Text information recommendation method and device, electronic equipment and storage medium

Abstract

Embodiments of the present application provide a text information recommendation method and device, electronic equipment and storage medium, relating to the technical field of data analysis, the method comprises: obtaining a data table, the data table comprising a patent data table and a user data table; obtaining a heterogeneous graph convolution recommendation model and a natural language model, the heterogeneous graph convolution recommendation model comprising at least a semantic convolution layer and a heterogeneous convolution layer; inputting the patent data table and the user data table into the semantic convolution layer for semantic extraction to obtain a first semantic representation and a second semantic representation; inputting the first semantic representation and the second semantic representation into the heterogeneous convolution layer for keyword recommendation to obtain a first keyword combination; inputting the first keyword combination into the natural language model for keyword recommendation to obtain a second keyword combination; performing text retrieval according to the first keyword combination and the second keyword combination to output corresponding target patent text, improving the comprehensiveness and accuracy of the retrieval.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data analysis, in particular to a text information recommendation method, a text information recommendation device, an electronic device and a computer readable storage medium. BACKGROUND

[0002] With the increasingly fierce competition of world economy and technology, patents have become an important lever to promote the progress of modern society and the development of economic technology as an important form and carrier of scientific and technological innovation. The path for enterprises to obtain innovative technology includes only two kinds of self-research and technology introduction. Whether it is self-research or technology introduction, it is necessary to first obtain the relevant patent achievements of the innovative technology. For example, when choosing the self-research path, patent avoidance design is needed to bypass the protection scope of the relevant patent to prevent infringement. When choosing the technology introduction path, it is necessary to analyze the relevant patents to select the most suitable and advanced patent technology for targeted introduction.

[0003] However, how to obtain enough patent texts related to innovative technology from a large number of patent databases is a relatively complex and professional work, and most small and medium-sized enterprises do not have the relevant ability. In related technologies, the retrieval method of patent texts often has a high missing rate, and it is easy to miss a large number of important and high-technology-similarity patent texts, which makes it difficult to achieve satisfactory patent recommendation effect. SUMMARY

[0004] The embodiments of the present application provide a text information recommendation method, device, electronic device and computer readable storage medium to solve or partially solve the problem of high missing rate in the process of recommending patent texts.

[0005] The embodiments of the present application disclose a text information recommendation method, comprising:

[0006] Obtain a data table, the data table at least including a patent data table and a user data table;

[0007] Obtain a heterogeneous graph convolution recommendation model and a natural language model, the heterogeneous graph convolution recommendation model at least including a semantic convolution layer and a heterogeneous convolution layer;

[0008] Input the patent data table and the user data table into the semantic convolution layer for semantic extraction to obtain a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table;

[0009] Input the first semantic representation and the second semantic representation into the heterogeneous convolution layer for keyword recommendation to obtain at least one first keyword combination corresponding thereto;

[0010] inputting the first keyword combination into the natural language model for keyword recommendation to obtain at least one second keyword combination corresponding thereto;

[0011] performing text retrieval according to the first keyword combination and the second keyword combination to output target patent text corresponding thereto.

[0012] In some possible implementation manners, the data table further comprises a project data table, and the inputting the patent data table and the user data table into the semantic convolution layer for semantic extraction to obtain the first semantic representation corresponding to the patent data table and the second semantic representation corresponding to the user data table comprises:

[0013] extracting employee information from the user data table, and taking the employee information as semantic data;

[0014] extracting project information from the project data table and extracting patent information from the patent data table, and taking the project information and the patent information as auxiliary semantic data corresponding to the semantic data;

[0015] extracting a plurality of patent keywords from the patent information;

[0016] training the semantic data, the auxiliary semantic data, and each of the patent keywords to obtain the first semantic representation corresponding to the patent data table and the second semantic representation corresponding to the user data table.

[0017] In some possible implementation manners, the semantic convolution layer is a graph convolution network, and the graph convolution network at least comprises an input layer, a plurality of execution convolution layers, an activation function, a pooling layer, and an output layer; wherein the graph convolution network further comprises a first user node and a first patent node, and the training the semantic data, the auxiliary semantic data, and each of the patent keywords to obtain the first semantic representation corresponding to the patent data table and the second semantic representation corresponding to the user data table comprises:

[0018] converting the semantic data and the auxiliary semantic data into a first embedding vector;

[0019] converting the plurality of patent keywords into a second embedding vector;

[0020] taking the first embedding vector as node features of the first user node and taking the second embedding vector as node features of the first patent node to obtain a first feature vector of the patent data table and a second feature vector of the user data table;

[0021] The first feature vector and the second feature vector are iteratively updated by the execution of the convolution layer to aggregate information of neighbor nodes, to obtain a node representation fused with neighborhood semantics;

[0022] The node representation fused with neighborhood semantics is nonlinearly transformed according to the activation function, to obtain a transformed node representation;

[0023] The transformed node representation is reduced in dimension by the pooling layer, to obtain a compressed node representation, and the first semantic representation corresponding to the patent data table and the second semantic representation corresponding to the user data table are generated by linear transformation and task adaptation processing of the output layer.

[0024] In some possible implementation manners, the employee information at least includes one of user basic information, professional background information, post data, research direction, and historical achievements; the project information at least includes one of project basic information, technical content, and achievement data; and the patent information at least includes one of a title, an abstract, a text of a claim, a patent type, classification information, a text of a technical feature, conversion data, pledge information, financing information, and text derivative information.

[0025] In some possible implementation manners, the inputting of the first semantic representation and the second semantic representation into the heterogeneous convolution layer to perform keyword recommendation to obtain at least one first keyword combination includes:

[0026] A heterogeneous graph for the first semantic representation and the second semantic representation is constructed, the heterogeneous graph includes several different types of graph nodes, and the different types of graph nodes in the heterogeneous graph are connected through corresponding node relationships;

[0027] Node attributes corresponding to each graph node are obtained;

[0028] Weight values corresponding to each node relationship are obtained;

[0029] The first semantic representation and the second semantic representation are input into the heterogeneous convolution layer, and keyword recommendation is performed according to the node attributes, the node relationships, and the weight values, to obtain at least one first keyword combination.

[0030] In some possible implementation manners, the inputting of the first semantic representation and the second semantic representation into the heterogeneous convolution layer to perform keyword recommendation to obtain at least one first keyword combination includes:

[0031] The weight value and the node relationship are used for path finding in the heterogeneous graph, to obtain a candidate edge sequence corresponding to the first semantic representation and the second semantic representation;

[0032] A weight-sensitive random walk is performed on the candidate edge sequence, to obtain a node sequence for the candidate edge sequence, and a path total weight corresponding to the node sequence is calculated according to the weight value;

[0033] A node sequence with a path total weight greater than or equal to a preset threshold is taken as an initial meta-path;

[0034] The initial meta-path is enhanced using the node attribute, to obtain a semantic-enhanced meta-path;

[0035] The semantic-enhanced meta-path is hierarchically sampled, to obtain a meta-path neighbor dictionary corresponding to each graph node;

[0036] According to the heterogeneous convolution layer, keyword recommendation is performed on the meta-path neighbor dictionary, to obtain at least one first keyword combination corresponding to the first semantic representation and the second semantic representation.

[0037] In some possible implementation manners, the keyword recommendation performed on the meta-path neighbor dictionary according to the heterogeneous convolution layer, to obtain at least one first keyword combination corresponding to the first semantic representation and the second semantic representation, includes:

[0038] A parameter matrix for the meta-path neighbor dictionary is called from the heterogeneous convolution layer;

[0039] The meta-path neighbor dictionary is subjected to feature transformation using the parameter matrix and the weight value, and a weighted aggregation is performed on the transformed output result, to output a group aggregation result corresponding to the meta-path neighbor dictionary;

[0040] The group aggregation result is calculated through the trained attention mechanism parameter, to output a corresponding calculation result, and the calculation result is subjected to feature weighted aggregation and residual connection, to output a first target semantic representation of the first semantic representation and a second target semantic representation corresponding to the second semantic representation;

[0041] A similarity algorithm for the first target semantic representation and the second target semantic representation is obtained;

[0042] The first target semantic representation and the second target semantic representation are subjected to keyword recommendation using the similarity algorithm, to generate at least one first keyword combination.

[0043] The embodiment of the application further discloses a text information recommendation device, which comprises:

[0044] a data acquisition module configured to acquire data tables, the data tables comprising at least a patent data table and a user data table;

[0045] a model acquisition module configured to acquire a heterogeneous graph convolution recommendation model and a natural language model, the heterogeneous graph convolution recommendation model comprising at least a semantic convolution layer and a heterogeneous convolution layer;

[0046] a semantic extraction module configured to input the patent data table and the user data table into the semantic convolution layer to perform semantic extraction, and obtain a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table;

[0047] a first recommendation module configured to input the first semantic representation and the second semantic representation into the heterogeneous convolution layer to perform keyword recommendation, and obtain at least one first keyword combination;

[0048] a second recommendation module configured to input the first keyword combination into the natural language model to perform keyword recommendation, and obtain at least one second keyword combination;

[0049] a text retrieval module configured to perform text retrieval according to the first keyword combination and the second keyword combination, and output a target patent text corresponding thereto.

[0050] In some possible implementation manners, the data tables further comprise a project data table, and the semantic extraction module is specifically configured to:

[0051] extract employee information from the user data table, and take the employee information as semantic data;

[0052] extract project information from the project data table and extract patent information from the patent data table, and take the project information and the patent information as auxiliary semantic data corresponding to the semantic data;

[0053] extract a plurality of patent keywords from the patent information;

[0054] train the semantic data, the auxiliary semantic data and the patent keywords to obtain the first semantic representation corresponding to the patent data table and the second semantic representation corresponding to the user data table.

[0055] In some possible implementation manners, the semantic convolution layer is a graph convolution network, the graph convolution network comprising at least an input layer, a plurality of execution convolution layers, an activation function, a pooling layer and an output layer, and further comprising a first user node and a first patent node, and the semantic extraction module is specifically configured to:

[0056] convert the semantic data and the auxiliary semantic data into a first embedding vector;

[0057] convert the plurality of patent keywords into a second embedding vector;

[0058] obtain a first feature vector of the patent data table and a second feature vector of the user data table by taking the first embedding vector as a node feature of the first user node and taking the second embedding vector as a node feature of the first patent node;

[0059] update the first feature vector and the second feature vector iteratively by aggregating information of neighbor nodes through the execution of the convolution layer, to obtain a node representation fused with neighborhood semantics;

[0060] perform a nonlinear transformation on the node representation fused with neighborhood semantics according to the activation function, to obtain a transformed node representation;

[0061] perform dimension reduction on the transformed node representation through the pooling layer, to obtain a compressed node representation, and generate a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table through linear transformation and task adaptation processing of the output layer.

[0062] In some possible implementation manners, the employee information at least includes one of user basic information, professional background information, post data, research direction, and historical achievements; the project information at least includes one of project basic information, technical content, and achievement data; and the patent information at least includes one of a title, an abstract, a text of a claim, a patent type, classification information, a text of a technical feature, conversion data, pledge information, financing information, and text derivative information.

[0063] In some possible implementation manners, the first recommendation module is specifically configured to:

[0064] construct a heterogeneous graph for the first semantic representation and the second semantic representation, the heterogeneous graph including a plurality of different types of graph nodes, and the different types of graph nodes in the heterogeneous graph being connected through corresponding node relationships;

[0065] obtain node attributes corresponding to each of the graph nodes;

[0066] obtain weight values corresponding to each of the node relationships;

[0067] input the first semantic representation and the second semantic representation into the heterogeneous convolution layer, and perform keyword recommendation according to the node attributes, the node relationships, and the weight values, to obtain at least one first keyword combination.

[0068] In some possible implementation manners, the first recommendation module is specifically configured to:

[0069] perform path finding in the heterogeneous graph by using the weight values and the node relationships, to obtain a candidate edge sequence corresponding to the first semantic representation and the second semantic representation;

[0070] perform weight-sensitive random walk on the candidate edge sequence, to obtain a node sequence corresponding to the candidate edge sequence, and calculate a path total weight corresponding to the node sequence according to the weight values;

[0071] take a node sequence with a path total weight greater than or equal to a preset threshold as an initial meta-path;

[0072] enhance the initial meta-path by using the node attributes, to obtain a semantic-enhanced meta-path;

[0073] perform hierarchical sampling on the semantic-enhanced meta-path, to obtain a meta-path neighbor dictionary corresponding to each graph node;

[0074] perform keyword recommendation on the meta-path neighbor dictionary according to the heterogeneous convolution layer, to obtain at least one first keyword combination corresponding to the first semantic representation and the second semantic representation.

[0075] In some possible implementation manners, the first recommendation module is specifically configured to:

[0076] invoke a parameter matrix for the meta-path neighbor dictionary from the heterogeneous convolution layer;

[0077] perform feature transformation on the meta-path neighbor dictionary by using the parameter matrix and the weight values, and perform weighted aggregation on an output result after the transformation, to output a group aggregation result corresponding to the meta-path neighbor dictionary;

[0078] perform calculation on the group aggregation result by using trained attention mechanism parameters, to output a corresponding calculation result, and perform feature weighted aggregation and residual connection on the calculation result, to output a first target semantic representation of the first semantic representation and a second target semantic representation corresponding to the second semantic representation;

[0079] obtain a similarity algorithm for the first target semantic representation and the second target semantic representation;

[0080] perform keyword recommendation on the first target semantic representation and the second target semantic representation by using the similarity algorithm, to generate at least one first keyword combination corresponding to the first target semantic representation and the second target semantic representation.

[0081] This invention also discloses an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0082] The memory is used to store computer programs;

[0083] When the processor executes a program stored in the memory, it implements the method described in the embodiments of the present invention.

[0084] This invention also discloses a computer-readable storage medium storing instructions that, when executed by one or more processors, cause the processors to perform the methods described in this invention.

[0085] The embodiments of the present invention have the following advantages:

[0086] In this embodiment of the invention, when a user wants to search for relevant text, a data table is obtained, which includes at least a patent data table and a user data table. Then, a heterogeneous graph convolutional recommendation model and a natural language model are obtained. The heterogeneous graph convolutional recommendation model includes at least a semantic convolutional layer and a heterogeneous convolutional layer. The patent data table and the user data table are then input into the semantic convolutional layer for semantic extraction, obtaining a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table. The first and second semantic representations are then input into the heterogeneous convolutional layer for keyword recommendation, obtaining at least one corresponding first keyword combination. The first keyword combination is then input into the natural language model for keyword recommendation, obtaining at least one corresponding second keyword combination. Finally, text retrieval is performed according to the first and second keyword combinations, outputting the corresponding target patent text. Thus, keyword recommendation processing through the semantic convolutional layer and the heterogeneous convolutional layer enriches the information and makes the recommendation results more accurate. Simultaneously, by using the natural language model to generate new keyword combinations based on the recommended keyword combinations, users are provided with more diverse choices, reducing the probability of users getting trapped in information cocoons. Furthermore, text retrieval based on rich keyword combinations effectively ensures the comprehensiveness and accuracy of the retrieval results. Attached Figure Description

[0087] Figure 1 This is a flowchart illustrating the steps of a text information recommendation method provided in an embodiment of the present invention;

[0088] Figure 2 This is a schematic diagram of the recommendation process provided in an embodiment of the present invention;

[0089] Figure 3 This is a schematic diagram of an application scenario provided in an embodiment of the present invention;

[0090] Figure 4 is a structural block diagram of a text information recommendation device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0091] In order to make the above objectives, characteristics and advantages of the present application more apparent, comprehensible and easier to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0092] As an example, how to obtain sufficient patent texts related to innovative technologies from a mass of patent databases is a relatively complex and professional work, and most small and medium-sized enterprises do not have the relevant capabilities. In related technologies, the retrieval method of patent texts often has a high missing rate, and a large amount of important patent texts with high technical similarity are easily missed, and it is difficult to achieve satisfactory patent recommendation effect.

[0093] Reference Figure 1 , a step flowchart of a text information recommendation method provided in an embodiment of the present application is shown, which can specifically include the following steps:

[0094] Step 101, obtaining a data table, the data table at least including a patent data table and a user data table;

[0095] For enterprise employees, especially employees responsible for technology research and development, in the process of researching related technologies, it is inevitable to obtain relevant patent achievements related to the innovative technology. For example, when choosing a self-research and development path, patent avoidance design is needed to bypass the protection range of related patents to prevent infringement. When choosing a technology introduction path, technical analysis of related patents is needed to select the most suitable and advanced patent technology for targeted introduction. Based on this, retrieving comprehensive and highly accurate patent texts can effectively assist research and development personnel in technological innovation.

[0096] In the process of retrieving related patent texts by users (such as research and development personnel, etc.), corresponding data tables can be constructed to extract associated information from the data tables for patent text retrieval. The data tables at least include a patent data table, a user data table, a project data table and an enterprise data table, etc. The patent data table at least includes patent abstracts, patent titles and information of the technical field to which the patents belong corresponding to different patent texts; the user data table can include the professional, post, research direction and other information of the employees in the enterprise; the project data table can include project title, research content, key technology and project difficulty information and other information corresponding to the research and development projects in the enterprise; the enterprise data table can include the main business information and product information of the enterprise, so that by constructing corresponding data tables, a good data basis can be provided for subsequent text retrieval, and the comprehensiveness and accuracy of text retrieval are ensured.

[0097] At step 102, an heterogeneous graph convolution recommendation model and a natural language model are obtained, the heterogeneous graph convolution recommendation model at least including a semantic convolution layer and a heterogeneous convolution layer.

[0098] After the corresponding data table is constructed, the heterogeneous graph convolution recommendation model and the natural language model for text recommendation can also be obtained. The heterogeneous graph convolution recommendation model can be a convolution model for content recommendation based on an heterogeneous graph. The model is designed for an heterogeneous graph (including a graph of multiple types of nodes and edges) based on a graph convolution network. The heterogeneous graph convolution recommendation model can effectively process heterogeneity and retain multi-relation information, and can effectively perform feature fusion and representation learning.

[0099] For example, the nodes and edges in the heterogeneous graph have multiple types, and different types of nodes and edges represent different semantic information, so that the heterogeneous graph convolution model can fully utilize these heterogeneous information and distinguish and process different types of nodes and edges. Meanwhile, there are multiple relationships in the heterogeneous graph, which reflect the interaction and connection between different entities. The heterogeneous graph convolution recommendation model can retain these multi-relation information and consider the influence of different relationships in the information transmission process. In the process of feature fusion, the heterogeneous graph convolution recommendation model can effectively fuse the features of different types of nodes. In the heterogeneous graph, different types of nodes can have features of different dimensions and meanings. The heterogeneous graph convolution recommendation model can integrate these features through convolution operation to generate node representations with better expressiveness. In the process of representation learning, the heterogeneous graph convolution recommendation model can learn low-dimensional vector representations of nodes and edges in the heterogeneous graph. These representations can capture the structure and semantic information of the graph. Through convolution operation on the heterogeneous graph, the model can automatically learn the feature representation of different types of nodes and edges, so that nodes with similar structure and semantics are closer in the vector space. Therefore, the heterogeneous graph convolution recommendation model can effectively perform content recommendation.

[0100] In some possible implementation manners, the heterogeneous graph convolution recommendation model can include a semantic convolution layer and a heterogeneous convolution layer. The semantic convolution layer can be used to extract corresponding semantic representations from the data table and input the semantic representations to the heterogeneous convolution layer for further data processing (such as cooperative processing), so as to make the information more rich and improve the comprehensiveness and accuracy of the recommendation result.

[0101] In addition, for the natural language model, it can be an artificial intelligence model for processing and generating natural language, which learns and models patterns, structures and semantics in natural language, etc., so as to be able to understand, generate and process human language. In the embodiments of the present application, by inputting the corresponding keyword combination into the natural language model, the natural language model can output another keyword combination associated but not identical, thereby providing additional recommended options for the user and providing more diversified choices for the user.

[0102] In step 103, the patent data table and the user data table are input into a semantic convolution layer for semantic extraction to obtain a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table.

[0103] In the embodiments of the present application, after the data table and the heterogeneous convolution recommendation model are determined, the patent data table and the user data table in the data table can be input into the semantic convolution layer for semantic extraction to obtain the first semantic representation corresponding to the patent data table and the second semantic representation corresponding to the user data table. By extracting the corresponding semantic representation, the user information and the patent information can be better understood, so as to recommend more suitable content for the user, thereby improving the accuracy and personalization degree of the recommendation.

[0104] In some possible implementation manners, in the process of extracting the semantic representation, the data table further includes a project data table, the employee information can be extracted from the user data table, the employee information is taken as semantic data, the project information is extracted from the project data table, and the patent information is extracted from the patent data table, the project information and the patent information are taken as auxiliary semantic data corresponding to the semantic data, a plurality of patent keywords are extracted from the patent information, and the semantic data, the auxiliary semantic data and the patent keywords are trained to obtain the first semantic representation corresponding to the patent data table and the second semantic representation corresponding to the user data table.

[0105] The employee information at least includes one of user basic information, professional background information, post data, research direction and historical achievement; the project information at least includes one of project basic information, technical content and achievement data; and the patent information at least includes one of title, abstract, right claim text, patent type, classification information, technical feature text, conversion data, pledge information, financing information and text derivative information, so that more characteristic information is extracted from the data table to be input into the semantic convolution layer for semantic understanding to obtain the corresponding semantic representation, so as to recommend more suitable content for the user, thereby improving the accuracy and personalization degree of the recommendation.

[0106] Optionally, the semantic convolution layer is a graph convolution network, and the graph convolution network at least comprises an input layer, a plurality of convolution layers, an activation function, a pooling layer, and an output layer; wherein the graph convolution network further comprises a first user node and a first patent node; the semantic data and the auxiliary semantic data can be converted into first embedding vectors, and the plurality of patent keywords can be converted into second embedding vectors; the first embedding vectors are taken as node features of the first user node, and the second embedding vectors are taken as node features of the first patent node, to obtain a first feature vector of the patent data table and a second feature vector of the user data table; the information of neighbor nodes is aggregated through the convolution layers to iteratively update the first feature vector and the second feature vector, to obtain node representations fused with neighborhood semantics; the node representations fused with neighborhood semantics are subjected to nonlinear transformation according to the activation function, to obtain transformed node representations; the transformed node representations are subjected to dimension reduction through the pooling layer, to obtain compressed node representations; and the first semantic representation corresponding to the patent data table and the second semantic representation corresponding to the user data table are generated through linear transformation and task adaptation processing of the output layer, so that more characteristic information is extracted from the data table to be input into the semantic convolution layer for semantic understanding, to obtain corresponding semantic representations, so as to recommend more suitable content for the user, and thus improve the accuracy and personalization degree of the recommendation.

[0107] In some examples, the corresponding processing procedure of semantic understanding through the semantic convolution layer can be as follows:

[0108] Firstly, user information is extracted from the user data table as semantic data thereof, and item information, patent information and other information associated with the user are extracted as auxiliary semantic information thereof.

[0109] Then, the summary, title, technical field and other information of a patent are extracted from the patent data table, and a plurality of keywords are extracted from each patent information and marked with a category through an artificial marking assisted supervised learning algorithm such as LDA or TF-IDF. The keywords can be divided into three categories, such as technical keywords, scene keywords and application keywords, which are not limited by the present application.

[0110] For example, there is a patent about smart home, and the summary and title thereof are as follows:

[0111] Title: A smart home control system based on Internet of Things

[0112] Summary: The present application proposes a smart home control system based on Internet of Things, which collects home environment data through sensors and transmits the data to a central controller for processing through a wireless network. The system can automatically adjust the working state of home devices according to user-set rules and home environment data, to realize automatic control of smart home.

[0113] The extracted keywords and their categories can be as follows: technical keywords: Internet of Things, sensor, wireless network, central controller; scenario keywords: smart home; application keywords: automation control, environmental data collection, device state adjustment.

[0114] Then, the user and keyword information can be further trained using BERT or other encoders to obtain initial embedding information corresponding to the first user node u and the first keyword node i;

[0115] The initial embedding information of the first user node u and the first keyword node i is used as node features to further enrich the semantic information of the nodes using a multi-layer graph convolution model. The corresponding process can include:

[0116] Input layer: the feature vector of the node is used as the input of the graph convolution network;

[0117] Convolution layer: apply convolution operation on the graph structure to extract the neighborhood information of the node. The graph convolution network updates the representation of the current node by aggregating the information of the neighbor nodes. This usually involves multiple convolution layers, each of which further integrates the information of the neighbor nodes;

[0118] Activation function: apply a nonlinear activation function (such as ReLU) after convolution operation to increase the nonlinear representation ability of the model;

[0119] Pooling layer: use pooling operation to reduce the size of the graph while preserving important information;

[0120] Output layer: obtain the final semantic representation of each node, i.e., the first semantic representation corresponding to the user data table and the second semantic representation corresponding to the patent data table.

[0121] Step 104, input the first semantic representation and the second semantic representation into the heterogeneous convolution layer for keyword recommendation, and obtain at least one corresponding first keyword combination;

[0122] After obtaining the corresponding semantic representations through the semantic convolution layer, the corresponding first semantic representation and second semantic representation can be input into the heterogeneous convolution layer for keyword recommendation to obtain the corresponding keyword combination, so that the user can perform text retrieval based on the recommended keyword combination to obtain the corresponding retrieval result.

[0123] In some possible implementation manners, in the process of keyword recommendation based on the semantic representations, a heterogeneous graph for the first semantic representation and the second semantic representation can be constructed, the heterogeneous graph includes several different types of graph nodes, and the different types of graph nodes in the heterogeneous graph are connected through corresponding node relationships, then node attributes corresponding to each graph node are obtained, and weight values corresponding to each node relationship are obtained, finally the first semantic representation and the second semantic representation are input into a heterogeneous convolution layer, keyword recommendation is performed according to the node attributes, the node relationships and the weight values, and at least one first keyword combination corresponding to the first semantic representation and the second semantic representation is obtained.

[0124] For the heterogeneous graph, the heterogeneous graph can be composed of four different types of nodes, including a second user node, an item node, a second patent node and a keyword node, and the different types of nodes jointly constitute the heterogeneous graph, and the different types of nodes can be connected with each other through different node relationships, for example, a user and an item, an item and a patent, and a patent and a keyword are connected, these relationships can be constructed based on a project participated by a user, a patent contained by an item and a keyword mentioned in a patent, and the present application does not limit this.

[0125] After the corresponding heterogeneous graph is constructed, weight values and node relationships can be used to find a path in the heterogeneous graph, to obtain a candidate edge sequence corresponding to the first semantic representation and the second semantic representation, then weight-sensitive random walk is performed on the candidate edge sequence to obtain a node sequence corresponding to the candidate edge sequence, the path total weight corresponding to the node sequence is calculated according to the weight values, then a node sequence with a path total weight greater than or equal to a preset threshold is taken as an initial meta-path, the initial meta-path is enhanced using node attributes to obtain a semantic enhanced meta-path, the semantic enhanced meta-path is hierarchically sampled to obtain a meta-path neighbor dictionary corresponding to each graph node, finally keyword recommendation is performed on the meta-path neighbor dictionary by the heterogeneous convolution layer to obtain at least one first keyword combination corresponding to the first semantic representation and the second semantic representation.

[0126] It should be noted that for the weight sensitive random walk, it refers to an algorithm for selecting a walk path according to an edge weight (that is, a weight value corresponding to a node relationship), and the higher the weight of the edge, the greater the possibility of being selected; the meta path refers to the path model of the semantic relationship between different nodes in the heterogeneous graph, such as "patent-keyword-user-patent" and the like; the semantic enhanced meta path refers to the expansion of the meta path combined with the node attribute (the node attribute can be the static feature data carried by the node itself, which is independent of the network topology structure, and these static feature data can be used to enhance the semantic information of the node representation, which can effectively help the model better understand the internal characteristics of the node in the process of embedding learning); the hierarchical sampling refers to sampling the neighbor nodes according to the meta path type, balancing the calculation efficiency and semantic coverage, etc.; the attention mechanism parameter can be used to learn the importance weight of different meta paths or neighbor nodes; and the similarity algorithm can be used to calculate the correlation between node representations, such as the cosine similarity algorithm, the inner product algorithm, etc., which are not limited by the present application.

[0127] In the keyword recommendation process, the parameter matrix for the meta path neighbor dictionary can be called from the heterogeneous convolution layer first, then the parameter matrix and the weight value are used to perform feature transformation on the meta path neighbor dictionary, the transformed output result is weighted and aggregated, the grouping aggregation result corresponding to the meta path neighbor dictionary is output, the grouping aggregation result is calculated through the trained attention mechanism parameter, the corresponding calculation result is output, and the calculation result is feature weighted and aggregated and residual connected, the first target semantic representation of the first semantic representation and the second target semantic representation corresponding to the second semantic representation are output, the similarity algorithm for the first target semantic representation and the second target semantic representation is obtained, and finally the similarity algorithm is used to recommend the first target semantic representation and the second target semantic representation, and at least one first keyword combination is generated.

[0128] In the specific implementation, in the process of path discovery and meta path generation, in the heterogeneous graph, the possible path is searched according to the edge weight and the node relationship, and the candidate edge sequence connecting the "first semantic representation" (such as a patent) and the "second semantic representation" (such as a user) is obtained. Then, the candidate edge sequence can be converted into a node sequence through the weight sensitive random walk, and the total path weight (such as calculated by edge weight accumulation) is calculated, and after the total path weight is obtained, the path with a weight greater than or equal to a preset threshold value can be selected as an initial meta path. Then, the node attribute (such as a patent title, user information, etc.) is used to enhance the initial meta path, and a meta path with richer semantics (such as "patent A->keyword 'AI'->user B->patent C") is generated.

[0129] In the process of neighbor sampling and feature aggregation, semantic-enhanced meta-paths are hierarchically sampled, and a meta-path neighbor dictionary is constructed for each node (e.g., "Patent A's neighbors: {'keywords': ['AI', 'blockchain'], 'users': ['User B']}"). The parameter matrix is called through the heterogeneous convolution layer, and the neighbor dictionary is transformed in feature (e.g., linear projection) combined with edge weights. The grouping results (e.g., "keyword group aggregation result", "user group aggregation result") are obtained through weighted aggregation.

[0130] In the process of attention mechanism and keyword recommendation, the pre-trained attention mechanism can be used to calculate the weights of different meta-paths or neighbors (e.g., "The attention score of keyword 'AI' is higher than 'blockchain'"). After weighted aggregation, the residual connection is added to output the enhanced target semantic representation, and the similarity between "Patent" and "User" target representations is calculated (e.g., cosine value, etc.). Based on the similarity calculation result, the keyword combination with the highest similarity (e.g., "AI, machine learning") is recommended.

[0131] In some examples, the corresponding processing process of keyword recommendation through the heterogeneous convolution layer can be as follows:

[0132] 1. Build a heterogeneous graph:

[0133] First, define four types of nodes: users, projects, patents, and keywords. These nodes together form a heterogeneous graph, where different types of nodes are connected to each other through different relationships. There are connections between users and projects, projects and patents, and patents and keywords. These relationships may be based on user participation in projects, projects containing patents, and patents mentioning keywords.

[0134] 2. Obtain initial embedding:

[0135] The user and keyword embeddings obtained by the semantic convolution layer are used as the initial embedding of the collaborative convolution layer to better reflect the position and relationship of the nodes in the heterogeneous graph. At the same time, the attribute information of the nodes is input into the model as features, so that these additional information can be considered in the embedding process.

[0136] 3. Introduce weighted edges:

[0137] In order to distinguish the meaning of different edges, each edge is assigned a weight according to specific weight rules. These rules are based on company level, project level and personal level considerations. Company level: for example, if the patent layout strategy of the company is a patent wall of the same type, then the keywords similar to the user's current keywords increase the weight. Project level: if a certain keyword is mentioned in the project attribute project difficulty, and the keyword also exists in the patent associated with the project, then the weight of this path is increased. Personal level: if a user's patent is completed because of a certain keyword, then the weight of the path is increased. In the heterogeneous graph, each edge is automatically assigned a weight according to these rules, and finally a weighted heterogeneous graph is formed.

[0138] For the above weight rules, the edge weight (i.e. the weight value corresponding to the node relationship) can be dynamically adjusted by the degree of association between the keywords and enterprise information, project information, personal information, etc. At the same time, it can also be adjusted according to time decay (i.e. the weight is lower the further away from the current time), achievement transformation multiplier (e.g. if the keyword exists in a commercialized patent, the weight value is doubled; if the keyword does not exist in a commercialized patent, the weight value remains unchanged), and strategy matching degree (i.e. when matching the research direction of the enterprise, the weight value can be multiplied by 1.5 times, etc.). In practical applications, it can be adjusted according to actual needs, and the present application does not limit this.

[0139] Automatic learning and definition of meta-path: In the heterogeneous graph, a meta-path is a sequence that connects different types of nodes. For example, user-project-patent-keyword can be a meaningful meta-path. According to the weight rules and node attribute information, the model automatically learns and defines meaningful meta-paths. Then through R-GCNh or HetGNN, etc. Heterogeneous graph GCN model is used to learn the representation of user nodes and keyword nodes combined with semantic information and collaborative information.

[0140] 4. Recommendation:

[0141] Finally, top n keyword combinations are recommended to each user through cosine similarity or other similarity algorithms, each combination can at least include a technical keyword, a scene keyword and an application keyword.

[0142] For example, assuming that according to the data table, the "AI-based medical image analysis patent" recommendation multi-dimensional associated keywords are to be implemented, the corresponding process can include:

[0143] 1. Heterogeneous graph construction and path discovery:

[0144] Node type:

[0145] Patent P: "A CT image lung cancer detection method based on deep learning"

[0146] Technical keywords: K1 ("Convolutional Neural Network"), K2 ("Attention Mechanism")

[0147] Scenario keywords: S1 ("Medical Imaging"), S2 ("Top-tier Hospital")

[0148] Application keywords: A1 ("Early Lung Cancer Screening"), A2 ("Remote Diagnosis")

[0149] User U: R&D Manager of a Medical AI Company

[0150] Edge relationships and corresponding weights in the heterogeneous graph can be:

[0151] P→K1 (Technology, weight=0.9), P→S1 (Scenario, weight=0.8), P→A1 (Application, weight=0.85), K1→U (User interest in technology, weight=0.7), S1→U (User demand for scenario, weight=0.75)

[0152] Process of candidate path discovery: Through random walk, the following paths can be obtained:

[0153] P→K1→U→S1→P (Technology→User→Scenario→Patent, total weight=2.15)

[0154] P→A1→U→K2→P (Application→User→Technology→Patent, total weight=2.0)

[0155] 2. Semantic enhancement and meta-path generation:

[0156] Initial meta-path:

[0157] Path 1: Patent→Technology→User→Scenario→Patent

[0158] Path 2: Patent→Application→User→Technology→Patent

[0159] Attribute enhancement:

[0160] Expand path semantics with patent attributes ("Deep Learning", "CT Imaging"):

[0161] Enhanced Path 1:

[0162] "Patent (Deep Learning + CT Imaging)→Technology (CNN)→User (Medical AI Company)→Scenario (Top-tier Hospital + Medical Imaging)→Patent"

[0163] Enhanced Path 2:

[0164] "Patent (Lung Cancer Detection)→Application (Early Screening)→User (R&D Manager)→Technology (Attention Mechanism)→Patent"

[0165] 3. Neighbor aggregation and attention calculation:

[0166] Meta-path Neighbor Dictionary for Patent P:

[0167] {

[0168] "Techniques": ["Convolutional Neural Network", "Attention Mechanism"],

[0169] "Scenarios": ["Medical Imaging", "Top-tier Hospital"],

[0170] "Applications": ["Lung Cancer Early Screening", "Remote Diagnosis"]

[0171] }

[0172] Heterogeneous Convolutional Layer Processing:

[0173] Technique Group Aggregation: CNN (Weight 0.6) + Attention Mechanism (Weight 0.4)

[0174] Scenario Group Aggregation: Medical Imaging (Weight 0.7) + Top-tier Hospital (Weight 0.3)

[0175] Application Group Aggregation: Lung Cancer Screening (Weight 0.8) + Remote Diagnosis (Weight 0.2)

[0176] Attention Mechanism Output:

[0177] Technique Keyword Importance: CNN > Attention Mechanism

[0178] Scenario Keyword Importance: Medical Imaging > Top-tier Hospital

[0179] Application Keyword Importance: Lung Cancer Screening > Remote Diagnosis

[0180] 4. Multi-Dimensional Keyword Recommendation

[0181] Target Semantic Representation Similarity Calculation:

[0182] Representation Similarity between Patent P and User U:

[0183] Technique Matching Degree: 0.9 (CNN)

[0184] Scenario Matching Degree: 0.85 (Medical Imaging)

[0185] Application Matching Degree: 0.88 (Lung Cancer Screening)

[0186] Final Recommended Keyword Combination:

[0187] Technique Keywords: Convolutional Neural Network (CNN), Attention Mechanism

[0188] Scenario Keywords: Medical Imaging, Top-tier Hospital

[0189] Application keywords: early screening of lung cancer, remote diagnosis

[0190] In the above process, a weighted heterogeneous graph is used, and different rules are defined to weight the edges in the heterogeneous graph, so that the relationship strength between different nodes can be more accurately described to reflect the similarity, correlation strength, interaction frequency and other factors between nodes, thereby improving the accuracy of the recommendation result.

[0191] Step 105, inputting the first keyword combination into the natural language model for keyword recommendation to obtain at least one corresponding second keyword combination;

[0192] After obtaining the corresponding first keyword combination through the heterogeneous convolution recommendation model, the first keyword combination can be further processed using the natural language model to obtain at least one associated and different second keyword combination, so that the natural language model is used to generate new keyword combinations according to the recommended keyword combinations, thereby providing users with more diversified choices and reducing the probability of users falling into the information cocoon.

[0193] The natural language model can be a large model trained using patent titles, abstracts and other data as a corpus. By inputting the first keyword combination into the natural language model, the natural language model can generate a new set of keyword combinations according to the learning experience, thereby obtaining two different sets of keyword combinations.

[0194] For example, in the process of keyword recommendation, on the one hand, the first keyword combination can be obtained by processing historical data according to the heterogeneous convolution recommendation model, and on the other hand, the first keyword combination can be input into the natural language model to obtain the second keyword combination, so that new keywords related to the original list but not completely the same are generated through generative recommendation, and these new keywords are used as additional recommended options to provide users with more diversified choices and reduce the probability of users falling into the information cocoon.

[0195] Step 106, performing text retrieval according to the first keyword combination and the second keyword combination, and outputting the corresponding target patent text.

[0196] After the first keyword combination and the second keyword combination are obtained through the above process, text retrieval can be performed according to the first keyword combination and the second keyword combination, and the corresponding target patent text is output based on the retrieval result, so that the keyword recommendation processing is performed through the semantic convolution layer and the heterogeneous convolution layer, the information is more rich, the recommendation result is more accurate, meanwhile, a new keyword combination is generated according to the recommended keyword combination by using the natural language model, more diversified choices are provided for the user, the probability of the user being trapped in the information cocoon is reduced, and text retrieval is performed based on the rich keyword combination, so that the comprehensiveness and accuracy of the retrieval result are effectively ensured.

[0197] In some examples, the various keywords (such as technical keywords, scene keywords, and application keywords) involved in the first keyword combination and the second keyword combination recommended to the user can be taken as Prompt prompt words, and input to the language large model to assist the model to generate patent titles and abstracts that combine related technical information, scene information, and application information, etc. Based on the corresponding text recommendation, more rich, comprehensive and accurate text content can be effectively provided for the user, the recommendation effect is guaranteed, so that the user can obtain and study the technical content associated with the research direction based on the generated patent title and abstract text information, thereby assisting the user to better understand the technical content of the related technical direction.

[0198] It should be noted that the embodiments of the present application include but are not limited to the above examples, and it can be understood that those skilled in the art can also set according to actual needs under the guidance of the idea of the embodiments of the present application, and the present application does not limit this.

[0199] In the embodiment of the present application, when a user wants to search for related text, a data table is obtained, the data table at least including a patent data table and a user data table, then a heterogeneous graph convolution recommendation model and a natural language model are obtained, the heterogeneous graph convolution recommendation model at least including a semantic convolution layer and a heterogeneous convolution layer, then the patent data table and the user data table are input into the semantic convolution layer for semantic extraction, obtaining a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table, and the first semantic representation and the second semantic representation are input into the heterogeneous convolution layer for keyword recommendation, obtaining at least one first keyword combination corresponding thereto, and the first keyword combination is input into the natural language model for keyword recommendation, obtaining at least one second keyword combination corresponding thereto, and finally, text retrieval is performed according to the first keyword combination and the second keyword combination, and the corresponding target patent text is output, so that the keyword recommendation processing is performed through the semantic convolution layer and the heterogeneous convolution layer, the information is more abundant, the recommendation result is more accurate, at the same time, the natural language model is used to generate new keyword combinations according to the recommended keyword combinations, more diversified choices are provided for the user, the probability of the user falling into the information cocoon is reduced, and text retrieval is performed based on the rich keyword combinations, effectively ensuring the comprehensiveness and accuracy of the retrieval result.

[0200] In order for those skilled in the art to better understand the technical solutions in the embodiments of the present application, the following exemplary examples are used for illustration:

[0201] As an example, referring to Figure 2 , a schematic diagram of a recommendation process provided in the embodiment of the present application is shown, which can at least include the following processes:

[0202] For the process of data acquisition and data storage, the patent data is obtained, the abstract, title, technical field and transformation information are decomposed, and stored in the patent data table; the title, research content, key technology and project difficulty of the project in the enterprise are obtained, and stored in the project data table; the professional, work position, research direction and achievement information of the user in the enterprise are obtained, and stored in the user data table; the main business and product layout information of the enterprise is obtained, and stored in the enterprise data table.

[0203] After obtaining the corresponding data content, on the one hand, the title and abstract data of the patent can be used as a corpus to train a natural language model, and on the other hand, it can be used for data analysis, and the keyword recommendation is performed through a recommendation algorithm (a heterogeneous graph convolution recommendation model combining semantic structure and collaborative structure), to obtain a top n ternary keyword candidate list recommended in combination with historical data.

[0204] In the top n ternary keyword candidate list is obtained, it can be input to the natural language model for keyword expansion, generating recommended top n ternary keyword candidate list.

[0205] Through the above two groups of top n ternary keyword candidate list of keywords obtained by patent text retrieval, and based on the retrieval result output corresponding patent title and abstract list, etc., so that the semantic structure and heterogeneous structure are used for keyword recommendation processing, so that the information is more rich, the recommendation result is more accurate, at the same time, through using natural language model according to the recommended keyword combination to generate new keyword combination, provide more diversified selection for user, reduce the probability of user trapped in information cocoon, and based on the rich keyword combination for text retrieval, effectively guarantee the comprehensiveness and accuracy of the retrieval result.

[0206] For example, referring to Figure 3 , a schematic diagram of the application scenario provided in the embodiment of the application is shown, the heterogeneous convolution recommendation model can include semantic module (graph convolution aggregation) and collaborative module (weighted heterogeneous graph convolution), etc., the heterogeneous convolution recommendation model can output corresponding keyword candidate list according to the constructed data table, such as:

[0207] 1. Intelligent home, Internet of things technology, intelligent control;

[0208] 2. Online education, cloud computing platform, remote teaching;

[0209] 3. Automatic driving, deep learning algorithm, vehicle navigation;

[0210] 4. Wisdom agriculture, Internet of things sensor, precision planting……, etc.

[0211] Then, the top n keyword combinations in these keyword candidate lists can be input to the natural language model for keyword expansion, obtaining associated and different keyword candidate lists, such as:

[0212] 1. Intelligent home appliance, sensing interconnection technology, scene linkage;

[0213] 2. Online classroom, cloud service architecture, asynchronous teaching;

[0214] 3. Driverless, neural network algorithm, path planning;

[0215] 4. Digital farmland, environment perception device, crop optimization……, etc.

[0216] Further, based on the obtained keyword candidate list, the patent title and patent abstract are recommended, such as:

[0217] Recommendation list 1:

[0218] Patent title: XXXXXXXX;

[0219] Patent abstract: XXXXXXXXXX;

[0220] Recommended list 2:

[0221] Patent title: YYYYYYY;

[0222] Patent abstract: YYYYYYYYYYY;

[0223] , etc.

[0224] In the above process, the keyword recommendation processing is performed through the semantic convolution layer and the heterogeneous convolution layer, so that the information is more rich and the recommendation result is more accurate. Meanwhile, new keyword combinations are generated according to the recommended keyword combinations by using the natural language model, so as to provide more diversified choices for the user, reduce the probability of the user being trapped in the information cocoon, and perform text retrieval based on the rich keyword combinations, so as to effectively ensure the comprehensiveness and accuracy of the retrieval result.

[0225] It should be noted that, for the method embodiments, in order to simply describe, they are all described as a series of action combinations, but those skilled in the art should know that the embodiments of the present application are not limited by the action sequence described, because according to the embodiments of the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions involved are not necessarily necessary for the embodiments of the present application.

[0226] Referring to Figure 4 , a structural block diagram of a text information recommendation device provided in the embodiments of the present application is shown, which can specifically include the following modules:

[0227] The data acquisition module 401 is configured to acquire a data table, wherein the data table at least includes a patent data table and a user data table;

[0228] The model acquisition module 402 is configured to acquire a heterogeneous graph convolution recommendation model and a natural language model, wherein the heterogeneous graph convolution recommendation model at least includes a semantic convolution layer and a heterogeneous convolution layer;

[0229] The semantic extraction module 403 is configured to input the patent data table and the user data table into the semantic convolution layer for semantic extraction, so as to obtain a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table;

[0230] The first recommendation module 404 is configured to input the first semantic representation and the second semantic representation into the heterogeneous convolution layer to perform keyword recommendation, and obtain at least one first keyword combination corresponding thereto.

[0231] The second recommendation module 405 is configured to input the first keyword combination into the natural language model to perform keyword recommendation, and obtain at least one second keyword combination corresponding thereto.

[0232] The text retrieval module 406 is configured to perform text retrieval according to the first keyword combination and the second keyword combination, and output target patent text corresponding thereto.

[0233] In some possible implementation manners, the data table further includes a project data table, and the semantic extraction module is specifically configured to:

[0234] extract employee information from the user data table, and take the employee information as semantic data;

[0235] extract project information from the project data table and extract patent information from the patent data table, and take the project information and the patent information as auxiliary semantic data corresponding to the semantic data;

[0236] extract a plurality of patent keywords from the patent information;

[0237] train the semantic data, the auxiliary semantic data, and the patent keywords to obtain a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table.

[0238] In some possible implementation manners, the semantic convolution layer is a graph convolution network, and the graph convolution network at least includes an input layer, a plurality of execution convolution layers, an activation function, a pooling layer, and an output layer; wherein the graph convolution network further includes a first user node and a first patent node, and the semantic extraction module is specifically configured to:

[0239] convert the semantic data and the auxiliary semantic data into a first embedding vector;

[0240] convert the plurality of patent keywords into a second embedding vector;

[0241] take the first embedding vector as a node feature of the first user node, and take the second embedding vector as a node feature of the first patent node, to obtain a first feature vector of the patent data table and a second feature vector of the user data table;

[0242] aggregate information of neighbor nodes through the execution convolution layer, and iteratively update the first feature vector and the second feature vector to obtain a node representation fused with neighborhood semantics.

[0243] performing a nonlinear transformation on the fused neighborhood semantics of the node representation according to the activation function, to obtain a transformed node representation;

[0244] performing dimension reduction on the transformed node representation through the pooling layer, to obtain a compressed node representation, and generating a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table through linear transformation and task adaptation processing of the output layer.

[0245] In some possible implementation manners, the employee information at least includes one of user basic information, professional background information, post data, research direction, and historical achievements; the project information at least includes one of project basic information, technical content, and achievement data; and the patent information at least includes one of a title, an abstract, a claim text, a patent type, classification information, technical feature text, conversion data, pledge information, financing information, and text derivative information.

[0246] In some possible implementation manners, the first recommendation module 404 is specifically configured to:

[0247] construct a heterogeneous graph for the first semantic representation and the second semantic representation, the heterogeneous graph including several different types of graph nodes, and the different types of graph nodes in the heterogeneous graph being connected through corresponding node relationships;

[0248] obtain node attributes corresponding to each graph node;

[0249] obtain weight values corresponding to each node relationship;

[0250] input the first semantic representation and the second semantic representation into the heterogeneous convolution layer, and perform keyword recommendation according to the node attributes, the node relationships, and the weight values, to obtain at least one first keyword combination corresponding thereto.

[0251] In some possible implementation manners, the first recommendation module 404 is specifically configured to:

[0252] perform path finding in the heterogeneous graph by using the weight values and the node relationships, to obtain a candidate edge sequence corresponding to the first semantic representation and the second semantic representation;

[0253] perform weight-sensitive random walk on the candidate edge sequence, to obtain a node sequence corresponding to the candidate edge sequence, and calculate a total path weight corresponding to the node sequence according to the weight values;

[0254] take a node sequence with a total path weight greater than or equal to a preset threshold as an initial meta-path;

[0255] The initial metapath is enhanced using the node attributes to obtain a semantically enhanced metapath;

[0256] The semantically enhanced meta-paths are sampled hierarchically to obtain the meta-path neighbor dictionary corresponding to each graph node;

[0257] Based on the heterogeneous convolutional layer, keyword recommendations are performed on the meta-path neighbor dictionary to obtain at least one combination of first keywords corresponding to the first semantic representation and the second semantic representation.

[0258] In some feasible implementations, the first recommendation module 404 is specifically used for:

[0259] The parameter matrix for the meta-path neighbor dictionary is called from the heterogeneous convolutional layer;

[0260] The parameter matrix and the weight values ​​are used to perform feature transformation on the meta-path neighbor dictionary, and the transformed output results are weighted and aggregated to output the grouping aggregation results corresponding to the meta-path neighbor dictionary;

[0261] The grouping aggregation result is calculated using the trained attention mechanism parameters, and the corresponding calculation result is output. The calculation result is then subjected to feature weighting aggregation and residual connection to output the first target semantic representation of the first semantic representation and the second target semantic representation corresponding to the second semantic representation.

[0262] An algorithm for obtaining the similarity between the first target semantic representation and the second target semantic representation;

[0263] The similarity algorithm is used to recommend keywords to the first target semantic representation and the second target semantic representation, generating at least one corresponding first keyword combination.

[0264] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0265] In addition, embodiments of the present invention also provide an electronic device, including: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the various processes of the above-described recommended method embodiments for text information and achieves the same technical effect. To avoid repetition, it will not be described again here.

[0266] This invention also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the various processes of the recommended method embodiment for the aforementioned text information and achieves the same technical effect. To avoid repetition, it will not be described again here. The computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0267] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0268] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, EEPROM, Flash, and eMMC, etc.) containing computer-usable program code.

[0269] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0270] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0271] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0272] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

[0273] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0274] The foregoing has provided a detailed description of a text information recommendation method and a text information recommendation device provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for recommending text information, characterized in that, include: Obtain data tables, which include at least a patent data table and a user data table; Obtain a heterogeneous graph convolutional recommendation model and a natural language model, wherein the heterogeneous graph convolutional recommendation model includes at least a semantic convolutional layer and a heterogeneous convolutional layer; The patent data table and the user data table are input into a semantic convolutional layer for semantic extraction to obtain a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table. The first semantic representation and the second semantic representation are input into the heterogeneous convolutional layer for keyword recommendation to obtain at least one corresponding first keyword combination; The first keyword combination is input into the natural language model for keyword recommendation to obtain at least one corresponding second keyword combination; The text is retrieved based on the first keyword combination and the second keyword combination, and the corresponding target patent text is output. The step of inputting the first semantic representation and the second semantic representation into the heterogeneous convolutional layer for keyword recommendation to obtain at least one corresponding first keyword combination includes: Construct a heterogeneous graph for the first semantic representation and the second semantic representation, wherein the heterogeneous graph has several different types of graph nodes, and the different types of graph nodes in the heterogeneous graph are connected through corresponding node relationships; Obtain the node attributes corresponding to each of the graph nodes; Obtain the weight value corresponding to each of the node relationships; Using the weight values ​​and the node relationships, path finding is performed in the heterogeneous graph to obtain candidate edge sequences corresponding to the first semantic representation and the second semantic representation; Perform a weight-sensitive random walk on the candidate edge sequence to obtain a node sequence on the candidate edge sequence, and calculate the total path weight corresponding to the node sequence based on the weight value. The weight-sensitive random walk is an algorithm that selects a walk path based on the weight value corresponding to the node relationship. The higher the weight of the node relationship, the greater the probability of it being selected. The node sequence with a total path weight greater than or equal to a preset threshold is used as the initial metapath; The initial metapath is enhanced using the node attributes to obtain a semantically enhanced metapath; The semantically enhanced meta-paths are sampled hierarchically to obtain the meta-path neighbor dictionary corresponding to each graph node; The parameter matrix for the meta-path neighbor dictionary is called from the heterogeneous convolutional layer; The parameter matrix and the weight values ​​are used to perform feature transformation on the meta-path neighbor dictionary, and the transformed output results are weighted and aggregated to output the grouping aggregation results corresponding to the meta-path neighbor dictionary; The grouping aggregation result is calculated using the trained attention mechanism parameters, and the corresponding calculation result is output. The calculation result is then subjected to feature weighting aggregation and residual connection to output the first target semantic representation of the first semantic representation and the second target semantic representation corresponding to the second semantic representation. An algorithm for obtaining the similarity between the first target semantic representation and the second target semantic representation; The similarity algorithm is used to recommend keywords to the first target semantic representation and the second target semantic representation, generating at least one corresponding first keyword combination.

2. The method according to claim 1, characterized in that, The data table also includes a project data table. The step of inputting the patent data table and the user data table into a semantic convolutional layer for semantic extraction to obtain a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table includes: Employee information is extracted from the user data table and used as semantic data. Project information is extracted from the project data table, and patent information is extracted from the patent data table. The project information and the patent information are used as auxiliary semantic data corresponding to the semantic data. Extract multiple patent keywords from the patent information; The semantic data, the auxiliary semantic data, and each of the patent keywords are trained to obtain a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table.

3. The method according to claim 2, characterized in that, The semantic convolutional layer is a graph convolutional network, which includes at least an input layer, multiple execution convolutional layers, an activation function, a pooling layer, and an output layer. The graph convolutional network also includes a first user node and a first patent node. The step of training the semantic data, the auxiliary semantic data, and each of the patent keywords to obtain a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table includes: The semantic data and the auxiliary semantic data are converted into a first embedding vector; Convert the plurality of patent keywords into a second embedding vector; The first embedding vector is used as the node feature of the first user node, and the second embedding vector is used as the node feature of the first patent node to obtain the first feature vector of the patent data table and the second feature vector of the user data table. By aggregating the information of neighboring nodes through the convolutional layer, the first feature vector and the second feature vector are iteratively updated to obtain a node representation that integrates neighborhood semantics; The node representation of the fused neighborhood semantics is nonlinearly transformed according to the activation function to obtain the transformed node representation; The transformed node representation is reduced in dimensionality by the pooling layer to obtain a compressed node representation. Then, through linear transformation and task adaptation processing of the output layer, a first semantic representation corresponding to the patent data table and a second semantic representation corresponding to the user data table are generated.

4. The method according to claim 2 or 3, characterized in that, The employee information includes at least one of the following: basic user information, professional background information, job data, research direction, and historical achievements; the project information includes at least one of the following: basic project information, technical content, and achievement data; the patent information includes at least one of the following: title, abstract, claim text, patent type, classification information, technical feature text, transformation data, pledge information, financing information, and text-derived information.

5. A text information recommendation device, characterized in that, include: A data acquisition module is used to acquire data tables, which include at least a patent data table and a user data table. The model acquisition module is used to acquire a heterogeneous graph convolutional recommendation model and a natural language model, wherein the heterogeneous graph convolutional recommendation model includes at least a semantic convolutional layer and a heterogeneous convolutional layer; The semantic extraction module is used to input the patent data table and the user data table into the semantic convolutional layer for semantic extraction, and obtain the first semantic representation corresponding to the patent data table and the second semantic representation corresponding to the user data table. The first recommendation module is used to input the first semantic representation and the second semantic representation into the heterogeneous convolutional layer to recommend keywords and obtain at least one corresponding first keyword combination. The second recommendation module is used to input the first keyword combination into the natural language model to recommend keywords and obtain at least one corresponding second keyword combination. The text retrieval module is used to perform text retrieval according to the first keyword combination and the second keyword combination, and output the corresponding target patent text; In some feasible implementations, the first recommendation module is specifically used for: Construct a heterogeneous graph for the first semantic representation and the second semantic representation, wherein the heterogeneous graph has several different types of graph nodes, and the different types of graph nodes in the heterogeneous graph are connected through corresponding node relationships; Obtain the node attributes corresponding to each of the graph nodes; Obtain the weight value corresponding to each of the node relationships; Using the weight values ​​and the node relationships, path finding is performed in the heterogeneous graph to obtain candidate edge sequences corresponding to the first semantic representation and the second semantic representation; Perform a weight-sensitive random walk on the candidate edge sequence to obtain a node sequence on the candidate edge sequence, and calculate the total path weight corresponding to the node sequence based on the weight value. The weight-sensitive random walk is an algorithm that selects a walk path based on the weight value corresponding to the node relationship. The higher the weight of the node relationship, the greater the probability of it being selected. The node sequence with a total path weight greater than or equal to a preset threshold is used as the initial metapath; The initial metapath is enhanced using the node attributes to obtain a semantically enhanced metapath; The semantically enhanced meta-paths are sampled hierarchically to obtain the meta-path neighbor dictionary corresponding to each graph node; The parameter matrix for the meta-path neighbor dictionary is called from the heterogeneous convolutional layer; The parameter matrix and the weight values ​​are used to perform feature transformation on the meta-path neighbor dictionary, and the transformed output results are weighted and aggregated to output the grouping aggregation results corresponding to the meta-path neighbor dictionary; The grouping aggregation result is calculated using the trained attention mechanism parameters, and the corresponding calculation result is output. The calculation result is then subjected to feature weighting aggregation and residual connection to output the first target semantic representation of the first semantic representation and the second target semantic representation corresponding to the second semantic representation. An algorithm for obtaining the similarity between the first target semantic representation and the second target semantic representation; The similarity algorithm is used to recommend keywords to the first target semantic representation and the second target semantic representation, generating at least one corresponding first keyword combination.

6. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; The memory is used to store computer programs; When the processor executes a program stored in the memory, it implements the method as described in any one of claims 1-4.

7. A computer-readable storage medium having instructions stored thereon that, when executed by one or more processors, cause the processors to perform the method as described in any one of claims 1-4.

Images