Deep learning-based intelligent matching method, device and medium for supply and demand of technology services

Abstract

The application discloses a science and technology service supply and demand intelligent matching method and device based on deep learning and a medium, relates to the technical field of deep learning, and comprises the following steps: constructing a multidimensional data set covering the field of science and technology services, post-training a pre-trained language model based on multi-task learning, and constructing a knowledge graph auxiliary sample based on a knowledge graph; comparing and learning a first optimization model based on the knowledge graph auxiliary sample to obtain a second optimization model; collecting historical interaction data of real users to optimize the second optimization model to obtain a rearrangement model, then performing multidimensional correlation scoring on candidate services, and obtaining a final recommendation result based on weighted sorting. Through the construction of a field-specific data set, the design of an innovative multi-task transition training and a field-specific task, the introduction of an adaptive multi-task weight optimization mechanism, the proposal of a knowledge-enhanced contrast learning method, and the combination of a two-stage retrieval and rearrangement strategy, efficient, intelligent and accurate matching of science and technology service supply and demand data is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a technology service supply and demand intelligent matching method, device and medium based on deep learning. BACKGROUND

[0002] With the rapid development of technology service industry, the information asymmetry problem between technology demand side and service provider is increasingly prominent. Traditional technology service matching mainly relies on manual retrieval or simple keyword matching method, which has the defects of low matching precision, insufficient efficiency and inability to fully excavate the deep semantic association between demand and service. In recent years, the progress of deep learning technology provides new possibilities for intelligent matching, especially the wide application of pre-trained language models in natural language processing field, which significantly improves the performance of semantic understanding and text matching. However, the existing matching methods based on pre-trained language models are designed for general field, lacking specific adaptation to technology service field specific semantics, industry terms and complex demand scenarios, resulting in limited model performance in processing professional and complex context technology service supply and demand matching tasks.

[0003] In addition, in the multi-task joint training process of the prior art, a static weight distribution strategy is often adopted, which is difficult to balance the convergence speed and optimization target between different tasks, resulting in low training efficiency or poor model performance. At the same time, the existing matching system generally has the problem that the retrieval efficiency and matching precision are difficult to balance when facing large-scale data, especially in the application scenarios with high real-time requirements, the system response speed becomes the key bottleneck restricting user experience. On the other hand, the knowledge update speed in technology service field is fast, and the existing model lacks dynamic adaptation and knowledge enhancement mechanism, which is difficult to meet the diversified demands across industries and scenes.

[0004] In summary, the prior art has significant deficiencies in the field adaptation, training efficiency, matching precision and system scalability of technology service supply and demand matching, and an efficient and intelligent matching method designed for the characteristics of technology service field is needed to solve the above problems and improve the platform service capability. SUMMARY

[0005] The technical problem to be solved by the present application is that the prior art has significant deficiencies in the field adaptation, training efficiency, matching precision and system scalability of technology service supply and demand matching, and the purpose is to provide a technology service supply and demand intelligent matching method, device and medium based on deep learning, which realizes efficient and accurate supply and demand matching through field specific dataset construction, multi-task joint training, self-adaptive weight optimization, knowledge enhancement contrast learning and two-stage retrieval rearrangement strategy, can realize accurate and efficient docking of technology demand and service, and significantly improve the matching precision, system efficiency and user experience.

[0006] The application is realized by the following technical solutions:

[0007] The first aspect of the application provides a technology service supply and demand intelligent matching method based on deep learning, comprising the following specific steps:

[0008] Obtain user demand and technology service interaction data accumulated by a technology service platform, and construct a technology service supply and demand matching data set;

[0009] Construct a pre-training language model, and post-train the pre-training language model based on multi-task learning to obtain a first optimization model;

[0010] Use the path distance between entities in the knowledge graph as a measure of sample difficulty, and perform knowledge graph assisted sample pair classification;

[0011] Based on the knowledge graph assisted sample pair, the first optimization model is subjected to comparative learning to obtain a second optimization model;

[0012] Collect real user historical interaction data to optimize the second optimization model to obtain a rearrangement model;

[0013] Based on the rearrangement model, a multi-dimensional correlation score is given to the candidate service, and a final recommendation result is obtained based on weighted sorting.

[0014] Further, the obtaining of the user demand and technology service interaction data accumulated by the technology service platform and the construction of the technology service supply and demand matching data set specifically comprises:

[0015] Obtain user demand and technology service interaction data accumulated by the technology service platform, extract technical demand text, technology service text and correlation score as the technology service supply and demand matching data set; wherein,

[0016] The technical demand text is a technical service demand description submitted by the user;

[0017] The technology service text is a technology service content description provided by the platform;

[0018] The correlation score is generated by combining artificial annotation and a scoring model based on user historical interaction behavior.

[0019] Further, in the multi-task learning, the multi-task specifically comprises a mask task, a term prediction task, a regression task and a sorting task.

[0020] Further, the post-training of the pre-training language model based on multi-task learning specifically comprises:

[0021] Based on the mask task, the input text of the pre-trained language model is randomly masked; the pre-trained language model performs semantic understanding of the text according to the remaining visible text; the masked words are obtained according to the semantic understanding of the text; the predicted words of the model are calculated with the actual masked words to obtain the prediction accuracy; based on the prediction accuracy, the model parameters are optimized using the cross-entropy loss function;

[0022] Based on the term prediction task, the key technical terms are extracted from the input text using the field dictionary; the extracted key technical terms are randomly selected and replaced with a mark in the text; the key technical terms replaced with the mark are predicted according to the context information; the difference between the predicted key technical terms and the actual key technical terms is calculated using the cross-entropy loss function, and the model parameters are optimized according to the difference;

[0023] Based on the regression task, the demand text and service text pairs are extracted according to the input text; each pair of texts is labeled with a correlation score; the pre-trained language model is used to encode the demand text and service text pairs, and the semantic feature vectors of the encoded demand text and service text pairs are extracted; the correlation score is predicted based on the extracted feature vectors; the loss function is constructed according to the correlation score prediction, and the model parameters are optimized;

[0024] The sorting task: according to the correlation score of the demand text and service text pair, the correlation score is used to construct the loss function to optimize the model sorting ability.

[0025] Further, when training the pre-trained model, it also includes dynamically adjusting the task weight of multi-task learning, the process of dynamic adjustment includes:

[0026] Set the training round, get the current training round;

[0027] According to the training round and the current training round, the progress ratio is obtained;

[0028] According to the progress ratio, the transition factor is obtained;

[0029] According to the transition factor, the weights of the mask task, the term prediction task, the regression task and the sorting task are set.

[0030] Further, the knowledge graph of the science and technology service field is obtained, and the knowledge graph auxiliary sample pair is constructed, which specifically includes:

[0031] According to the science and technology service supply and demand matching data set, the knowledge graph of the science and technology service field is generated;

[0032] In the knowledge graph of the technology service field, the demand-service pairs belonging to the same technical subfield or having strong semantic correlation are defined as positive samples, and the demand-service pairs across technical fields or irrelevant in semantics are defined as negative samples.

[0033] Further, the first optimization model is subjected to contrastive learning based on the knowledge graph auxiliary sample pair to obtain a second optimization model, specifically including:

[0034] The nodes of the knowledge graph of the technology service field are acquired, and the node path distance is extracted.

[0035] The nodes with a node path distance higher than a set threshold are extracted and defined as difficult-to-classify samples.

[0036] A contrastive learning loss function is constructed, the difficult-to-classify samples are weighted, the sample representation is optimized, and the second optimization model is obtained.

[0037] Further, the second optimization model is optimized based on the historical interaction data of real users to obtain a rearrangement model, specifically including:

[0038] The historical interaction data of real users are acquired,

[0039] The second optimization model is trained based on the historical interaction data of real users to obtain the rearrangement model.

[0040] The second aspect of the application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to realize the deep learning-based intelligent matching method for supply and demand of technology services.

[0041] The third aspect of the application provides a computer-readable storage medium having a computer program stored thereon, wherein the program is executed by a processor to realize the deep learning-based intelligent matching method for supply and demand of technology services.

[0042] Compared with the prior art, the application has the following advantages and beneficial effects:

[0043] 1. Improved field adaptability: In view of the professionalism and complexity of the technology service field, a field-specific dataset construction method and pre-training task are designed, and the understanding ability of the model for industry terms, technical background and demand context is improved through semantic enhancement and knowledge graph integration, thereby solving the limitations of general pre-training models in professional field applications.

[0044] 2. Optimized training efficiency and performance: An adaptive multi-task weight optimization mechanism is introduced to dynamically adjust the weight distribution of regression, sorting and other multi-tasks, reduce the interference between tasks, improve the training efficiency and overall performance of the model, and ensure the stability and reliability of the system in complex matching tasks.

[0045] 3. Enhancing semantic matching accuracy: Propose a knowledge-enhanced contrastive learning method, use domain knowledge graph to construct high-quality sample pairs, optimize the model's ability to distinguish fine-grained semantic relationships between technical needs and services, and achieve more accurate supply and demand matching.

[0046] 4. Balance efficiency and accuracy: Design a two-stage retrieval and rearrangement strategy, filter candidate services through efficient semantic retrieval, and combine a refined rearrangement model for multi-dimensional relevance scoring to ensure matching accuracy while significantly improving system response speed, meeting real-time needs in large-scale data scenarios.

[0047] 5. Realize cross-scene versatility: Build a modular and extensible technical framework that supports flexible adjustment of task design and optimization strategies according to different industries and application scenarios, and through dynamic updating of user interaction data and knowledge graph, ensure the adaptability and continuous optimization ability of the model in cross-domain and cross-scene scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the example embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments, it should be understood that the following drawings only show some embodiments of the present application, therefore should not be regarded as a limitation on the scope, for those skilled in the art, without creative labor, can also obtain other related drawings according to these drawings. In the drawings:

[0049] Figure 1 The flowchart of the matching method in the embodiment of the present application. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical scheme and advantages of the present application more clear and obvious, the following will combine with examples and drawings to make further detailed description of the present application, the illustrative embodiments of the present application and its description are only used to explain the present application, and not as a limitation on the present application.

[0051] As a possible implementation, as Figure 1As shown, the embodiment provides a deep learning-based intelligent matching method for technology service supply and demand, which includes the following specific steps: obtaining user demand and technology service interaction data accumulated by a technology service platform over time, and constructing a technology service supply and demand matching dataset; constructing a pre-trained language model, and performing post-training on the pre-trained language model based on multi-task learning to obtain a first optimized model; using the path distance between entities in the knowledge graph as a measure of sample difficulty to classify knowledge graph-assisted sample pairs; performing contrastive learning on the first optimized model based on the knowledge graph-assisted sample pairs to obtain a second optimized model; collecting historical interaction data of real users to optimize the second optimized model to obtain a rearrangement model; and based on the rearrangement model, performing multi-dimensional correlation scoring on candidate services to obtain a final recommendation result based on weighted ranking.

[0052] In the embodiment, by constructing a multi-dimensional dataset covering the technology service field, and enhancing technical terms, industry background and semantic relationships with the help of the field knowledge graph, a field-specific technical term prediction task (TTP) is designed, which significantly improves the understanding ability of the pre-trained language model for complex semantics in the technology field. An adaptive multi-task weight optimization mechanism is introduced, and through dynamic weighted averaging and reinforcement learning assisted optimization strategy, the weight distribution of regression, sorting and other multi-tasks is adjusted in real time according to the task convergence speed and difficulty, effectively reducing the interference between tasks, improving the training efficiency and model performance. High-quality positive and negative sample pairs are constructed using the technology service field knowledge graph, and difficult samples are optimized preferentially through the weighted InfoNCE loss function, which enhances the model's ability to distinguish fine-grained semantic relationships between technical requirements and services. Through a two-stage retrieval and rearrangement strategy, the first stage efficiently retrieves and filters candidate services through semantic vectors, and the second stage uses a refined rearrangement model to perform multi-dimensional correlation scoring, and combines a weighted ranking algorithm to output the final matching result. This strategy significantly improves system efficiency while ensuring matching accuracy. Through dynamic updating of user historical interaction data and field knowledge graph, the universality of the model is ensured. The embodiment significantly reduces user search costs and improves user satisfaction by accurately matching technical requirements and technology services; at the same time, through efficient retrieval and intelligent recommendation, the efficiency of resource allocation of the technology service platform is optimized, and the service capability and market competitiveness of the platform are enhanced. Compared with the prior art, the embodiment has significant advantages in field adaptability, semantic understanding ability, training efficiency, matching accuracy, system performance and application scalability. The innovative technical solution not only fills the technical gap in the field of intelligent matching of technology service supply and demand, but also demonstrates outstanding results in practical applications, and has important theoretical value and business prospects.

[0053] The specific implementation steps include:

[0054] 1. Construction of technology service supply and demand matching dataset:

[0055] This embodiment is based on the user demand and technology service interaction data accumulated by the technology service platform, and constructs high-quality and field-specific training data sets. The data set takes (technical demand text, technology service text, and relevance score) as the basic unit, where the technical demand text is the technical service demand description submitted by the user, the technology service text is the technology service content description provided by the platform, and the relevance score is generated by combining artificial annotation and score model based on user historical interaction behavior (such as clicking, consulting, and transaction). The matching degree is quantified in the form of 0-1 range value. To improve data quality, this embodiment introduces a field knowledge graph to enhance the semantics of the data, annotates the key entities (such as technical field, application scenario, and core technology point) and their relationships in the technical demand and service text, and ensures the accuracy and representativeness of the data at the semantic level. At the same time, through data cleaning, denoising and diversity sampling, training data covering multiple industries and scenarios are constructed, laying a solid foundation for subsequent model training.

[0056] 2. Training phase of pre-training language model based on multi-task learning: multi-task joint training and field-specific task design:

[0057] Based on the pre-training language model (such as BERT), post-training is performed to reduce the difference between pre-training and field fine-tuning. Through multi-task joint optimization and field-specific task design in the post-training phase, the understanding and representation ability of the model for the complex semantic relationship between technology services and technical demands is improved. The specific task design is as follows:

[0058] 2.1 Mask task (Masked Language Modeling task, MLM):

[0059] Mask some words in the input text, and require the model to predict the masked content based on the context. Cross-entropy loss function is used for optimization, which further enhances the language understanding ability of the model.

[0060] Masked language modeling task loss function : ;

[0061] N represents the number of samples;

[0062] M i is the set of masked words in the sample i ;

[0063] w t is the true word at the masked position t ;

[0064] p ( wt |context i ) is the probability of the model predicting the word given the context context i w t , computed by the output layer of a pre-trained language model (e.g., BERT) via softmax.

[0065] The Masked Language Modeling task loss function optimizes the model's understanding of contextual semantics by minimizing the difference in probability between the predicted word and the true word.

[0066] 2.2 Technical Term Prediction (TTP) Task:

[0067] An innovative task is proposed to address the characteristics of professional terms and contextual semantics in the field of technology services. In the input text of technology services, key technical terms are randomly masked (extracted through a domain dictionary 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 relevance between terms in the input text (such as the hierarchical relationship between "artificial intelligence" and "deep learning"), optimized through a cross-entropy loss function to improve the model's sensitivity to specific semantics in the technology field.

[0068] Loss function for Term Prediction Task :

[0069] ;

[0070] N represents the number of samples;

[0071] T i is the set of key technical terms masked in the sample i ;

[0072] w t is the true word at the masked position t ;

[0073] p ( w t |context i ) is the probability of the model predicting the word given the context context i w t , computed by the output layer of a pre-trained language model (e.g., BERT) via softmax.

[0074] ​​Loss function of Relation Classification Task :

[0075]

[0076] N represents the number of samples;

[0077] is the set of relation classes of the term pair in sample i (such as the hyponymy relation, the association relation);

[0078] is the true label of the term pair in relation class r (0 or 1, using one-hot encoding);

[0079] is the probability of the model predicting that the term pair belongs to the relation class r, which is calculated by the feature vector through the fully connected layer and the softmax.

[0080] Total loss function : ;

[0081] is the weight coefficient for balancing the two sub-tasks, which can be adjusted according to the importance of the task or the convergence speed.

[0082] This loss function optimizes the term prediction and relation prediction jointly, enhancing the model's deep understanding of specific semantics in the field of science and technology. Term selection and relation labeling can be combined with domain knowledge graph or expert dictionary to ensure the quality and domain adaptability of the training data.

[0083] 2.3 Regression task:

[0084] Take the demand text and service text as input, predict the relevance score, and use the mean square error (MSE) loss function to optimize the model, so that the model can accurately quantify the matching degree.

[0085] Regression loss function : ;

[0086] N represents the number of samples;

[0087] s i is the true relevance score of sample pair i (annotated by humans or generated by rules);

[0088] is the relevance score predicted by the model, usually obtained by mapping the feature vector encoded by the pre-trained language model through the fully connected layer.

[0089] The loss function optimizes the model's ability to predict continuous values related to supply-demand matching by minimizing the squared difference between predicted scores and true scores.

[0090] 2.4 Ranking task:

[0091] The correlation of different services under the same demand is ranked to ensure that services highly related to the demand are ranked in the front position. Based on loss function optimization, the ranking ability of the model in the recommendation scenario is improved.

[0092] Ranking loss function :

[0093] ;

[0094] N represents the number of samples;

[0095] are the model prediction scores of samples i and j, respectively;

[0096] are the true correlation scores of samples i and j, respectively;

[0097] is an indicator function, which is 1 when , and 0 otherwise.

[0098] The loss function optimizes the ranking result to ensure that services with higher correlation obtain higher predicted scores.

[0099] The loss functions of the above tasks are jointly optimized by weighted summation to form a multi-task learning framework. Compared with the general task design of traditional pre-training models, the TTP task of the present embodiment focuses on technical field terms and relationships, significantly enhancing the model's ability to model deep semantic content of technology, and obtaining a preliminary optimized model (first optimized model).

[0100] Total loss function :

[0101] ;

[0102] where α, β, γ, δ are hyperparameters, which are adjusted through experiments to balance task weights, is the mask language modeling task loss function, is the total loss function of the relationship classification task, is the regression loss function, is the ranking loss function.

[0103] 3. Adaptive multi-task weight optimization mechanism:

[0104] To solve the problem of low training efficiency or task interference caused by manual adjustment of loss weight (alpha, beta, gamma, delta) in multi-task training, the embodiment proposes an adaptive weight adjustment mechanism based on task convergence speed and difficulty.

[0105] The weight is dynamically adjusted based on a function of the training progress (epoch or step). Assuming the total training rounds is T (in epoch units), the current training round is t, define a progress ratio , which ranges from [0, 1]. Then define a transition factor : ;

[0106] Where:

[0107] k controls the steepness of the function (i.e. transition speed), usually takes a value of 10~20;

[0108] is the midpoint of the function, representing the middle position of the transition, usually takes a value of 0.5;

[0109] is a natural constant (approximately equal to 2.71828), used as the base number in the transition factor formula to smooth the transition scenario; the core goal of the transition factor is to smoothly control the dynamic change of task weight, solving the pain point of manual parameter tuning;

[0110] The weights of the mask task and the term prediction task gradually decrease with the training progress, using ;

[0111] The weights of the regression task and the ranking task gradually increase with the training progress, using .

[0112] In model training, the dynamic weight of each task loss function is:

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] ;

[0118] ;

[0119] ;

[0120] ;

[0121] ;

[0122] wherein α, β, γ, δ are hyperparameters, ω is the sum of hyperparameters, and is the total weight.

[0123] 4. Knowledge-enhanced contrastive learning improves model representation ability:

[0124] Based on the first optimization model, this embodiment further proposes a "knowledge-enhanced contrastive learning" (KECL) method, which significantly improves the model's ability to distinguish semantic relationships and the fine-grained distinction of matching by deeply integrating the domain knowledge graph and the contrastive learning framework. The specific approach is as follows:

[0125] 4.1 Knowledge graph assisted sample pair construction

[0126] KECL innovatively introduces a knowledge graph in the field of scientific services as an auxiliary tool to guide the construction of sample pairs. The knowledge graph contains structured information such as technical field classification, hierarchical and associative relationships between terms, and expert resources. Based on this information, KECL closely integrates the construction process of sample pairs with domain knowledge:

[0127] Positive sample pairs: Define demand-service pairs that belong to the same technical subfield or have strong semantic associations as positive samples. For example, "artificial intelligence algorithm development demand" and "machine learning model optimization service" in the same technical subfield are considered positive samples because they share similar technical classifications or term paths in the knowledge graph.

[0128] Negative sample pairs: Define demand-service pairs that cross technical fields or have no semantic relevance as negative samples. For example, "biomedical technology demand" and "cloud computing service" are considered negative samples because they belong to different technical branches in the knowledge graph.

[0129] With the guidance of the knowledge graph, the construction of sample pairs is no longer limited to surface-level text features, but fully considers the semantic hierarchy and entity relationships within the domain, enabling the model to learn more domain-specific representations.

[0130] 4.2 Knowledge-weighted contrastive loss function optimization:

[0131] Knowledge-Enhanced Contrastive Learning (KECL) utilizes the path distance between entities in the knowledge graph as a measure of sample difficulty. The path distance reflects the degree of semantic association between two entities in the knowledge graph: entity pairs with shorter path distances usually have stronger semantic associations and are easier to distinguish; entity pairs with longer path distances have weaker semantic associations and are "hard-to-distinguish samples". Based on this, KECL designs a weighted InfoNCE loss function, which assigns higher weights to hard-to-distinguish samples (i.e., positive sample pairs with longer path distances or negative sample pairs with shorter path distances), and prioritizes optimizing the representations of these samples.

[0132] The direct effect of this weighting mechanism is that the model can better pull the distance between positive sample pairs in the semantic space, while significantly pushing away the distance between negative sample pairs, thereby improving the ability to distinguish fine-grained semantic relationships.

[0133] ;

[0134] ;

[0135] wherein, represents the weight between samples i and samples j , represents the distance between samples i and samples j , represents the sum of distances between all sample pairs, represents the feature representation of sample i , represents the feature representation of positive sample , represents the sensitivity of the loss function, represents the similarity of the sample, represents the positive sample pair, represents the knowledge embedding contrastive loss function.

[0136] After optimization, a further optimized model, i.e., a second optimized model, is obtained.

[0137] 5. Model fine-tuning based on real user historical data:

[0138] To make the model more suitable for practical application scenarios, the embodiment fine-tunes the second optimization model using real user historical interaction data. By collecting user behavior data such as real clicks, consultations, transactions, etc. on the platform, combined with user feedback information, the model parameters are further optimized using supervised fine-tuning. This stage of fine-tuning not only improves the adaptability of the model to real user needs, but also effectively reduces the cold start and generalization risk of the model in practical application, obtaining the final rearrangement model (rearrangement model).

[0139] 6. Two-stage retrieval and rearrangement mechanism for technology service supply and demand matching:

[0140] In actual application, the embodiment adopts a two-stage retrieval and rearrangement strategy, taking into account matching efficiency and accuracy:

[0141] First stage: Efficient semantic vector retrieval: Use the second optimization model as a semantic vector retrieval model to vectorize the user's input technical requirement text and the service text in the technology service library. Through efficient vector similarity calculation (such as cosine similarity or ANN neighbor search), candidate technology services are quickly filtered out to achieve coarse sorting in a large-scale service library.

[0142] Second stage: Fine rearrangement and multi-dimensional sorting: Use the rearrangement model as a fine rearrangement model to score the candidate services in multiple dimensions (including semantic relevance, user preference matching degree, field adaptability, etc.), and output the final recommendation results using a weighted sorting algorithm. This mechanism effectively improves the retrieval efficiency and recommendation accuracy of the system, meeting the actual needs of large-scale technology service supply and demand scenarios.

[0143] In summary, the embodiment significantly improves the intelligence level of technology service platform supply and demand matching by constructing high-quality field-specific datasets, innovatively designing technical term prediction tasks (TTP) and multi-task transition training, introducing an adaptive multi-task weight optimization mechanism, proposing a knowledge-enhanced contrast learning (KECL) method, and combining a two-stage retrieval and rearrangement strategy. This technical solution not only achieves breakthroughs in matching accuracy and efficiency, but also has significant advantages in field adaptability, technical innovation, and training efficiency. It can be widely applied to intelligent recommendation and resource allocation scenarios in various technology service platforms, and has important practical application value and promotion prospects.

[0144] As a possible implementation, the embodiment provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements a deep learning-based technology service supply and demand intelligent matching method when executing the program.

[0145] As a possible implementation, the embodiment provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the deep learning-based intelligent matching method for supply and demand of technology services.

[0146] The above detailed description further illustrates the purpose, technical solutions and beneficial effects of the present application. It should be understood that the above description is only a specific implementation of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A deep learning-based intelligent matching method for supply and demand of technology services, characterized in that, The method comprises the following specific steps: Obtain the user demand and technology service interaction data accumulated by the technology service platform, and construct a technology service supply and demand matching data set; Construct a pre-training language model, and perform post-training on the pre-training language model based on multi-task learning to obtain a first optimization model; Specifically comprising: Based on the mask task, randomly mask the input text of the pre-training language model; the pre-training language model performs semantic understanding of the text according to the remaining visible text; the masked words are obtained according to the semantic understanding of the text; the predicted words and the actual masked words are calculated for relevance 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, extract key technical terms from the input text using a domain dictionary; randomly select the extracted key technical terms and replace the key technical terms with a marker in the text; predict the key technical terms replaced with the marker according to the context information; calculate the difference between the predicted key technical terms and the actual key technical terms using the cross-entropy loss function, and optimize the model parameters according to the difference; Based on the regression task, extract the demand text and service text pair according to the input text; label each pair of texts with a correlation score; encode the demand text and service text pair using the pre-training language model, extract the semantic feature vector of the encoded demand text and service text pair; predict the correlation score based on the extracted feature vector; construct a loss function based on the correlation score prediction to optimize the model parameters; Sorting task: based on the correlation score of the demand text and service text pair, construct a loss function based on the correlation score to optimize the model sorting ability; Obtain the knowledge graph of the technology service field, and construct a knowledge graph auxiliary sample pair; Specifically comprising: Generate a knowledge graph of the technology service field according to the technology service supply and demand matching data set; In the knowledge graph of the technology service field, define the demand-service pair belonging to the same technical subfield or having strong semantic correlation as a positive sample, and define the demand-service pair across technical fields or having no semantic correlation as a negative sample; Based on the knowledge graph auxiliary sample pair, perform contrastive learning on the first optimization model to obtain a second optimization model; Collect the historical interaction data of real users to optimize the second optimization model to obtain a rearrangement model; Based on the rearrangement model, perform multi-dimensional correlation scoring on the candidate services, and obtain the final recommendation result based on weighted sorting. 2.The deep learning-based intelligent matching method for supply and demand of technology services according to claim 1, characterized in that, The technology service supply and demand matching data set is constructed by obtaining the user demand and technology service interaction data accumulated by the technology service platform, and specifically comprising: Obtain the user demand and technology service interaction data accumulated by the technology service platform, extract the technical demand text, technology service text and correlation score as the technology service supply and demand matching data set; wherein, The technical demand text is the technical service demand description submitted by the user; The technology service text is the description of the technology service content provided by the platform; The correlation score is generated by combining the artificial annotation and the scoring model based on the user historical interaction behavior. 3.The deep learning-based intelligent matching method for supply and demand of technology services according to claim 1, characterized in that, The multi-task learning includes a mask task, a term prediction task, a regression task, and a ranking task. 4.The deep learning based intelligent matching method for supply and demand of technology services according to claim 1, characterized in that, The training of the pre-training model further includes dynamically adjusting the task weights in the multi-task learning, and the adjusting process includes: setting a training round, and obtaining a current training round number; obtaining a progress ratio according to the training round and the current training round number; obtaining a transition factor according to the progress ratio; setting the weights of the mask task, the term prediction task, the regression task, and the ranking task according to the transition factor. 5.The deep learning based intelligent matching method for supply and demand of technology services according to claim 1, characterized in that, The first optimization model is subjected to contrast learning based on the knowledge graph assisted sample pair to obtain a second optimization model, and the process specifically includes: obtaining nodes of a knowledge graph in a scientific and technological service field, and extracting node path distances; extracting nodes with a node path distance higher than a set threshold value, and defining the nodes as difficult-to-classify samples; constructing a contrast learning loss function, weighting the difficult-to-classify samples, optimizing sample representation, and obtaining the second optimization model. 6.The deep learning based intelligent matching method for supply and demand of technology services according to claim 1, characterized in that, The second optimization model is optimized based on the historical interaction data of real users to obtain a rearrangement model, and the process specifically includes: obtaining the historical interaction data of real users, training the second optimization model based on the historical interaction data of real users to obtain the rearrangement model.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the deep learning-based intelligent matching method for supply and demand of scientific and technological services according to any one of claims 1 to 6.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the deep learning-based intelligent matching method for supply and demand of scientific and technological services according to any one of claims 1 to 6.

Images