CN118484527A - A data processing method and system based on multi-source heterogeneous matching - Google Patents

Abstract

A data processing method and system based on multi-source heterogeneous matching relates to the field of data processing, and the method comprises the following steps: the method comprises the steps of carrying out text splitting processing on demand information of product research and development in a target company to obtain research and development demand label information, acquiring expert label information in an expert database, patent label information in a patent database and paper label information in a paper database, respectively calculating first similarity of the research and development demand label information and the patent label information and second similarity of optimal patent label information and paper label information, selecting the patent and paper label information with the largest similarity as an optimal matching result, and then selecting the patent, paper and expert with the largest similarity as the optimal matching result, so that a user can acquire the most relevant information of the research and development demand in three dimensions of the patent, paper and expert at the same time, comprehensive inquiry and global decision of the user are facilitated, and user experience is improved.

Description

A data processing method and system based on multi-source heterogeneous matching

Technical Field

The present application relates to the field of data processing, and in particular to a data processing method and system based on multi-source heterogeneous matching.

Background Art

With the rapid development of science and technology, the number of patent applications has increased year by year, and competition in various industries has become increasingly fierce. Enterprises must not only quickly discover and utilize relevant patented technologies, but also learn from cutting-edge academic achievements and cooperate with outstanding experts and scholars in the industry.

Existing information service systems usually manage and retrieve different types of data such as patents, papers, and experts separately. In terms of patents, the main search method is based on keyword matching, and the keyword index is used to achieve rapid search and relevance sorting of patent texts. In terms of papers, it mainly relies on academic literature databases to support conditional search of structured information such as paper titles, abstracts, and keywords. In terms of experts, it mainly browses and queries through manually compiled expert catalogs and profiles. However, these methods are independent of each other and lack fusion and association, and cannot achieve cross-domain and cross-resource comprehensive search and association recommendations. This makes it difficult for users to discover the internal connections between patents, papers, and experts from massive heterogeneous data and obtain targeted information services, which reduces the user experience.

Summary of the invention

The present application provides a data processing method and system based on multi-source heterogeneous matching, which is used to obtain R&D demand label information by performing text splitting processing on product R&D demand information within a target company, and then obtain expert label information in an expert database, patent label information in a patent database, and paper label information in a paper database. By respectively calculating the first similarity between the R&D demand label information and the patent label information, and the second similarity between the optimal patent label information and the paper label information, the patent and paper label information with the greatest similarity are selected as the optimal matching result, and the semantic association between label information from different sources is fully utilized. The third similarity between the expert patent label of each expert and the optimal patent label, and the fourth similarity between the expert paper label and the optimal paper label are calculated, and the third and fourth similarities are weighted averaged to select the patent, paper, and expert with the greatest similarity as the optimal matching result, so that users can simultaneously obtain the most relevant information in the three dimensions of patents, papers, and experts for a certain R&D demand, which facilitates users' comprehensive query and global decision-making and improves user experience.

In a first aspect, the present application provides a data processing method based on multi-source heterogeneous matching, which is applied to a data processing system based on multi-source heterogeneous matching. The method includes:

Perform text splitting processing on the demand information of product development within the target company to obtain the research and development demand label information of the product development within the target company;

Obtain expert label information in an expert database, patent label information in a patent database, and paper label information in a paper database, wherein the expert database contains the expert label information of each expert and the expert patent label information and expert paper label information corresponding to the expert;

Calculate the first similarity between the R&D demand label information and the patent label information respectively, and select the optimal patent label information corresponding to the maximum first similarity;

Calculate the second similarity between the optimal patent label information and the paper label information respectively, and select the optimal paper label information corresponding to the maximum second similarity;

Calculate the third similarity between each expert's expert patent label information and the optimal patent label information and the fourth similarity between the expert's paper label information and the optimal paper label information respectively, and take the expert with the highest weighted average similarity as the optimal target expert, where the weighted average similarity is the weighted average of the third similarity and the fourth similarity;

The optimal expert label information, the optimal patent label information and the optimal paper label information corresponding to the optimal target expert are sent to the client.

In the above embodiment, by performing text splitting processing on the demand information of product R&D within the target company, R&D demand label information is obtained, and then the expert label information in the expert database, the patent label information in the patent database and the paper label information in the paper database are obtained. By respectively calculating the first similarity between the R&D demand label information and the patent label information, and the second similarity between the optimal patent label information and the paper label information, the patent and paper label information with the greatest similarity are selected as the optimal matching result, and the semantic association between label information from different sources is fully utilized. The third similarity between the expert patent label of each expert and the optimal patent label, and the fourth similarity between the expert paper label and the optimal paper label are calculated, and the third and fourth similarities are weighted averaged to select the patent, paper and expert with the greatest similarity as the optimal matching result, so that users can simultaneously obtain the most relevant information in the three dimensions of patents, papers and experts for a certain R&D demand, which facilitates users' comprehensive query and global decision-making.

In conjunction with some embodiments of the first aspect, in some embodiments, the step of respectively calculating the first similarity between the R&D demand label information and the patent label information and selecting the optimal patent label information corresponding to the maximum first similarity specifically includes:

Perform text preprocessing on the R&D demand label information to obtain a set of R&D demand keywords;

Perform text preprocessing on the patent tag information to obtain a patent keyword set;

Calculate the first similarity of the R&D demand keyword set and the patent keyword set respectively to obtain a first similarity collection;

The first similarity collection is sorted from large to small according to the first similarity value, and the patent label information with the largest first similarity value is used as the optimal patent label information.

In the above embodiment, text preprocessing is performed on the R&D demand label information and the patent label information respectively, and the respective keyword sets are extracted, and then the similarity of the two keyword sets is calculated to obtain a similarity collection. By sorting the similarity collection from large to small according to the similarity value, the patent label information with the greatest similarity is selected as the optimal matching result, making full use of text mining and similarity calculation technology, and realizing accurate matching of R&D demand and patents by refining keywords and comparing their similarities. At the same time, the most similar patents are sorted and screened, providing high-quality data support for subsequent patent analysis and utilization.

In combination with some embodiments of the first aspect, in some embodiments, the step of respectively calculating the third similarity between the expert patent label information of each expert and the optimal patent label information and the fourth similarity between the expert paper label information and the optimal paper label information, taking the expert with the highest weighted average similarity as the optimal target expert, and the weighted average similarity being the weighted average of the third similarity and the fourth similarity, specifically includes:

The optimal patent label information, the optimal paper label information, the expert patent label information of each expert, and the expert paper label information are respectively vectorized to obtain the optimal patent feature vector, the optimal paper feature vector, the expert patent feature vector, and the expert paper feature vector;

Calculate the third similarity between the optimal patent feature vector and each of the expert patent feature vectors respectively to obtain a third similarity set;

Calculate the fourth similarity between the optimal paper feature vector and each of the expert paper feature vectors to obtain a fourth similarity set;

Performing weighted summation on the third similarity and the fourth similarity corresponding to each expert to obtain a weighted average similarity set of each expert, wherein the weighted similarity set includes a plurality of weighted average similarities, and the weighted average similarity is a weighted average of the third similarity and the fourth similarity;

The expert with the highest weighted average similarity in the weighted average similarity set is determined as the optimal target expert.

In the above embodiment, the best patent label, the best paper label, and the expert patent label and expert paper label of each expert are vectorized to obtain their respective feature vectors. Then, the similarity between the best patent feature vector and each expert patent feature vector, and the similarity between the best paper feature vector and each expert paper feature vector are calculated respectively, and then the two similarities of each expert are weighted and summed, and the expert with the highest weighted average similarity is selected as the optimal target expert. The similarity of the two dimensions of patents and papers is integrated and assigned a certain weight, so that the comprehensive matching degree of experts can be evaluated more comprehensively and balancedly, and the rationality and credibility of expert recommendations are improved.

In combination with some embodiments of the first aspect, in some embodiments, after the step of sending the optimal expert label information, the optimal patent label information, and the optimal paper label information corresponding to the optimal target expert to the client, the method further includes:

Receive the requirement document information sent by the client;

Perform text mining on the requirement document information to obtain technical keywords, product feature keywords, and application scenario keywords;

Searching the technical keyword, the product feature keyword, and the application scenario keyword in the patent database to obtain a patent document collection, wherein the patent document collection includes no less than one patent document;

Calculate the risk value according to the patent document keywords of each patent document and the information of the demand document to obtain a risk patent collection, wherein the risk patent collection contains at least one patent document with a risk greater than a preset risk threshold;

The risk patent collection is sent to the client.

In the above embodiment, the client uploads the demand document, performs text mining to extract technical keywords, product feature keywords and application scenario keywords, uses these keywords to search in the patent database, obtains a batch of relevant patent documents, and then calculates the risk value of each patent based on the similarity between the keywords and the demand document, screens out high-risk patents to form a risk patent collection and feeds back to the customer. By deeply mining the multi-dimensional key information of the demand document and conducting patent search and risk assessment based on this, high-risk patents related to the demand can be found more accurately and comprehensively, so that customers can have a clear understanding of the potential infringement risks, provide reference and guidance for the formulation and implementation of their R&D strategies, and reduce the legal risks of enterprises.

In combination with some embodiments of the first aspect, in some embodiments, after the step of sending the risky patent collection to the client, the method further includes:

The candidate patent documents that are not greater than the preset risk threshold are stored in the preset patent database.

In the Alternative Patents column.

In the above embodiment, while identifying high-risk patents, the solution also stores other patent documents whose risk values do not exceed the preset threshold in the alternative patent column. Although these alternative patents have lower risks, they are also relevant to the demand and still have reference value for the company's R&D work. By establishing an alternative patent library, companies can avoid risks while learning and drawing on the relevant technical content in these patents in a targeted manner to guide and optimize their own R&D directions and technical solutions.

In conjunction with some embodiments of the first aspect, in some embodiments, after the step of obtaining expert label information in the expert database, patent label information in the patent database, and paper label information in the paper database, the following steps are specifically included:

Obtaining a user satisfaction score fed back by the client, the user satisfaction score including a first satisfaction score for the optimal expert label information, a second satisfaction score for the optimal patent label information, and a third satisfaction score for the optimal paper label information;

If the user satisfaction score is lower than the preset score threshold, the feedback information sent by the client is obtained;

Extracting label filtering conditions in the feedback information based on a large language model;

The initial expert label information in the expert database is screened according to the label screening condition to obtain the expert label information, the initial patent label information in the patent database is screened to obtain the patent label information, and the initial paper label information in the paper database is screened to obtain the paper label information.

In the above embodiment, by obtaining the user's satisfaction score for the recommendation results, it is determined whether there is a score below the threshold. If the satisfaction is low, the user's feedback is further obtained, and the label screening conditions are extracted from it using the large language model. The initial labels of experts, patents, and papers are screened accordingly to obtain more refined label information that is closer to user needs. The label screening strategy is dynamically adjusted according to the user's subjective evaluation and opinions to make the label more in line with the user's actual needs and preferences. At the same time, the key information in the user feedback is intelligently extracted through the large language model, which can effectively improve the personalized service capabilities, continuously optimize and improve the knowledge matching effect, and enhance user stickiness.

In conjunction with some embodiments of the first aspect, in some embodiments, after the step of obtaining a user satisfaction score fed back by the client, the user satisfaction score including a first satisfaction score for the optimal expert label information, a second satisfaction score for the optimal patent label information, and a third satisfaction score for the optimal paper label information, the method further includes:

If there is no user satisfaction score lower than the preset score threshold, the detailed introduction information of the optimal target expert, the patent details corresponding to the optimal patent label information, and the paper details corresponding to the optimal paper label information are pushed to the client.

In the above embodiment, by fully considering another situation of user satisfaction, that is, all scores reach the preset threshold, indicating that the user is relatively satisfied with the current recommendation results, more detailed information is further pushed to the user, including a detailed introduction of the best experts, detailed content of the best patents and papers, etc. By providing these additional detailed information, users can have a more comprehensive and in-depth understanding of the acquired knowledge resources, which helps them better understand and utilize the recommendation results and improve the overall service experience.

In a second aspect, an embodiment of the present application provides a data processing system based on multi-source heterogeneous matching, the data processing system based on multi-source heterogeneous matching comprising: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, the one or more processors call the computer instructions to enable the data processing system based on multi-source heterogeneous matching to perform the method described in the first aspect and any possible implementation manner of the first aspect.

In a third aspect, an embodiment of the present application provides a computer program product comprising instructions. When the above-mentioned computer program product is run on a data processing system based on multi-source heterogeneous matching, the above-mentioned data processing system based on multi-source heterogeneous matching executes the method described in the first aspect and any possible implementation method of the first aspect.

In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, comprising instructions. When the above instructions are executed on a data processing system based on multi-source heterogeneous matching, the above data processing system based on multi-source heterogeneous matching executes the method described in the first aspect and any possible implementation method of the first aspect.

It can be understood that the data processing system based on multi-source heterogeneous matching provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the method provided in the embodiment of the present application. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding method, which will not be repeated here.

One or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

1. This application obtains R&D demand label information by performing text splitting processing on the product R&D demand information within the target company, and then obtains expert label information in the expert database, patent label information in the patent database, and paper label information in the paper database. By respectively calculating the first similarity between the R&D demand label information and the patent label information, and the second similarity between the optimal patent label information and the paper label information, the patent and paper label information with the greatest similarity are selected as the optimal matching results. The semantic association between label information from different sources is fully utilized to calculate the third similarity between each expert's expert patent label and the optimal patent label, and the fourth similarity between the expert paper label and the optimal paper label. The third and fourth similarities are weighted averaged, and the patents, papers, and experts with the greatest similarity are selected as the optimal matching results, so that users can simultaneously obtain the most relevant information in the three dimensions of patents, papers, and experts for a certain R&D demand, which facilitates users' comprehensive query and global decision-making and improves user experience.

2. This application is to have the client upload the demand document, perform text mining on it to extract technical keywords, product feature keywords and application scenario keywords, use these keywords to search in the patent database, obtain a batch of relevant patent documents, and then calculate the risk value of each patent based on the similarity between the keywords and the demand document, screen out high-risk patents to form a risk patent collection and feedback to the customer. By deeply mining the multi-dimensional key information of the demand document and conducting patent search and risk assessment based on this, high-risk patents related to the demand can be found more accurately and comprehensively, so that customers can have a clear understanding of the potential infringement risks, provide reference and guidance for the formulation and implementation of their R&D strategies, and reduce the legal risks of enterprises.

3. This application obtains the user's satisfaction score for the recommendation results to determine whether there is a score below the threshold. If the satisfaction is low, it will further obtain user feedback, use the large language model to extract label screening conditions, and filter the initial labels of experts, patents, and papers accordingly to obtain more refined label information that is closer to user needs. The label screening strategy is dynamically adjusted according to the user's subjective evaluation and opinions to make the label more in line with the user's actual needs and preferences. At the same time, the large language model is used to intelligently extract key information from user feedback, which can effectively improve the system's personalized service capabilities, continuously optimize and improve knowledge matching effects, and enhance user stickiness.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a data processing method based on multi-source heterogeneous matching in an embodiment of the present application;

FIG2 is another flow chart of a data processing method based on multi-source heterogeneous matching in an embodiment of the present application;

FIG3 is another flow chart of a data processing method based on multi-source heterogeneous matching in an embodiment of the present application;

FIG4 is another flow chart of a data processing method based on multi-source heterogeneous matching in an embodiment of the present application;

FIG. 5 is a schematic diagram of a physical device structure of a data processing system based on multi-source heterogeneous matching in an embodiment of the present application.

DETAILED DESCRIPTION

The terms used in the following embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to be used as limitations to the present application. As used in the specification of the present application, the singular expressions "a", "a kind", "above", "the" and "this" are intended to also include plural expressions, unless there is a clear indication to the contrary in the context. It should also be understood that the term "and/or" used in the present application refers to any or all possible combinations comprising one or more of the listed items.

In the following, the terms "first" and "second" are used for descriptive purposes only and should not be understood as suggesting or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, unless otherwise specified, the meaning of "plurality" is two or more.

For ease of understanding, the application scenarios of the embodiments of the present application are introduced below.

In related technologies, keyword retrieval and manual screening can be used to find the required technical information from databases such as patents and papers. The following describes the scenarios in which the data processing method based on multi-source heterogeneous matching in related technologies is used:

Company A, an electronic equipment manufacturer, has been committed to developing a new generation of smart watch products in recent years. In order to improve the innovation and competitiveness of its products, Company A urgently needs to understand the latest technological trends and research results in this field, and hopes to find relevant patents, papers and industry experts to provide guidance and support for its R&D work. Company A tried to use the existing information service system to solve the above problems. They first used the patent search system to find relevant patents by entering keywords. However, since the technical points involved in smart watches are relatively scattered, simple keyword matching often misses some important but differently expressed patents. At the same time, the search results are also mixed with a large number of irrelevant or low-quality patents, which require manual screening and judgment. These limitations have prevented Company A from efficiently and coherently using various information resources, greatly affecting the progress and quality of its R&D innovation.

The data processing method based on multi-source heterogeneous matching in the embodiment of the present application is adopted to extract structured label information from patent libraries, paper libraries and expert libraries, and calculate its semantic similarity with R&D needs to achieve cross-domain and cross-object correlation matching, which not only greatly improves the efficiency of information retrieval and screening, but also can discover implicit knowledge that is expressed differently but is inherently related. The following introduces the scenarios in which the data processing method based on multi-source heterogeneous matching in the present application is used:

Company A decided to adopt a data processing solution based on multi-source heterogeneous matching to solve the above problems. First, they extracted the key technical requirements of the smart watch R&D project and converted them into structured label information. Then, the corresponding patent labels, paper labels and expert labels were automatically extracted from the patent library, paper library and expert library. By calculating the semantic similarity between the demand label and the patent label and paper label, the most relevant patents and papers were quickly found. Then, the correlation between the expert and his paper and patent was comprehensively considered, and the target expert who best matched the demand was calculated by weighted summation. Finally, the best matched patent, paper and expert information was fed back to the R&D team of Company A. Through this solution, R&D personnel can quickly obtain the most critical and authoritative technical reference materials and fully understand the principles, applications and expert opinions of related technologies. The high correlation and aggregation of this information enables R&D personnel to have a clear understanding of the technical context and development trend of smart watches, greatly improving the pertinence and foresight of R&D. At the same time, through the intelligent matching of industry experts, Company A has also established a deep industry-university-research cooperation relationship, providing intellectual support for subsequent continuous innovation.

It can be seen that the data processing method based on multi-source heterogeneous matching in the embodiment of the present application can not only realize the rapid acquisition of relevant knowledge resources across fields and objects, but also effectively solve the problems of information fragmentation and low correlation existing in traditional keyword retrieval and manual screening modes, thereby realizing the deep integration and efficient utilization of knowledge resources, and providing all-round support for the research and development innovation of enterprises.

For ease of understanding, the following describes the process of the method provided by this embodiment in combination with the above scenario. Please refer to Figure 1, which is a flow chart of a data processing method based on multi-source heterogeneous matching in an embodiment of the present application.

S101. Perform text splitting processing on the demand information of product development within the target company to obtain research and development demand label information of product development within the target company.

Among them, the target company refers to an enterprise that needs to conduct product research and development and seek relevant knowledge resource support. Product research and development demand information refers to the technical requirements proposed by the target company for a specific product research and development project, which is usually presented in text form and includes descriptions of research and development goals, key technical points, application scenarios, etc.; text splitting processing refers to the process of converting unstructured demand text into structured information units using natural language processing technology; research and development demand label information represents keywords or phrases extracted from the demand text, which are used to characterize the core content elements of the research and development demand, facilitating subsequent label matching.

This step is usually performed when the target company proposes new product development requirements and is the starting point of the entire multi-source heterogeneous matching process. Specifically, the target company is first required to provide a detailed description of the product development requirements in the form of a document. Then, the document is preprocessed, such as word segmentation, part-of-speech tagging, named entity recognition, etc. On this basis, text mining technologies such as keyword extraction and topic models are used to automatically identify keywords, phrases or entities that can represent core R&D requirements and store them in the form of tags.

S102, obtaining expert label information in an expert database, patent label information in a patent database, and paper label information in a paper database, wherein the expert database contains the expert label information of each expert and the expert patent label information and expert paper label information corresponding to the expert.

Among them, the expert database refers to a structured data set that stores information about experts in various fields, including the basic attributes, research directions, and achievements of experts; the patent database refers to a data set that includes patent documents and their metadata. The paper database refers to a data set that includes academic papers and their metadata; expert label information refers to a keyword description of the expert's research field, technical direction, and other characteristics; patent label information and paper label information refer to keyword descriptions of the core content elements of patents and papers, respectively; expert patent label information and expert paper label information refer to the label information corresponding to each expert's patents and papers.

This step is to prepare data for subsequent multi-source heterogeneous matching after obtaining the R&D demand labels. Specifically, first connect the databases of experts, patents, and papers respectively, and extract the structured representation of each type of data according to the pre-defined database schema. For expert data, it is necessary to obtain the expert's basic information, research field, main achievements and other attribute values, as well as label words that reflect their research characteristics. For patent and paper data, it is necessary to obtain metadata information such as the title, abstract, keywords, classification number, etc. of each record, and extract the label words that best represent the core content. At the same time, it is also necessary to obtain the corresponding relationship data between each expert and the patents and papers he published, in order to establish a mapping between experts and their achievements.

In some embodiments, expert label information in the expert database, patent label information in the patent database, and paper label information in the paper database can be obtained in a variety of ways: Optionally, for the patent database, the text content of key parts such as the title, abstract, and claims of the patent can be obtained, and then a text mining method similar to the demand label extraction can be used to identify keywords that can represent the core technical points of the patent to form patent labels; Optionally, with experts as the center, various types of attribute information are labeled separately, such as matching the predefined domain vocabulary according to the research field name to generate domain labels, generating research direction labels according to paper keywords, etc., and finally summarizing the labels of each dimension to form complete expert label information. It is understandable that other knowledge engineering and natural language processing technologies such as ontology construction and semantic analysis can also be used to achieve labeled representation and organization of expert, patent, and paper data, which is not limited here.

S103, respectively calculating the first similarity between the R&D demand label information and the patent label information, and selecting the optimal patent label information corresponding to the maximum first similarity.

Among them, the first similarity indicates the semantic similarity between the R&D demand label and the patent label, which is used to measure the matching degree between the two in terms of technical themes, functional applications, etc. The best patent label information refers to those patent labels with the highest first similarity to the R&D demand label, reflecting the patent content that best meets the R&D demand.

This step is performed after obtaining the R&D demand labels and patent labels, with the aim of finding the patent knowledge most relevant to the R&D demand. Specifically, it is necessary to first select a suitable similarity calculation method, such as cosine similarity, Jaccard similarity, etc., and represent the label in the form of word frequency vector, TF-IDF vector, etc. Then, using this similarity method, calculate the similarity score of each R&D demand label and each patent label respectively to obtain a similarity matrix. Next, for each R&D demand label, select the Top-N patent labels with the highest similarity as the candidate optimal patent labels. Finally, the optimal patent labels of all R&D demand labels are combined, and the optimal matching patent label set for the entire R&D demand is decided through weighted average, voting and other strategies.

In some embodiments, the similarity calculation between the R&D demand label and the patent label and the selection of the optimal patent label can be achieved in a variety of ways: Optionally, the label word is embedded with a pre-trained word vector model such as Word2Vec, GloVe, etc., and then the average value of the label word vector is calculated as the embedding vector of the entire label, and then the similarity of the two labels is calculated by vector dot product, cosine similarity and other methods; Optionally, a semantic relationship network between label words is constructed based on an external knowledge base such as WordNet, and the semantic similarity between labels is calculated by the shortest path on the graph, random walk and other methods. When selecting the optimal patent label, you can first sort based on similarity, set a similarity threshold, filter out patent labels with low similarity, and then select the top N from the remaining patent labels in descending order of similarity as the optimal matching result. It is understandable that other semantic matching and sorting algorithms such as deep learning and topic models can also be used to achieve similarity calculation and optimization between R&D demand labels and patent labels, which are not limited here.

S104, respectively calculating the second similarity between the optimal patent label information and the paper label information, and selecting the optimal paper label information corresponding to the maximum second similarity.

Among them, the second similarity indicates the semantic similarity between the best patent label and the paper label, which is used to measure the consistency between the two in terms of technical themes, innovations, etc. The best paper label information refers to those paper labels with the highest second similarity to the best patent label, reflecting the academic research results that best fit the patent content.

This step is performed after obtaining the optimal patent label, with the aim of finding the scientific research paper knowledge that is most relevant to the patent technology. Specifically, a calculation method similar to the first similarity is used to calculate the second similarity between each optimal patent label and each paper label, and a new similarity matrix is obtained. Then, for each optimal patent label, the top-M paper labels with the highest second similarity are selected as candidate optimal paper labels. Finally, the optimal paper labels of all optimal patent labels are combined, and the optimal matching paper label set for the entire R&D needs is decided through weighted averaging, voting and other strategies.

S105. Calculate the third similarity between each expert's expert patent label information and the optimal patent label information and the fourth similarity between the expert's paper label information and the optimal paper label information respectively, and take the expert with the highest weighted average similarity as the optimal target expert. The weighted average similarity is the weighted average of the third similarity and the fourth similarity.

Among them, expert patent label information refers to the label information set of patents published by each expert, reflecting the patent research and development characteristics of the expert. Expert paper label information refers to the label information set of papers published by each expert, reflecting the academic research characteristics of the expert. The third similarity is used to measure the degree of match between the expert's patent label and the optimal patent label in terms of technical themes, innovations, etc. The fourth similarity is used to measure the consistency between the expert's paper label and the optimal paper label in terms of research fields, key technologies, etc. The optimal target expert refers to the expert who combines the characteristics of both patents and papers and has the highest match with R&D needs. The weighted average similarity refers to the weighted average of the third similarity and the fourth similarity, which is used to balance the contribution of the two types of labels, patents and papers, to the expert matching degree.

In some embodiments, the similarity calculation of expert patent labels, expert paper labels and optimal labels and the selection of the optimal target expert can be realized in a variety of ways: Optionally, the same method as the first and second similarity calculations is used to calculate the third and fourth similarities, such as word vector similarity, topic similarity, etc., and then the weighted average is used to obtain the comprehensive matching degree according to the preset weights such as the ratio of 0.6 for patents and 0.4 for papers; Optionally, the patent labels and paper labels are respectively embedded in Doc2Vec documents, and then an end-to-end expert matching model, such as a dual-tower DNN model, is jointly trained to learn the matching relationship between expert features and demand features, and finally the optimal expert is obtained by sorting the matching probability values output by the model. When determining the optimal target expert, the Top-1 expert with the highest weighted average similarity (or matching probability) can be directly selected; or the Top-K experts can be taken, and then combined with other factors such as the influence of the experts, willingness to cooperate, etc., to determine the final optimal expert through a multi-index decision-making method. It is understandable that other algorithms such as collaborative filtering and representation learning based on graph networks can also be used to achieve the matching calculation and optimization of experts and needs, which are not limited here.

S106. Send the optimal expert label information, the optimal patent label information and the optimal paper label information corresponding to the optimal target expert to the client.

The client refers to the terminal device used by enterprise users to submit R&D requirements and receive matching results, such as PC, mobile APP, etc. The optimal expert label information represents the characteristic labels such as research fields and technical directions related to the optimal target expert.

This step is performed after the optimal target expert and its related patents and papers are determined, with the purpose of feeding back the matching results to the enterprise user. Specifically, the background encapsulates the information of the optimal target expert, including basic information such as name, unit, and feature tags that match R&D needs, such as the optimal patent tag, the optimal paper tag, etc., into a structured data object. At the same time, the literature information of related patents and papers, such as titles, keywords, abstracts, etc., can also be used as a supplementary description of the matching results. Then, these result data are sent to the client of the enterprise user through the network, and the client renders the matching results to the user according to the predefined visualization template.

In some embodiments, the label information corresponding to the optimal target expert can be sent to the client in a variety of ways: Optionally, static web pages and documents containing matching results are pre-generated on the server side, such as HTML, PDF and other formats, and then the web page URL or document download link is sent to the client, and the client accesses the online web page or downloads the document for viewing. In the interface design of the client, a variety of layout methods such as lists, cards, and graphs can be used, and visual elements such as expert portraits and technology roadmaps can be provided to facilitate enterprise users to browse and understand. It is understandable that other information interaction methods such as mobile push and email sending can also be used to achieve the distribution of matching results to enterprise users, which is not limited here.

The following is a more detailed description of the process of the method provided by this embodiment. Please refer to Figure 2, which is another flowchart of the data processing method based on multi-source heterogeneous matching in the embodiment of this application.

S201. Perform text splitting processing on the demand information of product development within the target company to obtain research and development demand label information of product development within the target company.

It can be understood that this step is similar to step S101 and will not be described in detail here.

S202, obtaining expert label information in an expert database, patent label information in a patent database, and paper label information in a paper database, wherein the expert database contains the expert label information of each expert and the expert patent label information and expert paper label information corresponding to the expert.

It can be understood that this step is similar to step S102 and will not be described in detail here.

S203: Perform text preprocessing on the R&D demand tag information to obtain a set of R&D demand keywords.

Specifically, the text preprocessing of R&D demand label information can be divided into the following steps: first, text cleaning, removing meaningless characters, punctuation marks, HTML tags and other noise data in the R&D demand description, and unifying the text format; second, word segmentation and part-of-speech tagging, using natural language processing tools to divide the R&D demand text into independent words, and mark the part of speech of each word, such as noun, verb, adjective, etc.; third, stop word filtering, removing common stop words in the R&D demand text, such as "的", "是", "了", etc., to improve the accuracy of keyword extraction; finally, stem extraction and word form restoration, restoring the words in the R&D demand text to their original form, eliminating the impact of word form changes on keyword extraction.

For example, if the label information of a research and development requirement is "We need to develop a new type of positive electrode material for lithium batteries, hoping to improve the energy density and cycle life of the material", after text preprocessing, it is converted into the following structured information: "lithium battery/n positive electrode/n material/n new/a development/v improvement/v energy density/n cycle life/n", and keywords such as "lithium battery", "positive electrode material", "energy density", and "cycle life" are extracted to construct a standardized set of research and development requirement keywords for subsequent analysis and application.

S204: Perform text preprocessing on the patent tag information to obtain a patent keyword set.

Similar to R&D demand tag information, the original patent tag information also needs to be preprocessed to extract key words and build a standardized patent keyword set. Since the patent text content is usually more professional and standardized, the preprocessing process of patent tag information is slightly different from that of R&D demand tag information, and needs to be optimized and adjusted in a targeted manner.

Specifically, the text preprocessing of patent label information can be divided into the following steps: first, patent field parsing, extracting the contents of key fields such as invention name, abstract, claims, and instructions from the original patent text, and storing them in a structured manner; second, professional vocabulary recognition, using the terminology dictionary and named entity recognition tools in the patent field, identifying technical keywords, product names, company organizations and other professional vocabulary from the patent text; third, keyword sorting, calculating the importance weight of each vocabulary based on the frequency, position, co-occurrence relationship and other characteristics of the professional vocabulary in the patent text, and arranging them in descending order according to the weight; finally, keyword indexing, referring to standards such as patent classification numbers and field vocabularies, semantically indexing the sorted professional vocabulary, mapping them to standardized technical keywords, and constructing a standardized patent keyword set.

For example, if the label information of a patent is "a positive electrode material for lithium-ion batteries and its preparation method", it contains multiple fields such as patent name, abstract, and claims. After text preprocessing, key words such as "positive electrode material for lithium-ion batteries", "precursor", "coating", "doping", and "sintering" are extracted from the patent abstract, and professional terms such as "spinel lithium manganese oxide", "spherical particles", "coating layer", "doping elements", and "sintering temperature" are extracted from the claims. The weights are calculated based on the word frequency and co-occurrence relationship, and the Top 10 are selected as the keyword labels of the patent, mapped to the standard technical keyword vocabulary, such as "lithium-ion battery", "positive electrode material", "spinel lithium manganese oxide", "coating modification", "element doping", "sintering process", etc., to construct a standardized patent keyword set.

S205 , respectively calculating the first similarities of the R&D demand keyword set and the patent keyword set to obtain a first similarity set.

Specifically, a variety of similarity calculation methods can be used, such as cosine similarity, Jaccard similarity, edit distance, etc., and the most suitable algorithm can be selected according to the characteristics of the keyword set.

Taking cosine similarity as an example, first, we need to express both the R&D demand keyword set and the patent keyword set in the form of vectors. Each keyword corresponds to a dimension in the vector, and the importance weight of the keyword is used as the component value of the vector. Then, calculate the cosine value of the angle between the two vectors as the similarity score between them. The smaller the angle, the closer the cosine value is to 1, indicating that the two keyword sets are more similar; conversely, the larger the angle, the closer the cosine value is to 0, indicating that the two keyword sets are less similar.

For example, suppose the keyword set of a research and development requirement is {"lithium battery": 0.6, "positive electrode material": 0.8, "energy density": 0.5, "cycle life": 0.4}, and the keyword set of a patent is {"lithium-ion battery": 0.7, "positive electrode material": 0.9, "spinel lithium manganese oxide": 0.6, "coating modification": 0.5}. These two keyword sets are represented as vectors, and their cosine similarity is calculated, and the similarity score is 0.85.

Repeat the above process to calculate the first similarity between the R&D demand keyword set and each patent keyword set, and finally obtain a first similarity collection, which reflects the correlation between the R&D demand and each patent.

S206: Sort the first similarity collection from large to small according to the first similarity values, and use the patent label information with the largest first similarity value as the optimal patent label information.

After obtaining the first similarity collection between the R&D requirements and each patent, it is necessary to sort and filter these similarity values in order to quickly find the patent label information that is most relevant and matches the R&D requirements. Specifically, all patents in the first similarity collection can be sorted in descending order according to the first similarity value from large to small, with the patent with the highest similarity at the front and the patent with the lowest similarity at the end.

For example, suppose the first similarity values between a research and development requirement and patents A, B, C, D, and E are 0.85, 0.76, 0.92, 0.68, and 0.81, respectively. The first similarity collection is {(A, 0.85), (B, 0.76), (C, 0.92), (D, 0.68), (E, 0.81)}. This collection is sorted in descending order according to the first similarity value, and the sorted result is {(C, 0.92), (A, 0.85), (E, 0.81), (B, 0.76), (D, 0.68)}. It can be seen that patent C has the highest correlation with the research and development requirement, and the first similarity value reaches 0.92, which is much higher than other patents. Therefore, the label information of patent C can be used as the optimal patent label information and recommended to R&D personnel.

S207, respectively calculating the second similarity between the optimal patent label information and the paper label information, and selecting the optimal paper label information corresponding to the maximum second similarity.

It can be understood that this step is similar to step S105 and will not be described in detail here.

S208. Vectorize the optimal patent label information, the optimal paper label information, the expert patent label information of each expert, and the expert paper label information to obtain the optimal patent feature vector, the optimal paper feature vector, the expert patent feature vector, and the expert paper feature vector.

Specifically, the optimal patent label information, the optimal paper label information, the patent label information and the paper label information of each expert can all be represented as feature vectors in a high-dimensional vector space.

This vectorized representation usually uses methods such as One-Hot encoding or TF-IDF weighting. Taking One-Hot encoding as an example, you first need to build a dictionary containing all tag keywords. Each keyword in the dictionary corresponds to a dimension in the vector space. Then, for each tag information, if it contains a keyword, assign a value of 1 to the corresponding dimension, otherwise assign a value of 0. Finally, each tag information is represented as a high-dimensional sparse vector composed of 0 and 1, and the dimension of the vector is equal to the number of keywords in the dictionary.

For example, suppose the dictionary contains five keywords ["lithium battery", "positive electrode material", "solid electrolyte", "energy density", "cycle life"], the label information of the best patent is "lithium battery, positive electrode material, energy density", the label information of the best paper is "solid electrolyte, lithium battery, cycle life", the patent label information of expert A is "lithium battery, positive electrode material", and the paper label information of expert A is "positive electrode material, solid electrolyte, cycle life". After One-Hot encoding, the following feature vectors are obtained: best patent feature vector: [1, 1, 0, 1, 0], best paper feature vector: [1, 0, 1, 0, 1], expert A patent feature vector: [1, 1, 0, 0, 0], expert A paper feature vector: [0, 1, 1, 0, 1].

It can be seen that these feature vectors depict the similarities and differences between different tag information in an intuitive way, providing a mathematical basis for subsequent similarity calculations and expert recommendations. Of course, in addition to One-Hot encoding, more complex vectorization methods such as TF-IDF weighting can also be used, which is not limited here.

S209, respectively calculating the third similarity between the optimal patent feature vector and each of the expert patent feature vectors to obtain a third similarity set.

After obtaining the optimal patent feature vector and each expert patent feature vector, the matching degree between each expert and the optimal patent can be evaluated by calculating the similarity between them.

Taking cosine similarity as an example, it measures the similarity between two vectors by calculating the cosine value of the angle between them. The smaller the angle, the closer the cosine value is to 1, indicating that the two vectors are more similar; conversely, the larger the angle, the closer the cosine value is to 0, indicating that the two vectors are less similar. The specific calculation formula is as follows: cos(θ) = (A·B) / (||A|| ||B||);

Among them, A and B are two eigenvectors, A·B represents the dot product of the vectors, and ||A|| and ||B|| represent the modulus of the vector (i.e., the square root of the sum of the squares of the components of each dimension of the vector).

For example, assuming that the optimal patent feature vector is [1, 1, 0, 1, 0], the patent feature vector of expert A is [1, 1, 0, 0, 0], the patent feature vector of expert B is [0, 1, 1, 0, 1], and the patent feature vector of expert C is [1, 0, 0, 1, 1]. Calculate the cosine similarity between the optimal patent feature vector and the patent feature vectors of these three experts respectively, and get:

The third similarity with expert A: cos(θ_A)=(11+11+00+10+0*0)/(sqrt(3)*sqrt(2)) =0.82, the third similarity with expert B: cos(θ_B)=(10+11+01+10+0*1)/(sqrt(3)*sqrt(3))=0.33, the third similarity with expert C: cos(θ_C)=(11+10+00+11+0*1)/(sqrt(3)*sqrt(3))=0.67.

It can be seen that the cosine similarity between the patent feature vector of expert A and the optimal patent feature vector is the highest, which is 0.82, indicating that the patent research direction of expert A is closest to the optimal patent and he may be an authoritative expert in this field. The cosine similarity between the patent feature vector of expert C and the optimal patent feature vector is second, which is 0.67, and also has a certain correlation. The cosine similarity between the patent feature vector of expert B and the optimal patent feature vector is the lowest, which is only 0.33, indicating that the patent research direction of expert B is quite different from the optimal patent and the matching degree is low.

S210, respectively calculating the fourth similarity between the optimal paper feature vector and each of the expert paper feature vectors to obtain a fourth similarity set.

Specifically, the same method as the third similarity calculation, such as cosine similarity, can be used to measure the cosine value of the angle between the optimal paper feature vector and each expert paper feature vector. The smaller the angle, the closer the cosine value is to 1, indicating that the two vectors are more similar and the expert's paper research direction is closer to the optimal paper; conversely, the larger the angle, the closer the cosine value is to 0, indicating that the two vectors are less similar and the expert's paper research direction is more different from the optimal paper.

For example, suppose the optimal paper feature vector is [1, 0, 1, 0, 1], the paper feature vector of expert A is [0, 1, 1, 0, 1], the paper feature vector of expert B is [1, 1, 0, 1, 0], and the paper feature vector of expert C is [0, 0, 1, 1, 1]. Calculate the cosine similarity between the optimal paper feature vector and the feature vectors of these three experts' papers, and you get:

The fourth similarity with expert A: cos(θ_A)=(10+01+11+00+1*1)/(sqrt(3)*sqrt(3))=0.67, the fourth similarity with expert B: cos(θ_B)=(11+01+10+01+1*0)/(sqrt(3)*sqrt(3))=0.33, the fourth similarity with expert C: cos(θ_C)=(10+00+11+01+1*1)/(sqrt(3)*sqrt(3))=0.67;

It can be seen that the cosine similarity between the paper feature vectors of experts A and C and the feature vector of the best paper is 0.67, indicating that their paper research directions are close to the best paper and have a certain correlation. However, the cosine similarity between the paper feature vector of expert B and the feature vector of the best paper is lower, only 0.33, indicating that the research direction of expert B's paper is quite different from that of the best paper.

In this way, the fourth similarity between the optimal paper feature vector and each expert paper feature vector can be calculated to obtain a fourth similarity set, such as {(A, 0.67), (B, 0.33), (C, 0.67)}.

S211. Perform weighted summation on the third similarity and the fourth similarity corresponding to each expert to obtain a weighted average similarity set for each expert, wherein the weighted similarity set includes a plurality of weighted average similarities, and the weighted average similarity is a weighted average of the third similarity and the fourth similarity.

After obtaining the third similarity (matching degree with the best patent) and the fourth similarity (matching degree with the best paper) of each expert, it is necessary to consider these two similarities comprehensively to comprehensively evaluate the overall relevance of each expert to the R&D needs. A common method is to perform a weighted summation of the third similarity and the fourth similarity to obtain a weighted average similarity as the final score of the expert matching degree.

Specifically, the weight coefficients of the third and fourth similarities can be reasonably set according to the focus of R&D needs. For example, if the R&D needs place more emphasis on the innovation and practicality of patented technology, a higher weight can be given to the third similarity; if the R&D needs place more emphasis on theoretical basis and academic value, a higher weight can be given to the fourth similarity. Then, the third and fourth similarities of each expert are weighted and summed to obtain the weighted average similarity of the expert.

For example, suppose the weight of the third similarity is 0.6 and the weight of the fourth similarity is 0.4. For expert A, the third similarity is 0.82 and the fourth similarity is 0.67; for expert B, the third similarity is 0.33 and the fourth similarity is also 0.33; for expert C, the third similarity is 0.67 and the fourth similarity is 0.67. According to the weighted summation formula, the weighted average similarity of each expert is calculated:

The weighted average similarity of expert A = 0.6*0.82+0.4*0.67=0.76, the weighted average similarity of expert B = 0.6*0.33+0.4*0.33=0.33, and the weighted average similarity of expert C = 0.6*0.67+0.4*0.67=0.67.

It can be seen that expert A has the highest weighted average similarity of 0.76, which means that considering both patents and papers, expert A has the highest overall match with R&D needs. Expert C has a weighted average similarity of 0.67, which also has a high match. Expert B has the lowest weighted average similarity of only 0.33, which means that its overall correlation with R&D needs is weak.

S212: Determine the expert with the highest weighted average similarity in the weighted average similarity set as the optimal target expert.

Specifically, we only need to sort all the experts in the weighted average similarity set from high to low according to the weighted average similarity value, and select the expert with the highest similarity as the optimal target expert. If there are multiple experts with the same weighted average similarity and they are all the highest values, we can prioritize them according to the specific situation and consider other factors, such as the academic reputation and cooperation history of the experts, or recommend multiple optimal target experts for R&D personnel to choose from.

S213, sending the optimal expert label information, the optimal patent label information and the optimal paper label information corresponding to the optimal target expert to the client.

It can be understood that this step is similar to step S106 and will not be described in detail here.

The following is a more detailed description of the process of the method provided by this embodiment. Please refer to Figure 3, which is another flowchart of the data processing method based on multi-source heterogeneous matching in the embodiment of this application.

S301, sending the optimal expert label information, the optimal patent label information and the optimal paper label information corresponding to the optimal target expert to the client.

It can be understood that this step is similar to step S106 and will not be described in detail here.

S302: Receive the demand document information sent by the client.

When enterprise users complete the editing and input of R&D requirements on the client side, they can submit the formed requirements document to the server side. Usually, the requirements document exists in the form of a structured electronic document, and the document content usually includes a background description of the R&D project, a description of technical requirements, a list of product features, an analysis of application scenarios, and other aspects.

S303: Perform text mining on the requirement document information to obtain technical keywords, product feature keywords, and application scenario keywords.

After obtaining the R&D requirements documents submitted by enterprise users, it is necessary to automatically extract keyword information that accurately describes the requirements from the unstructured text description to provide data support for subsequent matching of requirements and resources. Text mining is an analytical method that uses natural language processing, machine learning and other technologies to discover implicit knowledge and patterns from large-scale text data. In this scenario, the goal of text mining is to identify three types of keywords that can reflect the core semantics of the requirements document: technical keywords, product feature keywords, and application scenario keywords. Among them, technical keywords reflect the key technical points involved in the requirements, such as algorithms, materials, processes, etc.; product feature keywords reflect the requirements for product functions, performance, appearance and other characteristics; application scenario keywords describe the product's target application areas, usage environment, user groups, etc.

S304: Search the technical keyword, the product feature keyword and the application scenario keyword in the patent database to obtain a patent document collection, wherein the patent document collection contains no less than one patent document.

After obtaining the keyword representation of the R&D requirements document, it is necessary to perform information retrieval in the patent database to discover existing patent technologies related to the requirements. The patent database is a structured database that contains a large number of full-text patent documents and metadata, covering patent information in various technical fields and regions. In this step, technical keywords, product feature keywords, and application scenario keywords are used as search terms, and full-text searches are performed in the title, abstract, claims, instructions and other fields of the patent database.

Specifically, classic information retrieval models such as Boolean retrieval and vector space retrieval can be used to combine keywords into complex query statements through logical operators such as AND, OR, NOT, etc., and then quickly locate patent documents containing the search terms with the support of data structures such as inverted indexes. Considering that matching only a single keyword may not fully reflect the relevance of patents and needs, recall rate optimization technologies such as term expansion, synonym expansion, and hyponym expansion can also be used to expand the semantic scope of the original search terms to match more potential related patents.

S305. Calculate the risk value according to the patent document keywords of each patent document and the demand document information to obtain a risk patent collection, wherein the risk patent collection contains at least one patent document whose number is greater than a preset risk threshold.

After generating a collection of patent documents, it is necessary to further evaluate the risk level of each patent document to R&D needs in order to identify high-risk patents that may lead to intellectual property infringement or hinder project progress. The core of risk assessment is to calculate the similarity between patent documents and demand documents. The higher the similarity, the greater the overlap between the patent-protected technical solution and the R&D content, and the higher the potential risk.

Specifically, for each patent in the patent document collection, its key technical information is first extracted to form a structured patent element representation, such as the purpose of the invention, technical field, key issues, solutions, claims, etc. The extraction method can use natural language processing technologies such as rule template matching and conditional random field sequence labeling, or it can use the domain knowledge of patent review to manually index the elements. Then, the patent element representation is converted into a vector form, usually a high-dimensional sparse vector, such as a TF-IDF weight vector, a topic distribution vector, etc. Similarly, the R&D requirements document can also be extracted and represented by a vector.

After obtaining the vector representation of patents and requirements, vector distance measurement methods such as cosine similarity and KL divergence are used to calculate the similarity score between each patent and the requirement as the risk value of the patent. The higher the risk value, the closer the technical solution of the patent is to the requirement, and the greater the potential obstacle to R&D posed by the patent layout. By setting an empirical risk threshold, such as 0.75, a subset of patents with higher risk values can be quickly filtered out to form a collection of risky patents. These patents are likely to cover the key technical points of R&D requirements, which need to attract the attention and avoidance of enterprises.

S306: Send the risk patent collection to the client.

After identifying a collection of risky patents, information on these high-risk patents needs to be fed back to corporate users in a timely manner to support them in making intellectual property decisions and R&D plans.

Specifically, a structured patent summary report is automatically generated for each patent document in the risk patent collection. The report is based on the key paragraphs and claims in the patent specification, extracting and condensing the core content elements of the patent, such as technical field, problem background, invention purpose, technical solution, claim scope, etc. It can also include metadata such as the patent's bibliographic information, legal status, citation relationship, etc. The patent summary report is generated in rich text formats such as HTML and PDF, and supports rich visual elements such as mixed text and image layout, table display, link jump, etc.

After generating the patent summary report, all reports are packaged and compressed and sent as attachments to designated corporate users via email, instant messaging, etc. At the same time, risk patent information can also be displayed in a more interactive way on the client's graphical interface. For example, the titles and summaries of risk patents are listed in a tree structure or list on the left side of the interface. After clicking on a patent, the corresponding patent summary report pops up on the right for users to review. In addition, the interface can also provide auxiliary functions such as patent sorting, screening, annotation, and sharing to facilitate users to manage and collaborate on risk patents.

S307. Store the candidate patent documents that are not greater than the preset risk threshold into the candidate patent column in the preset patent database.

In addition to identifying patent documents with higher risks for R&D needs, some patent documents with lower risks but related to the required technical direction may also be found during the patent search process. Although such patents do not currently pose a significant risk, they still have certain reference value for enterprises to understand the latest technological trends in the industry, inspire R&D ideas, and avoid patent layout. Therefore, these alternative patent documents are stored separately for subsequent review and analysis by enterprise users.

Specifically, first, patents with risk values lower than a preset threshold (such as 0.75) are screened out from the patent document collection to form an alternative patent subset. Since the correlation between these patents and demand is relatively weak, the number may be large, and the information noise is relatively large if the full amount is directly presented to the user. Therefore, it is also necessary to organize and manage the alternative patent subset to improve its searchability and readability.

In some embodiments, the alternative patent subsets can be classified according to technical fields, functional applications and other dimensions to construct a hierarchical classification system. For example, existing technical classification standards such as the IPC International Patent Classification Number and the CPC Cooperative Patent Classification Number can be referred to to automatically classify and index each alternative patent. Afterwards, the alternative patent subsets are written into the preset patent database, and a corresponding alternative patent column is established in the database. The column is organized in the form of a directory tree, with the top level being the major technical fields, each technical field having several subcategories, and each subcategory corresponding to one or more alternative patents.

The following is a more detailed description of the process of the method provided by this embodiment. Please refer to Figure 4, which is another flowchart of the data processing method based on multi-source heterogeneous matching in the embodiment of this application.

S401, performing text splitting processing on the demand information of product development within the target company to obtain research and development demand label information of the product development within the target company.

It can be understood that this step is similar to step S101 and is not limited here.

S402, obtaining a user satisfaction score fed back by the client, wherein the user satisfaction score includes a first satisfaction score for the optimal expert label information, a second satisfaction score for the optimal patent label information, and a third satisfaction score for the optimal paper label information.

After pushing information such as experts, patents, and papers related to R&D needs to enterprise users, it is also necessary to obtain the user's satisfaction with the push results. Satisfaction evaluation is usually conducted in the form of subjective scoring, with users scoring different types of push results on the client interface to express their subjective feelings.

After the user completes a round of satisfaction rating, the rating results uploaded by the client are automatically obtained, which include rating values in three dimensions: expert label information satisfaction, patent label information satisfaction, and paper label information satisfaction. These three rating values are recorded in the background database respectively, and associated with the corresponding enterprise user account, R&D demand number, push batch and other information for subsequent satisfaction analysis.

S403: If the user satisfaction score is lower than a preset score threshold, feedback information sent by the client is obtained.

In addition to obtaining user satisfaction scores, it is also necessary to pay attention to abnormal situations with low scores and actively collect detailed feedback from users in order to improve the recommendation effect in a targeted manner. Since the satisfaction score itself cannot reflect the specific reasons for user dissatisfaction, it is necessary to supplement it with open feedback channels to encourage users to provide more constructive opinions.

Specifically, a satisfaction score threshold is first set, such as 3 points (out of 5 points) or 6 points (out of 10 points), as a standard for judging whether the user satisfaction meets the standard. After obtaining the scoring results submitted by the user, it will automatically determine whether the score value of each dimension is lower than the preset threshold. If the satisfaction score of any of the three types of information, experts, patents, and papers, does not meet the standard, it is considered that the user is dissatisfied with the push result this time, and the reason for his dissatisfaction needs to be further understood.

At this time, a satisfaction reminder will be sent to the client to inform the user that there is a low satisfaction score and ask the user to provide more feedback. After receiving the reminder, the client will automatically pop up a feedback input box in the interface to guide the user to fill in their opinions and suggestions on the push results, such as the expert's research direction is not consistent, the number of recommended patents is too small, and the paper is not innovative enough. After the user enters the feedback and submits it, the client will package the feedback content and send it to the backend.

S404: extracting label screening conditions from the feedback information based on the large language model.

After obtaining user feedback on the recommendation results, it is necessary to automatically understand the feedback content and extract the implicit label filtering conditions for subsequent label information optimization. Since user feedback is usually provided in the form of natural language text, it is necessary to use natural language processing technology, especially large language models, to semantically understand the feedback information and extract key information.

Specifically, the text content of user feedback is first input into a pre-trained large language model, such as BERT, GPT, etc. A large language model is a language representation model trained based on massive text corpus and deep learning algorithms. It can learn the contextual semantics of text and generate high-quality word embedding representations. Through the encoder part of the large language model, the feedback text can be converted into a high-dimensional semantic vector, which contains the key semantic information of the feedback content.

After obtaining the semantic vector of the feedback text, we further use named entity recognition, keyword extraction and other technologies to identify keywords and phrases related to label screening from the semantic vector.

For example, in response to the R&D needs of "new lithium battery separator materials", a user filled in the feedback: "The number of recommended patents is too small, and most of them are in English, which is difficult to understand. I hope to recommend more relevant Chinese patents and provide translations of patent technical key points. In addition, the recommended experts seem to be mostly in the field of polymer materials, which is not very relevant to the field of lithium batteries. I hope to recommend some experts with more research experience in the field of lithium batteries."

The feedback text is input into the pre-trained large language model, and the text is converted into a semantic vector through the encoder part of the model. Then, named entity recognition and keyword extraction technology are used to identify keywords and phrases related to label screening from the semantic vector, such as "Chinese", "patent", "lithium battery", "research experience", etc.

Next, the extracted keywords and phrases are matched with the predefined tag system to identify the corresponding tag type and value. For example, the keyword "Chinese" is mapped to the tag "Patent Language" with the value "Chinese"; the keyword "Lithium Battery" is mapped to the tags "Patent Field" and "Expert Field", both with the value "Lithium Battery"; the keyword "Research Experience" is mapped to the tag "Expert Title", with the value "Professor" or "Researcher" and other senior titles.

After the above steps, a set of structured label filtering conditions were extracted from user feedback, mainly including: {"patent language": "Chinese"}, {"patent field": "lithium battery"}, {"expert field": "lithium battery"}, {"expert title": "professor/researcher"}.

These label screening conditions clearly reflect the specific demands of users for recommendation results, that is, they hope to obtain more Chinese patents and senior expert information related to the field of lithium batteries. These conditions can be used to conduct a secondary screening of the original patent label library and expert label library to obtain more accurately matched recommendation results.

Of course, since the expression of user feedback may not be standardized and complete, the extracted label screening conditions may contain omissions or ambiguities. For example, although the user mentioned "translation of key points of patent technology", it was not clear whether it was machine translation or manual translation. Therefore, after extracting the label conditions, it is still necessary to manually perform necessary verification and completion, such as refining "translation of key points of patent technology" into "patent machine translation: yes" or "patent manual translation: yes" and other clearer label values.

S405. Filter the initial expert label information in the expert database according to the label screening condition to obtain the expert label information, filter the initial patent label information in the patent database to obtain the patent label information, and filter the initial paper label information in the paper database to obtain the paper label information.

After extracting the implicit label screening conditions in user feedback, it is necessary to conduct a secondary screening of the original expert, patent, and paper label information to obtain a more accurate label subset that meets user needs. Through label screening, the most relevant and highest quality labels for user R&D needs can be filtered out from the massive amount of initial label information, improving the pertinence and effectiveness of the recommendation results.

S406. If there is no user satisfaction score lower than the preset score threshold, the detailed introduction information of the optimal target expert, the patent detailed information corresponding to the optimal patent label information, and the paper detailed information corresponding to the optimal paper label information are pushed to the client.

Once it is determined that the user satisfaction scores meet the standards, the detailed information push process will be automatically triggered. The detailed introductions and key contents corresponding to the optimal tag information will be screened out from the three dimensions of experts, patents, and papers, forming a set of personalized and precise knowledge packages, which will be actively pushed to users through the client.

For example, if the user's satisfaction with the previously recommended lithium battery expert A is 4.5 points, the satisfaction with the lithium battery separator patent B is 4.2 points, and the satisfaction with the lithium battery material review paper C is 4.7 points, all of which are higher than the preset 4-point threshold, then in this round of recommendations, the user will automatically be given a detailed introduction to expert A's educational background, research results, contact information, etc., key information such as the abstract, claims, and similar patents of patent B, as well as key content such as the core ideas, research methods, and references of paper C.

The following describes the data processing system based on multi-source heterogeneous matching in the embodiment of the present application from the perspective of hardware processing. Please refer to Figure 5, which is a schematic diagram of the physical device structure of the data processing system based on multi-source heterogeneous matching in the embodiment of the present application.

It should be noted that the structure of the data processing system based on multi-source heterogeneous matching shown in FIG5 is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present invention.

As shown in FIG5 , the data processing system based on multi-source heterogeneous matching includes a central processing unit (CPU) 501, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 502 or the program loaded from the storage part 508 to the random access memory (RAM) 503, such as executing the method described in the above embodiment. Various programs and data required for system operation are also stored in the RAM 503. The CPU 501, the ROM 502, and the RAM 503 are connected to each other through the bus 504. The input/output (I/O) interface 505 is also connected to the bus 504.

The following components are connected to the I/O interface 505: an input section 506 including an audio input device, a button switch, etc.; an output section 507 including a liquid crystal display (LCD) and an audio output device, an indicator light, etc.; a storage section 508 including a hard disk, etc.; and a communication section 509 including a network interface card such as a LAN (Local Area Network) card, a modem, etc. The communication section 509 performs communication processing via a network such as the Internet. A drive 510 is also connected to the I/O interface 505 as needed. A removable medium 511, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 510 as needed so that a computer program read therefrom is installed into the storage section 508 as needed.

In particular, according to an embodiment of the present invention, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present invention includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a computer program for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication part 509, and/or installed from a removable medium 511. When the computer program is executed by a central processing unit (CPU) 501, various functions defined in the present invention are performed.

It should be noted that specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present invention, a computer-readable storage medium may be any tangible medium containing or storing a program that may be used by or in combination with an instruction execution system, apparatus, or device.

The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. Each box in the flowchart or block diagram may represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions marked in the box may also occur in an order different from that marked in the accompanying drawings.

Specifically, the data processing system based on multi-source heterogeneous matching of this embodiment includes a processor and a memory, and a computer program is stored in the memory. When the computer program is executed by the processor, the data processing method based on multi-source heterogeneous matching provided in the above embodiment is implemented.

As another aspect, the present invention further provides a computer-readable storage medium, which may be included in the data processing system based on multi-source heterogeneous matching described in the above embodiment; or may exist independently without being assembled into the data processing system based on multi-source heterogeneous matching. The above storage medium carries one or more computer programs, and when the above one or more computer programs are executed by a processor of the data processing system based on multi-source heterogeneous matching, the data processing system based on multi-source heterogeneous matching implements the data processing method based on multi-source heterogeneous matching provided in the above embodiment.

As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

As used in the above embodiments, the term "when..." may be interpreted to mean "if..." or "after..." or "in response to determining..." or "in response to detecting...", depending on the context. Similarly, the phrases "upon determining..." or "if (the stated condition or event) is detected" may be interpreted to mean "if determining..." or "in response to determining..." or "upon detecting (the stated condition or event)" or "in response to detecting (the stated condition or event)", depending on the context.

A person skilled in the art can understand that to implement all or part of the processes in the above-mentioned embodiments, the processes can be completed by a computer program to instruct the relevant hardware, and the program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above-mentioned method embodiments. The aforementioned storage medium includes: ROM or random access memory RAM, magnetic disk or optical disk and other media that can store program codes.

Claims (10)

1. A data processing method based on multi-source heterogeneous matching, characterized in that it is applied to a patent recommendation system, and the method comprises: Perform text splitting processing on the demand information of product development within the target company to obtain research and development demand label information of product development within the target company; Obtain expert label information in an expert database, patent label information in a patent database, and paper label information in a paper database, wherein the expert database contains the expert label information of each expert and the expert patent label information and expert paper label information corresponding to the expert; Calculate the first similarity between the R&D demand label information and the patent label information respectively, and select the optimal patent label information corresponding to the maximum first similarity; Calculate the second similarity between the optimal patent label information and the paper label information respectively, and select the optimal paper label information corresponding to the maximum second similarity; Calculate the third similarity between the expert patent label information of each expert and the optimal patent label information and the fourth similarity between the expert paper label information and the optimal paper label information respectively, and take the expert with the highest weighted average similarity as the optimal target expert, where the weighted average similarity is the weighted average of the third similarity and the fourth similarity; The optimal expert label information, the optimal patent label information and the optimal paper label information corresponding to the optimal target expert are sent to the client. 2. The method according to claim 1 is characterized in that the step of respectively calculating the first similarity between the R&D demand label information and the patent label information and selecting the optimal patent label information corresponding to the maximum first similarity specifically comprises: Performing text preprocessing on the research and development demand tag information to obtain a research and development demand keyword set; Performing text preprocessing on the patent tag information to obtain a patent keyword set; Calculating first similarities of the R&D demand keyword set and the patent keyword set respectively to obtain a first similarity collection; The first similarity collection is sorted from large to small according to the first similarity value, and the patent label information with the largest first similarity value is used as the optimal patent label information. 3. The method according to claim 1 is characterized in that the steps of respectively calculating the third similarity between the expert patent label information of each expert and the optimal patent label information and the fourth similarity between the expert paper label information and the optimal paper label information, taking the expert with the highest weighted average similarity as the optimal target expert, and the weighted average similarity being the weighted average of the third similarity and the fourth similarity, specifically include: The optimal patent label information, the optimal paper label information, the expert patent label information of each expert, and the expert paper label information are respectively vectorized to obtain the optimal patent feature vector, the optimal paper feature vector, the expert patent feature vector, and the expert paper feature vector; Calculating the third similarity between the optimal patent feature vector and each of the expert patent feature vectors respectively to obtain a third similarity set; Calculating the fourth similarity between the optimal paper feature vector and each of the expert paper feature vectors respectively to obtain a fourth similarity set; Performing weighted summation on the third similarity and the fourth similarity corresponding to each expert to obtain a weighted average similarity set of each expert, wherein the weighted similarity set includes a plurality of weighted average similarities, and the weighted average similarity is a weighted average of the third similarity and the fourth similarity; The expert with the highest weighted average similarity in the weighted average similarity set is determined as the optimal target expert. 4. The method according to claim 1, characterized in that after the step of sending the optimal expert label information, the optimal patent label information and the optimal paper label information corresponding to the optimal target expert to the client, the method further comprises: Receiving the demand document information sent by the client; Performing text mining on the demand document information to obtain technical keywords, product feature keywords, and application scenario keywords; Searching the technical keywords, the product feature keywords, and the application scenario keywords in the patent database to obtain a patent document collection, wherein the patent document collection includes no less than one patent document; Calculate the risk value according to the patent document keywords of each of the patent documents and the demand document information to obtain a risk patent collection, wherein the risk patent collection contains at least one patent document with a value greater than a preset risk threshold; The risk patent collection is sent to the client. 5. The method according to claim 4, characterized in that after the step of sending the risk patent collection to the client, the method further comprises: The alternative patent documents whose risk is not greater than a preset risk threshold are stored in the alternative patent column in the preset patent database. 6. The method according to claim 1 is characterized in that the step of obtaining expert label information in the expert database, patent label information in the patent database, and paper label information in the paper database specifically comprises: Obtaining a user satisfaction score fed back by the client, wherein the user satisfaction score includes a first satisfaction score for the optimal expert label information, a second satisfaction score for the optimal patent label information, and a third satisfaction score for the optimal paper label information; If the user satisfaction score is lower than a preset score threshold, obtaining feedback information sent by the client; Extracting label screening conditions in the feedback information based on a large language model; The initial expert label information in the expert database is screened according to the label screening conditions to obtain the expert label information, the initial patent label information in the patent database is screened to obtain the patent label information, and the initial paper label information in the paper database is screened to obtain the paper label information. 7. The method according to claim 6 is characterized in that after the step of obtaining the user satisfaction score fed back by the client, the user satisfaction score including a first satisfaction score for the optimal expert label information, a second satisfaction score for the optimal patent label information and a third satisfaction score for the optimal paper label information, the method further comprises: If there is no user satisfaction score lower than the preset score threshold, the detailed introduction information of the optimal target expert, the patent details corresponding to the optimal patent label information and the paper details corresponding to the optimal paper label information are pushed to the client. 8. A data processing system based on multi-source heterogeneous matching, characterized in that the data processing system based on multi-source heterogeneous matching comprises: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code comprises computer instructions, and the one or more processors call the computer instructions to enable the data processing system based on multi-source heterogeneous matching to execute the method as described in any one of claims 1-7. 9. A computer-readable storage medium, comprising instructions, characterized in that when the instructions are executed on a data processing system based on multi-source heterogeneous matching, the data processing system based on multi-source heterogeneous matching executes the method as described in any one of claims 1 to 7. 10. A computer program product, characterized in that when the computer program product is run on a data processing system based on multi-source heterogeneous matching, the data processing system based on multi-source heterogeneous matching executes the method according to any one of claims 1 to 7.