A domain-adaptive method and system for assessing the value of scientific and technological intelligence

Abstract

This application discloses a domain-adaptive method and system for evaluating the value of scientific and technological intelligence. The method includes the following steps: acquiring scientific and technological intelligence data to be analyzed; preprocessing the data to be analyzed, extracting multi-dimensional features, and generating a domain feature vector; inputting the domain feature vector into a domain automatic classifier to determine the target scientific and technological domain to which the data belongs; dynamically loading a domain feature library and an association rule library corresponding to the target scientific and technological domain; performing multi-level value quantification evaluation on the data to be analyzed based on the loaded domain feature library and the association rule library, wherein the multi-level value quantification evaluation includes surface indicator calculation, domain depth analysis, and cross-domain association value evaluation; and fusing the results of the multi-level value quantification evaluation to generate a comprehensive value score for the data to be analyzed.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a field-adaptive method and system for assessing the value of scientific and technological intelligence. Background Technology

[0002] With increasingly fierce global technological competition, accurately and efficiently extracting high-value intelligence from massive amounts of scientific literature, patents, and technical reports has become a key support for corporate R&D decisions, government science and technology planning, and investment institutions' technology deployment. However, the inherent structure, value dimensions, and evaluation logic of data in different technological fields (such as biomedicine, semiconductors, new energy, and artificial intelligence) differ significantly. For example, when evaluating the value of patents in the biomedicine field, core indicators often involve the clinical trial stage, target novelty, and breadth of indications; while when evaluating the technological value in the semiconductor field, attention needs to be paid to process nodes, device structures, material systems, and performance parameters. These differences between fields pose a significant challenge to the accurate value assessment of interdisciplinary scientific and technological intelligence.

[0003] Currently, most existing algorithms for assessing the value of scientific and technological intelligence are of general design. These algorithms typically employ a unified indicator system to quantify the value of various types of scientific and technological intelligence, mainly relying on general indicators such as citation frequency, publication time, journal impact factor, and patent family size. Existing algorithms for assessing the value of scientific and technological intelligence generally suffer from problems such as coarse analytical granularity, poor domain adaptability, and limited evaluation dimensions in interdisciplinary scenarios, making it difficult to meet the growing demand for refined and intelligent scientific and technological intelligence analysis. Summary of the Invention

[0004] The first aspect of this application discloses a domain-adaptive method for evaluating the value of scientific and technological intelligence, comprising the following steps: acquiring scientific and technological intelligence data to be analyzed; preprocessing the scientific and technological intelligence data to be analyzed, extracting multi-dimensional features of the data to be analyzed, and generating a domain feature vector; inputting the domain feature vector into a domain automatic classifier to determine the target scientific and technological domain to which the data to be analyzed belongs; dynamically loading a domain feature library and an association rule library corresponding to the target scientific and technological domain; performing a multi-level value quantification evaluation on the data to be analyzed based on the loaded domain feature library and the association rule library, wherein the multi-level value quantification evaluation includes surface indicator calculation, domain depth analysis, and cross-domain association value evaluation; and fusing the results of the multi-level value quantification evaluation to generate a comprehensive value score for the data to be analyzed.

[0005] Based on the domain-adaptive scientific and technological intelligence value assessment method proposed in this application, an automatic domain classifier is used to determine the target scientific and technological domain to which the intelligence data to be analyzed belongs. A feature library and association rule library corresponding to that domain are dynamically loaded, achieving precise adaptation of the assessment index system to the domain to which the intelligence belongs. This solves the technical problem of the "one-size-fits-all" assessment standards of traditional general algorithms across different domains. Furthermore, by setting up a three-layer quantitative architecture of surface index calculation, domain-deep analysis, and cross-domain association value assessment, general citation influence, domain-specific weights, and cross-domain potential application value are organically integrated. This ensures that the final comprehensive value score can comprehensively and accurately reflect the true value of the intelligence within its domain, significantly improving the accuracy and comprehensiveness of interdisciplinary scientific and technological intelligence value assessment and providing reliable data support for scientific and technological decision-making.

[0006] In the method disclosed in this application, the multi-dimensional features include at least one of the following: a technical indicator parameter vector, obtained by standardizing and transforming specific types of parameters in the scientific and technological intelligence data to be analyzed; a text semantic vector, obtained by extracting the distribution of high-frequency keywords from the summary of the scientific and technological intelligence data to be analyzed, generating a semantic vector, and reducing its dimensionality; and a cooperation network relationship vector, constructed by statistically analyzing the number of types of cooperative institutions and the number of cross-domain cooperations.

[0007] According to the method disclosed in this application, multi-dimensional features of the scientific and technological intelligence data to be analyzed are extracted from three dimensions: technical indicator parameters, textual semantics, and cooperative network relationships. This forms a domain feature vector that comprehensively represents the characteristics of the field to which the scientific and technological intelligence belongs. Specifically, the technical indicator parameter vector captures the domain-specific numerical technical characteristics (such as process nanometers in the semiconductor field and energy density in the new energy field); the textual semantic vector reflects the distribution of technical keywords and thematic clustering characteristics of the field; and the cooperative network relationship component reflects the domain differences at the social network level, such as industry-academia-research cooperation models. The organic combination of these three elements significantly improves the accuracy of subsequent domain classification, thereby enabling the dynamic loading of the correct domain feature library and rule library.

[0008] In the method disclosed in this application, the domain automatic classifier adopts a three-layer structure, including: a feature extraction layer, used to perform feature enhancement processing on the domain feature vector and output a dense feature vector of fixed dimensions; a domain classification layer, used to receive the dense feature vector, combine it with the distribution of high-frequency keywords, calculate the probability distribution of the scientific and technological intelligence data to be analyzed belonging to each preset scientific and technological domain, and select the category with a probability higher than a preset threshold as the preliminary label; and a weight generation layer, used to output the evaluation dimension weight combination of the corresponding domain according to the preliminary label.

[0009] According to the method disclosed in this application, the domain automatic classifier achieves an automated process from feature enhancement and domain discrimination to evaluation weight output through a three-layer architecture design consisting of a feature extraction layer, a domain classification layer, and a weight generation layer. Specifically, the feature extraction layer enhances the robustness and expressive power of input features by adding noise and performing feature crossing; the domain classification layer effectively addresses the boundary ambiguity problem of cross-disciplinary intelligence by calculating probability distributions and setting threshold judgment mechanisms; and the weight generation layer directly outputs a combination of evaluation dimension weights that match the domain, providing domain-adaptive weight configurations for subsequent multi-layer value quantification and avoiding the subjectivity and inefficiency of manually setting evaluation indicators.

[0010] In the method disclosed in this application, the weight generation layer adopts a domain adversarial training mechanism to optimize the robustness of feature representation through the game between the generator and the discriminator.

[0011] According to the method disclosed in this application, a domain adversarial training mechanism is introduced into the weight generation layer. The generator attempts to generate features that can confuse domain discrimination, and the discriminator attempts to distinguish domain categories. This game process prompts the feature representation to learn a domain-independent robust expression, thereby solving the problem of decreased model generalization ability caused by the large differences in data distribution across different domains.

[0012] In the method disclosed in this application, the surface index calculation is used to calculate the general value index of the scientific and technological intelligence data to be analyzed, the general value index including citation frequency and / or publication time; the domain depth analysis includes: calling the loaded domain feature library to extract domain-specific entities and relationships from the scientific and technological intelligence data to be analyzed; calling the loaded association rule library to calculate domain weights; generating a domain depth analysis score based on the domain-specific entities and relationships and the domain weights; the cross-domain association value assessment includes: based on a knowledge graph, discovering the potential application value of the technologies involved in the scientific and technological intelligence data to be analyzed in other scientific and technological fields; and adding cross-domain points based on the discovered potential application value.

[0013] According to the method disclosed in this application, a comprehensive and multi-level assessment of the value of scientific and technological intelligence is achieved through the collaborative work of a three-layer value quantification evaluation system. The surface-level indicator calculation utilizes common indicators such as citation frequency and publication time, providing a basic value dimension comparable across domains. Domain-level in-depth analysis extracts specific entities and relationships by calling a domain feature library and calculates domain weights by combining them with an association rule library, enabling accurate identification of implicit high-value signals within the domain. Cross-domain association value assessment, based on knowledge graph discovery technology, identifies the potential application value in other fields and adds cross-domain bonuses, effectively solving the technical problem that traditional methods easily overlook interdisciplinary application potential, leading to lower evaluation values.

[0014] In the method disclosed in this application, the domain feature library contains a value feature dictionary specific to each scientific and technological domain, and the value feature dictionary includes domain-specific entity type definitions and their corresponding value assessment dimensions.

[0015] According to the method disclosed in this application, the domain feature library contains a dictionary of value features specific to each technology field and a definition of domain-specific entity types. This enables the system to accurately identify the unique high-value intelligence features of each field during the in-depth analysis stage, thereby giving reasonable value bonuses to technical intelligence with breakthrough significance in the field during the evaluation process, and avoiding the problem of underestimation of core technology breakthroughs in the field when evaluating general indicators.

[0016] The second aspect of this application discloses a domain-adaptive science and technology intelligence value assessment system, comprising: a data acquisition module for acquiring science and technology intelligence data to be analyzed; a feature extraction module for preprocessing the science and technology intelligence data to be analyzed, extracting multi-dimensional features of the science and technology intelligence data to be analyzed, and generating a domain feature vector; a domain classification module for inputting the domain feature vector into a domain automatic classifier to determine the target science and technology domain to which the science and technology intelligence data to be analyzed belongs; a dynamic loading module for dynamically loading a domain feature library and an association rule library corresponding to the target science and technology domain; a multi-layer value quantification assessment module for performing multi-layer value quantification assessment on the science and technology intelligence data to be analyzed based on the loaded domain feature library and the association rule library, wherein the multi-layer value quantification assessment includes a surface index calculation layer, a domain depth analysis layer, and a cross-domain association value assessment layer; and a fusion output module for fusing the results of the multi-layer value quantification assessment to generate a comprehensive value score for the science and technology intelligence data to be analyzed.

[0017] The domain-adaptive science and technology intelligence value assessment system of this application achieves fully automated assessment from raw intelligence data input to comprehensive value score output through the coordinated operation of data acquisition module, feature extraction module, domain classification module, dynamic loading module, multi-layer value quantification assessment module and fusion output module. It solves the technical problem of "one-size-fits-all" assessment standards in traditional general algorithms across different domains.

[0018] The system in this application also includes: a domain feature library construction module, which is used to extract entity-relationship-entity triples from multi-source heterogeneous data, perform entity alignment and fusion, and use association rule mining algorithms to mine frequent itemsets from historical data to generate association rules, so as to construct domain feature libraries and association rule libraries corresponding to various scientific and technological fields.

[0019] According to the system of this application, by setting up a domain feature library construction module, the entire process of automatically extracting entity-relationship-entity triples from multi-source heterogeneous data, performing entity alignment and fusion, and generating domain association rules using association rule mining algorithms is realized.

[0020] In the system of this application, the domain feature library construction module includes: a multi-source data aggregation unit, used to collect multi-source heterogeneous data and perform word segmentation, stop word removal, and part-of-speech tagging; a knowledge extraction unit, used to extract entity-relation-entity triples from core documents by referring to the domain ontology library and using text mining tools; an entity alignment and fusion unit, used to perform strong rule deduplication using unique identifiers, and weak rule alignment by calculating edit distance, co-occurrence frequency, and semantic similarity, and to resolve entity attribute conflicts; an association rule mining unit, used to mine frequent itemsets from historical data and generate association rules; and a dynamic update unit, used to periodically recalculate entity weights and association rules based on real-time data streams.

[0021] According to the system in this application, the domain feature library construction module achieves full lifecycle management of the domain feature library from construction to continuous evolution through a collaborative architecture of a multi-source data aggregation unit, a knowledge extraction unit, an entity alignment and fusion unit, an association rule mining unit, and a dynamic update unit. Specifically, the entity alignment and fusion unit effectively solves the data fusion challenges of homonyms, synonyms, and attribute conflicts in multi-source heterogeneous data by combining strong rules (unique identifiers) with weak rules (edit distance, co-occurrence frequency, semantic similarity). The dynamic update unit periodically recalculates entity weights and association rules based on real-time data streams, ensuring that the feature library and rule library can be updated in a timely manner with technological development, always reflecting the latest research hotspots and value assessment standards in various fields, and guaranteeing the timeliness and accuracy of the assessment results.

[0022] In the system of this application, the domain classification module uses a domain automatic classifier that is trained under supervision using the cross-entropy loss function and the hyperparameters are adjusted using a validation set.

[0023] According to the system of this application, by employing the cross-entropy loss function to supervise the training of the domain automatic classifier and by using an independent validation set to adjust the hyperparameters, the classifier is ensured to have good generalization performance and classification accuracy. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating an adaptive scientific and technological intelligence value assessment method for the field covered by this application. Detailed Implementation

[0025] The present application will be further described below with reference to specific embodiments and accompanying drawings. It is to be understood that the illustrative embodiments of this disclosure are merely for explaining the present application and not for limiting it. Furthermore, for ease of description, the accompanying drawings show only the parts relevant to the present application, and not all of the structures or processes.

[0026] The following specific embodiments illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Although the description of this application is presented in conjunction with preferred embodiments, this does not mean that the features of this invention are limited to this embodiment. On the contrary, the purpose of describing the invention in conjunction with embodiments is to cover other options or modifications that may be derived based on the claims of this application. To provide a thorough understanding of this application, many specific details will be included in the following description. This application may also be implemented without using these details. Furthermore, to avoid confusion or obscuring the focus of this application, some specific details will be omitted in the description. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0027] Unless the context otherwise specifies, the terms “contains,” “has,” and “includes” are synonyms. The phrase “A / B” means “A or B.” The phrase “A and / or B” means “(A and B) or (A or B).”

[0028] It should be understood that although terms such as "first," "second," etc., may be used herein to describe various components, units, or data, these components, units, or data should not be limited by these terms. These terms are used merely to distinguish one feature from another. For example, without departing from the scope of the exemplary embodiments, a first feature may be referred to as a second feature, and similarly, a second feature may be referred to as a first feature.

[0029] It should be understood that although directional terms such as "up," "down," "left," and "right" may be used here to describe the positional relationship between the various components, these directional terms are only used for ease of understanding to indicate the directions in the accompanying drawings and cannot be used to limit the scope of protection of this application.

[0030] It should be noted that in this specification, similar reference numerals and letters in the accompanying drawings indicate similar items. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0031] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0032] Figure 1This is a flowchart illustrating a scientific and technological information value assessment method adapted to the field of this application. Figure 1 As shown, the domain-adaptive scientific and technological intelligence value assessment method provided in this application includes the following steps:

[0033] Step S1: Obtain the science and technology intelligence data to be analyzed.

[0034] The scientific and technological intelligence data to be analyzed includes, but is not limited to, patent documents, academic papers, technical reports, clinical trial reports, funding project reports, and technical standard documents. This data can originate from multiple heterogeneous data sources, such as patent databases, academic literature databases, clinical trial registration platforms, and science and technology project management systems.

[0035] Step S2: Preprocess the scientific and technological intelligence data to be analyzed, extract multi-dimensional features of the data, and generate domain feature vectors.

[0036] Preprocessing includes: text cleaning (removing HTML tags, special characters, redundant whitespace, etc.) of the scientific and technological intelligence data to be analyzed, word segmentation, stop word removal, and part-of-speech tagging, converting unstructured text into structured data tables and unstructured data mapping tables (key-value pair format).

[0037] Multidimensional features include at least one of the following:

[0038] (1) Technical indicator parameter vector: Extract specific types of numerical technical parameters from the scientific and technological intelligence data to be analyzed. For example, for intelligence data in the semiconductor field, extract parameters such as process nanometer number (e.g., 3nm, 5nm, 7nm), leakage current reduction ratio (e.g., 40%), and operating frequency (e.g., 5GHz); for intelligence data in the new energy field, extract parameters such as energy density (e.g., 300Wh / kg), cycle life (e.g., 2000 cycles), and ionic conductivity (e.g., 10⁻³S / cm); for intelligence data in the biomedical field, extract parameters such as clinical stage (Phase I, Phase II, Phase III), number of targets, and number of indications. Standardize the extracted numerical parameters (e.g., Z-score standardization or Min-Max normalization) and perform logarithmic transformation to form a numerical component vector.

[0039] (2) Text Semantic Vector: Extract the distribution of high-frequency technical keywords from the titles and abstracts of the scientific and technological intelligence data to be analyzed. Specifically, use the TF-IDF (Term Frequency-Inverse Document Frequency) algorithm or the TextRank algorithm to extract the top N high-frequency keywords (e.g., N=5), and count the frequency of each keyword to form a keyword frequency distribution vector. Furthermore, for patent documents, the number of claims can be combined, or for papers, the number of references can be used to generate auxiliary semantic features. The above keyword frequency distribution vector is reduced to a preset dimension using dimensionality reduction methods such as PCA (Principal Component Analysis) or t-SNE to obtain text feature components.

[0040] (3) Collaboration network relationship vector: Statistically count the number of types of collaborating institutions (such as universities, research institutes, enterprises, hospitals, etc.) and the number of cross-domain collaborations (such as joint R&D between biopharmaceutical companies and artificial intelligence companies) involved in the scientific and technological intelligence data to be analyzed, and construct relationship feature components. Specifically, the collaboration network relationship vector may include feature dimensions such as the total number of collaborating institutions and the number of cross-domain collaborations.

[0041] The aforementioned technical indicator parameter vectors, text semantic vectors, and / or cooperation network relationship vectors are concatenated to form a fixed-dimensional domain feature vector. For example, when the technical indicator parameter vector is 32-dimensional, the text semantic vector is 128-dimensional, and the cooperation network relationship vector is 16-dimensional, the concatenated vector will be 176-dimensional. Missing dimensions are filled using zero-padding or mean-padding strategies to ensure that all input samples have a uniform vector dimension.

[0042] Step S3: Input the domain feature vector into the domain automatic classifier to determine the target technology domain to which the technology intelligence data to be analyzed belongs.

[0043] The domain automatic classifier adopts a three-layer structure, specifically including:

[0044] (1) Feature Extraction Layer: Receives the domain feature vector as input and performs feature enhancement processing on the input vector. Feature enhancement processing includes, but is not limited to: adding Gaussian noise to the input vector (standard deviation σ can be set to 0.01~0.05) to improve robustness; performing feature crossing (such as second-order feature crossing) on ​​different dimensions of the input vector to capture the nonlinear interaction between features. Nonlinear mapping is performed through a fully connected layer to output a dense feature vector of fixed dimensions.

[0045] (2) Domain Classification Layer (Discriminator): Receives dense feature vectors and, combined with the high-frequency keyword distribution extracted in step S2, calculates the probability distribution of the scientific and technological intelligence data to be analyzed belonging to various preset scientific and technological domains using, for example, the Softmax function. Preset scientific and technological domains include, but are not limited to, 16 major scientific and technological domains: advanced materials, artificial intelligence, biomedicine, semiconductors / microelectronics, new energy / energy storage, aerospace, quantum information, intelligent manufacturing, agricultural technology, environmental science, marine engineering, nuclear technology, information security, transportation, construction engineering, and basic physics.

[0046] The determination logic is as follows: Categories with a probability value higher than a preset threshold (e.g., 70%) in the probability distribution are selected as initial labels. If multiple categories have probabilities higher than the preset threshold, the category with the highest probability is selected as the primary label, and the remaining categories are used as auxiliary labels. If the highest probability in the probability distribution does not reach the preset threshold, the scientific and technological intelligence data to be analyzed is marked as "interdisciplinary," or a manual review process is triggered for final determination by domain experts.

[0047] (3) Weight Generation Layer (Generator): Based on the preliminary labels output by the domain classification layer, output the weight combination of the evaluation dimensions for the corresponding domain. For example, when the preliminary label is "Biomedicine", the weight combination output has the following weights: "Clinical Stage" weight is 0.35, "Target Novelty" weight is 0.25, "Indication Breadth" weight is 0.20, "Citation Frequency" weight is 0.10, and "Cross-Domain Association" weight is 0.10; when the preliminary label is "Semiconductor / Microelectronics", the weight combination output has the following weights: "Process Advancement" weight is 0.30, "Device Architecture Innovation" weight is 0.25, "Performance Improvement" weight is 0.20, "Patent Family Size" weight is 0.15, and "Cross-Domain Association" weight is 0.10.

[0048] The weight generation layer employs a domain adversarial training mechanism. Specifically, during the training phase, the weight generation layer acts as a generator, attempting to generate feature representations capable of obfuscating domain discrimination; the domain classification layer acts as a discriminator, attempting to accurately distinguish between different domain categories. Through the adversarial game between the generator and the discriminator (i.e., a mini-max optimization problem), the feature extraction layer is prompted to learn domain-independent robust feature representations, thereby addressing the problem of decreased model generalization ability caused by significant differences in data distribution across different domains. The loss function for adversarial training can be expressed as:

[0049] L_total=L_classification+λ·L_adversarial

[0050] Where L_classification is the classification cross-entropy loss, L_adversarial is the adversarial loss, and λ is the balance coefficient (e.g., λ=0.1).

[0051] The domain-specific automatic classifier is trained under supervised conditions using a cross-entropy loss function. Specifically, historical labeled data (i.e., intelligence data labeled by domain experts with their respective scientific and technological domains) is used as the training set, and supervised learning is performed using the cross-entropy function as the loss function. During training, hyperparameters are adjusted using independent validation sets (e.g., training, validation, and test sets are divided in an 8:1:1 ratio).

[0052] Step S4: Based on the target technology field, dynamically load the domain feature library and association rule library corresponding to the target technology field.

[0053] The system pre-constructs domain-specific feature libraries and association rule libraries for each preset technology field. After step S3 determines the target technology field to which the technology intelligence data to be analyzed belongs, the system dynamically loads the domain feature library and association rule library corresponding to the target technology field from multiple pre-stored domain feature libraries and association rule libraries into the core analysis engine.

[0054] The domain feature library contains a value feature dictionary specific to each technology domain. This dictionary includes domain-specific entity type definitions and their corresponding value assessment dimensions. A specific example is shown below:

[0055] The biomedical field feature library includes entity types such as "clinical stage" (Phase I / II / III / IV), "target" (e.g., PD-1, CD19, BCMA), "indication" (e.g., solid tumors, hematological malignancies, autoimmune diseases), "cell lines", "experimental reagents", and "disease targets". The corresponding value assessment dimensions include: clinical stage weight, target novelty score, indication breadth coefficient, and combination therapy potential.

[0056] The semiconductor field feature library includes entity types such as "process node" (e.g., 3nm, 5nm, 7nm), "device structure" (e.g., Planar, FinFET, GAA), "material system" (e.g., High-k / Metal Gate, SiGe, InGaAs), and "lithography technology" (e.g., EUV, DUV). The corresponding value assessment dimensions include: process generation weight, architectural innovation coefficient, performance improvement, and supply chain criticality.

[0057] The new energy field feature library includes entity types such as "electrolyte type" (e.g., sulfides, oxides, polymers), "electrode materials" (e.g., silicon-carbon anode, high-nickel cathode), and "performance indicators" (e.g., energy density, cycle life, rate performance). The corresponding value assessment dimensions include: practical application threshold attainment, degree of resolution of core pain points, cost advantage coefficient, and breadth of application scenarios.

[0058] The association rule base contains association rules of the form {entity A, entity B} → {relation C}, used to represent the mapping relationship between entity combinations and value judgments within a domain. A specific example is as follows:

[0059] Rules in the biopharmaceutical field: {"Specific target", "Phase II clinical trial"} → {"High commercial value"}; {"Successful Phase I clinical trial", "Good safety"} → {"Value weight + 0.3"}; {"Combination therapy regimen"} → {"Value weight + 0.2"}.

[0060] Semiconductor industry rules: {"EUV lithography", "multi-layer stacking"} → {"advanced process"}; {"node shrinkage by one generation"} → {"basic weight + 0.4"}; {"architectural innovation (FinFET to GAA)"} → {"additional weight + 0.5"}.

[0061] Rules for the new energy sector: {"Cycle life > 1000 cycles"} → {"Practical application threshold met"}; {"Interface modification", "Core pain points"} → {"High innovation score"}; {"Does not contain rare metals"} → {"Low cost bonus"}.

[0062] Step S5: Based on the loaded domain feature library and association rule library, perform multi-level value quantification assessment of the scientific and technological intelligence data to be analyzed.

[0063] Multi-level value quantification assessment includes the following three levels:

[0064] (a) Calculation of surface indicators

[0065] Surface index calculation is used to calculate the general value indicators of the scientific and technological intelligence data to be analyzed. General value indicators may include citation frequency and / or publication time, etc.

[0066] Citation frequency score: This score is calculated by taking the total number of citations of the scientific and technological intelligence data to be analyzed since its publication / publication, and then normalizing it using a logarithmic normalization method. For example, citation frequency score = log(1 + number of citations) / log(1 + maximum number of citations in the field).

[0067] Publication Time Score: The timeliness score is calculated based on the difference between the publication time / application time of the scientific and technological intelligence data to be analyzed and the current time. For example, Publication Time Score = max(0, 1 - number of years / preset decay period), where the preset decay period can be set according to the characteristics of the field (e.g., 5 years for the biomedical field and 3 years for the semiconductor field).

[0068] Other general metrics may include journal impact factor (for papers), patent family size (for patents), downloads / reads, etc.

[0069] Surface index scores are calculated using general weights (e.g., citation frequency weight 0.5, publication time weight 0.3, other general index weights 0.2).

[0070] (II) In-depth analysis of the field

[0071] Domain-level in-depth analysis includes the following sub-steps:

[0072] The loaded domain feature library is invoked to extract domain-specific entities and relationships from the scientific and technological intelligence data to be analyzed. Specifically, Named Entity Recognition (NER) technology and relationship extraction models are used, referencing domain ontology libraries (such as UMLS Unified Medical Language System and STKOS Scientific Knowledge Organization System), to identify and extract "entity-relationship-entity" triples from the text of the scientific and technological intelligence data to be analyzed. For example, from a biomedical paper, {"CAR-T cells", "treatment", "solid tumor"} and {"CAR-T", "combination", "PD-1 inhibitor"} are extracted; from a semiconductor technology report, {"3nm", "adopted", "GAA structure"} and {"GAA", "dependent on", "EUV lithography"} are extracted.

[0073] The loaded association rule library is invoked, and the extracted entities and relationships are matched with the rules in the association rule library to calculate the domain weight. Specifically, all rules in the association rule library are traversed. When an entity combination in the scientific and technological intelligence data to be analyzed matches the antecedent of a rule, the consequent of that rule is triggered, and the corresponding weight is added. For example, when a biomedical paper is identified to contain "completed Phase I clinical trials" and "no serious side effects", the rule {"Phase I clinical trials successful", "good safety"} → {"value weight +0.3"} is matched, and a weight of 0.3 is added.

[0074] Based on domain-specific entities and relationships, and domain weights, a domain depth resolution score is generated. Specifically, the weighted scores of all hit rules are accumulated, and then combined with the evaluation dimension weights output from the weight generation layer in step S3 for a weighted sum to obtain the domain depth resolution score. For example:

[0075] Domain depth analysis score = Σ(weighted score of rule i × corresponding dimension weight i)

[0076] (III) Cross-domain linkage value assessment

[0077] Cross-domain association value assessment includes the following sub-steps:

[0078] Based on knowledge graphs, the potential application value of the technologies involved in the analyzed science and technology intelligence data in other science and technology fields is discovered. Specifically, the extracted entities and relationships are mapped to a pre-constructed cross-domain knowledge graph. Graph traversal algorithms (such as Breadth-First Search or random walk algorithms) are used to search for nodes in other domains that are related to the current entity. For example, for the entity "CAR-T cell therapy," the knowledge graph reveals a strong correlation with the field of "gene editing (CRISPR)" (shared technology path), and it can be transferred to the field of "autoimmune disease" treatment.

[0079] Cross-domain scoring is based on the discovered potential application value. Specifically, cross-domain scoring is calculated based on the strength of the cross-domain association (such as the weight of the associated edges) and the number of associated domains. For example:

[0080] Cross-domain bonus = Σ(Relationship strength of related domain j × Application potential coefficient of domain j)

[0081] The strength of association can be represented by the weight of the edges in the knowledge graph, and the application potential coefficient can be preset according to factors such as the market size and technological maturity of the target field.

[0082] Step S6: Integrate the results of the multi-level value quantification assessment to generate a comprehensive value score for the scientific and technological intelligence data to be analyzed.

[0083] The evaluation results from the three levels in step S5 are weighted and summed according to preset fusion weights to generate a comprehensive value score. The fusion weights can be set according to actual needs, for example:

[0084] Overall Value Score = α × Surface Indicator Score + β × Domain Depth Analysis Score + γ × Cross-Domain Relationship Value Score

[0085] Where α+β+γ=1. In a preferred embodiment, α=0.2, β=0.5, and γ=0.3, meaning the domain depth analysis score has the highest weight, reflecting the core design concept of "domain adaptation" in this application. The comprehensive value score can be normalized to a range of 0-100.

[0086] After generating the comprehensive value score, the comprehensive value score and the results of the multi-level value quantification assessment can be visualized based on the domain characteristics of the target technology field. The visualization includes at least one of the following:

[0087] (1) Technology Readiness Level (TRL) curve: applicable to fields such as biomedicine, it labels the scientific and technological intelligence data to be analyzed on the Technology Readiness Level (TRL 1~9), and intuitively shows the R&D stage where the intelligence is located (such as basic research period, pre-clinical verification period, clinical verification period, commercialization period, etc.).

[0088] (2) Technology evolution roadmap: applicable to fields such as semiconductors, it takes the time axis as the horizontal axis to show the evolution path from early technologies to current technologies and then to future technologies (such as Planar→FinFET→GAA→CFET), and marks the scientific and technological intelligence data to be analyzed as key nodes on the evolution path.

[0089] (3) Industrial chain radiation heat map: applicable to fields such as new energy and intelligent manufacturing, it takes the upstream and downstream links of the industrial chain as dimensions, and uses the shade of color to represent the radiation influence degree of the technology involved in the scientific and technological intelligence data to be analyzed in each industrial chain link.

[0090] (4) Knowledge graph differentiated path highlight map: in the cross-domain knowledge graph, it highlights the differentiated path between the scientific and technological intelligence data to be analyzed and the existing technologies, and intuitively shows its innovation points and potential application directions.

[0091] The construction method of the domain feature library and the association rule library involved in the present application will be further described below.

[0092] The construction process of the domain feature library and the association rule library includes the following steps:

[0093] Step T1: Multi-source data aggregation and structurization.

[0094] Collect multi-source heterogeneous data such as patents, papers, clinical trial reports, and fund projects. Perform word segmentation processing on the collected original data, for example, use jieba word segmentation or LAC word segmentation tool for Chinese, and use NLTK or spaCy word segmentation tool for English, perform stop word removal processing (remove words with no actual semantics such as "de", "shi", "the", "a") and part-of-speech tagging processing (tag part-of-speech labels such as nouns, verbs, adjectives). Convert unstructured text into a structured data table (including fields such as title, author, date, and classification number) and an unstructured data mapping table (storing text content such as abstract and full text in the form of key-value pairs).

[0095] Step T2: Ontology-based knowledge extraction.

[0096] Referencing domain ontology libraries (e.g., UMLS Unified Medical Language System for the biomedical field, STKOS Science and Technology Ontology for the semiconductor field, and relevant IEC standard classification systems for the new energy field), text mining tools (such as BERT-based named entity recognition models and dependency parsing-based relation extraction models) are used to extract "entity-relation-entity" triples from core literature.

[0097] Entity types include not only general entities (such as "author", "institution", "journal", "date" etc.), but also domain-specific entities.

[0098] Specific entity types in the biomedical field include: "cell lines" (such as HEK293, Jurkat), "experimental reagents" (such as antibodies, culture media), "disease targets" (such as PD-1, CD19, BCMA), "clinical stages" (such as Phase I, Phase II), and "routes of administration" (such as intravenous injection, oral administration), etc.

[0099] Specific entity types in the semiconductor field include: "process nodes" (such as 3nm, 5nm), "device structures" (such as FinFET, GAA, Nanosheet), "material systems" (such as High-k, Metal Gate, SiGe), and "process steps" (such as photolithography, etching, deposition), etc.

[0100] Specific entity types in the new energy field include: "electrolyte type" (such as sulfides and oxides), "electrode material" (such as LiCoO2 and LiFePO4), and "battery configuration" (such as pouch, cylindrical, and prismatic).

[0101] Step T3: Entity alignment and blending.

[0102] The entities extracted from the multi-source data in step T2 are aligned and merged to eliminate redundancy and conflicts. Specifically, this includes:

[0103] (1) Strong rule deduplication: Using unique identifiers such as DOI (Digital Object Identifier), ORCID (Open Researcher and Contributor ID), patent, PubMed ID, etc., entities with the same identifier are directly merged to remove duplicates.

[0104] (2) Weak rule alignment: For entities that cannot be directly matched by unique identifiers (such as Chinese and English variants, abbreviations and full names, spelling variants, etc.), the following similarity index is calculated for fuzzy matching:

[0105] Edit distance (Levenshtein Distance): Calculates the minimum number of edit operations between two entity name strings, and obtains the string similarity after normalization;

[0106] Co-occurrence frequency: The frequency at which two entities appear together in the same document; the higher the frequency, the more likely they are to point to the same entity.

[0107] Semantic similarity: Entity names are encoded into semantic vectors using pre-trained language models (such as BERT and SciBERT), and cosine similarity is calculated.

[0108] When the combined score of the above similarity indicators exceeds a preset threshold (e.g., 0.85), the two entities are determined to be the same entity and are merged.

[0109] (3) Conflict resolution: When there are attribute conflicts among the merged entities (e.g., a technology is classified as "semiconductor" in data source A and "advanced materials" in data source B), a comprehensive decision is made based on the reliability weight of the data source (e.g., the weight of authoritative databases is set to 1.0, and the weight of ordinary sources is set to 0.6) and the frequency of information occurrence to determine the final attribute value. Specifically, the final attribute value = argmax(data source reliability weight × frequency of occurrence of the attribute value).

[0110] Step T4: Association rule mining.

[0111] Use association rule mining algorithms (such as the Apriori algorithm or the FP-Growth algorithm) to mine frequent itemsets from historical data. Set a minimum support threshold (e.g., min_support=0.01, meaning that the entity combination is present in at least 1% of the literature) and a minimum confidence threshold (e.g., min_confidence=0.6, meaning that when entity A appears, the probability of entity B appearing at the same time is at least 60%).

[0112] Based on the frequent itemsets obtained from mining, association rules of the form {entity A, entity B} → {relation C} are generated.

[0113] In the biopharmaceutical field: {"specific target", "phase II clinical trial"} → {"high commercial value"} (support = 0.03, confidence = 0.78);

[0114] In the semiconductor field: {"EUV lithography", "multilayer stacking"} → {"advanced process"} (support = 0.05, confidence = 0.85);

[0115] In the field of new energy: {"sulfide electrolyte", "interface modification"} → {"solving core pain points"} (support = 0.02, confidence = 0.72).

[0116] Step T5: Dynamic update mechanism.

[0117] Establish a long-term update mechanism to recalculate entity weights and association rules regularly (e.g., monthly or quarterly) based on real-time data streams (such as newly published papers daily / weekly, newly disclosed patent applications, and newly granted patents).

[0118] New entity recognition: Identify and store new entities appearing in the new data stream;

[0119] Weight decay and update: Decay the weight of entities that have not been referenced for a long time (e.g., decay by 10% per year), and increase the weight of newly emerging high-frequency entities.

[0120] Rule update: Rerun the association rule mining algorithm, update frequent itemsets and association rules, eliminate outdated rules that no longer meet the minimum support / confidence threshold, and add new rules that meet the conditions.

[0121] Through the aforementioned dynamic update mechanism, the domain feature library and association rule library are ensured to always reflect the latest research hotspots and value assessment standards in each domain, guaranteeing the timeliness and accuracy of the assessment results.

[0122] This application also provides a domain-adaptive scientific and technological intelligence value assessment system, which specifically includes the following modules.

[0123] The data acquisition module is used to acquire scientific and technological intelligence data to be analyzed. It can obtain various types of scientific and technological intelligence data, such as patent documents, academic papers, technical reports, and clinical trial reports, from multiple heterogeneous data sources through API interfaces, file import, and other methods. The module also performs format standardization and preliminary cleaning on the acquired data.

[0124] The feature extraction module is used to preprocess the scientific and technological intelligence data to be analyzed, extracting multi-dimensional features from the data and generating domain feature vectors. Internally, the feature extraction module integrates a word segmentation engine, a TF-IDF calculator, a PCA dimensionality reduction component, and a feature concatenation tool, enabling automatic completion of the entire process from raw text to fixed-dimensional domain feature vectors.

[0125] The domain classification module is used to input domain feature vectors into the domain automatic classifier to determine the target scientific and technological domain to which the scientific and technological intelligence data to be analyzed belongs. Internally, the domain classification module deploys a pre-trained three-layer domain automatic classifier model, supporting real-time inference. When the classification confidence level falls below a preset threshold, the domain classification module automatically triggers a manual review interface or cross-disciplinary labeling logic.

[0126] The dynamic loading module is used to dynamically load the domain feature library and associated rule library corresponding to the target technology domain. The dynamic loading module maintains an index mapping table of feature libraries and rule libraries for each technology domain, and can complete the loading and initialization of the corresponding feature library and rule library after receiving the target technology domain tag.

[0127] The multi-layered value quantification assessment module is used to perform multi-layered value quantification assessments on the scientific and technological intelligence data to be analyzed, based on the loaded domain feature library and association rule library. This module comprises three sub-modules: a surface-level indicator calculation sub-module (calculating common indicators such as citation frequency and publication time), a domain-level in-depth analysis sub-module (calling the domain feature library and association rule library for entity extraction, rule matching, and weight calculation), and a cross-domain association value sub-module (discovering cross-domain associations and calculating bonuses based on knowledge graphs).

[0128] The fusion output module is used to integrate the results of multi-level value quantification assessments to generate a comprehensive value score for the scientific and technological intelligence data to be analyzed. The fusion output module is also used to visualize the comprehensive value score and the results of the multi-level value quantification assessments based on the domain characteristics of the target scientific and technological field, generating visualizations such as technology maturity curves, technology evolution roadmaps, industrial chain radiation heat maps, or knowledge graph differentiated path highlighting maps.

[0129] The adaptive scientific and technological intelligence value assessment system involved in this application may further include the following modules.

[0130] The domain feature library construction module is used to extract entity-relation-entity triples from multi-source heterogeneous data, perform entity alignment and fusion, and use association rule mining algorithms to mine frequent itemsets from historical data to generate association rules, so as to build domain feature libraries and association rule libraries corresponding to various scientific and technological fields.

[0131] The domain feature library construction module specifically includes:

[0132] The multi-source data aggregation unit is used to collect heterogeneous data from multiple sources, perform word segmentation, stop word removal and part-of-speech tagging, and convert unstructured text into structured data tables and unstructured data mapping tables.

[0133] The knowledge extraction unit is used to reference domain ontology libraries (such as UMLS, STKOS, etc.) and extract entity-relation-entity triples from core documents using text mining tools. Entity types include general entities and domain-specific entities.

[0134] The entity alignment and fusion unit is used to perform strong rule deduplication using unique identifiers (such as DOI, ORCID, and patent number), and weak rule alignment by calculating edit distance, co-occurrence frequency, and semantic similarity. It also comprehensively resolves entity attribute conflicts based on the reliability weight of the data source and the frequency of information occurrence.

[0135] The association rule mining unit is used to mine frequent itemsets from historical data and generate association rules in the form of {entity A, entity B} → {relation C}.

[0136] The dynamic update unit is used to periodically recalculate entity weights and association rules based on real-time data streams, ensuring the timeliness of the feature library and rule library.

[0137] The following uses a patent and related papers on "novel CAR-T cell therapy for solid tumors" as input data to further illustrate the complete evaluation process of the adaptive scientific and technological information value assessment method in the field involved in this application.

[0138] Original data input: The patent title contains keywords such as "CAR-T", "solid tumor" and "PD-1 knockout"; the abstract mentions "completed Phase I clinical trial" and "no serious side effects"; the collaborating institutions include a top hospital and a biotechnology company.

[0139] Automatic domain classification and dynamic loading

[0140] (1) Feature extraction: High-frequency technical keywords (CAR-T, PD-1, solid tumors, immune checkpoints, cell therapy) were extracted from the title and abstract, and domain-specific entities (cell therapy, immune checkpoints, clinical staging) were identified. Technical indicator parameters (Phase I clinical trials, safety data) were extracted, and a collaborative network relationship vector was constructed (collaborating institution type: hospital + enterprise, number of cross-domain collaborations: 1). The above features were concatenated into a domain feature vector.

[0141] (2) Classification determination: Input the domain feature vector into the domain automatic classifier. The classifier calculates that the probability of "biomedicine" is 96% (>70% threshold), and the target technology domain is determined to be "biomedicine".

[0142] (3) Dynamic loading: The system automatically loads the biomedical feature library (including entity type definitions such as clinical stage, target novelty, indication breadth, cell line, experimental reagent, disease target, etc.) and the association rule library (including rules such as: {"Phase I clinical success", "good safety"} → {"value weight + 0.3"}; {"combination drug regimen"} → {"value weight + 0.2"}, etc.).

[0143] Multi-level value quantification assessment

[0144] (1) Surface index calculation: The citation frequency is at a medium level (e.g., 15 citations within 2 years of publication), and the publication time score is relatively high (most recently published). The overall surface score is calculated as 62 / 100.

[0145] (2) In-depth domain analysis:

[0146] "Completion of Phase I clinical trial": According to biomedical regulations, it is given a high weight (compared to preclinical studies, the weight of successful Phase I clinical trial is +0.3).

[0147] "PD-1 knockout" was identified: it was matched with the "target novelty" dictionary and confirmed as a popular and groundbreaking modification technique with a target novelty score of 0.85;

[0148] Entity relation extraction: Construct knowledge graph subgraphs for "CAR-T cells - [treatment] -> solid tumor" and "CAR-T - [combination] -> PD-1 inhibitor";

[0149] The domain depth analysis score is calculated to be 88 / 100.

[0150] (3) Cross-domain association value assessment: The knowledge graph found that the technology has a strong association with the field of "gene editing (CRISPR)" (association strength 0.8), and the technology can be transferred to the field of "autoimmune disease" treatment (application potential coefficient 0.7), triggering cross-domain bonus. The cross-domain association value score is calculated as 75 / 100.

[0151] Value integration and visualization output

[0152] (1) Comprehensive score: Calculated according to the fusion weight (α=0.2, β=0.5, γ=0.3): Comprehensive value score = 0.2×62+0.5×88+0.3×75=12.4+44+22.5=78.9.

[0153] (2) Visual presentation: Generate a technology maturity curve (TRL) to mark the project as being in the "clinical validation phase"; highlight its differentiated path from existing therapies (solving the problem of penetration in solid tumors) in the knowledge graph, and show the innovation points in the form of a differentiated path highlight map.

[0154] The following uses a technical report on "3nm GAA (Gate All Around) Transistor Structure" as input data to demonstrate the complete evaluation process of the method in this application.

[0155] Original data input: The technical report mentions "Gate-All-Around", "3nm node", and "40% reduction in leakage current"; the material involved is "High-k / Metal Gate"; the applicant is a leading wafer foundry.

[0156] Automatic domain classification and dynamic loading

[0157] (1) Feature extraction: Extract technical parameters (3nm process node, 40% reduction in leakage current), extract high-frequency technical keywords (GAA, FinFET evolution, Gate-All-Around, High-k), and construct a cooperative network relationship vector.

[0158] (2) Classification determination: The classifier calculates that the probability of "semiconductor / microelectronics" is 98% (>70% threshold), and the target technology field is determined to be "semiconductor / microelectronics".

[0159] (3) Dynamic loading: Load the semiconductor field feature library (including entity type definitions such as process node, device architecture, material system, lithography technology, etc.) and the association rule library (including rules such as: {"node downsizing by one generation"} → {"basic weight + 0.4"}; {"architectural innovation (FinFET to GAA)"} → {"extra weight + 0.5"}; {"EUV lithography", "multilayer stacking"} → {"advanced process"}, etc.).

[0160] Multi-level value quantification assessment

[0161] (1) Surface index calculation: The patent family size is large (covering multiple countries such as US, EP, CN, JP, KR, etc.), and IEEE standard documents are cited. The surface comprehensive score is calculated as 75 / 100.

[0162] (2) In-depth domain analysis:

[0163] Identifying "3nm" and "GAA": Matching the "process node" and "device architecture" dictionaries. According to the rules, the transition from FinFET to GAA is a generational revolution, triggering the rule {"architectural innovation"} → {"extra weight +0.5"}, assigning it a high weight;

[0164] Performance metrics analysis: "40% reduction in leakage current" is quantified into a specific performance improvement score (performance improvement score = 40% / industry benchmark improvement value).

[0165] Dependency chain analysis: The knowledge graph shows that the technology is heavily reliant on "EUV lithography" and "atomic layer deposition (ALD)" equipment, identifying key nodes in the supply chain;

[0166] The domain depth analysis score is calculated to be 95 / 100.

[0167] (3) Cross-domain correlation value assessment: It was found that the structure optimization algorithm can be applied to the field of "high performance computing (HPC)" chip design, increasing the application potential score. The cross-domain correlation value score is calculated to be 80 / 100.

[0168] Value integration and visualization output

[0169] (1) Overall score: Overall value score = 0.2×75+0.5×95+0.3×80=15+47.5+24=86.5.

[0170] (2) Visual presentation: Generate a technology evolution roadmap, clearly showing the evolution path from Planar→FinFET→GAA, and mark the project as a key node of the "next generation mainstream architecture"; use the industry chain radiation heat map to show its radiation range in the upstream of the industry chain (EUV lithography equipment, ALD equipment, High-k materials), and the color depth indicates the degree of influence.

[0171] The following uses a research paper on "interface modification of sulfide electrolyte in solid-state batteries" as input data to demonstrate the complete evaluation process of the method in this application.

[0172] Raw data input: The paper mentions "sulfide electrolyte", "interfacial impedance" and "cycle life of 2000 cycles"; experimental data includes ionic conductivity increased to the level of 10⁻³ S / cm.

[0173] Automatic domain classification and dynamic loading

[0174] (1) Feature extraction: Extract material name (sulfide), performance index parameters (cycle life 2000 times, ionic conductivity 10⁻³ S / cm), extract high-frequency keywords (solid-state battery, interface modification, electrolyte), and construct cooperative network relationship vector.

[0175] (2) Classification determination: The classifier calculates that the probability of "new energy / energy storage" is 94% (>70% threshold), and the target technology field is determined to be "new energy / energy storage".

[0176] (3) Dynamic loading: Load the feature library of the new energy field (including entity type definitions such as energy density, cycle life, safety, material cost, electrolyte type, electrode material, etc.) and the association rule library (including rules such as: {"cycle life>1000 times"}→{"practical application threshold met"}; {"interface modification", "core pain point"}→{"high innovation score"}; {"does not contain rare metals"}→{"low cost bonus"}, etc.).

[0177] Multi-level value quantification assessment

[0178] (1) Surface index calculation: The journal has a high impact factor (e.g., IF>15) and a high number of downloads. The surface comprehensive score is calculated as 70 / 100.

[0179] (2) In-depth domain analysis:

[0180] Pain point matching: "High interfacial impedance" is a recognized core pain point of sulfide batteries in the knowledge base. This paper proposes an "interface modification" solution, which matches the rule {"interface modification", "core pain point"} → {"high innovation score"}, resulting in a significant increase in weight;

[0181] Data quantification: Compare "cycle life of 2000 cycles" with the industry average (usually <500 cycles) to calculate a significant performance advantage score (performance advantage score = 2000 / 500 = 4 times the industry average);

[0182] Cost-related: Analyzing the material composition confirms that it does not contain rare metals (such as cobalt or germanium), triggering the rule of {"not containing rare metals"} → {"low cost bonus"}.

[0183] The domain depth analysis score is calculated as 90 / 100.

[0184] (3) Cross-domain association value assessment: The knowledge graph is associated with the range anxiety problem of "electric vehicles (EV)" (association strength 0.85) and the miniaturization demand of "consumer electronics" (association strength 0.6), confirming its wide range of application scenarios. The cross-domain association value score is calculated to be 82 / 100.

[0185] Step 3: Value Integration and Visualization Output

[0186] (1) Comprehensive score: Comprehensive value score = 0.2×70+0.5×90+0.3×82=14+45+24.6=83.6.

[0187] (2) Visual presentation: Generate a technology maturity curve to mark the technology as being in the "laboratory verification to pilot production" stage (TRL 4~5); use a heat map of the industrial chain to show the coverage and impact of the technology in downstream application scenarios such as electric vehicles, consumer electronics, and energy storage power stations.

[0188] This application effectively solves the technical bottleneck of mismatch between cross-domain evaluation standards in traditional general algorithms. By leveraging a three-layer domain classifier and adversarial training mechanism, it achieves accurate domain determination and adaptive weight output. Combined with the dynamic loading of domain feature libraries and association rule libraries, it comprehensively captures the domain-specific value and cross-disciplinary transfer potential of intelligence through three-layer quantitative fusion of surface indicators, in-depth domain analysis, and cross-domain association value. Moreover, the ontology-based knowledge extraction and automatic association rule mining mechanism realizes full lifecycle management of the feature library, eliminating the need for experts to write rules one by one. With the differentiated visualization output adapted to the domain, it achieves full-process closed-loop automation, which can be seamlessly adapted to various scientific and technological fields, significantly improving the accuracy, automation level, and decision support of the evaluation.

[0189] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Where there is no conflict, the embodiments and features described in the embodiments of this application can be combined with each other. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A domain-adaptive method for assessing the value of scientific and technological intelligence, characterized in that, Includes the following steps: Acquire the science and technology intelligence data to be analyzed; The scientific and technological intelligence data to be analyzed is preprocessed to extract multi-dimensional features of the data and generate a domain feature vector. The domain feature vector is input into a domain automatic classifier to determine the target technology domain to which the technology intelligence data to be analyzed belongs. Based on the target technology field, dynamically load the domain feature library and association rule library corresponding to the target technology field; Based on the loaded domain feature library and the association rule library, the scientific and technological intelligence data to be analyzed is subjected to multi-level value quantification assessment, which includes surface indicator calculation, domain in-depth analysis and cross-domain association value assessment. The results of the multi-level value quantification assessment are integrated to generate a comprehensive value score for the scientific and technological intelligence data to be analyzed.

2. The method according to claim 1, characterized in that, The multidimensional features include at least one of the following: The technical indicator parameter vector is obtained by standardizing and transforming specific types of parameters in the scientific and technological intelligence data to be analyzed; The text semantic vector is obtained by extracting the distribution of high-frequency keywords from the summary of the scientific and technological intelligence data to be analyzed, generating a semantic vector, and then reducing its dimensionality. The collaborative network relationship vector is constructed by statistically analyzing the number of types of collaborating institutions and the number of cross-domain collaborations.

3. The method according to claim 1, characterized in that, The domain automatic classifier adopts a three-layer structure, including: The feature extraction layer is used to perform feature enhancement processing on the domain feature vectors and output dense feature vectors of fixed dimensions. The domain classification layer is used to receive the dense feature vector, combine it with the high-frequency keyword distribution, calculate the probability distribution of the scientific and technological intelligence data to be analyzed belonging to each preset scientific and technological domain, and select the category with a probability higher than a preset threshold as the initial label. The weight generation layer is used to output the weight combination of the evaluation dimensions for the corresponding domain based on the initial labels.

4. The method according to claim 3, characterized in that, The weight generation layer adopts a domain adversarial training mechanism, which optimizes the robustness of feature representation through the game between the generator and the discriminator.

5. The method according to claim 1, characterized in that, The surface index calculation is used to calculate the general value index of the scientific and technological intelligence data to be analyzed, and the general value index includes citation frequency and / or publication time; The domain-specific analysis includes: calling the loaded domain feature library to extract domain-specific entities and relationships from the scientific and technological intelligence data to be analyzed; calling the loaded association rule library to calculate the domain weight; and generating a domain-specific analysis score based on the domain-specific entities and relationships and the domain weight. The cross-domain correlation value assessment includes: based on knowledge graphs, discovering the potential application value of the technologies involved in the scientific and technological intelligence data to be analyzed in other scientific and technological fields; and awarding cross-domain bonuses based on the discovered potential application value.

6. The method according to claim 5, characterized in that, The domain feature library contains a value feature dictionary specific to each technology domain, and the value feature dictionary includes domain-specific entity type definitions and their corresponding value assessment dimensions.

7. A domain-adaptive scientific and technological information value assessment system, characterized in that, include: The data acquisition module is used to acquire the scientific and technological intelligence data to be analyzed. The feature extraction module is used to preprocess the scientific and technological intelligence data to be analyzed, extract multi-dimensional features of the scientific and technological intelligence data to be analyzed, and generate domain feature vectors. The domain classification module is used to input the domain feature vector into the domain automatic classifier to determine the target technology domain to which the technology intelligence data to be analyzed belongs. The dynamic loading module is used to dynamically load the domain feature library and association rule library corresponding to the target technology domain according to the target technology domain; The multi-layer value quantification assessment module is used to perform multi-layer value quantification assessment on the scientific and technological intelligence data to be analyzed based on the loaded domain feature library and the association rule library. The multi-layer value quantification assessment includes a surface indicator calculation layer, a domain depth analysis layer, and a cross-domain association value assessment layer. The fusion output module is used to fuse the results of the multi-level value quantification assessment to generate a comprehensive value score for the scientific and technological intelligence data to be analyzed.

8. The system according to claim 7, characterized in that, Also includes: The domain feature library construction module is used to extract entity-relation-entity triples from multi-source heterogeneous data, perform entity alignment and fusion, and use association rule mining algorithms to mine frequent itemsets from historical data to generate association rules, so as to build domain feature libraries and association rule libraries corresponding to various scientific and technological fields.

9. The system according to claim 8, characterized in that, The domain feature library construction module includes: The multi-source data aggregation unit is used to collect heterogeneous data from multiple sources and perform word segmentation, stop word removal, and part-of-speech tagging. The knowledge extraction unit is used to reference domain ontology libraries and extract entity-relation-entity triples from core literature using text mining tools. The entity alignment and fusion unit is used to perform strong rule deduplication using unique identifiers, and weak rule alignment by calculating edit distance, co-occurrence frequency and semantic similarity, and to resolve entity attribute conflicts. The association rule mining unit is used to mine frequent itemsets from historical data and generate association rules; The dynamic update unit is used to periodically recalculate entity weights and association rules based on real-time data streams.

10. The system according to claim 7, characterized in that, The domain classification module employs a domain autoclassifier trained under supervision using the cross-entropy loss function, and adjusts hyperparameters using a validation set.

Images