Scientific achievement transformation method based on graph knowledge large model

Abstract

The invention provides a scientific achievement transformation method based on a graph knowledge large model, and belongs to the technical field of artificial intelligence. The method comprises the following steps: constructing a scientific research achievement conversion method data set through multi-source data acquisition; a large scientific research achievement conversion model is constructed based on parser-generator stacking, a conversion scene guiding module is introduced, and the large scientific research achievement conversion model comprises a coding vector module, a conversion scene guiding module, a parser, a generator and a method output module; a segmented training strategy of preliminary training, scene guiding training and conversion strengthening training is adopted, and the model is optimized through a joint loss function; and converting the trained scientific achievements into large model deployment application. According to the invention, the robustness and generalization ability of the scientific achievement transformation method are improved; the model training cost is reduced, the matching between the model and the actual conversion environment is enhanced, and the practicability and accuracy of the scientific achievement conversion method are improved.

Description

Technical Field

[0001] The present invention relates to a method for transforming scientific research results based on a large graph knowledge model, and belongs to the field of artificial intelligence technology. Background Art

[0002] In terms of scientific research, the number of scientific research results obtained is increasing year by year, and the transformation of scientific research results is a key link in connecting scientific research to the formation of industry. The existing scientific research results transformation methods are weak in data integration and complex correlation analysis capabilities, and are unable to obtain the invisible correlation between scientific research results data and the scientific research results transformation environment. The transformation of scientific research results often relies on fixed rules, and cannot take into account the dynamic factors in the technology transformation process. It lacks the ability to perceive the actual environment of scientific research transformation. Moreover, the existing scientific research results transformation process is too dependent on manual labor, and the level of automation of the scientific research results transformation process is low. The vigorous development of artificial intelligence big model technology provides new possibilities for the generation of scientific research results transformation methods. Based on the powerful generation and feature extraction capabilities of big models, it makes up for the problems of poor data processing capabilities, high manual dependence, and poor complex decision-making capabilities of traditional scientific research results transformation methods. By constructing an intelligent big model system that can evolve by itself to adapt to complex display scenes, it can further fill the gap between technology research and development and industrial implementation, and ensure that scientific research results can be released efficiently and with high quality. To this end, the present invention proposes a scientific research results transformation method based on a graph knowledge big model. Existing methods for transforming scientific research results include the following: (1) Research achievement transformation method based on research achievement transformation specialists: Research achievement transformation specialists manually evaluate the technical maturity and market value of research achievements, and formulate complete research achievement transformation solutions based on the actual research achievement transformation environment. Based on the formulated research achievement transformation solutions, they directly communicate with the docking enterprises on needs, screening achievements, technology pricing, signing agreements and implementation guidance, and finally complete the actual transformation of research achievements; (2) Research achievement transformation method based on deep learning model: Based on historical research achievement data, enterprise demand data, achievement transformation method data and market feedback data, a feature extraction module is constructed to generate standardized feature vectors. The model is trained with historical research achievement transformation method data and market feedback data to capture the semantic association input features between research achievements and achievement transformation methods. The trained model is then used to automatically generate research achievement transformation methods. (3) A reinforcement learning-based method for transforming scientific research results: The technical parameters of scientific research results (maturity, patent value), enterprise demand characteristics (industry, budget scale), and market environment variables (policy support, competitive product dynamics) are encoded as a state space, and the transformation strategy (licensing model, pricing plan, resource investment scale) is defined as an action space. A multi-dimensional reward function is constructed based on transformation revenue, implementation efficiency, market penetration, etc., forming a "state-action-reward" interactive closed loop. A deep reinforcement learning algorithm is used to train the policy network. By simulating the trial and error process of different transformation paths, the optimal strategy combination that maximizes long-term benefits is learned. Ultimately, the effective generation of a scientific research result transformation method is achieved.

[0003] However, the aforementioned methods have the following shortcomings and disadvantages when generating research results transformation methods: Methods based on research results transformation specialists rely heavily on their personal experience and subjective judgment, making technology maturity and market value assessments susceptible to cognitive limitations, leading to biased or inefficient assessments. Methods based on deep learning models overly focus on historical transformation patterns, have weak predictive capabilities for adaptability to emerging technology fields or breakthroughs, and struggle to dynamically respond to changes in market demand. Reinforcement learning-based methods require integrating multiple variables, such as technology, demand, and market, when constructing the state space. This results in high parameter complexity, and reinforcement learning training requires extensive trial-and-error simulations, resulting in high computational costs. Furthermore, the model requires continuous manual calibration after implementation. Summary of the Invention

[0004] The purpose of this invention is to provide a scientific research results transformation method based on a large graph knowledge model, enhance the matching between the model and the actual transformation environment, and improve the practicality and accuracy of the scientific research results transformation method.

[0005] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions: S1. Construct a dataset for scientific research achievement transformation methods through multi-source data collection, including a scientific research achievement text familiarization dataset, a scientific research achievement transformation question-answering dataset, and a scenario guidance module dataset; S2. Construct a large model for the transformation of scientific research results based on the parser-generator stack, and introduce a transformation scenario guidance module; the large model for the transformation of scientific research results includes a coding vector module, a transformation scenario guidance module, a parser, a generator, and a method output module. The coding vector module converts text data into a weighted coding vector. The transformation scenario guidance module generates a multi-level transformation condition background feature vector based on the scenario guidance module data set. The method output module includes two fully connected layers connected in sequence and a Softmax activation function. S3. A segmented training strategy of scenario-guided training, preliminary training, and conversion-enhanced training is adopted to optimize the model through a loss function; in the scenario-guided training phase, the conversion scenario-guided module is trained separately using a scenario-guided module dataset, and the parameters of the conversion scenario-guided module after training are frozen; in the preliminary training phase, the large model for conversion of scientific research results is trained using a scientific research result text familiarization dataset; in the conversion-enhanced training phase, the large model for conversion of scientific research results is trained using a scientific research result conversion question-answering dataset; S4. Transform the trained scientific research results into large-scale models for deployment and application.

[0006] Preferably, constructing a dataset for scientific research achievement transformation methods through multi-source data collection includes: Collect examples of scientific research results transformation and extract text data covering the entire process of scientific research results transformation; Using sliding window mechanism to construct scientific research results text familiarity dataset , including: segmenting the scientific research results text data, extracting the first half of each segment as the input data of the scientific research results text familiarity data, and the second half as the output data of the scientific research results text familiarity data, together forming a set of scientific research results text familiarity data, and the scientific research results text familiarity data composed of all text segments as the scientific research results text familiarity data set ; Constructing a question-answering dataset for the transformation of scientific research results , including: extracting the text of the technical solution to be converted from the scientific research results text data as the scientific research results text data , extract the corresponding conversion requirement data , together as input data for scientific research results , based on the text data of scientific research results Mark the corresponding scientific research results transformation method data , input scientific research results into data Data on methods and methods for transforming scientific research results Constructing a question-answering dataset for the transformation of scientific research results ; Constructing a scene guidance module dataset via non-negative matrix factorization and semantic expansion , including: extracting transformation requirement data through non-negative matrix factorization algorithm Keywords in the research results to be enriched , according to the data of scientific research results transformation method Keywords for enriching scientific research results Perform semantic expansion to obtain enriched keyword data for the entire process , to be enriched with scientific research results keywords and enriched keyword data throughout the entire process Together they constitute the scene guidance module dataset .

[0007] Preferably, the conversion scenario guidance module includes: a graph generation module, a conversion environment feature sensor and a task semantics reconstructor; The graph generation module converts the input data set into a transformation scene guidance graph that can represent complex topological relationships through a scene graph embedding method, including an input guidance graph and an output guidance graph; The input guidance image is fed into the transformation environment feature sensor for feature extraction and is processed by three transformation environment feature sensors in sequence; The extracted features are sequentially passed through ten task semantic reconstructors to perform multi-level transformation condition background feature reconstruction operations. Each task semantic reconstructor extracts the transformation condition background feature vector and gradually generates an output guidance map that fits the actual scientific research results transformation background.

[0008] Specifically, the data processing method during the training phase of the conversion scenario guidance module is as follows: The graph generation module converts the keyword text into graph data that can represent complex topological relationships through the scene graph embedding method, wherein the keyword structure of the scientific research results to be enriched is Building a basic semantic graph , enriched full-process keyword data Built as an extended knowledge graph with more complete contextual semantics ; The basic semantic graph The input guide map is fed into the transformation environment feature sensor for feature extraction and processed by three transformation environment feature sensors in sequence; The extracted features are sequentially passed through ten task semantic reconstructors to perform multi-level transformation condition background feature reconstruction operations. Each task semantic reconstructor extracts the transformation condition background feature vector and gradually generates an output guidance graph that fits the actual scientific research results transformation background. The output guidance graph is an extended knowledge graph. .

[0009] Preferably, the scene graph embedding method is specifically as follows: Each keyword in the input data set is treated as a node in the graph data, and each keyword text is converted into a feature vector through the word embedding layer; The cosine similarity between the feature vectors of each node is calculated. If the cosine similarity exceeds the set threshold, an edge is constructed between the two nodes to realize the construction of the transformation scene guidance graph.

[0010] Preferably, the conversion environment feature sensor processes the input guidance graph as follows: The input guide graph is extracted through two sequentially connected graph convolutional layers, and the features are nonlinearly activated using the Relu activation function. The activated features are sequentially processed by the Dropout layer and the graph maximum pooling layer, and then two sequentially connected graph convolutional layers are used to extract features through the Relu activation function. The task semantic reconstructor processes the input features as follows: The feature data extracted by the perceptron module is upsampled through three sequentially connected graph deconvolution layers, and the upsampled features are integrated and transformed using three sequentially connected fully connected layers. The features are nonlinearly activated by the Mish activation function, and further upsampled by three sequentially connected graph deconvolution layers. An attention factor is introduced through a graph attention layer, and the final activation of the features is completed by the Simgoid activation function.

[0011] Preferably, the data processing method for converting scientific research results into large-scale models is as follows: The input data in the input data set (scientific research results input data or scientific research results text familiar data input data) is sent to the encoding vector module to obtain the encoding vector corresponding to the input data , the input data in the input data set is fed into the transformation scene guidance module to obtain multi-level transformation condition background features , ; The encoding vector The output of the upper parser is sent to the stacked scientific research results transformation parser for deep processing. In the parser stacking structure, the output of the upper parser is not only used as the input of the lower parser, but also the output of the same level transformation condition background feature obtained by the transformation scenario guidance module. Perform channel splicing operations and use the output vector after channel splicing as the input of the generator of scientific research results transformation methods at the same level; In the stacked structure of scientific research results transformation method generators, the output of the lower-level generator is used as the input of the upper-level generator, and the output of the top-level generator is sent to the method output module for final processing to obtain the output scientific research results transformation method.

[0012] Preferably, the encoding vector module converts text data into encoding vectors in the following specific manner: Divide the input data into word , Indicates the total number of words in the input data set, and for each word The random initialization vector is encoded as , the vector dimension is 50; Calculate the context influence weight vector for each word , and add it to the vector code corresponding to the word itself to obtain the weighted code vector corresponding to the word , the weighted coding vector corresponding to each word is combined to obtain the coding vector corresponding to the text ; Context influence weight vector The calculation method is as follows: Get the word to be encoded about The initialization vectors corresponding to the words are Initialize the encoding vectors and assign an adaptive adjustment weight to each encoding vector ; Multiply each initialization code vector by its corresponding adaptive adjustment weight, and obtain The word to be encoded is obtained by adding the products The corresponding context influence weight vector .

[0013] Preferably, the scientific research results conversion parser converts the encoding vector The in-depth analysis method is as follows: Encoded vector The features are extracted through two local connection layers and then sent to the Relu activation function for nonlinear activation to obtain the features. ; By improving the multi-head attention mechanism to Further processing to obtain the final features ; The final feature Through two fully connected layers and Dropout layers connected in sequence, and then sent to the Simgoid activation function to activate, the feature vector processed by the scientific research results conversion parser is obtained. ; The improved multi-head attention mechanism is as follows: The features that will be obtained Perform random cutting to obtain Cut features , Cut feature Send it to the polymorphic perceptron for parameter extraction and processing to obtain cutting features The corresponding cutting feature weight , the multi-state perceptron is composed of 5 layers of interconnected fully connected layers and Silu activation functions; Cut feature Feed into the matrix The calculation function calculates the The feature matrix corresponding to the attention head , cut feature Feed into the matrix The calculation function calculates the The feature matrix corresponding to the attention head , cut feature Feed into the matrix The calculation function calculates the The feature matrix corresponding to the attention head , , is the total number of heads included in the multi-head attention mechanism; based on 、 , multiply the three feature matrices and divide by the normalization factor , send the result to the softmax activation function to activate and get the cutting feature The corresponding Attention scores of attention heads , and this attention score and Multiply to get the cutting feature The corresponding The output features of the attention head ; The output features Send it to the fully connected layer for fusion processing to obtain its multi-head attention output features , and output this feature and its corresponding cutting feature weight Multiply to get the multi-head attention weighted features , will get Multi-head attention weighted features Perform sum operation to get the feature Final features after processing by the improved multi-head attention mechanism .

[0014] Preferably, the scientific research results transformation method generator includes, in sequence: a local connection layer, two fully connected layers connected in sequence, a Relu activation function, an improved multi-head attention mechanism, a BN normalization layer, a global average pooling layer and a Relu function.

[0015] Preferably, the scenario guided training is as follows: transforming scientific research results into question answering datasets based on the scene graph embedding method Converting to a scene graph guided dataset (including basic semantic graph and extended knowledge graph ), the transformation scene guidance module is trained in combination with the reconstruction loss function, and the model parameters in the transformation scene guidance module are optimized using the SGD optimizer. The reconstruction loss function is as follows: in, Represents the Huber loss value calculation function, represents the output guidance graph, represents the extended knowledge graph, Represents the binary cross entropy loss value calculation function; The initial training is as follows: Using scientific research results text to familiarize yourself with the dataset The large model for transforming scientific research results is trained based on the overall model loss function, and the Adam optimizer is used to update the model parameters. The overall model loss function is as follows: , in, represents the overall loss function, Represents the Kl divergence loss value calculation function, Indicates the method of transforming the results. Indicates the data of scientific research results transformation methods, Represents the mean square error loss calculation function; The transformation intensive training is as follows: unfreeze the parameters of the large model for transforming scientific research results after the initial training is completed, and use the scientific research results to transform the question and answer dataset Combined with the overall loss function of the model, the structural parameters of the large model for the transformation of scientific research results are further strengthened through training, and the SGD optimizer is used to optimize the model parameters during model training.

[0016] The advantages of the present invention are: (1) Design of transformation scenario guidance module: This paper designs a transformation scenario guidance module based on graph knowledge, which can generate transformation condition background features that fit the dynamic actual transformation scenario according to the ever-changing achievement transformation environment. The transformation condition background features are integrated into the subsequent scientific research achievement transformation model, providing the current optimal transformation path for the subsequent scientific research achievement transformation method, further clarifying the transformation direction, and directly improving the feasibility of the implementation of the scientific research achievement transformation plan; (2) Design of a large-scale model architecture for the transformation of scientific research results: In the large-scale model architecture for the transformation of scientific research results designed by the present invention, the term associations in technical texts are effectively captured by designing an encoding vector module, and the improved multi-head attention mechanism is used to further improve the efficiency of feature interaction in complex transformation scenarios. In addition, the designed parser and generator stacking architecture takes into account both the technical documents of scientific research results and the actual transformation scenario condition data, so that the model can still maintain a stable transformation method generation capability in the emerging technology field, thereby improving the robustness and generalization ability of the scientific research results transformation method; (3) Design of a segmented model training strategy: This paper adopts a multi-stage training strategy for the transformation of scientific research results. First, the model is trained to understand the semantics of the research results text through preliminary training, then undergoes preliminary transformation training to enable it to have generation capabilities, and finally introduces scene information training to integrate the background characteristics of the transformation conditions. This approach can reduce the cost of model training, enhance the compatibility of the model with the actual transformation environment, and improve the practicality and accuracy of the scientific research results transformation method. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0018] Figure 1 It is a schematic diagram of the overall technical route flow of the present invention.

[0019] Figure 2 This is a schematic diagram of the conversion scenario guidance module structure.

[0020] Figure 3 Schematic diagram of the transformation environment feature perception module.

[0021] Figure 4 Schematic diagram of the task semantic reconstructor model architecture.

[0022] Figure 5 Schematic diagram of the large model for transforming scientific research results.

[0023] Figure 6 This is a schematic diagram of the architecture of the scientific research results transformation method analyzer.

[0024] Figure 7 This is a schematic diagram of the structure of the scientific research results transformation method generator. DETAILED DESCRIPTION

[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0026] A scientific research results transformation method based on a large graph knowledge model enhances the matching between the model and the actual transformation environment, and improves the practicality and accuracy of the scientific research results transformation method.

[0027] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions: S1: Construct a dataset of scientific research achievement transformation methods through multi-source data collection, including a scientific research achievement text familiarization dataset, a scientific research achievement transformation question and answer dataset, and a scenario guidance module dataset.

[0028] As a refinement of the above embodiment, the process includes: S1-1: Collection of Examples of Scientific Research Achievement Transformation: Comprehensive and feasible examples of scientific research achievement transformation can ensure the quality of the scientific research transformation methods proposed in the model. To this end, the present invention collects examples of scientific research achievement transformation from various channels, including the National Science and Technology Achievement Transformation Project Database, local government achievement transformation case databases, university technology transfer centers, school-local government cooperation platforms, listed companies, industrial parks, and third-party scientific research achievement transformation service platforms, to ensure the comprehensiveness of the collection of examples of scientific research achievement transformation.

[0029] S1-2: Collection of initial text data of scientific research results: Based on the collection method of scientific research results transformation examples described in S1-1, collect the intellectual property description documents, R&D process documents, business planning documents, policy and resource docking documents, transaction and legal documents, and project execution and acceptance documents involved in each scientific research results transformation example, and define the collected document data as scientific research results text data The above scientific research results text data It covers the entire process of scientific research results transformation and provides data support for subsequent models to provide excellent scientific research results transformation methods.

[0030] S1-3: Construction of a dataset for familiarizing with scientific research results text: This dataset is used for preliminary training of the model; specifically, the following steps are included: (1) Complete scientific research results text data collected based on S1-2 , the scientific research results text data Divide The length of the text segment is Text data ,and ; (2) Then extract each specific text The front texts constitute a scientific research results text. ,in ; (3) Extraction Middle Text to Words constitute a scientific research result text. Familiar with the output data in the data ; (4) Familiarize the text with the input data and output data from text familiarity data Together they constitute a set of scientific research results text familiarity datasets; (5) Based on the above dataset preparation method, we get A dataset of scientific research results text familiarity .

[0031] S1-4: Construction of a question-answering dataset for the transformation of scientific research results: Based on the complete scientific research results text data collected in S1-2 ,extract The text of the technical solution to be transformed in the scientific research results text data , and extract the corresponding conversion requirement data , and Jointly input data as scientific research results In addition, according to Mark the corresponding scientific research results transformation method data Specifically include the following: A description of the research results to be transformed, such as a composite catalyst for atmospheric-pressure and low-temperature ammonia synthesis and its preparation method. This invention uses mechanochemical activation and high-pressure hydrogen-nitrogen activation processes to construct a ternary composite catalyst system, breaking through the limiting relationship between the adsorption state and transition state energy of traditional catalysts, and realizing efficient ammonia synthesis reaction under atmospheric-pressure and low-temperature conditions.

[0032] This is a description of a specific conversion scenario. For example, a coal chemical enterprise with an annual synthetic ammonia production capacity of 200,000 to 300,000 tons would build a new catalyst production line according to the above-mentioned catalyst preparation method, and transform the heat exchange system of the original reaction unit during the enterprise's annual overhaul. In addition, the technical transformation is required to pass the acceptance of the national energy-saving inspection center.

[0033] Develop a specific implementation plan for the conversion of results in this scenario, such as the existing synthesis tower drawings, process flow charts, energy consumption data, measured tower wall corrosion rates, and scientific research results input data provided by the company , compiling a catalyst production line process package and designing a heat exchange system modification plan. Subsequently, the company completed specific equipment procurement and infrastructure preparation, followed by the construction of the catalyst production line and the modification of the reaction unit. Finally, the company conducted a coordinated trial run and commissioning, and submitted an energy-saving report for the technical modification project within one month of passing the trial run.

[0034] Scientific research results input data Data on methods and methods for transforming scientific research results Together they constitute a set of scientific research results transformation question-answering datasets, and based on this method, a total of Group data, together constitute the scientific research results transformation question answering dataset ; S1-5: Scenario-guided module dataset construction: Conversion requirement data in S1-4 It contains keyword information of specific scene conditions in the process of results conversion. However, the semantic content carried by each keyword text itself is limited, and the semantic associations between them are difficult to directly reflect as structured complex semantic relationships, making it difficult for large models to meet the requirements of deep semantic understanding and reasoning; Therefore, the present invention constructs a data set dedicated to the scene guidance module to systematically enrich the semantic information of scene keywords, and ultimately enhance the context expression ability and knowledge association ability of the large model; This data set includes the scientific research results keywords to be enriched , and enriched full-process keyword data The specific collection process includes: Keywords for scientific research results to be enriched First, the conversion demand data The extraction process uses the non-negative matrix factorization (NMF) algorithm, and initially obtains a set of keywords such as "ternary composite catalyst", "ammonia synthesis reaction", "annual production capacity of 200,000 to 300,000 tons", "coal chemical enterprise", "enterprise annual overhaul", and "energy-saving inspection center acceptance". Although the above keywords can describe certain scene elements, it is still difficult to fully represent the semantic information of the entire process involved in the conversion of results.

[0035] Further utilize more detailed and upstream and downstream specific results conversion implementation plans , semantically expand the above keywords to form complete full-process keyword data ;therefore Through specific implementation plans The technical information in the document was expanded and extracted, and keywords such as "synthesis tower drawings", "process flow chart", "energy consumption data", "inner wall corrosion rate", "compilation of catalyst production line process package", "heat exchange system transformation plan design", "equipment procurement", "infrastructure preparation", "production line construction", "linked trial run and debugging", "submission of energy-saving report", etc. were obtained.

[0036] Keywords of initial scientific research results The corresponding enriched full-process keyword data Together they constitute a set of scene guidance data, which are produced by repeating the above data collection and expansion process. Group data to obtain the conversion scenario guidance module data set .

[0037] S2: A large model for the transformation of scientific research results is constructed based on the parser-generator stack, and a module based on transformation scenario guidance is introduced; the large model for the transformation of scientific research results includes a coding vector module, a transformation scenario guidance module, a parser, a generator and a method output module. The coding vector module converts text data into a weighted coding vector, and the transformation scenario guidance module generates a multi-level transformation condition background feature vector based on the scenario guidance module data set. The method output module includes two fully connected layers connected in sequence and a Softmax activation function.

[0038] As a refinement of the above embodiment, in order to improve the effectiveness of the scientific research results transformation method and ensure that the large model can accurately meet the complex and diverse conditional constraints and semantic requirements in the actual scientific research results transformation scenario, the present invention constructs a transformation scenario guidance module based on graph knowledge; the core goal of this module is to achieve a deep understanding and multi-level guidance of the transformation scenario semantics through graph structure modeling. Therefore, firstly, a scene graph embedding method is designed to convert the input data set into a transformation scenario guidance graph that can represent complex topological relationships, including an input guidance graph and an output guidance graph. In the training stage of this module, the initial scientific research results keywords Building a basic semantic graph , enriched full-process keyword data Built as an extended knowledge graph with more complete contextual semantics .

[0039] This module takes the input guidance graph as input, learns to generate the output guidance graph containing upstream and downstream logic, causal relationships and task dependency structures, and extracts and generates multi-level transformation condition background feature vectors in the process. , Then, it will be integrated into the subsequent large-scale model architecture for the transformation of scientific research results to strengthen the understanding and adaptation capabilities of the transformation context and improve the feasibility of the large-scale model generation solution. The overall architecture diagram of the transformation scenario guidance module based on graph knowledge is as follows: Figure 2 As shown, specifically including: S2-1: Design of scene graph embedding method and construction of transformation scene guidance graph: First, the scene graph embedding method takes each keyword in the input data set as a node in the transformation scene guidance graph, converts each keyword text into a feature vector through the word embedding layer, and then calculates the cosine similarity between the feature vectors of each node. If the cosine similarity exceeds the set threshold, Then an edge is constructed between these two nodes, thereby realizing the construction of the transformation scene guide graph. The value is 0.8.

[0040] Specifically, in the training phase of this module, based on the scene guidance module dataset constructed in S1-5 , the initial scientific research results keywords are embedded into the scene graph Building a basic semantic graph , enriched full-process keyword data Built as an extended knowledge graph ; and the scene guide module dataset Converting to a scene graph guided dataset .

[0041] S2-2: Transformation environment feature sensor construction: Transformation environment feature perception model architecture such as Figure 3 As shown, the construction process is as follows: The transformation environment feature sensor is used to further process the features of the generated transformation scenario input guidance graph extracted in S2-1. To this end, the transformation environment feature sensor constructed in this invention first uses two sequentially connected graph convolutional layers to perform preliminary feature extraction, followed by a Relu activation function to achieve nonlinear feature activation. Subsequently, to enhance the ability to fit the input graph data during feature extraction, a dropout layer and a graph max pooling layer are added to the sensor module. Feature extraction is then performed using two sequentially connected graph convolutional layers and a Relu activation function to complete the construction of the transformation environment feature sensor.

[0042] S2-3: Task semantics reconstructor construction: The task semantics reconstructor model architecture is as follows Figure 4 As shown, the construction process is as follows: To generate a scientific research achievement transformation map that is more aligned with the real-world environment, a task semantics reconstructor is also needed to complete the overall design of the transformation scenario guidance module. In the task semantics reconstructor, three sequentially connected graph deconvolution layers are first used to upsample the feature data extracted by the perceptron module. Three sequentially connected fully connected layers are then used to integrate and transform the upsampled features. The features are then nonlinearly activated using the Mish activation function. Three sequentially connected graph deconvolution layers are then used to further upsample the features. Finally, a graph attention layer introduces attention factor judgment, and the Simgoid activation function is used to complete the feature activation, completing the construction of the task semantics reconstructor.

[0043] S2-4: Transformation Scenario Guidance Module Architecture Design: Utilize the transformation environment feature sensor constructed in S2-2 and the task semantics reconstructor constructed in S2-3 to complete the sensor-reconstructor architecture design used in the transformation scenario guidance module; In the transformation scenario guidance module, three interconnected transformation environment feature sensors are first used to extract features from the input guidance graph to obtain the intrinsic feature relationship between each keyword variable in the scientific research results input data.

[0044] Ten sequentially connected task semantic reconstructors are used to perform multi-level transformation condition background feature reconstruction operations to gradually generate an output guidance map that fits the actual scientific research results transformation background. . In the obtained multi-level transformation condition background characteristics In particular, the background characteristics of the transformation conditions This is the output guide map for the final reconstruction of the model Based on the different levels of conversion condition background features extracted by this module ,The background features at each level contain the complex topological structure information between ,the various transformation conditions when actual scientific research results are transformed.

[0045] This multi-level transformation condition background feature Introducing it into the large model of scientific research results transformation will help the model correctly understand the complex influence relationship between specific transformation scenario conditions and the scientific research results to be transformed when the model is carrying out actual scientific research results transformation, further improve the fit between the proposed scientific research results transformation method and the actual transformation environment, and ensure the implementation ability of the scientific research results transformation method proposed by the model in actual transformation operations.

[0046] As a refinement of the above embodiment, the overall architecture of the scientific research results transformation model is as follows: Figure 5 The specific process is as follows: S3-1: Construction of encoding vector module: In order to convert text data into a data format that can be directly processed by the large model, the present invention constructs an encoding vector module to perform vector encoding on text data and convert the text data into vector data. In the vector encoding module, the input data in the input data set (scientific research results input data or scientific research results text familiar data input data) is first divided into characters one by one. , and for each word Perform random initialization vector encoding with a vector dimension of 50, and define the initialization encoding vector corresponding to each word as .

[0047] After the random initialization vector encoding is completed, the context influence weight vector of each word is calculated , and add it to the vector code corresponding to the word itself to obtain the weighted code vector corresponding to the word Then, the weighted coding vectors corresponding to each word are combined to obtain this text The corresponding encoding vector .Context influence weight vector The calculation method is as follows: For the word to be encoded In terms of The initialization vectors corresponding to the words are Initialization code vector , and assign an adaptive adjustment weight to each encoding vector , in the subsequent training of the model, the SGD optimizer is used to adaptively adjust the weights Gradually optimize as the model is trained to achieve the best weight distribution. And each initialization code vector and its corresponding adaptive adjustment weight Multiply and get The products are added together to get the word to be encoded The corresponding context influence weight vector .

[0048] In the subsequent model training process, the initialization encoding vector of the above process And adaptively adjust weights It will be adaptively updated according to the update rules set in the model optimizer to obtain the optimal encoding vector representation corresponding to the text; S3-2: Design of the parser module for scientific research results transformation methods: The structural diagram of the parser module for scientific research results transformation methods is as follows: Figure 6 As shown, the construction process is as follows: The text encoding vector obtained based on the above S3-1 process In order to ensure the effectiveness of the scientific research results transformation method of the present invention, a parser module is designed to encode the vector To achieve further feature extraction. In the parser module, the vector Perform feature extraction processing, and send the feature vector obtained after processing the two local connection layers into the Relu activation function for nonlinear activation to obtain the feature In order to further improve the model’s ability to analyze various scientific research documents, an improved multi-head attention mechanism is designed in the parser module to Do further processing.

[0049] The improved multi-head attention mechanism is constructed as follows: (1) The features obtained Perform random cutting to obtain Cut features , , then cut the feature Send it to the polymorphic perceptron for parameter extraction and processing to obtain cutting features The corresponding cutting feature weight , this multi-state perceptron is composed of 5 layers of interconnected fully connected layers and Silu activation functions.

[0050] (2) Cutting features Feed into the matrix The calculation function calculates the The feature matrix corresponding to the attention head , cut feature Feed into the matrix The calculation function calculates the The feature matrix corresponding to the attention head , cut feature Feed into the matrix The calculation function calculates the The feature matrix corresponding to the attention head , , is the total number of heads included in the multi-head attention mechanism.

[0051] (3) Based on 、 , multiply the three feature matrices and divide by the normalization factor , send the result to the softmax activation function to activate and get the cutting feature The corresponding Attention scores of attention heads , and this attention score and Multiply to get the cutting feature The corresponding The output features of the attention head .

[0052] (4) Output features Send it to the fully connected layer for fusion processing to obtain its multi-head attention output features , and output this feature and its corresponding cutting feature weight Multiply to get the multi-head attention weighted features , will get Multi-head attention weighted features Perform sum operation to get the feature Final features after processing by the improved multi-head attention mechanism .

[0053] Then, two fully connected layers and Dropout layers are used to further process the features, and the feature processing results are sent to the Simgoid activation function for final activation to obtain the feature vector processed by the parser unit. ; S3-3: Design of the scientific research results transformation method generator module: The structural diagram of the scientific research results forwarding method generator module is as follows: Figure 7 As shown, the text description is as follows: After using the parser to extract the features of the input text feature vector, the generator needs to be used to further process the output of the parser in the constructed scientific research achievement transformation model to obtain the final scientific research achievement forwarding method. The scientific research achievement transformation method generator module first includes a local connection layer, and then uses two fully connected layers and Relu activation functions connected in sequence to achieve further feature processing. In order to ensure the effectiveness of the generated scientific research achievement transformation method, the multi-head attention mechanism described in S2-2 is also introduced in the generator module. After the output of the multi-head attention mechanism is finally processed using the BN normalization layer, the global average pooling layer and the Relu function, the construction of the generator module is completed.

[0054] S3-4: Overall architecture design of the large model for transformation of scientific research results: Based on the encoding vector module described in the S3-1 process, the scientific research result transformation parser module described in S3-2, and the scientific research result transformation method generator module described in S3-3, the overall architecture design of the large model for transformation of scientific research results is completed.

[0055] The input data in the input data set (scientific research results input data or scientific research results text familiar data input data) is sent to the encoding vector module to obtain the encoding vector corresponding to the input text , input the scientific research results input data into the transformation scenario guidance module to obtain multi-level transformation condition background characteristics .

[0056] The encoding vector The output is sent to 10 stacked scientific research results transformation parsers for deep processing. In the parser stacking structure, the output of the upper parser is not only used as the input of the lower parser, but also the output of the same level transformation condition background features obtained by the transformation scenario guidance module. Perform channel splicing operation and use the output vector after channel splicing as the input of the generator at the same level.

[0057] In addition, the number of stacking layers of the scientific research results transformation method generator is also 10. In the generator stacking structure, the output of the lower-level generator is used as the input of the upper-level generator. After the output of the top-level generator is sent to the method output module for final processing, the final large model output of the scientific research results transformation method is obtained. In the method output module, it includes two fully connected layers connected in sequence and a Softmax activation function.

[0058] S3: A segmented training strategy of scenario-guided training, preliminary training and conversion-enhanced training is adopted, and the model is optimized through a joint loss function; the scenario-guided stage training uses the scenario-guided module dataset to train the conversion scenario-guided module separately, and freezes the conversion scenario-guided module parameters after the training is completed; the preliminary training stage uses the scientific research results text familiarization dataset to train the scientific research results conversion large model; the conversion-enhanced training stage uses the scenario-guided module dataset to train the scientific research results conversion large model.

[0059] As a refinement of the above embodiment, when the model is actually used, the transformation of scientific research results is related to the actual transformation environment and transformation conditions. These factors will affect the practicality of the scientific research results transformation method. In order to match the ever-changing scientific research results transformation environment and transformation conditions, the present invention adopts a segmented training strategy. First, the transformation scenario guidance training is carried out to ensure that the model can smoothly obtain multi-level transformation condition background features. After that, the initial training of the large model is carried out to ensure that the model has a preliminary understanding of the context of scientific research results. Finally, the scientific research results transformation intensive training is carried out to transform the multi-level transformation condition background features. Integrating this into model training further enhances the compatibility between scientific research achievement transformation methods and the actual transformation environment. This multi-level training approach not only reduces model training costs but also further improves the integration of the model with the actual scientific research achievement transformation environment.

[0060] Specifically include: S4-1: Loss function construction: In order to adapt to the overall architecture of the scientific research achievement transformation model proposed by the present invention, which integrates the transformation scenario guidance module, an overall loss function for evaluating the scientific research achievement transformation method output by the large model is designed to calculate the loss value predicted by the model method. In addition, a reconstruction loss function is designed to evaluate the conversion scene output guidance graph generated by the conversion scene guidance module to calculate the graph reconstruction loss value. ; in, represents the overall loss function, Represents the Kl divergence loss value calculation function, Indicates the method of transforming the results. Indicates the data of scientific research results transformation methods, Represents the mean square error loss calculation function; Represents the Huber loss value calculation function, represents the output guidance graph, represents the extended knowledge graph, Represents the binary cross entropy loss calculation function.

[0061] S4-2: Transformation scenario guidance training: Since the transformation scenario guidance module in the scientific research results transformation model contains a graph neural network layer, in order to ensure the model training effect of the transformation scenario guidance module, the model can smoothly output multi-level transformation condition background features. The present invention uses the scene graph guidance data set constructed in S2-1 to transform the scene guidance module Pre-training was carried out. The transformation scenario guidance module was trained using the reconstruction loss function constructed in S4-1, and the model parameters in the transformation scenario guidance module were optimized using the SGD optimizer. Through separate transformation scenario guidance module training, it was ensured that the transformation scenario guidance module could accurately learn the complex topological relationship between the actual transformation environment of scientific research results and the scientific research results to be transformed. This is conducive to accurately extracting the specific background features of the scientific research results transformation scenario, and helping to achieve a more effective combination of the scientific research results to be transformed and the scientific research results transformation environment. After the transformation scenario guidance training is completed, the model parameters of the transformation scenario guidance module are frozen.

[0062] S4-3: Initial model training: After the conversion scenario guidance training is completed, in order to make the model familiar with the text environment of scientific research results, the present invention carries out initial model training to help the model familiarize itself with the contextual connection relationship of scientific research results. The scientific research results text familiarization dataset constructed in S1-3 is used. The model is trained using the overall loss function designed in S4-1, and the Adam optimizer is used to update the model parameters. After the initial training is completed, the model is guaranteed to understand the contextual semantics of the scientific research results. After the initial training is completed, the large model parameters obtained from the training are frozen.

[0063] S4-4: Strengthening training for transformation of scientific research results: Unfreeze the transformation scenario guidance module parameters after the transformation scenario guidance training in S4-1 to introduce specific scientific research transformation environment characteristics into the scientific research transformation model designed by this invention, and help the model achieve efficient output of scientific research results transformation solutions based on the initial training of the S4-3 model. This stage uses the scientific research results transformation question and answer dataset constructed in S1-4 Further intensive training of the overall structural parameters of the large model was carried out. The SGD optimizer was used to optimize the model parameters during model training. After this training was completed, the final large model for scientific research results conversion was obtained.

[0064] S4. Transform the trained scientific research results into large-scale models for deployment and application.

[0065] As a refinement of the above embodiment, a scientific research achievement transformation environment computing server is deployed to provide necessary hardware equipment support for the specific deployment of the scientific research achievement transformation model. The trained scientific research achievement transformation model is then deployed to the server terminal, and the corresponding software equipment debugging is performed to ensure that the data input interface and data output interface of the model can smoothly realize data reading and output. After the model is deployed, the specific description text of the scientific research achievement to be transformed and the scientific research achievement requirements are input into the scientific research achievement transformation model to obtain an adapted scientific research achievement transformation method, and the actual scientific research achievement transformation operation is carried out according to this method.

[0066] To further verify the effectiveness of the output guidance module proposed in this paper in improving model performance, we conducted performance tests using and not using the conversion scenario guidance module. Each output of the model for the research achievement conversion method was scored based on professionalism, practicality, conversion cost, and solution implementation. The final score, calculated by averaging a large number of scores, is shown in Table 1 below, with a score range of [0, 1].

[0067] Table 1 Performance test of the model with and without the conversion scenario guidance module During the model performance test, performance comparison tests were carried out in four different scientific research fields, namely agriculture, intelligent manufacturing, engineering construction and computer. The scores are shown in the table above. By analyzing the table data, it can be found that when the transformation scenario guidance module is used, the scientific research result transformation method proposed by the model is more professional, practical and feasible, and the transformation cost is lower. This is because the transformation scenario guidance module proposed by the present invention can accurately capture the upstream and downstream logic, causal relationship and task dependency between the transformation environment and the scientific research results to be transformed during the transformation of scientific research results, helping the model to better understand the semantics of complex scenarios. In addition, this module converts the transformation scenario semantics into computable structured information, and integrates the multi-level features containing complex topological structure information into the large model, providing accurate scenario guidance for the large model, thereby maintaining a good ability to generate scientific research results transformation methods under complex constraints.

[0068] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for transforming scientific research results based on a large graph knowledge model, characterized in that: The following steps are involved: S1. Construct a dataset for scientific research achievement transformation methods through multi-source data collection, including a scientific research achievement text familiarization dataset, a scientific research achievement transformation question-answering dataset, and a scenario guidance module dataset; S2. Construct a large model for the transformation of scientific research results based on the parser-generator stack, and introduce a transformation scenario guidance module; the large model for the transformation of scientific research results includes a coding vector module, a transformation scenario guidance module, a parser, a generator, and a method output module. The coding vector module converts text data into a weighted coding vector. The transformation scenario guidance module generates a multi-level transformation condition background feature vector based on the scenario guidance module data set. The method output module includes two fully connected layers connected in sequence and a Softmax activation function. S3, adopting a segmented training strategy of scenario-guided training, preliminary training and transformation-enhanced training, and optimizing the model through the loss function; The scenario guidance training phase uses the scenario guidance module dataset to train the conversion scenario guidance module separately, and freezes the parameters of the conversion scenario guidance module after training. The preliminary training phase uses the scientific research results text familiarization dataset to train the scientific research results conversion model. The conversion intensive training phase uses the scientific research results conversion question and answer dataset to train the scientific research results conversion model. S4. Transform the trained scientific research results into large-scale models for deployment and application.

2. The method for transforming scientific research results based on a graph knowledge large model according to claim 1 is characterized in that: The datasets for scientific research results transformation methods are constructed through multi-source data collection, including: Collect examples of scientific research results transformation and extract text data covering the entire process of scientific research results transformation; Using sliding window mechanism to construct scientific research results text familiarity dataset , including: segmenting the scientific research results text data, extracting the first half of each segment as the input data of the scientific research results text familiarity data, and the second half as the output data of the scientific research results text familiarity data, together forming a set of scientific research results text familiarity data, and the scientific research results text familiarity data composed of all text segments as the scientific research results text familiarity data set ; Constructing a question-answering dataset for the transformation of scientific research results , including: extracting the text of the technical solution to be converted from the scientific research results text data as the scientific research results text data , extract the corresponding conversion requirement data , together as input data for scientific research results , based on the text data of scientific research results Mark the corresponding scientific research results transformation method data , input scientific research results into data Data on methods and methods for transforming scientific research results Constructing a question-answering dataset for the transformation of scientific research results ; Constructing a scene guidance module dataset via non-negative matrix factorization and semantic expansion , including: extracting transformation requirement data through non-negative matrix factorization algorithm Keywords in the research results to be enriched , according to the data of scientific research results transformation method Keywords for enriching scientific research results Perform semantic expansion to obtain enriched keyword data for the entire process , to be enriched with scientific research results keywords and enriched keyword data throughout the entire process Together they constitute the scene guidance module dataset .

3. The method for transforming scientific research results based on a graph knowledge large model according to claim 2 is characterized in that: The transformation scenario guidance module includes: a graph generation module, a transformation environment feature sensor and a task semantics reconstructor; The graph generation module converts the input data set into a transformation scene guidance graph that can represent complex topological relationships through a scene graph embedding method, including an input guidance graph and an output guidance graph; The input guidance image is fed into the transformation environment feature sensor for feature extraction and is processed by three transformation environment feature sensors in sequence; The extracted features are sequentially passed through ten task semantic reconstructors to perform multi-level transformation condition background feature reconstruction operations. Each task semantic reconstructor extracts the transformation condition background feature vector and gradually generates an output guidance map that fits the actual scientific research results transformation background.

4. The method for transforming scientific research results based on a graph knowledge large model according to claim 3 is characterized in that: The scene graph embedding method is specifically as follows: Each keyword in the input data set is treated as a node in the graph data, and each keyword text is converted into a feature vector through the word embedding layer; The cosine similarity between the feature vectors of each node is calculated. If the cosine similarity exceeds the set threshold, an edge is constructed between the two nodes to realize the construction of the transformation scene guidance graph.

5. The method for transforming scientific research results based on a graph knowledge large model according to claim 3 is characterized in that: The transformation environment feature sensor processes the input guidance graph as follows: The input guide graph is extracted through two sequentially connected graph convolutional layers, and the features are nonlinearly activated using the Relu activation function. The activated features are sequentially processed by the Dropout layer and the graph maximum pooling layer, and then two sequentially connected graph convolutional layers are used to extract features through the Relu activation function. The task semantic reconstructor processes the input features as follows: The feature data extracted by the perceptron module is upsampled through three sequentially connected graph deconvolution layers, and the upsampled features are integrated and transformed using three sequentially connected fully connected layers. The features are nonlinearly activated by the Mish activation function, and further upsampled by three sequentially connected graph deconvolution layers. An attention factor is introduced through a graph attention layer, and the final activation of the features is completed by the Simgoid activation function.

6. The method for transforming scientific research results based on a graph knowledge large model according to claim 3 is characterized in that: The data processing method of the large model for transforming scientific research results is as follows: The input data in the input data set is fed into the encoding vector module to obtain the encoding vector corresponding to the input data. , the input data in the input data set is fed into the transformation scene guidance module to obtain multi-level transformation condition background features , ; The encoding vector The output of the upper parser is sent to the stacked scientific research results transformation parser for deep processing. In the parser stacking structure, the output of the upper parser is not only used as the input of the lower parser, but also the output of the same level transformation condition background feature obtained by the transformation scenario guidance module. Perform channel splicing operations and use the output vector after channel splicing as the input of the generator of scientific research results transformation methods at the same level; In the stacked structure of scientific research results transformation method generators, the output of the lower-level generator is used as the input of the upper-level generator, and the output of the top-level generator is sent to the method output module for final processing to obtain the output scientific research results transformation method.

7. The method for transforming scientific research results based on a graph knowledge large model according to claim 6 is characterized in that: The specific way in which the encoding vector module converts text data into encoding vectors is as follows: Divide the input data into word , Indicates the total number of words in the input data set, and for each word The random initialization vector is encoded as , the vector dimension is 50; Calculate the context influence weight vector for each word , and add it to the vector code corresponding to the word itself to obtain the weighted code vector corresponding to the word , the weighted coding vector corresponding to each word is combined to obtain the coding vector corresponding to the text ; Context influence weight vector The calculation method is as follows: Get the word to be encoded about The initialization vectors corresponding to the words are Initialize the encoding vectors and assign an adaptive adjustment weight to each encoding vector ; Multiply each initialization code vector by its corresponding adaptive adjustment weight, and obtain The word to be encoded is obtained by adding the products The corresponding context influence weight vector .

8. The method for transforming scientific research results based on a graph knowledge large model according to claim 6 is characterized in that: The research results are converted into parsers for encoding vectors The in-depth analysis method is as follows: Encoded vector The features are extracted through two local connection layers and then sent to the Relu activation function for nonlinear activation to obtain the features. ; By improving the multi-head attention mechanism to Further processing to obtain the final features ; The final feature Through two fully connected layers and Dropout layers connected in sequence, and then sent to the Simgoid activation function to activate, the feature vector processed by the scientific research results conversion parser is obtained. ; The improved multi-head attention mechanism is as follows: The features that will be obtained Perform random cutting to obtain Cut features , Cut feature Send it to the polymorphic perceptron for parameter extraction and processing to obtain cutting features The corresponding cutting feature weight , the multi-state perceptron is composed of 5 layers of interconnected fully connected layers and Silu activation functions; Cut feature Feed into the matrix The calculation function calculates the The feature matrix corresponding to the attention head , cut feature Feed into the matrix The calculation function calculates the The feature matrix corresponding to the attention head , cut feature Feed into the matrix The calculation function calculates the The feature matrix corresponding to the attention head , , is the total number of heads included in the multi-head attention mechanism; based on 、 , multiply the three feature matrices and divide by the normalization factor , send the result to the softmax activation function to activate and get the cutting feature The corresponding Attention scores of attention heads , and this attention score and Multiply to get the cutting feature The corresponding The output features of the attention head ; The output features Send it to the fully connected layer for fusion processing to obtain its multi-head attention output features , and output this feature and its corresponding cutting feature weight Multiply to get the multi-head attention weighted features , will get Multi-head attention weighted features Perform sum operation to get the feature Final features after processing by the improved multi-head attention mechanism .

9. The method for transforming scientific research results based on a graph knowledge large model according to claim 8 is characterized in that: The scientific research results transformation method generator includes, in sequence: a local connection layer, two fully connected layers connected in sequence, a Relu activation function, an improved multi-head attention mechanism, a BN normalization layer, a global average pooling layer and a Relu function.

10. The method for transforming scientific research results based on a graph knowledge large model according to claim 4 is characterized in that: The scenario-guided training is as follows: Based on the scene graph embedding method, the scientific research results are transformed into question-answering datasets Converting to a scene graph guided dataset , combined with the reconstruction loss function to train the conversion scene guidance module, and use the SGD optimizer to optimize the model parameters in the conversion scene guidance module. The reconstruction loss function is as follows: in, Represents the Huber loss value calculation function, represents the output guidance graph, represents the extended knowledge graph, Represents the binary cross entropy loss value calculation function; The initial training is as follows: Using scientific research results text to familiarize yourself with the dataset The large model for transforming scientific research results is trained based on the overall model loss function, and the Adam optimizer is used to update the model parameters. The overall model loss function is as follows: ， in, represents the overall loss function, Represents the Kl divergence loss value calculation function, Indicates the method of transforming the results. Indicates the data of scientific research results transformation methods, Represents the mean square error loss calculation function; The transformation intensive training is as follows: unfreeze the parameters of the large model for transforming scientific research results after the initial training is completed, and use the scientific research results to transform the question and answer dataset Combined with the overall loss function of the model, the structural parameters of the large model for the transformation of scientific research results are further strengthened through training, and the SGD optimizer is used to optimize the model parameters during model training.

Images