Text feature-based supply and demand matching method and device for supply service resources

Abstract

The invention provides a text feature-based supply and demand matching method and device for supply service resources. The method provided by the invention comprises the following steps: acquiring a plurality of themes from a demand description text of service resource supply, and matching description types of service resource supply; obtaining description data corresponding to each description type, and obtaining a plurality of supply service resource samples through combination; extracting a first feature of each supply service resource sample and a second feature of the demand description text based on a text representation model, and determining a first type of supply service resource samples based on the similarity of the first feature and the second feature; generating a candidate demand set according to the first type of supply service resource samples; extracting a third feature of each candidate enterprise demand description text based on the text representation model, and determining a second type of supply service resource sample based on the second feature and the third feature; and determining an optimal supply service resource based on the matching degree of the second type supply service resource sample and the demand description text.

Description

Technical Field

[0001] This application relates to the technical field of natural language processing, and in particular, to a method and device for matching supply service resources based on text features. Background Art

[0002] In modern supply chain management, the matching of supply service resources is a key link to ensure the efficient operation of the supply chain. Supply service resources cover the basic information of suppliers, supply product specifications, and technical patents that may be related to the supply products. Precise and efficient matching of supply service resources can help enterprises quickly determine the suppliers and product resources that meet their needs, thereby shortening the procurement cycle, reducing costs, and improving the overall operation efficiency of the supply chain. In the context of the increasingly large scale of the supply chain, how to quickly determine the optimal supply service resources is an important challenge faced by modern supply chain management.

[0003] Traditional supply service resource matching methods mainly rely on manual review and manual matching. Specifically, the demand description text is usually parsed and compared manually, and by referring to the basic information, product specifications, and relevant patent documents provided by suppliers, the supply resources that meet the requirements are screened out. In recent years, with the development of natural language processing (NLP) technology, some automated matching systems have begun to be applied, using keyword matching or simple text similarity calculation for preliminary matching. These systems calculate the similarity between the demand description text and the supply resource text to provide partial automated support. However, the current matching methods still have the following main defects: Traditional manual matching is time-consuming and laborious, and it is difficult to cope with large-scale supply service resource databases; existing methods based on keyword or simple similarity calculation, although they can improve part of the efficiency, are still difficult to quickly process massive information; manual matching depends on manual experience and is prone to deviation; automated systems usually rely on keyword matching or shallow semantic analysis and are difficult to understand the background, intention of user needs and the complex semantic information of supply service resources, resulting in limited relevance and accuracy of the matching results; existing general large language models lack in-depth understanding of the background knowledge in the supply chain field and are difficult to accurately capture the supplier background, product characteristics and their actual relevance to the demand, resulting in incomplete or inaccurate results.

[0004] Therefore, there is an urgent need for a method to accurately extract demand information through semantic parsing and reasoning of the demand description text, and combine the characteristics of supply service resources, and use a large model to quickly, accurately and reliably match supply service resources to improve the quality and efficiency of supply service resource matching. Summary of the Invention

[0005] In view of this, the present application provides a supply service resource supply-demand matching method and device based on text features, which are used to accurately extract demand information by semantic parsing and reasoning of demand description texts, and combine the features of supply service resources to quickly, accurately and reliably match supply service resources by using a large model, so as to improve the quality and efficiency of supply service resource matching.

[0006] Specifically, the present application is implemented through the following technical solutions:

[0007] In the first aspect of the present application, a supply service resource supply-demand matching method based on text features is provided, and the method includes:

[0008] Obtain multiple topics from the demand description text of supply service resources, and match the description types of supply service resources;

[0009] Obtain the description data corresponding to each description type, combine the multiple description data of each supply service resource as a sample, and obtain multiple supply service resource samples;

[0010] Extract the first feature of each supply service resource sample and the second feature of the demand description text based on a text representation model, screen the multiple supply service resource samples based on the similarity between the first feature and the second feature, and determine the first type of supply service resource samples; the text representation model is arranged in a computer;

[0011] Generate a candidate demand set according to each supply service resource sample in the first type of supply service resource samples;

[0012] Extract the third feature of each candidate enterprise demand description text in the candidate demand set based on the text representation model, and screen the first type of supply service resource samples based on each second feature and each third feature to determine the second type of supply service resource samples;

[0013] Determine the optimal supply service resource based on the matching degree between the second type of supply service resource samples and the demand description text.

[0014] In the second aspect of the present application, a supply service resource supply-demand matching device based on text features is provided, and the device includes an acquisition module, an extraction module, a generation module and a determination module;

[0015] Among them, the acquisition module is used to obtain multiple topics from the demand description text of supply service resources and match the description types of supply service resources;

[0016] The acquisition module is further used to obtain the description data corresponding to each description type, combine the multiple description data of each supply service resource as a sample, and obtain multiple supply service resource samples;

[0017] The extraction module is configured to extract the first feature of each of the supply service resource samples and the second feature of the demand description text based on a text representation model, screen the multiple supply service resource samples based on the similarity between the first feature and the second feature, and determine the first type of supply service resource samples; the text representation model is arranged in a computer;

[0018] The generation module is configured to generate a candidate demand set according to each supply service resource sample in the first type of supply service resource samples;

[0019] The determination module is configured to extract the third feature of each candidate enterprise demand description text in the candidate demand set based on a text representation model, screen the first type of supply service resource samples based on each of the second features and each of the third features, and determine the second type of supply service resource samples;

[0020] The determination module is further configured to determine an optimal supply service resource based on the matching degree between the second type of supply service resource samples and the demand description text.

[0021] The supply service resource supply-demand matching method and device based on text features provided by this application gradually optimize the accuracy and efficiency of supply service resource matching through two rounds of screening. Each screening targets different objects. The first screening starts from the demand description text and matches the demand with the samples. The second screening starts from the samples that have been matched and generates the most matching predicted demand according to the samples. By matching the demand with the demand, the most matching samples are determined. Different strategies are adopted to avoid insufficient model matching accuracy caused by insufficient feature extraction and model matching capabilities. It has hierarchy and coordination, and improves the accuracy of demand and service resource matching. First of all, the first screening targets all supply service resource samples. Based on the first features of all supply service resource samples and the second features of the demand description text, samples that are basically relevant to the demand description text are quickly screened out through semantic similarity calculation, reducing the screening scope and calculation cost, and shrinking the sample set from the global scale to the initially relevant sample set. At the same time, the combination of fusion features and auxiliary features is used to ensure that the results of the initial screening take into account the overall semantic relevance and multi-dimensional detail matching, laying the foundation for the screening accuracy. On the basis of the first screening, the second screening further screens the first type of supply service resource samples obtained from the first screening. The second screening generates a candidate demand set based on the first type of supply service resource samples, and further uses the features of the candidate demand set and the features of the demand description text for in-depth semantic matching. Through the dynamic generation of the candidate demand set, this step can be closer to the enterprise demand semantics and supplement the semantic details that may be missed in the first screening. Using the candidate demand set to refine the multi-dimensional semantic features of the resource samples improves the matching accuracy and avoids the interference of samples with high correlation but low accuracy that may be caused by the first screening. To sum up, this two-round screening method combines efficiency and accuracy. The first screening reduces the sample size through preliminary filtering, solving the problem of calculation cost in global screening; the second screening ensures the high accuracy of the final matching result through the dynamic generation and in-depth matching of the candidate demand set. The overall design realizes the progressive resource optimization matching layer by layer, and establishes an efficient and accurate matching mechanism between supply service resources and enterprise demands. Brief Description of the Drawings

[0022] Figure 1 It is a flowchart of the supply service resource supply-demand matching method based on text features provided by Embodiment 1 of this application;

[0023] Figure 2 It is a schematic structural diagram of the supply service resource supply-demand matching device based on text features provided by Embodiment 2 of this application. Detailed Embodiments

[0024] Exemplary embodiments will be described in detail herein, and examples thereof are shown in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application.

[0025] The terms used in this application are for the purpose of describing particular embodiments only and are not intended to limit the present application. The singular forms "a", "the", and "said" used in this application are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0026] It should be understood that although the terms first, second, third, etc. may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other. For example, without departing from the scope of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0027] Specific embodiments are given below to introduce the technical solutions of the present application in detail.

[0028] Figure 1 It is a flowchart of a supply service resource supply-demand matching method based on text features provided for Embodiment 1 of the present application. Please refer to Figure 1 , the method provided in this embodiment may include:

[0029] S101. Obtain multiple topics from the demand description text of the supply service resource and match the description type of the supply service resource.

[0030] Specifically, the demand description text refers to the natural language text content used by an enterprise to describe its specific needs, including information such as the products, service types, functional characteristics, technical specifications, budgets, application scenarios, etc. that the enterprise needs. The demand description text is used to clarify the specific needs of the enterprise in the supply chain so as to match the most suitable supply service resources.

[0031] Furthermore, the requirement description text contains multiple themes. Each theme is the main semantic unit or key element extracted from the requirement description text. A theme is an abstract representation that reflects the core information of the enterprise's requirements in the text content. A theme is the refinement and decomposition of the requirement description text, used to help the matching system transform the complex requirement description text into structured semantic information, facilitating subsequent matching with supply service resources. The multiple themes can include service objects, service contents, product-related themes, technical requirement themes, functional requirement themes, application scenario themes, budget or procurement limit themes, time requirement themes, etc.

[0032] Specifically, the supply service resources refer to all the information and data related to the supply chain provided by suppliers. The supply service resources are used to help enterprises meet specific requirements, such as providing products or technical support that meet the requirements. The description types of supply service resources are the classifications of different attributes or information in the supply service resources, and each description type corresponds to a specific attribute or feature of the supply service resources. There are multiple description types, including basic supplier information, supply product specifications, and technical patent information. Among them, the basic supplier information includes the enterprise name, enterprise profile, business scope, honorary qualifications, address, and contact information, etc.; the supply product specifications include the product name, product model, technical parameters, application scenarios, product function descriptions, etc.; the technical patent information includes the patent name, patent number, patent technology description, patent application field, etc.

[0033] It should be noted that each theme of the requirement description text corresponds to supply service resources of different description types. The description types corresponding to different themes may be different, and the supply service resource data corresponding to different themes may also be different.

[0034] In specific implementation, the requirement description text is obtained through relevant enterprise employees or from the enterprise management system, and the requirement description text is preprocessed to remove stop words (such as "de", "he", and other words without practical meaning) in the requirement description text, and the requirement description text is normalized and segmented. The pre-trained large language model is used to parse the preprocessed requirement description text, refine and decompose the requirement description text, identify the semantic structure in the requirement description text, and combined with the semantic representation ability of the large prediction model, multiple themes are extracted from the requirement description text. Furthermore, the multiple extracted themes are classified and screened, and the similar themes are merged or refined into more distinguishable specific themes to ensure that the themes have clear semantic meanings. Finally, the extracted and processed themes are determined as the multiple themes of the requirement description text.

[0035] Furthermore, pre-define supply service resources of multiple description types according to the attributes of the supply service resources (such as supplier basic information, supply product specifications, technical patent information, etc.). According to the defined description types, determine the specific attributes or information categories of the supply service resources corresponding to each description type. Analyze the content corresponding to each theme in the demand description text to determine the demand attributes mainly expressed by each theme. Semantically match the demand attributes mainly expressed by each theme with the specific attributes of the supply service resources corresponding to each description type to determine the description type most relevant to each theme.

[0036] For example: If a certain theme involves the technical parameter requirements of the supply product, then match the description type as "supply product specifications"; if a certain theme involves the service capabilities of the supplier, then match the description type "supplier basic information"; if a certain theme involves the patent name, then match the description type "technical patent information".

[0037] Optionally, obtaining multiple themes from the demand description text of the supply service resources and matching the description types of the supply service resources includes: performing semantic segmentation on the demand description text of the supply service resources to obtain multiple themes; obtaining the supply service history of each service object, and determining the supply service resource usage habits based on the supply service history; predicting the constraint conditions of the supply service resources based on the service content and the supply service resource usage habits; and matching the description types based on the constraint conditions.

[0038] In specific implementation, first, use a large model to parse the demand description text. After standard preprocessing steps, including text cleaning (removing stop words, punctuation), word segmentation, and syntactic analysis, input the processed text into the large language model to generate a context-aware vector representation. Then, adopt a clustering algorithm based on the embedded vector (such as K-Means) to automatically divide the text content into several semantically close sub-paragraphs, and each sub-paragraph represents a potential theme, such as "service object", "service content", "budget requirement", "delivery time", etc. For example, if the input demand description text is "Our company needs to find a battery management system supplier for the new energy vehicle manufacturing project, requiring fast charging function support and a delivery cycle within three months", after processing, it is automatically divided into Theme 1 "Service object: New energy vehicle manufacturing project", Theme 2 "Service content: Battery management system", Theme 3 "Function requirement: Fast charging support", and Theme 4 "Time requirement: Within three months".

[0039] Further, for the extracted "service object" theme, retrieve the supply service records of previous similar projects by calling the internal historical database or third-party APIs. For example, search the database for past procurement records according to the keywords "new energy vehicle" and "BMS project" to obtain information such as historical suppliers, transaction prices, delivery cycles, and performance evaluations. Then, based on the obtained historical supply service data, perform feature extraction through data analysis techniques. For example, use the DBSCAN clustering algorithm to perform unsupervised clustering on the supplier attributes in the supply records to determine usage habits, or extract common supplier features through frequency statistical analysis. Further, for the extracted "service content" theme, use a semantic analysis model (such as RoBERTa or T5) to extract key constraint words, such as "fast charging function", "battery protection strategy", "CAN bus interface compatibility", etc. Subsequently, fuse the key constraint words with the preference features extracted from historical data (such as supplier certification status, product technical parameters, etc.) to construct a feature vector and input it into the prediction model. The prediction model outputs specific constraint conditions for supply service resources based on the input features, such as "the battery management system needs to support a fast charging current of ≥150A", "the delivery period shall not exceed 90 days", "the supplier needs to have ISO9001 and ISO26262 certifications". Finally, according to the generated constraint conditions, establish a mapping relationship between the theme and the description type. For example, map the functional constraint "fast charging function" to the supply product specification description type, map the qualification requirements to the supplier basic information description type, and map the technical detail requirements to the technical patent information description type. Through a semantic matching engine (such as the Siamese Network text matching model), identify the description type that best matches each theme constraint condition in the supply resource data to achieve an accurate match between the theme and the description type.

[0040] It should be noted that the theme and the description type are in one-to-one correspondence. Each theme-limited requirement description text describes what aspects of information, such as who the initiating object is, who the service object is, and what the specific content of the requirement is; each description type is about how the supply service resource performs in this theme aspect. For example, if a semantic segment in the requirement description text describes who the initiating object is, the corresponding type of the supply service resource is the attribute description of which type of customer it is more suitable for, and the description data under the corresponding description type is the specific content of which type of customer this supply service resource is more suitable for. That is to say, each supply service resource sample includes multiple description data, each description data corresponds to a description type, and the number of description types in the supply service resource sample is the same as the number of themes in the requirement description text.

[0041] S102. Obtain the description data corresponding to each description type, combine the multiple description data of each supply service resource as a sample, and obtain multiple supply service resource samples.

[0042] Specifically, each supply service resource sample includes the description data of the supply service resources under each description type, that is, each supply service resource sample includes the supply service resource data under each description type. Combining the above description, each supply service resource sample includes the supply service resource data corresponding to the basic supplier information, the supply service resource data corresponding to the supply product specification, and the supply service resource data corresponding to the technical patent information.

[0043] In specific implementation, according to the determined description types, collect the supply service resource data corresponding to all description types. For each description type, collect the supply service resource data corresponding to that description type. Combine the supply service resource data corresponding to each description type into a complete supply service resource sample. Repeat the operation according to the above steps, gradually collect and construct multiple supply service resource samples to ensure that each supply service resource sample covers the supply service resource data of all description types.

[0044] For example, for the description type of "basic supplier information", obtain the supply service resource data such as the company profile and business scope of the supplier; for the description type of "supply product specification", obtain the supply service resource data such as product specifications and performance indicators. Each supply service resource sample contains the supply service resource data of multiple description types such as basic supplier information, supply product specification, and technical patent information. Each supply service resource sample can be expressed as , where the is the supply service resource sample, the is the basic supplier information, the is the supply product specification, and the is the technical patent information.

[0045] S103. Extract the first feature of each of the supply service resource samples and the second feature of the demand description text based on the text representation model, and screen the multiple supply service resource samples based on the similarity between the first feature and the second feature to determine the first type of supply service resource samples.

[0046] Specifically, the first feature includes a first fusion feature and multiple first auxiliary features corresponding to each description data. The second feature includes a second fusion feature and multiple second auxiliary features corresponding to each theme.

[0047] The first fusion feature is the fusion feature of supply service resource samples, and each first auxiliary feature is the feature corresponding to a single description data of the supply service resource sample. That is, the first fusion feature is the fusion feature of supply service resource data of multiple description types. The first fusion feature is single, and the first auxiliary feature is the single feature of the supply service resource data of each description type. The first auxiliary features are multiple, and the number of the first auxiliary features is the same as the number of description types. Among them, the information content of the same description data in the first fusion feature and the first auxiliary feature is different.

[0048] In specific implementation, extracting the first feature of each of the supply service resource samples based on the text representation model includes: extracting multiple features of the supply service resource data of different description types in each supply service resource sample at the first depth based on the text representation model; extracting multiple features of the supply service resource data of different description types in each supply service resource sample at the second depth based on the text representation model, and determining them as multiple first auxiliary features; the second depth is greater than the first depth.

[0049] Specifically, the text representation model is a model based on natural language processing (NLP) technology, which is used to convert text data into a numerical feature representation form that can be processed by machines. Common text representation models include large models based on deep learning, such as BERT (Bidirectional Encoder Representations from Transformers), GPT (Generative Pre-trained Transformer), RoBERTa, etc. The text representation model can generate high-quality text vector representations by understanding language information such as semantics and context. The text representation model supports feature extraction at different depths (such as the first depth and the second depth), enabling the text representation model to capture semantic information from coarse-grained to fine-grained, and providing support for modeling the relationships between description type data.

[0050] Furthermore, the text representation model is arranged in a computer. During the process of feature extraction by the text representation model, the text representation model relied on is usually a large model with a huge number of parameters and extremely high computing requirements, and its operation poses high requirements on computer hardware. Specifically, computer hardware mainly supports the efficient operation of the large model through high-performance computing resources, including but not limited to key components such as graphics processing units (GPUs), tensor processing units (TPUs), and high-bandwidth memories (such as HBM). Among them, GPUs have powerful parallel computing capabilities and can accelerate matrix operations and vector calculations in deep learning models. They are the current mainstream large model inference and training platforms; TPUs are dedicated hardware optimized for deep neural networks and can further improve computing efficiency. In addition, CPUs are used to coordinate system operations and process non-parallel tasks, caches and memories ensure high-speed reading and writing and temporary storage of data during the inference and training of large-scale models, and hard disk systems are used to store model weights and training data. Therefore, the efficient operation of large models not only depends on advanced algorithm designs but also highly depends on the comprehensive performance guarantee of underlying computer hardware.

[0051] In specific implementation, the text representation model is deployed in a server, and the central processing unit of the server is responsible for coordinating the scheduling of each hardware component, including task distribution, memory management, and data preprocessing. Subsequently, the core computing tasks of the large model are assigned to graphics processing units or tensor processing units, and these computing accelerators operate in a highly parallel architecture, and a large number of computing cores simultaneously execute operations such as matrix multiplication, activation functions, and self-attention mechanisms in neural networks. For example, in the Transformer architecture, a large number of vector dot products and weight matrix multiplications are divided into tiny tasks and are respectively executed in parallel by hundreds of GPU cores to accelerate the efficiency of forward propagation and backward propagation. During the computing process, caches and high-bandwidth video memories are used to quickly access intermediate activation values, weight parameters, and gradient information to ensure that data can flow efficiently between each computing unit without becoming a bottleneck. The memory controller manages the data reading and writing operations between the main memory and the video memory and optimizes the scheduling of data in different hierarchical storage structures. Some platforms also adopt data pipelining and tensor parallel mechanisms to split different computing layers or tensor dimensions of the model among multiple graphics processing units. Finally, the executed feature vectors are sent back from the accelerator to the host system and are output or further processed through the input / output interface controlled by the central processing unit. The entire process, from the control logic of the central processing unit, to the highly parallel computing of the graphics processing unit / tensor processing unit, and then to the efficient data transmission of the memory, each hardware module works together through clear role divisions and communication mechanisms to jointly complete the operation task of the large model.

[0052] It should be noted that after the text representation model is trained, features can be extracted from the input text based on the model. The input data is a supply service resource sample. According to different requirements for the information content of the output features, features are extracted from different layers of output. Once the parameters and structure of the model are determined, the relationship between the model input and output is corresponding. That is to say, there is a corresponding relationship between the input supply service resource sample and the output features, and the corresponding relationship between the two is restricted by natural laws. Since the text data itself contains stable semantic structures, language logics, and context rules, and the text representation model precisely learns and captures these language rules through a large amount of corpora to achieve the vectorized expression of the input text. Therefore, each input text will be mapped to a definite, high-dimensional semantic feature representation, and this mapping relationship is jointly determined by the model structure and training data. That is to say, the text representation model does not "create" features out of thin air, but through modeling at the language levels such as semantics, morphology, and syntax, under the constraints of computational rules and model weights, it extracts features that conform to language rules from the input text. Therefore, the content, structure, and semantics of the input text determine the manifestation form of the output features. This is a deterministic process jointly restricted by language rules, the large model training mechanism, and computational structure, with a clear one-to-one correspondence.

[0053] Furthermore, the first depth is less than the second depth. The first depth focuses on extracting the macroscopic global semantic features of the supply service resource sample for fusing features of different description types. The second depth focuses on extracting refined local semantic features for capturing the differences of various description features.

[0054] In specific implementation, first, the supply service resource data of different description types in each supply service resource sample is processed. Specifically, assume there is a supply service resource sample that contains text data of two description types: "data analysis ability requirements" and "service time limit requirements". At the first depth, the text representation model will extract multiple feature vectors from the text of these two description types. The first depth focuses on extracting macroscopic global semantic features. That is to say, by understanding the text content of these description types, the text representation model can extract the coarse-grained information of the entire text, such as the relationship between "data analysis ability" and "service time limit", reflecting the overall meaning of the text. For example, when the text "It is necessary to have the data analysis ability and be able to deliver results within two weeks" is processed, the first-depth feature vectors generated by the text representation model through BERT may contain the following information: The service described in this text is a "data analysis" task and has a "time limit requirement". These feature vectors synthesize the global semantics of the text and provide a basis for further matching and analysis.

[0055] Next, when further processing other types of descriptive text in the same supply service resource sample, at the second depth, the text representation model will further extract more fine-grained feature vectors. The feature extraction at the second depth pays more attention to local semantic information and is specifically used to capture the subtle differences between descriptive features. For example, after processing at the second depth, the text of "data analysis ability requirements" may extract more specific information, such as the specific requirements of "being able to analyze big data", "being proficient in data modeling", and "being familiar with specific analysis tools". For the text of "service time limit requirements", the second-depth features may identify different time limit types, such as "urgent requirements" and "regular requirements", and their corresponding processing methods. These more fine-grained feature vectors are multiple first auxiliary features, which can reflect the specific differences of various descriptive features in the text. Finally, based on the global features extracted at the first depth and the local features extracted at the second depth, the first fusion features generated by combining this information will be used to describe the comprehensive features of the entire supply service resource sample, and these auxiliary features will be further used for precise matching and screening of resource samples. For example, assume that a supply service resource sample contains the text "having five years of data analysis experience and being able to deliver an analysis report within three months". The feature vector at the first depth may indicate that this is a resource requirement for data analysis and there is a clear delivery time. The feature vector at the second depth can reveal the specific technical details of the resource sample, such as "being proficient in a specific analysis tool" and "delivery time is three months" and other information, so as to provide more details for subsequent precise matching.

[0056] Optionally, the extracting the first feature of each supply service resource sample based on the text representation model includes: constructing a cross-feature constraint relationship based on the logical relevance between supply service resource data of different description types; determining the first weight of supply service resource data of different description types based on the matching degree between supply service resource data of different description types and the requirement description text; extracting features of supply service resource data of different description types based on the text representation model; performing weighted fusion on the features based on the first weight to obtain a first fusion feature; and adjusting the first fusion feature based on the cross-feature constraint relationship to obtain an adjusted first fusion feature.

[0057] In specific implementation, first, preprocess the supply service resource data, and divide the supply resource information into different categories according to the preset description types. Then, analyze the logical relevance between different description types. Specifically, construct a logical relationship graph of supply service resources, and use knowledge graph technology to model each description type as a node and model the business logic relationship between them as an edge. For example, a positive association relationship is established between the "supplier scale" node and the "delivery capacity" node, and a dependency relationship is established between the "product specification" node and the "application scenario" node. The relationship strength of the dependency relationship is assigned according to historical business rules or expert experience (for example, an association weight of 0.8 indicates a high correlation). Then, use a large language model (such as BERT) to encode the supply service resource data under each description type into a semantic vector of a fixed length, and at the same time encode the demand description text into a semantic vector. For each description type, use the cosine similarity formula to calculate the similarity between its vector and the demand description text vector, and normalize the obtained similarity score. For example, use the Min-Max normalization method to linearly transform the similarity value to the interval [0, 1]. The normalized score is the first weight of the corresponding description type.

[0058] Furthermore, based on the text representation model, extract the data feature vectors under each description type respectively. Taking the supply product manual as an example, deep semantic features at the sentence level or paragraph level can be extracted through the BERT model (such as extracting the context relationship of product function descriptions, entity attributes of technical parameters, etc.) to form a feature matrix. The feature vectors of all description types are weighted and fused according to the first weight. That is, multiply the feature vector extracted for each description type by its corresponding weight, and then splice or average all the weighted feature vectors to form the first fusion feature of the supply service resource sample. For example, the weight of the supply product manual is 0.6, the weight of the supplier basic information is 0.3, and the weight of the technical patent information is 0.1. The corresponding features are weighted and fused respectively to obtain a comprehensive representation. Subsequently, based on the previously constructed logical association knowledge graph, further adjust the first fusion feature. The specific adjustment methods include: if there is a strong positive association between two description types logically (such as supplier scale and delivery capacity), then enhance their correlation in the fusion feature. For example, linearly combine and weight their feature vectors and then inject them into the fusion feature. If there is a negative or conflicting relationship (such as "low price" and "high-end technical indicators"), then introduce an inhibitory factor in the feature fusion to reduce the feature contribution of the conflicting part between the two. Process through feature normalization and regularization methods (such as L2 regularization). If the feature of a certain description type causes an out-of-reasonable threshold after overall fusion (such as abnormal exaggeration of product specifications), then apply normalization processing to the first fusion feature to ensure a reasonable distribution of the final feature space. Finally, through the above comprehensive processing methods, the adjusted first fusion feature that takes into account the importance of each description type and conforms to the cross-feature logical constraints is obtained.

[0059] The method provided in this embodiment, based on the feature fusion method of logical association, matching degree weighting, and feature adjustment, can effectively solve the conflicts and redundancies of information of different description types, enhance the accuracy and relevance of feature representation, thereby improving the matching degree between supply service resources and enterprise requirements, and ultimately achieving the purpose of optimizing the matching result. First, a cross-feature constraint relationship is constructed based on the logical relevance between supply service resource data of different description types, ensuring that features of different description types can maintain consistency and effectiveness during fusion. For example, the "price" and "specification" of supply service resources are usually interrelated, and a reasonable cross-feature constraint relationship can effectively avoid unreasonable conflicts or invalid combinations between these features, thereby improving the accuracy and representativeness of feature fusion. Secondly, weights are determined based on the matching degree between supply service resource data of different description types and the demand description text, enabling the feature fusion process to prioritize description types highly relevant to the demand text, which helps improve the matching accuracy of the model to the target demand. Through weighted fusion, the proportion of features of different description types in the fused feature vector will be reasonably adjusted according to their importance, thereby enhancing the ability of the final feature to reflect the true attributes of supply service resources. In addition, by extracting features of each description type through a text representation model and then performing weighted fusion on these features, different information of each description type can be integrated, avoiding the bias of a single feature on the matching result. For example, a supply resource may match the demand in terms of "specification" but not in terms of "price". Adopting weighted fusion can comprehensively consider these two aspects, thereby reflecting a more comprehensive resource feature in the fused feature. Finally, by adjusting the first fused feature based on the cross-feature constraint relationship, the quality of the feature is further improved, ensuring that the fused feature not only maintains semantic consistency but also conforms to the actual logical relationship between each description type. This adjustment can effectively optimize feature representation, making the final feature more accurate and practically valuable, thereby enhancing the effect of the entire resource matching system.

[0060] The method provided in this embodiment can effectively improve the accuracy and efficiency of supply service resource matching by extracting the first fusion feature and the first auxiliary feature of the supply service resource sample. The first fusion feature provides a global expression of the supply service resource by averaging the features extracted from data of different description types at the first depth. It can condense complex semantic information, simplify the feature dimension, and reduce the computational complexity, while ensuring the integrity of the overall semantics. The first auxiliary feature, on the other hand, retains the detailed information of different description types by extracting multi-dimensional features at the second depth, further enriching the hierarchical nature of the semantic expression and enabling each attribute of the supply service resource to be described and characterized in detail. This way of expressing that takes both the global and the details into account can quickly filter out irrelevant samples during the supply service resource matching process and provide a higher discrimination degree in the refined screening, ensuring the accuracy and rationality of the final matching result. The first fusion feature and the first auxiliary feature together constitute the basis for the multi-level matching of the supply service resource sample and the demand description text. The first fusion feature, as the core of the preliminary screening, quickly completes the matching of the enterprise demand and the global semantics of the resource, greatly improving the screening efficiency; while the first auxiliary feature further precisely locates the preliminary screening result through the fine-grained similarity calculation between description types in the subsequent stage. Such a design not only adapts to the diverse characteristics of enterprise demands from macro to micro, but also takes into account the computational efficiency and the final effect of the matching process. In addition, the hierarchical extraction method of the first fusion feature and the first auxiliary feature in the text representation model enables the matching system to achieve a comprehensive semantic understanding of the supply service resource and the enterprise demand from the global to the local, further enhancing the intelligence level and adaptability of the matching process.

[0061] Specifically, in combination with the above description, each theme of the demand description text corresponds to supply service resource data of multiple description types, that is, the supply service resource data of each description type is determined according to the theme of the demand description text. There may be many description types. In this embodiment, only three description types (basic information of suppliers, supply product specifications, technical patent information) are taken as examples for introduction. Structured supply enterprise information refers to the basic information of suppliers or enterprises organized in an orderly and standardized manner, including industry, location, main products, honorary qualifications, enterprise scale, output, etc. Functional supply product information refers to the detailed description of supply products, which focuses on the characteristics, functions, and uses of the products, including name, model, main technical parameters, application scenarios, etc.

[0062] Similarly, the second feature includes a second fusion feature and multiple second auxiliary features corresponding to each topic. The second fusion feature is the fusion feature of each topic in the requirement description text, and each second auxiliary feature is the feature of each topic in the requirement description text. That is, the second fusion feature is the fusion feature of multiple topics, the second fusion feature is single, the second auxiliary feature is the single feature of each topic, and the second auxiliary features are multiple. The first type of supply service resource sample refers to the supply service resource samples that are more matched with the requirement description text among multiple supply service resource samples.

[0063] In specific implementation, using a text representation model, multiple feature vectors are extracted for different topics in the requirement description text at the first depth, and the average value of these feature vectors is calculated and determined as the second fusion feature. Further, using the text representation model again, more fine-grained multiple feature vectors are extracted for different topics in the requirement description text at the second depth to obtain multiple second auxiliary features.

[0064] Further, screening the multiple supply service resource samples based on the similarity between the first feature and the second feature to determine the first type of supply service resource samples includes: calculating the first similarity between each first fusion feature and the second fusion feature; establishing a matching relationship between the description type and the topic; calculating the first auxiliary similarity between the first auxiliary feature and the second auxiliary feature according to the matching relationship; determining the second weight based on the influence degree of the semantic of the text segment corresponding to each topic in the requirement description text on the comprehensive semantic of the requirement description text; weighting the first auxiliary similarity based on the second weight to obtain the second similarity between the first auxiliary feature of each supply service resource sample and the second auxiliary feature of the requirement description text; correcting the first similarity based on the second similarity; and screening the first type of supply service resource samples according to the size of the corrected first similarity.

[0065] Specifically, combined with the above description, the supply service resource data of each description type corresponds to a topic of the requirement description text. Therefore, the first auxiliary feature of the supply service resource data of each description type also corresponds to a topic of the requirement description text.

[0066] It should be noted that the first preset quantity is not a fixed value but dynamically changes. The size of the first preset quantity is closely related to the requirement description text, and the first preset quantity is related to the service content and time limit requirements in the requirement description text. Among them, the service content is the key factor in determining the quantity of service resource samples to be supplied. The service content includes the professionalism, complexity, and diversity required for the service, etc. For services with high professionalism, such as technical support or high-end manufacturing, more precise supply of service resource samples is required. Therefore, the matching quantity can be smaller, but the quality requirements for each sample are higher. For services with high diversity requirements, such as comprehensive services or distributed support, more samples are needed to cover the requirements of different scenarios. For composite service scenarios, when a theme involves multiple related service contents, more resource samples need to be matched to ensure full satisfaction of the requirements. The more urgent the time limit requirement, the greater the demand for the quantity of resources to ensure sufficient selectivity. For urgent service requirements, more candidate resource samples are needed to complete the service within a short time for quick screening and adjustment. For loose service requirements, if the time limit requirement is relatively loose, the quantity of initial resource samples can be reduced, and the focus is on precise matching.

[0067] Specifically, first, use semantic analysis techniques (such as BERT) to preprocess the requirement description text, including operations such as removing stop words, word segmentation, and syntactic dependency analysis. Subsequently, extract the main semantic themes, such as "service content" and "time limit requirement", through clustering analysis or keyword extraction methods. Each theme corresponds to a text fragment. After vectorizing using BERT, determine its semantic weight by calculating the cosine similarity between each theme fragment and the overall requirement description text. After normalization, obtain the proportion of each theme in the overall requirement description. For example, for a requirement description: "We need to purchase a batch of communication chips supporting high-speed data transmission within two months for use in 5G base station equipment", the extracted themes include: "Communication chip procurement requirement" (service content theme, accounting for 70% of the overall text weight), "Two-month delivery cycle" (time limit requirement theme, accounting for 30% of the overall text weight).

[0068] After obtaining the subject, determine the initial value of the first preset quantity based on the complexity and diversity of the service content. Specifically, measure the complexity and diversity by analyzing the number of key technical indicators, the density of professional terms, and the span of technical fields in the service content. For example, when the service content involves multiple fields (such as "communication chips", "5G protocols", "high-speed interface standards") and has rich technical details, it is determined that both the complexity and diversity of the service content are high, so a relatively high initial preset quantity is set (such as an initial value of 100). If the service content is relatively simple, such as "purchasing standard model power modules", a lower initial value is set (such as an initial value of 30). Further, extract the time limit requirements, and judge the urgency of the delivery time through a rule engine or keyword recognition (such as "urgent", "deliver within 2 months", etc.). Establish an urgency level. For example: "Delivery time ≤ 2 months" is highly urgent, and the first preset quantity needs to be increased by 20%; "Delivery time is 3 - 6 months" is moderately urgent and remains unchanged; "Delivery time ≥ 6 months" is a loose requirement, and the first preset quantity is reduced by 10%. Taking the above example as an example, since "deliver within two months" is highly urgent, on the basis of the initial value of 100, 20% is increased, and the adjusted preset quantity is 120.

[0069] Finally, comprehensively consider the adjusted initial values of each subject, summarize all subject contributions, and calculate the weighted average according to the weights to obtain the final first preset quantity. For example, the subject contribution weight of the service content is 0.7, and the adjusted sample quantity is 120; the subject contribution weight of the time limit requirement is 0.3, and the adjusted sample quantity is 120. Then the comprehensive preset quantity = 120×0.7 + 120×0.3 = 120. Finally, determine that the first preset quantity in this demand matching process is 120 supply service resource samples, which serves as the initial quantity for subsequent sample screening.

[0070] The method provided in this embodiment can dynamically determine the first preset quantity based on the service content and time limit requirements in the demand description text, and can effectively balance the accuracy and calculation efficiency of matching during the resource screening process. First, setting the basic quantity in combination with the complexity and diversity of the service content can ensure that the selected resource samples have the ability to cover all key requirements; second, adding the time limit requirement as a dynamic adjustment factor enables the screening to adapt to the high-priority processing of urgent needs, avoiding delays caused by insufficient resources or waste caused by resource redundancy. At the same time, the dynamic adjustment mechanism can also improve the system's adaptability to different enterprise demand scenarios, provide a sufficient sample basis for subsequent refined screening, ensure that the screening results have both quality and breadth, and thus improve the accuracy of the final matching and the availability of services.

[0071] In specific implementation, the main text features are extracted from the requirement description text through the above steps. For example, "technical support" is extracted from the requirement description text as the service content and "urgent" as the time limit requirement. Subsequently, for each supply service resource sample, relevant features are extracted from its description, such as the service type (e.g., "technical support") and the time limit requirement (e.g., "urgent"). Then, the similarity between the first fusion feature of each supply service resource sample and the second fusion feature of the requirement description text is calculated. For example, using the cosine similarity algorithm, by taking "technical support" and "urgent" as keywords, each resource sample is calculated to obtain a preliminary similarity score. Next, a matching relationship is established based on the first auxiliary features (such as service type, field, etc.) of each resource sample and the corresponding topics in the requirement description text. For example, if "technical support" is mentioned in the requirement description text, all supply service resource samples related to "technical support" are grouped together. For these samples, the first auxiliary similarity is obtained by calculating the similarity between each first auxiliary feature and the second auxiliary feature. Suppose "technical support" is mentioned in the requirement text, and in a certain supply service resource sample, the first auxiliary feature of this sample is "computer technical support", calculate the similarity between it and "technical support" in the requirement text, for example, using word embedding technology to calculate the similarity between the two, and obtain the corresponding first auxiliary similarity value.

[0072] Furthermore, after obtaining the first auxiliary similarity, the semantic proportion of each topic in the requirement description text is analyzed. For example, in the requirement text, "technical support" may occupy a relatively large proportion, while the time limit requirement (such as "urgent") may occupy a relatively small proportion. The weight of each topic is calculated through semantic analysis technology, and the first auxiliary similarity is weighted based on this weight. For example, if "technical support" accounts for 70% importance in the requirement text and "urgent" accounts for 30% importance, the similarity is adjusted according to this ratio to obtain the weighted second similarity value. Based on the second similarity, the preliminarily calculated similarity is further corrected. For example, if a certain supply service resource sample has a relatively high similarity with the requirement description text in the first step, but it is found that the importance of the "urgent" time limit requirement is higher after the weighting process, the matching score of this resource sample is adjusted to correct the preliminary similarity. Finally, all supply service resource samples are sorted in ascending order according to the corrected similarity value to ensure that the resource samples that best meet the requirements are ranked at the front.

[0073] The method provided in this embodiment, firstly, the process of calculating similarity helps to quantify the matching degree between the supply service resources and the demand description text, making the matching result more accurate and measurable. By comparing the similarity between the first fusion feature and the second fusion feature, this application can first capture the overall semantic fit between the two main elements, so as to ensure that the resource samples closest to the enterprise's needs are selected from the very beginning. In addition, calculating the similarity between each first auxiliary feature and the second auxiliary feature corresponding to the theme in the enterprise's needs further refines the matching degree evaluation, emphasizing the refined processing of specific description types and themes. The weighted similarity calculation ensures that the importance of different description types in the overall matching is appropriately reflected, and the weights are dynamically adjusted according to the degree of semantic influence. In this way, the influence of some irrelevant or low-correlated features on the final result can be avoided, thereby improving the accuracy and relevance. Secondly, based on this similarity calculation, the most matching supply service resource samples can be effectively screened out, and the resource samples are sorted according to the similarity values, ensuring that the first type of supply service resource samples finally selected have the highest degree of fit with the enterprise's needs. Through the process of correcting the similarity, the screening priority of the samples can also be dynamically adjusted, making the matching process more flexible and accurate. Finally, this similarity calculation mechanism not only enhances the intelligence of the method, but also greatly improves the accuracy and efficiency of the matching process, avoiding the rough screening in the traditional method, being able to accurately reflect the specific requirements of the enterprise's needs, and providing high-quality resource matching results. These steps work together to ensure the selection of the optimal supply service resources and the satisfaction of the enterprise's needs.

[0074] S104. Generate a candidate demand set according to each supply service resource sample in the first type of supply service resource samples.

[0075] Specifically, the candidate demand set includes candidate enterprise demand description texts generated according to the description data of each description type of a supply service resource sample.

[0076] When specifically implemented, generating a candidate demand set according to each supply service resource sample in the first type of supply service resource samples includes: determining a first prompt word of a large language model based on each supply service resource sample in the first type of supply service resource samples and the feature extraction requirements; the large language model generating multiple initial segments based on the first prompt word and the first fusion feature of each supply service resource sample; the large language model generating multiple modification segments based on the first prompt word and the multiple first auxiliary features of each supply service resource sample; matching the corresponding multiple target modification segments based on the initial segments, and processing the initial segments based on the multiple target modification segments to generate multiple processed candidate demand texts, so as to obtain the candidate demand set.

[0077] Specifically, the initial segment is a text segment generated by a large language model based on the first fusion feature of the supply service resource sample. The initial segment is preliminary and basic text information used to describe the relevant features of the supply service resource sample and its preliminary matching. The modification segment is a text segment generated by a large language model based on the first auxiliary feature of the supply service resource sample. The modification segment contains further expansion, supplementation, and optimization of the initial segment. The modification segment further refines the features of the supply service resource sample, emphasizing its matching degree with enterprise requirements or specific attributes.

[0078] In specific implementation, first, extract data of multiple description types for each sample from historical supply service resource samples, including "service type", "service time limit", "resource availability", and "service field", etc. Based on the data of these description types, combined with the requirement description submitted by the demander, generate the "prompt words" of the large language model. For example, assume a demander needs cybersecurity technical support and requires a problem to be solved within 24 hours. The generated prompt words will include "Provide cybersecurity technical support, with an urgent response time limit, required to be solved within 24 hours, the service field includes cloud computing and cybersecurity, available 7x24 hours". Then, input this prompt word together with the fusion feature of each supply service resource sample into the large language model. The large language model generates multiple initial segments based on these inputs, and these segments contain basic descriptions that match the requirements. For example, the generated initial segment may be "Provide cybersecurity technical support services to ensure a response and solution within 24 hours". In addition, based on multiple auxiliary features of each supply service resource sample (such as service advantages, historical successful cases, etc.), the large language model will also generate multiple modification segments, which are intended to supplement and expand the initial segment. The generated modification segments may be "Have a globally leading technical team that can quickly identify and solve customers' cybersecurity problems" or "Provide round-the-clock technical support to ensure that customers' problems can be quickly responded to and solved at any time".

[0079] Next, through semantic matching algorithms such as cosine similarity or BERT model, calculate the semantic similarity between the initial fragment and the modifying fragment, and select the most suitable modifying fragment to supplement the initial fragment according to the similarity. For example, the modifying fragment "has a globally leading technical team that can quickly identify and solve customers' network security problems" has a high matching degree with the initial fragment "provides network security technical support services to ensure response and solution within 24 hours", so it is processed as the target modifying fragment. After the target modifying fragment is combined with the initial fragment, the text content will be optimized to form a processed fragment. For example, "We provide efficient network security technical support services to ensure response and solution to all customers' problems within 24 hours. Our technical team has globally leading technical capabilities, can quickly identify and solve various network security problems, and provides 7x24-hour real-time monitoring and technical support". Finally, multiple processed fragments are assembled into a candidate requirement set, and these text fragments are sorted according to the specific requirements of the requester. Each text fragment in the candidate requirement set contains accurate service content and time limit requirements, meets the actual needs of the requester, and finally generates a set of supply service resource samples for the requester to choose from.

[0080] Optionally, perform semantic parsing on each generated initial fragment, combine the structure and content of the initial fragment, and determine the target sub-fragment in the initial fragment corresponding to the modifying fragment to ensure that the semantics of the located sub-fragment in the initial fragment are related to the semantics of the modifying fragment. Further, perform in-depth semantic analysis on each modifying fragment, extract its core information, including supplementary descriptions, limiting conditions, or additional details, etc. Obtain the refined information of the modifying fragment through semantic feature extraction technology. Integrate the semantic content of the modifying fragment into the corresponding target sub-fragment, integrate the refined information in the modifying fragment with the core content of the target sub-fragment, perform semantic extension or semantic enhancement, and use a language generation model to generate a new target sub-fragment that is fluent and conforms to the context logic. Replace the regenerated target sub-fragment with the corresponding position in the original initial fragment, keep the overall sentence structure intact, and at the same time optimize the expression integrity and semantic accuracy of the initial fragment. Finally, splice or adjust all the processed initial fragments according to the coherence of the semantic content to generate multiple processed candidate requirement texts and form a candidate requirement set.

[0081] In the method provided in this embodiment, on the one hand, when generating the candidate requirement set, a large language model is used to generate multiple initial segments and modification segments. These segments construct a multi-angle description of the supply service resources based on different feature dimensions (such as fusion features and auxiliary features). The initial segment provides a basic requirement description, while the modification segment further refines the specific attributes and details of this description. This multi-level generation process can ensure that the requirement description is more complete and comprehensive, while avoiding the deviation or incomplete information that may be caused by a single description method. Secondly, continuously processing the initial segment with the modification segment can not only ensure that the expression of the candidate requirement is more accurate and fluent, but also increase the context adaptability of the text, making it more in line with the specific enterprise requirements. This processing can be customized based on the context of the actual requirements, thereby optimizing the matching quality and reducing the matching of irrelevant resources. On the other hand, this method of generating the candidate requirement set not only improves the intelligence and adaptability of the system, but also provides a rich set of candidates for the subsequent screening and matching processes. By further screening based on the candidate requirement set (such as determining the optimal supply service resources by comparing the similarity with the requirement description text), the supply service resources highly consistent with the actual requirements can be determined more efficiently. In summary, this setting of the candidate requirement set based on generation and processing enables the system to be more flexible in the description of supply resources and can provide multiple candidate texts that meet the requirements according to different input features (such as fusion features and auxiliary features). This diversity provides a rich basis for the screening of the second type of supply service resource samples, and the optimal matching supply resources can be more accurately screened according to further similarity calculations, further improving the overall performance of the system.

[0082] S105. Extract the third feature of each candidate enterprise requirement description text in the candidate requirement set based on the text representation model, and screen the first type of supply service resource samples based on each of the second features and each of the third features to determine the second type of supply service resource samples.

[0083] Specifically, in combination with the above description, the third feature includes a third fusion feature and multiple third auxiliary features. The third fusion feature is the fusion feature of each candidate enterprise requirement description text in the candidate requirement set, and each third auxiliary feature is the feature of each candidate enterprise requirement description text in the candidate requirement set.

[0084] In specific implementation, using a text representation model, multiple feature vectors are extracted from each candidate enterprise demand description text in the candidate demand set at the first depth, and the average value of these feature vectors is calculated and determined as the third fusion feature. Further, using the text representation model again, multiple more fine-grained feature vectors are extracted from each candidate enterprise demand description text in the candidate demand set at the second depth to obtain multiple third auxiliary features. Calculate the third similarity between the third fusion feature of each candidate enterprise demand description text and the second fusion feature of the demand description text; determine the corresponding theme based on the description type corresponding to each third auxiliary feature, calculate the fourth auxiliary similarity between each third auxiliary feature and the second auxiliary feature of the corresponding theme, and weight the fourth auxiliary similarity based on the third weight corresponding to each theme to obtain the fourth similarity between each candidate enterprise demand description text and the demand description text; correct the third similarity based on the fourth similarity; sort each supply service resource sample in the first type of supply service resource sample in ascending order according to the magnitude of the corrected third similarity, and determine the supply service resource samples ranked in the top preset number as the second type of supply service resource samples.

[0085] S106. Determine the optimal supply service resource based on the matching degree between the second type of supply service resource samples and the demand description text.

[0086] In specific implementation, the determining the optimal supply service resource based on the matching degree between the second type of supply service resource samples and the demand description text includes: the large language model determines the supply service resource and the corresponding first confidence level based on the first fusion feature of each supply service resource sample and the second fusion feature of the demand description text; determine each first fusion feature and the first auxiliary feature in the second type of supply service resource samples, and establish a matching relationship between the description type and the theme; the large language model calculates the second auxiliary confidence level between the corresponding first auxiliary feature and the second auxiliary feature based on the matching relationship; determine the fourth weight based on the influence degree of the semantic of each text segment corresponding to each theme in the demand description text on the comprehensive semantic of the demand description text; weight the second auxiliary confidence level based on the fourth weight to obtain the second confidence level between the second type of supply service resource samples and the demand description text; correct the first confidence level based on the second confidence level; determine the optimal supply service resource according to the magnitude of the corrected first confidence level.

[0087] Specifically, first, extract the first fusion feature of each supply service resource sample from the second type of supply service resource samples and compare it with the second fusion feature of the demand description text. Process the input first fusion feature and second fusion feature through the large language model, return the matching situation between each supply service resource sample and the demand description text, and assign a first confidence level to each sample.

[0088] Next, for each of the multiple first auxiliary features in the supply service resource sample, a matching relationship between the description type and the topic is established based on the description type corresponding to each first auxiliary feature and the relationship between the description type and the topic in the demand description text. For example, suppose a supply service resource sample contains descriptions of "data processing capability" and "service response time", which correspond to different description types. A matching relationship is established between each description type and a topic in the demand description text (such as "technical requirements" and "time requirements"). This process matches the first auxiliary feature with the relevant part (second auxiliary feature) in the demand description text through the training of a large language model, and generates the second auxiliary confidence between each pair of corresponding features. On this basis, the semantic influence of each topic is identified through in-depth analysis of the demand description text (through keyword extraction and topic modeling). Assuming that the demand description text contains multiple topics, such as "functional requirements" and "time requirements", after analysis, it is found that "time requirements" has a greater semantic influence in the entire text, then the semantic influence of this topic will be given a higher weight. This weight is the fourth weight, which is used to adjust the second auxiliary confidence. The fourth weight is combined with the second auxiliary confidence, and through weighted calculation, the second confidence between each supply service resource sample and the demand description text is obtained. By correcting the first confidence between the second type of supply service resource sample and the demand description text, the corrected first confidence can be obtained. The correction process adds the second confidence to the first confidence, and adjusts the weights of the two through the algorithm model generated by the large language model to form a final confidence value. This value reflects the matching degree of each supply service resource sample and the demand description text.

[0089] Finally, all the supply service resource samples are sorted in ascending order according to the corrected first confidence, and the samples in the front are regarded as the optimal supply service resource samples, and the first second preset number of supply service resource samples are selected from them. Finally, the sorted and screened supply service resource samples are determined to be the optimal supply service resources, which can meet the requirements in the demand description text.

[0090] Optionally, the fourth weight is determined based on the degree of influence of the semantics of the text fragment corresponding to each topic in the requirement description text on the comprehensive semantics of the requirement description text, including: extracting the comprehensive feature representation of the requirement description text and the feature representation of the text fragment corresponding to each topic based on the text representation model; calculating the similarity between the feature representation of the text fragment corresponding to each topic and the comprehensive feature representation of the requirement description text using a semantic similarity calculation model, and determining the similarity as the initial weight; adjusting the initial weight based on the semantic correlation between each topic and other topics, and determining the adjusted weight as the weight corresponding to each topic.

[0091] In specific implementation, first, a text representation model is used to process the requirement description text and extract its comprehensive feature representation, that is, the semantic vector of the whole text. For example, assume the requirement description text is "We need an expert with strong data analysis capabilities who can provide a data report within two weeks and train the team." In this process, the text representation model converts this text into a vector representing its overall semantics. Next, the text fragments corresponding to each theme in the requirement description text are separated. For example, for the above text, it can be divided into two theme fragments: "an expert with strong data analysis capabilities" and "able to provide a data report within two weeks and train the team". The same text representation model is used to process each fragment and extract its semantic vector representation. On this basis, a semantic similarity calculation model (such as cosine similarity or Euclidean distance) is used to calculate the similarity between the feature representation of the text fragment corresponding to each theme and the comprehensive feature representation of the requirement description text. For example, assume the similarity between the semantic vector of the first fragment and the semantic vector of the whole requirement description text is 0.85, and the similarity of the second fragment is 0.90. At this time, the obtained similarity values will be used as the initial weights of each theme fragment, indicating the strength of the correlation between each theme fragment and the overall text semantics. In this way, the initially obtained weights can reflect the importance of each theme in the requirement description. To further optimize the weights, the semantic relevance between each theme is then considered. Assume there is a certain overlap in semantics between the first theme fragment "an expert with strong data analysis capabilities" and the second theme fragment "able to provide a data report within two weeks and train the team", especially both are related to "data analysis capabilities" and "training capabilities". For this reason, the semantic similarity between each theme fragment can be calculated again to quantify their degree of association. For example, the semantic similarity between the two fragments is 0.75, indicating a strong association between them. Based on the above analysis, a weight adjustment function (such as a weight factor based on semantic association) is set to adjust the initial weights. If the semantic association degree between two theme fragments is relatively high, their weights can be appropriately increased, indicating that these fragments have a greater impact on the overall semantics of the requirement description text. For example, if the initial weights are 0.85 and 0.90, in the case of strong semantic association, the adjusted weights may be 0.95 and 1.00 respectively. Finally, the adjusted weights will be used as the weights indicating the degree of influence of the text fragment corresponding to each theme on the comprehensive semantics of the requirement description text.

[0092] Corresponding to the foregoing embodiment of a method for matching supply and demand of supply service resources based on text features, the present application also provides an embodiment of a device for matching supply and demand of supply service resources based on text features.

[0093] Figure 2This is a schematic structural diagram of the supply service resource supply-demand matching device based on text features provided in the second embodiment of this application. Please refer to Figure 2 The device provided in this embodiment includes an acquisition module 210, an extraction module 220, a generation module 230, and a determination module 240;

[0094] Among them, the acquisition module 210 is used to obtain multiple topics from the demand description text of supply service resources and match the description types of supply service resources;

[0095] The acquisition module 210 is further used to obtain the description data corresponding to each description type, combine the multiple description data of each supply service resource, and use it as a sample to obtain multiple supply service resource samples;

[0096] The extraction module 220 is used to extract the first feature of each supply service resource sample and the second feature of the demand description text based on a text representation model, screen the multiple supply service resource samples based on the similarity between the first feature and the second feature, and determine the first type of supply service resource samples; the text representation model is arranged in a computer;

[0097] The generation module 230 is used to generate a candidate demand set according to each supply service resource sample in the first type of supply service resource samples;

[0098] The determination module 240 is used to extract the third feature of each candidate enterprise demand description text in the candidate demand set based on a text representation model, screen the first type of supply service resource samples based on each of the second features and each of the third features, and determine the second type of supply service resource samples;

[0099] The determination module 240 is further used to determine the optimal supply service resource based on the matching degree between the second type of supply service resource samples and the demand description text.

[0100] The device of this embodiment can be used to execute Figure 1 the steps of the method embodiment shown, and the specific implementation principle and process are similar, which will not be elaborated here.

[0101] For the specific implementation process of the functions and roles of each unit in the above device, please refer to the implementation process of the corresponding steps in the above method, which will not be elaborated here.

[0102] For the apparatus embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the descriptions of the method embodiments. The apparatus embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this application. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0103] The above are only the preferred embodiments of this application and are not intended to limit this application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of this application shall be included within the scope of protection of this application.

Claims

1. A supply service resource supply-demand matching method based on text features, characterized in that The method includes: Obtaining multiple topics from the demand description text of the supply service resources, and matching the description types of the supply service resources; Obtaining the description data corresponding to each description type, combining the multiple description data of each supply service resource as a sample, and obtaining multiple supply service resource samples; Extracting the first features of each of the supply service resource samples and the second features of the demand description text based on a text representation model, screening the multiple supply service resource samples based on the similarity between the first features and the second features, and determining the first type of supply service resource samples; the text representation model is arranged in a computer; Generating a candidate demand set according to each supply service resource sample in the first type of supply service resource samples; Extracting the third features of each candidate enterprise demand description text in the candidate demand set based on the text representation model, screening the first type of supply service resource samples based on each of the second features and each of the third features, and determining the second type of supply service resource samples; Determining the optimal supply service resource based on the matching degree between the second type of supply service resource samples and the demand description text.

2. The method according to claim 1, wherein The first features include a first fusion feature and multiple first auxiliary features corresponding to each description data; The second features include a second fusion feature and multiple second auxiliary features corresponding to each topic.

3. The method according to claim 2, wherein The extracting the first features of each of the supply service resource samples based on the text representation model includes: Constructing a cross-feature constraint relationship based on the logical relevance between the supply service resource data of different description types; Determining the first weights of the supply service resource data of different description types based on the matching degree between the supply service resource data of different description types and the demand description text; Extracting the features of the supply service resource data of different description types based on the text representation model; Performing weighted fusion on the features based on the first weights to obtain a first fusion feature; Adjusting the first fusion feature based on the cross-feature constraint relationship to obtain an adjusted first fusion feature.

4. The method according to claim 2, wherein The screening the multiple supply service resource samples based on the similarity between the first features and the second features and determining the first type of supply service resource samples includes: Calculating a first similarity between each first fusion feature and the second fusion feature; Establishing a matching relationship between the description type and the topic; Calculating a first auxiliary similarity between the first auxiliary feature and the second auxiliary feature according to the matching relationship; Determining a second weight based on the influence degree of the text segment semantics corresponding to each topic in the demand description text on the comprehensive semantics of the demand description text; Performing weighting on the first auxiliary similarity based on the second weight to obtain a second similarity between the first auxiliary feature of each supply service resource sample and the second auxiliary feature of the demand description text; Correcting the first similarity based on the second similarity; Screening the first type of supply service resource samples according to the magnitudes of the corrected first similarities.

5. The method according to claim 1, wherein The generating a candidate demand set according to each supply service resource sample in the first type of supply service resource samples includes: Determine the first prompt word of the large language model based on each supply service resource sample in the first type of supply service resource samples and the feature extraction requirements; The large language model generates multiple initial segments based on the first prompt word and the first fusion features of each supply service resource sample; The large language model generates multiple modification segments based on the first prompt word and the multiple first auxiliary features of each supply service resource sample; Match the corresponding multiple target modification segments based on the initial segments, and process the initial segments based on the multiple target modification segments to generate multiple processed candidate requirement texts, obtaining a candidate requirement set.

6. The method according to claim 1, wherein The extracting of the first features of each of the supply service resource samples based on the text representation model includes: Extract multiple features at the first depth of the supply service resource data of different description types in each supply service resource sample based on the text representation model; Extract multiple features at the second depth of the supply service resource data of different description types in each supply service resource sample based on the text representation model, and determine them as multiple first auxiliary features; the second depth is greater than the first depth.

7. The method according to claim 2, characterized in that, The determining of the optimal supply service resource based on the matching degree between the second type of supply service resource samples and the requirement description text includes: The large language model determines the supply service resource and the corresponding first confidence level based on the first fusion feature of each supply service resource sample and the second fusion feature of the requirement description text; Determine each first fusion feature and first auxiliary feature in the second type of supply service resource samples, and establish a matching relationship between the description type and the theme; The large language model calculates the second auxiliary confidence level between the corresponding first auxiliary feature and the second auxiliary feature based on the matching relationship; Determine the fourth weight based on the influence degree of the semantic of the text segment corresponding to each theme in the requirement description text on the comprehensive semantic of the requirement description text; Weight the second auxiliary confidence level based on the fourth weight to obtain the second confidence level between the second type of supply service resource samples and the requirement description text; Correct the first confidence level based on the second confidence level; Determine the optimal supply service resource according to the magnitude of the corrected first confidence level.

8. The method according to claim 7, characterized in that, The determining of the fourth weight based on the influence degree of the semantic of the text segment corresponding to each theme in the requirement description text on the comprehensive semantic of the requirement description text includes: Extract the comprehensive feature representation of the requirement description text and the feature representation of the text segment corresponding to each theme based on the text representation model; Calculate the similarity between the feature representation of the text segment corresponding to each theme and the comprehensive feature representation of the requirement description text using the semantic similarity calculation model, and determine the similarity as the initial weight; Adjust the initial weight based on the semantic relevance between each theme and other themes, and determine the adjusted weight as the weight corresponding to each theme.

9. The method according to claim 1, characterized in that, Obtain multiple themes from the requirement description text of the supply service resource and match the description type of the supply service resource, including: Perform semantic segmentation on the requirement description text of the supply service resource to obtain multiple themes; Obtain the supply service history of each service object, and determine the supply service resource usage habits based on the supply service history; Predict the constraint conditions of the supply service resources based on the service content and the supply service resource usage habits; Match the description types based on the constraint conditions.

10. A supply service resource supply-demand matching device based on text features, characterized in that, The device includes an acquisition module, an extraction module, a generation module, and a determination module; Among them, the acquisition module is used to obtain multiple topics from the demand description text of the supply service resources and match the description types of the supply service resources; The acquisition module is further used to obtain the description data corresponding to each description type, and combine the multiple description data of each supply service resource as a sample to obtain multiple supply service resource samples; The extraction module is used to extract the first feature of each supply service resource sample and the second feature of the demand description text based on the text representation model, and screen the multiple supply service resource samples based on the similarity between the first feature and the second feature to determine the first type of supply service resource samples; the text representation model is arranged in a computer; The generation module is used to generate a candidate demand set according to each supply service resource sample in the first type of supply service resource samples; The determination module is used to extract the third feature of each candidate enterprise demand description text in the candidate demand set based on the text representation model, and screen the first type of supply service resource samples based on each of the second features and each of the third features to determine the second type of supply service resource samples; The determination module is further used to determine the optimal supply service resource based on the matching degree between the second type of supply service resource samples and the demand description text.

Images