmark library respectively, and the deviation value of the two quantization results is calculated. If the deviation exceeds the standard, the benchmark library mapping stability is deemed insufficient, requiring re-screening of benchmark library samples or optimization of text embedding model parameters. After completing the adversarial verification, if the deviation values ​​of both types of verification meet the acceptable threshold, and the results of three consecutive repeated verifications are consistent, then the feature extraction and benchmark library mapping and quantization relationship are confirmed to meet the requirements, and the subsequent dimension mapping and quantization stage can proceed. If the deviation in any verification stage exceeds the standard, iterative optimization of the corresponding stage is required until all verifications meet the standards.

[0029] If a Generative Adversarial Network (GAN) is used, a feature extraction GAN sub-network and a benchmark database mapping GAN sub-network need to be constructed and validated separately. For feature extraction, a feature generator-discriminator sub-network is constructed. The generator takes the full text of the target document as input, learns the true feature extraction rules, and generates adversarial feature samples that are semantically similar but lack key information and structurally consistent but with terminological deviations by adjusting the loss function. The discriminator takes true feature samples, adversarial feature samples, and invalid samples as input and learns to distinguish feature effectiveness. The adversarial samples are input into the dimensionality quantization model, and the deviation from the true samples is compared. Simultaneously, the discriminator scores the effectiveness of the true features to determine the quality of feature extraction; if it fails to meet the standard, the original feature extraction model is optimized. For benchmark database mapping, a benchmark sample generator-mapping discriminator sub-network is constructed. The generator takes the original benchmark database vectors as input, learns the vector distribution rules, and generates adversarial benchmark samples that are close to the original distribution but contain subtle perturbations, as well as replacement samples of the same dimension but with small technical differences. The discriminator takes the quantization results of the original benchmark database and original features, and the quantization results of the adversarial benchmark database and original features as input, and learns to distinguish the stability of the quantization results. The stability of the benchmark library mapping is determined by scoring the stability of the benchmark library quantization results using a discriminator. If the stability is not met, the benchmark library sample selection or embedding model is optimized.

[0030] Step S400: After dynamic verification, the literature information extracted from each dimension of the technical concept, core technology, public coverage, and differentiated increment is quantified by using the benchmark library of each dimension to obtain the quantification results of each dimension.

[0031] Specifically, after dynamic verification, the target document features extracted in step S100 are compared and calculated with the benchmark database constructed in step S200 to obtain a score.

[0032] In one possible implementation, the literature information extracted from the dimensions of technical concept, core technology, public coverage, and differentiated increment is quantified by utilizing the benchmark libraries for each dimension. Step S400 further includes step S410, constructing a sample dataset according to the preset dimensions of technical concept, core technology, public coverage, and differentiated increment. Specifically, training data is prepared for the automated fitting of the quantization formula, and a batch of literature samples are collected and labeled manually or semi-automatically. For each sample, in addition to its original text, there are scores assigned to it by domain experts according to established rules in four dimensions, such as 0-25 points. The set of these literature-score pairs constitutes the sample dataset.

[0033] Step S420: Based on the sample dataset, perform quantitative calculation formula fitting, wherein a positive example sample set and a negative example sample set are established, and both the positive and negative example sample sets include a training set and a test set. Specifically, the sample dataset is divided into subsets for training and testing to optimize and validate the formula. In the program, the sample dataset is randomly split, for example, 70% as the training set and 30% as the test set. For each evaluation dimension, a positive example and a negative example sample set for machine learning are constructed. Positive examples refer to the set of literature features that perform well in a specific dimension, while negative examples refer to the set of literature features that perform poorly in that dimension. Specifically, for the technical concept dimension, positive examples refer to literature with novel invention concepts that match actual needs; negative examples, conversely, refer to literature with outdated invention concepts that are divorced from actual needs. For the core technology dimension, positive examples refer to literature with novel technology and consistent logic; negative examples refer to literature with outdated technology and confused logic. For the publication coverage dimension, positive examples refer to literature that has not been publicly covered; negative examples refer to literature that has been widely published. For the dimension of differential increment, positive examples refer to literature with large overall differences and many breakthrough improvements; negative examples refer to literature with small overall differences and few breakthrough improvements.

[0034] Step S430 involves training the quantization operation relationships for each preset dimension using the training sets of the positive and negative sample sets, respectively. The trained quantization parameters are then validated using the test sets of the positive and negative sample sets until a bidirectional convergence objective is achieved, yielding the quantization operation formulas for each preset dimension. Specifically, machine learning methods are used to automatically determine the weight parameters and structure in each formula in steps S431-S434. Optimization algorithms such as linear regression or gradient descent can be employed. Various extracted feature values, such as similarity and sentence count, are used as input features, and the expert scoring dimension score is used as the target output label. The process iterates repeatedly on the training set, adjusting the weights in the formulas to make the calculation results as close as possible to the expert scores. Then, the generalization ability of the trained formulas is evaluated using the test set. The bidirectional convergence objective refers to a dual validation standard: first, the loss function on the training set no longer decreases significantly, indicating that the model has learned sufficiently; second, the performance indicators on the test set simultaneously reach stability and meet expectations, ensuring that the trained formulas do not overfit the training data and possess good generalization ability. Training stops when both conditions are met in consecutive training rounds, and the resulting formula is the final model.

[0035] In one possible implementation, step S430 further includes step S431, where the quantification formula for the dimension corresponding to the technical idea includes: quantifying the data of technical idea sentences in the literature, and calculating the proportion using the quantified number of technical idea sentences and a preset number threshold to obtain the quantified proportion of idea sentences; segmenting the benchmark library to obtain a literature idea benchmark library and an unsolved problem benchmark library; quantifying the mean cosine similarity between technical idea sentences and the literature idea benchmark library to obtain a first similarity; quantifying the mean cosine similarity between technical idea sentences and the unsolved problem benchmark library to obtain a second similarity; performing a weighted calculation on the first similarity and the second similarity, and using the quantified proportion of idea sentences as a coefficient to multiply and quantify the weighted result to obtain a comprehensive evaluation result for the technical idea dimension; and normalizing the comprehensive evaluation result to obtain the final quantified evaluation result for the technical idea dimension, wherein the preset number threshold for the technical idea sentences does not exceed 8.

[0036] Specifically, the program counts the number of technical idea sentences identified in the target document, then divides this number by a preset threshold, for example, 8. If the number of technical idea sentences is greater than or equal to 8, the quantification percentage of the idea sentence is directly set to 1. Each technical idea sentence in the target document is converted into a vector using the same model, and then the cosine similarity between each vector and all vectors in the document's idea benchmark library is calculated, and the average is taken to obtain the first similarity. Similarly, the average cosine similarity with the unresolved problem benchmark library is calculated to obtain the second similarity. The formula is used accordingly. Calculations are performed, in which, It is the final quantitative evaluation result of the technical concept dimension. It is the number of technical concept sentences. This means setting the preset quantity threshold upper limit to 8 to avoid quantity accumulation. This refers to the first similarity score. The lower the first similarity score, the greater the difference from existing literature, and the higher the novelty. The second similarity score indicates that the idea closely matches the real needs of the field and is more necessary. The 25-point normalization adjustment is used to ensure... It falls within the range of 0 to 25 points.

[0037] In one possible implementation, step S430 further includes step S432, wherein the quantitative calculation formula for the core technology corresponding dimension includes: quantifying the technical innovation points including the technical contribution characteristics of the literature, and calculating the proportion using the number of technical innovation points and a preset number threshold to obtain the quantitative proportion of contribution points; quantifying the average cosine similarity between the technical innovation points and the innovation features in the benchmark library of the corresponding field to obtain the intra-library similarity; quantifying the average cosine similarity between the technical innovation points to obtain the intra-point similarity; weighting the intra-library similarity and the intra-point similarity, and then multiplying the quantitative proportion of contribution points as a coefficient with the weighted calculation result, and normalizing the calculation result to obtain the quantitative evaluation result of the core technology dimension, wherein the preset number threshold of the technical innovation points does not exceed 10.

[0038] Specifically, this step is similar to step S431. The program counts the number of sentences representing technological innovation points and calculates their proportion relative to a preset threshold. It then calculates the average similarity between the vectors of these sentences and the benchmark database of technological innovation points to obtain the intra-database similarity. Simultaneously, it calculates the average cosine similarity between each pair of these sentence vectors to obtain the intra-point similarity. Finally, the results are substituted into the formula... Calculations are performed, in which, It is the final quantitative evaluation result of the core technology dimension. It refers to the number of sentences highlighting technological innovations in the literature. This means setting the preset quantity threshold upper limit to 10 to avoid quantity accumulation. This refers to intra-library similarity. The lower the intra-library similarity, the greater the difference from existing technologies, and the higher the novelty. This refers to intra-point similarity. Higher intra-point similarity indicates consistency and logical inconsistency in the technical solutions, demonstrating greater rigor and avoiding ineffective innovation driven by the pursuit of differences. The 25 refers to normalization adjustments used to ensure... It falls within the range of 0 to 25 points.

[0039] In one possible implementation, step S430 further includes step S433, where the quantitative calculation formula for the dimension corresponding to the disclosure coverage includes: segmenting the relevant field benchmark library according to the disclosure of technical theory, the disclosure of industry application, and the project goal to obtain a theoretical technical benchmark library, an industry technical application benchmark library, and a project goal benchmark library; quantifying the mean cosine similarity between the technical concept sentence combined with the technical innovation sentence and the theoretical technical benchmark library to obtain a third similarity, where a higher third similarity reflects a higher disclosure coverage; quantifying the mean cosine similarity between the technical concept sentence combined with the technical innovation sentence and the industry technical application benchmark library to obtain a fourth similarity; quantifying the mean cosine similarity between the technical concept sentence combined with the technical innovation sentence and the project goal benchmark library to obtain a fifth similarity; weighting the third, fourth, and fifth similarities to obtain a similarity weighted result, and calculating the undisclosed coverage based on the difference between the similarity weighted result and 1; weighting the second similarity and the undisclosed coverage, and normalizing the calculation result to obtain a quantitative evaluation result for the disclosure coverage dimension.

[0040] Specifically, the vectors of the technical concept sentences and technical innovation points in the target document are merged into a single query vector. The mean cosine similarity of this query vector with the theoretical technology benchmark database, the industry technology application benchmark database, and the project target benchmark database are calculated respectively to obtain the third, fourth, and fifth similarity scores. Then, the formula is applied... Calculations are performed, in which, This is the final quantitative evaluation result of the public coverage dimension. The third similarity score indicates that the technology is not covered by existing theories, and thus has high theoretical novelty. The fourth similarity score indicates that existing applications do not cover the technology, and the application is highly novel. The fifth similarity score indicates that existing projects do not cover the technology, suggesting high novelty at the application level. The 25 refers to normalization adjustments used to ensure... It falls within the range of 0 to 25 points.

[0041] In one possible implementation, step S430 further includes step S434, where the quantitative calculation formula for the dimension corresponding to the differentiated increment includes: quantifying the cosine similarity between the technological innovation sentence and the mean vector of all innovation features in the benchmark library of the relevant field to obtain the technological innovation similarity, wherein the lower the technological innovation similarity, the greater the differentiation; using a large model, comparing and identifying the technological innovation sentence with all innovation features in the benchmark library of the relevant field to extract breakthrough improvement technological sentences; comparing the number of identified and extracted breakthrough improvement technological sentences with a preset increment threshold to obtain the incremental breakthrough feature coefficient; and calculating the quantitative evaluation result of the differentiated increment dimension using the formula: S4 = (1 - technological innovation similarity) × (1 + incremental breakthrough feature coefficient) × 12.5, wherein the preset increment threshold does not exceed 2.

[0042] Specifically, the mean vector of all vectors in the benchmark database of technological innovation points is pre-calculated. Then, the cosine similarity between the vector of technological innovation point sentences in the target document and the mean vector is calculated to obtain the technological innovation similarity. An LLM with finely tuned or carefully designed prompt words is used to identify breakthrough improvement sentences. For example, the prompt word could be: Please judge whether the following technical point description represents a breakthrough improvement relative to the existing technology: {technical innovation point sentence}. Only yes or no answers are allowed. The program counts the number of sentences that answer yes, obtaining the number of breakthrough improvement technical sentences identified and extracted. Substitute the results into the formula S4 = (1 - technological innovation similarity) × (1 + incremental breakthrough feature coefficient) × 12.5 for calculation. The lower the technological innovation similarity, the further the technical solution deviates from the average level of the field, and the stronger the overall differentiation. The incremental breakthrough feature coefficient = number of breakthrough improvement technical sentences / preset incremental threshold. The more breakthrough improvement technical sentences there are, the more breakthroughs in local key technologies there are, and the higher the value of the innovation increment.

[0043] Step S440: Establish a quantization mapping relationship between the quantization calculation formulas for each dimension and the corresponding benchmark library. Specifically, at the system design level, the quantization calculation formulas, data, and calculation process are solidified.

[0044] In one possible implementation, according to the quantization calculation formulas for each dimension, a quantization mapping relationship is established between the corresponding benchmark libraries. Step S440 further includes step S441, which involves parsing the calculation parameters according to the quantization calculation formulas, identifying the calculation input data and calculation methods, and establishing the calculation mapping relationship between each dimension and the benchmark libraries based on the calculation mapping relationship between the calculation parameters and the benchmark libraries. Specifically, configuration management is performed; in the code, a configuration class or dictionary is created for each dimension. This configuration explicitly specifies which benchmark libraries are needed to calculate the score for that dimension, which features need to be extracted from the target literature, and which function is specifically used to perform the calculation.

[0045] Step S442: According to the aforementioned operational mapping relationship, the quantization calculation formula is precisely paired with the corresponding benchmark library data to construct a quantization mapping framework, wherein the quantization calculation formula is quantized and mapped to the corresponding operational benchmark library. Specifically, a runtime connection is executed. When the system needs to evaluate a new document, the scheduler, based on the mapping relationship established in step S441, calls the corresponding benchmark library data from the vector database and sends it, along with the corresponding calculation function and target document feature data, into the calculation engine, thus binding data query with business logic.

[0046] Step S450: Based on the quantization mapping relationship of each dimension, construct quantization modules for each dimension. Each quantization module embeds the corresponding quantization calculation formula, and a unique dimension identifier is assigned to each quantization module. Specifically, software modularization is implemented; in object-oriented programming, an independent module or class is created for each dimension. Each class encapsulates the quantization calculation formula for that dimension and possesses a unique dimension identifier.

[0047] Step S460: Based on the dimension identifier, the corresponding literature information is mapped to the corresponding dimension's quantization module. Combined with data from the benchmark library, the quantization calculation formula is used for processing, and the quantization results for each dimension are output. Specifically, the pipeline is executed, and the system main program holds a list of all quantization modules and mapping rules. When processing a document, feature extraction from step S100 is completed first, then each dimension identifier is traversed. For example, when processing the technical idea dimension, the scoring method of the technical idea dimension calculation module is called, and the extracted technical idea sentence vector and the required benchmark library connection information are passed as parameters to this method. This method internally performs the calculation and returns a score. The main program collects the scores returned by all modules.

[0048] Step S500: Perform knowledge structure entropy change calculation based on the quantification results of each dimension, quantify the knowledge structure entropy change value, and perform document value fusion evaluation based on the quantification results of each dimension and the knowledge structure entropy change value to generate document value evaluation results.

[0049] Specifically, based on the quantitative results of four dimensions—technical concept, core technology, public coverage, and differentiated increment—a knowledge structure entropy change operation is performed to obtain the knowledge structure entropy change value. This operation focuses on the changes in the topology, technical path distribution, and topic concentration of the knowledge network within the target document before and after its introduction. The quantitative results of the four dimensions are then weighted and fused with the knowledge structure entropy change value to generate a comprehensive evaluation result of the document's value. The fusion evaluation establishes a mapping function between the dimensional quantitative results and the entropy change components, assigning different dimensional scores a moderating weight to the entropy change components. Finally, a linear weighted summation formula is used to calculate the comprehensive score.

[0050] In one possible implementation, based on the quantization results of each dimension, an entropy change operation of the knowledge structure is performed to quantify the entropy change value of the knowledge structure. Step S500 further includes step S510 of constructing a representation of the state of the domain knowledge structure in the time period before the publication of the target document, extracting the core document set in the time period of the target document in the domain, and constructing a concept co-occurrence network. Specifically, determine the publication time node of the target document, and filter out the domain document data within a preset time period before the publication of the target document based on the publication timestamp of the document. For example, take the domain documents in the three years before the publication of the target document as the core document set, and the document types include patents, academic papers, industry technical reports, etc. Use word segmentation tools such as Jieba Segmentation or HanLP to perform word segmentation on the key texts such as the title, abstract, invention content (patent), and conclusion (paper) of the core document set, filter out meaningless words such as "of", "and", "a kind of" through the stop word list, and then use the TF-IDF algorithm to calculate the domain weights of the words, and screen out the words with the top-ranked weights as the core concepts of the domain. Construct a concept co-occurrence network based on the co-occurrence relationship of the core concepts in a single document to represent the baseline state of the domain knowledge structure in this time period. Among them, the nodes of the network are the core concepts of the domain, and the edge weights between the nodes are the co-occurrence frequencies of the two concepts in the core document set. Use the NetworkX graph calculation library to store the node set, edge set, and edge weight data of the network to complete the construction of the baseline state of the domain knowledge structure.

[0051] Step S520, simulate the knowledge structure reorganization process after introducing the target document. Among them, inject the concept set of the target document into the concept co-occurrence network to form a new concept co-occurrence network, recalculate the weights of all edges in the concept co-occurrence network based on the content similarity of the documents, identify the new technology path branches generated by the introduction of the concept set, and construct an extended path set; evaluate the reconstruction degree of the target document on the common knowledge structure based on the extended path set. Specifically, extract the core concept set of the target document according to the same method in step S510, add these core concepts as new nodes to the original concept co-occurrence network, and directly reuse them if the concepts already exist. Then calculate the co-occurrence frequencies of the core concepts of the target document and the node concepts of the original network, and update the weights of the corresponding edges. If there is no co-occurrence relationship between the new node and the original node, the edge weight is set to zero. Based on the updated concept co-occurrence network, use the shortest path algorithm to calculate the paths between any two core technical concepts, and screen out the paths containing the new concepts added by the target document as the new technology path branches to form an extended path set. Quantify the reconstruction degree of the target document on the knowledge structure by calculating two indicators: the proportion of the number of extended paths to the total number of original paths and the proportion of the weight of the newly added edges to the total edge weight.

[0052] Step S530: Calculate the multi-dimensional information entropy change components based on the degree of knowledge structure reconstruction, and perform collaborative weighted fusion of the entropy change components based on the quantification results of each dimension to calculate the final knowledge structure entropy change value. Specifically, three entropy change components are calculated: First, network structure entropy change, calculated as: Network structure entropy change = Node degree distribution entropy of the new concept co-occurrence network - Node degree distribution entropy of the original concept co-occurrence network. The larger this value, the more significant the change in network topology. The node degree distribution entropy is calculated as: Node degree distribution entropy = -(Degree percentage of node 1 × Logarithm of the degree percentage of node 1 (base 2) + Degree percentage of node 2 × Logarithm of the degree percentage of node 2 (base 2) + ... + Degree percentage of the nth node × Logarithm of the degree percentage of the nth node (base 2)). The degree percentage of a single node... The first is the ratio of the degree of the node to the sum of the degrees of all nodes in the network. The degree of a node refers to the number of edges connecting it to other nodes. The second is the path diversity entropy change, calculated as: Path diversity entropy change = Shannon entropy of the expanded path set - Shannon entropy of the original path set. The larger this value, the more significant the improvement in technical path diversity. The Shannon entropy is calculated as: Shannon entropy = -(Proportion of path type 1 × Logarithm of the proportion of path type 1 to base 2 + Proportion of path type 2 × Logarithm of the proportion of path type 2 to base 2 + ... + Proportion of the m-th path type × Logarithm of the proportion of path type 1 to base 2) The logarithm of the proportion of the m-th path type), where the proportion of a single path type = the number of paths of that type ÷ the total number of all paths in the path set; thirdly, the cognitive concentration entropy change, calculated as: cognitive concentration entropy change = 1 - cosine similarity of the core topic distributions of the original network and the new network. The larger this value, the more significant the change in the technical topic distribution. Cosine similarity = (proportion of original topic 1 × proportion of new topic 1 + proportion of original topic 2 × proportion of new topic 2 + ... + proportion of original topic k × proportion of new topic k) ÷ (the magnitude of the original topic distribution vector × the proportion of the new topic distribution vector) The vector modulus is √(the square of the proportion of topic 1 + the square of the proportion of topic 2 + ... + the square of the proportion of topic k), where k is the number of core topics. The vector modulus is obtained by training the LDA topic model on the document sets corresponding to the original concept co-occurrence network and the document sets corresponding to the new concept co-occurrence network. The core topic distribution vectors of the two networks are as follows: Original network topic distribution vector = [original topic 1 proportion, original topic 2 proportion, ..., original topic k proportion], New network topic distribution vector = [new topic 1 proportion, new topic 2 proportion, ..., new topic k proportion].

[0053] A collaborative mapping relationship between dimensional quantification results and entropy change components is established. The quantification score of the technical concept dimension is divided by a maximum score of 25 to obtain a normalized coefficient, which serves as an input factor for the cognitive concentration entropy change and is multiplied by the cognitive concentration entropy change. Similarly, the quantification score of the core technology dimension is divided by a maximum score of 25 to obtain a normalized coefficient, which serves as an adjustment coefficient for the path diversity entropy change and is multiplied by the path diversity entropy change. The quantification score of the differentiated increment dimension is divided by a maximum score of 25 to obtain a normalized coefficient, which serves as a weight adjustment factor for the network structure entropy change and is multiplied by the network structure entropy change. Finally, the quantification score of the public coverage dimension is divided by a maximum score of 25 to obtain a normalized coefficient, which serves as a global correction coefficient when the three adjusted entropy change components are weighted and summed. The three adjusted entropy change components are then weighted and summed according to weight coefficients such as 0.3, 0.4, and 0.3. The summation result is then multiplied by the normalized coefficient of the public coverage dimension, and the resulting product is the final knowledge structure entropy change value.

[0054] This application employs several technical means to address the existing problems of strong subjectivity and incomplete and inaccurate evaluation of literature value. These methods include extracting multi-dimensional information from target documents using a large model based on preset dimensions, constructing benchmark libraries for each dimension according to the target document's field, verifying the effectiveness of the mapping and quantification relationship between feature extraction and the benchmark libraries through a two-way adversarial verification mechanism, mapping and quantifying the literature information extracted from each dimension using the benchmark libraries, obtaining quantification results for each dimension, performing knowledge structure entropy change calculations based on the quantification results, and integrating the quantification results with the entropy change values ​​to generate the final literature value evaluation conclusion. This approach solves the technical problems of strong subjectivity and incomplete and inaccurate evaluation in existing literature value evaluation methods, achieving an objective, comprehensive, and accurate evaluation of literature value.

[0055] In the above text, refer to Figure 1 A multi-dimensional quantitative method for evaluating the value of documents according to embodiments of the present invention is described in detail. Next, reference will be made to... Figure 2 This invention describes a multi-dimensional quantitative document value evaluation device according to an embodiment of the present invention.

[0056] A multi-dimensional quantitative document value evaluation device according to an embodiment of the present invention addresses the technical problems of strong subjectivity and incomplete and inaccurate evaluation in existing document value evaluation methods, achieving the technical effect of objectively, comprehensively, and accurately evaluating document value. The multi-dimensional quantitative document value evaluation device includes: a feature extraction module 10, a benchmark database construction module 20, a dynamic verification module 30, a dimension mapping and quantification module 40, and a document value fusion evaluation module 50.

[0057] The feature extraction module 10 is used to extract features from the target document in preset dimensions using a large model to obtain multi-dimensional document information. The preset dimensions include at least technical concepts, core technologies, public coverage, and differentiated increments. The benchmark library construction module 20 is used to construct benchmark libraries for each preset dimension based on the field to which the target document belongs. The dynamic verification module 30 is used to introduce a two-way adversarial verification mechanism to dynamically verify the mapping and quantification relationship between feature extraction and benchmark libraries. The dimension mapping and quantification module 40 is used to perform dimension mapping and quantification on the document information extracted in each dimension (technical concepts, core technologies, public coverage, and differentiated increments) using the benchmark libraries for each dimension after dynamic verification, and obtain the quantification results for each dimension. The document value fusion and evaluation module 50 is used to perform knowledge structure entropy change calculation based on the quantification results of each dimension, quantify the knowledge structure entropy change value, and perform document value fusion and evaluation based on the quantification results of each dimension and the knowledge structure entropy change value to generate a document value evaluation result.

[0058] The detailed configuration of the dimension mapping quantization module 40 is explained below: As mentioned above, the module utilizes the benchmark libraries for each dimension to perform dimension mapping quantization on the literature information extracted from the dimensions of technical concept, core technology, public coverage, and differentiated increment. The dimension mapping quantization module 40 may further include: a quantization mapping relationship establishment unit for establishing a quantization mapping relationship with the corresponding benchmark library according to the quantization calculation formula for each dimension; a quantization module construction unit for constructing quantization modules for each dimension according to the quantization mapping relationship for each dimension, wherein the quantization module embeds the corresponding quantization calculation formula and assigns a unique dimension identifier to each quantization module; and a calculation processing unit for mapping the corresponding literature information to the quantization module of the corresponding dimension based on the dimension identifier, combining the data in the benchmark library, performing calculation processing through the quantization calculation formula, and outputting the quantization results for each dimension.

[0059] Specifically, according to the quantization calculation formulas for each dimension, a quantization mapping relationship is established with the corresponding benchmark library. The quantization mapping relationship establishment unit may further include: a calculation mapping relationship establishment subunit, which is used to parse the calculation parameters according to the quantization calculation formulas, identify the calculation input data and calculation methods, and establish the calculation mapping relationship between each dimension and the benchmark library according to the calculation mapping relationship between the calculation parameters and the benchmark library; and a quantization mapping framework construction subunit, which is used to accurately pair the quantization calculation formulas with the corresponding benchmark library data according to the calculation mapping relationship, and construct a quantization mapping framework, wherein the quantization calculation formulas are quantized and mapped with the corresponding calculation benchmark library.

[0060] Before establishing the quantization mapping relationship with the corresponding benchmark library according to the quantization calculation formula of each dimension, the dimension mapping quantization module 40 may further include: a sample dataset construction unit for constructing a sample dataset according to the aforementioned technical concept, core technology, public coverage, and differentiated incremental preset dimensions; a quantization calculation formula fitting unit for fitting quantization calculation formulas based on the sample dataset, wherein a positive sample set and a negative sample set are established, and both the positive sample set and the negative sample set include a training set and a test set; and a quantization calculation formula acquisition unit for training the quantization calculation relationship of each preset dimension using the training sets of the positive sample set and the negative sample set respectively, and then verifying the trained quantization parameters through the test sets of the positive sample set and the negative sample set until the bidirectional convergence target is reached, thereby obtaining the quantization calculation formula of each preset dimension.

[0061] The quantitative calculation formula acquisition unit may further include: a subunit for acquiring the quantitative calculation formula of the technical idea dimension, used to quantify the technical idea sentence data in the literature, and to calculate the proportion of the quantified technical idea sentence data with a preset quantity threshold to obtain the quantitative proportion of the idea sentences; segmenting the benchmark library to obtain a literature idea benchmark library and an unsolved problem benchmark library; quantifying the mean cosine similarity between the technical idea sentences and the literature idea benchmark library to obtain a first similarity; quantifying the mean cosine similarity between the technical idea sentences and the unsolved problem benchmark library to obtain a second similarity; performing a weighted calculation on the first similarity and the second similarity, and using the quantitative proportion of the idea sentences as a coefficient to multiply and quantify the weighted result to obtain a comprehensive evaluation result of the technical idea dimension; and normalizing and adjusting the comprehensive evaluation result to obtain the final quantitative evaluation result of the technical idea dimension.

[0062] The quantitative calculation formula acquisition unit may further include: a core technology dimension quantitative calculation formula acquisition subunit for quantifying technical innovation points including the technical contribution characteristics of the literature, and calculating the proportion of contribution points by using the number of technical innovation points and a preset number threshold; quantifying the average cosine similarity between the technical innovation points and the innovation features in the benchmark library of the relevant field to obtain the intra-library similarity; quantifying the average cosine similarity between the technical innovation points to obtain the intra-point similarity; weighting the intra-library similarity and the intra-point similarity, and then multiplying the contribution point quantitative proportion as a coefficient with the weighted calculation result, and normalizing the calculation result to obtain the quantitative evaluation result of the core technology dimension.

[0063] The quantitative calculation formula acquisition unit may further include: a subunit for acquiring the quantitative calculation formula of the public coverage dimension, used to segment the benchmark library of the relevant field according to the public disclosure of technical theory, public disclosure of industry application, and project objectives, to obtain a theoretical technical benchmark library, an industry technical application benchmark library, and a project objective benchmark library; quantifying the technical concept sentence combined with the technical innovation point sentence and the mean cosine similarity of the theoretical technical benchmark library to obtain a third similarity, the higher the third similarity, the higher the public coverage; quantifying the technical concept sentence combined with the technical innovation point sentence and the mean cosine similarity of the industry technical application benchmark library to obtain a fourth similarity; quantifying the technical concept sentence combined with the technical innovation point sentence and the mean cosine similarity of the project objective benchmark library to obtain a fifth similarity; weighting the third similarity, the fourth similarity, and the fifth similarity to obtain a similarity weighted result, and calculating the undisclosed coverage based on the difference between the similarity weighted result and 1; weighting the second similarity and the undisclosed coverage, and normalizing the calculation result to obtain the quantitative evaluation result of the public coverage dimension.

[0064] The quantitative calculation formula acquisition unit may further include: a differential incremental dimension quantitative calculation formula acquisition subunit, which is used to quantify the cosine similarity between the technological innovation point and the mean vector of all innovation features in the benchmark library of the corresponding field, to obtain the technological innovation similarity, wherein the lower the technological innovation similarity, the greater the differentiation; through a large model, based on the technological innovation point and all innovation features in the benchmark library of the corresponding field, to identify and extract breakthrough improvement technology sentences; based on the number of identified and extracted breakthrough improvement technology sentences and a preset incremental threshold, to obtain the incremental breakthrough feature coefficient; and through the formula: S4=(1-technical innovation similarity)×(1+incremental breakthrough feature coefficient)×12.5, to calculate the quantitative evaluation result of the differential incremental dimension.

[0065] The detailed description of the specific configuration of the document value fusion evaluation module 50 is as follows: As mentioned above, based on the quantitative results of each dimension, the knowledge structure entropy change calculation is performed to quantify the knowledge structure entropy change value. The document value fusion evaluation module 50 may further include: a concept co-occurrence network construction unit for constructing a domain knowledge structure state representation for the time period before the publication of the target document, extracting the core document set of the target document within the domain for the time period, and constructing a concept co-occurrence network; a reconstruction degree evaluation unit for simulating the knowledge structure reorganization process after the introduction of the target document, wherein the concept set of the target document is injected into the concept co-occurrence network to form a new concept co-occurrence network, the weights of all edges of the concept co-occurrence network are recalculated based on the content similarity of the document, the new technology path branches generated by the introduction of the concept set are identified, and an extended path set is constructed; the reconstruction degree of the target document on the common knowledge structure is evaluated based on the extended path set; and a collaborative weighted fusion unit for calculating multi-dimensional information entropy change components according to the knowledge structure reconstruction degree, and performing collaborative weighted fusion of the entropy change components based on the quantitative results of each dimension to calculate the final knowledge structure entropy change value.

[0066] The multi-dimensional quantitative document value evaluation device provided in this embodiment of the invention can execute the multi-dimensional quantitative document value evaluation method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0067] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A multi-dimensional quantitative method for evaluating the value of documents, characterized in that, include: By extracting features from target documents in a pre-defined dimension using a large model, multi-dimensional document information can be obtained. The pre-defined dimensions include at least technical concepts, core technologies, public coverage, and differentiated increments. Based on the field to which the target documents belong, a benchmark library is constructed for each dimension according to preset dimensions; A two-way adversarial verification mechanism is introduced to dynamically verify the mapping and quantization relationship between feature extraction and the benchmark library; After dynamic verification, the literature information extracted from each dimension of the technical concept, core technology, public coverage, and differentiated increment is quantified by using the benchmark library of each dimension to obtain the quantification results of each dimension. Based on the quantification results of each dimension, knowledge structure entropy change calculation is performed to quantify the knowledge structure entropy change value. Then, based on the quantification results of each dimension and the knowledge structure entropy change value, a literature value fusion evaluation is conducted to generate a literature value evaluation result, including: Construct a domain knowledge structure state representation for the time period prior to the publication of the target literature, extract the core literature set of the target literature within the domain for the time period, and construct a concept co-occurrence network; The process of knowledge structure reorganization after the introduction of target literature is simulated. Specifically, the concept set of the target literature is injected into the concept co-occurrence network to form a new concept co-occurrence network. The weights of all edges in the concept co-occurrence network are recalculated based on the content similarity of the literature. New technology path branches generated by the introduction of the concept set are identified, and an extended path set is constructed. The degree of reconstruction of the common knowledge structure by the target literature is evaluated based on the extended path set. The multi-dimensional information entropy change components are calculated based on the degree of knowledge structure reconstruction, and the entropy change components are synergistically weighted and fused based on the quantification results of each dimension to calculate the final knowledge structure entropy change value.

2. The multi-dimensional quantitative method for evaluating the value of documents according to claim 1, characterized in that, Using the benchmark databases for each dimension, the literature information extracted from the dimensions of technical concept, core technology, public coverage, and differentiated increment is quantified through dimensional mapping, including: Based on the quantization calculation formulas for each dimension, establish a quantization mapping relationship with the corresponding benchmark library; Based on the quantization mapping relationship of each dimension, quantization modules for each dimension are constructed. Each quantization module embeds the corresponding quantization calculation formula and assigns a unique dimension identifier to each quantization module. Based on the dimension identifier, the corresponding literature information is mapped to the corresponding dimension quantization module. Combined with the data in the benchmark library, the quantization calculation formula is used to perform calculations and output the quantization results for each dimension.

3. The multi-dimensional quantitative method for evaluating the value of documents according to claim 2, characterized in that, Based on the quantization calculation formulas for each dimension, establish a quantization mapping relationship with the corresponding benchmark library, including: The calculation parameters are analyzed according to the quantization calculation formula, the calculation input data and calculation method are identified, and the calculation mapping relationship between each dimension and the benchmark library is established according to the calculation mapping relationship between the calculation parameters and the benchmark library. According to the aforementioned operational mapping relationship, the quantization operation formula is precisely paired with the corresponding benchmark library data to construct a quantization mapping framework, wherein the quantization operation formula is quantized and mapped to the corresponding operational benchmark library.

4. The multi-dimensional quantitative method for evaluating the value of documents according to claim 2, characterized in that, Before establishing the quantization mapping relationship with the corresponding benchmark library according to the quantization calculation formulas for each dimension, the following steps are included: A sample dataset is constructed based on the aforementioned technical concept, core technologies, public coverage, and differentiated incremental preset dimensions; The quantization calculation formula is fitted based on the sample dataset, wherein a positive sample set and a negative sample set are established, and both the positive sample set and the negative sample set include a training set and a test set. The quantization operation relationships for each preset dimension are trained using the training sets of the positive and negative sample sets, respectively. The trained quantization parameters are then verified using the test sets of the positive and negative sample sets until the bidirectional convergence target is achieved, thereby obtaining the quantization operation formulas for each preset dimension.

5. The multi-dimensional quantitative method for evaluating the value of documents according to claim 4, characterized in that, The quantitative calculation formulas for the corresponding dimensions of the technical concept include: The data of technical idea sentences in the literature are quantified, and the proportion of the quantified technical idea sentences is calculated by comparing the number of quantified technical idea sentences with a preset threshold. The benchmark database is segmented to obtain a literature concept benchmark database and an unresolved problem benchmark database; The first similarity is obtained by quantifying the average cosine similarity between the technical conception sentences and the literature conception benchmark library; The second similarity is obtained by quantifying the average cosine similarity between the technical conception sentences and the unsolved benchmark library; The first similarity and the second similarity are weighted and calculated, and the weighted result is multiplied and quantified using the quantified proportion of the conceptual sentence as a coefficient to obtain a comprehensive evaluation result of the technical concept dimension. The comprehensive evaluation results are normalized and adjusted to obtain the final quantitative evaluation results for the technical concept dimension.

6. The multi-dimensional quantitative method for evaluating the value of documents according to claim 5, characterized in that, The quantization formulas for the corresponding dimensions of the core technology include: The document quantifies the technical innovation points that contribute to the document's technical features, and calculates the proportion of contribution points by comparing the number of technical innovation points with a preset threshold. The average cosine similarity between the stated technological innovation points and the innovation features in the benchmark library of the relevant field is quantified to obtain the intra-library similarity. The mean cosine similarity between the points of technological innovation is quantified to obtain the intra-point similarity; The similarity within the library and the similarity within each point are weighted and calculated. Then, the quantified proportion of the contributing sentence is used as a coefficient and multiplied with the weighted calculation result. The calculation result is then normalized and adjusted to obtain the quantitative evaluation result of the core technology dimension.

7. The multi-dimensional quantitative method for evaluating the value of documents according to claim 6, characterized in that, The quantification formula for the dimension corresponding to the public coverage includes: The benchmark library in the relevant field is divided into a theoretical technical benchmark library, an industry technical application benchmark library, and a project objective benchmark library according to the public disclosure of technical theories, the public disclosure of industry applications, and the project objective benchmark library. The third similarity is obtained by quantifying the technical concept sentence, combining it with the technical innovation sentence, and comparing it with the average cosine similarity of the theoretical technical benchmark library. The higher the third similarity, the higher the disclosure coverage. The fourth similarity is obtained by quantifying the technical concept sentence and the technical innovation point sentence with the mean cosine similarity of the industry technology application benchmark library; The fifth similarity score is obtained by quantifying the technical concept sentence, the technical innovation point sentence, and the mean cosine similarity with the project target benchmark library. The third similarity, the fourth similarity, and the fifth similarity are weighted to obtain a similarity weighted result, and the undisclosed coverage is calculated based on the difference between the similarity weighted result and 1. The second similarity and the undisclosed coverage are weighted and calculated, and the calculation results are normalized to obtain a quantitative evaluation result of the public coverage dimension.

8. The multi-dimensional quantitative method for evaluating the value of documents according to claim 6, characterized in that, The quantification formulas for the dimensions corresponding to the differential increments include: The cosine similarity between the technological innovation points and the mean vectors of all innovation features in the benchmark library of the relevant field is quantified to obtain the technological innovation similarity. The lower the technological innovation similarity, the greater the difference. By using a large model, the technological innovation points are compared and identified with all innovation features in the benchmark library of the relevant field to extract breakthrough improvement technology points; The incremental breakthrough feature coefficient is obtained by comparing the number of sentences identified and extracted by the breakthrough improved technology with a preset incremental threshold. The quantitative evaluation results of the differentiated incremental dimension are obtained by using the formula: S4=(1-similarity of technological innovation)×(1+incremental breakthrough feature coefficient)×12.

5.

9. A multi-dimensional quantitative document value evaluation device, characterized in that, The apparatus is used to implement the multi-dimensional quantitative document value evaluation method according to any one of claims 1-8, the apparatus comprising: The feature extraction module is used to extract features from target documents in preset dimensions using a large model to obtain multi-dimensional document information. The preset dimensions include at least technical concepts, core technologies, public coverage, and differentiated increments. The benchmark library construction module is used to construct benchmark libraries for each dimension according to preset dimensions based on the field to which the target document belongs; The dynamic verification module is used to introduce a two-way adversarial verification mechanism to dynamically verify the mapping and quantization relationship between feature extraction and the benchmark library. The dimension mapping and quantification module is used to perform dimension mapping and quantification on the literature information extracted from each dimension of the technical concept, core technology, public coverage, and differentiated increment after dynamic verification, and obtain the quantification results of each dimension. The document value fusion evaluation module is used to perform knowledge structure entropy change calculation based on the quantitative results of each dimension, quantify the knowledge structure entropy change value, and perform document value fusion evaluation based on the quantitative results of each dimension and the knowledge structure entropy change value to generate document value evaluation results. The document value fusion evaluation module further includes: a concept co-occurrence network construction unit for constructing a domain knowledge structure state representation for the period before the target document is published, extracting the core document set of the target document within the domain for that period, and constructing a concept co-occurrence network; a reconstruction degree evaluation unit for simulating the knowledge structure reorganization process after the introduction of the target document, wherein the concept set of the target document is injected into the concept co-occurrence network to form a new concept co-occurrence network, the weights of all edges of the concept co-occurrence network are recalculated based on the content similarity of the documents, new technology path branches generated by the introduction of the concept set are identified, and an extended path set is constructed; the reconstruction degree of the target document on the common knowledge structure is evaluated based on the extended path set; and a collaborative weighted fusion unit for calculating multi-dimensional information entropy change components according to the degree of knowledge structure reconstruction, and performing collaborative weighted fusion of the entropy change components based on the quantitative results of each dimension to calculate the final knowledge structure entropy change value.

Images