A scientific literature recommendation method integrating multi-dimensional features of metadata

Abstract

The invention discloses a method for recommending scientific and technological documents by integrating multi-dimensional features of metadata, including: obtaining document interaction record data and document metadata, and performing preprocessing; constructing a knowledge graph based on the preprocessed document metadata, and linking the preprocessed document interaction record data with entities in the knowledge graph to obtain a final knowledge graph; outputting a first embedding vector of each document in the document interaction record data through a knowledge graph embedding model; outputting a second embedding vector of each document in the document interaction record data through a language model; fusing the first embedding vector and the second embedding vector to obtain a final embedding vector of each document in the document interaction record data; obtaining a user research intention representation vector according to the final embedding vector; and obtaining a recommended document set according to the user research intention representation vector and a recommended document list. The invention can improve the accuracy of document recommendation.

Description

Technical Field

[0001] The present invention belongs to the field of document recommendation, relates to a recommendation method, and specifically relates to a scientific and technological document recommendation method integrating multi-dimensional features of metadata. Background Art

[0002] In the modern society with developed Internet, the information generated every day is growing explosively. When faced with this information, users need to spend a lot of attention and energy to filter the information they want. For researchers, the amount of literature published is also increasing rapidly. It takes a lot of energy to find the literature needed for research, which reduces the efficiency of research. Traditionally, literature is stored in paper form. With the development of information technology, literature is stored in electronic form, which improves the efficiency of literature retrieval and speeds up the dissemination of literature. In the context of information technology, literature recommendation systems can help researchers better match the literature needed for research.

[0003] Traditional content-based recommendation methods need to calculate project-related features through feature engineering. Different features need to be extracted for different types of projects, which increases the workload. The basic idea of ​​collaborative filtering-based algorithms is to find users similar to the current user. The projects that these similar users are interested in are also the projects that the current user is interested in, avoiding the extraction of project features alone, but there are problems such as cold start and sparsity. In recent years, recommendation systems based on knowledge graphs have effectively alleviated the cold start and sparsity problems in recommendation systems. However, existing recommendation systems based on knowledge graphs mostly use information extracted from text to construct knowledge graphs. There is a certain probability that wrong information will be extracted during the extraction process, and the extracted information is too single and cannot well represent the implicit relationship between documents. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for recommending scientific and technological documents by integrating multi-dimensional features of metadata, which solves the problem of inaccurate scientific and technological document recommendation to users in the prior art.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0006] A method for recommending scientific and technological literature by integrating multi-dimensional features of metadata comprises the following steps:

[0007] S1. Obtain document interaction record data and document metadata, and preprocess the document interaction record data and document metadata to obtain preprocessed document interaction record data and preprocessed document metadata;

[0008] Among them, document interaction record data is obtained from users, and document metadata is obtained from third-party document databases;

[0009] S2. Based on the preprocessed document metadata, a knowledge graph is constructed, and the preprocessed document interaction record data and entities in the knowledge graph are linked to obtain the final knowledge graph;

[0010] S3, input the final knowledge graph into the knowledge graph embedding model, and output the first embedding vector of each document in the document interaction record data through the knowledge graph embedding model;

[0011] S4, inputting the title and abstract of each document in the document interaction record data into the language model, and outputting the second embedding vector of each document in the document interaction record data through the language model;

[0012] S5, fusing the first embedding vector and the second embedding vector to obtain a final embedding vector for each document in the document interaction record data;

[0013] S6. Obtain the user research intention representation vector based on the final embedding vector;

[0014] S7. Obtain a recommended document set based on the user research intention representation vector and the recommended document list.

[0015] The beneficial effects of the above scheme are:

[0016] The present invention utilizes high-quality metadata in third-party document databases to enrich the representation of documents in the document system, further expands the semantic representation of documents using long-text semantic representations of titles and abstracts, and uses document records of user interactions to construct user representation vectors that contain research intent, thereby improving the accuracy of document recommendations.

[0017] Furthermore, in step S1, obtaining document interaction record data specifically includes:

[0018] S11, obtaining the document interaction record of the user, where the document interaction record includes document download record, document collection record and document browsing history record;

[0019] S12. The user ID, document title, document author and first document identifier in each document interaction record are used as document interaction record data.

[0020] Furthermore, in step S1, obtaining document metadata specifically includes:

[0021] S13, obtaining initial document metadata from a third-party document database;

[0022] S14. The author keywords, extended keywords, research field, author, source journal and publisher information in the initial document metadata are used as document metadata, and the document metadata is obtained by sorting according to the second document identifier and the number of citations in the initial document metadata.

[0023] The beneficial effect of the above further solution is that by obtaining document interaction record data and document metadata, the diversity of initial data can be guaranteed, and when the data is subsequently processed, the information of the representation vector obtained is richer.

[0024] Furthermore, in step S2, a knowledge graph is constructed based on the preprocessed document metadata, specifically including:

[0025] S21, obtaining a first node according to the source journal and publisher information in the preprocessed document metadata;

[0026] S22, splitting the author keywords, extended keywords, research fields and authors in the preprocessed document metadata to obtain a second node;

[0027] S23, removing duplicates from the first node and the second node to obtain a final node;

[0028] S24, according to the final node, establish the inArea many-to-many relationship, the researchBy many-to-many relationship, the masterIn many-to-many relationship and the publishBy many-to-many relationship respectively;

[0029] Among them, the inArea many-to-many relationship is established based on the relationship between the author keywords, extended keywords and research fields in the final node, the researchBy many-to-many relationship is established based on the relationship between the research field and the author in the final node, the masterIn many-to-many relationship is established based on the relationship between the research field and the source journal in the final node, and the publishBy many-to-many relationship is established based on the relationship between the source journal and the publisher information in the final node;

[0030] S25. Establish a knowledge graph based on the inArea many-to-many relationship, the researchBy many-to-many relationship, the masterIn many-to-many relationship and the publishBy many-to-many relationship.

[0031] The beneficial effects of the above further scheme are: many-to-many relationships allow complex information interaction patterns, which can capture richer and more precise associations between entities, making the knowledge graph deeper and broader; through many-to-many relationships, more detailed and in-depth data analysis can be performed, such as discovering patterns and trends hidden behind the complex interactions between document nodes; many-to-many relationships can provide more reasoning paths and logic, enhancing the performance of the graph in reasoning and intelligent query.

[0032] Further, in step S3, the knowledge graph embedding model is a TransE model, and the data input form of the knowledge graph embedding model is a triple;

[0033] The activation function of the knowledge graph embedding model is:

[0034]

[0035] in, represents the first embedding vector of the document, n represents the total number of nodes related to the current document, and x i represents the feature vector of the i-th node, W i represents the weight of the feature vector of the i-th node, sigmoid represents the sigmoid activation function, and b1 represents the regular term when learning the first embedding vector.

[0036] Furthermore, in step S4, the language model is a Bert model, and the activation function of the language model is:

[0037]

[0038] in, represents the second embedding vector of the document, W(x title ,x abstract ) represents the weight value assigned to the concatenation of the feature vector of the title and the feature vector of the summary, and b2 represents the regularization term when learning the second embedding vector.

[0039] Further, in step S5, the activation function used to fuse the first embedding vector and the second embedding vector is:

[0040]

[0041] Among them, agg P represents the final embedding vector of the document, represents the weight value assigned to the concatenation of the first embedding vector and the second embedding vector, and b3 represents the regularization term when the first embedding vector and the second embedding vector are fused into the final embedding vector.

[0042] Furthermore, in step S6, the activation function used to obtain the user research intention representation vector is obtained according to the final embedding vector:

[0043] agg U =sigmoid(W down P down +W star P star +W click P click +b4)

[0044] Among them, agg U represents the user research intention representation vector, P down Represents the final embedding vector corresponding to the download interaction behavior, Wdown represents the weight of the final embedding vector corresponding to the download interaction behavior, P star Represents the final embedding vector corresponding to the collection interaction behavior, W star represents the weight of the final embedding vector corresponding to the collection interaction behavior, P click Represents the final embedding vector corresponding to the click interaction behavior, W click represents the weight of the final embedding vector corresponding to the click interaction behavior, and b4 represents the regularization term when learning the user research intention representation vector.

[0045] The beneficial effect of the above further scheme is that the user research intention representation vector contains the relationship between the user in the process of interacting with the literature and the many-to-many relationship between the nodes of the knowledge graph, which can well represent complex relationships through vectors and improve the data conversion and representation capabilities.

[0046] Further, step S7 specifically includes:

[0047] S71. Determine, according to the recommended document list, a final embedding vector of the recommended document for each piece of data in the recommended document list;

[0048] S72, calculating the cosine similarity between the user's research intention representation vector and the final embedding vector of each recommended document, and obtaining the recommendation value of each data in the recommended document list;

[0049] S73, determining an initial document recommendation set according to data whose recommendation value is greater than a preset threshold;

[0050] S74. According to the number of citations of each piece of data in the initial document recommendation set, the initial document recommendation set is sorted to obtain a recommended document set.

[0051] The beneficial effect of the above further scheme is: by calculating the similarity between the vector of each data in the recommended literature list and the vector representing the user's research intention, the recommended literature list can be better screened, a collection of literature that better meets the user's intention can be determined, and the accuracy of scientific and technological literature recommendations can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 The figure is a flowchart of a scientific literature recommendation method integrating multi-dimensional features of metadata. DETAILED DESCRIPTION

[0053] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0054] like Figure 1 As shown, a method for recommending scientific and technological literature by integrating multi-dimensional features of metadata includes the following steps:

[0055] S1. Obtain document interaction record data and document metadata, and preprocess the document interaction record data and document metadata to obtain preprocessed document interaction record data and preprocessed document metadata;

[0056] Among them, the document interaction record data is obtained from users, and the document metadata is obtained from a third-party document database.

[0057] In this embodiment, in step S1, obtaining document interaction record data specifically includes:

[0058] S11, obtaining the document interaction record of the user, where the document interaction record includes document download record, document collection record and document browsing history record;

[0059] S12. The user ID, document title, document author and first document identifier in each document interaction record are used as document interaction record data.

[0060] Furthermore, in step S1, obtaining document metadata specifically includes:

[0061] S13, obtaining initial document metadata from a third-party document database;

[0062] S14. The author keywords, extended keywords, research field, author, source journal and publisher information in the initial document metadata are used as document metadata, and the document metadata is obtained by sorting according to the second document identifier and the number of citations in the initial document metadata.

[0063] Exemplarily, the document interaction record data and the document metadata are preprocessed to remove null values ​​from the document interaction record data and the document metadata, respectively.

[0064] For example, obtaining the user's document interaction records and preprocessing the document interaction record data can be: exporting the user interaction records, i.e., the download records, collections, and history records of the user's reading of documents; selecting the user ID, document title, document author, and document DOI columns from the exported records, and removing records containing empty values.

[0065] For another example, obtaining document metadata and preprocessing the document metadata may include: exporting the document metadata from a third-party document database and removing records containing null values.

[0066] S2. Based on the preprocessed document metadata, a knowledge graph is constructed, and the preprocessed document interaction record data and entities in the knowledge graph are linked to obtain the final knowledge graph.

[0067] In this embodiment, in step S2, a knowledge graph is constructed based on the preprocessed document metadata, specifically including:

[0068] S21, obtaining a first node according to the source journal and publisher information in the preprocessed document metadata;

[0069] S22, splitting the author keywords, extended keywords, research fields and authors in the preprocessed document metadata to obtain a second node;

[0070] S23, removing duplicates from the first node and the second node to obtain a final node;

[0071] S24, according to the final node, establish the inArea many-to-many relationship, the researchBy many-to-many relationship, the masterIn many-to-many relationship and the publishBy many-to-many relationship respectively;

[0072] Among them, the inArea many-to-many relationship is established based on the relationship between the author keywords, extended keywords and research fields in the final node, the researchBy many-to-many relationship is established based on the relationship between the research field and the author in the final node, the masterIn many-to-many relationship is established based on the relationship between the research field and the source journal in the final node, and the publishBy many-to-many relationship is established based on the relationship between the source journal and the publisher information in the final node;

[0073] S25. Establish a knowledge graph based on the inArea many-to-many relationship, the researchBy many-to-many relationship, the masterIn many-to-many relationship and the publishBy many-to-many relationship.

[0074] Exemplarily, for all data records, fields such as source journals and publisher information that do not require further operations can be directly used as nodes; information that can be subjected to regular operations such as author keywords, extended keywords, research fields, and authors can be processed; based on data characteristics, the data can be split according to the delimiters of the content in the corresponding fields; and each keyword, research field, and author after the split can be used as a node in the knowledge graph.

[0075] Exemplarily, for each data record, for each keyword in the author keyword and extended keyword fields, they can be pointed to each research field in the research field field in the current data record, forming a many-to-many relationship, and the relationship type between the keyword and the research field can be set to inArea; similarly, multiple research fields can be pointed to each author in the author field in the current data record, forming a many-to-many relationship, and the relationship type between the research field and the author field can be set to researchBy; multiple research fields can be pointed to the source journal in the current data record, forming a many-to-many relationship, and the relationship type between the research field and the source journal can be set to masterIn; the source journal can be pointed to the publisher information in the current data record, forming a many-to-one relationship, and the relationship type between the source journal and the publisher information can be set to publishBy.

[0076] S3. Input the final knowledge graph into the knowledge graph embedding model, and output the first embedding vector of each document in the document interaction record data through the knowledge graph embedding model.

[0077] In this embodiment, in step S3, the knowledge graph embedding model is a TransE model, and the data input form of the knowledge graph embedding model is a triple;

[0078] The activation function of the knowledge graph embedding model is:

[0079]

[0080] in, represents the first embedding vector of the document, n represents the total number of nodes related to the current document, and x i represents the feature vector of the i-th node, W i represents the weight of the feature vector of the i-th node, sigmoid represents the sigmoid activation function, and b1 represents the regular term when learning the first embedding vector.

[0081] Exemplarily, the constructed document metadata knowledge graph is input into the knowledge graph embedding model, namely the TransE model. The input of the model is the triple of the constructed knowledge graph, which may include: {keyword, inArea, research field}, {research field, researchBy, author}, {research field, masterIn, source journal}, {source journal, publishBy, publisher information}; the triple form is as follows:

[0082] S={h,r,t}

[0083] Among them, h represents a node in the knowledge graph, r represents a relationship of the h node, and t represents the tail node corresponding to the h node in the r relationship.

[0084] S4. Input the title and abstract of each document in the document interaction record data into the language model, and output the second embedding vector of each document in the document interaction record data through the language model.

[0085] In this embodiment, in step S4, the language model is a Bert model, and the activation function of the language model is:

[0086]

[0087] in, represents the second embedding vector of the document, W(x title ,x abstract ) represents the weight value assigned to the concatenation of the feature vector of the title and the feature vector of the summary, and b2 represents the regularization term when learning the second embedding vector.

[0088] S5. The first embedding vector and the second embedding vector are merged to obtain a final embedding vector for each document in the document interaction record data.

[0089] In this embodiment, in step S5, the activation function used to fuse the first embedding vector and the second embedding vector is:

[0090]

[0091] Among them, agg P represents the final embedding vector of the document, represents the weight value assigned to the concatenation of the first embedding vector and the second embedding vector, and b3 represents the regularization term when the first embedding vector and the second embedding vector are fused into the final embedding vector.

[0092] S6. Based on the final embedding vector, obtain the user research intention representation vector.

[0093] In this embodiment, in step S6, the activation function used to obtain the user research intention representation vector according to the final embedding vector is:

[0094] agg U =sigmoid(W down P down +W star P star +W click P click +b4)

[0095] Among them, agg U represents the user research intention representation vector, P down Represents the final embedding vector corresponding to the download interaction behavior, W downrepresents the weight of the final embedding vector corresponding to the download interaction behavior, P star Represents the final embedding vector corresponding to the collection interaction behavior, W star represents the weight of the final embedding vector corresponding to the collection interaction behavior, P click Represents the final embedding vector corresponding to the click interaction behavior, W click represents the weight of the final embedding vector corresponding to the click interaction behavior, and b4 represents the regularization term when learning the user research intention representation vector.

[0096] Exemplarily, based on the output final embedding vector, the user research intention representation vector is calculated: since there are different types of user interaction behaviors, each behavior has a different impact on the user research intention. Therefore, the user research intention representation vector is calculated for different interaction behaviors and the final embedding vector corresponding to each behavior type, and different weights are assigned to different behaviors.

[0097] In this embodiment, the first embedding vector and the second embedding vector are fused to obtain the final embedding vector of each document in the document interaction record data, which can be performed by a trained fusion network model, and when the fusion network is trained, three weights W down , W star , W click Assign different initial values ​​according to W down >W star >W click The principle of giving them initial values, and W down +W star +W click Should be 1.

[0098] S7. Obtain a recommended document set based on the user research intention representation vector and the recommended document list.

[0099] In this embodiment, step S7 specifically includes:

[0100] S71. Determine, according to the recommended document list, a final embedding vector of the recommended document for each piece of data in the recommended document list;

[0101] S72, calculating the cosine similarity between the user's research intention representation vector and the final embedding vector of each recommended document, and obtaining the recommendation value of each data in the recommended document list;

[0102] S73, determining an initial document recommendation set according to data whose recommendation value is greater than a preset threshold;

[0103] S74. According to the number of citations of each piece of data in the initial document recommendation set, the initial document recommendation set is sorted to obtain a recommended document set.

[0104] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the invention.

Claims

1. A scientific literature recommendation method integrating multi-dimensional features of metadata, characterized in that: The method comprises: S1. Acquire document interaction record data and document metadata, and preprocess the document interaction record data and the document metadata to obtain preprocessed document interaction record data and preprocessed document metadata; Wherein, the document interaction record data is obtained from the user, and the document metadata is obtained from a third-party document database; S2. Based on the preprocessed document metadata, a knowledge graph is constructed, and the preprocessed document interaction record data and entities in the knowledge graph are linked to obtain a final knowledge graph; S3, inputting the final knowledge graph into a knowledge graph embedding model, and outputting a first embedding vector of each document in the document interaction record data through the knowledge graph embedding model; S4, inputting the title and abstract of each document in the document interaction record data into a language model, and outputting a second embedding vector of each document in the document interaction record data through the language model; S5, fusing the first embedding vector and the second embedding vector to obtain a final embedding vector for each document in the document interaction record data; S6. Obtaining a user research intention representation vector according to the final embedding vector; S7. Obtain a recommended document set according to the user research intention representation vector and the recommended document list.

2. The method according to claim 1, characterized in that In the step S1, obtaining document interaction record data specifically includes: S11, obtaining the document interaction record of the user, wherein the document interaction record includes document download record, document collection record and document browsing history record; S12: taking the user ID, document title, document author and first document identifier in each document interaction record as the document interaction record data.

3. The method according to claim 1, characterized in that In step S1, obtaining document metadata specifically includes: S13, obtaining initial document metadata from the third-party document database; S14. The author keywords, extended keywords, research field, author, source journal and publisher information in the initial document metadata are used as the document metadata, and the document metadata is obtained by sorting according to the second document identifier and the number of citations in the initial document metadata.

4. The method according to claim 1, characterized in that: In step S2, constructing a knowledge graph based on the preprocessed document metadata specifically includes: S21, obtaining a first node according to the source journal and publisher information in the preprocessed document metadata; S22, splitting the author keywords, extended keywords, research fields and authors in the preprocessed document metadata to obtain a second node; S23, removing duplicates from the first node and the second node to obtain a final node; S24. According to the final node, establish an inArea many-to-many relationship, a researchBy many-to-many relationship, a masterIn many-to-many relationship, and a publishBy many-to-many relationship respectively; Among them, the inArea many-to-many relationship is established based on the relationship between the author keywords, the extended keywords and the research fields in the final node, the researchBy many-to-many relationship is established based on the relationship between the research field and the author in the final node, the masterIn many-to-many relationship is established based on the relationship between the research field and the source journal in the final node, and the publishBy many-to-many relationship is established based on the relationship between the source journal and the publisher information in the final node; S25. Establish the knowledge graph according to the established inArea many-to-many relationship, the researchBy many-to-many relationship, the masterIn many-to-many relationship and the publishBy many-to-many relationship.

5. The method according to claim 1, characterized in that In step S3, the knowledge graph embedding model is a TransE model, and the data input form of the knowledge graph embedding model is a triple; The activation function of the knowledge graph embedding model is: in, represents the first embedding vector of the document, n represents the total number of nodes related to the current document, and x i represents the feature vector of the i-th node, W i represents the weight of the feature vector of the i-th node, sigmoid represents the sigmoid activation function, and b1 represents the regular term when learning the first embedding vector.

6. The method according to claim 1, characterized in that In step S4, the language model is a Bert model, and the activation function of the language model is: in, represents the second embedding vector of the document, W(x title ,x abstract ) represents the weight value assigned to the concatenation of the feature vector of the title and the feature vector of the summary, and b2 represents the regularization term when learning the second embedding vector.

7. The method according to claim 1, characterized in that In step S5, the activation function used to fuse the first embedding vector and the second embedding vector is: Among them, agg P represents the final embedding vector of the document, represents the weight value assigned to the concatenation of the first embedding vector and the second embedding vector, and b3 represents the regularization term when the first embedding vector and the second embedding vector are fused into the final embedding vector.

8. The method according to claim 1, characterized in that In step S6, the activation function used to obtain the user research intention representation vector based on the final embedding vector is: agg U =sigmoid(W down P down +W star P star +W click P click +b4) Among them, agg U represents the user research intention representation vector, P down Represents the final embedding vector corresponding to the download interaction behavior, W down represents the weight of the final embedding vector corresponding to the download interaction behavior, P star Represents the final embedding vector corresponding to the collection interaction behavior, W star represents the weight of the final embedding vector corresponding to the collection interaction behavior, P click Represents the final embedding vector corresponding to the click interaction behavior, W click represents the weight of the final embedding vector corresponding to the click interaction behavior, and b4 represents the regularization term when learning the user research intention representation vector.

9. The method according to claim 1, characterized in that: The step S7 specifically includes: S71, determining a final embedding vector of the recommended document for each piece of data in the recommended document list according to the recommended document list; S72, calculating the cosine similarity between the user research intention representation vector and the final embedding vector of each recommended document, and obtaining the recommendation value of each data in the recommended document list; S73, determining an initial document recommendation set according to the data whose recommendation value is greater than a preset threshold; S74. Sort the initial document recommendation set according to the number of citations of each piece of data in the initial document recommendation set to obtain the recommended document set.

Images