An AI speech recognition-based language processing system for an elderly care service terminal

Abstract

This invention relates to the field of smart elderly care technology and discloses a language processing system for elderly care service terminals based on AI speech recognition. The system includes: a semantic understanding and contextual fusion module for fusing recognized text and real-time data to generate contextualized semantic representations; a multi-domain parallel parsing module with multiple semantic parsers for parallel collaborative parsing of the contextualized semantic representations and generating a list of candidate subtasks; a structured plan arbitration module for resolving conflicts among all candidate subtasks and performing logical and temporal relationship reasoning to generate a structured service plan; and a resource scheduling and execution module for maintaining a dynamic service resource library and generating instruction sequences to drive the execution of specific service resources. This invention enables the automatic and reliable conversion of unstructured, fuzzy speech commands from the elderly into a safe, coherent, and executable integrated elderly care service solution.

Description

Technical Field

[0001] This invention relates to the field of smart elderly care technology, and in particular to a language processing system for elderly care service terminals based on AI voice recognition. Background Technology

[0002] Voice interaction is a crucial entry point for seniors to access digital elderly care services. Current technologies typically employ general speech recognition and intent classification models to process seniors' voice commands. However, this approach has technical limitations in practical applications of elderly care services.

[0003] First, due to accents, dialects, colloquialisms, and cognitive expression habits, elderly people's voice commands often exhibit high ambiguity and non-standardity. General-purpose speech models show a significant decrease in accuracy in recognizing and extracting intent from such expressions. Simultaneously, processing voice signals in isolation, ignoring real-time physical context information collected by terminal sensors, leads to a severe deficiency in the system's assessment of the user's true state and the level of urgency.

[0004] Secondly, elderly care needs are inherently cross-domain and complex. A simple statement like "I feel dizzy, and there's no food at home" simultaneously involves medical health and daily care. Existing systems, employing sequential or single-domain intent classification models, can only process one explicit intent or require users to express it multiple times. Their architecture cannot support multi-domain parallel parsing and collaborative analysis of a single voice input, resulting in fragmented service responses and failing to meet the holistic needs of the elderly.

[0005] More importantly, elderly care services have extremely high requirements for safety and continuity. When generating service plans, existing technologies only consider the simple logical sequence of task execution, and their resource scheduling is mostly based on simple matching of static skill tags or efficiency-first optimization algorithms. They do not consider dynamic reputation factors such as the real-time status of service personnel, historical performance quality, and suitability for specific elderly people, resulting in high risks and low acceptance of the plans in actual implementation.

[0006] To address the aforementioned technical problems, this invention proposes a language processing system for elderly care service terminals based on AI speech recognition. Summary of the Invention

[0007] The purpose of this invention is to solve the problems of semantic ambiguity, intent fragmentation, risk oversight, and rigid scheduling in the prior art, and to propose a language processing system for elderly care service terminals based on AI speech recognition.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: a language processing system for elderly care service terminals based on AI speech recognition, comprising: The semantic understanding and context fusion module is used to fuse recognized text from the voice interface and real-time data from terminal sensors to generate contextualized semantic representations that include user intent, emotional state, and environmental context. The multi-domain parallel parsing module contains multiple semantic parsers corresponding to different elderly care service domains. It is used to call one or more semantic parsers to perform parallel collaborative parsing of contextualized semantic representations and combine the outputs of each parser to generate a list of candidate subtasks. The structured plan arbitration module is used to resolve conflicts based on confidence comparison for all received candidate subtasks, and to reason about the logical and temporal relationships between tasks according to the preset domain knowledge graph, generating a structured service plan consisting of multiple dependent subtasks. The resource scheduling and execution module is used to maintain a dynamic service resource library and, based on the structured service plan and the dynamic status information of the service resources obtained from the resource library, optimize and schedule the execution of each sub-task in the plan, and generate a sequence of instructions to drive the execution of specific service resources.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention, by setting up a semantic understanding and contextual fusion module and employing cross-modal reliability assessment and contextualized semantic synthesis technology, can achieve deep and accurate fusion understanding of non-standard speech and multi-source sensor information of the elderly. It can not only parse the literal intent, but also perceive the user's emotions and environmental context, generating high-quality contextualized semantic representations, fundamentally improving the naturalness of interaction and the robustness of intent recognition.

[0010] This invention, by setting up a multi-domain parallel parsing module and adopting a context-aware routing and inter-parser collaborative communication mechanism, can efficiently perform parallel analysis and collaborative judgment on complex and ambiguous user requests containing elements from multiple domains. Like an expert consultation, it can mobilize parsers from different domains to work together and exchange information, thereby comprehensively generating a comprehensive and accurate list of candidate subtasks, effectively solving the problem of limited processing capacity of a single parsing model.

[0011] This invention, by setting up a structured plan arbitration module and introducing risk avoidance decision-making and continuity constraint reasoning based on risk transmission analysis, can intelligently combine a series of sub-tasks into a safe and reliable structured service plan. This not only clarifies the logical sequence of tasks but also proactively avoids health and safety risks, ensures the continuity of nursing services, and provides alternative paths for core tasks, thereby generating personalized and resilient service solutions that truly fit the physical and mental characteristics of the elderly.

[0012] This invention establishes a resource scheduling execution module and constructs a collaborative commitment scheduling mechanism based on multi-dimensional resource profiles and a reputation database. This mechanism enables efficient, flexible, and reliable matching of service resources. It replaces rigid assignment with negotiated commitments, fully respects the wishes and status of resource providers, and optimizes matching by combining historical service quality. This significantly improves the practical feasibility and execution quality of the scheduling scheme, and enhances the predictability of service delivery and user trust.

[0013] This invention constitutes a complete technical closed loop from accurate perception, intelligent analysis, risk perception planning to reliable scheduling. It can automatically transform a vague and complex spoken request from an elderly person into a safe, reliable, and executable overall service solution, significantly improving the intelligence level, personalization, and system reliability of elderly care services, and providing core technical support for achieving high-quality smart elderly care with services available at the request of an elderly person. Attached Figure Description

[0014] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 This is a schematic diagram of the system configuration provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system flow provided for an embodiment of the present invention. Detailed Implementation

[0016] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a language processing system for an elderly care service terminal based on AI speech recognition proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0017] 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 invention pertains.

[0018] The following examples are for illustrative purposes and are not intended to limit the scope of the invention.

[0019] The following description, in conjunction with the accompanying drawings, details a specific solution for a language processing system for elderly care service terminals based on AI speech recognition, provided by this invention.

[0020] Please see Figure 1 and Figure 2 It illustrates a system configuration diagram and a system flow diagram of a language processing system for an elderly care service terminal based on AI speech recognition, according to an embodiment of the present invention, including: The semantic understanding and context fusion module is used to fuse recognized text from the voice interface and real-time data from terminal sensors to generate contextualized semantic representations that include user intent, emotional state, and environmental context. The multi-domain parallel parsing module contains multiple semantic parsers corresponding to different elderly care service domains. It is used to call one or more semantic parsers to perform parallel collaborative parsing of contextualized semantic representations and combine the outputs of each parser to generate a list of candidate subtasks. The structured plan arbitration module is used to resolve conflicts based on confidence comparison for all received candidate subtasks, and to reason about the logical and temporal relationships between tasks according to the preset domain knowledge graph, generating a structured service plan consisting of multiple dependent subtasks. The resource scheduling and execution module is used to maintain a dynamic service resource library and, based on the structured service plan and the dynamic status information of the service resources obtained from the resource library, optimize and schedule the execution of each sub-task in the plan, and generate a sequence of instructions to drive the execution of specific service resources.

[0021] It should be noted that the recognized text refers to the digitized text content obtained by converting the elderly user's voice signal through automatic speech recognition technology via the voice interface. It is the basic data for extracting the user's intent, and its generation quality directly affects the accuracy of subsequent semantic parsing.

[0022] Real-time data from terminal sensors refers to the environmental and status data collected at the current moment by various sensors mounted on elderly care service terminals. This includes spatial information acquired by location sensors, physiological data collected by health monitoring sensors, environmental parameters recorded by temperature and humidity sensors, and activity status captured by motion sensors, providing objective data support for contextual understanding.

[0023] Contextualized semantic representation refers to a unified semantic carrier formed by integrating user intent, emotional state, and environmental context. It can comprehensively depict the scenario background and personalized characteristics of user needs, enabling the system to move beyond literal semantic interpretation and achieve a deep understanding of scenario-based needs.

[0024] User intent refers to the core needs and goals expressed by elderly users through voice, including medical and health consultations, requests for daily care, safety alerts and assistance, and emotional companionship. It is the core basis for the system to provide services.

[0025] Emotional state refers to the user's emotional tendency obtained through the analysis of voice features and semantic content, including anxiety, calmness, loneliness, and eagerness, which provides a reference for the system to provide personalized service responses that match the emotions.

[0026] Contextual information refers to objective situational elements composed of sensor data and scene information, including location, time period, environmental parameters, and activity stage, which are used to help determine the scene adaptability to user needs.

[0027] The elderly care service sector refers to the professional service category related to the lives of elderly users, including core areas such as medical and health care, daily living care, safety warnings, and emotional companionship, covering the main demand scenarios for elderly care services.

[0028] A semantic parser is a semantic processing tool optimized for a specific service domain. It has a domain-specific vocabulary and requirement parsing logic, and can accurately identify user needs in the corresponding domain and transform them into standardized sub-tasks.

[0029] Parallel collaborative parsing refers to the simultaneous operation of multiple semantic parsers. Through data sharing and interactive collaboration, they jointly handle complex cross-domain requirements, avoiding the functional blind spots of a single parser and achieving comprehensive coverage of requirements.

[0030] The candidate subtask list refers to the collection of tasks formed by the parsing results of each semantic parser. Each subtask contains key information such as task type, priority, triggering conditions, and required resource types, providing basic materials for subsequent plan generation.

[0031] Confidence comparison refers to the process of quantitatively evaluating and ranking the reliability of each candidate subtask. The confidence level is calculated based on factors such as parser matching degree and collaborative activity, providing a basis for task selection and priority ranking.

[0032] Conflict resolution refers to the process of selecting the optimal solution through risk assessment for problems such as resource competition, timing conflicts, and state constraint conflicts among candidate subtasks. The core objective is to reduce service execution risks.

[0033] Domain knowledge graphs are structured knowledge bases that store entities, relationships, and attributes related to elderly care services. They include core entities such as users, health risks, service resources, and task types, as well as the constraints and rules between entities, providing knowledge support for logical reasoning.

[0034] Logical and temporal reasoning refers to mining the dependencies and execution order requirements between subtasks based on domain knowledge graphs, including the order of execution, state preconditions, resource usage rules, etc., to ensure the logical rationality of the service plan.

[0035] A structured service plan is a standardized execution plan that includes multiple sub-tasks and dependencies. It clarifies the execution logic, timing nodes, constraints, and alternative paths of each task, providing a clear basis for resource scheduling.

[0036] Dependency refers to the logical relationship between subtasks, including temporal dependency, state dependency, resource dependency, etc., which determines the order of task execution and preconditions.

[0037] The dynamic service resource database refers to a database that stores information on various elderly care service resources, including nursing staff, service robots, smart home devices, medical service platforms, and terminals for children and family members, providing data support for resource allocation.

[0038] Dynamic status information of service resources refers to the real-time operational status data of various service resources, including the location and busy status of nursing staff, the online and operational status of equipment, and the responsiveness of the service platform. It is a key basis for optimizing resource matching.

[0039] Optimized matching and scheduling refers to the process of selecting the best resources and allocating execution time windows based on task requirements and resource status through algorithms. The core objective is to improve resource utilization and service response efficiency.

[0040] Instruction sequences refer to a standardized set of execution instructions adapted to different resource interface protocols. They adopt corresponding formats and parameters according to the differences in resource types to ensure that resources can accurately execute tasks.

[0041] This application's solution constructs a full-link technical architecture of "contextual fusion - parallel parsing - plan arbitration - resource scheduling". Specifically, firstly, the semantic understanding and contextual fusion module breaks through the limitations of single text interpretation, deeply integrating speech recognition text with real-time sensor data and scene information to generate contextualized semantic representations containing intent, emotion, and environment, enabling the system to accurately capture the contextual needs and potential demands of elderly users. Next, the multi-domain parallel parsing module, through built-in professional parsers for medical health, daily care, etc., achieves synchronous parsing and collaborative communication of cross-domain needs, responding to complex needs without multiple user interactions, reducing the operational threshold for elderly users. Then, the structured plan arbitration module, based on a domain knowledge graph, resolves conflicts through confidence comparison, adds constraints to subtasks by combining logic and temporal reasoning, and generates alternative execution paths, ensuring service security and continuity while addressing resource unavailability issues in unexpected scenarios. Finally, the resource scheduling and execution module, relying on a dynamic service resource library, optimizes matching based on real-time resource status and historical performance, transforming the structured service plan into instruction sequences adapted to different resources, achieving efficient utilization and precise scheduling of service resources. The entire process forms a closed-loop processing mechanism from understanding needs to implementing services, which fully adapts to the usage habits of elderly users and their high safety and personalized needs for elderly care services.

[0042] In one specific implementation, this system is applied to home-based elderly care scenarios, adapting to the daily service needs of elderly users living alone. The equipment used includes an elderly care service terminal host, an array microphone, a multi-sensor kit, an edge computing module, a 5G communication module, and interfaces with service robots, caregiver mobile terminals, smart home devices, community medical service platforms, and children's mobile terminals. The elderly care service terminal host serves as the core control device, integrating an edge computing module for data processing and algorithm execution; the array microphone acts as a voice interface to collect the elderly user's voice; the multi-sensor kit includes position sensors, health monitoring sensors, temperature and humidity sensors, and motion sensors to collect environmental and physiological data; and the 5G communication module enables real-time communication between the terminal and various service resources.

[0043] Semantic understanding and context fusion stage: An elderly user says in the bedroom, "I feel dizzy and want to check my blood pressure, and then I want to tell my son." The array microphone collects the voice signal and converts it into recognized text. The multi-sensor kit collects real-time data such as "bedroom location, heart rate 98 beats / minute, 3 pm, user activity stage is afternoon rest." The semantic understanding and context fusion module uses a BERT pre-trained model to extract semantic feature vectors, processes the sensor data through a convolutional neural network to obtain context feature vectors, and then fuses the two types of vectors through a dynamic weight allocation algorithm. Combined with the scene prior vector, a contextualized semantic representation is generated: "afternoon - bedroom - dizziness - high heart rate - need to check blood pressure - need to contact children - calm emotions."

[0044] Multi-domain parallel parsing stage: The parser library of the multi-domain parallel parsing module includes semantic parsers for four domains: medical health, daily care, safety warning, and emotional companionship. The context-aware routing unit extracts the scene triggering elements "dizziness," "blood pressure measurement," and "contacting son." The matching degree of the parser in the medical health domain is calculated to be 0.95, and the matching degree of the parser in the emotional companionship domain is 0.90, both of which are higher than the activation threshold of 0.3, triggering the two parsers. The inter-parser collaborative communication unit shares user demand data between the two parsers based on a collaborative weighted connection table. The competitive result synthesis unit calculates the comprehensive confidence of each parser and generates a list of candidate sub-tasks: ① Start the smart blood pressure monitor to measure blood pressure (medical health, confidence 0.96); ② Send a voice notification to the child's mobile terminal (emotional companionship, confidence 0.93).

[0045] In the structured plan arbitration phase: the domain knowledge graph of the structured plan arbitration module stores entities and relationships related to elderly care services; the risk perception and transmission analysis unit identifies the health risk entity "abnormal blood pressure" associated with the sub-task and constructs an analysis model; there are no candidate sub-task conflicts, so no conflict resolution is required; the continuity constraint reasoning unit infers the temporal dependency relationship through the domain knowledge graph, determining that "measuring blood pressure" is executed before "sending notification," with a minimum time interval of 5 minutes; the initial plan synthesis unit generates an initial structured service plan, clarifying the serial execution logic and timing of the two tasks; the plan encapsulation unit generates alternative paths for "starting the smart blood pressure monitor to measure blood pressure" (if the smart blood pressure monitor malfunctions, the community medical service platform will be dispatched to remotely guide the measurement), forming the final structured service plan.

[0046] Resource scheduling and execution phase: The dynamic service resource library of the resource scheduling and execution module stores resource information such as smart blood pressure monitors, children's mobile terminals, and community medical service platforms. The resource profile includes static capabilities, real-time status, historical ratings, and user feedback. The collaborative commitment scheduling unit issues task proposals to the smart blood pressure monitor and children's mobile terminals and receives autonomous commitment feedback: smart blood pressure monitor (estimated execution time 1 minute, confidence level 0.98), children's mobile terminals (estimated response time 3 minutes, confidence level 0.99). The global constraint coordination and arbitration unit confirms no conflicts and generates a scheduling plan. The instruction sequence generation and distribution unit generates instructions adapted to different resources: sending MQTT protocol control commands to the smart blood pressure monitor to start the measurement function; sending RESTful API format interaction instructions to the children's mobile terminals, including user status and blood pressure measurement plan; distributing instructions through the 5G communication module and receiving real-time execution status feedback to ensure service implementation.

[0047] I. Semantic Understanding and Context Fusion Module The semantic understanding and context fusion module includes a multimodal feature extraction unit, a cross-modal reliability assessment unit, a feature fusion unit, and a contextualized semantic synthesis unit; The multimodal feature extraction unit is used to parse the speech recognition text to extract semantic feature vectors representing the user's intent and emotional state, and to process the terminal sensor data stream to extract context feature vectors representing the physical environment state. A cross-modal reliability assessment unit is used to calculate confidence scores for semantic feature vectors and quality scores for context feature vectors. The feature fusion unit is used to dynamically adjust the weights of the semantic feature vector and the context feature vector in the fusion process based on the confidence score and the quality score, and perform weighted fusion to output a preliminary fusion representation. The contextualized semantic synthesis unit is used to synthesize the preliminary fusion representation with a scenario prior vector jointly encoded by the current time, user activity stage, and device geographic location, and output a contextualized semantic representation.

[0048] It should be noted that speech recognition text refers to the digitized text content obtained by converting the speech signals of elderly users through a voice interface. It is the core text data for mining user intentions and emotions.

[0049] Semantic feature vectors are digital feature carriers extracted from speech recognition text. They can quantitatively represent user intent and emotional state, providing data support for semantic understanding.

[0050] Terminal sensor data stream refers to the raw data stream continuously collected by various sensors mounted on elderly care service terminals, including continuous data in dimensions such as location, health, environment, and movement.

[0051] Context feature vectors are digital feature carriers extracted from sensor data streams that can quantitatively characterize the real-time state of the physical environment and provide objective basis for context understanding.

[0052] The physical environment state refers to the objective environmental conditions perceived by sensors, including specific parameters such as location, temperature and humidity, lighting, and spatial scene.

[0053] The confidence score is a numerical value that quantifies the reliability of semantic feature vectors. It is calculated based on factors such as speech recognition accuracy and semantic parsing consistency. The score ranges from 0 to 1, with a higher score indicating a more reliable semantic feature.

[0054] The quality score is a numerical value that quantifies the effectiveness of context feature vectors. It is calculated based on factors such as sensor data integrity, sampling frequency, and noise level. The score ranges from 0 to 1, with a higher score indicating a more effective context feature.

[0055] Dynamic weight adjustment refers to optimizing the contribution ratio of the two types of features in the fusion process in real time based on the confidence score and quality score, so as to ensure that the fusion result is more in line with the actual needs of the scenario.

[0056] Weighted fusion refers to the process of mathematically integrating two types of feature vectors according to adjusted weights, achieving the organic combination of features through weighted summation, attention mechanisms, and other methods.

[0057] Preliminary fusion representation refers to the intermediate feature carrier obtained after weighted fusion, which integrates the core information of semantics and environment, but has not yet incorporated prior scene information, laying the foundation for subsequent scene-based synthesis.

[0058] Current time refers to the specific moment information at the time of feature synthesis, including time nodes such as hours and minutes, used to distinguish different time periods such as morning, afternoon, and night.

[0059] User activity stages refer to the habitual activity states derived from mining historical user behavior data, including morning exercise, afternoon rest, dinner preparation, etc., which are used to understand user needs by aligning with their daily routines.

[0060] The geographical location of the equipment refers to the specific spatial location information of the elderly care service terminal, including indoor areas such as the living room, bedroom, and bathroom, to clarify the spatial scenario of user needs.

[0061] Scene prior vectors are feature carriers obtained by quantifying and encoding the current time, user activity stage, and device geographical location, and can centrally represent the core attributes of the scene.

[0062] In one specific implementation, this module is applied to emergency nighttime scenarios for home-based elderly care, adapting to the needs of elderly users living alone when they suddenly experience physical discomfort.

[0063] Multimodal feature extraction stage: At 22:30 at night, an elderly user said "I feel tightness in my chest" in the bedroom. The array microphone collected the voice signal and converted it into speech recognition text. The multi-sensor kit collected data simultaneously. The position sensor identified that the device was in the bedroom area. The health monitoring sensor detected that the user's heart rate was 102 beats / minute and blood oxygen saturation was 93%. The temperature and humidity sensor recorded the indoor temperature as 26℃ and humidity as 55%. The multimodal feature extraction unit used a BERT pre-trained model to segment the speech recognition text and perform semantic role labeling, extracting a 128-dimensional semantic feature vector (representing the intent as health emergency help and the emotional state as anxiety). At the same time, the sensor data stream was filtered and normalized through a convolutional neural network to extract a 64-dimensional contextual feature vector (representing the physical environment as night - bedroom - warm and humid and the physiological state as high heart rate - low blood oxygen).

[0064] Cross-modal reliability assessment phase: The cross-modal reliability assessment unit calculates the confidence score of the semantic feature vector. Due to the speech recognition accuracy of 98% and the semantic parsing consistency of 95%, the confidence score is 0.96. The quality score of the context feature vector is calculated. Due to the completeness of the sensor data, the sampling frequency meets the standard, and the noise level is low, the quality score is 0.94.

[0065] Feature fusion stage: The feature fusion unit calculates the fusion weight based on the confidence score and the quality score. The semantic feature vector weight is 0.505 and the context feature vector weight is 0.495. An attention mechanism is used to fuse the two types of feature vectors after weighting, and outputs a preliminary fusion representation of 256 dimensions. This representation contains core information such as the user's health help intention, anxiety, bedroom scene at night, and abnormal physiological state.

[0066] Contextualized semantic synthesis stage: The contextualized semantic synthesis unit performs one-hot encoding and embedding vector combination on the current time (22:30 – nighttime rest period), user activity stage (nighttime rest), and device geographical location (bedroom) to generate a 32-dimensional scenario prior vector. The preliminary fusion representation and the scenario prior vector are then added dimension-by-dimensionally to synthesize the final 256-dimensional contextualized semantic representation: "Nighttime rest period – Bedroom – Living alone – Chest tightness – High heart rate – Low blood oxygen – Anxiety – Health emergency help request," providing comprehensive contextual support for subsequent multi-domain analysis.

[0067] II. Multi-domain Parallel Parsing Module The multi-domain parallel parsing module includes a parser library unit, a context-aware routing unit, an inter-parser cooperative communication unit, and a competitive result synthesis unit; The parser library unit is used to store and manage a collection of semantic parsers configured for the domains of healthcare, daily care, safety alerts, and emotional companionship; The context-aware routing unit is used to receive contextualized semantic representations, extract scene triggering elements from them, and calculate the scene matching degree between the scene triggering elements and each parser to generate a list of target parser triggers. The inter-parser collaborative communication unit is used to manage and maintain an inter-parser communication link network, which supports the exchange of structured messages between triggered parsers. The competitive result synthesis unit is used to aggregate and evaluate the parsing results of all triggered parsers, and output a list of candidate subtasks according to a preset synthesis strategy.

[0068] Furthermore, the context-aware routing unit is configured to perform the following operations: Extract one or more scene triggering elements from the received contextualized semantic representation; For each semantic parser in the parser library unit, a scene matching degree value is calculated based on the scene triggering element and the domain focus label of the parser. The scene matching score of each parser is compared with the preset activation threshold of that parser. Add the identifiers of all semantic parsers whose scene matching scores are greater than or equal to their own activation thresholds to the target parser trigger list.

[0069] Furthermore, the inter-parser cooperative communication unit is configured to perform the following operations: Maintain a predefined collaboration weight join table, which records the collaboration weight values ​​between any two semantic parsers; When any triggered semantic parser generates a parsing result, the collaboration weight connection table is queried to obtain a list of associated parsers with a collaboration weight greater than zero. Filter from the list of associated resolvers to find other resolvers that are currently in a triggered state; Send a collaborative consultation message containing the parsing results to all other selected parsers.

[0070] Furthermore, the competitive results synthesis unit is configured to perform the following operations: Maintain a cooperative activity counter for each triggered parser to record the total number of cooperative consultation messages sent and received by the parser during cooperative communication. For each parsing result, the scene matching degree of its source parser and the value of its collaborative activity counter are weighted and summed to calculate the overall confidence score of the parsing result; All analysis results are sorted in descending order based on the overall confidence score, and the top N results are selected to generate a candidate subtask list.

[0071] It should be noted that the semantic parser set refers to a collection of dedicated parsers for four domains: medical and health care, daily care, safety warning, and emotional companionship. Each parser is optimized for the vocabulary and requirement logic of its corresponding domain and has its own exclusive parsing capabilities.

[0072] The healthcare field refers to the service scope centered around the physiological health needs of elderly users, including disease consultation, health monitoring, medical assistance, and related needs analysis.

[0073] The field of daily care refers to the service scope that revolves around the daily needs of elderly users, including the analysis of related needs such as diet, living arrangements, housekeeping services, and travel assistance.

[0074] The field of safety early warning refers to the service scope surrounding the safety protection needs of elderly users, including emergency assistance, risk prevention, and alarm for abnormal conditions.

[0075] The field of emotional companionship refers to the service scope centered around the psychological and emotional needs of elderly users, including needs analysis related to chatting, emotional comfort, and interest-based interaction.

[0076] Scene triggering elements refer to the core features extracted from contextualized semantic representation that can characterize the domain to which the demand belongs, including domain keywords, scene features, user state features, etc.

[0077] Scene matching degree refers to the quantitative value of the degree of fit between scene triggering elements and semantic parser’s domain focus tags. It ranges from 0 to 1. The higher the score, the stronger the parser’s adaptability to the current needs.

[0078] Domain-specific tags refer to unique domain identifiers assigned to each semantic parser. These tags are used to clarify the core service scope of the parser and serve as the basis for calculating the scenario matching degree.

[0079] The activation threshold is a preset trigger threshold for each semantic parser, used to determine whether the scene matching degree meets the minimum requirements for starting the parser, and to ensure the effectiveness of the parser triggering.

[0080] The target parser trigger list is a list that compiles all semantic parser identifiers that meet the activation conditions, clearly defining the parser object that needs to be called for the current parsing requirement.

[0081] The inter-parser communication link network refers to a dedicated data transmission network that connects various semantic parsers, providing a stable channel for information exchange between parsers.

[0082] Structured messages refer to parsed result data organized in a standardized format, including core information such as subtask type, triggering conditions, and related data, to ensure the accuracy of data interaction between parsers.

[0083] The collaboration weighted connection table is a table that records the strength of the collaboration association between any two semantic parsers. The weight values ​​range from 0 to 1 and are used to determine whether the parsers need to communicate collaboratively.

[0084] The associated parser list refers to the set of parsers that have a collaborative association (collaboration weight greater than zero) with the parser that currently generates the parsing result, and identifies the objects whose parsing results need to be synchronized.

[0085] Collaborative consultation messages are standardized notification messages that contain parsing results. They are used to share requirement parsing information among associated parsers and support collaborative parsing.

[0086] The collaborative activity counter is a counting tool that records the frequency of each triggered parser's participation in collaborative communication. It reflects the degree of collaborative participation of the parser by counting the total number of collaborative consultation messages sent and received.

[0087] The comprehensive confidence score is a quantitative value of the reliability of the analysis results calculated by combining the scene matching degree and the collaborative activity level. It is used to measure the quality and credibility of the analysis results.

[0088] The pre-defined synthesis strategy refers to the rule system used to integrate the analysis results, including core elements such as the comprehensive confidence calculation method, result ranking rules, and the number of candidate results to be selected.

[0089] N refers to the preset threshold for the number of candidate results to be filtered. It is set according to the actual application scenario and is used to control the size of the candidate subtask list to ensure the efficiency of subsequent processing.

[0090] In one specific implementation, this module is applied to the day care scenario of a community elderly care center, adapting to the complex needs of elderly users who want to "measure their blood sugar and then book housekeeping to clean their room, and then video chat with their spouse."

[0091] Context-aware routing phase: The community elderly care service terminal receives the contextualized semantic representation "daytime - community care center - blood glucose monitoring needs - housekeeping service needs - video call needs - calming emotions"; the context-aware routing unit extracts the scene trigger elements "blood glucose measurement", "housekeeping", and "video chat" from this semantic representation; for the four semantic parsers in the parser library, the scene matching degree is calculated: the parser matching degree for the medical and health domain is 0.92, the parser matching degree for the daily care domain is 0.90, the parser matching degree for the emotional companionship domain is 0.88, and the parser matching degree for the safety warning domain is 0.25; the preset activation threshold for each parser is 0.3, so the identifiers of the first three parsers are added to the target parser trigger list.

[0092] Inter-parser collaborative communication phase: In the collaboration weight connection table maintained by the inter-parser collaborative communication unit, the collaboration weight of the medical health and daily care parser is 0.6, the collaboration weight of the daily care and emotional companionship parser is 0.5, and the collaboration weight of the medical health and emotional companionship parser is 0.4. The medical health domain parser first generates the parsing result "start blood glucose monitoring equipment", and after querying the collaboration weight connection table, sends a collaborative consultation message to the daily care and emotional companionship parser that are in the triggered state. The daily care domain parser generates the parsing result "dispatch domestic service personnel" and sends a collaborative consultation message to the medical health and emotional companionship parser. The emotional companionship domain parser generates the parsing result "connect family member video terminal" and sends a collaborative consultation message to the medical health and daily care parser.

[0093] Competitive Result Synthesis Phase: The competitive result synthesis unit consists of three triggered parsers, each maintaining a collaborative activity counter. The medical and health parser sends and receives two messages, resulting in a counter value of 4; the daily care parser sends and receives two messages, resulting in a counter value of 4; and the emotional companionship parser sends and receives two messages, resulting in a counter value of 4. The overall confidence score is calculated using a scenario matching weight of 0.7 and a collaborative activity weight of 0.3 (the collaborative activity counter value is normalized before being included in the calculation). The overall confidence scores are 0.926 for the medical and health parser, 0.906 for the daily care parser, and 0.886 for the emotional companionship parser. With a preset N value of 3, after sorting the results in descending order of overall confidence, three parser results are selected to generate a candidate sub-task list: ① Activate blood glucose monitoring equipment (medical and health); ② Dispatch home care service personnel for in-home cleaning (daily care); ③ Connect to family members' video terminals for video calls (emotional companionship).

[0094] III. Structured Program Arbitration Module The structured plan arbitration module includes a risk perception and transmission analysis unit, a risk avoidance guidance unit, a continuous constraint reasoning unit, an initial plan synthesis unit, and a plan encapsulation unit; The risk perception and transmission analysis unit, based on a pre-set domain knowledge graph, identifies the health risk entities and safety status entities associated with each candidate sub-task, and constructs an analysis model representing the relationship between tasks and entities. The risk avoidance guidance unit is used to deduce the execution consequences of different conflict solutions based on the analysis model when there are conflicting candidate subtasks, and select candidate subtasks according to the principle of risk minimization. The continuity constraint reasoning unit is used to automatically reason for candidate subtasks and add logical dependencies and temporal constraints based on the continuity guarantee rules related to elderly care services. The initial plan synthesis unit is used to synthesize candidate subtasks that have undergone conflict resolution and have been given constraints into an initial structured service plan, which defines the execution logic and timing of each task. The planning and encapsulation unit is used to receive the initial structured service plan, identify the key task nodes in it, dynamically generate one or more alternative execution paths for each key task node, and encapsulate the initial plan and all alternative paths together into the final output structured service plan.

[0095] Furthermore, the risk aversion-oriented unit is specifically configured to perform the following operations to deduce the consequences of execution: Based on the analysis model, a virtual task execution subgraph is constructed for each conflict solution; In the pre-defined domain knowledge graph, forward propagation traversal is performed based on the causal relationship edges between health risk entities and safety status entities; The comprehensive risk quantification value of the scheme is calculated based on the predefined severity level weight of the terminal risk entity reached by the traversal path in the domain knowledge graph, and the preset occurrence probability of causal edges in the path. Based on the principle of minimizing risk according to the comprehensive risk quantification value, the solution with the lowest comprehensive risk quantification value is selected as the candidate sub-task.

[0096] Furthermore, the continuity guarantee rules in the continuity constraint reasoning unit are a special type of relation edge in the domain knowledge graph, including: Temporal dependency edges are used to represent the necessary order between two tasks, and their attributes include minimum time interval and maximum time interval. State premise edges are used to represent the user's health status or environmental conditions necessary for performing a certain task. A resource-exclusive edge is used to indicate that two tasks cannot be executed concurrently because they are competing for the same physical resource or caregiver. The continuity constraint reasoning unit adds logical dependencies and temporal constraints to candidate subtasks by querying and interpreting temporal dependency edges, state premise edges, and resource mutual exclusion edges.

[0097] It should be noted that health risk entities refer to health and safety-related objects associated with the execution of subtasks, including entities that may affect the user's health, such as high blood pressure, fall risk, and postoperative rehabilitation risk.

[0098] A security status entity refers to an environment and state-related object associated with the execution of a subtask, including entities that may affect service security, such as being alone, nighttime scenarios, and device security status.

[0099] An analytical model is a mathematical model that characterizes the relationship between candidate subtasks and health risk entities and safety status entities, and is used to support risk transmission simulation and conflict solution evaluation.

[0100] Candidate subtask conflict refers to the execution contradictions between multiple candidate subtasks, including resource competition conflicts, timing conflicts, state constraint conflicts, etc., which may lead to service execution failure or security risks.

[0101] Conflict resolution refers to solutions proposed to resolve conflicts in subtasks, including specific strategies such as adjusting task timing, reallocating resources, and dynamically adjusting priorities.

[0102] Execution consequence simulation refers to simulating the execution process of different conflict solutions based on analytical models, and predicting the potential health and safety risks and service effects.

[0103] The risk minimization principle refers to the decision criterion of selecting the conflict solution with the lowest comprehensive risk quantification value as the candidate sub-task. The core objective is to reduce the health and safety risks of service execution.

[0104] A virtual task execution subgraph is a visual task execution logic diagram built for each conflict solution, which clearly presents the task execution sequence, related entities, and risk transmission paths.

[0105] Causal association edges refer to the relationship links connecting health risk entities and safety status entities in a domain knowledge graph, representing the causal influence relationship between entities.

[0106] Forward propagation traversal refers to the process of starting from the initial entity corresponding to the conflict solution and traversing subsequent related entities layer by layer along the causal relationship edges, which is used to discover potential risk transmission paths.

[0107] The final risk entity refers to the last risk-related entity reached during forward propagation traversal, and its severity level directly affects the calculation of the comprehensive risk quantification value.

[0108] The severity level weight refers to the pre-defined risk level quantification value for each risk entity in the domain knowledge graph, ranging from level 1 to level 5, with level 5 being the highest risk level.

[0109] The preset occurrence probability refers to the risk transmission probability preset for each causal relationship edge in the domain knowledge graph. It ranges from 0 to 1 and is used to quantify the possibility of risk transmission.

[0110] The comprehensive risk quantification value refers to the risk quantification result of conflict solution obtained by weighted summation calculation. It comprehensively reflects the potential risk level of the solution, and the lower the value, the lower the risk.

[0111] Continuity assurance rules refer to a system of rules that ensure the continuous and orderly execution of elderly care services. They are predefined as special relation edges in a domain knowledge graph, including temporal dependency edges, state premise edges, and resource mutual exclusion edges.

[0112] Logical dependency refers to the logical association between subtasks based on rules, including constraints such as preconditions and execution order, to ensure the rationality of task execution.

[0113] Timing constraints refer to time-related restrictions on the execution of subtasks, including requirements on the order of execution and time intervals, to ensure the orderly execution of tasks.

[0114] Temporal dependency edges are relational edges that represent the order in which two tasks are executed. Their attributes include minimum time interval and maximum time interval, which explicitly define the time window constraints for task execution.

[0115] State premise edges refer to relational edges that represent the necessary state conditions for task execution, specifying the user's health state or environment state required to execute a certain task.

[0116] A resource-exclusive edge is a relationship edge that represents the resource competition relationship between two tasks. It clarifies the combination of tasks that cannot be executed concurrently, thus avoiding resource conflicts.

[0117] The initial structured service plan refers to the basic service plan that does not include alternative paths, and defines the execution order, dependencies and time nodes of each subtask.

[0118] Execution logic refers to the rules governing how tasks are executed, including specific modes such as serial execution and parallel execution.

[0119] Timing refers to the time arrangement for task execution, including time parameters such as start time, end time, and execution duration.

[0120] Critical mission nodes refer to core tasks that are essential to achieving service objectives, including emergency medical tasks and high-priority daily care tasks.

[0121] Alternative execution paths refer to pre-set alternative execution schemes for critical task nodes, used to cope with resource unavailability or state changes in unexpected scenarios.

[0122] The final output structured service plan refers to a complete execution plan that includes the initial plan and alternative paths, providing a comprehensive and reliable basis for resource scheduling.

[0123] In one specific implementation, this module is applied to a multi-task concurrent scenario for home-based elderly care. The candidate sub-task list includes three tasks: "in-home care", "rehabilitation training" and "household cleaning". There are resource competition (for the same caregiver) and timing conflicts.

[0124] Risk perception and transmission analysis phase: The risk perception and transmission analysis unit identifies the associated entities of each subtask from the domain knowledge graph. "Home care" is associated with the health risk entity "postoperative infection" and the safety status entity "being alone at home"; "rehabilitation training" is associated with the health risk entity "muscle strain" and the safety status entity "normal blood pressure"; "household cleaning" is associated with the safety status entity "environmental safety"; and an analysis model is built based on the entity association relationship to clarify the mapping relationship between tasks and risk entities.

[0125] Risk Aversion-Oriented Phase: It was identified that "home care" and "rehabilitation training" were competing for the same caregiver (resource conflict), and both were scheduled to be performed at 9:00 AM (time conflict), resulting in two conflict resolution solutions: Solution 1, "Home care first, rehabilitation training at 10:00 AM," and Solution 2, "Rehabilitation training first, home care at 10:00 AM." Virtual task execution subgraphs were constructed for each solution, and the process was traversed forward along causal edges. In Solution 1, the terminal risk entity "postoperative infection" had a severity level weight of 3, a preset occurrence probability of 0.2 for the causal edge, and a comprehensive risk quantification value of 0.6. In Solution 2, the terminal risk entity "muscle strain" had a severity level weight of 2, a preset occurrence probability of 0.3 for the causal edge, and a comprehensive risk quantification value of 0.6. Because the quantification values ​​were the same, and considering the user's postoperative rehabilitation priority, Solution 1 was selected as the final conflict resolution solution.

[0126] Continuity-constrained reasoning stage: The continuity-constrained reasoning unit queries the continuity guarantee rules in the domain knowledge graph, adds temporal dependency edges for "home care" and "rehabilitation training", sets a minimum time interval of 60 minutes and a maximum time interval of 90 minutes; adds a state prerequisite edge for "rehabilitation training", specifying the prerequisite condition as "blood pressure ≤ 140 / 90 mmHg"; adds resource mutual exclusion edges for "home care" and "household cleaning", specifying that they cannot be executed concurrently; and adds logical dependencies and temporal constraints for the three sub-tasks based on the rules.

[0127] Initial plan synthesis phase: The initial plan synthesis units are synthesized into an initial structured service plan in the order of "first home care (9:00-9:30), then rehabilitation training (10:00-10:30), and finally housekeeping (14:00-15:00)," clarifying the execution logic, time nodes and constraints of each task.

[0128] In the planning and encapsulation phase: the planning and encapsulation unit identifies "in-home nursing care" and "rehabilitation training" as key task nodes, generates alternative paths for "in-home nursing care" (if the nursing staff is temporarily unable to come to the site, a backup nursing staff will be dispatched), and generates alternative paths for "rehabilitation training" (if the user's blood pressure is not up to standard, remote rehabilitation guidance will be provided); the initial plan and alternative paths are encapsulated together to output the final structured service plan.

[0129] IV. Resource Scheduling and Execution Module The resource scheduling and execution module includes a resource profiling and reputation database unit, a collaborative commitment scheduling unit, a global constraint coordination and arbitration unit, and an instruction sequence generation and distribution unit; The Resource Profile and Reputation Database unit is used to maintain the dynamic service resource database. Each resource node in the dynamic service resource database is associated with a dynamically updated multi-dimensional resource profile, which includes static capabilities, real-time status, historical performance quality scores, and user feedback tags. The collaborative commitment scheduling unit is used to publish subtasks in the structured service plan to a subset of resources in the resource pool in the form of task proposals, and to receive autonomous commitment feedback from each resource based on its current multidimensional resource profile. The autonomous commitment feedback includes the expected execution time and confidence level. The global constraint coordination and arbitration unit is used to perform a global consistency check on all collected commitment feedback, arbitrate conflicting commitments, and generate a final scheduling scheme. The scheduling scheme defines the binding relationship between each subtask and specific resources and execution time windows. The instruction sequence generation and distribution unit is used to convert and encapsulate the task and resource binding relationship in the scheduling scheme into an executable instruction sequence that adapts to the interface protocol of different resource types, and drive the instruction sequence to distribute to the corresponding service resources. The instruction sequence includes a set of action instructions issued to the service robot, task notifications and navigation information pushed to the caregiver's mobile terminal, and control commands sent to smart home devices.

[0130] It should be noted that a resource node refers to the basic unit representing a single service resource in the dynamic service resource library. Each node corresponds to a specific service provider (such as a caregiver or a service robot).

[0131] A multidimensional resource profile is a set of digital features that comprehensively characterizes resource nodes. It includes four dimensions: static capabilities, real-time status, historical performance quality scores, and user feedback tags, enabling a three-dimensional evaluation of resources.

[0132] Static capabilities refer to the inherent, long-term, and unchanging characteristics of resources, including the professional qualifications of nursing staff, the functional scope of service robots, and the technical parameters of equipment. They serve as the basis for resource adaptation tasks.

[0133] Real-time status refers to the current operational status and availability of resources, including the location and busyness of nursing staff, the online and operational status of equipment, and the responsiveness of the service platform. It serves as the basis for dynamic decision-making in resource scheduling.

[0134] Historical performance quality score is a quantitative evaluation value calculated based on the past task completion status of resources. It ranges from 0 to 5 points. The higher the score, the better the historical service quality of the resource, providing a reference for resource selection.

[0135] User feedback tags are qualitative evaluation tags added to resources based on actual user experience, including dimensions such as service attitude, response speed, and adaptability, supplementing the deficiencies of quantitative scoring.

[0136] A task proposal refers to standardized resource recruitment information generated based on sub-task requirements. It includes core elements such as task type, execution requirements, time window, and constraints, and clarifies the specific requirements for resource execution.

[0137] A resource subset refers to a set of candidate resources selected from the dynamic service resource library that match the subtask requirements, avoiding indiscriminately issuing proposals to all resources and improving scheduling efficiency.

[0138] Autonomous commitment feedback refers to the task execution commitment given by candidate resources based on their own multi-dimensional resource profile, including the expected execution time and confidence level, reflecting the resource's assessment of its ability and willingness to complete the task.

[0139] The estimated execution time refers to the time period required for resources to complete a task, calculated from the time the resource receives the instruction, clearly defining the expected execution time of the task.

[0140] Confidence level refers to the quantitative value of a resource’s confidence in its ability to complete tasks on time and with quality. It ranges from 0 to 1, and the higher the score, the stronger the reliability of the resource in completing the task.

[0141] Global consistency check refers to a comprehensive verification of the autonomous commitment feedback of all candidate resources to check for problems such as time window conflicts, resource capacity mismatch, and constraint violation, so as to ensure the effectiveness of the feedback.

[0142] Conflicting commitments refer to contradictions between commitments from different resources or between the feedback and task requirements, including time conflicts, resource allocation conflicts, and capability mismatch conflicts, which need to be resolved through arbitration.

[0143] Conflict arbitration refers to the process of selecting the optimal resources and implementation plan based on preset rules when there are existing conflicting commitments. Arbitration principles include confidence priority, proximity, highest historical score, and user preference priority.

[0144] The final scheduling scheme refers to the standardized resource allocation scheme formed after conflict arbitration, which clarifies the binding relationship between each subtask and specific resources and execution time windows, providing a clear basis for instruction generation.

[0145] An execution time window refers to the specific time interval allocated to a subtask, which clarifies the time boundaries between the start and end of the task and ensures the orderly execution of the task.

[0146] Binding relationship refers to the fixed association between subtasks, service resources, and execution time windows, which determines that a certain task will be completed by a specific resource within a specified time.

[0147] An executable instruction sequence is a standardized set of instructions organized according to the requirements of a resource interface protocol. Each sequence corresponds to the complete execution requirements of a subtask, ensuring that the resource can accurately understand and execute it.

[0148] Adapting to different resource types and interface protocols means adjusting the format and transmission method of instructions according to the hardware characteristics and communication standards of the resources to ensure the compatibility of instructions on different resources.

[0149] Service robots are intelligent devices that have the ability to move and perform operations autonomously, and are used to perform physical operation tasks (such as food delivery, cleaning, and assisting in movement).

[0150] A motion instruction set refers to the execution instructions designed for service robots, which include parameters such as motion paths, operation steps, and safety thresholds, and adopt robot-compatible communication protocols.

[0151] A caregiver's mobile terminal refers to a smart terminal (such as a smartphone or tablet) carried by caregivers to receive task notifications, navigation information, and feedback on execution status.

[0152] Task notifications and navigation information refer to standardized information sent to caregivers' mobile terminals, including task details, user location, time requirements, urgency level, etc., to help caregivers perform tasks efficiently.

[0153] Smart home devices refer to home devices with intelligent control functions (such as smart blood pressure monitors, smart lights, smart door locks, etc.) used to assist in tasks such as daily care and health monitoring.

[0154] Control commands are operation instructions designed for smart home devices. They include information such as device identification, control parameters, and execution duration, and use device-compatible communication protocols.

[0155] In one specific implementation, this module is applied to a community-based home care integrated elderly care scenario. The structured service plan includes three sub-tasks: home care (wound dressing) from 9:00 to 9:30 am, rehabilitation training (physical activity guidance) from 10:00 to 10:30 am, and housekeeping (indoor dust removal) from 2:00 to 3:00 pm.

[0156] Resource profile maintenance phase: Resource profiles and reputation database units update resource information in real time. Caregiver A's multi-dimensional resource profile includes static capabilities (qualified to change wound dressings), real-time status (available in the community, 1.2 km from the user), historical performance quality score of 4.8, and user feedback tags (attentive, responsive); Caregiver B's static capabilities include rehabilitation guidance qualifications, real-time status (available in the community, 0.8 km from the user), historical score of 4.7, and user feedback tags (professional, patient); Cleaning Robot No. 1's static capabilities include indoor dust removal, real-time status (available online), historical score of 4.6, and user feedback tags (thorough cleaning); Smart Blood Pressure Monitor's static capabilities include blood pressure measurement, real-time status (available online), and historical score of 4.9; All resource information is synchronized to the dynamic service resource database every 5 minutes.

[0157] Collaborative commitment scheduling phase: The collaborative commitment scheduling unit will target and publish the task proposals of the three sub-tasks to the corresponding resource subsets. The home care task proposal will be sent to caregivers A and B, the rehabilitation training task proposal will be sent to caregivers B and C, and the house cleaning task proposal will be sent to cleaning robots 1 and 2. Receive autonomous commitment feedback: Caregiver A (home care estimated execution time 30 minutes, confidence level 0.98), Caregiver B (rehabilitation training estimated execution time 30 minutes, confidence level 0.97), Cleaning robot 1 (house cleaning estimated execution time 60 minutes, confidence level 0.95), Smart blood pressure monitor (measuring blood pressure in conjunction with rehabilitation training, estimated execution time 5 minutes, confidence level 0.99).

[0158] Global Constraint Coordination and Arbitration Phase: The global constraint coordination and arbitration unit performs consistency checks and finds no time conflicts or resource competition; based on the principles of confidence priority and nearest proximity, a final scheduling plan is generated: the home care task is performed by caregiver A from 9:00 to 9:30, the rehabilitation training task is performed by caregiver B from 10:00 to 10:30 (with simultaneous use of a smart blood pressure monitor to measure blood pressure), and the housekeeping task is performed by cleaning robot No. 1 from 14:00 to 15:00; the plan clarifies the binding relationship between each sub-task and resources and time windows.

[0159] Instruction generation and distribution phase: The instruction sequence generation and distribution unit generates instruction sequences adapted to different resources, pushes JSON format task notifications (including user address and wound dressing precautions) and navigation information (based on the optimal route generated by Gaode Map) to caregiver A's mobile terminal, pushes the same format notifications (including rehabilitation training action guidance) to caregiver B's mobile terminal, issues ROS-compatible protocol action instruction sets (including cleaning path and dust removal intensity parameters) to cleaning robot No. 1, and sends MQTT protocol control commands (including measurement time and data upload requirements) to the smart blood pressure monitor; distributes all instructions through the 5G communication module and establishes a real-time feedback channel to synchronously receive resource execution status (such as caregiver arrival confirmation and robot cleaning progress).

[0160] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A language processing system for elderly care service terminals based on AI speech recognition, characterized in that, include: The semantic understanding and context fusion module is used to fuse recognized text from the voice interface and real-time data from terminal sensors to generate contextualized semantic representations that include user intent, emotional state, and environmental context. The multi-domain parallel parsing module contains multiple semantic parsers corresponding to different elderly care service domains. It is used to call one or more semantic parsers to perform parallel collaborative parsing of contextualized semantic representations and combine the outputs of each parser to generate a list of candidate subtasks. The structured plan arbitration module is used to resolve conflicts based on confidence comparison for all received candidate subtasks, and to reason about the logical and temporal relationships between tasks according to the preset domain knowledge graph, generating a structured service plan consisting of multiple dependent subtasks. The resource scheduling and execution module is used to maintain a dynamic service resource library and, based on the structured service plan and the dynamic status information of the service resources obtained from the resource library, optimize and schedule the execution of each sub-task in the plan, and generate a sequence of instructions to drive the execution of specific service resources.

2. The language processing system for elderly care service terminals based on AI speech recognition according to claim 1, characterized in that, The semantic understanding and context fusion module includes: The multimodal feature extraction unit is used to parse the speech recognition text to extract semantic feature vectors representing the user's intent and emotional state, and to process the terminal sensor data stream to extract context feature vectors representing the physical environment state. A cross-modal reliability assessment unit is used to calculate confidence scores for semantic feature vectors and quality scores for context feature vectors. The feature fusion unit is used to dynamically adjust the weights of the semantic feature vector and the context feature vector in the fusion process based on the confidence score and the quality score, and to perform weighted fusion to output a preliminary fusion representation. The contextualized semantic synthesis unit is used to synthesize the preliminary fusion representation with a scenario prior vector generated by encoding the current time, user activity stage, and device geographic location, and output a contextualized semantic representation.

3. The language processing system for elderly care service terminals based on AI speech recognition according to claim 1, characterized in that, The multi-domain parallel parsing module includes: The parser library unit is used to store and manage a collection of semantic parsers configured for the domains of healthcare, daily care, safety alerts, and emotional companionship. The context-aware routing unit is used to receive contextualized semantic representations, extract scene triggering elements from them, and calculate the scene matching degree between the scene triggering elements and each parser to generate a list of target parser triggers. An inter-parser collaborative communication unit is used to manage and maintain an inter-parser communication link network, which supports the exchange of structured messages between triggered parsers. The competitive result synthesis unit is used to aggregate and evaluate the parsing results of all triggered parsers, and output a list of candidate subtasks according to a preset synthesis strategy.

4. The language processing system for elderly care service terminals based on AI speech recognition according to claim 3, characterized in that, The context-aware routing unit is configured to perform the following operations: Extract one or more scene triggering elements from the received contextualized semantic representation; For each semantic parser in the parser library unit, a scene matching degree value is calculated based on the scene triggering element and the domain focus label of the parser. The scene matching score of each parser is compared with the preset activation threshold of that parser. Add the identifiers of all semantic parsers whose scene matching scores are greater than or equal to their own activation thresholds to the target parser trigger list.

5. The language processing system for elderly care service terminals based on AI speech recognition according to claim 3, characterized in that, The inter-parser cooperative communication unit is configured to perform the following operations: Maintain a predefined collaboration weight join table, which records the collaboration weight values ​​between any two semantic parsers; When any triggered semantic parser generates a parsing result, the collaboration weight connection table is queried to obtain a list of associated parsers with a collaboration weight greater than zero. Filter from the list of associated resolvers to find other resolvers that are currently in a triggered state; Send a collaborative consultation message containing the parsing results to all other selected parsers.

6. The language processing system for elderly care service terminals based on AI speech recognition according to claim 3, characterized in that, The competitive result synthesis unit is configured to perform the following operations: Maintain a cooperative activity counter for each triggered parser to record the total number of cooperative consultation messages sent and received by the parser during cooperative communication. For each parsing result, the scene matching degree of its source parser and the value of its collaborative activity counter are weighted and summed to calculate the overall confidence score of the parsing result; All analysis results are sorted in descending order based on the overall confidence score, and the top N results are selected to generate a candidate subtask list.

7. The language processing system for elderly care service terminals based on AI speech recognition according to claim 1, characterized in that, The structured program arbitration module includes: The risk perception and transmission analysis unit, based on a pre-set domain knowledge graph, identifies the health risk entities and safety status entities associated with each candidate sub-task, and constructs an analysis model representing the relationship between tasks and entities. The risk avoidance guidance unit is used to deduce the execution consequences of different conflict solutions based on the analysis model when there are conflicting candidate subtasks, and select candidate subtasks according to the principle of risk minimization. The continuity constraint reasoning unit is used to automatically reason for candidate subtasks and add logical dependencies and temporal constraints based on the continuity guarantee rules related to elderly care services. The initial plan synthesis unit is used to synthesize candidate subtasks that have undergone conflict resolution and have been given constraints into an initial structured service plan, which defines the execution logic and timing of each task. The planning and encapsulation unit is used to receive the initial structured service plan, identify the key task nodes in it, dynamically generate one or more alternative execution paths for each key task node, and encapsulate the initial plan and all alternative paths together into the final output structured service plan.

8. The language processing system for elderly care service terminals based on AI speech recognition according to claim 7, characterized in that, The risk avoidance guidance unit is specifically configured to perform the following operations to deduce the consequences of execution: Based on the analysis model, a virtual task execution subgraph is constructed for each conflict solution; In the pre-defined domain knowledge graph, forward propagation traversal is performed based on the causal relationship edges between health risk entities and safety status entities; The comprehensive risk quantification value of the scheme is calculated based on the predefined severity level weight of the terminal risk entity reached by the traversal path in the domain knowledge graph, and the preset occurrence probability of the causal edge in the path. Based on the principle of minimizing risk according to the comprehensive risk quantification value, the solution with the lowest comprehensive risk quantification value is selected as the candidate sub-task.

9. The language processing system for elderly care service terminals based on AI speech recognition according to claim 7, characterized in that, The continuity guarantee rules in the continuity constraint reasoning unit are a special type of relation edge in the domain knowledge graph, including: Temporal dependency edges are used to represent the necessary order between two tasks, and their attributes include minimum time interval and maximum time interval. State premise edges are used to represent the user's health status or environmental conditions necessary for performing a certain task. A resource-exclusive edge is used to indicate that two tasks cannot be executed concurrently because they are competing for the same physical resource or caregiver. The continuity constraint reasoning unit adds logical dependencies and temporal constraints to candidate subtasks by querying and interpreting temporal dependency edges, state premise edges, and resource mutual exclusion edges.

10. The language processing system for elderly care service terminals based on AI speech recognition according to claim 1, characterized in that, The resource scheduling and execution module includes: The resource profiling and reputation database unit is used to maintain a dynamic service resource database. Each resource node in the dynamic service resource database is associated with a dynamically updated multi-dimensional resource profile, which includes static capabilities, real-time status, historical performance quality scores, and user feedback tags. The collaborative commitment scheduling unit is used to publish subtasks in the structured service plan to a subset of resources in the resource pool in the form of task proposals, and to receive autonomous commitment feedback from each resource based on its current multidimensional resource profile. The autonomous commitment feedback includes the expected execution time and confidence level. The global constraint coordination and arbitration unit is used to perform a global consistency check on all collected commitment feedback, arbitrate conflicting commitments, and generate a final scheduling scheme. The scheduling scheme defines the binding relationship between each subtask and specific resources and execution time windows. The instruction sequence generation and distribution unit is used to convert and encapsulate the task and resource binding relationship in the scheduling scheme into an executable instruction sequence that adapts to the interface protocol of different resource types, and drive the instruction sequence to distribute to the corresponding service resources. The instruction sequence includes a set of action instructions issued to the service robot, task notifications and navigation information pushed to the caregiver's mobile terminal, and control commands sent to smart home devices.

Images