A media asset playing intention type determination method and system based on multi-signal correction

Abstract

The application discloses a media playing intention type determination method and system based on multi-signal correction, and relates to the technical field of artificial intelligence. The method comprises the following steps: generating an initial playing intention instruction according to a text input by a user through a first large language model; identifying a key entity with a target semantic from the initial playing intention instruction; identifying a target media type and a confidence degree of the target media type as structured data according to the initial playing intention instruction through a second large language model; and performing multi-signal correction on a media playing intention type according to the initial playing intention instruction, the key entity and the structured data to obtain a target playing intention. The application performs semantic analysis on the text input by the user through the large language model, generates the initial playing intention instruction, and corrects the instruction from multiple dimensions to obtain the target playing intention conforming to the real intention of the user, thereby solving the problem of inaccurate intention recognition caused by semantic ambiguity and insufficient context perception.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and system for determining the type of media asset playback intent based on multi-signal correction. Background Technology

[0002] With the rapid development of artificial intelligence, natural language processing, and voice interaction technologies, smart terminals (such as smart speakers, smart TVs, and in-vehicle voice systems) have widely entered the daily lives of the general public. Controlling media content playback via voice commands has become one of the mainstream interaction methods. Against this backdrop, accurately identifying the user's true playback intent is a core challenge for a user-friendly human-computer interaction experience. Currently, most smart terminals employ intent recognition methods based on rule matching or traditional machine learning models. These methods rely on predefined keyword libraries and limited state machine logic, and their accuracy drops significantly when faced with ambiguous user expressions, contextual dependencies, or implicit intents. Summary of the Invention

[0003] The main objective of this application is to propose a method and system for determining media asset playback intent type based on multi-signal correction, so as to improve the accuracy of media asset playback intent recognition.

[0004] To achieve the above objectives, one aspect of this application proposes a method for determining the media asset playback intent type based on multi-signal correction, the method comprising the following steps: The initial playback intent instruction is generated based on the user's input text using the first large-scale language model; Key entities with target semantics are identified from the initial playback intent command; The target media asset type and the confidence level of the target media asset type are identified as structured data by the second large-scale language model based on the initial playback intent command; Based on the initial playback intent instruction, the key entity, and the structured data, the media asset playback intent type is modified using multiple signals to obtain the target playback intent.

[0005] In some embodiments, generating the initial playback intent instruction based on the user-input text using a first large language model includes the following steps: The text obtained by transing the user's speech is input into the first large-scale language model; The initial playback intent instruction is generated based on the text using the first large-scale language model.

[0006] In some embodiments, identifying key entities with target semantics from the initial playback intent instruction includes the following steps: In the media asset playback scenario, the named entity recognition model is used to identify the person entity, media asset name entity, and media asset type entity from the initial playback intent instruction based on the existing media asset information, which are then used as the key entities.

[0007] In some embodiments, the step of identifying the target media asset type and the confidence level of the target media asset type as structured data based on the initial playback intent instruction using a second large-scale language model includes the following steps: The network retrieval information is passed through to the second large language model through the service interface of the second large language model; The second large-scale language model is used to analyze the network retrieval information to obtain the multidimensional attributes of the media assets; The structured data is obtained by using the second large-scale language model based on the initial playback intent instruction and the multidimensional attributes of the media asset; wherein, the structured data includes the target media asset type and the confidence level of the target media asset type.

[0008] In some embodiments, the step of performing multi-signal correction on the media asset playback intent type based on the initial playback intent instruction, the key entity, and the structured data to obtain the target playback intent includes the following steps: The rule engine corrector uses a user syntax constraint template library to identify and classify the initial playback intent instruction and the key entities using regular expressions or pattern matching methods, thereby generating a first comprehensive score and a first weight. By using a context correction signaler, and combining the current foreground and background application status of the terminal device with the historical dialogue content within the effective session period, the target media asset type and the confidence level of the target media asset type are corrected, thereby generating a second comprehensive score and a second weight. Using a long-term memory correction signal generator, a user profile model indexed by user identifier and device unique ID is constructed based on desensitized user behavior data. This model records user media asset preferences, frequently used application configurations, and historical playback trajectories, thereby generating a third comprehensive score and a third weight. The target playback intent is obtained by weighting and fusing the corresponding comprehensive scores according to each weight.

[0009] In some embodiments, each of the comprehensive scores is calculated through the following steps: Each comprehensive score is calculated according to the comprehensive score calculation formula for a single corrector: ; in, The overall score for the i-th corrected signal device. , , , These are the basic coefficients set according to business requirements. The confidence level is the output of the signal corrector itself. To calculate the historical calibration accuracy based on sliding time window statistics. For context fit factor, For personalized factors; Each of the aforementioned weights is calculated through the following steps: Each weight is calculated according to the weighting formula for a single corrector signal: ; in, The weights of the i-th corrected signal device, For temperature coefficient, This refers to the sequence number of the corrector signal. The number of the correction signalers.

[0010] In some embodiments, the method further includes the following steps: The target playback intent is converted into executable control instructions that conform to the target platform interface specifications; The executable control command is executed to play the media asset corresponding to the target playback intent.

[0011] To achieve the above objectives, another aspect of this application proposes a media asset playback intent type determination system based on multi-signal correction, the system comprising: The instruction recognition unit is used to generate initial playback intent instructions based on the text input by the user through a first large-scale language model; An entity recognition unit is used to identify key entities with target semantics from the initial playback intent command; The media asset type identification unit is used to identify the target media asset type and the confidence level of the target media asset type as structured data based on the initial playback intent instruction using a second large-scale language model. The intent correction unit is used to perform multi-signal correction on the media asset playback intent type based on the initial playback intent instruction, the key entity, and the structured data to obtain the target playback intent.

[0012] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0013] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0014] To achieve the above objectives, another aspect of this application provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0015] The embodiments of this application include at least the following beneficial effects: This application provides a method and system for determining media asset playback intent type based on multi-signal correction. The solution generates an initial playback intent command based on user-input text using a first large-scale language model; identifies key entities with target semantics from the initial playback intent command; uses a second large-scale language model to identify the target media asset type and its confidence level as structured data based on the initial playback intent command; and performs multi-signal correction on the media asset playback intent type based on the initial playback intent command, key entities, and structured data to obtain the target playback intent. This application uses a large-scale language model to semantically parse the user-input text, generate an initial playback intent command, and corrects the initial playback intent command from multiple dimensions to obtain a target playback intent that matches the user's true intent, solving the problem of inaccurate intent recognition caused by semantic ambiguity and insufficient context awareness in existing technologies. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating a method for determining media asset playback intent type based on multi-signal correction, provided in an embodiment of this application; Figure 2 A schematic diagram of a media asset playback intent type determination system based on multi-signal correction provided in this application embodiment; Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of systems and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0020] Before providing a detailed description of the embodiments of this application, some related technologies involved in the embodiments of this application will be described first, as follows: Intelligent terminal: refers to an electronic device with an independent system, network connectivity, a human-computer interaction interface, and cloud-based intelligent processing capabilities. It is capable of running applications, sensing user input, understanding user intent, and providing intelligent services. In this text, the intelligent terminal is exemplified by a smart speaker, but is not limited to this form.

[0021] Media asset playback intent type refers to the subjective purpose expressed by users during interaction with a media playback system regarding what media content to play and what playback method to use. In this article, media asset playback intent types are mainly divided into live streaming, on-demand, replay, music playback, and audio content playback.

[0022] Named entity recognition (NER) is a fundamental task in natural language processing (NLP). It can automatically identify entity names with specific meanings from unstructured text and classify them into predefined categories, such as personal names, place names, organization names, time, numbers, proper nouns, etc.

[0023] A rule template engine is a component based on predefined rules and template structures that can match, parse, or generate text, instructions, or business logic. By separating variable parameters from fixed logic, it achieves efficient identification and processing of specific patterns and can be applied in intent recognition scenarios.

[0024] Large models refer to Large Language Models (MLMs), a type of deep learning model that typically has billions or even more parameters and demonstrates superior performance in natural language processing tasks. In this paper, we train intent recognition and media asset recognition based on an 8B large model.

[0025] Multi-signal correction: This refers to a technical method that designs multiple signal correctors from different dimensions and analyzes them during the user intent recognition process. These correctors dynamically calibrate and optimize the media playback intent initially determined by the large model. Each signal corrector evaluates its impact on the current playback intent and outputs the corrected intent and media asset slotting. The system integrates the intent scores from each corrector to adjust and rank the initial intent with confidence, thereby outputting a final playback intent that better meets the user's actual needs. This improves the accuracy of intent understanding and service adaptation capabilities of smart terminals in complex scenarios.

[0026] Intelligent agent: refers to a functional module or logical entity running in the system in this article. It can make logical judgments and decisions by sensing user input, system status and historical dialogue context, and construct corresponding operation instructions to achieve specific task objectives.

[0027] Existing technologies suffer from inaccurate intent recognition. For example, when a user commands "play 'YYY'", since "YYY" could refer to a musical work or a TV series, the system struggles to determine whether the user intends to play audio or video on demand based solely on the literal meaning. In such cases, the system may repeatedly call different types of media asset query interfaces for trial matching, increasing response latency and consequently slowing down service response and impacting user experience. In recent years, with the rapid iteration of large-scale language model technology, some systems have begun to introduce large models for semantic understanding and intent parsing. Large models, with their powerful contextual modeling and knowledge generalization capabilities, have achieved significant improvements in initial intent recognition. However, relying solely on the initial intent output by large models still presents some problems, such as the difficulty in dynamically integrating multi-dimensional contextual information like real-time device status and user historical behavior, and the inability to dynamically calibrate and re-evaluate the initial intent due to the lack of media asset information. Furthermore, although some research has attempted to introduce user profiles or simple feedback mechanisms for intent adjustment, these methods lack systematic architectural design and cannot achieve collaborative correction of multi-dimensional signals. Especially on resource-constrained smart terminals, achieving high-precision intent optimization while ensuring low latency remains a pressing technical challenge. Therefore, this application aims to propose an intent optimization method based on multi-signal correction, combining large-model semantic understanding capabilities with a dynamic correction mechanism for multi-source contextual signals. This comprehensive approach aims to accurately identify and continuously optimize user media asset playback intent, thereby improving the intent understanding capabilities and service response accuracy of smart terminals in complex real-world scenarios.

[0028] The core objective of this application is to propose a method and system for determining media asset playback intent type based on multi-signal correction, addressing the inaccurate intent recognition issues caused by semantic ambiguity and insufficient context awareness in existing technologies. The method utilizes a large language model to semantically parse user commands, generating an initial playback intent. Simultaneously, media types with conflicting intents are categorized into "music," "audio content," and "video." Multiple independent signal correctors are introduced, employing rule template engines, named entity recognition (NER), contextual information, and long-term user behavioral memory to assess the confidence level and dynamically adjust the initial intent. The system further integrates the results from each corrector, designing an intent decision evaluation mechanism to generate the final playback intent. This method achieves refined calibration of the large model output, effectively reducing the misjudgment rate in scenarios with ambiguous commands (such as "play something interesting") or unclear commands (such as "play YYY"). Furthermore, by designing an efficient and user-friendly system workflow, redundant calls to media asset interfaces are reduced, response latency is lowered, and system operating efficiency is improved. Ultimately, this will enable smart terminals to respond to user needs more accurately, quickly, and intelligently, significantly improving the human-computer intelligent interaction experience.

[0029] This application provides a method and system for determining media asset playback intent type based on multi-signal correction, relating to the field of artificial intelligence technology. The method and system provided in this application can be applied to a terminal, a server, or software running on a terminal or server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or in-vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing 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, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing a method for determining media asset playback intent type based on multi-signal correction, but is not limited to the above forms.

[0030] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0031] Reference Figure 1 This application provides a method for determining the media asset playback intent type based on multi-signal correction. This method may include, but is not limited to, steps S100 to S130, as follows: S100: Generates initial playback intent instructions based on the user's input text using the first large-scale language model; S110: Identify key entities with target semantics from the initial playback intent command; S120: The target media asset type and the confidence level of the target media asset type are identified as structured data by the second large-scale language model based on the initial playback intent instruction; S130: Perform multi-signal correction on the media asset playback intent type according to the initial playback intent instruction, the key entity, and the structured data to obtain the target playback intent.

[0032] For example, embodiments of this application can use a large language model to semantically parse user commands, generate an initial playback intent, and classify media types with conflicting intents into "music, audio content, and video." Multiple independent signal correctors are introduced, each based on a rule template engine, named entity recognition (NER), contextual information, and long-term user behavioral memory, to assess the confidence level of the initial intent and dynamically adjust it. The system further integrates the results of each corrector signal to design an intent decision evaluation mechanism and generate the final playback intent.

[0033] Optionally, generating the initial playback intent instruction based on the user-input text using a first large-scale language model includes the following steps: The text obtained by transing the user's speech is input into the first large-scale language model; The initial playback intent instruction is generated based on the text using the first large-scale language model.

[0034] Specifically, the input of the first large-scale language model is the textualized sentence forwarded by the smart terminal (text translated after speech recognition), and the output is the initial intention instruction that predicts the user's expected task goal, while extracting key information slots from the text.

[0035] Optionally, identifying key entities with target semantics from the initial playback intent instruction includes the following steps: In the media asset playback scenario, the named entity recognition model is used to identify the person entity, media asset name entity, and media asset type entity from the initial playback intent instruction based on the existing media asset information, which are then used as the key entities.

[0036] Specifically, the named entity recognition model is used to automatically identify and classify key entities with specific semantic meanings from the initial playback intent command, providing structured semantic support for the parsing and correction of media asset playback intent. The named entity recognition model references existing media resource information. In the media resource playback scenario involved in the embodiments of this application, the named entity recognition model focuses on identifying three types of entities: person entities, corresponding to the creators or performers of the content; media asset name entities (such as "YYY"), referring to specific programs, films, music, or works; and media asset type entities (such as "music," "TV series," "live broadcast"), used to distinguish content formats. The recognition results of the named entity recognition model are output in the form of structured slots and passed as feature inputs to the multi-signal intent correction module to assist in evaluating the rationality of the initial intent.

[0037] Optionally, the step of identifying the target media asset type and the confidence level of the target media asset type as structured data based on the initial playback intent command using a second large-scale language model includes the following steps: The network retrieval information is passed through to the second large language model through the service interface of the second large language model; The second large-scale language model is used to analyze the network retrieval information to obtain the multidimensional attributes of the media assets; The structured data is obtained by using the second large-scale language model based on the initial playback intent instruction and the multidimensional attributes of the media asset; wherein, the structured data includes the target media asset type and the confidence level of the target media asset type.

[0038] Specifically, the second large-scale language model is used to extract fine-grained information slots related to media resources from user commands and perform semantic inference. It is a large language model based on an 8B parameter scale and has undergone lightweight fine-tuning. The service interface of the second large-scale language model supports online plugins, allowing information retrieved from the internet to be passed to the model for reference. The second large-scale language model further analyzes multi-dimensional attributes such as content type (e.g., music, audio content, video), media asset tags (e.g., pop music, humanities, suspense), director, actors, singers, and production year, achieving accurate characterization of the target media asset. The input of the second large-scale language model is the text command converted from speech recognition, and the output is structured data containing the recognition results and their confidence levels.

[0039] Optionally, the step of performing multi-signal correction on the media asset playback intent type based on the initial playback intent instruction, the key entity, and the structured data to obtain the target playback intent includes the following steps: The rule engine corrector uses a user syntax constraint template library to identify and classify the initial playback intent instruction and the key entities using regular expressions or pattern matching methods, thereby generating a first comprehensive score and a first weight. By using a context correction signaler, and combining the current foreground and background application status of the terminal device with the historical dialogue content within the effective session period, the target media asset type and the confidence level of the target media asset type are corrected, thereby generating a second comprehensive score and a second weight. Using a long-term memory correction signal generator, a user profile model indexed by user identifier and device unique ID is constructed based on desensitized user behavior data. This model records user media asset preferences, frequently used application configurations, and historical playback trajectories, thereby generating a third comprehensive score and a third weight. The target playback intent is obtained by weighting and fusing the corresponding comprehensive scores according to each weight.

[0040] Specifically, the multi-signal intent correction module in this application is a core functional module that corrects the initial PLAY intent to the user's desired intent. This module is used to fine-tune the initial PLAY intent output by the intent recognition model and is the core functional unit for achieving high-precision intent understanding. This module integrates multiple independently operating signal correctors, each corresponding to different dimensions of contextual information sources, including rule-engine-based grammatical structure correction, contextual correction based on the current dialogue and device state, and memory-based correction based on the user's long-term behavioral preferences. Each signal corrector analyzes the support or interference level of its perceived signals to the initial intent according to preset rules or lightweight evaluation criteria, and outputs a corresponding confidence adjustment value or weight coefficient. The system synthesizes the output results of each corrector through weighted fusion, ranking decision, or voting mechanisms to generate the final corrected intent.

[0041] Optionally, each of the aforementioned comprehensive scores is calculated through the following steps: Each comprehensive score is calculated according to the comprehensive score calculation formula for a single corrector: ; in, The overall score for the i-th corrected signal device. , , , These are the basic coefficients set according to business requirements. The confidence level is the output of the signal corrector itself. To calculate the historical calibration accuracy based on sliding time window statistics. For context fit factor, For personalized factors; Each of the aforementioned weights is calculated through the following steps: Each weight is calculated according to the weighting formula for a single corrector signal: ; in, The weights of the i-th corrected signal device, For temperature coefficient, This refers to the sequence number of the corrector signal. The number of the correction signalers.

[0042] It is understood that the basic coefficients, context adaptation factors, and personalized factors in the embodiments of this application are not fixed parameter scores, but are dynamically generated and adjusted based on the following accumulated business experience to ensure that the strategy for correcting intentions is reasonable and adaptable.

[0043] Optionally, the method further includes the following steps: The target playback intent is converted into executable control instructions that conform to the target platform interface specifications; The executable control command is executed to play the media asset corresponding to the target playback intent.

[0044] It is understood that the embodiments of this application can convert the final playback intent output by the multi-signal intent correction module into executable control instructions that conform to the target platform interface specifications, thereby achieving adaptation to different media asset service platform protocol formats and calling methods. After receiving the correction intent, the module dynamically decides the optimal execution scheme by combining the real-time terminal status of the terminal device (such as whether a player is currently running in the foreground or background), user preference settings, and default player configuration.

[0045] The following sections will provide a detailed description and explanation of some optional embodiments of this application, using specific application examples.

[0046] This application's embodiments primarily design the interaction process of seven core modules: intelligent terminal, intent hub service, multi-signal intent correction, intent recognition model, NER model, media asset recognition model, and business intelligent agent. The responsibilities of each module are described below: (1) Smart Terminal Module: This refers to a device with computing power and a rich user interface, capable of running various applications and providing a rich user interaction experience. Smart terminals possess powerful processors, large-capacity storage, and high-resolution screens, enabling them to perform complex data processing and graphics display tasks. Smart terminals typically run operating systems (such as Android, iOS, Windows, etc.) and support the installation and operation of third-party applications. Common smart terminals include: smartphones, tablets, smart TVs, smart speakers, etc.

[0047] (2) Intent Central Service Module: This module is responsible for coordinating and processing the interaction between various intent recognition and processing functions in this system. It receives raw user commands from the smart terminal and distributes them in parallel to various intent recognition models for initial parsing. Simultaneously, it integrates the model recognition results to arbitrate the initial intent. When the arbitrated initial intent is a media playback type user intent (hereinafter referred to as "PLAY" intent), it can trigger the multi-signal intent correction module to execute the corresponding correction process; at the same time, it can delegate the corrected intent (hereinafter referred to as "corrected intent") to the corresponding business intelligence module. Furthermore, the Intent Central Service is also responsible for maintaining the context state during user sessions, supporting multi-round interactions and intent caching, and supporting synchronous and asynchronous call operation modes.

[0048] (3) Intent Recognition Model Module: This refers to a Large Language Model (LLM) based on an 8-byte parameter scale and lightweight fine-tuning. Supervised model fine-tuning is performed using a large amount of labeled intent recognition training data accumulated from historical business experience. The input to the model module is a textual sentence forwarded by the smart terminal (text translated after speech recognition). The output is the initial intent command predicting the user's expected task objective, while simultaneously extracting key information slots from the text. For example, if the user's input is "Play YYY's interlude," the model can output "PLAY, music, YYY interlude." This model possesses strong semantic understanding and sentence generalization capabilities, accurately parsing diverse and non-standardized natural language expressions from users.

[0049] (4) NER Model Module: This refers to a named entity recognition component based on natural language processing technology, used to automatically identify and classify key entities with specific semantic meanings from the user's original commands, providing structured semantic support for the parsing and correction of media asset playback intent. This model references existing media resource information. In the media resource playback scenario involved in the embodiments of this application, this model focuses on identifying three types of entities: person entities, corresponding to the creators or performers of the content; media asset name entities, referring to specific programs, films, music, or works; and media asset type entities (such as "music," "TV series," "live broadcast"), used to distinguish content forms. The recognition results of the NER model are output in the form of structured slots and passed as feature inputs to the multi-signal intent correction module to assist in evaluating the rationality of the initial intent. For example, when the user inputs "play YYY," the NER model identifies "YYY" as a media asset name entity, which, combined with contextual information, can further participate in the intent disambiguation of media resource content, becoming an important basis for the multi-signal intent correction module.

[0050] (5) Media Asset Recognition Model Module: This module is specifically designed to extract fine-grained information slots related to media resources from user commands and perform semantic inference. It is a large language model based on an 8B parameter scale and has undergone lightweight fine-tuning. The model's service interface supports online plugins, allowing information retrieved from the internet to be passed to the model for reference. The model further analyzes multi-dimensional attributes such as content type (e.g., music, audio content, video), media asset tags (e.g., pop music, humanities, suspense), director, actors, singers, and production year, achieving accurate characterization of the target media asset. Its input is the text command converted from speech recognition, and its output is structured data containing the recognition results and their confidence levels. For example, for "play song YYY", it generates {"type": "music", "name": "YYY", "confidence": 0.96}. This structured output is then passed to the multi-signal intent correction module as a key basis for media asset content matching and initial intent correction.

[0051] (6) Multi-signal intention correction module: It refers to a core functional module in the system of the embodiment of this application that corrects the initial PLAY intention to the user's expected intention. This module is used to finely calibrate the initial play (PLAY) intention output by the intention recognition model and is the core functional unit for achieving high-precision intention understanding. This module integrates multiple independently operating signal correctors, corresponding to different-dimensional context information sources respectively, including grammar structure correction based on a rule engine, context correction based on the current conversation and device status, and memory correction based on the user's long-term behavior preferences, etc. Each signal corrector analyzes the support or interference degree of the signals it perceives for the initial intention according to preset rules or lightweight evaluation criteria, and outputs the corresponding confidence adjustment value or weight coefficient. The system generates the final corrected intention by weighted fusion, sorting decision, or voting mechanism, integrating the output results of each corrector. This process effectively alleviates the problem of misjudgment of the initial intention caused by semantic ambiguity, naming ambiguity (such as "play YYY"), expression diversity, or lag in media asset information update, and significantly improves the intention recognition accuracy and robustness of the system in complex real-world scenarios. The main logics of each signal corrector are introduced below: First, the rule engine correction signaler, based on the user grammar constraint template library constructed (and continuously updated and iterated) through long-term business accumulation, uses regular expressions or pattern matching techniques to accurately identify and classify the sentence patterns of user instructions. This corrector establishes a rule set for high-frequency and enumerable expression patterns. For example, in the music playback scenario, it matches typical sentence patterns such as "verb + object +'s'+ music type" (such as "play the songs of XXX star"), "verb +'s'+ album / song" (such as "play the album of XXX star"), "play a song of XXX", "want to listen to the songs of XXX", and the independent phrase "the songs of XXX". Through this mechanism, instructions with clear semantic structures can be quickly identified, improving the intention determination accuracy and response speed in rule-based scenarios.

[0052] Second, the context correction signaler combines the current foreground and background application states of the terminal device and the historical conversation content within the effective conversation cycle to analyze the execution environment on which the user's operation intention depends. For example, the prerequisite for playing music is that the music player is in an active state; if there is no active player process on the current terminal, it is determined that the application launch operation needs to be executed first; if the foreground application is already a music player, when the user enters a fuzzy instruction (such as "play something interesting"), the context corrector infers that the user is more likely to expect to play audio content based on the current running environment, thus clearly correcting the initial PLAY intention to a "music category" intention, achieving semantic disambiguation based on the runtime context; Meanwhile, the long-term memory correction signal generator constructs a user profile model indexed by user identifiers and unique device IDs based on anonymized user behavior data, recording their media asset preferences, frequently used application configurations, and historical playback trajectories. When a user issues an ambiguous command (such as "play YYY"), the correction generator retrieves the user's recent viewing behavior (e.g., having watched the TV series "YYY" up to episode 4), combines content popularity and usage frequency, outputs a tendency judgment, and corrects the initial intent to "video playback". By introducing a personalized memory mechanism, the system can better adapt to user habits and improve the accuracy of intent recognition in cross-modal ambiguous scenarios.

[0053] The three types of corrective signal generators mentioned above operate in parallel, and each corrective signal generator obtains a comprehensive score after completing the analysis. Then, the weighted fusion mechanism of the dynamic weight decision model (i.e., the normalized exponential weighting mechanism) is used to calculate the comprehensive weight of each signaler. Finally, the scores of each signal device are accumulated based on the results of different categories, and compared to obtain the corrected intention result. The main calculation methods involved are as follows: Overall score of a single signal device: Then, the weighted fusion is adjusted to incorporate the overall weights: The final score for the signaler's intention to correct is: .

[0054] in: It is the confidence level output by the corrector itself (such as rule matching strength, NER recognition score); 2. It is the historical calibration accuracy based on the sliding time window statistics, reflecting the reliability of the corrector in recent interactions (e.g., whether the user invokes a multi-turn interaction mechanism to correct errors after the response result is sent to the end). 3. It is the context adaptability factor, which can be dynamically calculated based on the current device state, session stage, and user behavior pattern. It represents the degree of support for the signal by the current device state or session stage (for example, when the foreground application is a video player, this value can be set to a high value (such as above 0.9); the long-term memory corrector has a lower value at the beginning of a new session and gradually increases as context accumulates). It is a personalized factor that is dynamically adjusted based on user profiles (e.g., the proportion of user preferences for different modal content (e.g., if a user consumes 80% of their content in video, then the factor score of video-related signals will be increased)). Currently, the basic coefficients are set based on business parameters (for example, the rule engine assigns higher basic weights in high-frequency scenarios). 6. It is a temperature coefficient that can be dynamically adjusted to adapt to different ambiguous business scenarios. For example, when the instruction is clear (such as "play 'XXX song'"), the temperature value is set to a higher value (such as 1.0) to enhance the dominance of high-confidence signals; when the instruction is ambiguous (such as "play something interesting"), it is reduced (such as 0.2) to make the calculation results more conservative and avoid misjudgment when scoring the comprehensive intent category.

[0055] In the embodiments of this application, the base coefficient, context adaptation factor, and personalized factor are not fixed parameter scores, but are dynamically generated and adjusted based on the following accumulated business experience to ensure that the strategy for correcting intentions is reasonable and adaptable.

[0056] (7) Business Intelligence Module: This refers to a functional module in the system of this application embodiment that controls operation instructions. This module is responsible for converting the final playback intent output by the multi-signal intent correction module into executable control instructions that conform to the target platform interface specifications, thereby achieving adaptation to different media asset service platform protocol formats and calling methods. After receiving the correction intent, this module dynamically decides the optimal execution scheme based on the real-time terminal status of the terminal device (such as whether a player is currently running in the foreground or background), user preference settings, and default player configuration. For example, when the intent is determined to be playing music content and the terminal has no active player, the business intelligence agent first generates an instruction to start the specified music application (such as "open NetEase Cloud Music"), and then generates the corresponding playback control instructions; if the target player is already running, it directly generates operation instructions such as play, pause, or skip tracks supported by the player. All instructions are returned to the intelligent terminal in a streaming manner, and the terminal side is responsible for the specific execution, ensuring real-time response and operational reliability.

[0057] An optional implementation example is as follows: The technical solutions in the embodiments of this application will be clearly and completely described below with reference to implementation examples. Obviously, the described embodiments are only some embodiments of the embodiments of this application, and not all embodiments.

[0058] The device or system module objects involved in the implementation examples of this application include: smart speaker, intent hub service, multi-signal intent correction module, intent recognition model 8B service, named entity recognition (NER) service, media asset recognition model 8B service, and business intelligent agents (such as music agents, video agents, and media asset conflict multi-turn interaction agents). This application embodiment describes the process of system calling the model to identify the user's initial intent and correct it. The main steps are described below: Step S1: The user wakes up the smart speaker with voice, initiating the voice interaction process; the speaker collects audio data and uploads it to the intent hub service, which calls the ASR speech recognition module to convert the audio content into text information, completing the initial parsing from speech to text. For example, if the user says "play YYY", the system obtains the corresponding text after ASR recognition and records the current device status (such as whether a music player is running in the background), providing basic data support for subsequent contextual analysis.

[0059] Step S2: After receiving the text input, the Intent Central Service calls the Intent Recognition Model 8B service and the NER service in parallel. The Intent Recognition Model 8B, based on the semantic understanding capabilities of the large model, initially determines that the user's action is a media playback intent (i.e., PLAY intent). At the same time, the NER service extracts key entities from the text and identifies the keyword "YYY". However, since this word may refer to a song or a TV series, it returns multiple candidate media asset entities and their corresponding content types (music, video). The system combines the two outputs to perform preliminary intent arbitration, confirming that the initial intent is PLAY, but there is modal ambiguity, which requires further calibration.

[0060] Step S3 (optional): If the NER service fails to clearly distinguish the entity category or extract complete media asset information, the intent hub service will call the media asset recognition model 8B service for supplementary recognition. This model has online network retrieval capabilities and can obtain detailed information about "YYY" by accessing third-party knowledge sources, including structured data such as content type (e.g., period drama), tags (e.g., fantasy, romance), lead actors, and director, and infer the most likely media asset semantics based on this. Subsequently, the system calls the Entity-Linking interface of the NER service again to align the parsing results with the standard entities in the business knowledge base, verify and confirm the high-confidence media asset entities and their unique identifiers, thereby providing a more accurate input basis for subsequent intent correction.

[0061] Step S4: After obtaining the initial intent and media asset entity information, the intent hub service sends the data to the multi-signal intent correction module and initiates the concurrent correction mechanism. This module simultaneously activates the rule engine signaler, context signaler, and long-term memory signaler: the rule engine signaler performs pattern matching based on preset sentence templates (such as "play + work name" commonly used in music scenarios) and outputs a support score; the context signaler analyzes the current device status (such as whether a player is running in the background) and the historical dialogue content within the effective session period to determine whether the current execution environment is more inclined to audio or video playback; the long-term memory signaler queries the anonymized user profile built based on user behavior. If it is found that the user has previously watched "YYY" up to episode 2, the support weight of the "video" intent is significantly increased; each signaler outputs its own confidence adjustment value or weight coefficient, and the system completes multi-signal score arbitration through weighted fusion or sorting decision to generate the final corrected intent.

[0062] Step S5: When the system completes the intent correction and determines that there is no modal conflict, the intent hub service schedules the corrected intent to the corresponding business intelligence agent module. If the corrected intent is music-related, it is forwarded to the music agent (for example, this agent can construct instructions that meet the format requirements of the NetEase Cloud Music application and send them to the terminal for execution), which generates control instructions and response bodies that conform to the platform interface specifications. If the corrected intent is video-related or live-streaming-related, it is forwarded to the video agent (for example, this agent can construct instructions that meet the requirements of the China Telecom Ultra HD application) and encapsulates the corresponding playback instructions. The response body returned by the business intelligence agent is forwarded to the smart speaker terminal via the intent hub service, where the terminal performs specific operations, such as launching the application or starting playback.

[0063] Step S6 (optional): If the system still determines that there is a possibility of intent conflict after correction (e.g., music and video scores are close, making a clear decision impossible), the playback action is not executed directly. Instead, the task is dispatched to the media asset conflict multi-round interaction agent. This agent generates natural language follow-up questions based on the candidate media asset information (e.g., "Do you want to listen to YYY's songs or watch the TV series of the same name?") and constructs an interactive information card containing multiple candidate media assets (including cover image, type, and description). This response is returned to the smart speaker via the intent hub service and rendered on the screen while being broadcast via voice, guiding the user to make a secondary selection, thus achieving intent clarification and precise service under multi-round interaction. At the same time, the clarified intent information is cached in long-term memory.

[0064] This application presents a method and system for determining media asset playback intent type based on multi-signal correction. Its core technology involves generating an initial playback intent through a large language model and classifying conflict-prone media asset types into music, audio content, and video categories. It then introduces multiple parallel signal correctors, including a rule engine, contextual information, named entity recognition (NER), and user long-term behavioral memory, to perform multi-dimensional confidence assessment and dynamic adjustment of the initial intent. By designing a weighted fusion mechanism that includes contextual adaptability factors and personalized factors, it achieves intelligent weight allocation for each corrected signal, generating the final playback intent. This effectively solves the problem of inaccurate intent recognition caused by semantic ambiguity, naming ambiguity, and insufficient contextual awareness in existing technologies, significantly reducing the misjudgment rate under ambiguous commands such as "play YYY". Simultaneously, the system improves entity recognition completeness through a collaborative completion mechanism of NER and an online media asset recognition model, and triggers multi-round interactive agents for user clarification when intent conflicts occur, balancing decision accuracy and user-friendliness. Overall, it reduces redundant interface calls, lowers response latency, and improves the response speed and human-computer interaction experience of smart terminals in complex scenarios. Compared to existing technologies, this method offers higher accuracy, stronger interpretability, more efficient response, a more intelligent interactive system, and a better user experience.

[0065] Reference Figure 2 This application also provides a media asset playback intent type determination system based on multi-signal correction, which can implement the above-mentioned media asset playback intent type determination method based on multi-signal correction. The system includes: The instruction recognition unit is used to generate initial playback intent instructions based on the text input by the user through a first large-scale language model; An entity recognition unit is used to identify key entities with target semantics from the initial playback intent command; The media asset type identification unit is used to identify the target media asset type and the confidence level of the target media asset type as structured data based on the initial playback intent instruction using a second large-scale language model. The intent correction unit is used to perform multi-signal correction on the media asset playback intent type based on the initial playback intent instruction, the key entity, and the structured data to obtain the target playback intent.

[0066] It is understood that the content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0067] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method of this application. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0068] It is understood that the content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the methods of this application, and the beneficial effects achieved are the same as those achieved by the methods of this application.

[0069] Please see Figure 3 , Figure 3 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 301 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 302 can be implemented as a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM). The memory 302 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 302 and is called and executed by the processor 301. Input / output interface 303 is used to implement information input and output; The communication interface 304 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, network cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 305 transmits information between various components of the device (e.g., processor 301, memory 302, input / output interface 303, and communication interface 304); The processor 301, memory 302, input / output interface 303, and communication interface 304 are connected to each other within the device via bus 305.

[0070] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of this application.

[0071] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0072] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0073] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0074] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0075] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0076] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0077] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0078] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0079] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, 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 (item) 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 (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0080] In the embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between systems or units may be electrical, mechanical, or other forms.

[0081] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0082] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0083] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0084] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for determining the media asset playback intent type based on multi-signal correction, characterized in that, The method includes the following steps: The initial playback intent instruction is generated based on the user's input text using the first large-scale language model; Key entities with target semantics are identified from the initial playback intent command; The target media asset type and the confidence level of the target media asset type are identified as structured data by the second large-scale language model based on the initial playback intent command; Based on the initial playback intent instruction, the key entities, and the structured data, the media asset playback intent type is modified using multiple signals to obtain the target playback intent; The rule engine corrector uses a user syntax constraint template library to identify and classify the initial playback intent instruction and the key entities using regular expressions or pattern matching methods, thereby generating a first comprehensive score and a first weight. By using a context correction signaler, and combining the current foreground and background application status of the terminal device with the historical dialogue content within the effective session period, the target media asset type and the confidence level of the target media asset type are corrected, thereby generating a second comprehensive score and a second weight. Using a long-term memory correction signal generator, a user profile model indexed by user identifier and device unique ID is constructed based on desensitized user behavior data. This model records user media asset preferences, frequently used application configurations, and historical playback trajectories, thereby generating a third comprehensive score and a third weight. The target playback intent is obtained by weighting and fusing the comprehensive scores corresponding to each weight. Each comprehensive score is calculated according to the comprehensive score calculation formula for a single corrector: ； in, The overall score for the i-th corrected signal device. , , , These are the basic coefficients set according to business requirements. The confidence level is the output of the signal corrector itself. To calculate the historical calibration accuracy based on sliding time window statistics. For context fit factor, This indicates the level of signal support provided by the current device state or session phase. For personalized factors; Each weight is calculated according to the weighting formula for a single corrector signal: ； in, The weights of the i-th corrected signal device, For temperature coefficient, This refers to the sequence number of the corrector signal. The number of the correction signalers.

2. The method for determining media asset playback intent type based on multi-signal correction according to claim 1, characterized in that, The process of generating initial playback intent instructions based on user-input text using a first large-scale language model includes the following steps: The text obtained by transing the user's speech is input into the first large-scale language model; The initial playback intent instruction is generated based on the text using the first large-scale language model.

3. The method for determining media asset playback intent type based on multi-signal correction according to claim 1, characterized in that, The process of identifying key entities with target semantics from the initial playback intent command includes the following steps: In the media asset playback scenario, the named entity recognition model is used to identify the person entity, media asset name entity, and media asset type entity from the initial playback intent instruction based on the existing media asset information, which are then used as the key entities.

4. The method for determining media asset playback intent type based on multi-signal correction according to claim 1, characterized in that, The step of identifying the target media asset type and the confidence level of the target media asset type as structured data based on the initial playback intent command using a second large-scale language model includes the following steps: The network retrieval information is passed through to the second large language model through the service interface of the second large language model; The second large-scale language model is used to analyze the network retrieval information to obtain the multidimensional attributes of the media assets; The structured data is obtained by using the second large-scale language model based on the initial playback intent instruction and the multidimensional attributes of the media asset; wherein, the structured data includes the target media asset type and the confidence level of the target media asset type.

5. A method for determining media asset playback intent type based on multi-signal correction according to any one of claims 1 to 4, characterized in that, The method further includes the following steps: The target playback intent is converted into executable control instructions that conform to the target platform interface specifications; The executable control command is executed to play the media asset corresponding to the target playback intent.

6. A media asset playback intent type determination system based on multi-signal correction, characterized in that, The system is used to implement the media asset playback intent type determination method based on multi-signal correction as described in claim 1, the system comprising: The instruction recognition unit is used to generate initial playback intent instructions based on the text input by the user through a first large-scale language model; An entity recognition unit is used to identify key entities with target semantics from the initial playback intent command; The media asset type identification unit is used to identify the target media asset type and the confidence level of the target media asset type as structured data based on the initial playback intent instruction using a second large-scale language model. The intent correction unit is used to perform multi-signal correction on the media asset playback intent type based on the initial playback intent instruction, the key entity, and the structured data to obtain the target playback intent.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 5.

8. 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 described in any one of claims 1 to 5.

Images