Non-offensive news recommendation method, device and equipment and storage medium

Abstract

The invention discloses a non-offensive news recommendation method, device and equipment and a storage medium, and belongs to the technical field of information processing.The method comprises the steps that on the basis of a BERT-BiLSTM-CRM model and a first non-offensive knowledge graph, a title feature vector and a keyword feature vector of each non-offensive news in a non-offensive news set are obtained, the keyword feature vector is extracted from the text of the non-abandoned news; processing the title feature vector and the keyword feature vector of each piece of non-abandoned news by adopting an attention mechanism to obtain a feature vector of each piece of non-abandoned news; determining a user preference feature vector corresponding to each piece of candidate news based on the feature vectors of multiple pieces of non-abandoned news in the non-abandoned news set; and based on the user preference feature vector corresponding to each candidate news, pushing non-missing news to the user. The method can accurately recommend non-abandoned news according to user preferences.

Description

Technical Field

[0001] The present disclosure relates to the field of information processing technology, and in particular to a method, device, equipment and storage medium for recommending intangible cultural heritage news. Background Art

[0002] As an important part of China's excellent traditional culture, intangible cultural heritage (abbreviated as "ICH") carries rich historical and cultural connotations and Chinese cultural genes.

[0003] Since intangible cultural heritage news contains a large number of proper noun entities and special entity relationships, and there is a lack of intangible cultural heritage thematic corpus, it is difficult for traditional news recommendation algorithms to accurately recommend intangible cultural heritage news based on user preferences. Summary of the invention

[0004] The present disclosure provides a method, device, equipment and storage medium for recommending intangible cultural heritage news, which can accurately recommend intangible cultural heritage news according to user preferences. The technical solution at least includes the following solutions:

[0005] In a first aspect, a method for recommending intangible cultural heritage news is provided, comprising: based on a BERT-BiLSTM-CRM model and a first intangible cultural heritage knowledge graph, obtaining a title feature vector and a keyword feature vector of each intangible cultural heritage news in an intangible cultural heritage news set, wherein the keyword feature vector is extracted from the body of the intangible cultural heritage news; using an attention mechanism to process the title feature vector and the keyword feature vector of each intangible cultural heritage news to obtain a feature vector of each intangible cultural heritage news; based on the feature vectors of multiple intangible cultural heritage news in the intangible cultural heritage news set, determining a user preference feature vector corresponding to each candidate news, wherein the candidate news is the intangible cultural heritage news in the intangible cultural heritage news set that has not been clicked by the user; and based on the user preference feature vector corresponding to each candidate news, pushing the intangible cultural heritage news to the user.

[0006] Optionally, the method of obtaining a title feature vector and a keyword feature vector of each intangible cultural heritage news in the intangible cultural heritage news set based on the BERT-BiLSTM-CRM model and the first intangible cultural heritage knowledge graph includes: converting the title of the first intangible cultural heritage news into multiple first word vectors, where the first intangible cultural heritage news is any one of the multiple intangible cultural heritage news; obtaining multiple first entities in the title of the first intangible cultural heritage news and a first entity vector of each first entity based on the BERT-BiLSTM-CRM model; obtaining multiple first entity context vectors based on the first intangible cultural heritage knowledge graph and the multiple first entities, where each first entity context vector corresponds to one first entity; and inputting the multiple first word vectors, the multiple first entity vectors and the multiple entity context vectors into a knowledge-aware convolutional neural network KCNN to obtain the title feature vector of the first intangible cultural heritage news.

[0007] Optionally, the method of obtaining the title feature vector and keyword feature vector of each intangible cultural heritage news in the intangible cultural heritage news set based on the BERT-BiLSTM-CRM model and the first intangible cultural heritage knowledge graph also includes: obtaining a text keyword sequence of the first intangible cultural heritage news, wherein the text keyword sequence includes multiple keywords in the text of the first intangible cultural heritage news; converting the multiple keywords in the text keyword sequence into multiple second word vectors, and obtaining multiple second entities in the text keyword sequence of the first intangible cultural heritage news and a second entity vector of each second entity based on the BERT-BiLSTM-CRM model; obtaining multiple second entity context vectors based on the first intangible cultural heritage knowledge graph and the multiple second entities, each second entity context vector corresponding to one second entity; inputting the multiple second word vectors, the multiple second entity vectors and the multiple entity context vectors into a knowledge-aware convolutional neural network KCNN to obtain the keyword feature vector of the first intangible cultural heritage news.

[0008] Optionally, the using of the attention mechanism to process the title feature vector and the keyword feature vector of each intangible cultural heritage news to obtain the feature vector of each intangible cultural heritage news includes: using the following formula to determine the feature vector of each intangible cultural heritage news:

[0009] r n =α t r t +α k r k

[0010] Among them, r n is the feature vector of intangible cultural heritage news, r t is the title feature vector of the intangible cultural heritage news, r k is the keyword feature vector of intangible cultural heritage news, a t is the attention weight of the title feature vector, a k is the attention weight of the keyword feature vector.

[0011] Optionally, determining the user preference feature vector corresponding to each candidate news based on the feature vectors of multiple intangible cultural heritage news in the intangible cultural heritage news set includes: using the following formula to determine the user preference vector corresponding to the j-th candidate news:

[0012]

[0013] Among them, e(i) is the user preference vector of user i for the jth candidate news, is the feature vector of the kth intangible cultural heritage news clicked by the user i. The user i has clicked N times in total. i A piece of non-legacy news, is the weight matrix, t j is the j-th candidate news;

[0014] The weight matrix is ​​calculated using the following formula:

[0015]

[0016] in, is the weight matrix, softmax is the normalization function, e(t j ) is the feature vector of the j-th candidate news.

[0017] Optionally, the pushing of intangible cultural heritage news to the user based on the user preference feature vector corresponding to each candidate news includes: determining the user's preference degree for each candidate intangible cultural heritage news based on the user preference feature vector corresponding to each candidate news; sorting the candidate intangible cultural heritage news in descending order according to the degree of preference of each candidate news to obtain a first preference order; and pushing the first N candidate news in the first preference order to the user.

[0018] In the second aspect, a non-heritage news recommendation device is also provided, including: an acquisition module, used to acquire the title feature vector and keyword feature vector of each non-heritage news in the non-heritage news set based on the BERT-BiLSTM-CRM model and the first non-heritage knowledge graph, and the keyword feature vector is extracted from the text of the non-heritage news; an attention mechanism module, used to use the attention mechanism to process the title feature vector and the keyword feature vector of each non-heritage news to obtain the feature vector of each non-heritage news; a user preference feature vector calculation module, used to determine the user preference feature vector corresponding to each candidate news based on the feature vectors of multiple non-heritage news in the non-heritage news set, and the candidate news is the non-heritage news in the non-heritage news set that has not been clicked by the user; a push module, used to push the non-heritage news to the user based on the user preference feature vector corresponding to each candidate news.

[0019] Optionally, the acquisition module is also used to convert the title of the first intangible cultural heritage news into multiple first word vectors, where the first intangible cultural heritage news is any one of the multiple intangible cultural heritage news; based on the BERT-BiLSTM-CRM model, obtain multiple first entities in the title of the first intangible cultural heritage news and the first entity vector of each first entity; based on the first intangible cultural heritage knowledge graph and the multiple first entities, obtain multiple first entity context vectors, each first entity context vector corresponds to one first entity; input the multiple first word vectors, the multiple first entity vectors and the multiple entity context vectors into the knowledge-aware convolutional neural network KCNN to obtain the title feature vector of the first intangible cultural heritage news.

[0020] Optionally, the acquisition module is also used to obtain a text keyword sequence of the first intangible cultural heritage news, wherein the text keyword sequence includes multiple keywords in the text of the first intangible cultural heritage news; convert the multiple keywords in the text keyword sequence into multiple second word vectors, and based on the BERT-BiLSTM-CRM model, obtain multiple second entities in the text keyword sequence of the first intangible cultural heritage news and the second entity vector of each second entity; based on the first intangible cultural heritage knowledge graph and the multiple second entities, obtain multiple second entity context vectors, each second entity context vector corresponds to one second entity; input the multiple second word vectors, the multiple second entity vectors and the multiple entity context vectors into a knowledge-aware convolutional neural network KCNN to obtain the keyword feature vector of the first intangible cultural heritage news.

[0021] Optionally, the attention mechanism module is further used to determine the feature vector of each intangible cultural heritage news using the following formula:

[0022] r n =α t r t +α k r k

[0023] Among them, r n is the feature vector of intangible cultural heritage news, r t is the title feature vector of the intangible cultural heritage news, r k is the keyword feature vector of intangible cultural heritage news, a t is the attention weight of the title feature vector, a k is the attention weight of the keyword feature vector.

[0024] Optionally, the user preference feature vector calculation module is further used to determine the user preference vector corresponding to the j-th candidate news by using the following formula:

[0025]

[0026] Among them, e(i) is the user preference vector of user i for the jth candidate news, is the feature vector of the kth intangible cultural heritage news clicked by the user i. The user i has clicked N times in total. i A piece of non-legacy news, is the weight matrix, t j is the j-th candidate news;

[0027] The weight matrix is ​​calculated using the following formula:

[0028]

[0029] in, is the weight matrix, softmax is the normalization function, e(t j ) is the feature vector of the j-th candidate news.

[0030] Optionally, the push module is also used to determine the user's preference for each of the candidate intangible cultural heritage news based on the user preference feature vector corresponding to each of the candidate news; sort the candidate intangible cultural heritage news in descending order according to the degree of preference of each of the candidate news to obtain a first preference order; and push the first N candidate news in the first preference order to the user.

[0031] In a third aspect, a computer device is also provided, comprising: a memory and a processor, wherein at least one computer program is stored in the memory, and the at least one computer program is loaded and executed by the processor, thereby executing the intangible cultural heritage news push method described in the above embodiment.

[0032] In a fourth aspect, a computer-readable storage medium is also provided, in which at least one computer program is stored. The at least one computer program is loaded and executed by a processor, thereby executing the intangible cultural heritage news push method described in the above embodiment.

[0033] In a fifth aspect, a computer program product is provided, comprising a computer program / instruction, wherein when the computer program / instruction is executed by a processor, the method described in the first aspect is implemented.

[0034] The beneficial effects brought by the technical solution provided by the embodiments of the present disclosure include at least:

[0035] In the disclosed embodiment, although the news title is a high-level summary of the news content, some intangible cultural heritage news titles are relatively brief and contain less entity and semantic information. If only the news title is used for recommendation, the effect may not be ideal. Therefore, in the disclosed embodiment, based on the BERT-BiLSTM-CRM model and the first intangible cultural heritage knowledge graph, the title feature vector and keyword feature vector of each intangible cultural heritage news in the intangible cultural heritage news set are obtained, which is equivalent to using the title and the text of the intangible cultural heritage news for recommendation at the same time, so it is possible to accurately recommend intangible cultural heritage news according to user preferences. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0037] Figure 1 A flowchart of a method for pushing intangible cultural heritage news provided by an exemplary embodiment of the present disclosure is shown;

[0038] Figure 2 This is a schematic diagram of the first intangible cultural heritage knowledge graph;

[0039] Figure 3 A flowchart of a method for pushing intangible cultural heritage news provided by another exemplary embodiment of the present disclosure is shown;

[0040] Figure 4 This is a schematic diagram of the knowledge graph of the intangible cultural heritage;

[0041] Figure 5 This is a schematic diagram of the structure of KCNN;

[0042] Figure 6 This is a schematic diagram of the structure of the intangible cultural heritage news recommendation model;

[0043] Figure 7 A schematic diagram of the structure of a non-legacy news recommendation device provided by an exemplary embodiment of the present disclosure is shown;

[0044] Figure 8 It is a schematic diagram of the structure of a computer device provided in an embodiment of the present disclosure. DETAILED DESCRIPTION

[0045] Unless otherwise defined, the technical or scientific terms used herein shall have the usual meanings understood by persons of ordinary skill in the field to which the present disclosure belongs. The words "first", "second", "third" and similar words used in the patent application specification and claims of the present disclosure do not indicate any order, quantity or importance, but are only used to distinguish different components. Similarly, words such as "one" or "one" do not indicate a quantitative limitation, but rather indicate the presence of at least one. Words such as "include" or "comprise" and similar words mean that the elements or objects appearing before "include" or "comprises" cover the elements or objects listed after "include" or "comprises" and their equivalents, and do not exclude other elements or objects.

[0046] In order to make the objectives, technical solutions and advantages of the present disclosure more clear, the embodiments of the present disclosure will be further described in detail below with reference to the accompanying drawings.

[0047] Figure 1 A flowchart of a method for recommending intangible cultural heritage news provided by an exemplary embodiment of the present disclosure is shown, and the method can be executed by a computer device. Figure 1 , the method comprising:

[0048] In step 101, based on the BERT-BiLSTM-CRM model and the first intangible cultural heritage knowledge graph, the title feature vector and keyword feature vector of each intangible cultural heritage news in the intangible cultural heritage news set are obtained.

[0049] Among them, the keyword feature vector is extracted from the text of the intangible cultural heritage news.

[0050] The BERT-BiLSTM-CRM (Bidirectional Encoder Representations from Transformers-Bidirectional Long Short Term Memory-Conditional Random Fields) model can be divided into three parts: BERT, BiLSTM, and CRM. BERT is a bidirectional encoder representation from Transformer, which can understand complex relationships in text; BiLSTM has a strong context capture capability; and CRM has a sequence tagging advantage. Therefore, the BERT-BiLSTM-CRM model that combines the above three has a strong Chinese entity recognition capability, can reduce tagging errors, and effectively improve entity recognition accuracy.

[0051] The first intangible cultural heritage knowledge graph is a knowledge graph in the field of intangible cultural heritage. In the disclosed embodiment, the first intangible cultural heritage knowledge graph may be an existing intangible cultural heritage knowledge graph, or the first intangible cultural heritage knowledge graph is constructed by the following steps ac.

[0052] Step a, obtaining an intangible cultural heritage text data set from an intangible cultural heritage database, and screening out multiple intangible cultural heritage entities from the intangible cultural heritage text data set.

[0053] Optionally, the intangible cultural heritage database includes non-structured text information such as the "List of National Intangible Cultural Heritage Representative Items" and the "List and introduction of Representative Inheritors of National Intangible Cultural Heritage Representative Items" published on the official website "China Intangible Cultural Heritage Network·China Intangible Cultural Heritage Digital Museum" (referred to as China Intangible Cultural Heritage Network) sponsored by the Chinese Academy of Arts and the China Intangible Cultural Heritage Protection Center.

[0054] When obtaining the intangible cultural heritage text data set from the intangible cultural heritage database, all national project lists and inheritor lists, as well as a large amount of text data of intangible cultural heritage news can be crawled from the China Intangible Cultural Heritage Network based on a crawler framework, thereby obtaining the intangible cultural heritage text data set. The crawler framework can be Scrapy, a Python-based web crawler framework.

[0055] Before screening out multiple intangible cultural heritage entities from the intangible cultural heritage text data set, it is first necessary to define the types of intangible cultural heritage entities. In the disclosed embodiment, the intangible cultural heritage entities include intangible cultural heritage items, inheritors, time, geographical location, project category, project level, project batch, dynasty of origin, declared area and other types. Among them, intangible cultural heritage items, inheritors, time and geographical location can be first-level entities, and project category, project level, project batch, dynasty of origin, declared area and the like can be second-level entities. The importance of first-level entities is higher than that of second-level entities.

[0056] Optionally, the BERT-BiLSTM-CRM model is used to filter out multiple intangible cultural heritage entities from the intangible cultural heritage text data set.

[0057] Step b, obtaining entity relationships of multiple intangible cultural heritage entities.

[0058] Optionally, the Language Technology Platform (LTP) developed by the Harbin Institute of Technology team is used to obtain the entity relationships of multiple intangible cultural heritage entities. Finally, multiple entity relationship triples can be obtained, and the structure of the entity relationship triple is <entity, relationship, entity>.

[0059] Optionally, for any entity, you can also define the entity attributes of the entity. Click on the entity in the knowledge graph to view the entity attributes of the entity. Entity attributes can include data attributes and object attributes. For example, the entity of the type of intangible cultural heritage project includes data attributes such as project name, number, batch, category, level, application area, protection unit, dynasty of origin, and publication time; the object attributes of the entity of the type of intangible cultural heritage project include other related entities such as inheritors, project categories, project levels, and dynasty of origin, that is, the internal connection between entities, also known as relationship attributes.

[0060] Step c: constructing a first intangible cultural heritage knowledge graph based on multiple intangible cultural heritage entities and entity relationships between multiple intangible cultural heritage entities.

[0061] After obtaining the triplets of multiple entity relationships, the first intangible cultural heritage knowledge graph can be constructed based on these triplets.

[0062] For example, the open source graph database Neo4j is used to build the first intangible cultural heritage knowledge graph. Neo4j has the advantages of high performance, built-in visualization tools, efficient graph traversal, ability to process billions of nodes and relationships, and support for Cypher language.

[0063] When importing intangible cultural heritage entities, first store the identified intangible cultural heritage entities in CSV format files according to type classification, and then use the LOAD CSV command in the Cypher language to batch import the entities in the CSV format file into the Neo4j database to form a series of nodes. Then use the MATCH command in Cypher to query entities that meet specific conditions and create relationships between them, thereby obtaining the first intangible cultural heritage knowledge graph. For example, the statement "MATCH (from: project), (to: project_type) where from.type = to.type MERGE (from) - [r: project category] -> (to) return r" is to construct a relationship triplet of two types of entities, intangible cultural heritage projects and intangible cultural heritage categories, based on the common attribute "type" of the project entity "project" and the project category entity "project_type".

[0064] The collected text information of intangible cultural heritage items and inheritors was processed using the above steps ac, and then imported into the Neo4j graph database in batches to form an intangible cultural heritage knowledge graph, obtaining a total of 6731 nodes and 25523 relationships.

[0065] Figure 2 The first intangible cultural heritage knowledge graph is shown in Figure 2. Taking the intangible cultural heritage project “Legend of Mulan” in the database as the query node, the knowledge graph formed in Neo4j Browser is as follows: Figure 2 shown.

[0066] Optionally, the intangible cultural heritage news collection includes multiple published intangible cultural heritage news, some of which have been clicked by users and some have not been clicked by users. The embodiment of the present disclosure does not limit the number and source of the intangible cultural heritage news in the intangible cultural heritage news collection.

[0067] In step 102, the title feature vector and keyword feature vector of each intangible cultural heritage news are processed using an attention mechanism to obtain a feature vector of each intangible cultural heritage news.

[0068] In step 103, based on the feature vectors of multiple intangible cultural heritage news in the intangible cultural heritage news collection, a user preference feature vector corresponding to each candidate news is determined.

[0069] Among them, the candidate news is the intangible cultural heritage news in the intangible cultural heritage news collection that has not been clicked by the user.

[0070] In step 104, based on the user preference feature vector corresponding to each candidate news, the intangible cultural heritage news is pushed to the user.

[0071] In the disclosed embodiment, although the news title is a high-level summary of the news content, some intangible cultural heritage news titles are relatively brief and contain less entity and semantic information. If only the news title is used for recommendation, the effect may not be ideal. Therefore, in the disclosed embodiment, based on the BERT-BiLSTM-CRM model and the first intangible cultural heritage knowledge graph, the title feature vector and keyword feature vector of each intangible cultural heritage news in the intangible cultural heritage news set are obtained, which is equivalent to using the title and the text of the intangible cultural heritage news for recommendation at the same time, so it is possible to accurately recommend intangible cultural heritage news according to user preferences.

[0072] Figure 3 A flowchart of a method for recommending intangible cultural heritage news provided by another exemplary embodiment of the present disclosure is shown. The method can be executed by a computer device. Figure 3 , the method comprising:

[0073] In step 301, based on the BERT-BiLSTM-CRM model and the first intangible cultural heritage knowledge graph, the title feature vector and keyword feature vector of each intangible cultural heritage news in the intangible cultural heritage news set are obtained.

[0074] For the relevant contents of the BERT-BiLSTM-CRM model and the first intangible cultural heritage knowledge graph, please refer to the aforementioned step 101, and the detailed description is omitted here.

[0075] Optionally, the title feature vector of each intangible cultural heritage news is obtained using the following step dg.

[0076] Step d: convert the title of the first intangible cultural heritage news into multiple first word vectors.

[0077] The first intangible cultural heritage news is any one of the multiple intangible cultural heritage news.

[0078] Optionally, the title of the first intangible cultural heritage news is converted into multiple first word vectors based on the word2vec word vector representation method.

[0079] The word2vec word vector representation method includes the following four steps:

[0080] In the first step, a context window is selected to convert the target text into multiple text blocks.

[0081] The word2vec word vector representation method here is used to convert the title of the first intangible cultural heritage news into multiple first word vectors, so the target text here is the title text of the first intangible cultural heritage news.

[0082] Exemplarily, the context window is a sliding window of length 2C+1, which is convenient for capturing the context information of a word. Here, the context window or context window of a word is usually C, 2C is the sum of the context window and context window size of the word, and 1 represents the word itself.

[0083] In the second step, multiple text blocks are converted into word vectors to obtain multiple first context word vectors.

[0084] Optionally, the second step can be implemented by One-hot encoding. The core idea of ​​One-hot is to map each word into a pre-trained word vector space.

[0085] The third step is to obtain the vector of the hidden layer.

[0086] When implementing the third step, it is necessary to multiply multiple context word vectors by the shared weight matrix to obtain multiple intermediate vectors, and then perform weighted averaging on the multiple intermediate vectors to obtain the vector of the hidden layer.

[0087] The fourth step is to obtain the output vector based on the vector of the hidden layer and activate the output vector.

[0088] Optionally, the fourth step includes: performing a matrix multiplication operation on the vector of the hidden layer and the weight matrix to obtain an output vector.

[0089] Optionally, activation of the output vector can be implemented by a softmax function.

[0090] There are many implementation methods of weight matrix and shared weight matrix in the related technology, so the detailed description is omitted here.

[0091] The first to fourth steps above can exist in the form of a word vector model. For the word vector model, the gradient descent method can be used to optimize the minimum loss function by dynamically adjusting the model parameters (weight matrix). After the word vector model is trained, the word vector model can be used to generate word vectors.

[0092] Step e: Based on the BERT-BiLSTM-CRM model, multiple first entities in the title of the first intangible cultural heritage news and the first entity vector of each first entity are obtained.

[0093] The BERT-BiLSTM-CRM model can accurately identify Chinese entities, so based on the BERT-BiLSTM-CRM model, it is possible to obtain multiple first entities in the first intangible cultural heritage news title.

[0094] After obtaining multiple first entities, the multiple first entities can be converted into first entity vectors through TransD knowledge embedding technology.

[0095] There are many ways to implement TransD knowledge embedding technology in related technologies, so detailed description is omitted here.

[0096] Step f, based on the first intangible cultural heritage knowledge graph and multiple first entities, obtain multiple first entity context vectors.

[0097] Each first-entity context vector corresponds to a first entity.

[0098] After using the BERT-BiLSTM-CRM model to correctly identify the first entity in the first intangible cultural heritage news title, these first entities can be linked to the corresponding nodes of the first intangible cultural heritage knowledge graph through entity linking technology, thereby building a connection between the first intangible cultural heritage news title and the first intangible cultural heritage knowledge graph.

[0099] However, considering that using the complete intangible cultural heritage knowledge graph in intangible cultural heritage news recommendation will increase the model training cost and lead to a large amount of redundant information, the selective use of the knowledge graph is achieved through step f. In this case, step f includes the following two steps:

[0100] The first step is to construct a sub-intangible cultural heritage knowledge graph corresponding to each first entity in the first intangible cultural heritage knowledge graph to obtain multiple sub-intangible cultural heritage knowledge graphs.

[0101] The sub-intangible cultural heritage knowledge graph corresponding to any first entity is constructed in the following way: taking the first entity as the center and extending one layer outward, that is, considering all other entities and related edges with a distance of 1 from the entity, so as to obtain the sub-intangible cultural heritage knowledge graph corresponding to the first entity.

[0102] The entities other than the first entity in the sub-intangible cultural heritage knowledge graph corresponding to the first entity e can be expressed by formula (1).

[0103] context(e) = {e i |(e,r,e i )∈SG or(e i ,r,e)∈SG} (1)

[0104] In formula (1), context(e) is the entity other than the first entity in the sub-intangible cultural heritage knowledge graph corresponding to the first entity e, SG is the sub-intangible cultural heritage knowledge graph corresponding to the first entity e, and e i is an entity directly connected to the first entity e, and r represents a certain relationship between the head entity and the tail entity, for example (e, r, e i ) means that the head entity is the first entity e, and the tail entity is e i , the relationship between the head entity and the tail entity is r.

[0105] Figure 4 This is a schematic diagram of the sub-intangible cultural heritage knowledge graph. Figure 4 As shown in the figure, with the first entity "Huangmei Opera" as the center, one layer is extended outward, thus obtaining the sub-intangible cultural heritage knowledge graph corresponding to "Huangmei Opera" in the dotted line. This sub-intangible cultural heritage knowledge graph contains other entities with a distance of 1 from "Huangmei Opera", such as "traditional drama", "Huang Xinde", "Han Zaifen", "extension project", "Hefei City", "2011", and "national level".

[0106] The second step is to obtain multiple entity context vectors based on multiple sub-intangible cultural heritage knowledge graphs.

[0107] Among them, the first entity corresponding to the first entity context vector converted from any sub-intangible cultural heritage knowledge graph is the first entity at the center of the sub-intangible cultural heritage knowledge graph.

[0108] Optionally, an entity average method is used to generate an entity context vector representation, and the first entity context vector corresponding to the first entity e finally obtained can be expressed by formula (2).

[0109]

[0110] In formula (2), is the first entity context vector corresponding to the first entity e, e I For entity e i The vector form of entity e i The vector form of can be obtained through the TransD knowledge embedding technology, which is omitted here. The meanings of other parameters in formula (2) are the same as those in formula (1), which are omitted here.

[0111] Step g: input multiple first word vectors, multiple first entity vectors, and multiple entity context vectors into a knowledge-aware convolutional neural network KCNN to obtain a title feature vector of the first intangible cultural heritage news.

[0112] After obtaining multiple first entity vectors in the title of the first intangible cultural heritage news and multiple first entity context vectors corresponding to the multiple first entity vectors, it is necessary to organically combine the first entity vector, the first entity context vector and the first word vector through a splicing operation.

[0113] If the word vector is directly spliced ​​with the entity vector and the entity context vector, not only will the correlation between the two be destroyed, but the new sequence of word entities after splicing cannot express the alignment relationship between them. Therefore, the word vector, entity vector, and entity context vector need to be consistent in dimension during splicing. Although alignment can be achieved through Padding, it is not the best choice in practice. In order to overcome the limitations of traditional vector splicing, in the disclosed embodiment, the splicing operation is implemented by a convolutional neural network KCNN (Knowledge Convolutional Neural Networks) based on knowledge perception, and the multi-channel method of KCNN is used to solve the problem of inconsistent dimensions during splicing.

[0114] Figure 5 This is a schematic diagram of the KCNN structure. Figure 5 Step g is described.

[0115] Multiple first word vectors can be represented as the first word sequence News t =[s1,s2,s3,...s n ], where s1, s2, s3, ...s n They represent the first first word vector, the second first word vector, the third first word vector, and so on. The first word sequence is converted into a matrix form to obtain the first word vector matrix Each word They may all be related to the sub-intangible cultural heritage knowledge graph.

[0116] Multiple first entity vectors can be represented as a first entity vector sequence in They respectively represent the first first entity vector, the second first entity vector, the third first entity vector, and so on the nth first entity vector.

[0117] Accordingly, multiple first entity context vectors can be represented as a first entity context vector sequence in They respectively represent the first entity context vector corresponding to the first first entity vector, the first entity context vector corresponding to the second first entity vector, the first entity context vector corresponding to the third first entity vector,... the first entity context vector corresponding to the nth first entity vector.

[0118] The first entity vector sequence and the first entity context vector sequence can be scaled using formula (3), and the scales of the converted first entity vector sequence and the first entity context vector sequence are the same as the scale of the first word vector matrix.

[0119] g(x)=tanh(Mx+b) (3)

[0120] In formula (3), g(x) is the conversion function, M∈R d×k is the transformation matrix, b∈R d×1 is the bias parameter. Since the transformation function is continuous, it can be used to obtain the entity vector and entity context vector projection vector that retain the original spatial relationship.

[0121] When using the conversion function to convert the first entity vector sequence and the first entity context vector sequence, it is only necessary to substitute each item in the first entity vector sequence and the first entity context vector sequence into x in the conversion function in turn.

[0122] The converted first entity vector sequence finally obtained can be expressed by formula (4), and the converted first entity context vector sequence can be expressed by formula (5).

[0123]

[0124] In formula (4), g(e t ) is the first entity vector sequence after conversion, the first entity vector after conversion, the second first entity vector after conversion, and so on, the nth first entity vector after conversion.

[0125]

[0126] In formula (5), is the converted first entity context vector sequence, The converted 1st first entity context vector, the converted 2nd first entity context vector, ... the converted nth first entity context vector.

[0127] The converted first entity vector sequence, the converted first entity context vector sequence and the first word vector sequence are aligned and stacked to obtain the title feature matrix, and the title feature matrix is ​​used as the multi-channel input of KCNN. The title feature matrix can be expressed by formula (6).

[0128]

[0129] In formula (6), W t Represents the title feature matrix. The meanings of other parameters in formula (6) are the same as those of the parameters in formula (4), formula (5) and the first word vector sequence in step g, and detailed description is omitted here.

[0130] After obtaining multi-channel input, multiple convolution kernels with different window sizes are used in the convolution layer to extract features from the title feature matrix. The convolution process can be expressed by formula (7).

[0131]

[0132] In formula (7), h is a convolution kernel, is the convolution result of the title feature submatrix and the convolution kernel h, is the local sub-matrix of the title feature matrix, f is the activation function, and b is the bias vector.

[0133] After the title feature matrix is ​​convolved with the convolution kernel h, multiple convolution results are obtained. At this time, it is necessary to select important local features from the multiple convolution results through a pooling layer. In the embodiment of the present disclosure, the pooling layer is a maximum pooling layer, and the maximum pooling process can be expressed by formula (8).

[0134]

[0135] In formula (8), is the local title feature after maximum pooling, It is the convolution result obtained by convolving the title feature matrix with the convolution kernel h.

[0136] In addition, there are multiple convolution kernels in the convolution layer, and each convolution kernel needs to be convolved with the title feature matrix and perform the maximum pooling operation, so multiple local title features can be obtained in the end. These multiple local title features are concatenated to finally obtain the title feature vector of the first intangible cultural heritage news. The title feature vector can be expressed by formula (9).

[0137]

[0138] In formula (9), e(t) is the title feature vector, and n is the number of convolution kernels. The meanings of other parameters in formula (9) are the same as those in formula (8), and detailed description is omitted here.

[0139] For all the intangible cultural heritage news except the first intangible cultural heritage news in the intangible cultural heritage news set, the above step dg can be used to extract the title feature vector.

[0140] Although news titles are a high-level summary of news content, some intangible cultural heritage news titles are relatively brief and contain less entity and semantic information. Using only news titles for recommendation may not be ideal. Therefore, in the disclosed embodiment, a certain number of keywords are extracted from the news text to represent the semantic information of the full text, and they are integrated with the intangible cultural heritage knowledge graph as an important factor for recommendation.

[0141] Optionally, the keyword feature vector of each intangible cultural heritage news is obtained using the following steps hl.

[0142] Step h, obtaining the text keyword sequence of the first intangible cultural heritage news.

[0143] The text keyword sequence includes keywords in the text of multiple first intangible cultural heritage news.

[0144] Optionally, step h includes: constructing a dictionary in the field of intangible cultural heritage; using the Jieba word segmentation library that supports custom dictionaries to segment the text of the first intangible cultural heritage news, and filtering stop words to obtain multiple text words; selecting the first m words from the multiple text words to form a text keyword sequence.

[0145] Among them, the dictionary of intangible cultural heritage includes all project names and inheritors’ names. Each column of the dictionary represents a word in the intangible cultural heritage field, and the part of speech is marked after the word.

[0146] Optionally, the stop word list of Harbin Institute of Technology is used to filter stop words.

[0147] Considering the high cost of full vectorization of long texts, the value of m should not be too large, for example, it can be close to the number of words in the title of intangible cultural heritage news. For example, m is a positive integer, and the value range of m is between 7 and 12, for example, it can be 7, 10 or 12.

[0148] The keyword sequence in the body of the news not only helps to capture the core concepts of the news content, but also makes up for the situation where some titles are relatively simple or vague.

[0149] Step i: convert multiple keywords in the text keyword sequence into multiple second word vectors.

[0150] The keyword sequence can be expressed as Keys = [s1, s2, s3, ...s m], where s1, s2, s3, ...s m They respectively represent the first keyword, the second keyword, the third keyword, and so on the mth keyword.

[0151] Multiple second word vectors can be represented as a word vector matrix in They respectively represent the word vector corresponding to the first keyword, the word vector corresponding to the second keyword, the word vector corresponding to the third keyword, and so on, the word vector corresponding to the mth keyword.

[0152] Step j: Based on the BERT-BiLSTM-CRM model, multiple second entities in the text keyword sequence of the first intangible cultural heritage news and the second entity vector of each second entity are obtained.

[0153] Here, each keyword in the main text keyword is a second entity.

[0154] Step k, based on the first intangible cultural heritage knowledge graph and multiple second entities, obtain multiple second entity context vectors.

[0155] Each second entity context vector corresponds to a second entity.

[0156] Step 1: Input multiple second word vectors, multiple second entity vectors, and multiple entity context vectors into a knowledge-aware convolutional neural network KCNN to obtain a keyword feature vector of the first intangible cultural heritage news.

[0157] The implementation of step i to step 1 is similar to the implementation of step d to step g. It is only necessary to convert the news title in step d to step g into the news text keyword sequence, and the detailed description is omitted here.

[0158] Similar to step g, step 1 also includes the process of convolution, maximum pooling and concatenation of the keyword feature matrix, and finally the keyword feature vector of the first intangible cultural heritage news can be obtained. The keyword feature vector can be expressed by formula (10).

[0159]

[0160] In formula (10), e(k) is the keyword feature vector and n is the number of convolution kernels.

[0161] In step 302, the title feature vector and keyword feature vector of each intangible cultural heritage news are processed using an attention mechanism to obtain a feature vector of each intangible cultural heritage news.

[0162] Since different news contain different numbers of titles and keywords, their importance should be differentiated. In order to effectively fuse the news title representation and keyword representation vectors, in the disclosed embodiment, an attention mechanism at the news feature level is adopted. The sub-attention mechanism can dynamically adjust the weight distribution according to the importance of the input vector, so that the model pays different attention to different parts, thereby improving the model performance. Therefore, through the attention mechanism, the attention weight of the title feature vector and the attention weight of the keyword feature vector can be determined separately.

[0163] The attention weight of the title feature vector can be calculated using formula (11), and the attention weight of the keyword feature vector can be calculated using formula (12).

[0164] a t =q t tanh(V t r t +b t ) (11)

[0165] In formula (11), a t is the attention weight of the title feature vector, q t is the query vector of the title feature vector, V t and b t is a trainable parameter, r t is the title feature vector of the intangible cultural heritage news, which is e(t) in formula (9).

[0166] a k =q k tanh(V k r k +b k ) (12)

[0167] Similar to formula (11), in formula (12), a k is the attention weight of the keyword feature vector, q k is the query vector of the keyword feature vector, V k and b k is a trainable parameter, r k is the keyword feature vector of intangible cultural heritage news, which is e(k) in formula (10).

[0168] The attention weights of the title feature vector and the keyword feature vector need to be normalized using the softmax function. The normalization process can be expressed by formula (13).

[0169]

[0170] Optionally, formula (14) is used to determine the feature vector of each intangible cultural heritage news. a in formula (14) is t and a k is the normalized a in formula (13) t and a k .

[0171] r n =α t r t +α k r k (14)

[0172] In formula (14), r n is the feature vector of intangible cultural heritage news. The meanings of other parameters in formula (14) are the same as those in formula (11) and formula (12), and detailed description is omitted here.

[0173] Through the above steps 301 to 302, the feature vector of each intangible cultural heritage news in the intangible cultural heritage news set can be obtained.

[0174] In step 303, based on the feature vectors of multiple intangible cultural heritage news in the intangible cultural heritage news collection, a user preference feature vector corresponding to each candidate news is determined.

[0175] The candidate news is the intangible cultural heritage news in the intangible cultural heritage news collection that has not been clicked by the user.

[0176] Assume that user i has clicked N times in total i This is a non-legacy news. i Intangible cultural heritage news This i The feature vector corresponding to the intangible cultural heritage news is recorded as The feature vector of the j-th candidate news is recorded as e(t j ).

[0177] For this N i The feature vectors corresponding to the jth intangible cultural heritage news and the feature vectors of the jth candidate news are all convolved, and the results of the convolution operation are connected into a feature vector with attention. Then the feature vector with attention is input into the attention network layer to obtain a weighted matrix, and then the weighted matrix is ​​normalized to obtain a weight matrix.

[0178] The weight matrix can be expressed by formula (15).

[0179]

[0180] In formula (15), is the weight matrix, softmax is the normalization function, is the feature vector of the kth intangible cultural heritage news clicked by user i, e(t j ) is the feature vector of the j-th candidate news.

[0181] Optionally, formula (16) is used to determine the user preference vector corresponding to the j-th candidate news:

[0182]

[0183] In formula (15), e(i) is the user preference vector of user i for the jth candidate news. The meanings of other parameters in formula (16) are the same as those in formula (15), and detailed description is omitted here.

[0184] In this way, through step 303, the user preference vector of each candidate news can be obtained.

[0185] In step 304, based on the user preference feature vector corresponding to each candidate news, the intangible cultural heritage news is pushed to the user.

[0186] Optionally, step 304 includes the following three steps.

[0187] The first step is to determine the user's preference for each candidate intangible cultural heritage news based on the user preference feature vector corresponding to each candidate news.

[0188] Here, the user's preference for each candidate intangible cultural heritage news can be represented by the probability of the user clicking on each candidate intangible cultural heritage news. It is generally believed that the more similar the intangible cultural heritage news is to the user's preference, the more likely the user is to choose to read it. Therefore, when calculating the similarity between the feature vector of a candidate news and the user's preference vector for the candidate news, since both are represented in the form of vectors, the dot product operation of the vector is used, and then the reduce sum calculation is performed to obtain the probability of the user clicking on the candidate news. This process can be expressed by formula (17).

[0189]

[0190] In formula (17), is the user i’s preference for the jth candidate intangible cultural heritage news, The higher the value, the more likely user i is to click on the jth candidate intangible cultural heritage news; G(e(i), e(t j )) represents performing a dot product operation on the user preference vector of user i for the j-th candidate news and the feature vector of the j-th candidate news, and then performing summation and dimensionality reduction.

[0191] In this way, through the above formula (17), the degree of preference of user i for each candidate news can be calculated.

[0192] The second step is to sort each candidate intangible cultural heritage news item in descending order according to its degree of preference to obtain the first preference order.

[0193] The third step is to push the first N candidate news in the first favorite order to the user.

[0194] Considering the actual business scenario of online news, after the user refreshes the page, the news website will push the news that the user is interested in to the user based on the user's browsing history or click behavior. The first N candidate news in the embodiment of the present disclosure are the news pushed to the user by the news website after the user refreshes the page.

[0195] The above steps 301 to 304 can be established as an intangible cultural heritage news recommendation model. Figure 6 This is a schematic diagram of the structure of the intangible cultural heritage news recommendation model. The structure of the intangible cultural heritage news recommendation model is consistent with the above steps 301 to 304, and the detailed description is omitted here.

[0196] The following are device embodiments of the present application. For details not described in detail in the device embodiments, reference may be made to the above method embodiments.

[0197] Figure 7 A schematic diagram of the structure of a non-legacy news recommendation device provided by an exemplary embodiment of the present disclosure is shown. Figure 7 The intangible cultural heritage news recommendation device 700 includes: an acquisition module 701, an attention mechanism module 702, a user preference feature vector calculation module 703 and a push module 704.

[0198] The acquisition module 701 is used to obtain the title feature vector and keyword feature vector of each intangible cultural heritage news in the intangible cultural heritage news collection based on the BERT-BiLSTM-CRM model and the first intangible cultural heritage knowledge graph, and the keyword feature vector is extracted from the text of the intangible cultural heritage news.

[0199] The attention mechanism module 702 is used to process the title feature vector and keyword feature vector of each intangible cultural heritage news by using the attention mechanism to obtain the feature vector of each intangible cultural heritage news.

[0200] The user preference feature vector calculation module 703 is used to determine the user preference feature vector corresponding to each candidate news based on the feature vectors of multiple intangible cultural heritage news in the intangible cultural heritage news collection, and the candidate news is the intangible cultural heritage news in the intangible cultural heritage news collection that has not been clicked by the user.

[0201] The push module 704 is used to push intangible cultural heritage news to users based on the user preference feature vector corresponding to each candidate news.

[0202] Optionally, the acquisition module 701 is also used to convert the title of the first intangible cultural heritage news into multiple first word vectors, where the first intangible cultural heritage news is any one of the multiple intangible cultural heritage news; based on the BERT-BiLSTM-CRM model, obtain multiple first entities in the title of the first intangible cultural heritage news and the first entity vector of each first entity; based on the first intangible cultural heritage knowledge graph and the multiple first entities, obtain multiple first entity context vectors, each first entity context vector corresponds to a first entity; input the multiple first word vectors, the multiple first entity vectors and the multiple entity context vectors into the knowledge-aware convolutional neural network KCNN to obtain the title feature vector of the first intangible cultural heritage news.

[0203] Optionally, the acquisition module 701 is also used to obtain a text keyword sequence of the first intangible cultural heritage news, the text keyword sequence including multiple keywords in the text of the first intangible cultural heritage news; converting the multiple keywords in the text keyword sequence into multiple second word vectors, and based on the BERT-BiLSTM-CRM model, obtaining multiple second entities in the text keyword sequence of the first intangible cultural heritage news and the second entity vector of each second entity; based on the first intangible cultural heritage knowledge graph and the multiple second entities, obtaining multiple second entity context vectors, each second entity context vector corresponding to a second entity; inputting the multiple second word vectors, the multiple second entity vectors and the multiple entity context vectors into the knowledge-aware convolutional neural network KCNN to obtain the keyword feature vector of the first intangible cultural heritage news.

[0204] Optionally, the attention mechanism module 702 is further used to determine the feature vector of each intangible cultural heritage news using the following formula:

[0205] r n =α t r t +α k r k

[0206] Among them, r n is the feature vector of intangible cultural heritage news, r t is the title feature vector of intangible cultural heritage news, r k is the keyword feature vector of intangible cultural heritage news, a t is the attention weight of the title feature vector, a k is the attention weight of the keyword feature vector.

[0207] Optionally, the user preference feature vector calculation module 703 is further configured to determine the user preference vector corresponding to the j-th candidate news by using the following formula:

[0208]

[0209] Among them, e(i) is the user preference vector of user i for the jth candidate news, is the feature vector of the kth intangible cultural heritage news clicked by user i. User i has clicked N times in total. i A piece of non-legacy news, is the weight matrix, t j is the jth candidate news;

[0210] The weight matrix is ​​calculated using the following formula:

[0211]

[0212] in, is the weight matrix, softmax is the normalization function, e(t j ) is the feature vector of the j-th candidate news.

[0213] Optionally, the push module 704 is also used to determine the user's preference for each candidate intangible cultural heritage news based on the user preference feature vector corresponding to each candidate news; sort the candidate intangible cultural heritage news in descending order according to the preference of each candidate news to obtain a first preference order; and push the first N candidate news in the first preference order to the user.

[0214] It should be noted that: when the intangible cultural heritage news push device provided in the above embodiment pushes intangible cultural heritage news, only the division of the above functional modules is used as an example. In actual applications, the above functional distribution can be completed by different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the intangible cultural heritage news push device provided in the above embodiment and the intangible cultural heritage news push method embodiment belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0215] The division of modules in the embodiments of the present disclosure is schematic and is only a logical function division. There may be other division methods in actual implementation. In addition, each functional module in each embodiment of the present disclosure may be integrated into a processor, or may exist physically separately, or two or more modules may be integrated into one module. The above-mentioned integrated modules may be implemented in the form of hardware or in the form of software functional modules.

[0216] If the integrated module is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present disclosure is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a terminal device (which can be a personal computer, a mobile phone, or a communication device, etc.) or a processor (processor) to perform all or part of the steps of the method of each embodiment of the present disclosure. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program code.

[0217] Figure 8 Schematic diagram of the structure of the computer device provided by the embodiment of the present disclosure. Figure 8 As shown, the computer device 800 includes: a processor 801 and a memory 802 .

[0218] The processor 801 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 801 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 801 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 801 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 801 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.

[0219] The memory 802 may include one or more computer-readable storage media, which may be non-transitory. The memory 802 may also include a high-speed random access memory, and a non-volatile memory, such as one or more disk storage devices, flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 802 is used to store at least one instruction, which is used to be executed by the processor 801 to implement the non-legacy news push method provided in the embodiments of the present disclosure.

[0220] Those skilled in the art will understand that Figure 8 The structure shown in the figure does not constitute a limitation on the computer device 800, and the computer device 800 may include more or less components than those shown in the figure, or combine some components, or adopt a different arrangement of components.

[0221] The embodiment of the present disclosure also provides a non-temporary computer-readable storage medium. When the instructions in the storage medium are executed by the processor of a computer device, the computer device can execute the intangible cultural heritage news push method provided in the embodiment of the present disclosure.

[0222] The embodiments of the present disclosure also provide a computer program product, including a computer program / instruction, which, when executed by a processor, implements the intangible cultural heritage news push method provided in the embodiments of the present disclosure.

[0223] The above description is only an optional embodiment of the present disclosure and is not intended to limit the present disclosure. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present disclosure shall be included in the protection scope of the present disclosure.

Claims

1. A method for recommending intangible cultural heritage news, characterized in that: The method comprises: Based on the BERT-BiLSTM-CRM model and the first intangible cultural heritage knowledge graph, the title feature vector and keyword feature vector of each intangible cultural heritage news in the intangible cultural heritage news collection are obtained, and the keyword feature vector is extracted from the text of the intangible cultural heritage news; Using an attention mechanism to process the title feature vector and the keyword feature vector of each intangible cultural heritage news, to obtain a feature vector of each intangible cultural heritage news; Based on the feature vectors of multiple intangible cultural heritage news in the intangible cultural heritage news set, determining a user preference feature vector corresponding to each candidate news, wherein the candidate news is the intangible cultural heritage news in the intangible cultural heritage news set that has not been clicked by the user; Based on the user preference feature vector corresponding to each candidate news, the intangible cultural heritage news is pushed to the user.

2. The method according to claim 1, characterized in that: The method of obtaining the title feature vector and keyword feature vector of each intangible cultural heritage news in the intangible cultural heritage news set based on the BERT-BiLSTM-CRM model and the first intangible cultural heritage knowledge graph includes: Converting a title of a first intangible cultural heritage news into a plurality of first word vectors, wherein the first intangible cultural heritage news is any one of the plurality of intangible cultural heritage news; Based on the BERT-BiLSTM-CRM model, obtain multiple first entities in the title of the first intangible cultural heritage news and a first entity vector of each first entity; Based on the first intangible cultural heritage knowledge graph and the multiple first entities, obtain multiple first entity context vectors, each of which corresponds to one first entity; Input the multiple first word vectors, the multiple first entity vectors and the multiple entity context vectors into the knowledge-aware convolutional neural network KCNN to obtain the title feature vector of the first intangible cultural heritage news.

3. The method according to claim 2, characterized in that The method of obtaining the title feature vector and keyword feature vector of each intangible cultural heritage news in the intangible cultural heritage news set based on the BERT-BiLSTM-CRM model and the first intangible cultural heritage knowledge graph also includes: Acquire a text keyword sequence of the first intangible cultural heritage news, wherein the text keyword sequence includes a plurality of keywords in the text of the first intangible cultural heritage news; Converting multiple keywords in the text keyword sequence into multiple second word vectors; Based on the BERT-BiLSTM-CRM model, a plurality of second entities in the text keyword sequence of the first intangible cultural heritage news and a second entity vector of each second entity are obtained; Based on the first intangible cultural heritage knowledge graph and the plurality of second entities, obtaining a plurality of second entity context vectors, each of the second entity context vectors corresponding to one second entity; Input multiple second word vectors, multiple second entity vectors and multiple entity context vectors into a knowledge-aware convolutional neural network KCNN to obtain the keyword feature vector of the first intangible cultural heritage news.

4. The method according to any one of claims 1 to 3, characterized in that: The attention mechanism is used to process the title feature vector and the keyword feature vector of each intangible cultural heritage news to obtain the feature vector of each intangible cultural heritage news, including: The following formula is used to determine the feature vector of each intangible cultural heritage news: r n ＝α t r t +α k r k Among them, r n is the feature vector of intangible cultural heritage news, r t is the title feature vector of the intangible cultural heritage news, r k is the keyword feature vector of intangible cultural heritage news, a t is the attention weight of the title feature vector, a k is the attention weight of the keyword feature vector.

5. The method according to any one of claims 1 to 3, characterized in that: The step of determining a user preference feature vector corresponding to each candidate news based on the feature vectors of multiple pieces of intangible cultural heritage news in the intangible cultural heritage news set includes: The following formula is used to determine the user preference vector corresponding to the jth candidate news: Among them, e(i) is the user preference vector of user i for the jth candidate news, is the feature vector of the kth intangible cultural heritage news clicked by the user i. The user i has clicked N times in total. i A piece of non-legacy news, is the weight matrix, t j is the j-th candidate news; The weight matrix is ​​calculated using the following formula: in, is the weight matrix, softmax is the normalization function, e(t j ) is the feature vector of the j-th candidate news.

6. The method according to any one of claims 1 to 3, characterized in that: The method of pushing intangible cultural heritage news to the user based on the user preference feature vector corresponding to each candidate news includes: Determining the user's preference for each piece of candidate intangible cultural heritage news based on the user preference feature vector corresponding to each piece of candidate news; Sort the candidate intangible cultural heritage news in descending order according to the degree of preference of each candidate intangible cultural heritage news, and obtain a first preference order; Pushing the first N candidate news in the first favorite order to the user.

7. A device for recommending intangible cultural heritage news, characterized in that: The device comprises: An acquisition module, for acquiring a title feature vector and a keyword feature vector of each intangible cultural heritage news in the intangible cultural heritage news set based on a BERT-BiLSTM-CRM model and a first intangible cultural heritage knowledge graph, wherein the keyword feature vector is extracted from the body of the intangible cultural heritage news; An attention mechanism module, used for processing the title feature vector and the keyword feature vector of each intangible cultural heritage news by using an attention mechanism to obtain a feature vector of each intangible cultural heritage news; A user preference feature vector calculation module, for determining a user preference feature vector corresponding to each candidate news based on feature vectors of multiple pieces of intangible cultural heritage news in the intangible cultural heritage news set, wherein the candidate news is intangible cultural heritage news in the intangible cultural heritage news set that has not been clicked by the user; The push module is used to push the intangible cultural heritage news to the user based on the user preference feature vector corresponding to each candidate news.

8. A computer device, characterized in that: The computer device comprises: a memory and a processor, wherein at least one computer program is stored in the memory, and the at least one computer program is loaded and executed by the processor to implement the method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one computer program, and the at least one computer program is loaded and executed by a processor to implement the method according to any one of claims 1 to 6.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images