A news recommendation method and device

Abstract

The application relates to a news recommendation method and device. The news recommendation method comprises the following steps: performing cross-attention calculation on a news title self-attention representation and a news content self-attention representation of each candidate news to obtain corresponding news content cross-attention representations; performing element-by-element multiplication on a news title cross-attention representation and a news content cross-attention representation of each candidate news to obtain a news representation corresponding to each candidate news; performing dot product operation on the news representation corresponding to each candidate news and a user representation in a user representation database to obtain a non-normalized matching score of the candidate news and a corresponding user; and performing news recommendation on the corresponding user according to the high and low of the normalized matching score after sorting the candidate news according to the matching score. The news recommendation method has the advantages of mining potential semantic interaction information between news titles and news contents and improving the precision of news recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of user preference recommendation, in particular to the field of a news recommendation method. BACKGROUND

[0002] With the development of Internet technology, online news platforms have become an important way for users to obtain news information. Due to the rapid production and large accumulation of news, news portals have the problem of information overload, and it is difficult for people to quickly obtain the content they are interested in. In order to alleviate information overload and improve reading experience, personalized news recommendation has become an important solution for these news platforms. News recommendation is applied to news portals to obtain user interest preference features based on user usage records, provide personalized news recommendation services for users, and improve user experience.

[0003] Existing news modeling methods usually use convolutional neural networks, recurrent neural networks, etc. to encode news titles or news content. This method leads to single news source for news modeling, resulting in insufficient semantic information for news modeling and limited representation ability. Some news modeling methods independently encode news titles, news content, news topics, etc. and concatenate them as the final representation of the news. This method cuts off the connection between the title and the content, lacks the mining of semantic interaction information between the two, and results in limited accuracy of news representation, which fails to effectively recommend the news of interest to users. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a news recommendation method and device which can fully mine the potential semantic interaction information between news content and news title and improve the accuracy of news recommendation.

[0005] Specifically, the news recommendation method provided by the present application comprises the following steps:

[0006] obtaining a plurality of candidate news, processing the text content of the news title of each candidate news and the news content of each candidate news to obtain a news title self-attention representation of each candidate news and a news content self-attention representation of each candidate news;

[0007] cross-attention calculation is performed on the news title self-attention representation of each candidate news and the news content self-attention representation of each candidate news to obtain a corresponding news content cross-attention representation;

[0008] element-wise multiplication is performed on the news title cross-attention representation of each candidate news and the news content cross-attention representation of each candidate news to obtain a news representation corresponding to each candidate news;

[0009] Dot product operation is performed between the news representation corresponding to each candidate news and a user representation in the user representation database to obtain a non-normalized matching score of the candidate news and the corresponding user;

[0010] The candidate news is ranked according to the matching score, and news recommendation is performed for the corresponding user according to the high and low of the normalized matching score.

[0011] Further, cross-attention calculation is performed between the news title self-attention representation of each candidate news and the news content self-attention representation of each candidate news to obtain the corresponding news content cross-attention representation, which is obtained by the following formula:

[0012]

[0013]

[0014]

[0015]

[0016] wherein, α cross is the cross-attention weight of each query word, r t cross is the news title cross-attention representation, r c cross is the news content title cross-attention representation.

[0017] Further, element-wise multiplication is performed between the news title cross-attention representation of each candidate news and the news content cross-attention representation of each candidate news to obtain the news representation corresponding to each candidate news, which is obtained by the following formula:

[0018]

[0019] wherein, r n is the news representation, r t cross is the news title cross-attention representation, r c cross is the news content title cross-attention representation.

[0020] Further, dot product operation is performed between the news representation corresponding to each candidate news and a user representation in the user representation database to obtain a non-normalized matching score of the candidate news and the corresponding user, and the specific formula is as follows:

[0021] y = <r n , r u >

[0022] wherein, < > is a dot product operator, ru is a news representation, r u is a user representation.

[0023] Further, the user representation data in the user representation database is obtained by the following steps:

[0024] Obtaining historical news data browsed by the user, and performing text content processing on the news title of each historical news and the news content of each historical news to obtain a news title self-attention representation and a news content self-attention representation of each historical news;

[0025] Performing cross-attention calculation on the news title self-attention representation of each historical news and the news content self-attention representation of each historical news to obtain a corresponding news content cross-attention representation;

[0026] Element-wise multiplication of the news title cross-attention representation of each historical news and the news content cross-attention representation of each historical news to obtain a corresponding news representation of each historical news, and all news representations of the historical news constitute a user-news representation set.

[0027] Further, element-wise multiplication of the news title cross-attention representation of each historical news and the news content cross-attention representation of each historical news to obtain a corresponding news representation of each historical news, and all news representations of the historical news constitute a user-news representation set, and further comprising:

[0028] Calculating the user representation in the user-news set according to the self-attention mechanism to obtain a corresponding user self-attention representation; and calculating according to the multi-head attention mechanism to obtain a corresponding user multi-head attention representation;

[0029] Element-wise addition of the user self-attention representation and the user multi-head attention representation to obtain the final user representation.

[0030] Further, obtaining a plurality of candidate news, and performing text content processing on the news title of each candidate news and the news content of each candidate news to obtain a news title self-attention representation and a news content self-attention representation of each candidate news, specifically comprising:

[0031] Obtaining a plurality of candidate news, and performing word embedding processing on the news title of each candidate news and the news content of each candidate news to obtain a word vector representation of the news title of each candidate news and a word vector representation of the news content of each candidate news;

[0032] Performing feature extraction on the news title word vector of each candidate news and the news content word vector of each candidate news to obtain a corresponding news title embedding representation and a news content embedding representation;

[0033] The news title embedding representation of each candidate news and the news content embedding of each candidate news are respectively subjected to self-attention calculation to obtain corresponding self-attention representations of the news title and the news content.

[0034] Further, the news title embedding representation of each candidate news and the news content embedding of each candidate news are respectively subjected to self-attention calculation to obtain specific calculation formulas of the corresponding self-attention representations of the news title and the news content as follows:

[0035] α self,i = softmax(v T tanh(Wh i +b))

[0036]

[0037] wherein, wherein v T is a trainable weight vector, Wh is a parameter matrix, b is a bias term, and a self,i is the self-attention weight of each word query.

[0038] In another aspect, the application also provides a news recommendation device, which comprises:

[0039] a text extraction module configured to obtain a plurality of candidate news, and perform text content processing on the news title of each candidate news and the news content of each candidate news to obtain self-attention representations of the news title and the news content of each candidate news;

[0040] a cross calculation module configured to perform cross-attention calculation on the self-attention representations of the news title of each candidate news and the self-attention representations of the news content of each candidate news to obtain corresponding cross-attention representations of the news content;

[0041] a feature fusion module configured to perform element-wise multiplication on the cross-attention representations of the news title of each candidate news and the cross-attention representations of the news content of each candidate news to obtain corresponding news representations of each candidate news;

[0042] a matching degree calculation module configured to perform dot product operation on the corresponding news representations of each candidate news and a user representation in a user representation database to obtain a non-normalized matching score of the candidate news and the corresponding user;

[0043] a sorting module configured to sort the candidate news according to the matching score and recommend news to the corresponding user according to the high and low of the normalized matching score.

[0044] Further, the text extraction module comprises:

[0045] The word embedding module is configured to obtain a plurality of candidate news, perform word embedding processing on a news title of each candidate news and news content of each candidate news, and obtain a word vector representation of the news title of each candidate news and a word vector representation of the news content of each candidate news.

[0046] The feature extraction module is configured to perform feature extraction on the news title word vector of each candidate news and the news content word vector of each candidate news, respectively, to obtain a corresponding news title embedding representation and a news content embedding representation.

[0047] The self-attention calculation module is configured to perform self-attention calculation on the news title embedding representation of each candidate news and the news content embedding of each candidate news, respectively, to obtain a corresponding self-attention representation of the news title and a self-attention representation of the news content.

[0048] The news recommendation method and device provided by the present application adopt a double-channel feature extraction method to deeply mine the context information of news text, enhance the news representation ability, and fuse the self-attention representation of the news title and the self-attention representation of the news content through a cross-attention network to deeply mine the potential semantic interaction information between the two, further enrich the news representation, and improve the accuracy of news recommendation.

[0049] In order to better understand and implement, the present application will be described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 The structure block diagram of the news recommendation device provided by the present application is shown in the figure;

[0051] Figure 2 The flowchart of the news recommendation method provided by the present application is shown in the figure.

[0052] Reference signs: 11: word embedding module; 12: feature extraction module; 122: news title feature extraction module; 124: news content feature extraction module; 13: self-attention calculation module; 132: news title self-attention calculation module; 134: news content self-attention calculation module; 14: cross-attention calculation module; 15: feature fusion module; 16: matching degree calculation module; 17: sorting module. DETAILED DESCRIPTION

[0053] In view of the problem that the existing technology has insufficient news internal semantic interaction, resulting in limited representation accuracy; the present application starts from the composition of news, mines the relationship between the components of news, and deeply fuses the news title and the news content through the cross attention network, learns the potential semantic interaction information between the news title and the news content, enhances the news representation capability, and improves the news recommendation performance.

[0054] Please refer to Figure 1 and Figure 2 wherein, Figure 1 is a news recommendation device module schematic diagram of the present application, Figure 2 is Figure 1 a news recommendation method flowchart executed by the news recommendation device of the present application. The news recommendation device of the present application comprises a word embedding module 11, a feature extraction unit 12, a self-attention calculation unit 13, a cross-attention calculation module 14, a feature fusion module 15, a matching degree calculation module 16, a sorting module 17 and a user representation database.

[0055] The word embedding module 11 is used to execute step S10: obtaining a plurality of candidate news, performing word embedding processing on the news title of each candidate news and the news content of each candidate news, obtaining the word vector representation of the news title of each candidate news and the word vector representation of the news content of each candidate news.

[0056] Specifically, in the present application, the news title of each candidate news and the news content of each candidate news are processed by the Glove of the word embedding layer, and the news title of the candidate news and the news content of the candidate news are first subjected to text recognition and text content splitting; then the split text content is subjected to word vector construction to obtain the corresponding news title word vector representation and news content word vector representation wherein v is a matrix data, for example, [0.03, 0.06, 0.07; 0, 0.03, 0.05; 0.04, 0.06, 0].

[0057] Glove is a word representation tool based on global word frequency statistics, which can express a word as a vector composed of real numbers, and these vectors capture some semantic properties between words, such as similarity and analogy. Through the operation of the vector, the semantic similarity between two words can be calculated. Therefore, through Glove processing, the news title and the news content can be converted from text to word vector representation containing semantic properties.

[0058] The feature extraction unit 12 includes a news title feature extraction module 122 and a news content feature extraction module 124. The news title feature extraction module 122 is configured to perform step S202 of performing feature extraction on the news title word vector of each candidate news to obtain a corresponding news title embedding representation. The news content feature extraction module 124 is configured to perform step S204 of performing feature extraction on the news title word vector of each candidate news to obtain a corresponding news title embedding representation.

[0059] Specifically, the news title and the news content contain rich text sequence information. The application adopts a dual-channel mode of long short-term memory neural network (LSTM) and gated recurrent unit (GRU) to extract the text sequence information of the news title and the news content. The news title is obtained by LSTM, and the news content is obtained by LSTM. The news title is obtained by GRU, and the news content is obtained by GRU. The news title is obtained by GRU, and the news content is obtained by GRU. The news title is obtained by GRU, and the news content is obtained by GRU. The news title is obtained by GRU, and the news content is obtained by GRU. The news title is obtained by GRU, and the news content is obtained by GRU. The news title is obtained by GRU, and the news content is obtained by GRU. The news title is obtained by GRU, and the news content is obtained by GRU. The news title is obtained by GRU, and the news content is obtained by GRU. The news title is obtained by GRU, and the news content is obtained by GRU.

[0060] LSTM and GRU are time recurrent neural networks, which can effectively preserve historical information and learn long-term dependent information of the text. In the training of large amount of text features, the accuracy of LSTM is superior to that of GRU, but LSTM has the problem of overfitting due to too many parameters. Compared with LSTM, GRU is optimized on LSTM and has fewer parameters, reducing the risk of overfitting. Therefore, the application adopts a dual-channel mode of LSTM and GRU to reduce the risk of overfitting of the model while ensuring the accuracy of the model, thereby further improving the performance of the device.

[0061] The self-attention calculation unit 13 includes a news title self-attention calculation module 132 and a news content self-attention calculation module 134. The news title self-attention calculation module 132 is configured to perform step S302 of performing self-attention calculation on the news title embedding representation of each candidate news to obtain a corresponding self-attention representation of the news title. The news content self-attention calculation module 134 is configured to perform step S304 of performing self-attention calculation on the news content embedding representation of each candidate news to obtain a corresponding self-attention representation of the news content.

[0062] Specifically, the self-attention representation is calculated by the following formula:

[0063] α self,i= softmax(v T tanh(Wh i +b))

[0064] First, the self-attention weight of each word in the text content is calculated by the softmax function, and the importance of each word is obtained, and then the embedding representation corresponding to each word is summed to obtain the self-attention representation of the news title and the news content of the corresponding candidate news.

[0065]

[0066] wherein, wherein v T is a trainable weight vector, Whis a parameter matrix, b is a bias term, a self,i is the self-attention weight of each word, r self is the self-attention representation.

[0067] The cross-attention calculation module 14 is configured to perform step S40: performing cross-attention calculation on the self-attention representation of the news title of each candidate news and the self-attention representation of the news content of each candidate news to obtain the corresponding news content cross-attention representation.

[0068] Specifically, the calculation is performed by the following formula:

[0069]

[0070]

[0071] wherein, a cross is the cross-attention weight of each word, r t cross is the news title cross-attention representation, r c cross is the news content title cross-attention representation.

[0072] The self-attention representation of the title is taken as a query, and the content embedding h c is taken as a key value to build a cross-attention network to obtain the title cross-attention representation

[0073]

[0074]

[0075] The self-attention representation of the content is taken as a query, and the title embedding h t is taken as a key value to build a cross-attention network to obtain the content cross-attention representation

[0076] By constructing the key-value pair relationship between the self-attention representation of the news title and the embedding representation of the news content and the key-value pair relationship between the self-attention representation of the news content and the embedding representation of the news title through the cross-attention network, the fusion between the news title and the news content is realized, and the cross-attention representation of the corresponding news title and the news content is obtained.

[0077] The feature fusion module 15 is configured to perform step S50: element-wise multiplication of the news title cross-attention representation of each candidate news and the news content cross-attention representation of each candidate news to obtain the news representation corresponding to each candidate news.

[0078] Specifically, the calculation is performed by the following formula:

[0079]

[0080] wherein r n is the news representation, r t cross is the news title cross-attention representation, r c cross is the news content cross-attention representation.

[0081] The matching degree calculation module 16 is configured to perform step S60: dot product operation of the news representation corresponding to each candidate news and a user representation in the user representation database to obtain the non-normalized matching score of the candidate news and the corresponding user.

[0082] Specifically, the calculation is performed by the following formula:

[0083] y = <r n , r u >

[0084] wherein < > is a dot product operator, r u is the news representation, and r u is the user representation.

[0085] In the present application, the data stored in the user representation database is obtained by the following method:

[0086] The user's news browsing history data is obtained through the word embedding module 11, the feature extraction module 12, the self-attention calculation module 13, the cross-attention calculation module 14, the feature fusion module 15 to obtain the corresponding user-news representation set, and then the user-self-attention representation and the user-multi-head-attention representation are obtained through the self-attention mechanism and the multi-head-attention mechanism respectively. Finally, the user-self-attention representation and the user-multi-head-attention representation are added element by element to obtain the final user representation and store it in the user representation database.

[0087] Since different news has different information, using self-attention mechanism to learn the weight of different news is beneficial to mine user interest preference features. If only self-attention mechanism is used to encode the current news, the device will excessively focus attention on the news, lacking feature extraction of the user's overall interest preference. Multi-head attention mechanism can further mine the association between news and improve the completeness of the user representation. Through the parallel way of self-attention mechanism layer and multi-head attention mechanism layer, the model's ability to extract user features is enhanced, and the user interest representation is further mined.

[0088] The sorting module 17 is configured to perform step S70: sorting the candidate news according to the matching score, and recommending news to the corresponding user according to the high and low of the normalized matching score.

[0089] The matching score reflects the user's preference and the matching degree of the candidate news. According to the high and low of the matching score, the possibility of user reading can be predicted, and the higher the matching score, the greater the user's interest in reading. Based on this, the candidate news with high matching score is recommended to the user first.

[0090] The present application will be described below through a specific embodiment. The text content of the news title is as follows: Don't be fooled: Costco says $75 coupon is a scan. The text content of the news content is as follows: Wholesaleretailer Costco is warning customers about a new scam that is promisingshoppers a $75 coupon. Through the Glove processing of the embedding layer, first, the text is recognized and the news title is split into [Don't, be, fooled :, Costco, says, $75, coupon, is, a, scan,. ], and the news content is split into [Wholesale, retailer, Costco, is, warning, customers, about, a, new, scam, that, is, promising, shoppers, a, $75, coupon,. ]; Then the corresponding word vector is constructed If the title length is greater than M, the title is truncated to the maximum length M. If the title length is less than M, it is padded with 0 to length M; If the title length is greater than N, the title is truncated to the maximum length N. If the title length is less than N, it is padded with 0 to length N. Then the long short-term memory neural network (LSTM) and the gated recurrent unit (GRU) dual-channel method is used to extract the text sequence information of the news title and the news content. The news title is obtained through the LSTM The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU

[0091] The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU

[0092] The news content is obtained through the GRU The news content is obtained through the GRU The news content is obtained through the GRU

[0093] From the above table, it can be found that in the example, the text content of the news title and the news content can be effectively identified through the above steps. The word with a larger weight value carries more information. In the title word weight representation, words such as "coupon" and "scam" carry a large amount of information and are assigned a large weight value. In the content word weight representation, words such as "customers", "scam", "shoppers", and "coupon" have a large weight value and can be used to explain the title. Therefore, the news recommendation method of the present patent can mine words with strong characteristics between the title and the content, and can learn the association between them, capture news semantic interaction information, and improve the news modeling capability. Then the news title cross-attention representation and the news content cross-attention representation are multiplied element by element to obtain the news representation corresponding to each news. Finally, the user representation obtained according to the user's historical reading news is matched to calculate the matching score, and the matching degree of the candidate news and the user is obtained, and the related news is recommended to the user according to the matching degree.

[0094] Based on the same inventive concept, the present application also provides an electronic device, which can be a server, a desktop computing device or a mobile computing device (e.g., a laptop computer, a handheld computing device, a tablet computer, a netbook, etc.) or the like terminal device. The device includes one or more processors and a memory, wherein the processor is configured to execute a program to implement the news recommendation method described above; and the memory is configured to store a computer program executable by the processor.

[0095] Based on the same inventive concept, the present application also provides a computer readable storage medium, corresponding to the above-mentioned embodiments of the news recommendation method, the computer readable storage medium has a computer program stored thereon, which, when executed by a processor, implements the steps of the news recommendation method described in any of the above embodiments.

[0096] The present application can be in the form of a computer program product implemented on one or more storage media (including but not limited to a disk storage, a CD-ROM, an optical storage, etc.) containing program codes. The computer usable storage media includes permanent and non-permanent, removable and non-removable media, which can be implemented by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to: phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape, magnetic disk storage or other magnetic storage device, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0097] The above-described embodiments only express several embodiments of the present application, which are described in detail and specifically, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the inventive concept, a number of modifications and improvements can be made, and the present application also intends to include these modifications and improvements.

Claims

1. A news recommendation method, characterized in that, include: Multiple candidate news items are obtained. The news title and news content of each candidate news item are processed to obtain the self-attention representation of the news title and the self-attention representation of the news content of each candidate news item. Cross-attention calculation is performed on the self-attention representation of the news title and the self-attention representation of the news content of each candidate news item to obtain the corresponding cross-attention representation of the news content; wherein, the cross-attention representation of the news content is calculated by the following formula: Where, α cross It is the cross-attention weight for each word in the query, r t cross It is a cross-attention representation of news headlines, r c cross This is a cross-attention table for news headlines; The news representation for each candidate news item is obtained by element-wise multiplying the cross-attention representation of the news title of each candidate news item with the cross-attention representation of the news content of each candidate news item. Perform a dot product operation between the news representation corresponding to each candidate news item and a user representation in the user representation database to obtain the nonnormalized matching score between the candidate news item and the corresponding user. Candidate news items are sorted based on matching scores, and news is recommended to corresponding users based on the normalized matching scores.

2. The news recommendation method according to claim 1, characterized in that, The news representation for each candidate news item is obtained by element-wise multiplying the cross-attention representation of the news title and the cross-attention representation of the news content of each candidate news item, and then obtaining the news representation for each candidate news item through the following formula: Where r n The news report stated that... t cross It is a cross-attention representation of news headlines, r c cross It represents the cross-attention representation of news content headlines.

3. The news recommendation method according to claim 2, characterized in that, The specific formula for obtaining the nonnormalized matching score between each candidate news item and its corresponding user by performing a dot product operation between the news representation corresponding to each candidate news item and a user representation in the user representation database is as follows: y＝<r n ,r u > Where <> is the dot product operator, r u The news report stated that... u This is a user's statement.

4. A news recommendation method according to any one of claims 1-3, characterized in that, The user representation data in the user representation database is obtained through the following steps: Obtain the historical news data browsed by the user, process the text content of each historical news headline and each historical news content, and obtain the self-attention representation of the news headline and the self-attention representation of the news content for each historical news. Cross-attention calculation is performed on the self-attention representation of the news headline and the self-attention representation of the news content of each historical news item to obtain the corresponding cross-attention representation of the news content. The cross-attention representation of the news headline and the cross-attention representation of the news content of each historical news item are multiplied element-wise to obtain the news representation corresponding to each historical news item. All the news representations of historical news items form a user-news representation set.

5. The news recommendation method according to claim 4, characterized in that, The news representation for each historical news item is obtained by element-wise multiplying the cross-attention representation of its headline and content. The user-news representation set, composed of all historical news representations, also includes: The user representations in the user-news representation set are calculated using a self-attention mechanism to obtain the corresponding user self-attention representations; and the user multi-head attention representations are calculated using a multi-head attention mechanism. The user self-attention representation and the user multi-head attention representation are added element-wise to obtain the final user representation.

6. The news recommendation method according to claim 5, characterized in that, Multiple candidate news items are obtained. The news title and content of each candidate news item are processed to obtain a self-attention representation of the news title and a self-attention representation of the news content for each candidate news item. Specifically, this includes: Multiple candidate news items are obtained, and word embedding processing is performed on the news title and news content of each candidate news item to obtain the word vector representation of the news title and the word vector representation of the news content of each candidate news item. The news title word vector and the news content word vector of each candidate news are used to extract features to obtain the corresponding news title embedding representation and news content embedding representation. Self-attention calculations are performed on the news title embedding representation and the news content embedding representation of each candidate news item to obtain the corresponding self-attention representation of the news title and the self-attention representation of the news content.

7. A news recommendation method according to claim 6, characterized in that, The self-attention calculation is performed on the news title embedding and the news content embedding of each candidate news item, respectively. The specific calculation formulas for the corresponding self-attention representations of the news title and the news content are as follows: Among them, υ T Here, is a trainable weight vector, Wh is the parameter matrix, b is the bias term, and α is the weight vector. self,i It is the self-attention weight of each word in the query.

8. A news recommendation device, characterized in that, include: Text extraction module: used to obtain multiple candidate news articles, process the news title and news content of each candidate news article to obtain the self-attention representation of the news title and the self-attention representation of the news content of each candidate news article; Cross-attention module: This module performs cross-attention calculation on the self-attention representation of the news title and the self-attention representation of the news content for each candidate news item to obtain the corresponding cross-attention representation of the news content; wherein, the cross-attention representation of the news content is calculated using the following formula: Where, α cross It is the cross-attention weight for each word in the query, r t cross It is a cross-attention representation of news headlines, r c cross This is a cross-attention table for news headlines; Feature fusion module: used to multiply the cross-attention representation of the news title of each candidate news item and the cross-attention representation of the news content of each candidate news item element by element to obtain the news representation corresponding to each candidate news item; Matching score calculation module: Performs a dot product operation between the news representation corresponding to each candidate news and a user representation in the user representation database to obtain the non-normalized matching score between the candidate news and the corresponding user; The sorting module sorts candidate news based on matching scores and recommends news to corresponding users based on the normalized matching scores.

9. A news recommendation device according to claim 8, characterized in that, The text extraction module includes: Word embedding module: used to acquire multiple candidate news articles, process word embeddings for the news title and news content of each candidate news article, and obtain word vector representations for the news title and news content of each candidate news article. Feature extraction module: used to extract the corresponding news title embedding representation and news content embedding representation by performing feature extraction on the news title word vector and news content word vector of each candidate news item respectively; The self-attention calculation module is used to perform self-attention calculations on the news title embedding representation and the news content embedding representation of each candidate news item, respectively, to obtain the corresponding self-attention representation of the news title and the self-attention representation of the news content.

Images