Scientific and technological service supply and demand intelligent matching method and device based on deep learning and medium

Abstract

The invention discloses a science and technology service supply and demand intelligent matching method and device based on deep learning and a medium, and relates to the technical field of deep learning, and the method comprises the steps: constructing a multi-dimensional data set covering the science and technology service field, carrying out the post-pre-training of a pre-training language model based on multi-task learning, and constructing a knowledge graph auxiliary sample based on a knowledge graph; based on the knowledge graph auxiliary sample pair, performing comparative learning on the first optimization model to obtain a second optimization model; and historical interaction data of a real user is collected to optimize the second optimization model to obtain a rearrangement model, then multi-dimensional correlation scoring is performed on candidate services, and a final recommendation result is obtained based on weighted sorting. Efficient, intelligent and accurate matching of science and technology service supply and demand data is realized by constructing a field specific data set, designing innovative multi-task transition training and field specific tasks, introducing an adaptive multi-task weight optimization mechanism, providing a knowledge enhancement comparison learning method and combining a two-stage retrieval and rearrangement strategy.

Description

Technical Field

[0001] The present invention relates to the field of deep learning technology, and in particular to methods, devices, and media for intelligent matching of supply and demand of scientific and technological services based on deep learning. Background Art

[0002] With the rapid development of the science and technology service industry, the problem of information asymmetry between technology demanders and service providers has become increasingly prominent. Traditional science and technology service matching mainly relies on manual retrieval or simple keyword matching methods, which have the disadvantages of low matching accuracy, insufficient efficiency, and the inability to fully explore the deep semantic relationship between demand and service. In recent years, the advancement of deep learning technology has provided new possibilities for intelligent matching. In particular, the widespread application of pre-trained language models in the field of natural language processing has significantly improved the performance of semantic understanding and text matching. However, existing matching methods based on pre-trained language models are mostly designed for general fields, and lack targeted adaptation to the specific semantics, industry terminology and complex demand scenarios in the field of science and technology services. As a result, the model performance is limited when dealing with highly professional and contextually complex science and technology service supply and demand matching tasks.

[0003] Furthermore, existing technologies often employ static weight allocation strategies during multi-task joint training, making it difficult to balance the convergence speed and optimization objectives between different tasks, resulting in inefficient training or poor model performance. At the same time, existing matching systems generally struggle to balance retrieval efficiency and matching accuracy when faced with large-scale data. This is particularly true in application scenarios with high real-time requirements, where system response speed becomes a key bottleneck restricting user experience. Furthermore, knowledge in the field of scientific and technological services updates rapidly, and existing models lack dynamic adaptation and knowledge enhancement mechanisms, making it difficult to address diverse needs across industries and scenarios.

[0004] To sum up, the existing technologies have significant deficiencies in terms of domain adaptability, training efficiency, matching accuracy and system scalability in matching the supply and demand of scientific and technological services. There is an urgent need for an efficient and intelligent matching method that can be designed according to the characteristics of the scientific and technological service field to solve the above problems and enhance the platform service capabilities. Summary of the Invention

[0005] The technical problem to be solved by the present invention is that the existing technology has significant deficiencies in the field adaptability, training efficiency, matching accuracy and system scalability of technology service supply and demand matching. The purpose is to provide a method, equipment and medium for intelligent matching of technology service supply and demand based on deep learning, which can achieve efficient and accurate supply and demand matching through domain-specific data set construction, multi-task joint training, adaptive weight optimization, knowledge-enhanced comparative learning and two-stage retrieval and re-ranking strategy, and can realize accurate and efficient connection between technical requirements and services, significantly improving matching accuracy, system efficiency and user experience.

[0006] The present invention is achieved through the following technical solutions: The first aspect of the present invention provides a method for intelligent matching of supply and demand of scientific and technological services based on deep learning, comprising the following specific steps: Obtain the historical user demand and technology service interaction data accumulated by the technology service platform to build a technology service supply and demand matching data set; Constructing a pre-trained language model, and performing post-pretraining on the pre-trained language model based on multi-task learning to obtain a first optimized model; Use the path distance between entities in the knowledge graph as a metric for sample difficulty to perform knowledge graph-assisted sample pair classification; Based on the knowledge graph-assisted sample pairs, the first optimization model is compared and learned to obtain the second optimization model; Collect historical interaction data of real users to optimize the second optimization model and obtain a rearrangement model; Based on the re-ranking model, the candidate services are scored for multi-dimensional relevance, and the final recommendation results are obtained based on weighted sorting.

[0007] Furthermore, the acquisition of historical user demand and technology service interaction data accumulated by the technology service platform and the construction of a technology service supply and demand matching data set specifically include: Obtain user demand and technology service interaction data based on the historical accumulation of the technology service platform, extract technology demand text, technology service text and correlation score as the technology service supply and demand matching data set; The technical requirement text is a description of the technical service requirements submitted by the user; The technology service text is a description of the technology services provided by the platform; The relevance score is generated by combining manual annotation with predictions from a scoring model based on historical user interaction behaviors.

[0008] Furthermore, in the multi-task learning, the multiple tasks specifically include: a mask task, a term prediction task, a regression task, and a sorting task.

[0009] Furthermore, the post-pre-training of the pre-trained language model based on multi-task learning specifically includes: Based on the masking task, the input text of the pre-trained language model is randomly masked; the pre-trained language model performs semantic understanding of the text based on the remaining visible text; based on the semantic understanding of the text, the masked words are obtained; the correlation between the words predicted by the model and the actual masked words is calculated to obtain the prediction accuracy; based on the prediction accuracy, the model parameters are optimized using the cross-entropy loss function; Based on the term prediction task, the domain dictionary is used to extract key technical terms from the input text; the extracted key technical terms are randomly selected and replaced with tokens in the text; the token-replaced key technical terms are predicted based on the context information; the cross-entropy loss function is used to calculate the difference between the predicted key technical terms and the actual key technical terms, and the model parameters are optimized based on the difference; Based on the regression task, the demand text and service text pairs are extracted from the input text; a relevance score is assigned to each text pair; the demand text and service text pairs are encoded using a pre-trained language model, and the semantic feature vectors of the encoded demand text and service text pairs are extracted; the relevance score is predicted based on the extracted feature vectors; a loss function is constructed based on the relevance score prediction to optimize the model parameters; Sorting task: Based on the relevance score of the demand text and service text pairs, a loss function is constructed based on the relevance score to optimize the model's sorting ability.

[0010] Furthermore, when training the pre-trained model, the task weights during multi-task learning are dynamically adjusted. The dynamic adjustment process includes: Set the training round and get the current number of training rounds; Get the progress ratio based on the training rounds and the current number of training rounds; According to the progress ratio, the transition factor is obtained; The weights of the mask task, term prediction task, regression task, and ranking task are set according to the transition factor.

[0011] Furthermore, the acquisition of the knowledge graph in the field of scientific and technological services and the construction of knowledge graph auxiliary sample pairs specifically include: Generate a knowledge graph in the field of science and technology services based on the science and technology service supply and demand matching dataset; In the knowledge graph of the field of science and technology services, demand-service pairs belonging to the same technical sub-field or with strong semantic associations are defined as positive samples, and demand-service pairs that cross technical fields or are semantically unrelated are defined as negative samples.

[0012] Furthermore, the method of performing comparative learning on the first optimization model based on the knowledge graph-assisted sample pairs to obtain the second optimization model specifically includes: Obtain nodes of the knowledge graph in the field of science and technology services and extract node path distances; Nodes whose node path distance is higher than the set threshold are extracted and defined as hard-to-classify samples; Construct a contrastive learning loss function, weight the difficult samples, optimize the sample representation, and obtain the second optimization model.

[0013] Furthermore, the collecting of historical interaction data of real users to optimize the second optimization model to obtain a rearrangement model specifically includes: Get historical interaction data of real users, The second optimization model is trained based on historical interaction data of real users to obtain a rearrangement model.

[0014] A second aspect of the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a method for intelligently matching supply and demand of technology services based on deep learning is implemented.

[0015] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for intelligent matching of supply and demand of scientific and technological services based on deep learning.

[0016] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. Improve domain adaptability: In view of the professionalism and complexity of the scientific and technological service field, we design domain-specific dataset construction methods and pre-training tasks. Through semantic enhancement and knowledge graph integration, we improve the model's ability to understand industry terminology, technical background, and demand context, thereby addressing the limitations of general pre-training models in professional field applications.

[0017] 2. Optimize training efficiency and performance: Introduce an adaptive multi-task weight optimization mechanism to dynamically adjust the weight distribution of multiple tasks such as regression and sorting, reduce interference between tasks, improve training efficiency and overall model performance, and ensure the stability and reliability of the system in complex matching tasks.

[0018] 3. Enhanced semantic matching accuracy: A knowledge-enhanced comparative learning method is proposed, which uses domain knowledge graphs to construct high-quality sample pairs and optimizes the model's ability to discern fine-grained semantic relationships between technical requirements and services, thereby achieving more accurate supply and demand matching.

[0019] 4. Balancing efficiency and precision: Design a two-stage retrieval and re-ranking strategy, screen candidate services through efficient semantic retrieval, and combine it with a refined re-ranking model to perform multi-dimensional relevance scoring. This ensures that while ensuring matching accuracy, the system response speed is significantly improved to meet the real-time requirements in large-scale data scenarios.

[0020] 5. Achieve cross-scenario versatility: Build a modular and scalable technical framework that supports flexible adjustment of task design and optimization strategies according to different industries and application scenarios. At the same time, through the dynamic update of user interaction data and knowledge graphs, ensure the adaptability and continuous optimization capabilities of the model across domains and scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the following briefly introduces the drawings required for use in the examples. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be considered as limiting the scope. A person of ordinary skill in the art can also derive other relevant drawings based on these drawings without inventive effort. In the drawings: Figure 1 Schematic diagram of the matching method in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.

[0023] As a possible implementation, Figure 1 As shown, this embodiment provides a method for intelligent matching of supply and demand of technology services based on deep learning, including the following specific steps: obtaining user demand and technology service interaction data historically accumulated on the technology service platform, and constructing a technology service supply and demand matching data set; constructing a pre-trained language model, and post-pretraining the pre-trained language model based on multi-task learning to obtain a first optimization model; using the path distance between entities in the knowledge graph as a measure of sample difficulty, performing knowledge graph-assisted sample pair classification; based on the knowledge graph-assisted sample pairs, performing comparative learning on the first optimization model to obtain a second optimization model; collecting real users' historical interaction data to optimize the second optimization model to obtain a rearrangement model; based on the rearrangement model, performing multi-dimensional relevance scoring on the candidate services, and obtaining the final recommendation result based on weighted sorting.

[0024] In this example, a domain-specific Technical Term Prediction (TTP) task was designed by constructing a multidimensional dataset covering the science and technology service sector and enhancing technical terminology, industry context, and semantic relationships with a domain knowledge graph. This significantly improves the pre-trained language model's ability to understand complex semantics in the science and technology domain. An adaptive multi-task weight optimization mechanism was introduced. Through dynamic weighted averaging and reinforcement learning-assisted optimization strategies, the weight distribution of multiple tasks, such as regression and ranking, was adjusted in real time based on task convergence speed and difficulty, effectively reducing inter-task interference and improving training efficiency and model performance. High-quality positive and negative sample pairs were constructed using the science and technology service domain knowledge graph. A weighted InfoNCE loss function was used to prioritize difficult-to-classify samples, enhancing the model's ability to discern fine-grained semantic relationships between technical requirements and services. A two-stage retrieval and re-ranking strategy was implemented. In the first stage, semantic vectors were used to efficiently retrieve and screen candidate services. In the second stage, a refined re-ranking model was used to perform multi-dimensional relevance scoring, combined with a weighted ranking algorithm to output the final matching results. This strategy significantly improves system efficiency while ensuring matching accuracy. The versatility of the model is ensured by the dynamic update of user historical interaction data and domain knowledge graphs. This embodiment significantly reduces user search costs and improves user satisfaction by accurately matching technical needs with scientific and technological services; at the same time, through efficient retrieval and intelligent recommendation, it optimizes the resource allocation efficiency of the scientific and technological service platform and enhances the platform's service capabilities and market competitiveness. Compared with existing technologies, it has significant advantages in domain adaptability, semantic understanding capabilities, training efficiency, matching accuracy, system performance and application scalability. Its innovative technical solution not only fills the technical gap in the field of intelligent matching of supply and demand of scientific and technological services, but also demonstrates excellent results in practical applications, and has important theoretical value and commercial prospects.

[0025] Specific implementation steps include: 1. Construction of a data set for matching supply and demand of science and technology services: This embodiment constructs a high-quality, domain-specific training dataset based on historical user demand and technology service interaction data accumulated by the technology service platform. The dataset is based on the basic units of (technology requirement text, technology service text, and relevance score). The technology requirement text is the description of the technology service requirement submitted by the user, and the technology service text is the description of the technology service content provided by the platform. The relevance score is generated by combining manual annotation with a scoring model prediction based on historical user interaction behaviors (such as clicks, inquiries, and transactions), and uses a numerical value ranging from 0 to 1 to quantify the degree of match. To improve data quality, this embodiment introduces a domain knowledge graph to semantically enhance the data, annotating key entities (such as technology fields, application scenarios, and core technical points) in the technology requirement and service texts and their relationships to ensure the accuracy and representativeness of the data at the semantic level. Furthermore, through data cleaning, denoising, and diverse sampling, training data covering multiple industries and scenarios is constructed, laying a solid foundation for subsequent model training.

[0026] 2. Training phase of pre-trained language models based on multi-task learning: multi-task joint training and domain-specific task design: Post-pre-training is performed on the basis of pre-trained language models (such as BERT) to reduce the difference between pre-training and domain fine-tuning. In the post-pre-training stage, multi-task joint optimization and domain-specific task design are used to improve the model's understanding and representation of the complex semantic relationship between scientific and technological services and technical needs. The specific task design is as follows: 2.1 Mask task (Masked Language Modeling Task MLM): Mask some words in the input text, require the model to predict the masked content based on the context, and use the cross-entropy loss function for optimization to further enhance the model's language understanding ability.

[0027] Loss function for the masked language modeling task : ; N represents the number of samples; M i It is a sample i The set of masked words in ; w t is the masked position t The real word; p ( w t |context i ) is the model according to the context context i Predicted words w tThe probability of is obtained by calculating the softmax of the output layer of the pre-trained language model (such as BERT); The loss function of the masked language modeling task optimizes the model's ability to understand the contextual semantics by minimizing the probability difference between the predicted words and the actual words.

[0028] 2.2 Domain-specific Technical Term Prediction Task (TTP): In view of the professional terminology and contextual semantic characteristics in the field of scientific and technological services, an innovative task is proposed. Key technical terms are randomly masked in the input scientific and technological service text (extracted through domain dictionaries or TF-IDF), and the model is required to predict these terms. At the same time, a "term relationship prediction task" is designed to predict the hierarchical relationship or correlation between terms in the input text (such as the hierarchical relationship between "artificial intelligence" and "deep learning"). Through cross-entropy loss function optimization, the model's sensitivity to the specific semantics of the scientific and technological field is improved.

[0029] Loss Function for Term Prediction Task : ; N represents the number of samples; T i It is a sample i The set of key technical terms that are masked in ; w t is the masked position t The real word; p ( w t |context i ) is the model according to the context context i Predicted words w t The probability of is obtained by softmax calculation of the output layer of the pre-trained language model (such as BERT).

[0030] Loss Function for Relation Classification Task :

[0031] N represents the number of samples; is the set of relationship categories of term pairs in sample i (such as hyponymy and hyponymy, associativity); is the true label of the term pair on the relation category r (0 or 1, using one-hot encoding); It is the probability that the model predicts that the term pair belongs to the relation category r, which is calculated by passing the feature vector through the fully connected layer and softmax.

[0032] The total loss function : ; It is the weight coefficient for balancing the two subtasks and can be adjusted according to the importance of the task or the convergence speed.

[0033] This loss function jointly optimizes term prediction and relationship prediction, enhancing the model's deep understanding of specific semantics in scientific and technological fields. Term selection and relationship annotation can be combined with domain knowledge graphs or expert dictionaries to ensure the quality of training data and domain adaptability.

[0034] 2.3 Regression Task: Taking the demand text and service text as input, their relevance scores are predicted, and the mean square error (MSE) loss function is used to optimize the model so that the model can accurately quantify the degree of matching.

[0035] Regression loss function : ; N represents the number of samples; s i is the true relevance score of sample pair i (generated by manual annotation or rules); It is the relevance score predicted by the model, which is usually obtained by mapping the feature vector encoded by the pre-trained language model through a fully connected layer.

[0036] This loss function optimizes the model's ability to predict continuous values ​​of supply-demand matching correlation by minimizing the squared difference between the predicted score and the true score.

[0037] 2.4 Sorting task: Sort the relevance of different services under the same demand to ensure that services that are highly relevant to the demand are ranked at the top. Based on loss function optimization, the model's ranking ability in recommendation scenarios is improved.

[0038] Ranking loss function : ; N represents the number of samples; are the model prediction scores of samples i and j respectively; are the true correlation scores of samples i and j respectively; is the indicator function, when The value is 1 when , otherwise it is 0.

[0039] This loss function optimizes the ranking results, ensuring that services with higher relevance receive higher prediction scores.

[0040] The loss functions for these tasks are jointly optimized via a weighted summation approach, forming a multi-task learning framework. Compared to the general task design of traditional pre-trained models, the TTP tasks in this embodiment focus on technical terminology and relationships, significantly enhancing the model's ability to model the deep semantics of technical content, resulting in a preliminarily optimized model (the first optimized model).

[0041] Total loss function : ; Among them, α, β, γ, and δ are hyperparameters, which are tuned through experiments to balance the task weights. Loss function for the masked language modeling task, is the total loss function of the relation classification task, is the regression loss function, is the ranking loss function.

[0042] 3. Adaptive multi-task weight optimization mechanism: To address the problems of low training efficiency or inter-task interference caused by manual tuning of loss weights (α, β, γ, δ) in multi-task training, this embodiment proposes an adaptive weight adjustment mechanism based on task convergence speed and difficulty.

[0043] Dynamically adjust weights based on a function of training progress (epoch or step). Assuming the total number of training rounds is T (in epochs), the current number of training rounds is t, and a progress ratio is defined , which ranges from [0,1]. Then define a transition factor : ; in: k controls the steepness of the function (i.e., the transition speed), and is usually set to 10~20; It is the midpoint of the function, indicating the middle position of the transition, and is usually set to 0.5; It is a natural constant (approximately equal to 2.71828) and serves as the base in the transition factor formula to smooth transition scenarios. The core goal of the transition factor is to smoothly control the dynamic changes of task weights, solving the pain points of manual parameter adjustment.

[0044] The weights of the mask task and the term prediction task gradually decrease as the training progresses, using ; The weights of regression and sorting tasks gradually increase with the progress of training. .

[0045] During model training, the dynamic weights of the loss function for each task are: ; ; ; ; ; ; ; ; ; Among them, α, β, γ, and δ are hyperparameters, ω is the sum of hyperparameters, and is the total weight.

[0046] 4. Knowledge-enhanced contrastive learning improves model representation capabilities: Based on the first optimization model, this embodiment further proposes a "Knowledge-Enhanced Contrastive Learning" (KECL) method. By deeply integrating the domain knowledge graph with the contrastive learning framework, it significantly improves the model's ability to discriminate semantic relationships and the fine-grained discrimination of matching. The specific approach is as follows: 4.1 Knowledge Graph Assisted Sample Pair Construction KECL innovatively introduces a knowledge graph from the technology services sector as an auxiliary tool to guide the construction of sample pairs. The knowledge graph contains structured information such as technical field classifications, the hierarchical and interrelated relationships between terms, and expert resources. Based on this information, KECL closely integrates the sample pair construction process with domain knowledge: Positive pairs: Positive samples are defined as demand-service pairs that belong to the same technical subfield or have strong semantic connections. For example, "AI algorithm development requirements" and "machine learning model optimization services" in the same technical subfield are considered positive samples because they share similar technical classifications or terminology paths in the knowledge graph.

[0047] Negative sample pairs: Negative samples are defined as demand-service pairs that span technical fields or are semantically unrelated. For example, "biomedical technology demand" and "cloud computing services" are considered negative samples because they belong to different technical branches in the knowledge graph.

[0048] Guided by the knowledge graph, the construction of sample pairs is no longer limited to the surface features of the text, but fully considers the semantic hierarchy and entity relationships within the domain, enabling the model to learn more domain-specific representations.

[0049] 4.2 Knowledge Weighted Contrast Loss Function Optimization: Knowledge-Enhanced Contrastive Learning (KECL) uses the path distance between entities in the knowledge graph as a metric for sample difficulty. Path distance reflects the degree of semantic connection between two entities in the knowledge graph: entity pairs with shorter path distances are generally more semantically connected and easier to distinguish; whereas entity pairs with longer path distances are less semantically connected and are considered "difficult to distinguish." Based on this, KECL designs a weighted InfoNCE loss function that assigns higher weight to difficult examples (i.e., positive pairs with longer path distances or negative pairs with shorter path distances), prioritizing the optimization of the representation of these examples.

[0050] The direct effect of this weighting mechanism is that the model can better narrow the distance between positive sample pairs in the semantic space, while significantly pushing the distance between negative sample pairs further away, thereby improving the ability to distinguish fine-grained semantic relationships.

[0051] ; ; in, Representation sample i and samples j The weight between Representation sample i and samples j The distance between represents the sum of the distances between all pairs of samples, Representation sample i The feature representation of Represents a positive sample The feature representation of represents the sensitivity of the loss function, Indicates the similarity of samples, represents a positive sample pair, represents the knowledge embedding contrastive loss function.

[0052] After optimization, a further optimized model is obtained, namely the second optimized model.

[0053] 5. Model fine-tuning based on real user historical data: To make the model more relevant to real-world application scenarios, this embodiment fine-tunes the second optimized model using historical user interaction data. By collecting user behavior data on the platform, such as clicks, inquiries, and transactions, and combining it with user feedback, supervised fine-tuning is used to further optimize model parameters. This fine-tuning not only improves the model's adaptability to real-world user needs but also effectively reduces the model's cold start and generalization risks in real-world applications, resulting in the final re-ranked model (re-ranked model).

[0054] 6. Two-stage search and re-ranking mechanism for matching supply and demand of scientific and technological services: In actual applications, this embodiment adopts a two-stage search and re-ranking strategy to balance matching efficiency and accuracy: Phase 1: Efficient retrieval using semantic vectors: The second optimization model is used as the semantic vector retrieval model to vectorize the technical requirement text entered by the user and the service text in the technology service library. Through efficient vector similarity calculations (such as cosine similarity or ANN nearest neighbor search), candidate technology services are quickly screened out to achieve coarse sorting in a large-scale service library.

[0055] Phase 2: Refined Re-ranking and Multi-dimensional Sorting: Using the re-ranking model as a refined re-ranking model, candidate services are scored for multi-dimensional relevance (including semantic relevance, user preference matching, domain adaptability, etc.), and a weighted ranking algorithm is used to output the final recommendation results. This mechanism effectively improves the system's search efficiency and recommendation accuracy, meeting the actual needs of large-scale technology service supply and demand scenarios.

[0056] In summary, this embodiment significantly enhances the intelligent matching of supply and demand on technology service platforms by constructing high-quality domain-specific datasets, innovatively designing technical term prediction tasks (TTPs) and multi-task transition training, introducing an adaptive multi-task weight optimization mechanism, proposing a knowledge-enhanced contrastive learning (KECL) method, and combining it with a two-stage retrieval and re-ranking strategy. This technical solution not only achieves breakthroughs in matching accuracy and efficiency, but also offers significant advantages in domain adaptability, technological innovation, and training efficiency. It can be widely applied to intelligent recommendation and resource allocation scenarios across various technology service platforms, possessing significant practical application value and promising prospects for expansion.

[0057] As a possible implementation method, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, a method for intelligent matching of supply and demand of technology services based on deep learning is implemented.

[0058] As a possible implementation method, this embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, a method for intelligent matching of supply and demand of scientific and technological services based on deep learning is implemented.

[0059] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent matching method for technology service supply and demand, characterized by: The specific steps include: Obtain the historical user demand and technology service interaction data accumulated by the technology service platform to build a technology service supply and demand matching data set; Constructing a pre-trained language model, and performing post-pretraining on the pre-trained language model based on multi-task learning to obtain a first optimized model; Obtain knowledge graphs in the field of scientific and technological services and construct knowledge graph-assisted sample pairs; Based on the knowledge graph-assisted sample pairs, the first optimization model is compared and learned to obtain the second optimization model; Collect historical interaction data of real users to optimize the second optimization model and obtain a rearrangement model; Based on the re-ranking model, the candidate services are scored for multi-dimensional relevance, and the final recommendation results are obtained based on weighted sorting.

2. The deep learning-based intelligent matching method for technology service supply and demand according to claim 1 is characterized in that: The acquisition of historical user demand and technology service interaction data accumulated by the technology service platform and the construction of a technology service supply and demand matching dataset specifically includes: Obtain user demand and technology service interaction data based on the historical accumulation of the technology service platform, extract technology demand text, technology service text and correlation score as the technology service supply and demand matching data set; The technical requirement text is a description of the technical service requirements submitted by the user; The technology service text is a description of the technology services provided by the platform; The relevance score is generated by combining manual annotation with predictions from a scoring model based on historical user interaction behaviors.

3. The deep learning-based intelligent matching method for technology service supply and demand according to claim 1 is characterized in that: In the multi-task learning, the multiple tasks specifically include: a mask task, a term prediction task, a regression task, and a sorting task.

4. The method for intelligent matching of supply and demand of scientific and technological services based on deep learning according to claim 3 is characterized in that: The post-pre-training of the pre-trained language model based on multi-task learning specifically includes: Based on the masking task, the input text of the pre-trained language model is randomly masked; the pre-trained language model performs semantic understanding of the text based on the remaining visible text; based on the semantic understanding of the text, the masked words are obtained; the correlation between the words predicted by the model and the actual masked words is calculated to obtain the prediction accuracy; based on the prediction accuracy, the model parameters are optimized using the cross-entropy loss function; Based on the term prediction task, the domain dictionary is used to extract key technical terms from the input text; the extracted key technical terms are randomly selected and replaced with tokens in the text; the token-replaced key technical terms are predicted based on the context information; the cross-entropy loss function is used to calculate the difference between the predicted key technical terms and the actual key technical terms, and the model parameters are optimized based on the difference; Based on the regression task, the demand text and service text pairs are extracted from the input text; a relevance score is assigned to each text pair; the demand text and service text pairs are encoded using a pre-trained language model, and the semantic feature vectors of the encoded demand text and service text pairs are extracted; the relevance score is predicted based on the extracted feature vectors; a loss function is constructed based on the relevance score prediction to optimize the model parameters; Sorting task: Based on the relevance score of the demand text and service text pairs, a loss function is constructed based on the relevance score to optimize the model's sorting ability.

5. The method for intelligent matching of supply and demand of scientific and technological services based on deep learning according to claim 4 is characterized in that: When training the pre-trained model, the task weights during multi-task learning are also dynamically adjusted. The dynamic adjustment process includes: Set the training round and get the current number of training rounds; Get the progress ratio based on the training rounds and the current number of training rounds; According to the progress ratio, the transition factor is obtained; The weights of the mask task, term prediction task, regression task, and ranking task are set according to the transition factor.

6. The method for intelligent matching of supply and demand of scientific and technological services based on deep learning according to claim 1 is characterized in that: The acquisition of the knowledge graph in the field of scientific and technological services and the construction of knowledge graph auxiliary sample pairs specifically include: Generate a knowledge graph in the field of science and technology services based on the science and technology service supply and demand matching dataset; In the knowledge graph of the field of science and technology services, demand-service pairs belonging to the same technical sub-field or with strong semantic associations are defined as positive samples, and demand-service pairs that cross technical fields or are semantically unrelated are defined as negative samples.

7. The method for intelligent matching of supply and demand of scientific and technological services based on deep learning according to claim 1 is characterized in that: The method of performing comparative learning on the first optimization model based on the knowledge graph-assisted sample pairs to obtain the second optimization model specifically includes: Obtain nodes of the knowledge graph in the field of science and technology services and extract node path distances; Nodes whose node path distance is higher than the set threshold are extracted and defined as hard-to-classify samples; Construct a contrastive learning loss function, weight the difficult samples, optimize the sample representation, and obtain the second optimization model.

8. The method for intelligent matching of supply and demand of scientific and technological services based on deep learning according to claim 1 is characterized in that: The collecting of historical interaction data of real users to optimize the second optimization model to obtain a rearrangement model specifically includes: Get historical interaction data of real users, The second optimization model is trained based on historical interaction data of real users to obtain a rearrangement model.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, it implements the deep learning-based intelligent matching method for supply and demand of scientific and technological services as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, it implements the deep learning-based intelligent matching method for supply and demand of scientific and technological services as described in any one of claims 1 to 8.

Images