News content recommendation method based on double-layer attention network of value perception

Abstract

This disclosure presents a news content recommendation method based on a value-aware two-layer attention network. One specific implementation of the method includes: inputting candidate news information into a pre-trained two-layer attention network to generate news feature evaluation values; wherein the candidate news information is subjected to value perception to generate candidate news value features; attention feature extraction is performed on the candidate news value features to generate candidate news value attention features; the candidate news information is encoded to generate candidate news codes; feature evaluation is performed on the candidate news value attention features and candidate news codes to generate news feature evaluation values; based on the news feature evaluation values, the candidate news information is added to the corresponding position in a news information recommendation queue; and news information in the news information recommendation queue that meets preset push conditions is sent to the user terminal. This implementation can reduce the waste of transmission resources.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of news content recommendation, specifically to a news content recommendation method based on a value-perception-based two-layer attention network. Background Technology

[0002] Unfiltered news content will impact the connotation of news values, leading to widening differences in individual values, a fragmented news and public opinion landscape, and a serious misalignment between value rationality and technological rationality. Therefore, while accurately providing personalized recommendation services, news recommendation algorithms should also assume the social responsibility of value guidance, forming a dissemination path of "value identification - content distribution - interactive clicks - recognition."

[0003] Therefore, in practice, the following technical problems often arise when using the above methods for news content recommendation:

[0004] First, existing recommendation algorithms neglect news value and rely on tags: they primarily depend on features such as news popularity and user interests, making it difficult to comprehensively reflect news value and easily leading to the spread of vulgar content. Furthermore, existing methods rely on manually labeled single tags, making it difficult to model news value in real time and comprehensively. Consequently, it is difficult to accurately push valuable news content to users, requiring the push of even more news, thus wasting transmission resources.

[0005] Second, implicit feature selection and modeling are difficult: Although existing methods use implicit features such as dwell time to evaluate news quality, they are easily affected by misinformation and are prone to recommending low-quality news. Furthermore, they fail to fully capture the multi-dimensional value of news, thus easily sending too much low-quality news to users, resulting in wasted transmission resources.

[0006] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0008] Some embodiments of this disclosure propose a news content recommendation method based on a value-perceived two-layer attention network to address one or more of the technical problems mentioned in the background section above.

[0009] In a first aspect, some embodiments of this disclosure provide a news content recommendation method based on a value-perception-based two-layer attention network. The method includes: inputting candidate news information into a pre-trained two-layer attention network to generate news feature evaluation values. The two-layer attention network includes: a value feature extraction layer, a value feature attention layer, a news encoding layer, and a value category joint attention layer. The two-layer attention network generates news feature evaluation values ​​through the following steps: perceiving the value of candidate news information through the value feature extraction layer to generate candidate news value features; extracting attention features from the candidate news value features through the value feature attention layer to generate candidate news value attention features; encoding the candidate news information through the news encoding layer to generate candidate news codes; evaluating the candidate news value attention features and candidate news codes through the value category joint attention layer to generate news feature evaluation values; adding the candidate news information to the corresponding position in a news information recommendation queue based on the news feature evaluation values; and sending news information in the news information recommendation queue that meets preset push conditions to a user terminal.

[0010] Secondly, some embodiments of this disclosure provide a news content recommendation device based on a value-perception-based two-layer attention network. The device includes an input and generation unit configured to input candidate news information into a pre-trained two-layer attention network to generate news feature evaluation values. The two-layer attention network includes a value feature extraction layer, a value feature attention layer, a news encoding layer, and a value category joint attention layer. The two-layer attention network generates news feature evaluation values ​​through the following steps: value perception is performed on the candidate news information through the value feature extraction layer to generate candidate news value features; attention features are extracted from the candidate news value features through the value feature attention layer to generate candidate news value attention features; the candidate news information is encoded through the news encoding layer to generate candidate news encoding; and the candidate news value attention features and candidate news encoding are evaluated through the value category joint attention layer to generate news feature evaluation values. A news content recommendation unit is configured to add the candidate news information to the corresponding position in a news information recommendation queue based on the news feature evaluation values, and to send news information in the news information recommendation queue that meets preset push conditions to a user terminal.

[0011] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation of the first aspect above.

[0012] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0013] The various embodiments of this disclosure have the following beneficial effects: The news content recommendation method based on a value-aware dual-layer attention network, as described in some embodiments of this disclosure, can mine news value, accurately push valuable news content to users, and reduce the waste of transmission resources. Specifically, the reason for wasting transmission resources is: ignoring news value and relying on tags: Existing recommendation algorithms mainly rely on features such as news popularity and user interests, which are difficult to comprehensively reflect news value and easily lead to the spread of vulgar content. Furthermore, existing methods rely on manually labeled single tags, making it difficult to model news value in real time and comprehensively. Therefore, it is difficult to accurately push valuable news content to users, resulting in the need to push more news. Based on this, the news content recommendation method based on a value-aware dual-layer attention network, as described in some embodiments of this disclosure, introduces a dual-layer attention network to evaluate the news feature value of each candidate news item. Specifically, firstly, a value feature extraction layer is used to perceive the value of candidate news information, which can be used to determine news value features. Then, an attention feature extraction layer is used to mine deeper features within the value features. Afterwards, encoding processing facilitates the joint extraction of features. Next, through the value category joint attention layer, feature evaluation can be performed on the aforementioned candidate news value attention features and candidate news codes to generate news feature evaluation values. This completes the feature evaluation of candidate news information. Because of the two-layer attention mechanism of attention feature extraction and the value category joint attention layer, more in-depth and related information can be extracted from the features. Furthermore, value feature signals can be incorporated into user interests to achieve the evaluation of candidate news information. This can then be used to recommend news content based on each user's interests. Starting from the deep features of news value reduces reliance on news tags and comprehensively reflects news value. This greatly avoids continuously providing users with vulgar content or news information that users are not interested in. Thus, the goal of accurately pushing valuable news content to users is achieved, thereby reducing the waste of transmission resources. Attached Figure Description

[0014] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0015] Figure 1This is a flowchart of some embodiments of the news content recommendation method based on a value-perceived two-layer attention network according to the present disclosure;

[0016] Figure 2 This is a schematic diagram of the structure of some embodiments of a news content recommendation device based on a value-perception-based two-layer attention network according to the present disclosure;

[0017] Figure 3 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure;

[0018] Figure 4 This is a schematic diagram of the overall structure of the news content recommendation method based on a value-perception-based two-layer attention network disclosed herein. Detailed Implementation

[0019] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0020] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0021] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0022] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0023] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0024] Before performing any of the operations involving the collection, storage, and use of users' personal information (such as users' historical click information on news), the relevant organizations or individuals shall fulfill their obligations, including conducting personal information security impact assessments, informing personal information subjects, and obtaining prior authorization and consent from personal information subjects.

[0025] In practice, existing socially responsible recommendation models struggle to model news value comprehensively and in real-time. Some methods attempt to design responsible recommendation models using manually labeled features such as news authenticity to achieve goals beyond accuracy, but these methods suffer from strong label dependency. Unlike other attributes, news value attributes lack explicit definitions and labels, necessitating multi-faceted analysis and modeling of relationships between different features. Recent research indicates that, in addition to various explicit attribute labels that require prior acquisition, news items also contain rich implicit features, such as news popularity, recent click-through rates, and average user reading time. However, different users have varying degrees of acceptance of mainstream values. News being "popular" does not equate to high value. Therefore, the implicit features used to model news value should be carefully selected to avoid interference from user feedback signals.

[0026] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] Figure 1 A flow 100 illustrating some embodiments of a news content recommendation method based on a value-aware two-layer attention network according to this disclosure is shown. This news content recommendation method based on a value-aware two-layer attention network includes the following steps:

[0028] Step 101: Input candidate news information into a pre-trained two-layer attention network to generate news feature evaluation values.

[0029] In some embodiments, the implementer of the value-aware dual-layer attention network-based news content recommendation method can input candidate news information into a pre-trained dual-layer attention network to generate news feature evaluation values. The dual-layer attention network may include, but is not limited to, at least one of the following network layers: a value feature extraction layer, a value feature attention layer, a news encoding layer, and a value category joint attention layer. The aforementioned dual-layer attention network generates news feature evaluation values ​​through the following steps:

[0030] Step 1: Through the aforementioned value feature extraction layer, the candidate news information is value-perceived to generate candidate news value features.

[0031] Step 2: Through the aforementioned value feature attention layer, attention features are extracted from the aforementioned candidate news value features to generate candidate news value attention features.

[0032] Step three involves encoding the candidate news information using the aforementioned news encoding layer to generate candidate news codes. This news encoding layer can be a news encoder trained using the BERT (Bidirectional Encoder Representations from Transformers) natural language processing model. Furthermore, the news encoder can be used to capture the deep semantics of the news text. Encoding the candidate news information yields the candidate news codes.

[0033] Step four: Through the value category joint attention layer, perform feature evaluation on the above candidate news value attention features and the above candidate news codes to generate news feature evaluation values.

[0034] In some optional implementations of certain embodiments, the execution entity extracts value features from the candidate news information through the value feature extraction layer to generate candidate news value features, which may include the following steps:

[0035] The aforementioned value feature extraction layer extracts value features from the candidate news information to generate candidate news value features. This value feature extraction layer can perform value feature extraction and generate candidate news value features through the following steps:

[0036] Step one: Locate the title and body text within the candidate news information. This can be done using the text identifiers of the candidate news information. Here, the text identifiers can be the text attribute identifiers of the page description language corresponding to the news information at the time of publication.

[0037] Step two involves segmenting the title text and the body text into word groups to generate title word groups and body word groups. This can be done using a pre-defined word segmentation algorithm.

[0038] As an example, the above word segmentation algorithms may include, but are not limited to, at least one of the following: maximum matching word segmentation algorithm, shortest path word segmentation algorithm, CLM (Chinese Language Model), average perceptron word segmentation algorithm, CRF (conditional random field algorithm) word segmentation algorithm, etc.

[0039] Step 3: Determine the title text length value corresponding to the title word segment group and the body text length value corresponding to the body word segment group. Specifically, the number of each title word in the title word segment group can be determined as the title text length value (f1). The number of each body word in the body word segment group can be determined as the body text length value (f2).

[0040] Step four involves deduplicating each word segment in the above-mentioned main text word segmentation group to generate a deduplicated main text word segmentation group and a count of deduplicated main text word segments. Specifically, duplicate words in the above-mentioned main text word segmentation groups are removed to obtain the deduplicated main text word segmentation group. Then, the count of each deduplicated main text word segment in the deduplicated main text word segmentation group is determined as the count of deduplicated main text word segments (f3).

[0041] Step 5: Based on the title word segmentation, the above-mentioned body text word segmentation, the above-mentioned title text length value, the above-mentioned body text length value, and the above-mentioned deduplicated body text word segmentation count, generate candidate news value features.

[0042] Optionally, the execution entity above generates candidate news value features based on the title word segmentation group, the body word segmentation group, the title text length value, the body text length value, and the number of deduplicated body word segments, which may include the following steps:

[0043] The first step is to encode the title words in the aforementioned title word segmentation group to generate title word embedding vectors. This can be done using a pre-defined word vector encoding network.

[0044] As an example, the word vector encoding network mentioned above could be a Word2Vec (word to vector, a related model used to generate word vectors) encoding network.

[0045] The second step is to determine the text sentence embedding vector corresponding to each text word in the above text word segmentation group, thus obtaining a text sentence embedding vector group. Specifically, the text sentence embedding vector corresponding to each text word in the above text word segmentation group is determined through the word vector encoding network described above, thus obtaining a text sentence embedding vector group.

[0046] The third step involves associating the aforementioned title word segmentation embedding vectors and the aforementioned body text sentence embedding vector groups with news text features to generate news text feature values. The text similarity between the title word segmentation embedding vectors and each body text sentence embedding vector in the aforementioned body text sentence embedding vector group can be determined using the following formula, which serves as the news text feature value:

[0047] .

[0048] in, This represents the feature value of the news text. Indicates text similarity. This indicates the total number of sentences in the above candidate news information. Indicates the serial number. This represents the embedding vector for title segmentation. This represents the embedding vector of the main text statement. This indicates that the above text statement is embedded in the vector group of the first... Each text statement is embedded in a vector. The body text representing the candidate news information. This indicates the similarity between the current news headline and the main text.

[0049] Optionally, the above-mentioned execution entity generates candidate news value features based on the title word segmentation group, the body word segmentation group, the title text length value, the body text length value, and the number of deduplicated body word segments. This may further include the following steps:

[0050] The first step is to obtain the publication time and push time of the aforementioned candidate news information. The executing entity can obtain the publication time of the candidate news information via wired or wireless connection. Here, the push time can be the time when the news is displayed in the user's impression list.

[0051] It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G connection, WiFi connection, Bluetooth connection, WiMAX connection, Zigbee connection, UWB (ultra wideband) connection, and other currently known or future wireless connection methods.

[0052] The second step involves extracting news attribute features from the candidate news information based on the aforementioned publication and push times to generate news relevance feature values. This extraction can be performed using the following steps:

[0053] .

[0054] in, Indicates the recent time value, used to characterize the freshness of news (in hours). Indicates a point in time. Indicates the time of publication. Indicates the time point of the push notification. The formula represents the recent news characteristic value mentioned above. The recent value can be scaled down to further reduce the value attribute of old news.

[0055] In practice, the higher the news timeliness characteristic value, the fresher the news, and the more authentic and persuasive the valuable information it contains.

[0056] The third step involves conducting text analysis on the aforementioned candidate news information to generate a confidence score for the news text sentiment. This confidence score can be generated using sentiment analysis tools.

[0057] As an example, sentiment analysis tools may include, but are not limited to, at least one of the following: sentiment analysis tool Emotional Analysis, large language model interface for sentiment analysis, etc.

[0058] The fourth step is to generate candidate news value features based on the aforementioned news relevance feature values ​​and news text tendency confidence scores. Specifically, the aforementioned headline text length value, body text length value, number of unique words in the body text after deduplication, news text feature values, news relevance feature values, and news text tendency confidence scores can be added to the same feature set to construct candidate news value features.

[0059] The value perception process and related content and formulas of the aforementioned value feature extraction layer, as an inventive point of this disclosure, solve the second technical problem mentioned in the background art: "Difficulty in implicit feature selection and modeling: Although existing methods use implicit features such as dwell time to evaluate news quality, they are easily influenced by false information, easily recommending low-quality news, and failing to fully capture the multi-dimensional value of news. Therefore, it is easy to send too much low-quality news to users, resulting in wasted transmission resources." Factors leading to wasted transmission resources often include: difficulty in implicit feature selection and modeling: Although existing methods use implicit features such as dwell time to evaluate news quality, they are easily influenced by false information, easily recommending low-quality news, and failing to fully capture the multi-dimensional value of news. Therefore, it is easy to send too much low-quality news to users. To achieve this effect, firstly, by extracting news value attributes and modeling their contextual features, real-time extraction of value signals based on news content is achieved without relying on pre-completed manual annotation work. Furthermore, introducing value features into user interests not only helps in learning user interests from different perspectives. Then, considering that the concept of news value lacks a clear definition and label, this disclosure proposes to model its value features using various implicit information of news. To acquire news value features in real time without waiting for user interaction feedback, and to reduce the interference of widely disseminated news with low value content on value modeling tasks, this approach focuses only on features related to news content, ignoring user interaction information such as news popularity and recent click volume. Value attribute features of news content are extracted from both news text and news attributes perspectives. Textual features of news content can be used to express the readability and information richness of an article, facilitating the extraction of feature information from the news content. To mitigate the "clickbait" phenomenon, the semantic similarity between the news headline and the body text is evaluated. The closer the embeddings of the two, the more relevant the described topic, and the lower the probability that the news is using a packaged headline to deceive clicks. Therefore, news text feature values ​​are generated, indicating that the closer the semantics of the news headline and body text, the lower the probability of "clickbait" and the higher the news value. Simultaneously, news relevance feature values ​​are generated to represent the freshness of the news. Furthermore, scaling is applied using relevance features to further reduce the value attributes of outdated news. This avoids interference from misinformation and reduces the recommendation of low-quality news, while also enabling the comprehensive capture of the multi-dimensional value of news. Consequently, it significantly reduces the amount of low-quality news sent to users, thus avoiding wasted transmission resources.

[0060] In practice, more responsible news content recommendation methods are designed by introducing auxiliary information such as news quality. These recommendation models determine which news should be recommended to users by evaluating factors such as accuracy and authority. A common approach is to construct user reading history data from fake news detection datasets as input to the news recommendation model, and combine this with a candidate news authenticity prediction auxiliary task to mitigate the spread of fake news. Other common methods use explicit quality scores from the dataset, employing user-system bilateral correction to ensure that the distribution of high-quality news in the recommendation results meets requirements. However, these methods suffer from problems of manual labeling and dataset dependence, making it impossible to obtain value features in real time during the recommendation process. In addition, some methods indirectly predict news content quality based on implicit information such as user reading dwell time. Still other common methods propose weighting news clicks by users based on their reading dwell time during interest modeling. However, it is believed that shorter user reading dwell times usually indicate poor content quality, and user representations are learned from high-quality and low-quality news clicks using a combination of Transformer and additive attention networks. Alternatively, based on normalized reading speed, user reading preferences and satisfaction can be jointly learned to model user interests and train a recommendation model. Furthermore, reading time can be considered positive feedback, skipping articles negative feedback, and article publication and click times neutral feedback. Combining explicit behavior and implicit information can improve user reading satisfaction. This leads to the third technical problem that needs to be addressed: "Although these methods explore implicit features reflecting news quality, the user interaction signals they rely on may be interfered with by false information, malicious manipulation, etc. Simply increasing the probability of recommending highly popular and widely disseminated news may introduce fragmented, entertainment-oriented, and low-quality news. At the same time, these methods only use a single feature to evaluate news quality, failing to comprehensively extract the value attributes contained in the news content, thus wasting news transmission resources." Therefore, the following solution can be adopted to address the above-mentioned third technical problem.

[0061] In some optional implementations of certain embodiments, the execution entity extracts attention features from the candidate news value features through the aforementioned value feature attention layer to generate candidate news value attention features, which may include the following steps:

[0062] The first step is to obtain the historical information embedding vector sequence of user news clicks. This involves first acquiring the user's historical news clicks. Then, the aforementioned value feature extraction layer extracts the candidate news value features corresponding to each historical news item as historical information embedding vectors, resulting in the historical information embedding vector sequence.

[0063] The second step involves extracting attention features from the aforementioned candidate news value features and the aforementioned historical information embedding vector sequence through the aforementioned value feature attention layer, in order to generate candidate news value attention features.

[0064] The aforementioned value feature attention layer generates candidate news value attention features through the following steps:

[0065] Step 1: Through the embedding layer in the value feature attention layer, the candidate news value features and each of the historical information embedding vectors in the historical information embedding vector sequence are projected onto the same feature space to generate the information value feature embedding vector sequence.

[0066] Step two: Determine the relevance feature sequence corresponding to the above information value feature embedding vector sequence. The relevance feature sequence corresponding to the above information value feature embedding vector sequence can be determined using the following formula:

[0067] .

[0068] in, This represents the embedding vector of information value features. and This represents any two information value feature embedding vectors in the above information value feature embedding vector sequence. and This represents the projection matrix after training and adjustment. This represents the attention weight between two information value feature embedding vectors. Indicates inclusion Correlation characteristics with other characteristics.

[0069] Step 3: Determine the value feature embedding vector sequence corresponding to the above-mentioned relevance feature sequence as candidate news value attention features. The value feature embedding vector corresponding to the above-mentioned relevance feature sequence can be determined through the following steps:

[0070] .

[0071] in, This refers to the Multilayer Perceptron. This represents the value feature embedding vector in the candidate news value attention feature. This indicates the corresponding first among the above candidate news value attention features. Value feature embedding vector of a feature.

[0072] In some optional implementations of certain embodiments, the execution entity performs feature evaluation on the candidate news value attention features and the candidate news codes through the value category joint attention layer to generate news feature evaluation values, which may include the following steps:

[0073] The first step is to obtain the news category embedding vector sequence of user news clicks. This news category embedding vector sequence can be generated by pre-encoding each historical news item clicked by the user using the aforementioned news encoder. Then, the corresponding news category embedding vector can be determined using the aforementioned word vector encoding network, resulting in the news category embedding vector sequence.

[0074] The second step involves determining the news value attention weights corresponding to the candidate news value attention features and each historical information embedding vector in the historical information embedding vector sequence, based on the aforementioned news category embedding vector sequence. The corresponding news value attention weights can be determined by concatenating the candidate news value attention features with each historical information embedding vector in the aforementioned historical information embedding vector sequence using the following formula:

[0075] .

[0076] in, and This represents the learnable query vector and projection matrix in the attention network. This indicates a concatenation operation between the two vectors on either side of the semicolon. The value feature embedding vector corresponds to the first News value attention weights for each historical information embedding vector. This represents the transpose of a matrix. Indicates the serial number. This indicates the length of the historical information embedding vector sequence. This indicates the first element in the above news category embedding vector sequence. Each news category embedding vector. This represents the first of the above candidate news value attention features. Each value feature is embedded in a vector. This indicates that the above historical information is embedded in the vector sequence of the first... A historical information embedding vector. This indicates the news value and attention weight.

[0077] In practice, splicing operations can be used to capture the contextual features of clicked news sequences and introduce news value features into user interests.

[0078] The third step is to determine the candidate news category embedding vectors corresponding to the aforementioned candidate news value attention features. This can be achieved using the word vector encoding network described above.

[0079] The fourth step involves generating a news category-aware weight sequence based on the aforementioned candidate news category embedding vectors and the news category embedding vector sequence. This sequence can be generated using the following formula:

[0080] .

[0081] in, This represents the first element in the above news category embedding vector sequence. Each news category embedding vector. This represents the embedding vector of the above candidate news categories. This represents the first [number] in the above news category perception weight sequence. The perceived weight of each news category. This indicates the perceived weight of news categories.

[0082] The fifth step involves normalizing the aforementioned news value attention weight sequence and news category perception weight sequence to generate a value category perception attention weight sequence. This can be achieved through the following...

[0083] .

[0084] in, This represents the first value category in the above value category perception attention weight sequence. Each value category is perceived by attention weights. These are learnable parameters used during the training process to control the strength of the correction for the category-aware weights.

[0085] Step 6: Generate news feature evaluation values ​​based on the aforementioned news category perception weight sequence and value category perception attention weight sequence. Specifically, the news feature evaluation value can be determined by summing the products of the corresponding news category perception weights and value category perception attention weights in the aforementioned news category perception weight sequence and value category perception attention weight sequence.

[0086] The aforementioned process from attention feature extraction to feature evaluation, along with its related content and formulas, serves as an inventive point of this disclosure, addressing the third technical problem mentioned in the background: "Although these methods explore implicit features reflecting news quality, the user interaction signals they rely on may be interfered with by false information, malicious operations, and other behaviors. Simply increasing the probability of highly popular and widely disseminated news being recommended may instead introduce fragmented, entertaining, and vulgar low-quality news. Furthermore, these methods only use a single aspect of features to evaluate news quality, failing to comprehensively extract the value attributes contained in the news content, thus wasting news transmission resources." Solving these factors can reduce the waste of news resources. To achieve this effect, firstly, a value feature attention layer auxiliary model further learns the correlation between different combinations of various content value features, mining deep features of value features and outputting the value feature embedding of clicked news. Secondly, the value-category joint attention layer aims to introduce value feature signals into user interests. Using news content embedding and value feature embedding as input, it calculates the contextual weights of news value features and news content through an attention mechanism to reduce the impact of low-quality negative news on user interest generation. Meanwhile, for candidate news items, the model utilizes a value feature extraction layer to obtain their value feature set, and after evaluating the value score, incorporates this score into the final click probability calculation process to achieve the value-guided role of the recommendation model. Thus, by extracting news value attributes and modeling their contextual features, it achieves real-time extraction of value signals based on news content without relying on pre-completed manual annotation. Furthermore, incorporating value features into user interests not only helps learn user interests from different perspectives but also enables value-aware news recommendation algorithms. This reduces the waste of news transmission resources.

[0087] As an example, such as Figure 4 The diagram shows the overall structure of the news content recommendation method.

[0088] Step 102: Based on the news feature evaluation value, add the candidate news information to the corresponding position in the news information recommendation queue, and send the news information in the news information recommendation queue that meets the preset push conditions to the user terminal.

[0089] In some embodiments, the executing entity can add the candidate news information to the corresponding position in the news information recommendation queue based on the news feature evaluation value, and send news information in the news information recommendation queue that meets the preset push conditions to the user terminal. Specifically, news information in the news information recommendation queue that meets the preset push conditions can be sent to the user terminal corresponding to the user. Here, the preset push conditions can be the first N news information in the news information recommendation queue.

[0090] In some optional implementations of certain embodiments, the execution entity adds the candidate news information to the corresponding position in the news information recommendation queue based on the news feature evaluation value, which may include the following steps:

[0091] The candidate news items are added to their corresponding positions in a preset news recommendation queue according to the order of their news feature evaluation values, resulting in the added news recommendation queues. Each news recommendation queue can correspond to one user. The order can be from highest to lowest news feature evaluation value.

[0092] The various embodiments of this disclosure have the following beneficial effects: The news content recommendation method based on a value-aware dual-layer attention network, as described in some embodiments of this disclosure, can mine news value, accurately push valuable news content to users, and reduce the waste of transmission resources. Specifically, the reason for wasting transmission resources is: ignoring news value and relying on tags: Existing recommendation algorithms mainly rely on features such as news popularity and user interests, which are difficult to comprehensively reflect news value and easily lead to the spread of vulgar content. Furthermore, existing methods rely on manually labeled single tags, making it difficult to model news value in real time and comprehensively. Therefore, it is difficult to accurately push valuable news content to users, resulting in the need to push more news. Based on this, the news content recommendation method based on a value-aware dual-layer attention network, as described in some embodiments of this disclosure, introduces a dual-layer attention network to evaluate the news feature value of each candidate news item. Specifically, firstly, a value feature extraction layer is used to perceive the value of candidate news information, which can be used to determine news value features. Then, an attention feature extraction layer is used to mine deeper features within the value features. Afterwards, encoding processing facilitates the joint extraction of features. Next, through the value category joint attention layer, feature evaluation can be performed on the aforementioned candidate news value attention features and candidate news codes to generate news feature evaluation values. This completes the feature evaluation of candidate news information. Because of the two-layer attention mechanism of attention feature extraction and the value category joint attention layer, more in-depth and related information can be extracted from the features. Furthermore, value feature signals can be incorporated into user interests to achieve the evaluation of candidate news information. This can then be used to recommend news content based on each user's interests. Starting from the deep features of news value reduces reliance on news tags and comprehensively reflects news value. This greatly avoids continuously providing users with vulgar content or news information that users are not interested in. Thus, the goal of accurately pushing valuable news content to users is achieved, thereby reducing the waste of transmission resources.

[0093] Further reference Figure 2As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a news content recommendation device based on a value-perception two-layer attention network. These device embodiments are similar to... Figure 1 Corresponding to the method embodiments shown, the device can be specifically applied to various electronic devices.

[0094] like Figure 2 As shown, a news content recommendation device 200 based on a value-aware two-layer attention network in some embodiments includes: an input and generation unit 201 and a news content recommendation unit 202. The input and generation unit 201 is configured to input candidate news information into a pre-trained two-layer attention network to generate news feature evaluation values. The two-layer attention network includes a value feature extraction layer, a value feature attention layer, a news encoding layer, and a value category joint attention layer. The two-layer attention network generates news feature evaluation values ​​through the following steps: the value feature extraction layer performs value perception on the candidate news information to generate candidate news value features; the value feature attention layer extracts attention features from the candidate news value features to generate candidate news value attention features; the news encoding layer encodes the candidate news information to generate candidate news codes; and the value category joint attention layer evaluates the candidate news value attention features and the candidate news codes to generate news feature evaluation values. The news content recommendation unit 202 is configured to add the candidate news information to the corresponding position in the news information recommendation queue based on the news feature evaluation values, and send news information in the news information recommendation queue that meets preset push conditions to the user terminal.

[0095] It is understandable that the units described in the device 200 are related to the reference. Figure 1 The steps in the described method correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method also apply to the device 200 and the units contained therein, and will not be repeated here.

[0096] The following is for reference. Figure 3 It shows a schematic diagram of the structure of an electronic device (e.g., a computing device) 300 suitable for implementing some embodiments of the present disclosure. Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0097] like Figure 3As shown, the electronic device 300 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory 302 or a program loaded from a storage device 308 into a random access memory 303. The random access memory 303 also stores various programs and data required for the operation of the electronic device 300. The processing unit 301, the read-only memory 302, and the random access memory 303 are interconnected via a bus 304. An input / output interface 305 is also connected to the bus 304.

[0098] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic device 300 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 An electronic device 300 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 3 Each box shown can represent a device or multiple devices as needed.

[0099] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 309, or installed from a storage device 308, or installed from a read-only memory 302. When the computer program is executed by the processing device 301, it performs the functions defined in the methods of some embodiments of this disclosure.

[0100] It should be noted that, in some embodiments of this disclosure, the computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0101] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0102] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs. When the electronic device executes the aforementioned one or more programs, the electronic device causes the following steps: inputting candidate news information into a pre-trained two-layer attention network to generate news feature evaluation values. The two-layer attention network includes a value feature extraction layer, a value feature attention layer, a news encoding layer, and a value category joint attention layer. The two-layer attention network generates news feature evaluation values ​​through the following steps: performing value perception on candidate news information through the value feature extraction layer to generate candidate news value features; extracting attention features from the candidate news value features through the value feature attention layer to generate candidate news value attention features; encoding the candidate news information through the news encoding layer to generate candidate news encoding; evaluating the candidate news value attention features and the candidate news encoding through the value category joint attention layer to generate news feature evaluation values; adding the candidate news information to the corresponding position in the news information recommendation queue based on the news feature evaluation values; and sending news information in the news information recommendation queue that meets preset push conditions to the user terminal.

[0103] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0104] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0105] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including an input and generation unit and a news content recommendation unit. The names of these units are not necessarily limiting in certain circumstances; for example, a news content recommendation unit may also be described as "a unit that sends news information to a user terminal."

[0106] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0107] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A news content recommendation method based on a value-perception-based two-layer attention network, comprising: Candidate news information is input into a pre-trained two-layer attention network to generate news feature evaluation values. The two-layer attention network includes: a value feature extraction layer, a value feature attention layer, a news encoding layer, and a value category joint attention layer. The two-layer attention network generates news feature evaluation values ​​through the following steps: The value feature extraction layer performs value perception on candidate news information to generate candidate news value features. This process includes: The process involves locating the title text and body text in the candidate news information; performing word segmentation on the title text and body text to generate title word groups and body word groups; determining the title text length value corresponding to the title word group and the body text length value corresponding to the body word group; performing word deduplication on each body word in the body word group to generate deduplicated body word groups and the number of deduplicated body words; obtaining the publication time point and push time point of the candidate news information; extracting news attribute features from the candidate news information based on the publication time point and the push time point to generate news proximity feature values; performing text analysis on the candidate news information to generate news text tendency confidence scores; and adding the title text length value, body text length value, number of deduplicated body words, news text feature values, news proximity feature values, and news text tendency confidence scores to the same feature set to construct candidate news value features. Through the value feature attention layer, attention features are extracted from the candidate news value features to generate candidate news value attention features; The candidate news information is encoded through the news encoding layer to generate candidate news codes. The value category joint attention layer is used to evaluate the candidate news value attention features and the candidate news codes to generate news feature evaluation values. Based on the news feature evaluation value, the candidate news information is added to the corresponding position in the news information recommendation queue, and the news information in the news information recommendation queue that meets the preset push conditions is sent to the user terminal.

2. The method according to claim 1, wherein, The generated candidate news value features include: The title segments in the title segmentation group are encoded to generate title segmentation embedding vectors; Determine the text sentence embedding vector corresponding to each text word in the text word segmentation group to obtain the text sentence embedding vector group; The title word segmentation embedding vector and the body sentence embedding vector group are correlated with news text features to generate news text feature values.

3. The method according to claim 1, wherein, The step of extracting attention features from the candidate news value features through the value feature attention layer to generate candidate news value attention features includes: Obtain the embedded vector sequence of users' historical news clicks; The value feature attention layer extracts attention features from the candidate news value features and the historical information embedding vector sequence to generate candidate news value attention features. The value feature attention layer generates candidate news value attention features through the following steps: Through the embedding layer in the value feature attention layer, the candidate news value features and each of the historical information embedding vectors in the historical information embedding vector sequence are projected onto the same feature space to generate an information value feature embedding vector sequence. Determine the relevance feature sequence corresponding to the information value feature embedding vector sequence; The value feature embedding vector sequence corresponding to the relevance feature sequence is determined as the candidate news value attention feature.

4. The method according to claim 1, wherein, The step of adding the candidate news information to the corresponding position in the news information recommendation queue based on the news feature evaluation value includes: The candidate news information is added to the corresponding positions in the preset news information recommendation queue according to the order corresponding to the news feature evaluation values, so as to obtain the added news information recommendation queue, wherein each news information recommendation queue corresponds to one user.

5. A news content recommendation device based on a value-perception-based two-layer attention network, comprising: The input and generation units are configured to input candidate news information into a pre-trained two-layer attention network to generate news feature evaluation values. The two-layer attention network includes: a value feature extraction layer, a value feature attention layer, a news encoding layer, and a value category joint attention layer. The two-layer attention network generates news feature evaluation values ​​through the following steps: The value feature extraction layer performs value perception on candidate news information to generate candidate news value features. This process includes: The process involves locating the title text and body text in the candidate news information; performing word segmentation on the title text and body text to generate title word groups and body word groups; determining the title text length value corresponding to the title word group and the body text length value corresponding to the body word group; performing word deduplication on each body word in the body word group to generate deduplicated body word groups and the number of deduplicated body words; obtaining the publication time point and push time point of the candidate news information; extracting news attribute features from the candidate news information based on the publication time point and the push time point to generate news proximity feature values; performing text analysis on the candidate news information to generate news text tendency confidence scores; and adding the title text length value, body text length value, number of deduplicated body words, news text feature values, news proximity feature values, and news text tendency confidence scores to the same feature set to construct candidate news value features. Through the value feature attention layer, attention features are extracted from the candidate news value features to generate candidate news value attention features; The candidate news information is encoded through the news encoding layer to generate candidate news codes. The value category joint attention layer is used to evaluate the candidate news value attention features and the candidate news codes to generate news feature evaluation values. The news content recommendation unit is configured to add the candidate news information to the corresponding position in the news information recommendation queue based on the news feature evaluation value, and to send the news information in the news information recommendation queue that meets the preset push conditions to the user terminal.

6. An electronic device, comprising: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-4.

7. A computer-readable medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-4.

Images