Music resource recommendation method, program product, and electronic device

Abstract

The present disclosure provides a music resource recommendation method, program product and electronic device, and relates to the technical field of computers. The method comprises the following steps: obtaining associated corpus data of a target music resource, wherein the associated corpus data comprises text information related to the target music resource; inputting the associated corpus data into a large language model, performing semantic reasoning on the associated corpus data by the large language model, and generating an explanatory text corresponding to the target music resource, wherein the explanatory text is used to describe the content characteristics of the target music resource; encoding the explanatory text into a semantic vector; calculating the interest matching degree between the target music resource and a user based on the semantic vector and a historical behavior sequence of the user by using a recommendation model; determining target recommendation content recommended to the user according to the interest matching degree, and displaying the explanatory text corresponding to the target recommendation content as a recommendation reason. In this way, the cognitive load of the user can be reduced, and the recommendation effect of the music resource can be improved.

Description

Technical Field

[0001] Embodiments of this disclosure relate to the field of computer technology, and more specifically, embodiments of this disclosure relate to a method for recommending music resources, computer program products, and electronic devices. Background Technology

[0002] Currently, various digital platforms generally rely on intelligent recommendation mechanisms to distribute and push content, resources, and services to target audiences. By matching and filtering target content through algorithms and generating recommendation results, they improve the efficiency of content reach and the accuracy of distribution.

[0003] However, existing recommendation distribution schemes only output the final recommendation result in one direction, and the final recommendation result only provides basic display content of the recommended object itself, such as only displaying the entity name of the object or the literal meaning of the object. When users are faced with unfamiliar entity names, they are likely to be unable to form content expectations, which increases the user's cognitive load and leads to a lower willingness of users to click on the recommended object, thus reducing the recommendation effect. Summary of the Invention

[0004] In view of this, this disclosure provides a method for recommending music resources, a device for recommending music resources, a computer program product, and an electronic device, so as to reduce the cognitive load of users to a certain extent and improve click intention and recommendation effect.

[0005] According to a first aspect of this disclosure, a method for recommending music resources is provided, the method comprising: Obtain associated corpus data of the target music resource, wherein the associated corpus data includes text information related to the target music resource; The associated corpus data is input into a large language model, and semantic reasoning is performed on the associated corpus data through the large language model to generate explanatory text corresponding to the target music resource. The explanatory text is used to describe the content features of the target music resource. The explanatory text is encoded into a semantic vector; Based on the semantic vector and the user's historical behavior sequence, the matching degree between the target music resource and the user's interests is calculated through a recommendation model; Based on the interest matching degree, target recommended content is determined to be recommended to the user, and the explanatory text corresponding to the target recommended content is displayed as the recommendation reason.

[0006] In one possible implementation, the target music resource includes physical resources and / or aggregate resources; the physical resources include songs and / or artists; and the aggregate resources include playlists and / or podcasts.

[0007] In one possible implementation, the step of performing semantic reasoning on the associated corpus data using the large language model to generate explanatory text corresponding to the target music resource includes: If the target music resource is an entity resource, at least one of the lyrics, user comments, and music reviews in the target music resource is extracted from the associated corpus data. The extracted content is then subjected to semantic analysis and disambiguation processing based on the large language model using thought chain reasoning to generate explanatory text corresponding to the entity resource. If the target music resource is a collection-type resource, the title and / or description of the collection-type resource are subjected to noise filtering and semantic summarization processing by the large language model to generate explanatory text corresponding to the collection-type resource.

[0008] In one possible implementation, the semantic analysis and disambiguation processing of the extracted content based on the thought chain reasoning of the large language model includes: Based on the lyrics text, high-frequency words and imagery features in the lyrics are identified; Determine the emotional tone based on the user comments and / or music reviews; Based on the imagery features and the emotional tone, the content features of the target music resource are determined, and the content features include at least one of the following: musical style, emotion, scene, and genre.

[0009] In one possible implementation, the noise filtering and semantic summarization processing of the titles and / or descriptions of the aggregated resources using the large language model includes: Filter out invalid words of a preset type from the title and / or description of the collection-type resources; Extract at least one of the following information from the filtered title and / or description: core style information, scene information, and emotional information; The extracted information is compressed into standardized phrases, which serve as explanatory text for the aggregated resources.

[0010] In one possible implementation, before inputting the associated corpus data into the large language model, the method further includes: Based on the retrieval enhancement generation mechanism, related information concerning the target music resource is retrieved from the database and used as context input for the large language model. The related information includes at least one of the following: lyrics, user comments, and artist background information. Structured instructions are input to the large language model based on a preset prompt template, wherein the prompt template includes at least one of role settings, task constraints, and example information.

[0011] In one possible implementation, encoding the explanatory text into a semantic vector includes: Explanatory texts generated for different types of target music resources are encoded into semantic vectors in the same semantic vector space using the same vector encoder.

[0012] In one possible implementation, the step of calculating the match between the target music resource and the user's interests using a recommendation model based on the semantic vector and the user's historical behavior sequence includes: Obtain the user's historical behavior sequence, where each interaction item in the historical behavior sequence includes a resource identifier vector and a corresponding semantic vector; In the feature input layer of the recommendation model, the semantic vector of the target music resource is concatenated with the semantic vector of each interaction item in the historical behavior sequence, and the resource identifier vector of the target music resource is concatenated with the resource identifier vector of each interaction item in the historical behavior sequence to obtain a fused feature representation. In the attention layer of the recommendation model, based on the fusion feature representation, the attention weight between each interaction item in the user's historical behavior sequence and the target music resource is calculated; The user's interest matching degree for the target music resource is obtained by weighting the attention weights.

[0013] According to a second aspect of this disclosure, a music resource recommendation device is provided, the device comprising: The acquisition unit is used to acquire the associated corpus data of the target music resource, wherein the associated corpus data includes text information related to the target music resource; The generation unit is used to input the associated corpus data into a large language model, perform semantic reasoning on the associated corpus data through the large language model, and generate explanatory text corresponding to the target music resource. The explanatory text is used to describe the content features of the target music resource. An encoding unit is used to encode the explanatory text into a semantic vector; The processing unit is used to calculate the matching degree between the target music resource and the user's interests based on the semantic vector and the user's historical behavior sequence through a recommendation model; The recommendation unit is used to determine the target recommended content to be recommended to the user based on the interest matching degree, and to display the explanatory text corresponding to the target recommended content as the recommendation reason.

[0014] In one possible implementation, the target music resource includes physical resources and / or aggregate resources; the physical resources include songs and / or artists; and the aggregate resources include playlists and / or podcasts.

[0015] In one possible implementation, the generating unit is configured to: If the target music resource is an entity resource, at least one of the lyrics, user comments, and music reviews in the target music resource is extracted from the associated corpus data. The extracted content is then subjected to semantic analysis and disambiguation processing based on the large language model using thought chain reasoning to generate explanatory text corresponding to the entity resource. If the target music resource is a collection-type resource, the title and / or description of the collection-type resource are subjected to noise filtering and semantic summarization processing by the large language model to generate explanatory text corresponding to the collection-type resource.

[0016] In one possible implementation, the generating unit is configured to: Based on the lyrics text, high-frequency words and imagery features in the lyrics are identified; Determine the emotional tone based on the user comments and / or music reviews; Based on the imagery features and the emotional tone, the content features of the target music resource are determined, and the content features include at least one of the following: musical style, emotion, scene, and genre.

[0017] In one possible implementation, the generating unit is configured to: Filter out invalid words of a preset type from the title and / or description of the collection-type resources; Extract at least one of the following information from the filtered title and / or description: core style information, scene information, and emotional information; The extracted information is compressed into standardized phrases, which serve as explanatory text for the aggregated resources.

[0018] In one possible implementation, the device includes a supplementary unit for: Based on the retrieval enhancement generation mechanism, related information concerning the target music resource is retrieved from the database and used as context input for the large language model. The related information includes at least one of the following: lyrics, user comments, and artist background information. Structured instructions are input to the large language model based on a preset prompt template, wherein the prompt template includes at least one of role settings, task constraints, and example information.

[0019] In one possible implementation, the encoding unit is used for: Explanatory texts generated for different types of target music resources are encoded into semantic vectors in the same semantic vector space using the same vector encoder.

[0020] In one possible implementation, the processing unit is configured to: Obtain the user's historical behavior sequence, where each interaction item in the historical behavior sequence includes a resource identifier vector and a corresponding semantic vector; In the feature input layer of the recommendation model, the semantic vector of the target music resource is concatenated with the semantic vector of each interaction item in the historical behavior sequence, and the resource identifier vector of the target music resource is concatenated with the resource identifier vector of each interaction item in the historical behavior sequence to obtain a fused feature representation. In the attention layer of the recommendation model, based on the fusion feature representation, the attention weight between each interaction item in the user's historical behavior sequence and the target music resource is calculated; The user's interest matching degree for the target music resource is obtained by weighting the attention weights.

[0021] According to a third aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the method of the first aspect described above and possible implementations thereof.

[0022] According to a fourth aspect of this disclosure, an electronic device is provided, comprising: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to perform the method of the first aspect and possible implementations thereof by executing the executable instructions.

[0023] The technical solution disclosed herein has the following beneficial effects: In this embodiment, associated corpus data of the target music resource can be obtained, wherein the associated corpus data includes text information related to the target music resource; then, the associated corpus data is input into a large language model, and semantic reasoning is performed on the associated corpus data through the large language model to generate explanatory text corresponding to the target music resource, the explanatory text being used to describe the content features of the target music resource; and the explanatory text is encoded into a semantic vector. Further, based on the semantic vector and the user's historical behavior sequence, a recommendation model can calculate the interest matching degree between the target music resource and the user; thus, target recommended content can be determined to recommend to the user based on the interest matching degree, and the explanatory text corresponding to the target recommended content can be displayed as the recommendation reason.

[0024] In this embodiment of the disclosure, when recommending target content to a user, explanatory text corresponding to the target content is also displayed to the user, which can reduce the user's decision-making cost to a certain extent, thereby improving the user's click-through rate and user experience of the target content.

[0025] Other features and advantages of this disclosure will be set forth in the following description and will be apparent in part from the description or may be learned by practicing the disclosure. The objects and other advantages of this disclosure may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments of this disclosure will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A schematic diagram of an application scenario in this exemplary embodiment is shown.

[0028] Figure 2 A system diagram of a recommendation system according to this exemplary embodiment is shown.

[0029] Figure 3 This diagram illustrates a flowchart of a music resource recommendation method according to this exemplary embodiment.

[0030] Figure 4 This diagram illustrates one method of obtaining a semantic vector in this exemplary embodiment.

[0031] Figure 5 This illustration shows a flowchart of a process for determining the match between a target music resource and a user's interests in this exemplary embodiment.

[0032] Figure 6 This diagram illustrates one method of obtaining explanatory text in this exemplary embodiment.

[0033] Figure 7 This illustration shows a process diagram of a music resource recommendation method in this exemplary embodiment.

[0034] Figure 8 A schematic diagram of a music resource recommendation device is shown in this exemplary embodiment.

[0035] Figure 9 A schematic diagram of the structure of an electronic device in this exemplary embodiment is shown. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this disclosure clearer, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not all of them. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure. Unless otherwise specified, the embodiments and features in the embodiments of this disclosure can be arbitrarily combined with each other. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0037] The term "comprising" and any variations thereof in the specification and claims of this disclosure are intended to cover non-exclusive protection. For example, a process, method, system, product, or apparatus that comprises a series of steps or units is not limited to the steps or units listed, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus.

[0038] In this disclosure, there are one or more embodiments; "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be single or multiple.

[0039] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order, sequence, size, or priority. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0040] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, which are schematic illustrations of this disclosure and are not necessarily drawn to scale. Some block diagrams shown in the drawings may be functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in hardware modules or integrated circuits, or in networks, processors, or microcontrollers. Implementations can be carried out in various forms and should not be construed as limited to the examples set forth herein. The features, structures, or characteristics described in this disclosure can be combined in any suitable manner in one or more implementations. Numerous specific details are provided in the following description to give a thorough description of the embodiments of this disclosure. However, those skilled in the art will recognize that one or more specific details may be omitted when implementing the technical solutions of this disclosure, or other methods, components, apparatuses, steps, etc., may be used to replace one or more specific details. It should be noted that in the embodiments of this disclosure, certain software, components, models, and other existing solutions in the industry may be mentioned. These should be considered exemplary and their purpose is merely to illustrate the feasibility of implementing the technical solutions of this disclosure, but does not imply that the applicant has already used or necessarily used such solutions. In the embodiments disclosed herein, the collection, dissemination, use, and display of data all comply with relevant national laws and regulations. Invention Overview

[0042] Currently, the mainstream technical solutions for recommendation and distribution systems in related technologies can be mainly divided into the following two categories: Option 1: ID sequence recommendation technology based on collaborative filtering.

[0043] This is currently the most mainstream approach. Its core logic is to treat songs, artists, playlists, etc., as independent item IDs, and then construct a "user-item" interaction matrix or sequence to uncover implicit associations where "people who like item A also like item B." However, this approach is highly dependent on the scale of historical user behavior data.

[0044] Option 2: Text tag retrieval and matching technology based on metadata.

[0045] This approach relies on the original textual information of music resources (such as song titles, album names, and manually tagged labels). The system calculates text similarity through keyword matching or shallow semantic vectors (such as Word2Vec and BERT Embedding). For example, when a user searches for "sleep," the system retrieves songs whose titles contain "sleep" or whose tags contain "sleep aid."

[0046] While the above solutions meet the distribution and recommendation needs to some extent, they still have significant technical shortcomings when facing users' high demands for "content explainability" and "accurate semantic understanding." The deficiencies of existing technical solutions are as follows: (1) The recommended keywords are poorly explained, giving users a “black box” experience and resulting in a low click-through rate.

[0047] In related technologies, distribution and recommendation systems often only present entity names such as "similar songs" or "similar artists" in their recommendation reasons. However, many song titles themselves have strong artistic expressions and lack direct information gain. Relying solely on textual information, users find it difficult to judge the style and mood of the music before clicking, resulting in high decision-making costs and low click-through rates. At the same time, playlist and podcast titles often suffer from keyword stuffing (such as "featured / audiophile / Hi-Fi / must-listen / popular / featured") or excessive truncation, making it difficult to accurately convey the content style like natural language, further increasing cognitive load.

[0048] (2) The distribution and recommendation system lacks deep semantic reasoning ability, resulting in a serious “semantic gap”.

[0049] Specifically, in the processing of literal and figurative meanings, distribution and recommendation systems in related technologies often establish associations solely based on the literal meaning. For example, the song "Goodnight" is a typical Cantonese pop love song with a theme of emotional regret, but shallow semantic models are prone to misjudging it as a "sleep aid / bedtime" scenario because of its title, ultimately pushing the sad song to users who need to calm down and help them sleep, causing a significant distribution error.

[0050] Furthermore, songs and playlists have inherent differences in metadata structure. Existing technologies lack the ability to evaluate the similarity between "an R&B song" and "a city night tour playlist" in the same semantic space, which limits the efficiency of distribution and recommendation systems when dealing with multi-source heterogeneous resources.

[0051] (3) The system has insufficient support for long-tail and cold-start resources.

[0052] Distribution and recommendation systems in related technologies heavily rely on historical behavioral data. Therefore, these systems struggle to effectively distribute newly released or niche songs due to a lack of interaction records. Furthermore, current content analysis technologies cannot extract sufficiently precise "selling points" from brief lyrics or descriptions, resulting in many potentially valuable long-tail content pieces being overlooked due to a lack of semantic characterization.

[0053] In summary, existing technical solutions are insufficient in terms of text interpretability and deep semantic understanding, which increases the cognitive load on users, reduces their decision-making efficiency, and diminishes their user experience.

[0054] In view of this, embodiments of this disclosure provide a method for recommending music resources. This method involves acquiring associated corpus data of a target music resource, wherein the associated corpus data includes text information related to the target music resource; then, the associated corpus data is input into a large language model, and semantic reasoning is performed on the associated corpus data through the large language model to generate explanatory text corresponding to the target music resource. The explanatory text describes the content features of the target music resource; and the explanatory text is encoded into a semantic vector. Further, based on the semantic vector and the user's historical behavior sequence, a recommendation model can calculate the interest matching degree between the target music resource and the user. Thus, target recommended content can be determined to recommend to the user based on the interest matching degree, and the explanatory text corresponding to the target recommended content can be displayed as a recommendation reason.

[0055] In this embodiment of the disclosure, when recommending target content to a user, explanatory text corresponding to the target content is also displayed to the user, which can reduce the user's decision-making cost to a certain extent, thereby improving the user's click-through rate and user experience of the target content.

[0056] After introducing the basic principles of this disclosure, various non-limiting embodiments of this disclosure will be described in detail below.

[0057] Application Scenarios Overview

[0058] To better understand the technical solutions provided in the embodiments of this disclosure, the following is a brief introduction to the application scenarios applicable to the technical solutions provided in the embodiments of this disclosure. It should be noted that the application scenarios described below are only for illustrating the embodiments of this disclosure and are not intended to limit the scope. In specific implementation, the technical solutions provided in the embodiments of this disclosure can be flexibly applied according to actual needs.

[0059] To provide a clearer description of the embodiments of this disclosure, some key terms are introduced below: Deep learning models: As the core of semantic understanding and text generation, this model is based on the Transformer architecture and possesses powerful natural language processing and logical reasoning capabilities, used to perform semantic reconstruction tasks of music entities. For example, the deep learning model is DeepSeek (a new generation large language model developed by Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.).

[0060] Chain of Thought (CoT) refers to a prompting engineering technique used when generating explanatory text using deep learning models. By guiding the model to explicitly generate intermediate reasoning steps (e.g., first analyze the imagery of the lyrics > then determine the emotional tone -> finally summarize the musical style), it ensures that the generated music tags are accurate and avoids the model producing non-factual descriptions.

[0061] Retrieval-Augmented Generation (RAG): Before the deep learning model generates explanatory text, the system first retrieves external knowledge such as the song's lyrics, user comments, and artist background from the database as context input to the model, ensuring that the generated explanatory text is based on facts rather than fabrication.

[0062] Prompt engineering refers to the technique of designing specific input templates for deep learning models. In this application, it specifically refers to structured instructions that contain explanatory text for guiding the model to output text that conforms to the recommender system specifications.

[0063] Explainable Terms: These are short textual features, such as words, sentences, and phrases, generated through deep learning models that accurately describe the essence of musical content (e.g., style, mood, scene, genre). They are the core semantic carrier used in this application to replace raw metadata (e.g., meaningless song titles).

[0064] Explainable Text Integrated Recommendation System (ETIRS): This refers to the full name of the music resource recommendation system constructed in this application. Its core feature is the deep integration of unstructured natural language explanation information (generated by a large language model (LLM)) with structured ID features, thereby solving the problems of "black box" and "semantic missing" in traditional recommendation systems.

[0065] Deep Interest Network with Text (DIN): This disclosure is an improved version of the classic Deep Interest Network (DIN). Its innovation lies in the introduction of a text-based attention layer, which forces the model to focus on semantic features (i.e., semantic vectors) in the user's historical behavior sequence, rather than just focusing on the item ID.

[0066] In this embodiment of the disclosure, the music resource recommendation technology can be applied to various business scenarios that recommend music resources, such as business scenarios that recommend songs. This embodiment of the disclosure does not limit this application.

[0067] Please see Figure 1 As shown, Figure 1This is an application scenario to which the technical solution of the present disclosure embodiment can be applied. The schematic diagram of this scenario includes a terminal device 110 and a service device 120. The terminal device 110 can be one or more. Figure 1 One example is shown below. Terminal device 110 and service device 120 are connected via network 130.

[0068] In this embodiment of the disclosure, the terminal device 110 may be deployed with a platform (such as an application, website, etc.) that provides recommendation services, which can be logged in and used by different users with corresponding accounts and passwords.

[0069] In this embodiment, the service device 120 can acquire associated corpus data of the target music resource, including text information related to the target music resource; input the associated corpus data into a large language model, perform semantic reasoning on the associated corpus data through the large language model, and generate explanatory text corresponding to the target music resource, which describes the content features of the target music resource; encode the explanatory text into a semantic vector. Furthermore, when a user logs into the recommendation system through the terminal device 110, the terminal device 110 can determine the user's relevant information and send the user's relevant information to the service device 120. The service device 120 can then determine the user's historical behavior sequence based on the user's relevant information, and then calculate the interest matching degree between the target music resource and the user based on the semantic vector and the user's historical behavior sequence through a recommendation model; then, based on the interest matching degree, determine the target recommended content to recommend to the user, and display the explanatory text corresponding to the target recommended content as a recommendation reason.

[0070] In this embodiment of the disclosure, Figure 1 The service device 120 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server or cloud server cluster that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms, but is not limited to these.

[0071] In this embodiment of the disclosure, the service device 120 is equipped with an explanatory text fusion recommendation system, which can provide recommendation services to the platform in the terminal device.

[0072] For example, see Figure 2 As shown, the logical architecture of this explanatory text fusion recommendation system is divided into five layers from top to bottom: data source layer, semantic reconstruction layer, feature alignment layer, recommendation model layer, and application interaction layer.

[0073] In this exemplary embodiment, the data source layer is used to collect data, such as when recommending songs to users. This data may include raw audio, lyrics text, user behavior logs, and metadata of the playlist.

[0074] In this exemplary embodiment, the semantic reconstruction layer includes: a large language model based on a Mixture of Experts (MoE) architecture, which, for example, contains eight independent "expert" modules (i.e., Figure 2 The system employs Expert-1 to Expert-8 experts, and a Router / Gating Network dynamically selects the Top-K experts based on the input. The semantic reconstruction layer is responsible for deep semantic understanding of multi-source heterogeneous information such as original audio, lyrics, user behavior logs, and song unit data, generating semantic vectors (Embeddings) rich in semantic information. It solves the "semantic gap" problem in traditional recommendation systems, capturing deep features such as song emotion and style, providing interpretable basis for subsequent recommendations (e.g., "melancholy-style Cantonese pop songs"). In other words, the semantic reconstruction layer completes deep semantic understanding, transforming unstructured music information into computable vectors.

[0075] In this exemplary embodiment, the feature alignment layer includes a vector database, which stores semantic vectors (embeddings) of all music content generated by the large language model, serving as a vectorized knowledge base. The feature alignment layer enables efficient vector retrieval and matching, quickly finding music content semantically similar to the user's historical behavior vectors, thus providing an accurate candidate set for the upper-layer recommendation model.

[0076] In this exemplary embodiment, the recommendation model layer includes an Adaptive Attention Mechanism, a Rectified Linear Unit (ReLU) hidden layer 1, a Gaussian Error Linear Unit (GELU) hidden layer 2, and an output layer.

[0077] The attention unit calculates the attention weights of each interaction item in the user's historical behavior sequence and the target music resource, and performs a weighted summation of the historical behavior vectors. Hidden layer 1 performs a non-linear transformation on the vector after weighting by the attention unit, extracting more complex feature combinations, increasing the model's expressive power, and suppressing negative values ​​through the characteristics of ReLU to alleviate the gradient vanishing problem. Hidden layer 2 further performs a smooth non-linear transformation on the output of hidden layer 1. GELU is more robust than ReLU and can adaptively introduce random regularization effects to improve the model's generalization ability and prediction stability, providing more accurate feature support for the final click-through rate (CTR) prediction. Thus, the recommendation model layer dynamically captures the relevance between the user's historical behavior and the target music resource, allowing the model to focus more on the historical interests most relevant to the current recommendation, thereby improving the accuracy and personalization of the recommendation. The output layer outputs the CTR / preference score, combined with explanatory text generated by the large language model (such as "melancholy-style Cantonese pop songs"), and finally pushes it to the user.

[0078] In this exemplary embodiment, the application interaction layer is used to display a sorted list of recommendations to the user, along with explanatory text in natural language for each item in the list, thereby improving user experience and click-through rate.

[0079] In this embodiment of the disclosure, Figure 1 The terminal device 110 can be a mobile phone, tablet computer (PAD), personal computer, smart TV, smart watch, smart speaker, smart vehicle device, and wearable device, but is not limited to these.

[0080] Of course, the methods provided in this disclosure are not limited to... Figure 1 The application scenarios shown can also be used in other possible application scenarios, such as a music resource recommendation method executed only by the terminal device. This disclosure does not impose any limitations.

[0081] Exemplary methods

[0082] To further illustrate the technical solutions provided by the embodiments of this disclosure, a detailed description is provided below in conjunction with the accompanying drawings and specific implementation methods. Although the embodiments of this disclosure provide method operation steps as shown in the following embodiments or drawings, the method may include more or fewer operation steps based on conventional or non-inventive methods. In steps where there is no logically necessary causal relationship, the execution order of these steps is not limited to the execution order provided by the embodiments of this disclosure. In actual processing or when the device executes the method, it may be executed sequentially or in parallel according to the method shown in the embodiments or drawings.

[0083] Please see Figure 3 , Figure 3 This is a flowchart illustrating a method for recommending music resources according to an embodiment of this disclosure. The method flow can be, for example, carried out by an electronic device, such as... Figure 1 The service device 120 executes the method, and the specific implementation process is as follows: Step 301: Obtain the associated corpus data of the target music resource. The associated corpus data includes text information related to the target music resource.

[0084] In this embodiment of the disclosure, a target music resource can be determined first, wherein the target music resource refers to a set of music content to be recommended to the user.

[0085] In an exemplary embodiment, the target music resource includes entity-type resources and / or aggregate-type resources. Optionally, entity-type resources include songs and / or artists. Entity-type resources are characterized by fixed content, rich semantics, and ambiguity. Optionally, aggregate-type resources include playlists and / or podcasts. Aggregate-type resources are characterized by user-defined titles, which typically contain a large amount of marketing noise or meaningless characters.

[0086] In an exemplary embodiment, for entity-type resources, associated corpus data of the resource can be retrieved in parallel through an interface. This associated corpus data may include text-dimensional data and public opinion-dimensional data. Text-dimensional data may include, for example, full lyrics text, song descriptions, and artist background information, while public opinion-dimensional data may include, for example, highly rated user comments and professional music reviews.

[0087] In this exemplary embodiment, for aggregated resources, their associated corpus data can be obtained, including comprehensive corpus materials such as aggregate theme positioning, content introduction, column introduction, category tags, scene attributes, target audience, copywriting description, program introduction, content summary, aggregation features, scene keywords, column tone and overall content annotation, which fully cover the overall positioning and content characteristics of aggregated resources.

[0088] Step 302: Input the associated corpus data into the large language model, perform semantic reasoning on the associated corpus data through the large language model, and generate explanatory text corresponding to the target music resource. The explanatory text is used to describe the content features of the target music resource.

[0089] In this embodiment of the disclosure, after obtaining the associated corpus data, a retrieval enhancement generation mechanism can be used to retrieve related information from the database concerning the target music resource, which serves as contextual input for the large language model. The associated information includes at least one of the following: lyrics, user comments, and artist background information. The database refers to a real, callable external associated resource library that aggregates and stores various types of original, underlying music-related data.

[0090] In this embodiment, structured instructions can also be input into the large language model based on a preset prompt template. The prompt template includes at least one of the following: role settings, task constraints, and example information. Role settings clarify the model's identity, professional field, capabilities, and output tone, defining the overall stance, style, and business suitability of the response, allowing the model to operate in a way that aligns with the music recommendation business scenario. Task constraints clarify the core objectives, operational procedures, content requirements, boundary limitations, and format standards for this execution, limiting the data processing scope, content writing rules, prohibited and required fields, ensuring that the output content is compliant, consistent, and meets business needs. Example information provides standardized reference cases and demonstration examples, intuitively showcasing the input format, output structure, content granularity, expression logic, and writing standards, providing a reference standard for the model, reducing misunderstanding bias, and ensuring the consistency and standardization of the generated content.

[0091] In this embodiment of the disclosure, after obtaining the associated corpus data, determining the structured instructions based on the preset prompt template, and the context input, all the aforementioned content can be input into the large language model to generate explanatory text corresponding to the target music resource.

[0092] In one possible implementation, if the target music resource is an entity resource, at least one of the lyrics, user comments, and music reviews in the target music resource is extracted from the associated corpus data. The extracted content is then subjected to semantic analysis and disambiguation processing based on thought chain reasoning using a large language model to generate explanatory text corresponding to the entity resource.

[0093] In this exemplary embodiment, the large language model can identify high-frequency words and imagery features in the lyrics based on the lyrics text. It can also determine the emotional tone based on user comments and / or music reviews. Then, based on the high-frequency words, imagery features, and emotional tone, the content features of the target music resource are determined. These content features include at least one of the following: musical style, mood, scene, and genre. Specifically, musical style includes pop, folk, rock, traditional Chinese style, rap, and instrumental music; mood is categorized as healing, sadness, passion, soothing, cheerful, lonely, and languid; scene is adapted to diverse usage environments such as commuting, exercising, before bed, studying, partying, stress relief, and driving; and genre is further subdivided into different music systems such as classical, electronic, country, jazz, traditional Chinese style, and Western pop.

[0094] In this exemplary embodiment, a retrieval enhancement generation mechanism can also be used to retrieve relevant information related to the target music resource from the database, which can then be used as contextual input for the large language model. This relevant information includes at least one of the following: lyrics, user comments, and artist background information. This relevant information refers to real-world information related to the target music resource. Furthermore, structured instructions can be input into the large language model based on a preset prompt template, which includes at least one of the following: role settings, task constraints, and example information.

[0095] In this way, the large language model can generate explanatory text corresponding to entity-type resources based on structured instructions, combined with the aforementioned content features and related information.

[0096] For example, the process of obtaining interpretive text is introduced by performing deep semantic reasoning and disambiguation on the song "Goodnight".

[0097] First, the large language model identifies the resource title as "Goodnight," initially associating it with the scenarios of "sleep" and "sleep aid." Then, content validation and contradiction detection are performed: specifically, the large language model analyzes the lyrics text, identifying high-frequency words such as "regret," "insomnia," "missing," and "due to." Finally, sentiment analysis is performed: the large language model combines the emotional tendencies in user comments (such as "made me cry," "reminded my ex") to determine the actual emotional tone of the resource as "sadness" and "nostalgia." In this way, the large language model can determine the content characteristics.

[0098] Furthermore, the large language model performs logical correction processing, specifically: based on structured instructions, the large language model integrates content features and related information to determine that the song belongs to the "emotional catharsis" category, thereby outputting the final explanatory text, such as: "depressing voice | late at night alone | lyrical in Mandarin".

[0099] As can be seen, in this embodiment of the disclosure, semantic reconstruction of entity-type resources can be performed based on a large language model to obtain explanatory text corresponding to the entity-type resources.

[0100] In one possible implementation, if the target music resource is a collection-type resource, noise filtering and semantic summarization processing are performed on the title and / or description of the collection-type resource using a large language model to generate explanatory text corresponding to the collection-type resource.

[0101] In this exemplary embodiment, the large language model can filter out invalid words of a preset type from the title and / or description of a collection-type resource; then, extract at least one of the following information from the filtered title and / or description: core style information, scene information, and emotion information; and compress the extracted information into standardized phrases as explanatory text corresponding to the collection-type resource.

[0102] In this exemplary embodiment, the entity recognition capability of a large language model is utilized to scan the playlist titles and / or descriptions of aggregated resources. Then, relying on a pre-set filter lexicon, the large language model automatically identifies and marks invalid words in the playlist titles and / or descriptions that lack substantive semantics. These invalid words include, but are not limited to, exaggerated modifiers such as "hottest," "must-listen," "BGM," "TikTok hits," "shocking," and "ultimate." These words are assigned zero weight or directly removed during text reconstruction. Furthermore, after noise removal, the large language model extracts features from at least one of three core dimensions from the remaining text information, i.e., the filtered titles and / or descriptions (such as the description body and features of the included song list). These three core dimensions are, for example: Genre dimension: such as pop, rock, electronic; Scene dimension: such as commuting, studying, party; and Mood dimension: such as sad, healing, excited.

[0103] In this exemplary embodiment, a retrieval enhancement generation mechanism can also be used to retrieve relevant information related to the target music resource from the database, which can then be used as contextual input for the large language model. This relevant information includes at least one of the following: lyrics, user comments, and artist background information. This relevant information refers to real-world information related to the target music resource. Furthermore, structured instructions can be input into the large language model based on a preset prompt template, which includes at least one of the following: role settings, task constraints, and example information.

[0104] In this way, the large language model can process the features and related information after fusing at least one core dimension feature extracted according to structured instructions, and rewrite the originally long and messy title into a standardized phrase, thereby obtaining explanatory text of the collection resource.

[0105] For example, the original input is titled "[Must Listen] The Hottest Sad BGM of 2024, Every Song is a Nostalgic Trip, You'll Cry When You Hear It!!". The large language model then processes it as follows: filtering out noise such as "must listen," "hottest," "BGM," and "cry when you hear it"; extracting the core dimension features "sadness" (Mood) and "nostalgia" (Scene / Mood). Furthermore, the large language model can, based on structured instructions, integrate core dimension features and related information to output the final explanatory text, such as "Sad Healing Pop Music" or "A Selection of Nostalgic Sentiments."

[0106] Step 303: Encode the explanatory text into a semantic vector.

[0107] In this embodiment of the disclosure, in response to the problem that the metadata structures of multi-source heterogeneous resources such as singles, playlists, and podcasts are not uniform and cannot be directly measured for similarity, this disclosure provides a semantic alignment and high-dimensional vector mapping mechanism.

[0108] In this embodiment of the disclosure, explanatory texts generated for different types of target music resources can be encoded into semantic vectors in the same semantic vector space by the same vector encoder.

[0109] In an exemplary embodiment, the semantic vector is a fixed-dimensional semantic vector that is converted from explanatory text using a pre-trained text encoder (such as a Text Encoder, like Bert or Sentence-Transformer).

[0110] For example, see Figure 4 As shown, for discrete, unstructured data, interpretable text with semantic alignment can be obtained based on resource interpretability reconstruction techniques using large language models. Here, a large language model is, for example,... Figure 2 The large language model shown is then used. Next, the semantically aligned interpretable text is transformed into computer-processable semantic vectors using a text encoder, further improving the reliability of distance metrics in the semantic space.

[0111] Step 304: Based on semantic vectors and the user's historical behavior sequence, calculate the matching degree between the target music resource and the user's interests using a recommendation model.

[0112] In this embodiment of the disclosure, after obtaining the semantic vector of the music resource, the matching degree between the target music resource and the user's interests can be calculated by a recommendation model based on the semantic vector and the user's historical behavior sequence.

[0113] In an exemplary embodiment, see Figure 5 As shown, the following steps can be used, but are not limited to, to determine the match between the target music resource and the user's interests.

[0114] Step 501: Obtain the user's historical behavior sequence. Each interaction item in the historical behavior sequence includes resource identification features and corresponding semantic vectors.

[0115] In this embodiment, the user's historical behavior sequence can also be obtained. For example, the user is a target user who logs into a recommendation service platform. Then, each interaction item in the user's historical behavior sequence can be determined to establish the user's historical behavior sequence. An interaction item refers to historical behavioral information and content entries where the user has actively engaged in, touched upon, or interacted with the platform. It is core behavioral data reflecting the user's true listening preferences, habits, and content inclinations. For example, various music resources and behavioral records corresponding to a series of operations such as playing, collecting, liking, sharing, downloading, looping, skipping, pausing, commenting, following artists, subscribing to playlists, listening to podcasts, browsing music details, clicking on recommended content, and adding to a favorites list. All entity-type and aggregate-type resources that the user has actually interacted with fall under the category of interaction items. Furthermore, the four-tuple features of each interaction item can be determined; these features include resource identifier features, type features, context features, and semantic vectors.

[0116] In this embodiment, the resource identifier feature is a unique identifier for the resource, used to vectorize the resource through the embedding layer. Based on the vector results output by this layer, collaborative filtering vectors with collaborative similarity relationships are retrieved and filtered. This allows for the discovery of users' historical behavior sequences and target music resources, i.e., implicit associations between resources, providing vector support for subsequent personalized matching and content recommendation. Type features refer to features corresponding to the type of music resource, such as songs, artists, playlists, etc. Context features refer to traditional attribute features, such as time and scene.

[0117] In this embodiment of the disclosure, semantic vectors can be obtained based on a vector database acquired during the offline data processing phase. This vector database is, for example, […]. Figure 2 The vector database shown is an example. Alternatively, the corresponding semantic vectors can be obtained by processing the resource data to be processed online, but this embodiment does not limit this approach.

[0118] In one possible implementation, for each interaction item, the following operations may be performed: determining a vector database; the vector database includes semantic vectors obtained by vector transformation of explanatory text; and determining the semantic vector corresponding to the interaction item from the vector database.

[0119] In this exemplary embodiment, the vector database can continuously scan newly added and changed resources in the music library (local and / or external) via scheduled tasks (such as T+1) or real-time streaming interfaces, and automatically trigger a resource reconstruction service deployed in the cloud. This reconstruction service performs batch semantic disambiguation and reconstruction on the newly added and changed resources, i.e., the original metadata, to generate standardized explanatory text. Then, it calls the same-origin embedding encoder to convert the explanatory text into semantic vectors, which are stored along with the resource identifier (ID). The vector database supports Hierarchical Navigable Small World (HNSW) indexing, which facilitates the rapid acquisition of semantic vectors of interactive items and target music resources during subsequent object recommendation, improving recommendation efficiency.

[0120] In one possible implementation, determining the semantic vector of each interaction item includes: for each interaction item, performing the following operations: acquiring the resource data to be processed for the interaction item; performing preset processing on the resource data to be processed to obtain interpretable text corresponding to the resource data to be processed; and performing vector transformation processing on the interpretable text to obtain the semantic vector.

[0121] In this exemplary embodiment, the preset processing can be determined based on the resource attributes of the resource data to be processed, and this embodiment does not limit this. For example, if the resource attribute is that the content is fixed but the semantics are rich and ambiguous, then the preset processing is determined to be deep semantic analysis and disambiguation processing; if the resource attribute is user-defined and the title generally contains a lot of marketing noise or meaningless characters, then the preset processing is determined to be word filtering and extraction processing.

[0122] In this exemplary embodiment, once the preset processing is determined, the resource data to be processed can be subjected to preset processing to obtain the explanatory text corresponding to the resource data to be processed.

[0123] For example, see Figure 6 As shown, the resources to be processed are first determined to be songs and their lyrics, and then the songs and lyrics are input into a large language model (e.g., Figure 6 The DeepSeek-V3 model shown is based on a mixture of multi-head latent attention and expert input, thus obtaining interpretive text (e.g.) Figure 6 The tags "Popular Cantonese", "Emotional", and "Late Night Atmosphere" are shown.

[0124] In this embodiment of the disclosure, when the explanatory text corresponding to the resource data to be processed is obtained, a vector transformation can be performed on the explanatory text. This vector transformation is, for example, as described in [reference needed]. Figure 4 The implementation method shown is not detailed here. This allows us to obtain the semantic vectors of the interaction items, thereby obtaining the four-tuple features of the interaction items and determining the user's historical behavior sequence.

[0125] Step 502: In the feature input layer of the recommendation model, the semantic vector of the target music resource is concatenated with the semantic vector of each interaction item in the historical behavior sequence, and the resource identifier vector of the target music resource is concatenated with the resource identifier vector of each interaction item in the historical behavior sequence to obtain the fused feature representation.

[0126] Step 503: In the attention layer of the recommendation model, based on the fusion feature representation, calculate the attention weight between each interaction item in the user's historical behavior sequence and the target music resource.

[0127] Step 504: Obtain the user's interest matching degree for the target music resource based on the attention weight.

[0128] In this disclosure embodiment, see Figure 7 As shown, the input attention and feature fusion layer in the recommendation model not only calculates the relevance between the interaction item and the target music resource in terms of interaction behavior (i.e., concatenating the semantic vector of the target music resource with the semantic vector of each interaction item in the historical behavior sequence to obtain behavior-related attention), but also calculates the similarity between the interaction item and the target music resource in terms of text content and semantic style (i.e., concatenating the resource identifier vector of the target music resource with the resource identifier vector of each interaction item in the historical behavior sequence to obtain semantic matching attention). The weights of the behavior-related attention and semantic matching attention are then weighted and summed with the user's historical sequence vector to obtain semantic feature interest vector and structured feature interest vector. These two vectors are then fused using concatenation and a multilayer perceptron (MLP) to obtain a cross-feature vector, i.e., the aforementioned fused feature representation. Click-through rate (CTR) prediction is performed on the cross-feature vector, and finally, the click-through rate (CTR) of the user for the target music resource is output, measuring the user's interest matching degree for the target music resource based on the CTR.

[0129] In this exemplary embodiment, the traditional local activation unit is improved when calculating the user's interest matching degree for the target music resource. Specifically, a query-key pair is constructed, where Query(Q) is set as the representation vector of the target music resource (Candidate Item), and Key(K) is set as the representation vector of the interaction item in the user's historical behavior sequence. The calculation logic of the attention score is as follows: an outer product and element-wise subtraction operation is performed on Q and K, that is, an outer product and element-wise subtraction operation is performed on the quadruple features corresponding to the interaction item and the target music resource respectively to obtain the cross feature vector corresponding to each target music resource, so as to explicitly capture the cross features of the two in the semantic space and ID space. Then, the obtained cross feature vector is input into a multilayer perceptron (MLP) and a Dice activation function to output the contribution weight of the interaction item to the target music resource. Further, the calculated weights are used to perform a weighted summation of the user's historical behavior sequence to obtain the user's current interest vector, which is then fed into a fully connected layer for the final CTR (click-through rate) prediction.

[0130] Step 305: Determine the target recommended content to the user based on the interest matching degree, and display the explanatory text corresponding to the target recommended content as the recommendation reason.

[0131] In this embodiment of the disclosure, for example, the top N target music resources with the highest click-through rates are selected for recommendation. For example, recommendation information corresponding to the top 3 target music resources with the highest click-through rates (e.g., "sad and healing pop music" or "nostalgic emotional selection") is recommended.

[0132] The music resource recommendation method provided in this disclosure not only offers interpretable recommendation information to users, reducing their cognitive load, but also, by simultaneously capturing users' "behavioral habits" and "semantic preferences," is more comprehensive than relying solely on behavioral attention, thus improving recommendation accuracy. Furthermore, for music resources lacking behavioral data, or for alternative music resources with low popularity, a niche audience, limited exposure, or limited widespread appeal within the overall music resource pool, their semantic vectors can effectively supplement the lack of behavioral data, allowing the recommendation model to still accurately calculate interest matching and improve the recommendation effect for new or long-tail target music resources.

[0133] Exemplary device

[0134] Exemplary embodiments of this disclosure also provide a music resource recommendation device. (See reference...) Figure 8 As shown, the music resource recommendation device 800 includes the following program units: The acquisition unit 801 is used to acquire the associated corpus data of the target music resource, wherein the associated corpus data includes text information related to the target music resource; The generation unit 802 is used to input the associated corpus data into a large language model, perform semantic reasoning on the associated corpus data through the large language model, and generate explanatory text corresponding to the target music resource. The explanatory text is used to describe the content features of the target music resource. Encoding unit 803 is used to encode the explanatory text into a semantic vector; Processing unit 804 is used to calculate the matching degree between the target music resource and the user's interests based on the semantic vector and the user's historical behavior sequence through a recommendation model; The recommendation unit 805 is used to determine the target recommended content to be recommended to the user based on the interest matching degree, and to display the explanatory text corresponding to the target recommended content as the recommendation reason.

[0135] In one possible implementation, the target music resource includes physical resources and / or aggregate resources; the physical resources include songs and / or artists; and the aggregate resources include playlists and / or podcasts.

[0136] In one possible implementation, the generating unit 802 is configured to: If the target music resource is an entity resource, at least one of the lyrics, user comments, and music reviews in the target music resource is extracted from the associated corpus data. The extracted content is then subjected to semantic analysis and disambiguation processing based on the large language model using thought chain reasoning to generate explanatory text corresponding to the entity resource. If the target music resource is a collection-type resource, the title and / or description of the collection-type resource are subjected to noise filtering and semantic summarization processing by the large language model to generate explanatory text corresponding to the collection-type resource.

[0137] In one possible implementation, the generating unit 802 is configured to: Based on the lyrics text, high-frequency words and imagery features in the lyrics are identified; Determine the emotional tone based on the user comments and / or music reviews; Based on the imagery features and the emotional tone, the content features of the target music resource are determined, and the content features include at least one of the following: musical style, emotion, scene, and genre.

[0138] In one possible implementation, the generating unit 802 is configured to: Filter out invalid words of a preset type from the title and / or description of the collection-type resources; Extract at least one of the following information from the filtered title and / or description: core style information, scene information, and emotional information; The extracted information is compressed into standardized phrases, which serve as explanatory text for the aggregated resources.

[0139] In one possible implementation, the device includes a supplementary unit for: Based on the retrieval enhancement generation mechanism, related information concerning the target music resource is retrieved from the database and used as context input for the large language model. The related information includes at least one of the following: lyrics, user comments, and artist background information. Structured instructions are input to the large language model based on a preset prompt template, wherein the prompt template includes at least one of role settings, task constraints, and example information.

[0140] In one possible implementation, the encoding unit 803 is used for: Explanatory texts generated for different types of target music resources are encoded into semantic vectors in the same semantic vector space using the same vector encoder.

[0141] In one possible implementation, the processing unit 804 is configured to: Obtain the user's historical behavior sequence, where each interaction item in the historical behavior sequence includes a resource identifier vector and a corresponding semantic vector; In the feature input layer of the recommendation model, the semantic vector of the target music resource is concatenated with the semantic vector of each interaction item in the historical behavior sequence, and the resource identifier vector of the target music resource is concatenated with the resource identifier vector of each interaction item in the historical behavior sequence to obtain a fused feature representation. In the attention layer of the recommendation model, based on the fusion feature representation, the attention weight between each interaction item in the user's historical behavior sequence and the target music resource is calculated; The user's interest matching degree for the target music resource is obtained by weighting the attention weights.

[0142] The specific details of each part of the above-mentioned device have been described in detail in the method section of the implementation plan. For any undisclosed details, please refer to the implementation plan of the method section, and therefore will not be repeated here.

[0143] It should be noted that although several modules or units for the device used to perform actions have been mentioned in the detailed description above, this division is not mandatory. In fact, according to exemplary embodiments of this disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0144] Exemplary program product

[0145] Exemplary embodiments of this disclosure also provide a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the aforementioned method for recommending music resources.

[0146] In one implementation, the computer program product can be a tangible product containing a computer program, such as a computer-readable storage medium storing the computer program. The readable storage medium can be a storage medium based on electrical, magnetic, optical, electromagnetic, infrared, or other signals, including but not limited to: random access memory (RAM), read-only memory (ROM), magnetic tape, floppy disk, flash memory, hard disk drive (HDD), solid-state drive (SSD), etc. For example, the computer program product can be implemented as a non-volatile storage medium storing a computer program, such as read-only memory, NAND flash memory, etc.

[0147] In one implementation, the computer program product can be an intangible product containing a computer program. For example, the computer program product can be implemented as a virtual digital product, such as an executable file, installation package, or other digital file storing the computer program.

[0148] Computer program code can be written in one or more programming languages. Examples of programming languages ​​include C, Java, and C++. Program code can execute entirely on the user's computing device, partially on the user's computing device, or as a standalone software package. It can also execute partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, such as a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via an internet connection provided by a mobile network operator).

[0149] Computer programs can be carried or transmitted via signals such as electricity, magnetism, light, electromagnetic fields, and infrared radiation. Electronic devices can convert signals carrying computer programs into digital signals, thereby running the computer programs. When a computer program runs on an electronic device, its code is used to cause the electronic device to execute (more specifically, the processor of the electronic device to execute) the method steps of various exemplary embodiments of this disclosure, such as the music resource recommendation method described above, which includes the following steps: Step 301: Obtaining associated corpus data of the target music resource, the associated corpus data including text information related to the target music resource; Step 302: Inputting the associated corpus data into a large language model, performing semantic reasoning on the associated corpus data through the large language model to generate explanatory text corresponding to the target music resource, the explanatory text being used to describe the content features of the target music resource; Step 303: Encoding the explanatory text into a semantic vector; Step 304: Calculating the interest matching degree between the target music resource and the user based on the semantic vector and the user's historical behavior sequence through a recommendation model; Step 305: Determining the target recommendation content to recommend to the user based on the interest matching degree, and displaying the explanatory text corresponding to the target recommendation content as the recommendation reason.

[0150] By implementing the above method steps through a computer program, associated corpus data of the target music resource can be obtained. This associated corpus data includes textual information related to the target music resource. Then, the associated corpus data is input into a large language model, which performs semantic reasoning on the associated corpus data to generate explanatory text corresponding to the target music resource. This explanatory text describes the content features of the target music resource. Finally, the explanatory text is encoded into a semantic vector. Further, based on the semantic vector and the user's historical behavior sequence, a recommendation model can calculate the interest matching degree between the target music resource and the user. Thus, the target recommended content can be determined based on the interest matching degree, and the explanatory text corresponding to the target recommended content can be displayed as the recommendation reason.

[0151] In this embodiment of the disclosure, when recommending target content to a user, explanatory text corresponding to the target content is also displayed to the user, which can reduce the user's decision-making cost to a certain extent, thereby improving the user's click-through rate and user experience of the target content.

[0152] Exemplary electronic devices

[0153] Exemplary embodiments of this disclosure also provide an electronic device. The electronic device may include a processor and a memory. The memory stores executable instructions for the processor, such as computer programs. The processor executes these executable instructions to perform the method steps of various exemplary embodiments of this disclosure.

[0154] The following is for reference. Figure 9The electronic device is illustrated by way of a general-purpose computing device. It should be understood that... Figure 9 The electronic device 900 shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments disclosed herein.

[0155] like Figure 9 As shown, the electronic device 900 may include: a processor 910, a memory 920, a bus 930, an I / O (input / output) interface 940, and a network adapter 950.

[0156] The memory 920 may include volatile memory, such as RAM 921 and cache unit 922, and may also include non-volatile memory, such as ROM 923. The memory 920 may also include one or more program modules 924, including but not limited to: an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment. For example, program module 924 may include the units described in the above-described apparatus.

[0157] The processor 910 may include one or more processing units, such as an AP (Application Processor), a modem processor, a GPU (Graphics Processing Unit), an ISP (Image Signal Processor), a controller, an encoder, a decoder, a DSP (Digital Signal Processor), a baseband processor, and / or an NPU (Neural-Network Processing Unit).

[0158] The processor 910 can be used to execute executable instructions stored in the memory 920, such as the music resource recommendation method described above, which includes the following steps: Step 301: Obtain the associated corpus data of the target music resource, the associated corpus data including text information related to the target music resource; Step 302: Input the associated corpus data into a large language model, perform semantic reasoning on the associated corpus data through the large language model, and generate explanatory text corresponding to the target music resource, the explanatory text being used to describe the content features of the target music resource; Step 303: Encode the explanatory text into a semantic vector; Step 304: Based on the semantic vector and the user's historical behavior sequence, calculate the interest matching degree between the target music resource and the user through a recommendation model; Step 305: Determine the target recommendation content to be recommended to the user based on the interest matching degree, and display the explanatory text corresponding to the target recommendation content as the recommendation reason.

[0159] By executing the above method steps through processor 910, associated corpus data of the target music resource can be obtained. This associated corpus data includes text information related to the target music resource. Then, the associated corpus data is input into a large language model, which performs semantic reasoning on the associated corpus data to generate explanatory text corresponding to the target music resource. This explanatory text describes the content features of the target music resource. Furthermore, the explanatory text is encoded into a semantic vector. Further, based on the semantic vector and the user's historical behavior sequence, a recommendation model can calculate the interest matching degree between the target music resource and the user. Thus, the target recommended content can be determined based on the interest matching degree, and the explanatory text corresponding to the target recommended content can be displayed as the recommendation reason.

[0160] In this embodiment of the disclosure, when recommending target content to a user, explanatory text corresponding to the target content is also displayed to the user, which can reduce the user's decision-making cost to a certain extent, thereby improving the user's click-through rate and user experience of the target content.

[0161] Bus 930 is used to connect different components of electronic device 900 and may include data bus, address bus and control bus.

[0162] Electronic device 900 can communicate with one or more external devices 1000 (such as keyboard, mouse, external controller, etc.) through I / O interface 940.

[0163] Electronic device 900 can communicate with one or more networks via network adapter 950. For example, network adapter 950 can provide mobile communication solutions such as 3G / 4G / 5G, or wireless communication solutions such as wireless LAN, Bluetooth, and near-field communication. Network adapter 950 can communicate with other modules of electronic device 900 via bus 930.

[0164] although Figure 9 As not shown in the diagram, other hardware and / or software modules may also be configured in the electronic device 900, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0165] As can be seen from the above, the technical solutions disclosed herein can be implemented as methods, apparatus, systems, computer program products, storage media, electronic devices, etc. Those skilled in the art will understand that various aspects of this disclosure can be specifically implemented in the following forms: a completely hardware implementation, a completely software implementation (including firmware, microcode, etc.), or an implementation combining hardware and software aspects, which may be referred to as "circuit," "module," or "system," respectively.

[0166] It should be understood that this disclosure is not limited to the specific methods, steps, or structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. Those skilled in the art will readily conceive of other embodiments based on the specific implementations provided in this disclosure. Therefore, the specific implementations provided in this disclosure are merely exemplary, and the scope and spirit of this disclosure are indicated by the claims, and should cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary technical means in the art not disclosed in this disclosure.

Claims

1. A method for recommending music resources, characterized in that, include: Obtain associated corpus data of the target music resource, wherein the associated corpus data includes text information related to the target music resource; The associated corpus data is input into a large language model, and semantic reasoning is performed on the associated corpus data through the large language model to generate explanatory text corresponding to the target music resource. The explanatory text is used to describe the content features of the target music resource. The explanatory text is encoded into a semantic vector; Based on the semantic vector and the user's historical behavior sequence, the matching degree between the target music resource and the user's interests is calculated through a recommendation model; Based on the interest matching degree, target recommended content is determined to be recommended to the user, and the explanatory text corresponding to the target recommended content is displayed as the recommendation reason.

2. The method according to claim 1, characterized in that, The target music resources include physical resources and / or aggregated resources; the physical resources include songs and / or artists; the aggregated resources include playlists and / or podcasts.

3. The method according to claim 2, characterized in that, The step of performing semantic reasoning on the associated corpus data using the large language model to generate explanatory text corresponding to the target music resource includes: If the target music resource is an entity resource, at least one of the lyrics, user comments, and music reviews in the target music resource is extracted from the associated corpus data. The extracted content is then subjected to semantic analysis and disambiguation processing based on the large language model using thought chain reasoning to generate explanatory text corresponding to the entity resource. If the target music resource is a collection-type resource, the title and / or description of the collection-type resource are subjected to noise filtering and semantic summarization processing by the large language model to generate explanatory text corresponding to the collection-type resource.

4. The method according to claim 3, characterized in that, The semantic analysis and disambiguation processing of the extracted content based on the thought chain reasoning of the large language model includes: Based on the lyrics text, high-frequency words and imagery features in the lyrics are identified; Determine the emotional tone based on the user comments and / or music reviews; Based on the imagery features and the emotional tone, the content features of the target music resource are determined, and the content features include at least one of the following: musical style, emotion, scene, and genre.

5. The method according to claim 3, characterized in that, The step of performing noise filtering and semantic summarization processing on the titles and / or descriptions of the aggregated resources using the large language model includes: Filter out invalid words of a preset type from the title and / or description of the collection-type resources; Extract at least one of the following information from the filtered title and / or description: core style information, scene information, and emotional information; The extracted information is compressed into standardized phrases, which serve as explanatory text for the aggregated resources.

6. The method according to claim 1, characterized in that, Before inputting the associated corpus data into the large language model, the method further includes: Based on the retrieval enhancement generation mechanism, related information concerning the target music resource is retrieved from the database and used as context input for the large language model. The related information includes at least one of the following: lyrics, user comments, and artist background information. Structured instructions are input to the large language model based on a preset prompt template, wherein the prompt template includes at least one of role settings, task constraints, and example information.

7. The method according to claim 1, characterized in that, Encoding the explanatory text into a semantic vector includes: Explanatory texts generated for different types of target music resources are encoded into semantic vectors in the same semantic vector space using the same vector encoder.

8. The method according to claim 1, characterized in that, The step of calculating the match between the target music resource and the user's interests using a recommendation model based on the semantic vector and the user's historical behavior sequence includes: Obtain the user's historical behavior sequence, where each interaction item in the historical behavior sequence includes a resource identifier vector and a corresponding semantic vector; In the feature input layer of the recommendation model, the semantic vector of the target music resource is concatenated with the semantic vector of each interaction item in the historical behavior sequence, and the resource identifier vector of the target music resource is concatenated with the resource identifier vector of each interaction item in the historical behavior sequence to obtain a fused feature representation. In the attention layer of the recommendation model, based on the fused feature representation, the attention weight between each interaction item in the user's historical behavior sequence and the target music resource is calculated; The user's interest matching degree for the target music resource is obtained by weighting the attention weights.

9. An electronic device, characterized in that, include: processor; Memory for storing the executable instructions of the processor; The processor is configured to perform the method of any one of claims 1-8 by executing the executable instructions.

10. A computer program product storing a computer program thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-8.

Images