CN-122020264-A - Innovative capability assessment method based on large language model and multi-modal semantic understanding

Abstract

The invention discloses an innovation capability assessment method based on a large language model and multi-mode semantic understanding, belonging to the technical field of information processing; the method comprises the steps of collecting multisource heterogeneous data inside and outside enterprises to generate a unified feature matrix, achieving global feature interaction among the enterprises through a multi-head self-attention mechanism, outputting the enterprise feature matrix, designing a gating network to dynamically modulate the enterprise feature matrix element by element, extracting unchanged domain features through countertraining, dynamically generating domain-specific features through a super network, obtaining final adaptation features through gating aggregation, and obtaining enterprise innovation dynamic scores. According to the method, the DANN structure is used for carrying out domain invariance constraint on enterprise features, so that features of different technical domains are aligned in a bottleneck space, and HyperNetwork is used for combining features to guide soft domain routing, so that a domain-specific feature transformation matrix is dynamically generated for each enterprise, and the final features comprehensively reflect the innovative investment, output and potential value of the enterprise.

Inventors

• Bei jing
• ZHANG WEIXIANG
• WANG XUAN
• WANG CHAO
• YANG HUA

Assignees

• 江苏省新质生产力促进中心

Dates

Publication Date

20260512

Application Date

20260414

Claims (10)

1. 1. The innovation capability assessment method based on the large language model and the multi-mode semantic understanding is characterized by comprising the following steps of: step S1, collecting and classifying multisource heterogeneous data inside and outside an enterprise, and respectively adopting a BERT-base model and a RoBERTa-large model to process the classified multisource heterogeneous data to generate a unified feature matrix; s2, constructing a transducer encoder with adjacent matrix attention constraint, realizing global feature interaction among enterprises through a multi-head self-attention mechanism, and outputting an enterprise feature matrix; Step S3, designing a gating network, performing element-by-element dynamic modulation on the enterprise feature matrix, and performing residual error connection and layer normalization polymerization on the enterprise feature matrix and the unified feature matrix to form depth fusion features; s4, extracting domain-invariant features through countermeasure training, dynamically generating domain-specific features through a super network, and obtaining final adaptive features through gating aggregation; and S5, dynamically generating a reference value subjected to clutter correction and piecewise nonlinear mapping by suppressing abnormal values to obtain dynamic scores of enterprise innovation.
2. 2. The innovation ability assessment method based on large language model and multi-mode semantic understanding according to claim 1 is characterized in that in the step S1, multi-source heterogeneous data inside and outside an enterprise are collected and classified, and the classified multi-source heterogeneous data are processed by adopting a BERT-base model and a RoBERTa-large model respectively, specifically: Step S1-1, defining a sample enterprise set, acquiring multi-source heterogeneous data from the sample enterprise set, and classifying the multi-source heterogeneous data, wherein the method specifically comprises the following steps: Defining sample enterprise collections , To participate in the total number of enterprises to be evaluated, each enterprise is added with The acquired data are split into two types, namely structured data And text data The structured data is a structured information set taking a numerical value or a category as a carrier, and the text data is an unstructured information set taking natural language as a carrier; Step S1-2, adopting RoBERTa-large model to process text data: step S1-2-1, each enterprise is processed Is set of original text data of (a) Sequentially performing cleaning, de-duplication and quality filtering operations to obtain a preprocessed text data set Wherein, the method comprises the steps of, Is an enterprise Is a number of text documents; step S1-2-2, carrying out multi-document aggregation on the preprocessed text data set: Ordering by information priority The documents are arranged according to a predefined fixed priority; splicing and cutting, namely splicing the sequenced documents in turn, separating the documents by RoBERTa separators, adding a start mark < s > to the sequence head, cutting the total length to the maximum input length at the token level Tokens, the expression is: Wherein, the method comprises the steps of, Is an aggregate sequence; step S1-2-3, polymerizing the sequence Feature extraction is performed through RoBERTa-large model: the RoBERTa-large model is used as a fixed feature extractor, and training parameters of the RoBERTa-large model are frozen into fixed quantity by adopting a parameter freezing strategy, and the RoBERTa-large model extracts features, which are specifically expressed as follows: extracting a hidden state vector corresponding to a sequence start mark < s > as an aggregate semantic representation of the whole text sequence, and obtaining a text feature line vector: ; Wherein, the Representing a sequence matrix; All are put together Text feature line vectors of a home enterprise are stacked according to lines to form a text feature matrix, and the expression is as follows: ; Wherein, the ; S1-3, adopting BERT-base to process structured data; and S1-4, fusing the two paths of features into a unified feature matrix through a cross-modal attention weighted aggregation mechanism, and constructing an enterprise association adjacency matrix based on the fused enterprise-level feature matrix to serve as auxiliary structure information.
3. 3. The innovation ability assessment method based on the large language model and the multi-modal semantic understanding according to claim 2, wherein the step S1-3 adopts BERT-base to process structured data, and is specifically as follows: Every enterprise Structured field records of (a) , For the total number of fields to be structured, Uniformly preprocessing the structured data, wherein the uniform preprocessing is performed on the field value for the Mth key value and passes through a predefined field configuration table Driving, wherein, the driving device comprises a driving device, As a field type of the field, Is a unit text of a character, A number of significant digits reserved; Field name for the M-th key value pair; Step S1-3-2, converting the preprocessed field records into text input sequences accepted by the BERT-base according to a predefined template and a fixed sequence: the field arrangement sequence is that the fields are fixedly arranged according to the index level of innovation capability evaluation; The serialization template is characterized in that each field is converted into text fragments according to a field name and field value format, the fragments are connected by SEP, a CLS mark is added to the sequence head, and the expression is as follows: ; Step S1-3-3, performing dimension projection by using a BERT-base model: The BERT-base model is used as a word segmentation device, the BERT-base model freezes all training parameters by adopting a parameter freezing mode, and the BERT-base model extracts the final hidden state of the [ CLS ] position as an aggregation representation: ; Wherein, the ; Is an enterprise Is a structured data serialization text of (1); Representing aggregated semantic information; Semantic coding, namely finishing dimension alignment through a learnable linear projection layer: Wherein, the method comprises the steps of, Is a matrix of weights of the linear projection layer, Is a linear projection layer bias vector; All are put together Structured features of a home enterprise are stacked in rows into a structured feature matrix: ; in step S1-4, the text feature matrix and the structured feature matrix are fused into a unified enterprise-level feature matrix through a cross-modal attention weighted aggregation mechanism, specifically: respectively sending the text feature matrix and the structured feature matrix into a shared parameter attention scoring network, and projecting the text feature matrix and the structured feature matrix to the attention scoring network Dimension hiding space, warp Compression to scalar scores after activation; converting the original score into a fusion weight of which the sum is 1 through Softmax normalization, and converting the original score into the fusion weight of which the sum is 1 through Softmax normalization; Weighting and summing the text feature matrix and the structural feature matrix by using the attention weight, and adaptively adjusting the fusion proportion of the two modal information according to the specific situation of each enterprise; The weighting characteristics before fusion and the projection characteristics after fusion are added element by element, so that information attenuation in the fusion process is relieved, and original fusion information can be reserved through a straight-through path even if the learning effect of the projection layer is poor; All are put together The fusion features of the home enterprises are stacked according to rows to form a unified feature matrix: Wherein, the method comprises the steps of, Representing an enterprise The comprehensive characteristics of the text semantic information and the structured quantization information are fused after the cross-modal attention weighting aggregation; in a unified feature matrix Based on the method, an adjacency matrix reflecting the association structure between enterprises is constructed by adopting a cosine similarity calculation mode Wherein, the method comprises the steps of, 。
4. 4. The innovation ability assessment method based on large language model and multi-mode semantic understanding according to claim 1 is characterized in that a gating network is designed in the step S3, the enterprise feature matrix is dynamically modulated element by element, and then is polymerized into deep fusion features with a unified feature matrix through residual connection and layer normalization, specifically: step S3-1, generating gating weight: Splicing the enterprise feature matrix and the unified feature matrix along the feature dimension to form the input of the gating network: Wherein, the method comprises the steps of, In order to be an enterprise feature matrix, Is a unified feature matrix; For the total number of enterprises involved in the assessment; the unified feature matrix carries original multisource fusion information and forms a complementary view angle with the enterprise feature matrix, so that the gating network can comprehensively code the front information layer and the rear information layer to judge the importance of each dimension; mapping the splicing characteristics into element-by-element gating weights through a full connection layer and a Sigmoid activation function: ; Wherein, the Is a learnable weight matrix; Is a learnable bias vector; Compressing each output element to a Sigmoid function ; Representing an enterprise Is the first of (2) The degree of correlation of the dimensional depth coding features to the evaluation task, A near 1 indicates that the dimension carrying key signals should be preserved, a near 0 indicates that the dimension is noise or redundancy should be suppressed; Step S3-2, performing element-by-element modulation on the enterprise feature matrix by using gating weight, and performing residual connection and layer normalization aggregation to obtain a depth fusion feature matrix, wherein the depth fusion feature matrix specifically comprises the following steps: Using gating weights For enterprise feature matrix Independently performing on-off control for each element of (a): Wherein, the method comprises the steps of, For the element-by-element multiplication, Is an enterprise feature matrix subjected to gating refining; modulating the gating of the modulated features And unifying feature matrices Adding elements by elements, and performing layer normalization aggregation to obtain a depth fusion feature matrix: ; Wherein, the Is residual connection; normalizing the layers; And (5) depth fusion of the feature matrix.
5. 5. The innovation ability assessment method based on the large language model and the multi-modal semantic understanding according to claim 1, wherein in the step S4, domain-invariant features are extracted through countermeasure training, domain-specific features are dynamically generated through a super network, and final adaptation features are obtained through gating aggregation, specifically: step S4-1, constructing a DANN model to obtain domain-invariant features of cross-domain generalization: the DANN model comprises a feature encoder Gradient inversion layer GRL and domain discriminator Using feature encoders Extracting a generic feature representation shared across domains: Feature encoder Is a single-layer fully-connected network, will Depth feature compression of dimensions to Dimension bottleneck space: ; Wherein, the For the feature encoder weight matrix, A feature encoder bias vector; Is an enterprise A coded representation in bottleneck space; the behavior of the gradient inversion layer GRL is asymmetric in forward and backward propagation, with forward being identity mapped Reversing the direction of the gradient and then Control inversion strength: Wherein, the method comprises the steps of, As a function of the gradient tensor, Is a gradient inversion coefficient; Gradient inversion coefficient The S-shaped scheduling function proposed by DANN is adopted: ; Wherein, the Is the training progress proportion; for the current iteration step, The total iteration step number; Domain discriminator For two-layer fully connected network, receiving coding features transferred by gradient inversion layer GRL And predict the domain source of each sample, the first layer will Projection of dimension code features via nonlinear transformation Dimension discrimination subspace: ; Wherein, the For the first layer weight matrix of the arbiter, A first layer bias vector for the arbiter; The second layer linearly projects the distinguishing subspace features to Individual domain categories and normalized to probability distribution by softmax: ; Wherein, the For the second layer weight matrix of the arbiter, A second layer bias vector for the arbiter; Is the total number of domain categories; Is an enterprise Belongs to the prediction probability distribution in various fields, ; Activating a feature matrix for the hidden layer of the discriminator; Domain discrimination cross entropy loss, namely measuring deviation between domain probability distribution predicted by a discriminator and a real domain label by using the cross entropy loss, wherein the expression is as follows: ; Wherein, the A single-heat coding matrix for the tag in the real field; Indicating if and only if the enterprise Belonging to the first A number of fields; The encoder output is defined as a domain invariant feature: ; S4-2, dynamically generating domain-specific features based on HyperNetwork super-networks and feature-guided soft domain routes FSDR; and S4-3, adaptively polymerizing the domain-invariant features and the domain-specific features into final adaptive features through a gating network and dual-path projection.
6. 6. The innovation ability assessment method based on large language model and multi-modal semantic understanding according to claim 5, wherein the step S4-2 is based on HyperNetwork super network and feature guide soft domain route FSDR, and the dynamic generation domain-specific features are specifically as follows: step S4-2-1, calculating feature projection and soft route weight by using the feature guide soft domain route FSDR: Will be Domain invariant feature projection of dimensions Domain embedding space of dimensions, mapping enterprise features to the same semantic space as domain embedding to perform similarity calculations: ; Wherein, the Is a learnable route projection matrix; Is an enterprise Is a domain invariant feature vector of (1); Is an enterprise A routing query vector embedded in a domain space; Routing query vectors Embedding tables for queries in a learnable domain Calculating soft routing weights of enterprises to various fields through scaling dot product attention: ; Wherein, the Is the first A field-of-things learnable embedded vector; Scaling factors for soft routing weights; Representing an enterprise Belongs to the field of Soft weight of (2), softmax guarantees And is also provided with ; The domain embedding table is weighted and mixed by the soft routing weight, and the personalized domain embedding is synthesized for each enterprise: ; Wherein, the Is an enterprise Is embedded in the personalized domain of (1) When the heat vector is degenerated into a single heat vector, ; To prevent the routing weights from completely deviating from the known domain labels, regularization loss is introduced, providing weak anchor constraints for soft routing: ; Wherein, the Representing a loss penalty; step S4-2-2, embedding the personalized field Inputting the parameters of the super network to a parameter generation module, dynamically synthesizing the special characteristic transformation matrix of the enterprise, and generating a low-rank transformation matrix: Will be The complete transform matrix of the transform matrix is decomposed into left factors And the right factor Is a product of (2); the two factors are symmetrical in generation structure and expressed in a unified form Is provided with As the total number of elements of the corresponding factor, Is the target shape of the corresponding factor, wherein, For a low-rank decomposed rank, For domain-specific feature dimensions: ; ; Wherein, the Generating a matrix for factors of the super network; for ultra-network to enterprise Flat parameter vectors embedded in the personalized field of the system; representing a super network as an enterprise Dynamically generated first A domain-specific weight matrix of low rank factors; Combining the two low rank factors into a complete transform matrix: ; Wherein, the Is an enterprise Left factor of proprietary transformation matrix, capture slave Space-invariant dimension A compressed mapping of the dimensional intermediate space; As right factor, capture slave Space-invariant dimension A compressed mapping of the dimensional intermediate space; Is an enterprise A proprietary complete transformation matrix; and S4-2-3, projecting the domain invariant features to a domain feature space by using a dynamic transformation matrix, and extracting domain specific features: re-expressing the domain invariant features of each sample into domain feature forms according to the specific domain rules: ; Wherein, the Is an enterprise Is a domain invariant feature vector of (1); Is an enterprise Is a domain-specific feature vector of (1); Code enterprises Is a cross-domain generic model of (a), The characteristic transformation rule of the field where the enterprise is located is encoded, and the general mode is re-expressed into a field characteristic form by multiplying the characteristic transformation rule and the field transformation rule according to the field rule; For all of The individual samples perform the above projection and are stacked along the batch dimension: Wherein, the method comprises the steps of, Representing domain-specific features.
7. 7. The innovation ability assessment method based on large language model and multi-modal semantic understanding according to claim 5, wherein in the step S4-3, domain-invariant features and domain-specific features are adaptively aggregated into final adaptation features through gating network and dual-path projection, specifically: step S4-3-1, designing the aggregation weight of the self-adaptive gating network: generating scalar gating weights for each sample through a gating network Quantifying the relative dependence of the sample on the domain-invariant features and domain-specific features: ; Wherein, the For a learnable weight vector of the gating network, A learnable bias scalar for a gating network; Compression of output to Sigmoid function A section; Is an enterprise Is a gating weight scalar of (a), A value close to 1 indicates that the sample should focus on the domain invariant feature, and a value close to 0 indicates that the sample should focus on the domain specific feature; Step S4-3-2, dual-path projection and feature aggregation, namely firstly splicing the domain-invariant features and domain-specific features along feature dimensions: ; The final adaptation features are then generated by a two-path projection: ; Wherein, the For gating the modulated domain invariant feature, Broadcast to along characteristic dimension Rear and The multiplication is performed element by element, For element-by-element multiplication; in order to gate the generic feature projection matrix, The feature projection matrix is spliced for dual path projection, Projecting a bias vector for the dual path; the feature dimensions are finally adapted.
8. 8. The innovation ability assessment method based on the large language model and the multi-modal semantic understanding according to claim 1, wherein in the step S5, the dynamic score of the innovation of the enterprise is obtained by suppressing abnormal values, dynamically generating the reference values after the confounding correction and piecewise nonlinear mapping, specifically: Step S5-1, compressing Gao Weishi configuration features into continuous innovation capability predicted values of enterprises by using a scoring prediction head: The scoring prediction head adopts a two-layer fully-connected network, and the first layer is to Projection of dimension-adaptive features via nonlinear transformation to Dimension scoring subspace: ; Wherein, the In order to be a hidden layer weight matrix, Is a hidden layer bias vector; is the final adapting feature; The second layer compresses hidden layer characteristics into unconstrained one-dimensional real values through linear projection, and the unconstrained one-dimensional real values are used as continuous prediction of innovation capacity of each enterprise by a model: ; Wherein, the In order to output the weight vector(s), To output a bias scalar; for modeling enterprises Continuous predictive value of innovation ability, and scoring predictive vector is as follows: Wherein, the method comprises the steps of, Transpose the symbol; S5-2, eliminating dimension differences and inhibiting disturbance of extreme values to the population statistics through robust standardization and outlier self-adaptive attenuation based on quartiles; And S5-3, generating a dynamic reference value subjected to clutter correction based on the current population distribution, and converting the dynamic reference value into a final score through piecewise nonlinear mapping.
9. 9. The innovation ability assessment method based on the large language model and the multi-modal semantic understanding according to claim 8, wherein in the step S5-2, through robust normalization and outlier adaptive attenuation based on quartiles, dimension differences are eliminated and disturbance of extreme values to the population statistics is suppressed, which is specifically expressed as follows: step S5-2-1, calculating group quantile statistics: Population median: ; Quartered bit distance: ; Wherein, the As a position reference instead of the mean value, As a scale reference instead of standard deviation; step S5-2-2, calculating an outlier adaptive penalty intensity coefficient: quantifying the abnormality degree of each sample according to the degree that each sample exceeds the Tukey box diagram criterion, and providing self-adaptive penalty force for subsequent attenuation: ; Wherein, the Is a constant value of the stability of the numerical value, Penalty intensity coefficient for indication function Determined by the excess of samples from the median: step S5-2-3, robust standardization and outlier self-adaptive attenuation, wherein the expression is: ; When the value is abnormal And the indicator function takes 0, the normalized degradation is in classical robust form, when the abnormal value is And the indicator function takes 1, the molecule is Compression, the more extreme the deviation, the stronger the compression; For all of The home enterprise performs the above operations to form a robust normalized scoring vector: ; Each element of (3) Reflecting enterprises The relative position of the innovation ability prediction value in the current population has eliminated dimensional differences and suppressed amplitude disturbance of outliers.
10. 10. The method for evaluating innovation ability based on large language model and multi-modal semantic understanding according to claim 8, wherein the step S5-3 is characterized in that the dynamic benchmark value subjected to clutter correction is generated based on current population distribution and is converted into final score through piecewise nonlinear mapping, and the method is characterized in that: step S5-3-1, constructing a confounding factor characteristic matrix and eliminating systematic deviation through OLS regression: Constructing confounding factor feature vectors for each enterprise , wherein, For the purpose of industry independent thermal coding, To evaluate year single heat coding, stacking all samples as confounding factor characteristic matrix ; Regression coefficients for each confounding factor were estimated by OLS fit linear model: ; Wherein, the Indicating a scored baseline shift under conditions where all confounding factors are zero; is the first Regression coefficients for the confounding factors; the number of year categories; is the first of the confounding factor feature matrix Line 1 Column elements; Comprehensive quantification of industry attribution and year of evaluation for enterprises The systematic effect of normalizing scores; reject bias from normalized scores: ; Wherein, the Scoring the normalized score after correction; step S5-3-2, calculating relative reference position indexes: the relative reference position index comprises a median deviation index and a head distance index, and the median deviation index measures enterprises The relative degree of deviation from the median level in the population is expressed as: ; Head distance index measuring enterprise The relative distance from the head threshold is expressed as: ; Wherein, the Representing a weighing enterprise A relative degree of deviation from the median level in the population; Representing the relative distance from the head threshold; indicating the mid-level of the current population, Representing the current population lead level threshold; s5-3-3, calculating field semantic calibration items and synthesizing comprehensive dynamic reference values: embedding personalized fields And routing query vectors The degree of coincidence of the two is measured by cosine similarity: ; A high value indicates that the enterprise feature is highly consistent with the domain image thereof, and the standard is adjusted in a proper forward direction; Linearly combining the indexes of the three dimensions with fixed weights to synthesize a comprehensive dynamic reference value: ; Wherein, the , , ; Comprehensive reflection enterprise After correction for confounding factors, relative to the current population median level and head level performance, and calibrated for domain similarity; step S5-3-4, segmentation is carried out through a nonlinear mapping function: When (when) When the method is used for low segmentation, the super linear mapping is adopted, the enterprise is withheld and accelerated more backward, the distinguishing strength of the tail enterprise is enhanced, and the expression is as follows: ; When (when) When the enterprise score is used as the middle section, the main enterprise score is uniformly distributed among 40-60 partitions by using strict linear mapping, the fairness is maintained and evaluated, and the expression is as follows: ; When (when) When the sub-linear mapping is adopted, the head saturation effect is restrained, and the expression is as follows: ; Final evaluation set: , ; In order to be a level of a target, In order to raise the stage it is necessary, The score is generated based on current data, and the benchmark is dynamically adjusted along with the composition and the performance of the evaluation group, so that the hysteresis of static standards is overcome, and the latest situation of productivity development can be sensitively reflected.

Description

Innovative capability assessment method based on large language model and multi-modal semantic understanding Technical Field The invention belongs to the technical field of information processing, and particularly relates to an innovation capability assessment method based on a large language model and multi-mode semantic understanding. Background In the age background of development of digital economy and innovation drive, the innovation capability of enterprises has become an important scale for measuring the level of industrial competitiveness and productivity. The government department needs to classify the enterprises in layers according to innovation ability when making the technology supporting policy, financial prize compensation and tax reduction and exemption standard, the financial institution needs to quantitatively characterize the technical innovation level and continuous innovation ability of the enterprises when developing technology financial trust, equity investment and parallel purchase recombination, and the enterprises hope to identify the short plates and optimize resource allocation through quantitative evaluation results so as to realize continuous improvement of innovation efficiency. Therefore, how to perform scientific, objective and dynamic quantitative assessment on the innovation capability of enterprises has become a common requirement in policy making, resource allocation and enterprise management. The existing innovation capability assessment method mainly adopts traditional multi-index assessment means such as expert scoring, analytic Hierarchy Process (AHP), fuzzy comprehensive assessment and the like, generally depends on a small amount of structural indexes such as research and development input intensity, patent quantity, business income increasing rate and the like, and sets fixed weight and static scoring standard on the basis of the structural indexes. Such methods have the following limitations: 1. The association relation and group position among enterprises are difficult to model, the enterprises are mostly regarded as mutually independent evaluation objects by the traditional method, and network structures formed in the aspects of technical similarity, patent citation, industry chain upstream and downstream collaboration, joint research and development and the like among the enterprises are ignored only based on respective static index scores. In practical application, the relative position of enterprises in technical groups (such as heads, chasers or potential invisible champions) is critical for innovation capability judgment, but the traditional method lacks systematic modeling of group interaction information; 2. The evaluation standard is statically stiff and lacks dynamic adaptability, and the existing evaluation system often adopts a preset fixed scoring standard or threshold value grading, so that the changes of macro economic environment, industrial period, technical iteration and overall innovation level are difficult to reflect in time. The actual meaning of the same score at different times may drift significantly, resulting in a lack of reliability of the longitudinal comparison across years; 3. The cross-industry comparability is poor, the field difference and the innovation capability are mutually mixed, namely, a unified index system and a linear standardization method are simply adopted, and the inherent attribute of the industry is easily mistaken as the innovation quality, so that the scores of enterprises in different industries lack comparability, and the cross-industry resource allocation and the cross-area transverse alignment are difficult to support. With the development of a pre-trained large language model and the maturation of multi-mode semantic understanding technology, large-scale heterogeneous data is subjected to unified semantic coding by utilizing the large model, and a new technical path is provided for enterprise innovation ability assessment by combining graph structure modeling, countermeasure learning and dynamic scoring mechanisms. However, the existing disclosure technology is not an integrated method which simultaneously combines multi-source data fusion, enterprise group interaction modeling, field self-adaption and dynamic reference generation, and is difficult to realize fine depiction of enterprise innovation capability and quantitative comparison of cross-field fairness in a complex real scene. Therefore, a new innovation ability assessment method which can integrate multi-source heterogeneous data, jointly describe enterprise group relations and simultaneously has field self-adaptation ability and dynamic benchmark generation ability is urgently needed, so that scientific quantification of the innovation level of enterprises and fair comparison of across industries and periods are realized. Disclosure of Invention The invention aims to provide an innovation capability assessment method based on a large language mod