Document personalized recommendation method and system based on knowledge graph

Abstract

The invention provides a personalized document recommendation method and system based on a knowledge graph, and relates to the technical field of document recommendation, and the method comprises the steps: obtaining a local document, and preprocessing the text content of the document; inputting the preprocessed document text content into a bidirectional recurrent neural network, extracting text features, inputting the text features into a CRF model again, obtaining a classification result of each word, calculating scores of whole text sentences, obtaining entity types, performing entity relationship extraction on each entity, and constructing a text knowledge graph; the method comprises the steps of obtaining historical behavior data of a user, constructing an interest preference behavior analysis model of the user, establishing a semantic matching matrix based on a text knowledge graph, performing matching degree calculation on user interests and document resource entity attributes, obtaining an interest matching matrix corresponding to a document resource entity and a semantic relationship, and obtaining user interest indexes; and sorting different standard documents according to the user interest indexes to generate a personalized document recommendation list.

Description

Technical Field

[0001] The present disclosure relates to the technical field of document recommendation, and in particular to a method and system for personalized document recommendation based on a knowledge graph. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] Due to the rapid development of information technology, the generation of massive information has brought people a variety of document resources, but it has also made how to quickly and accurately obtain documents of interest to users from massive resources a problem to be optimized.

[0004] The traditional recommendation method is the text recommendation algorithm, which mainly includes content recommendation and collaborative filtering recommendation. Content recommendation is to make recommendations based on the content information of the document by calculating the matching degree, but the recommendation results are too homogeneous and lack diversity, and content-based recommendation has limitations and is not easy to extract the features of text content; collaborative filtering recommendation analyzes the user's behavior and compares it with other users to find similar user groups for recommendation. This method requires a large number of users as a basis, has a cold start problem, and the recommendation is highly dependent on user relationships. New users cannot get accurate recommendation resources and the accuracy of the recommendation results is poor. Summary of the invention

[0005] In order to solve the above problems, the present disclosure proposes a document personalized recommendation method and system based on knowledge graph, which predicts user behavior preferences, retrieves relevant documents based on knowledge graph, and recommends them to users after analyzing and filtering the relevant documents.

[0006] According to some embodiments, the present disclosure adopts the following technical solutions:

[0007] The document personalized recommendation method based on knowledge graph includes:

[0008] Obtain local documents, extract document text content, and preprocess the document text content;

[0009] The preprocessed document text content is input into the bidirectional recurrent neural network to extract text features, and the text features are then input into the CRF model to obtain the classification results of each word. The entity type is obtained by calculating the score of the entire text sentence, and the entity relationship is extracted for each entity to obtain the entity relationship and attributes, and construct a text knowledge graph;

[0010] Obtain the user's historical behavior data, build a user interest preference behavior analysis model, establish a semantic matching matrix based on the text knowledge graph, calculate the matching degree between the user's interests and the document resource entity attributes, obtain the interest matching matrix corresponding to the document resource entity and the semantic relationship, and obtain the user's interest index;

[0011] Different standard documents are sorted according to the user interest index to generate a personalized document recommendation list.

[0012] According to some embodiments, the present disclosure adopts the following technical solutions:

[0013] The document personalized recommendation system based on knowledge graph includes:

[0014] The data acquisition module is used to acquire local documents, extract document text content, and pre-process the document text content;

[0015] The knowledge graph construction module is used to input the preprocessed document text content into the bidirectional recurrent neural network, extract text features, and then input the text features into the CRF model to obtain the classification results of each word. By calculating the score of the entire text sentence, the entity type is obtained, and the entity relationship is extracted for each entity, the entity relationship and attributes are obtained to construct a text knowledge graph;

[0016] The prediction module is used to obtain the user's historical behavior data, build a user interest preference behavior analysis model, establish a semantic matching matrix based on the text knowledge graph, calculate the matching degree between the user's interests and the document resource entity attributes, obtain the interest matching matrix corresponding to the document resource entity and the semantic relationship, and obtain the user's interest index;

[0017] The recommendation module is used to sort different standard documents according to the user interest index and generate a personalized document recommendation list.

[0018] According to some embodiments, the present disclosure adopts the following technical solutions:

[0019] An electronic device comprises: a processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes the document personalized recommendation method based on the knowledge graph.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] The document personalized recommendation method based on knowledge graph disclosed in the present invention will encounter a cold start problem for new users due to the lack of historical data of the user. Popular information can be recommended first instead of personalized recommendation, or when the user logs in for the first time, the user customizes and selects personal interest tags, and documents are recommended based on the tag content; for users who have used it for a period of time, an interest preference behavior analysis model of the user is constructed based on the user's historical behavior data (including search keywords, reading time frequency, likes and comments interactive content, uploaded and downloaded document content types, browsing time period and other indicators), and a personalized recommendation list is generated based on the correlation between user interests and knowledge graphs. The method has strong scalability. With the continuous improvement and expansion of knowledge graphs, the recommendation effect will become better and better, and the application scenarios are wide, and it can be applied to search engines, personalized reading, online learning and other fields to help users quickly acquire the required knowledge and improve the efficiency and accuracy of information acquisition.

[0022] The document personalized recommendation method based on knowledge graph disclosed in the present invention constructs a document knowledge graph and organizes it in the form of a graph, wherein entities are the basic units constituting the graph, referring to things or concepts in the real world, and they are organized in the form of a graph. In this method, entities represent users and documents, nodes represent entities or concepts, and edges represent various relationships between entities or concepts. The structured representation of the knowledge graph makes the recommendation results more accurate and reliable; the efficiency of recommendation can be greatly improved through rapid query and matching of the knowledge graph; and a personalized recommendation list is generated based on the association between user interests and the knowledge graph. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings constituting a part of the present disclosure are used to provide a further understanding of the present disclosure. The illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation on the present disclosure.

[0024] Figure 1 The figure is a flow chart of the personalized recommendation method according to an embodiment of the present disclosure.

[0025] Figure 2 This is the structure diagram of the bidirectional recurrent neural network. DETAILED DESCRIPTION

[0026] The present disclosure is further described below in conjunction with the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present disclosure belongs.

[0028] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0029] Example 1

[0030] In one embodiment of the present disclosure, a document personalized recommendation method based on a knowledge graph is provided, comprising:

[0031] Step 1: Obtain a local document, extract the document text content, and preprocess the document text content;

[0032] Step 2: Input the preprocessed document text content into the bidirectional recurrent neural network to extract text features, and then input the text features into the CRF model to obtain the classification results of each word. By calculating the score of the entire text sentence, the entity type is obtained, and the entity relationship is extracted for each entity. The entity relationship and attributes are obtained to construct a text knowledge graph;

[0033] Step 3: Obtain the user's historical behavior data, build a user interest preference behavior analysis model, establish a semantic matching matrix based on the text knowledge graph, calculate the matching degree between the user's interests and the document resource entity attributes, obtain the interest matching matrix corresponding to the document resource entity and the semantic relationship, and obtain the user's interest index;

[0034] Step 4: Sort different standard documents according to the user interest index to generate a personalized document recommendation list.

[0035] As an embodiment, the present invention discloses a document personalized recommendation method based on knowledge graph, which predicts user behavior preferences, retrieves relevant documents based on knowledge graph, analyzes and filters document features, and finally recommends documents to users. The specific implementation process is as follows:

[0036] Step 1: Construction of knowledge graph

[0037] A knowledge graph is a semantic network that reveals the relationships between entities and can formally describe real-world objects and their relationships.

[0038] A knowledge graph is a heterogeneous graph with multiple types of nodes and edges. Entities are the basic units that make up a graph, and refer to things or concepts in the real world. They are organized in the form of a graph, where nodes represent entities or concepts, and edges represent various relationships between entities or concepts.

[0039] A common representation of knowledge graphs is triples, which can be defined as G = (E, R, S). 1 ,e 2 ,…} represents the set of all entities in the knowledge graph; R = {r 1 ,r 2 ,…} represents the set of all relations in the knowledge graph; S = {(h,r,t)} represents the set of triples, and a single triple consists of a head entity h, a relation r, and a tail entity t.

[0040] Furthermore, the steps of constructing the knowledge graph include:

[0041] 1) Upload local documents and extract the document text content;

[0042] 2) Preprocessing of document text content, including:

[0043] (1) Unified text format;

[0044] (2) Clean the text data and remove illegal characters in the text content through rules; use the HanLP word segmentation tool to segment the unstructured text data and delete characters that have no actual meaning;

[0045] (3) Data labeling: matching and labeling text data through automated scripts.

[0046] 3) Entity recognition: Entity recognition is a process of automatically identifying and extracting entities with specific meanings (such as names of people, places, and names of organizations) from unstructured texts through natural language processing technology. The task of entity recognition can be divided into two main aspects: one is to determine which parts of the text constitute entities, that is, entity boundary recognition; the other is to determine the types of these entities, such as names of people, places, etc. These two tasks are usually combined to achieve accurate entity recognition. Specifically include:

[0047] First, a bidirectional recurrent neural network is used to represent the input text and extract text features such as the number of characters, the number of sentences, the length of each word, and the number of punctuation marks. Then, the text features are input into the CRF model to classify each word in the sentence. Finally, the score of the entire sentence is calculated, where the classification label is usually a combination of entity type and BIO labels, and the final classification result is output.

[0048] A bidirectional recurrent neural network is a neural network that can process sequence data and consider both past and future information in a sequence data. The basic idea is to propose that each training sequence is two recurrent neural networks, one forward and one backward, and both are connected to an output layer. This structure provides the output layer with complete past and future context information for each point in the input sequence. The structure of a bidirectional recurrent neural network is as follows Figure 2 shown.

[0049] Among them, the input is a sequence, the forward RNN processes the input sequence from front to back, the reverse RNN processes the input sequence from back to front, and the output is the combination of the outputs of the two RNN networks.

[0050] In natural language processing, bidirectional recurrent neural networks can process left-to-right and right-to-left text sequences, improving the quality of text representation and thus improving model performance.

[0051] The CRF (Conditional Random Field) model is a basic model in natural language processing and is widely used in scenarios such as word segmentation, entity recognition, and part-of-speech tagging. 1 ,x 2 ,...,x n ) represents the observed input data sequence, y=(y 1 ,y,...,y n ) represents a state sequence. Given an input sequence, the linear chain CRF model defines the joint conditional probability formula of the state sequence as follows:

[0052]

[0053] where Z(x) is the probability normalization factor conditioned on the observed sequence x; f k (y,x) is an arbitrary characteristic function; λ k is the weight of each feature function.

[0054] By establishing a conditional probability distribution between the input sequence and the output series, problems such as labeling bias can be solved. Use the extracted features to train a conditional random field model. The model can learn the dependencies between words and how to classify words based on contextual information. After the model training is completed, the CRF model can be used to classify each word in a new sentence. The model will score the possibility of each word based on the contextual information and the structure of the entire sentence, and finally assign a category with the highest score, that is, the most likely category.

[0055] The BIO tagging method is to label each word in a sentence with a label. This label consists of two parts: one part is the position of the entity to which the word belongs, where B indicates that the word is the first word of the entity, I indicates that the word is the middle word of the entity, and O indicates that it is not an entity; the other part is the entity type corresponding to the word.

[0056] 3) Entity relationship extraction: Relationship extraction is to extract the relationship between entities and associate scattered knowledge. Relationship extraction is achieved through dependency syntactic analysis of text content. The results of word segmentation and part-of-speech tagging are input into the syntactic analysis model to obtain the relationship between all words in the sentence. Through the dependency syntactic analysis model, the subject-predicate relationship, verb-object relationship, etc. in the sentence can be identified, and matched with commonly used sentence relationship templates to obtain a triple set of entities, relationships, and attributes. For example, the relationship between the triples is as follows:

[0057] As an example, for example, xx University is located in a certain city, where the entity is xx University, the relationship is located in, and its attribute value is: a certain city.

[0058] 4) Entity disambiguation: Entity recognition and entity relationship extraction results are disambiguated. A clustering algorithm is used to make each category in the clustering result correspond to a target entity.

[0059] The specific clustering algorithm used is the K-means algorithm. The basic idea of ​​the algorithm is: select K categories and K initial cluster centers, assign each sample to one of the K categories according to the minimum distance principle, iteratively calculate the class center, adjust the sample category, and finally minimize the sum of the squares of the distances from each pattern to the center of the category. It can be expressed as:

[0060]

[0061] Among them, w ik is the prompt function. If word x i belongs to a cluster, then w ik =1, otherwise w ik =0, c k is the center point of the cluster, that is, the average value of all words belonging to the cluster. This objective function minimizes the sum of the squares of the distances of all word points to the center of the cluster to which they belong.

[0062] 5) Use the Neo4j graph database to import the generated text triple data to form the corresponding text knowledge graph.

[0063] Step 2: Predict user interests

[0064] Specifically, the steps to extract user interest preferences are as follows:

[0065] For new users, there will be a cold start problem due to the lack of historical data about the user. Instead of personalized recommendations, popular information can be recommended first, or when the user logs in for the first time, the user can customize the personal interest tags and recommend documents based on the tag content;

[0066] For users who have used the app for a period of time, we build an interest preference behavior analysis model based on their historical behavior data (including search keywords, reading time frequency, likes and comments, interactive content, uploaded and downloaded document content types, browsing time periods, and other indicators):

[0067] Step 1: Data collection and preprocessing

[0068] The data on user search keywords, reading time, likes and comments, document upload and download, browsing time periods, etc. are collected. Through data cleaning, invalid or duplicate data is removed, different types of data are standardized, and useful features are initially extracted.

[0069] Step 2: Feature extraction and selection

[0070] Use word embedding technology (such as World2Vector or BERT) to convert search keywords into vectors, generate user behavior features based on user reading time, likes and comment frequencies, analyze user browsing time periods, convert time into periodic features (such as hours, days of the week), classify uploaded and downloaded documents, and generate classification features.

[0071] Step 3: Model selection and training

[0072] Select a deep learning model (bidirectional recurrent neural network), train the model using historical behavior data, and select an appropriate loss function for optimization.

[0073] Step 4: Model evaluation and optimization

[0074] Use indicators such as accuracy and recall to evaluate the performance of the model, and use methods such as network search or random search to optimize the model.

[0075] Step 5: Model deployment and application

[0076] Deploy the trained model, monitor user behavior in real time and update the model, and push content that matches users' interests and preferences through the recommendation system.

[0077] Through the above steps, we can build an interest preference behavior analysis model based on user historical behavior data. This model not only helps to understand user needs more accurately, but also improves user experience and increases user stickiness.

[0078] Furthermore, a semantic matching matrix based on the knowledge graph is established. The triple of the knowledge graph is (h, r, t), and the vector representation of the resource entity in the semantic relationship space is: In relational space, by Mapping to the entity space and adjusting it, we can get the entity vector In the process of semantic matching, assuming that the number of hidden layers of the knowledge graph is k, the document resource entity at the kth layer is matched with the user's interest. The similarity between the document resource entity at the kth layer and the user's interest is h k The calculation formula is:

[0079] h k =f(h,t)k (2)

[0080] According to the document resource entity and semantic relationship, this method sets the weight of the feature vector, and obtains the interest matching matrix corresponding to the document resource entity and semantic relationship by calculating the matching degree between the user interest and the document resource entity attribute. The matching degree M between the user interest and the document resource entity attribute is calculated as:

[0081] M=αsim(h,t)+βsim(h,t)+χsim(r,t) (3)

[0082] Among them, α, β, χ are weight coefficients; sim(·,·) represents the interest similarity between the two. represents the interest similarity between the two. The above formula can be used to calculate the user's interest preference matching degree for different document resources, predict user behavior, and obtain the user's interest index at this time.

[0083] Step 3: Generate a personalized document recommendation list

[0084] Specifically, after obtaining the candidate standard documents with the highest interest index, different standard documents are sorted. For a large number of standard documents, different standard documents and related entities are divided into fields according to different standard types and professional fields. According to the probability distribution of different fields, standard documents in different fields are given weights respectively, and the document resources are scored in combination with their influence.

[0085] The scoring formula is:

[0086]

[0087] where w i is the weight of the ith field, f i It is the characteristic value of the document in this field (such as the number of visits, number of likes, etc.), and the final score of each document is calculated based on the comprehensive weight.

[0088] Resources are sorted by their final scores, and a personalized document recommendation list is generated for the top-ranked resources.

[0089] Example 2

[0090] In one embodiment of the present disclosure, a document personalized recommendation system based on a knowledge graph is provided, including:

[0091] The data acquisition module is used to acquire local documents, extract document text content, and pre-process the document text content;

[0092] The knowledge graph construction module is used to input the preprocessed document text content into the bidirectional recurrent neural network, extract text features, and then input the text features into the CRF model to obtain the classification results of each word. By calculating the score of the entire text sentence, the entity type is obtained, and the entity relationship is extracted for each entity, the entity relationship and attributes are obtained to construct a text knowledge graph;

[0093] The prediction module is used to obtain the user's historical behavior data, build a user interest preference behavior analysis model, establish a semantic matching matrix based on the text knowledge graph, calculate the matching degree between the user's interests and the document resource entity attributes, obtain the interest matching matrix corresponding to the document resource entity and the semantic relationship, and obtain the user's interest index;

[0094] The recommendation module is used to sort different standard documents according to the user interest index and generate a personalized document recommendation list.

[0095] Example 3

[0096] An embodiment of the present disclosure provides an electronic device, including: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the document personalized recommendation method based on the knowledge graph.

[0097] The present disclosure is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.

[0098] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0099] Although the above describes the specific implementation methods of the present disclosure in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present disclosure. Technical personnel in the relevant field should understand that on the basis of the technical solution of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative work are still within the scope of protection of the present disclosure.

Claims

1. Personalized document recommendation method based on knowledge graph, It is characterized in that include: Obtain local documents, extract document text content, and preprocess the document text content; The preprocessed document text content is input into the bidirectional recurrent neural network to extract text features, and the text features are then input into the CRF model to obtain the classification results of each word. The entity type is obtained by calculating the score of the entire text sentence, and the entity relationship is extracted for each entity to obtain the entity relationship and attributes, and construct a text knowledge graph; Obtain the user's historical behavior data, build a user interest preference behavior analysis model, establish a semantic matching matrix based on the text knowledge graph, calculate the matching degree between the user's interests and the document resource entity attributes, obtain the interest matching matrix corresponding to the document resource entity and the semantic relationship, and obtain the user's interest index; Different standard documents are sorted according to the user interest index to generate a personalized document recommendation list.

2. The document personalized recommendation method based on knowledge graph as claimed in claim 1, It is characterized in that Preprocessing of document text content includes: unifying text format, cleaning text data, cleaning illegal characters in text content through rules, using HanLP word segmentation tool to segment unstructured text data, deleting characters without actual meaning, and matching and annotating text content through automated scripts.

3. The document personalized recommendation method based on knowledge graph as claimed in claim 1, It is characterized in that The knowledge graph is represented by a triple, defined as G = (E, R, S), where E = {e 1 ,e 2 ,…} represents the set of all entities in the knowledge graph; R = {r 1 ,r 2 ,…} represents the set of all relations in the knowledge graph; S = {(h,r,t)} represents the set of triples, and a single triple consists of a head entity h, a relation r, and a tail entity t.

4. The document personalized recommendation method based on knowledge graph according to claim 1, It is characterized in that A bidirectional recurrent neural network is used to represent the input text content, obtain text features, and then input the text features into the CRF model to classify each word in the sentence. Finally, the score of the entire sentence is calculated, where the classification label is a combination of three labels: entity type and BIO, thereby outputting the final entity classification result.

5. The document personalized recommendation method based on knowledge graph according to claim 1, It is characterized in that Extract the relationships between entities, associate scattered knowledge, extract relationships through dependency syntactic analysis of text content, match them together, and obtain a ternary set of entities, relationships, and attributes.

6. The document personalized recommendation method based on knowledge graph according to claim 1, It is characterized in that A clustering algorithm is used for entity disambiguation, so that each category in the clustering results corresponds to a target entity, including entity referent similarity calculation based on surface features, entity referent similarity calculation based on extended features, and entity referent similarity calculation based on social networks.

7. The document personalized recommendation method based on knowledge graph according to claim 1, It is characterized in that A semantic matching matrix based on the knowledge graph is established. The triple of the knowledge graph is (h, r, t), and the vector of the resource entity in the semantic relationship space is represented as h→r. In the relationship space, h→r is mapped to the entity space for adjustment to obtain the entity vector t→r. In the process of semantic matching, the number of hidden layers of the knowledge graph is set to k, and the document resource entity at the kth layer is matched with the user interest.

8. The document personalized recommendation method based on knowledge graph according to claim 1, It is characterized in that According to the document entity and semantic relationship, the weight of the feature vector is set, and the interest matching matrix corresponding to the document resource entity and semantic relationship is obtained by calculating the matching degree between the user interest and the document entity attribute. The matching degree calculation formula between the user interest and the document entity attribute is: M=αsim(h,t)+βsim(h,t)+χsim(r,t) Among them, α, β, and χ are weight coefficients; sim(·,·) represents the interest similarity between the two; the user’s interest preference matching degree for different document resources is calculated, user behavior is predicted, and user interest indicators are obtained.

9. Personalized document recommendation system based on knowledge graph, It is characterized in that include: The data acquisition module is used to acquire local documents, extract document text content, and pre-process the document text content; The knowledge graph construction module is used to input the preprocessed document text content into the bidirectional recurrent neural network, extract text features, and then input the text features into the CRF model to obtain the classification results of each word. By calculating the score of the entire text sentence, the entity type is obtained, and the entity relationship is extracted for each entity, the entity relationship and attributes are obtained to construct a text knowledge graph; The prediction module is used to obtain the user's historical behavior data, build a user interest preference behavior analysis model, establish a semantic matching matrix based on the text knowledge graph, calculate the matching degree between the user's interests and the document resource entity attributes, obtain the interest matching matrix corresponding to the document resource entity and the semantic relationship, and obtain the user's interest index; The recommendation module is used to sort different standard documents according to the user interest index and generate a personalized document recommendation list.

10. An electronic device, It is characterized in that include: A processor, a memory and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory so that the electronic device executes the document personalized recommendation method based on the knowledge graph as described in any one of claims 1-8.

Images