Audio summary generation method, device and electronic equipment for subscription content
By generating personalized audio summaries through attention mechanisms and large language models, the problem of insufficient user experience in traditional content recommendation systems is solved, enabling efficient consumption and personalized presentation of information.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- MIGU MUSIC CO LTD
- Filing Date
- 2025-12-24
- Publication Date
- 2026-05-26
AI Technical Summary
Traditional content recommendation systems are too homogeneous in their presentation, resulting in users being unable to consume information in a timely and effective manner, leading to a poor user experience.
By calculating user interest levels through an attention mechanism, an audio summary is generated. Combined with a large language model and comment popularity, a personalized audio file is produced, which includes an article summary and insightful comments, tailored to user interests.
It improves the efficiency of content consumption, enhances the user experience, and reduces the interference of information overload.
Smart Images
Figure CN122087103A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of data processing technology, and in particular to a method, apparatus, and electronic device for generating audio summaries of subscribed content. Background Technology
[0002] Digital content recommendation systems, as a core technology for information distribution in the mobile internet, are widely used in personalized content services for news apps. With the fragmentation of content consumption scenarios, while traditional recommendation systems can achieve basic content distribution, their output format remains primarily text-based push notifications and waterfall-style displays. In this approach, users can independently enter corresponding channels to select their preferred content. However, the sheer volume of content, homogenized presentation formats, and limited content consumption methods can lead to users being unable to consume content promptly and effectively, resulting in a subpar experience. Summary of the Invention
[0003] This disclosure provides a method, apparatus, and electronic device for generating audio summaries of subscribed content. Its main purpose is to increase the presentation of information and simplify the complexity of information manipulation for users, allowing them to listen to subscribed audio without having to manually view the article content.
[0004] According to a first aspect of this disclosure, a method for generating audio summaries of subscription content is provided, comprising:
[0005] The weights of the content features of articles that users pay attention to are calculated using an attention mechanism to obtain a predetermined number of recommended articles with high user interest and the content features of the recommended articles. The text information of the illustrations in the recommended article, the text fragments of the underlined parts, the content features of the recommended article, and the main text are input into the large language model according to the preset prompt word template to generate the text summary corresponding to the recommended article; The most interesting comments on the recommended articles are obtained based on the semantic matching degree and comment popularity metrics. The text summary and insightful comments are combined into audio playback data, and an audio file is generated using text-to-speech technology. The audio file is then added to a dynamic audio playlist based on the playback parameters configured by the user.
[0006] Optionally, the step of calculating the weights of the content features of articles that the user focuses on through an attention mechanism to obtain a predetermined number of recommended articles with high user interest and the content features of the recommended articles includes: The attention mechanism is used to calculate the weights of each content feature that users pay attention to in each article created by the creator. The weights corresponding to each content feature of each article are weighted and summed to obtain the comprehensive weight of each article; The comprehensive weights are input into the interest level prediction model to evaluate the user's interest level; the higher the evaluation score, the higher the interest level. A predetermined number of articles are selected as recommended articles based on the degree of interest from high to low, and a predetermined number of content features that have a significant impact on the evaluation score are obtained as the content features of the recommended articles.
[0007] Optionally, the step of calculating the weights corresponding to each content feature of each article created by the creator through the attention mechanism includes: The acquired user features and the content features of each article created by the creators followed by the users are vectorized using embedding matrices. The weights of each content feature corresponding to the user's attention to each article created by the creator are obtained by performing an inner product operation with each vectorized user feature and the vectorized content feature.
[0008] Optionally, obtaining insightful comments on the recommended articles based on semantic matching degree and comment popularity metrics further includes: Obtain the user's interest tags, as well as the comment text information of the comments corresponding to the recommended articles; Input the user's interest tags and the comment text information into the large model, calculate the semantic relevance between the interest tags and each comment, and obtain the first normalized score of each comment; The comment text information is normalized based on the comment popularity index to obtain a normalized popularity value for each comment. The first score and popularity value are weighted and calculated to obtain the comprehensive comment value for each comment; A predetermined number of comments corresponding to the larger overall review values are selected as the featured comments of the recommended articles.
[0009] Optionally, the normalization processing of the comment text information based on the comment popularity index to obtain a normalized popularity value for each comment, wherein the comment popularity index includes at least the number of likes and replies, including: The number of likes and replies to the comments are normalized to obtain normalized like score and normalized reply score. The score for likes and the score for replies are weighted to obtain a normalized popularity value for each comment.
[0010] Optionally, the step of combining the personalized text summary with insightful comments into voice playback data, generating an audio file using text-to-speech technology, and adding the audio file to a dynamic audio playlist according to user-configured playback parameters further includes: The personalized text summaries and insightful comments undergo preprocessing operations including word segmentation, grammatical analysis, and semantic understanding to identify sentence structure, punctuation, and sentiment in the text. Based on the user-configured speech rate settings and speech playback preference parameters, the corresponding speech synthesis model is selected to synthesize the text into a speech signal. The speech signal is processed by a vocoder to optimize its sound quality.
[0011] Optional, also includes: Based on the user-configured subscription expiration time parameter, the system manages the expiration time of audio files in the dynamic audio playlist and automatically removes audio files that have exceeded the preset expiration time.
[0012] According to a second aspect of this disclosure, an audio summary generation apparatus for subscription content is provided, comprising: The first acquisition unit is used to calculate the weight of the content features of the articles that the user pays attention to through an attention mechanism, and to acquire a predetermined number of recommended articles with high user interest and the content features of the recommended articles. The first processing unit is used to input the text information of the illustrations in the recommended article, the text fragments of the underlined parts, the content features of the recommended article, and the main text of the article into a large language model according to a preset prompt word template to generate a text summary corresponding to the recommended article. The second acquisition unit acquires insightful comments about the recommended articles based on the semantic matching degree and comment popularity metrics of the recommended articles. The second processing unit is used to combine the text summary and the highlights into voice playback data, generate an audio file through text-to-speech technology, and add the audio file to the dynamic audio playlist according to the playback parameters configured by the user.
[0013] According to a third aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in the first aspect above.
[0014] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are configured to cause the computer to perform the method described in the first aspect above.
[0015] This disclosure provides a method, apparatus, and electronic device for generating audio summaries of subscribed content. By employing an attention mechanism, the method considers the weights corresponding to various content features of each article to select a predetermined number of articles recommended based on user interest levels and the content features of these recommended articles. Then, it adds text from images, underlined text fragments, content features, and the main text of the article to a preset prompt template. A large language model is used to determine the article summary. The large language model, combined with content features, scores each comment on the article. Each comment is scored based on its click and reply counts to determine its popularity. Selected comments are then filtered based on the large language model score and popularity score. This method, using the article summary and selected comments to generate audio summaries of the article, effectively alleviates the problem of information overload for users. By combining personalized recommendations, multimodal text summarization generation, and voice broadcasting, it improves content consumption efficiency and scenario adaptability, reduces user disturbance, and enhances the reading experience.
[0016] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description
[0017] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 A flowchart illustrating the method for generating audio summaries of subscription content provided in this embodiment of the disclosure; Figure 2 A flowchart illustrating another method for generating audio summaries of subscribed content provided in this disclosure embodiment; Figure 3 A flowchart illustrating another method for generating audio summaries of subscribed content provided in this disclosure embodiment; Figure 4 A list of sample examples used in an audio summary generation method for subscription content provided in this disclosure embodiment; Figure 5 This is a schematic diagram illustrating a content feature weight calculation process provided in an embodiment of the present disclosure; Figure 6 This is a schematic diagram illustrating the acquisition of recommended articles provided in an embodiment of the present disclosure; Figure 7 This is another schematic diagram illustrating the acquisition of recommended articles provided in this embodiment of the disclosure; Figure 8 This is a schematic diagram of text digest generation provided in an embodiment of the present disclosure; Figure 9 A schematic diagram of comment filtering provided in an embodiment of this disclosure; Figure 10 Another schematic diagram of comment filtering provided for embodiments of this disclosure; Figure 11 A schematic diagram of a summary voice subscription provided in an embodiment of this disclosure; Figure 12 Another schematic diagram of summary voice subscription provided for an embodiment of this disclosure; Figure 13 A schematic diagram of an audio summary generation device for subscription content provided in an embodiment of this disclosure; Figure 14 A schematic block diagram of an example electronic device 500 provided for embodiments of this disclosure. Detailed Implementation
[0018] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.
[0019] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0020] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.
[0021] This embodiment provides a method for generating audio summaries of subscribed content, such as... Figure 1 As shown, it includes: 101. Calculate the weights of the content features of articles that users pay attention to through an attention mechanism, and obtain a predetermined number of recommended articles with high user interest and the content features of the recommended articles.
[0022] The attention mechanism in this embodiment, a key technology in deep learning, simulates how human attention is allocated. It allows the model to focus more on crucial features when processing the interaction between user and content features, thereby improving the model's accuracy and interpretability.
[0023] When calculating the weights of article content features that users pay attention to using the attention mechanism, an estimated score can be given to the article based on these weights. The higher the estimated score, the higher the user's level of interest. When obtaining a predetermined number of recommended articles with high user interest, the weights of the article content features can be sorted, and the top-weighted articles can be selected as the reasons for recommendation. Specific implementation depends on the specific needs, and this embodiment does not impose any limitations on this.
[0024] 102. The text information of the illustrations in the recommended article, the text fragments of the underlined parts, the content features of the recommended article, and the main text are input into the large language model according to the preset prompt word template to generate the text summary corresponding to the recommended article.
[0025] The underline function in articles is a common feature in current text content presentation formats. Readers can underline passages they feel are well-written, important, or even controversial, and other readers can see these underlines. The underline function can be repeated. Essentially, it allows readers to vote on the key points of the article; the more underlines, the more attention that passage deserves and the more likely it is to be extracted as a prompt for generating personalized summaries.
[0026] Cue word engineering refers to the process of guiding artificial intelligence models (such as large language models) to generate outputs that meet user expectations by carefully designing prompts. In the field of artificial intelligence, especially in Natural Language Processing (NLP), prompts are text instructions that users input into the model. The model then uses these prompts to complete NLP tasks such as intelligent question answering, sentiment recognition, and text summarization.
[0027] A key task in this embodiment is to guide the big oracle model to generate personalized article summaries using a carefully constructed cue word template. The key to accomplishing this task is to input the cue content into the big language model according to the predetermined cue word template. This embodiment constructs a structured and semantically rich cue word template by integrating the main text of the article, images, and user behavior data (text fragments of the underlined parts), thereby guiding the big language model to generate summary content that matches the user's interests.
[0028] 103. Obtain insightful comments on the recommended articles based on the semantic matching degree and comment popularity metrics.
[0029] In this application, the semantic matching degree is calculated by a large model to determine the semantic similarity between user interest tags and comments, with an output range of [0,1]. The comment popularity is calculated using normalized numbers of likes and replies. Finally, a comprehensive score is awarded to each comment, and the comments are then ranked based on this score. A predetermined number of comments are selected as supplementary content for the audio summary.
[0030] 104. Combine the text summary and insightful comments into audio playback data, generate an audio file using text-to-speech technology, and add the audio file to a dynamic audio playlist according to the playback parameters configured by the user.
[0031] In this embodiment, an attention mechanism is used to select a predetermined number of articles based on user interest levels and their content features, considering the weights of various content features of each article. Then, text from images, underlined text fragments, content features, and the main text of the article are added to a preset prompt template. A large language model is used to determine the article summary. The large language model, combined with content features, scores each comment on the article. Each comment is scored based on its clicks and replies, and then selected comments are chosen based on both the large language model score and the popularity score. This method, using the article summary and selected comments to generate subscription audio, effectively alleviates the problem of information overload for users. By combining personalized recommendations, multimodal text summarization, and voice broadcasting, content consumption efficiency and scenario adaptability are improved, user disturbance is reduced, and the reading experience is enhanced.
[0032] Furthermore, when executing the method for generating audio summaries of subscribed content, one can refer to, for example... Figure 2 The content shown is elaborated as follows.
[0033] Furthermore, when calculating the weights of the content features of articles that users pay attention to through an attention mechanism, and obtaining a predetermined number of recommended articles with high user interest and the content features of those recommended articles, such as... Figure 3 As shown, this can be achieved using, but is not limited to, the following methods, which include: 201. Calculate the weights of each content feature of each article created by the creator based on the attention mechanism.
[0034] In this embodiment, when calculating the weights corresponding to various content features of each article created by a creator through the attention mechanism, it is necessary to prepare relevant samples of articles followed by the user. First, positive and negative samples are constructed based on user behavior data. Positive samples are articles clicked by the user, and negative samples are content not clicked. User features include gender, age, region, interest preference tags, and articles that the user frequently reads, such as financial and military articles. Content features include at least category features, keyword features, and high-frequency words in the article: category features include, for example, an article's primary category is financial and its secondary category is investment; keyword features include, for example, the author's name; and high-frequency words in the article include, for example, "Product A" and "Product B" in the article's images. After these user features and content features are extracted, such as... Figure 4 As shown, the user characteristics are: male, age 20, interests: finance, military. Content characteristics include: finance, investment, a certain studio, product A, product B, etc. Some of this information is optional, filled in by the author, while the other part is extracted through text mining, such as using, but not limited to, keyword extraction and topic mining techniques. Specifically, this embodiment does not limit the scope of these techniques.
[0035] In this embodiment, the attention mechanism can be, but is not limited to, a deep learning model of target attention.
[0036] When calculating the weights of various content features corresponding to user attention to each article created by a creator using the attention mechanism, the following methods can be used, but are not limited to: First, the acquired user features and the content features of each article created by the creators followed by the users are vectorized using embedding matrices.
[0037] It's important to note here that the vectorization process for user features and content features involves converting user features (usually discrete features) into dense vectors. This process can be simply understood as a "table lookup" process. This "table lookup" process can specifically include, but is not limited to, the following steps: Step 1: Enumerate all features, totaling vocab_size features; and define a unique number for each feature; Step 2: Randomly initialize the embedding matrix [vocab_size,emb_size] of the parameters, where emb_size is the length of the dense vector; The third step: During model training / inference, the dense vectors corresponding to the features in the sample are queried to enter the model training / inference process. During training, these vector parameters will also change as the model is optimized.
[0038] For example, such as Figure 5As shown, user characteristics are: male, 20 years old, military, finance; content characteristics are: finance, investment, a certain studio, product A, product B; the embedding matrix is used for quantification to obtain a user feature vector and a content feature vector respectively.
[0039] Secondly, the vectorized user features are individually multiplied with the vectorized content features to obtain the weights corresponding to each content feature of each article created by the creator that the user follows.
[0040] In this embodiment, after the interaction process values of user features and content features are vectorized, the inner product of each user feature and the content feature is calculated one by one. The larger the inner product, the more important the content feature is, and the more the model will pay attention to these features with higher weights.
[0041] like Figure 5 As shown, the corresponding content feature weight e is obtained by the inner product of the user vector and the content vector. Different content features correspond to different weight values, namely e1, e2, e3, e4, and e5.
[0042] In addition to the above, after obtaining the different weights corresponding to different content features, this embodiment can also perform softmax normalization on the content feature weights e to obtain the final normalized probability distribution w:
[0043] Formula 1 The larger the w value, the more important the corresponding content feature is, and the more likely it is to be used as a recommendation reason.
[0044] 202. The weights corresponding to each content feature of each article are summed in a weighted manner to obtain the comprehensive weight of each article.
[0045] In this embodiment, it is necessary to further explain that the weight w of each content feature is calculated through an attention mechanism. These weights reflect the importance of different features in the current context. Based on these weights w, the content features are weighted and summed to generate a comprehensive feature representation, namely the comprehensive weight of each article, i.e., last_embedding.
[0046] 203. Input the comprehensive weight into the interest level prediction model to evaluate the user's interest level. The higher the evaluation score, the higher the interest level.
[0047] In embodiments of the present invention, such as Figure 6 and Figure 7 As shown, when last_embedding is input into a fully connected neural network, the model outputs a predicted score to assess the user's level of interest in the content.
[0048] The model scores the candidate articles to be recommended to the user. Using an attention mechanism, the model identifies the three content features that have the greatest impact on the score. These three features not only have the highest weight, but in principle, they are the key factors the model considers when determining if a user might like an article. Therefore, these three features can be used as the recommendation reasons to provide the user with an intuitive and persuasive explanation, helping them better understand the connection between the recommended content and their own interests.
[0049] 204. Select a predetermined number of articles as recommended articles in descending order of interest level, and obtain a predetermined number of content features that have a significant impact on the evaluation score as the content features of the recommended articles.
[0050] For the trained model, we score each content candidate set based on a user's features, and obtain the three most weighted content features as recommendation reasons. As shown in Table 1 below: we select the three articles with the highest estimated evaluation scores and obtain the three most weighted content features as recommendation reasons.
[0051]
[0052] Table 1 Compared to traditional recommendation ranking models, this recommendation method, in addition to outputting personalized ranking results, can also output the reasons why users like the content.
[0053] Furthermore, when inputting the illustrated text information, underlined text fragments, content features of the recommended article, and the main text of the recommended article into a large language model to generate a text summary corresponding to the recommended article according to a preset prompt word template, the following methods can be used, but are not limited to: When generating text summaries, the text information in the illustrations of the recommended articles, the underlined text fragments, the content features of the recommended articles, and the main text are all crucial. In obtaining the text information in the illustrations of the recommended articles, OCR (Optical Character Recognition) technology can be used, but is not limited to, this technology. OCR is a technique that uses computers to automatically recognize and parse text information in images. It can extract text content from images into editable digital text and is widely used in document scanning, image processing, information extraction, and many other fields.
[0054] In this embodiment, when comparing image content, text in the article illustrations can be extracted using OCR technology to obtain a comparison between a genuine product and a counterfeit product.
[0055] In this embodiment, the underline function is a common feature in current text content presentation formats. Readers can underline passages they feel are well-written, important, or even controversial, and other readers can see this. They can also repeat the underlining. Essentially, it allows readers to vote on the key content of the article; the more underlines, the more attention that passage should receive and the more likely it is to be extracted as a prompt for generating a personalized summary.
[0056] Furthermore, cue word templates are also crucial when generating personalized text summaries using large language models. Cue word engineering refers to the process of guiding AI models (such as large language models) to generate outputs that meet user expectations through carefully designed prompts. In the field of artificial intelligence, especially in Natural Language Processing (NLP), cue words are text instructions that users input into the model, which then uses these prompts to complete NLP tasks such as intelligent question answering, sentiment recognition, and text summarization.
[0057] In this embodiment, a pre-constructed prompt word template is used to guide the Big Prophecy model in generating personalized article summaries. The process is as follows: Figure 8 As shown, it includes: This prompt template includes different prompts, such as basic prompt fields, body text prompt fields, recommendation reason prompt fields, illustration keyword prompt fields, and user-underlined content prompt fields.
[0058] In this embodiment, based on the above-mentioned prompt word template settings, the prompt word template table can be as shown in Table 2 below:
[0059] Table 2 Furthermore, in embodiments of the present invention, the function of generating extended comment filtering for the summary of subscribed content is also presented as audio at the end of the main text summary. Comments corresponding to the articles for which personalized summaries have been generated are filtered, taking into account user preference for comments and the popularity of the comments themselves. When obtaining outstanding comments for the recommended articles based on semantic matching and comment popularity metrics, such as... Figure 9 and Figure 10 As shown, this can be achieved using, but is not limited to, the following methods, which include: 301. Obtain the user's interest tags and the comment text information of the comments corresponding to the recommended articles.
[0060] 302. Input the user's interest tags and the comment text information into the large model, calculate the semantic relevance between the interest tags and each comment, and obtain the first normalized score value of each comment.
[0061] This embodiment requires explanation of the construction of a prompt word project, allowing a large model to perform matching and scoring. The prompt project is as follows: "You are a comment scoring expert. First, I give you a batch of user interest tags as follows: ..., then I give you a list of comments, namely: [Comment 1, Comment 2, Comment 3]. By calculating the semantic relevance between the interest tags and each comment, assign a score between 0 and 1 to each comment," which is defined as match.
[0062] 303. Normalize the comment text information based on the comment popularity index to obtain the normalized popularity value of each comment.
[0063] When performing normalization processing on the comment text information based on comment popularity metrics to obtain a normalized popularity value for each comment, where the comment popularity metrics include at least the number of likes and replies, the following methods can be used, but are not limited to: The number of likes and replies to the comments are normalized to obtain normalized score values for likes and replies.
[0064] The comment popularity metric includes at least the number of likes and replies, normalized to between 0 and 1; let the number of likes be c and the number of replies be r, the normalization formula is as follows:
[0065] in, This indicates the highest number of likes among all comments on this article. This represents the minimum number of likes among all comments on this article. Similarly, the normalized number of replies is calculated as follows:
[0066] The score for likes and the score for replies are weighted to obtain a normalized popularity value for each comment.
[0067] Will The popularity of a comment is defined as follows: 304. The first score and popularity value are weighted and calculated to obtain the comprehensive comment value for each comment.
[0068] 305. Select a predetermined number of comments corresponding to the larger comprehensive review values as the featured comments of the recommended articles.
[0069] The top three comments, based on a comprehensive weighted score, are concatenated into the summary backend for subsequent voice generation.
[0070] Furthermore, based on the personalized summaries and selected insightful comments generated in the above steps, audio files are generated through text-to-speech and added to the user's app playlist. Specifically, when combining the personalized text summaries and insightful comments into speech playback data, generating audio files through text-to-speech technology, and adding the audio files to the dynamic audio playlist according to the user-configured playback parameters, the following methods can also be used, but are not limited to: Step 1: Perform preprocessing operations such as word segmentation, grammatical analysis, and semantic understanding on the personalized text summary and insightful comments to identify sentence structure, punctuation, and sentiment in the text.
[0071] The input summary undergoes text preprocessing, including word segmentation, syntactic analysis, and semantic understanding, to identify sentence structure, punctuation, and sentiment in the text.
[0072] Step 2: Based on the user-configured speech rate settings and speech playback preference parameters, select the corresponding speech synthesis model to synthesize the text into a speech signal.
[0073] The system uses deep learning models (such as Tacotron or WaveNet) to synthesize speech from text. These models are trained on a large amount of speech data and are able to adjust intonation, pauses, and emotional expression according to the text content.
[0074] Step 3: Process the speech signal through a vocoder to optimize the sound quality of the speech signal.
[0075] The generated speech signal is processed by a vocoder to further optimize the sound quality, making it more natural and fluent.
[0076] Furthermore, it also includes: managing the expiration time of audio files in the dynamic audio playlist based on the user-configured subscription expiration time parameter, and automatically deleting audio files that have exceeded the preset expiration time.
[0077] In this embodiment, users can configure audio subscription settings when subscribing. These settings can be customized according to user needs, and the embodiments disclosed herein do not impose limitations on them. For example... Figure 11 and Figure 12 As shown: The setup instructions can be found in Table 3.
[0078]
[0079] Table 3 This proposal suggests a method that combines a personalized recommendation system with a large language model to provide users with personalized text summaries of content they are interested in. The text summarization process comprehensively considers the recommendation reasons generated by the recommendation system, key text of the article, and text information from the article's illustrations. The text summaries are combined with insightful comments to generate audio files, which are then added to the user's audio playlist. This approach provides users with a completely new way of presenting article content.
[0080] Embodiments of this disclosure also provide an audio summary generation apparatus for subscription content, such as... Figure 13 ,include: The first acquisition unit 401 is used to calculate the weight of the content features of the articles that the user pays attention to through an attention mechanism, and to acquire a predetermined number of recommended articles with high user interest and the content features of the recommended articles. The first processing unit 402 is used to input the illustration text information, underlined text fragments, content features of the recommended article, and the main text of the recommended article into a large language model according to a preset prompt word template to generate a text summary corresponding to the recommended article. The second acquisition unit 403 acquires insightful comments on the recommended articles based on the semantic matching degree and comment popularity metrics for the recommended articles. The second processing unit 404 is used to combine the text summary and the highlights into voice playback data, generate an audio file through text-to-speech technology, and add the audio file to the dynamic audio playlist according to the playback parameters configured by the user.
[0081] Furthermore, the first acquisition unit 401 also includes: The first calculation module is used to calculate the weights of each content feature that the user pays attention to in each article created by the creator through an attention mechanism. The second calculation module is used to perform a weighted summation of the weights corresponding to each content feature of each article to obtain the comprehensive weight of each article; The evaluation module is used to input the comprehensive weights into the interest level prediction model to evaluate the user's interest level. The higher the evaluation score, the higher the interest level. The selection module is used to select a predetermined number of articles as recommended articles in descending order of interest, and to obtain a predetermined number of content features that have a significant impact on the evaluation score as content features of the recommended articles.
[0082] This first calculation module is specifically used for: The acquired user features and the content features of each article created by the creators followed by the users are vectorized using embedding matrices. The weights of each content feature corresponding to the user's attention to each article created by the creator are obtained by performing an inner product operation with each vectorized user feature and the vectorized content feature.
[0083] Furthermore, the second acquisition unit 403 also includes: The acquisition module is used to acquire the user's interest tags and the comment text information of the comments corresponding to the recommended articles; The first calculation module is used to input the user's interest tags and the comment text information into the large model, calculate the semantic relevance between the interest tags and each comment, and obtain the normalized first score value of each comment; The second calculation module is used to normalize the comment text information based on the comment popularity index to obtain the normalized popularity value of each comment. The third calculation module is used to perform a weighted calculation on the first score and the popularity value to obtain the comprehensive comment value for each comment; The selection module is used to select a predetermined number of comments corresponding to the larger comprehensive review values as the featured comments of the recommended articles.
[0084] This second calculation module is specifically used for: The number of likes and replies to the comments are normalized to obtain normalized like score and normalized reply score. The score for likes and the score for replies are weighted to obtain a normalized popularity value for each comment.
[0085] Furthermore, the second processing unit 404 also includes: The processing module is used to perform preprocessing operations such as word segmentation, grammatical analysis and semantic understanding on the personalized text summary and insightful comments, and to identify sentence structure, punctuation and sentiment in the text; The synthesis module is used to select the corresponding speech synthesis model to synthesize the text into a speech signal based on the user-configured speech rate settings and speech playback preference parameters. The speech signal is processed by a vocoder to optimize its sound quality.
[0086] Furthermore, the device also includes: The third processing unit is used to manage the timeliness of audio files in the dynamic audio playlist according to the subscription expiration time parameter configured by the user, and automatically delete audio files that have exceeded the preset expiration time.
[0087] It should be noted that the foregoing explanation of the method embodiments also applies to the apparatus of the embodiments of this disclosure, and the principle is the same. Therefore, the embodiments of this disclosure are not limited thereto.
[0088] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.
[0089] According to embodiments of this disclosure, an electronic device is provided, comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform the aforementioned... Figure 1-14 The method described.
[0090] According to embodiments of this disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause the computer to perform the aforementioned... Figure 1-14 The method described.
[0091] According to embodiments of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the aforementioned... Figure 1-14 The method described.
[0092] Figure 14 A schematic block diagram of an example electronic device 500 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.
[0093] like Figure 14As shown, the electronic device 500 includes a computing unit 501, which can perform various appropriate actions and processes based on a computer program stored in ROM (Read-Only Memory) 502 or a computer program loaded from storage unit 508 into RAM (Random Access Memory) 503. The RAM 503 can also store various programs and data required for the operation of the electronic device 500. The computing unit 501, ROM 502, and RAM 503 are interconnected via a bus 504. An I / O (Input / Output) interface 505 is also connected to the bus 504.
[0094] Multiple components in electronic device 500 are connected to I / O interface 505, including: input unit 506, such as keyboard, mouse, etc.; output unit 507, such as various types of monitors, speakers, etc.; storage unit 508, such as disk, optical disk, etc.; and communication unit 509, such as network card, modem, wireless transceiver, etc. Communication unit 509 allows electronic device 500 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.
[0095] The computing unit 501 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 501 include, but are not limited to, CPUs (Central Processing Units), GPUs (Graphics Processing Units), various special-purpose AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, DSPs (Digital Signal Processors), and any suitable processor, controller, microcontroller, etc. The computing unit 501 performs the various methods and processes described above, such as virtual resource scheduling optimization methods. For example, in some embodiments, the virtual resource scheduling optimization method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 508. In some embodiments, part or all of the computer program may be loaded and / or installed on the electronic device 500 via ROM 502 and / or communication unit 509. When the computer program is loaded into RAM 503 and executed by the computing unit 501, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, computing unit 501 may be configured to perform the aforementioned virtual resource scheduling optimization method by any other suitable means (e.g., by means of firmware).
[0096] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application-Specific Standard Products), SOCs (System-on-Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.
[0097] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.
[0098] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
[0099] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).
[0100] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include LANs (Local Area Networks), WANs (Wide Area Networks), the Internet, and blockchain networks.
[0101] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service system that addresses the shortcomings of traditional physical hosts and VPS (Virtual Private Server) services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.
[0102] It's important to note that artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.
[0103] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.
[0104] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.
Claims
1. A method for generating audio summaries of subscribed content, characterized in that, include: The weights of the features of the articles that users pay attention to are calculated using an attention mechanism to obtain a predetermined number of recommended articles with high user interest and the content features of the recommended articles. The text information of the illustrations in the recommended article, the text fragments of the underlined parts, the content features of the recommended article, and the main text are input into the large language model according to the preset prompt word template to generate the text summary corresponding to the recommended article; The most interesting comments on the recommended articles are obtained based on the semantic matching degree and comment popularity metrics. The text summary and insightful comments are combined into audio playback data, and an audio file is generated using text-to-speech technology. The audio file is then added to a dynamic audio playlist based on the playback parameters configured by the user.
2. The method as described in claim 1, characterized in that, The step of calculating the weights of the content features of articles that users pay attention to through an attention mechanism, and obtaining a predetermined number of recommended articles with high user interest and the content features of the recommended articles, includes: The attention mechanism is used to calculate the weights of each content feature that users pay attention to in each article created by the creator. The weights corresponding to each content feature of each article are weighted and summed to obtain the comprehensive weight of each article; The comprehensive weights are input into the interest level prediction model to evaluate the user's interest level; the higher the evaluation score, the higher the interest level. A predetermined number of articles are selected as recommended articles based on the degree of interest from high to low, and a predetermined number of content features that have a significant impact on the evaluation score are obtained as the content features of the recommended articles.
3. The method as described in claim 2, characterized in that, The process of calculating the weights corresponding to various content features of each article created by a creator through an attention mechanism includes: The acquired user features and the content features of each article created by the creators followed by the users are vectorized using embedding matrices. The weights of each content feature corresponding to the user's attention to each article created by the creator are obtained by performing an inner product operation with each vectorized user feature and the vectorized content feature.
4. The method as described in claim 1, characterized in that, The process of obtaining insightful comments on the recommended articles based on semantic matching degree and comment popularity metrics also includes: Obtain the user's interest tags, as well as the comment text information of the comments corresponding to the recommended articles; Input the user's interest tags and the comment text information into the large model, calculate the semantic relevance between the interest tags and each comment, and obtain the first normalized score of each comment; The comment text information is normalized based on the comment popularity index to obtain a normalized popularity value for each comment. The first score and popularity value are weighted and calculated to obtain the comprehensive comment value for each comment; A predetermined number of comments corresponding to the larger overall review values are selected as the featured comments of the recommended articles.
5. The method as described in claim 4, characterized in that, The normalization process of the comment text information based on the comment popularity index yields a normalized popularity value for each comment. The comment popularity index includes at least the number of likes and replies. The number of likes and replies to the comments are normalized to obtain normalized like score and normalized reply score. The score for likes and the score for replies are weighted to obtain a normalized popularity value for each comment.
6. The method as described in claim 5, characterized in that, The process of combining the personalized text summary with insightful comments into audio playback data, generating an audio file using text-to-speech technology, and adding the audio file to a dynamic audio playlist according to user-configured playback parameters further includes: The personalized text summaries and insightful comments undergo preprocessing operations such as word segmentation, grammatical analysis, and semantic understanding to identify sentence structure, punctuation, and sentiment in the text. Based on the user-configured speech rate settings and speech playback preference parameters, the corresponding speech synthesis model is selected to synthesize the text into a speech signal. The speech signal is processed by a vocoder to optimize its sound quality.
7. The method according to any one of claims 1-6, characterized in that, Also includes: Based on the user-configured subscription expiration time parameter, the system manages the expiration time of audio files in the dynamic audio playlist and automatically removes audio files that have exceeded the preset expiration time.
8. An audio summary generation device for subscription content, characterized in that, include: The first acquisition unit is used to calculate the weight of the content features of the articles that the user pays attention to through an attention mechanism, and to acquire a predetermined number of recommended articles with high user interest and the content features of the recommended articles. The first processing unit is used to input the text information of the illustrations in the recommended article, the text fragments of the underlined parts, the content features of the recommended article, and the main text of the article into a large language model according to a preset prompt word template to generate a text summary corresponding to the recommended article. The second acquisition unit acquires insightful comments about the recommended articles based on the semantic matching degree and comment popularity metrics of the recommended articles. The second processing unit is used to combine the text summary and the highlights into voice playback data, generate an audio file through text-to-speech technology, and add the audio file to the dynamic audio playlist according to the playback parameters configured by the user.
9. An electronic device, characterized in that, include: processor; The memory stores executable instructions; when the processor executes the instructions, it implements the method as described in any one of claims 1-7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as claimed in any one of claims 1-7.