Touching method and device based on dynamic calendar knowledge graph, equipment and medium

Abstract

The application relates to the technical field of semantic parsing, and discloses a touch method and device based on a dynamic calendar knowledge graph, equipment and a medium, which comprises the following steps: performing text processing on an initial natural language sequence to generate a standard text sequence; performing intent recognition and emotion recognition based on the standard text sequence to generate a user state vector; combining benchmark time data and a dynamic calendar knowledge graph to analyze a fuzzy time entity and generate a standard time window; generating target wake-up time and touch strategy data based on the user state vector, the standard time window and task scheduling data; generating and storing to-be-woken-up task data, and performing touch processing when real-time time is not earlier than the target wake-up time. The application can be applied to business scenarios such as financial technology and medical health, target wake-up time and touch strategies are generated by identifying user states and combining dynamic calendar knowledge graph to analyze fuzzy time entities, the accuracy of fuzzy time analysis and task scheduling is improved, and touch deviation is reduced.

Description

Technical Field

[0001] This invention relates to the field of semantic parsing technology, and in particular to a method, apparatus, device, and medium for reaching others based on a dynamic calendar knowledge graph. Background Technology

[0002] Existing automatic tracking technologies struggle to convert ambiguous time expressions in natural language into explicit time ranges that can be directly used for task scheduling. For expressions that rely on the current date and context, such as "after a holiday," "in a few days," or "after finishing work," existing technologies typically only recognize surface-level time words and cannot further determine the specific trigger time. Furthermore, existing technologies lack an effective correlation between the recognition results of user intent and emotional state and subsequent task scheduling, making it difficult to adjust the reach time and method according to changes in user status.

[0003] In the fintech business, the auto insurance claims process typically involves a series of steps, including reporting and follow-up, damage assessment and confirmation, repair at the workshop, material replenishment, claims processing, and case closure. When customers use vague time expressions such as "I'll go after the holiday," "I'll check again next week," or "I'm too busy recently," existing claims robots struggle to establish clear follow-up time windows and adjust subsequent expedited repair tasks based on the customer's intention to postpone or their emotional state. This results in claims follow-up still relying on fixed time intervals, easily leading to a mismatch between the timing of contact and the customer's actual situation.

[0004] In the healthcare field, scheduling follow-up appointments, health visits, rehabilitation reminders, and check-up visits also require arranging subsequent tasks based on the patient's natural language responses. When patients indicate that they will contact the patient after a period of time, after they have recovered, or when they will be available next time, existing follow-up technologies struggle to determine an appropriate time to do so. They also find it difficult to adjust task arrangements based on the patient's intention to postpone or their anxiety, which can easily lead to follow-up tasks being triggered prematurely, delayed, or repeatedly. Summary of the Invention

[0005] The main objective of this invention is to provide a method, apparatus, device, and storage medium for reaching users based on a dynamic calendar knowledge graph, aiming to solve the technical problem that existing technologies struggle to accurately convert context-dependent fuzzy time representations into executable time windows and thereby dynamically schedule subsequent tasks.

[0006] To achieve the above objectives, this invention provides an outreach method based on a dynamic calendar knowledge graph, comprising: Obtain an initial natural language sequence, perform text processing on the initial natural language sequence, and generate a standard text sequence; Intent recognition and emotion recognition are performed based on the standard text sequence to generate a user state vector; Acquire baseline time data and dynamic calendar knowledge graph, extract fuzzy time entities from the standard text sequence, and perform time parsing on the fuzzy time entities based on the baseline time data and the dynamic calendar knowledge graph to generate a standard time window; Obtain task scheduling data, perform dynamic scheduling processing based on the user state vector, the standard time window, and the task scheduling data, and generate target wake-up time and reach strategy data; Based on the target wake-up time, the reach strategy data, and the task scheduling data, wake-up task data is generated and written into the task storage area; Acquire real-time date and time data. If the time represented by the real-time date and time data is not earlier than the target wake-up time, retrieve the task data to be woken up from the task storage area and perform outreach processing according to the outreach strategy data contained in the task data to be woken up.

[0007] Furthermore, to achieve the above objectives, the present invention provides an access device based on a dynamic calendar knowledge graph, comprising: The text preprocessing module is used to obtain an initial natural language sequence, perform text processing on the initial natural language sequence, and generate a standard text sequence. The user state recognition module is used to perform intent recognition and emotion recognition based on the standard text sequence and generate a user state vector. The fuzzy time parsing module is used to acquire reference time data and dynamic calendar knowledge graph, extract fuzzy time entities from the standard text sequence, and perform time parsing on the fuzzy time entities based on the reference time data and the dynamic calendar knowledge graph to generate a standard time window; The dynamic scheduling decision module is used to acquire task scheduling data, perform dynamic scheduling processing based on the user state vector, the standard time window and the task scheduling data, and generate target wake-up time and reach strategy data. The wake-up task storage module is used to generate wake-up task data based on the target wake-up time, the reach strategy data and the task scheduling data, and write the wake-up task data into the task storage area; The timed outreach module is used to acquire real-time date and time data. When the time represented by the real-time date and time data is not earlier than the target wake-up time, it acquires the task data to be woken up from the task storage area and performs outreach processing according to the outreach strategy data contained in the task data to be woken up.

[0008] Furthermore, to achieve the above objectives, the present invention also provides a computer device, the computer device including a memory, a processor, and a reach program based on a dynamic calendar knowledge graph stored in the memory and executable on the processor, wherein when the reach program based on the dynamic calendar knowledge graph is executed by the processor, it implements the steps of the reach method based on the dynamic calendar knowledge graph as described above.

[0009] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium storing a reach program based on a dynamic calendar knowledge graph, wherein the reach program based on the dynamic calendar knowledge graph, when executed by a processor, implements the steps of the reach method based on the dynamic calendar knowledge graph as described above.

[0010] Beneficial Effects: This invention relates to the field of semantic parsing technology and discloses a method, apparatus, device, and medium for reaching users based on a dynamic calendar knowledge graph. The method includes: processing an initial natural language sequence to generate a standard text sequence; performing intent and emotion recognition based on the standard text sequence to generate a user state vector; parsing fuzzy time entities using benchmark time data and a dynamic calendar knowledge graph to generate a standard time window; generating target wake-up time and reach strategy data based on the user state vector, standard time window, and task scheduling data; generating and storing wake-up task data; and performing reach processing when the real-time time is no earlier than the target wake-up time. This invention can be applied to business scenarios such as fintech and healthcare. By standardizing natural language sequences, it identifies user intent and emotion, and combines benchmark time data with a dynamic calendar knowledge graph to parsing fuzzy time entities. Then, based on user state, standard time window, and task scheduling data, it determines the target wake-up time and reach strategy, thereby improving the accuracy of fuzzy time parsing and dynamic scheduling, and reducing reach timing deviations. Attached Figure Description

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a schematic diagram of an application environment for an outreach method based on a dynamic calendar knowledge graph, according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating an embodiment of the outreach method based on a dynamic calendar knowledge graph according to the present invention. Figure 3 for Figure 2 A detailed flowchart of step S10; Figure 4 for Figure 2 A detailed flowchart of step S20; Figure 5 for Figure 2 A detailed flowchart of step S30; Figure 6 for Figure 2 A detailed flowchart of step S40; Figure 7 for Figure 2 A detailed flowchart of step S50; Figure 8 for Figure 2 A detailed flowchart of step S60; Figure 9 This is a schematic diagram of the functional modules of a preferred embodiment of the reach device based on a dynamic calendar knowledge graph of the present invention. Figure 10 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention; Figure 11 This is another structural schematic diagram of a computer device according to one embodiment of the present invention. Detailed Implementation

[0012] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0013] The outreach method based on a dynamic calendar knowledge graph provided in this invention can be applied to, for example... Figure 1 In this application environment, the client communicates with the server via a network. The server can process the initial natural language sequence through the client to generate a standard text sequence; based on the standard text sequence, it performs intent recognition and emotion recognition to generate a user state vector; it combines baseline time data and a dynamic calendar knowledge graph to resolve fuzzy time entities and generate a standard time window; based on the user state vector, the standard time window, and task scheduling data, it generates target wake-up time and reach strategy data; it generates and stores wake-up task data, and performs reach processing when the real-time time is not earlier than the target wake-up time. This invention can be applied to business scenarios such as fintech and healthcare. By standardizing natural language sequences, it identifies user intent and emotion, and combines baseline time data and a dynamic calendar knowledge graph to resolve fuzzy time entities. Then, based on user state, standard time window, and task scheduling data, it determines the target wake-up time and reach strategy, thereby improving the accuracy of fuzzy time resolution and dynamic scheduling, and reducing reach timing deviations. The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will be described in detail below through specific embodiments.

[0014] Please see Figure 2 , Figure 2This is a flowchart illustrating an embodiment of the outreach method based on a dynamic calendar knowledge graph provided by the present invention. It should be noted that although a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than that shown here.

[0015] like Figure 2 As shown, the outreach method based on dynamic calendar knowledge graph proposed in this invention includes the following steps: S10, Obtain an initial natural language sequence, perform text processing on the initial natural language sequence, and generate a standard text sequence; In this embodiment, the interactive interface, message queue, or speech-to-text component receives user input and organizes it into an initial natural language sequence according to the order of character appearance, message sending order, or speech-to-text order. The initial natural language sequence preserves word order and intra-sentence semantic relationships and may include a session identifier, input time, and source channel to prevent incorrect merging of different interactive content. The input content may contain character encoding differences, whitespace, invalid symbols, missing punctuation, word form differences, and transcription errors; these factors can cause the same semantic meaning to appear in different text forms.

[0016] Text processing involves organizing the character forms, word forms, and text boundaries in the initial natural language sequence. Character form organization unifies character encoding, full-width and half-width forms, number forms, and punctuation forms, and removes control characters and consecutive whitespace characters that do not carry semantic meaning. Word form organization corrects character combinations that deviate significantly from the context based on the relationship between the preceding and following text, and converts expressions with the same meaning but different forms into a unified expression. Text boundary organization determines the range of text units based on punctuation positions, input intervals, and semantic completeness, while maintaining the original arrangement order of the input content and preventing different semantic fragments from mixing.

[0017] A standard text sequence is formed by combining processed text units according to their original semantic order, and can be stored as a character sequence, word sequence, or sentence sequence. Each text unit has clear boundaries and a uniform character format, and retains the positional correspondence with the initial natural language sequence. When corrections are made, the position before correction, the position after correction, and the correction type can be recorded, making the standard text sequence traceable and reducing text deviations caused by format differences and input noise.

[0018] When the input content is short and the expression format is relatively fixed, a character mapping table, a lexical conversion table, and a finite-state converter are used to tidy up the text. The character mapping table converts different encoding forms into unified characters, the lexical conversion table merges expressions with consistent meanings, and the finite-state converter determines the text boundaries according to punctuation, whitespace, and character categories. For multiple consecutively input messages, they can be combined based on session identifiers and time intervals, and boundary markers can be written between messages to generate a standard text sequence.

[0019] When the input content is derived from speech-to-text and contains many homophones, weighted edit distance and local language scoring are used to correct words. For character segments with confidence scores below a set value, candidate words with similar pronunciations or shapes are selected from the domain vocabulary, written into their corresponding positions, and the target words are determined based on the combined scores of the preceding and following words. When processing car insurance repair communication content, the matching weight of words such as "repair," "workshop entry," and "damage assessment" can be increased; when processing health follow-up content, the matching weight of words such as "re-examination," "medication," and "rehabilitation" can be increased. After character correction is completed, the text units are recombined according to the correction positions, and a standard text sequence is output.

[0020] In the fintech business, when a car insurance claim customer inputs that they don't have time to process the vehicle repairs in the next two days and will do so after the holiday, the interface generates an initial natural language sequence based on the input order. The characters are then organized, the text boundaries are supplemented, and the word forms are standardized to generate a standard text sequence indicating that the vehicle repairs cannot be processed in the next two days and will be arranged after the holiday. The system also retains the vehicle repair details and the postponement.

[0021] In the healthcare business, when a health management platform receives a message that the patient has finished taking medication and will schedule a follow-up appointment, the platform identifies the semantic relationship between the end of medication use and the scheduled follow-up appointment, adds punctuation and standardizes word forms, and generates a standard text sequence for scheduling a follow-up appointment after the medication is finished, thus preserving the association between the medication status and the follow-up appointment.

[0022] This embodiment maintains the arrangement of natural language content, unifies the character form, word form, and text boundaries, and establishes the positional correspondence before and after processing. This reduces text inconsistencies caused by encoding differences, invalid characters, word form differences, and transcription deviations, resulting in a standard text sequence with a stable semantic order, clear text boundaries, and a unified data format.

[0023] S20, based on the standard text sequence, perform intent recognition and emotion recognition to generate a user state vector; In this embodiment, standard text sequences are converted into sequential encoded data at the character, word, or sub-word granularity and written into positional data, enabling the semantic differences of the same words in different contexts to be distinguished. Intent recognition extracts action tendency, time tendency, negation relation, and object relation based on the sequential encoded data, generating category data and intensity data to reflect interactive purposes such as continuing processing, postponing processing, refusing processing, or consultation confirmation. Emotion recognition extracts the combination relationships between emotion words, modifiers, negation modifiers, and degree modifiers from the sequential encoded data, generating category data and intensity data to reflect states such as calm, anxiety, resistance, or dissatisfaction. The scope of negation words is determined based on word position and syntactic association, and degree modifiers adjust the emotion intensity according to their distance from emotion words and modification direction, avoiding the identification of "not in a hurry" as an anxious state. Intent categories and emotion categories are converted into category codes, and intent intensity and emotion intensity are converted into a unified numerical range, then combined according to fixed feature positions to form a user state vector. Each position in the user state vector carries the intent category, intent intensity, emotion category, and emotion intensity respectively, enabling different natural language content to be converted into structurally consistent state data.

[0024] In scenarios where interactive text is short and the scope of business expression is relatively concentrated, a word feature extraction layer and a dual-output classification layer are used to complete the recognition. The word feature extraction layer extracts action words, negation words, time words, emotion words, and degree words from the standard text sequence and generates feature weights based on word positions. The intent classification layer outputs intent category and intent intensity based on action words, object words, and negation relationships, while the emotion classification layer outputs emotion category and emotion intensity based on emotion words, degree words, and the scope of negation. The training data consists of a standard text sequence, intent labels, emotion labels, and corresponding intensity markers. During training, intent classification loss and emotion classification loss are generated separately and then merged to update the model parameters according to the set weights.

[0025] In interactive text containing ellipsis, long-range modifiers, or multi-sentence context, an embedding layer, a positional encoding layer, a shared semantic encoding layer, an intent output layer, and a sentiment output layer are employed for recognition. The embedding layer receives word or sub-word identifiers, the positional encoding layer records their arrangement, the shared semantic encoding layer generates a semantic representation including contextual relationships, and the intent output layer and sentiment output layer are respectively connected to the output of the shared semantic encoding layer. During training, a batch input method is used, and the learning rate and the weights of the two-class losses are adjusted based on the category recognition deviations in the validation data. The corresponding output layer dimension is adjusted when the number of categories changes, while the feature positions of the user state vector remain fixed, ensuring a consistent structure for state data across different business environments.

[0026] In the fintech business, when a car insurance claims customer inputs "I'm very busy lately, please don't rush me to get it repaired," the intent is identified to determine the category of postponement and the strength of the delay. The emotion is identified based on the tone of voice and the degree of resistance to determine the category of resistance and the corresponding strength. These are combined to form a user state vector that reflects the intention to postpone and the state of resistance.

[0027] In the healthcare business, when a health management user inputs "I'm feeling anxious lately, but I still want to go for my follow-up appointment on time," the system identifies the intention to continue processing the appointment and the emotion to be anxious. It retains both the intention to process the appointment and the emotional state separately, thus avoiding misinterpreting the intention to follow up the appointment as a refusal or postponement due to anxiety-related words.

[0028] This embodiment extracts action tendencies, negation relationships, emotion words, and degree modification relationships from standard text sequences, which can form intent categories, intent intensity, emotion categories, and emotion intensity respectively. It also converts different types of data into user state vectors with fixed structures, reducing misjudgments of intent and deviations in emotion direction caused by single keywords, and improving the accuracy of structured expression of user interaction states.

[0029] S30, acquire reference time data and dynamic calendar knowledge graph, extract fuzzy time entities from the standard text sequence, and perform time parsing on the fuzzy time entities based on the reference time data and the dynamic calendar knowledge graph to generate a standard time window; In this embodiment, the reference time data consists of session reception time, task creation time, business area time zone, and system calibration time, used to determine the reference position for relative time expression. The dynamic calendar knowledge graph contains time information such as dates, weekdays, weekends, holidays, regional events, and time periods, and continuously updates the time relationships between them based on calendar adjustment data. Text fragments such as "after the holiday," "next week," "before the end of the month," and "after finishing this busy period" in the standard text sequence can determine the text boundaries of fuzzy time entities based on time words, range words, direction words, and adjacent action content. Time parsing uses the reference time data as a reference, matches fuzzy time entities with time information in the dynamic calendar knowledge graph, and determines the candidate date range along the time relationship. When multiple candidate ranges exist, the target range is determined based on the degree of text semantic matching, time sequence, and regional calendar status, generating a standard time window containing start date and time, end date and time, and time zone information.

[0030] In cases where time representation types are relatively concentrated, a time lexicon, a finite state converter, and a graph query are used to determine the standard time window. The time lexicon identifies relative dates, calendar events, and time period terms; the finite state converter determines the boundaries of fuzzy time entities; and the graph query filters the date range based on the baseline time and time relationships.

[0031] In situations where spoken expression varies considerably, an embedding layer, a positional encoding layer, a contextual encoding layer, and a sequence labeling layer are employed to identify ambiguous time entities. The training data includes standard text sequences, entity boundaries, and time type labels. Sequence labeling loss is used to update model parameters. The recognition results are semantically matched with a dynamic calendar knowledge graph, and a standard time window is generated based on the relative position of the candidate time range and the base time.

[0032] In the fintech business, when a car insurance claim customer replies before a holiday to have their car repaired after the holiday, the base time data provides the current date and business area time zone, and the dynamic calendar knowledge graph provides the start and end dates of the holiday and the working days after the holiday. Based on this, the period after the holiday is interpreted as the range of available dates after the holiday ends.

[0033] In the healthcare business, when a health management user replies on the weekend that they want to schedule a follow-up appointment for the following week in the morning, the baseline time data determines the current week, and the dynamic calendar knowledge graph provides information on the next week's date, workday, and morning time slot, thereby generating a standard time window corresponding to the morning time slot within the next work week.

[0034] This embodiment determines the relative time reference position through benchmark time data and supplements the information of holidays, weekdays, weekends and regional time periods by using a dynamic calendar knowledge graph. It can transform fuzzy time entities that depend on context and calendar events into standard time windows with clear start and end boundaries and time zone information, reducing the problem that fuzzy time expressions cannot form a definite time range.

[0035] S40, acquire task scheduling data, perform dynamic scheduling processing based on the user state vector, the standard time window and the task scheduling data, and generate target wake-up time and reach strategy data; In this embodiment, the task management interface reads data about the tasks to be scheduled. This data may include the available time range, urgency level, waiting status, resource status, and past interaction results, and uses a unified time format. User state vectors represent willingness to handle tasks, degree of delay, and emotional state, while a standard time window defines the start and end times and time zone range that can be scheduled.

[0036] Dynamic scheduling sets candidate time points within a standard time window and filters and sorts them based on user state vectors and task scheduling data. When a task is highly urgency, the candidate time is moved to the front of the window; when there is strong intention to postpone or resistance, the candidate time is moved to the back of the window; when resources are unavailable, the corresponding candidate time is deleted. After sorting, a target wake-up time that meets all conditions is determined, where wake-up indicates that the task to be scheduled has entered a processable state.

[0037] Outreach strategy data is used to describe outreach channels, content types, and outreach intervals. When there is a strong intention to postpone, the frequency of outreach can be reduced and low-interference channels can be selected; when there is a strong willingness to process, content types containing clear operational information can be generated; when there is strong resistance, urging content can be reduced.

[0038] When the number of scheduling conditions is limited, a candidate time scoring method is used to determine the target wake-up time. The standard time window is divided into candidate time points, and scores for time urgency, user acceptance, and resource availability are generated for each. These scores are then combined according to set weights, and the candidate time point with the best score is selected. Outreach strategy data can be generated through a state mapping table, which records the correspondence between different intent states, emotional states, and outreach channels, content types, and intervals.

[0039] When dealing with a variety of task types and available historical interaction data, a time-based ranking model and a strategy-based classification model are employed. The time-based ranking model receives user state vectors, candidate time features, and task scheduling data, and outputs the candidate time ranking results. The strategy-based classification model outputs the reach channel category, content type category, and interval range. Training data can consist of historical scheduling records, actual reach times, and interaction response results.

[0040] In the fintech business, auto insurance claim customers indicated that they were busy recently and hoped to arrange vehicle repairs after the holiday. Dynamic scheduling determines the target wake-up time based on the post-holiday time range, the customer's intention to postpone, and the status of repair resources, and generates low-frequency reach strategy data.

[0041] In the healthcare business, health management users want to schedule a follow-up examination after completing the current medication cycle. Dynamic scheduling determines the target wake-up time based on the expected medication end time, the available time slot for the follow-up examination, and the user's status, and generates outreach strategy data containing information on follow-up examination preparation.

[0042] This embodiment constrains the scheduling range through a standard time window, uses user state vectors to reflect the willingness and emotional state of the user, and filters candidate times based on task scheduling data. It can generate target wake-up time and reach strategy data that match the user's state and task conditions, reducing time deviations and reach mismatches caused by fixed-interval scheduling.

[0043] S50, generate wake-up task data based on the target wake-up time, the reach strategy data and the task scheduling data, and write the wake-up task data into the task storage area; In this embodiment, the target wake-up time uses a data format including date, time, and time zone as the time attribute of the task to be woken up. The reach strategy data records control information such as reach channel, content type, and reach interval. The task scheduling data comes from task management records, interaction records, or business status records, and may include task identifier, reach object, task topic, and current status. When generating the task data to be woken up, the target wake-up time, reach strategy data, and task scheduling data are combined according to a preset field structure to generate a unique task tag, while simultaneously verifying the time format, necessary fields, and data relationships. After successful verification, the task data to be woken up is converted into a structured record and written to the task storage area. The writing process can use transaction commit or atomic write methods to avoid only partially saving data for the same task; when the same task tag is detected, the old record can be overwritten or version information can be retained based on the task update time.

[0044] When the number of tasks is small and the field structure is stable, a relational data table is used to store the data of tasks to be woken up. The data table sets fields such as task tag, target wake-up time, reach strategy, reach object, task topic, and task status, using the task tag as a unique key and creating an index on the target wake-up time. Required fields are validated before writing, and transaction control is used to maintain record integrity during the writing process.

[0045] When dealing with a large number of tasks and varying data structures, key-value stores or document stores are used to store the data for tasks awaiting wake-up. The task tag is used as the storage key, and the remaining data is serialized into document objects. Storage partitions are then created based on the target wake-up time. Document objects are updated when task content changes, and storage partitions and time attributes are adjusted synchronously when the time changes.

[0046] In the fintech business, after a car insurance claim customer confirms the vehicle repair arrangements after the holiday, the confirmed follow-up time, SMS contact method, repair details and customer contact data are combined into a task data to be activated and written into the claims task storage area.

[0047] In the healthcare business, health management users are scheduled to receive follow-up reminders after the medication cycle ends. The follow-up reminder time, message delivery method, follow-up items, and user contact data are combined into wake-up task data and written into the health follow-up task storage area.

[0048] This embodiment combines the target wake-up time, reach strategy data, and task scheduling data into a unified wake-up task data structure, and saves it using integrity verification, unique task marking, and atomic writing methods. This reduces problems such as missing task data, duplicate writing, and inconsistencies between time attributes and task content, thereby improving the integrity and storage reliability of the wake-up task data.

[0049] S60, obtain real-time date and time data. If the time represented by the real-time date and time data is not earlier than the target wake-up time, obtain the task data to be woken up from the task storage area, and perform reach processing according to the reach strategy data contained in the task data to be woken up.

[0050] In this embodiment, the real-time date and time data is provided by the system clock, time synchronization service, or trusted time source, and includes the current date, current time, and time zone information. Before comparison, the real-time date and time data and the target wake-up time are converted to the same time format and time zone to avoid misjudgment due to time zone differences or different time precision. "Not earlier than the target wake-up time" includes the current time being equal to or later than the target wake-up time, used to determine that the task to be woken up has met the triggering conditions.

[0051] The task storage area stores the data of tasks to be woken up according to the target wake-up time or the corresponding time index. After the triggering condition is met, the data of the task to be woken up is located according to the target wake-up time, time index, or task identifier, and the target object, task status, task topic, and reach strategy data are read. During the reading process, the task status is verified to avoid duplicate operations on tasks that have already been completed, canceled, or reached.

[0052] The outreach strategy data is used to determine the outreach channel, content type, and sending control information. Based on the outreach channel, it calls SMS, application messaging, voice, or other information output interfaces to generate outreach content according to the task topic and content type, and outputs it to the receiving address corresponding to the target audience. Outreach processing can record the outreach time, outreach channel, and output status to distinguish whether the task has been completed.

[0053] When the number of tasks to be woken up is small and the target wake-up times are relatively dispersed, a periodic scanning method is used to determine the time. Real-time date and time data is read according to a set period, tasks to be woken up whose target wake-up time is no later than the current time are selected, the reach policy data is read, and the corresponding information output interface is called.

[0054] When there are a large number of tasks to be woken up and their trigger times are relatively concentrated, a time index and priority queue are used to locate tasks. Tasks are sorted according to their target wake-up time, with earlier tasks placed at the front of the queue. Once the real-time date and time reach the target wake-up time of the task at the front of the queue, the task content is read based on the task identifier, and the reach content is output according to the reach strategy. During concurrent reach, task locking tags can be used to prevent the same task from being read repeatedly.

[0055] In the fintech business, after the auto insurance claims follow-up task reaches the target wake-up time, the system reads vehicle repair details, customer contact information, and outreach strategy data from the task storage area, and sends a repair reminder to the customer through the corresponding message interface.

[0056] In the healthcare business, after the health management follow-up task reaches the target wake-up time, the system reads the review items, user receiving address, and reach strategy data, and outputs the review reminder content through the application message interface.

[0057] This embodiment unifies the format and time zone of real-time date and time data with the target wake-up time, accurately locates the task data to be woken up based on time conditions, and determines the reach channel and content according to the reach strategy data. This can reduce time misjudgment, task omission and duplicate reach, and improve the accuracy of timely reach and the stability of task processing.

[0058] In one embodiment, such as Figure 3 As shown, step S10 above includes: S101, Obtain the initial natural language sequence, divide the initial natural language sequence according to semantic pause positions, punctuation positions and time expression boundaries, and generate the initial text unit sequence and text boundary data; S102, Identify spoken omission text units, repeated text units, ambiguous word text units and non-standard time expression text units from the initial text unit sequence, and generate a set of text units to be processed; S103, extract the context text units located before and after each text unit to be processed from the initial text unit sequence to generate a context text unit set; S104, based on the context text unit set, perform semantic completion on the spoken omission text unit to generate a semantic completion text unit set and a semantic completion replacement relationship; based on the context text unit set, perform content merging on the repeated text unit to generate a content merged text unit set and a content merged replacement relationship. S105, for each ambiguous text unit, a candidate word set is generated based on the context text unit set, and the candidate words in the candidate word set are used to replace the corresponding ambiguous text units to generate a candidate text fragment set. Based on the semantic coherence of the candidate text fragment set, the target words corresponding to each ambiguous text unit are determined from the candidate word set to generate a target word set. The target words in the target word set are used to replace the corresponding ambiguous text units to generate a word correction text unit set and a word correction replacement relationship. S106, normalize the spoken time words in the non-standard time expression text units to generate a time-normalized text unit set and a time-normalized substitution relationship; S107, summarize the semantic completion text unit set, the content merging text unit set, the word correction text unit set, and the time normalization text unit set to generate a correction text unit set, and summarize the semantic completion replacement relationship, the content merging replacement relationship, the word correction replacement relationship, and the time normalization replacement relationship to generate a replacement correspondence relationship; S108, according to the replacement correspondence, the text units to be processed in the initial text unit sequence are replaced by the corrected text unit set to generate a corrected text unit sequence; S109, Reorganize the corrected text unit sequence according to the text boundary data, and unify the character format and punctuation form to generate a standard text sequence.

[0059] In this embodiment, the initial natural language sequence stores the input content according to the order of appearance of characters, words, or sub-words, and retains the positional markers of each content item. Semantic pause positions can be determined based on syntactic structure, tone changes, input intervals, or pause durations carried by speech transcription; punctuation positions are determined by the positions of periods, commas, question marks, and separators; time expression boundaries are determined based on time indicators, range modifiers, direction modifiers, and the relationship between adjacent words. Based on these three types of positions, continuous content is divided into multiple initial text units, and the start position, end position, boundary type, and adjacency relationship of each text unit are recorded, forming an initial text unit sequence and text boundary data. When time words and action words are adjacent, they can be retained within the same text unit to avoid separating the time meaning from the corresponding event.

[0060] Spoken omission text units are identified based on the missing sentence components. When a text unit lacks an action subject, action object, predicate component, or necessary modifier, and the missing content can be determined by adjacent text units, the text unit is marked as a spoken omission text unit. Repeated text units can be identified based on character overlap ratio, word arrangement similarity, and semantic similarity, including continuous repetition, spaced repetition, and semantic repetition. Ambiguous word text units can be identified based on the number of candidate meanings of the word in the current context, phonetic similarity, character similarity, and degree of collocational abnormality. Non-standard time expression text units include those lacking a complete date structure, using spoken time words, or containing time content with an uncertain range. The identification results are written into the set of text units to be processed according to their text position, and the category to which each text unit belongs is retained.

[0061] The set of contextual text units is extracted from the text units before and after each text unit to be processed. The extraction range can be determined by a fixed number of adjacent text units, or it can be dynamically adjusted based on sentence boundaries, the continuity of themes, and referential relationships. When transition words, conditional words, or new objects are present, outward expansion stops to avoid irrelevant content affecting the correction results. Each contextual text unit is linked to its corresponding text unit to be processed, enabling the same contextual content to serve omission completion, repetition merging, and ambiguity resolution respectively.

[0062] Semantic completion determines missing content based on the subject, action, object, and modifying relationships within the context of each text unit. For text units that omit the subject, the nearest semantically consistent subject is selected from the preceding text; for text units that omit the action object, the object matching the action is extracted from adjacent descriptions; for text units that omit the predicate, a predicate component that maintains the sentence's semantic integrity is added based on the preceding and following action relationships. The completed content forms a set of semantically completed text units, and the positions of the replaced text units, the added content, and the completed text units are recorded, forming a semantic completion substitution relationship.

[0063] Content merging determines the content to retain based on the character coverage and semantic inclusion relationships between duplicate text units. For complete duplication, the text unit that appears first and has a complete structure is retained, while the remaining duplicate content is deleted. For partial duplication, non-duplicated information is extracted from each text unit and combined to form a complete merged text unit. When semantics are repeated but wording differs, text units with greater information content and clear semantics are retained. The merging results form a set of merged text units, and the corresponding positions between multiple duplicate text units and the merged text units are recorded through content merging replacement relationships.

[0064] For ambiguous text units, candidate words are selected from a word dictionary, a list of near-homophones, a list of near-similar words, and contextual collocation data. Each candidate word replaces the ambiguous position, forming a set of candidate text segments. Semantic coherence is determined based on the degree of collocation between the candidate word and the words before and after it, the syntactic completeness of the candidate text segment, the consistency of the context theme, and the probability of word usage. After comparing the candidate text segments, the target word is determined from the candidate word set. The target word replaces the corresponding ambiguous word, forming a word-corrected text unit. The target word set, the word-corrected text unit set, and the word-correction replacement relationships jointly record the candidate selection results and position changes.

[0065] Spoken time words are standardized according to their time expression type. Words indicating relative dates are converted to a unified relative date form, words indicating time ranges are converted to time forms with range attributes, words indicating time periods are converted to unified time period names, and modifiers indicating uncertainty retain their range meaning. Word standardization only unifies the language form, not the specific date and time. The standardized text units form a time-standardized text unit set, and the time standardization substitution relationship records the positional correspondence between non-standard time expression text units and the standardization results.

[0066] The semantically complete text unit set, the content-merged text unit set, the word-corrected text unit set, and the time-normalized text unit set are aggregated into a corrected text unit set according to their initial text positions. Various substitution relationships are aggregated into substitution correspondence relationships based on the same positional benchmark. When multiple correction results cover the same text range, the retained result can be determined based on recognition confidence, coverage, and semantic integrity, and the substitution position is updated synchronously to avoid the same text unit being replaced repeatedly.

[0067] The text units to be processed are located in the initial text unit sequence according to the replacement correspondence. The corrected text units are written to the corresponding positions, and the unmarked text units retain their original content and arrangement, forming a corrected text unit sequence. Text boundary data is used to restore the connection relationship between text units, and punctuation is added or adjusted according to the boundary type. Character encoding, full-width and half-width forms, number forms, whitespace characters, and punctuation forms are converted into a unified format, finally generating a standard text sequence.

[0068] This embodiment subdivides natural language content by semantic pause positions, punctuation positions, and time expression boundaries, and performs omission completion, repetition merging, ambiguous word correction, and colloquial time word unification respectively. This can reduce semantic deviations caused by missing sentence components, content repetition, word meaning uncertainty, and differences in time expression forms. By maintaining the correspondence between the corrected content and the input position through text boundary data and replacement correspondence, it can avoid content misalignment or semantic mixing during text reconstruction, so that the standard text sequence has clear boundaries, a unified format, and relatively complete semantic content.

[0069] In one embodiment, such as Figure 4 As shown, step S20 above includes: S201, the standard text sequence is input into the shared semantic coding layer, and word semantic features, contextual semantic features and tone expression features are extracted from the standard text sequence through the shared semantic coding layer to generate a shared semantic feature sequence; S202, based on the shared semantic feature sequence, perform intent recognition to generate intent recognition data including intent category and intent intensity; S203, extract emotion word features, negation modification features and degree modification features from the shared semantic feature sequence to generate an emotion judgment feature sequence; S204, Based on the emotion determination feature sequence, perform emotion recognition to generate emotion recognition data including emotion category and emotion intensity; S205, perform category coding on the intent category in the intent recognition data and the emotion category in the emotion recognition data to generate intent category code and emotion category code; S206, normalize the intent intensity in the intent recognition data and the emotion intensity in the emotion recognition data to generate normalized intent intensity and normalized emotion intensity. S207, combine the intent category code, the normalized intent intensity, the emotion category code, and the normalized emotion intensity according to the preset feature positions to generate a user state vector.

[0070] In this embodiment, the standard text sequence is converted into a text tag sequence at the word or sub-word granularity and written into positional encoding, so that the meaning, order, and modification relationship of words can be distinguished. When multiple sentences are included, sentence segment encoding and effective position mask can be added to maintain the semantic relationship between sentences and eliminate the influence of padding characters on the encoding result.

[0071] The shared semantic coding layer can consist of self-attention coding units, bidirectional recurrent coding units, or convolutional coding units. Each coding unit receives text tags and positional codes, extracts the association between the current word and the words before and after it, and generates a semantic vector corresponding to the text position. Intent recognition and emotion recognition share the semantic vector to maintain consistency in the context on which the two types of recognition are based.

[0072] Word semantic features record the meaning of individual words and the meaning of word combinations; contextual semantic features record the subject, object, scope of negation, conditional relationship, adversative relationship, and temporal modification relationship; and tone expression features are extracted from modal particles, degree words, repeated words, interrogative forms, and punctuation combinations. These three types of features are arranged according to their position in the text, forming a shared semantic feature sequence.

[0073] Intent recognition extracts action tendency, negation relation, object relation, and time tendency from shared semantic feature sequences. The weighted convergence layer assigns weights to different text positions, the intent classification layer outputs the category score for each intent category and determines the intent category, and the intent strength output layer combines degree modification, repetition expression, and affirmation or negation degree to generate intent strength, avoiding directly using category confidence as intent strength.

[0074] Emotional word features are extracted from words with emotional connotations and their context. Negation modification features record the position, object, and scope of the negation word, while degree modification features record the direction of enhancement, weakening, and object of the degree word. These three types of features are combined according to their text position to form a sentiment determination feature sequence, preserving the modification relationships between emotional words and negation and degree words.

[0075] The emotion classification layer determines the emotion category based on the emotion judgment feature sequence, while the emotion intensity output layer generates the emotion intensity by combining the intensity of emotion words, the extent of degree modification, the scope of negation, and the degree of tone. When the text contains multiple emotional information, the main emotion category can be determined based on the category score, and other emotional information can be retained in the intensity generation.

[0076] The shared semantic encoding layer, intent classification layer, intent intensity output layer, emotion classification layer, and emotion intensity output layer can be jointly trained. Training data includes standard text sequences, intent category labels, intent intensity labels, emotion category labels, and emotion intensity labels. Category recognition uses classification loss, and intensity recognition uses regression loss; the losses are combined and the parameters are updated according to preset weights.

[0077] Intent and emotion categories are converted into fixed-dimensional category codes according to their encoding correspondence, which can be achieved using one-hot or dense coding. Intent and emotion intensities are normalized to a unified numerical range, with values ​​exceeding the set range truncated to boundary values. Category codes and normalized intensities are written into corresponding segments according to preset feature positions, generating user state vectors of fixed length and consistent structure.

[0078] This embodiment extracts word meaning, contextual relationships, and tone information uniformly through a shared semantic coding layer, enabling intent recognition and emotion recognition to be judged based on consistent semantic content. By generating category data and intensity data separately and retaining negation and degree modification relationships, it can reduce category bias and intensity bias caused by isolated word recognition. Through category coding, intensity normalization, and fixed position combination, the recognition results of different texts are converted into user state vectors with consistent dimensions and uniform numerical scale.

[0079] In one embodiment, such as Figure 5 As shown, step S30 above includes: S301, Obtain reference time data and dynamic calendar knowledge graph, and read date nodes, calendar event nodes, time period nodes and time relationship data between nodes from the dynamic calendar knowledge graph; S302, extract time indicator words, event indicator words, direction indicator words and range indicator words from the standard text sequence, and generate fuzzy time entities and time constraint data based on the time indicator words, event indicator words, direction indicator words and range indicator words; S303, determine the reference date node from the date node based on the reference time data; S304, semantically match the fuzzy time entity with the date node, the calendar event node and the time period node to generate node matching data, and generate a candidate time node set based on the node matching data, the reference date node and the time relationship data; S305, Based on the time constraint data, filter the candidate time node set to determine the target start time node and the target end time node; S306, extract start date and time data, end date and time data and time zone marker from the target start time node and the target end time node, and generate a standard time window based on the start date and time data, the end date and time data and the time zone marker.

[0080] In this embodiment, the reference time data can be composed of the current date, current time, time zone marker, and region marker, and is uniformly converted to the same time format. The current date and current time are used to determine the reference position for relative time, the time zone marker is used to eliminate date offsets between different time zones, and the region marker is used to select a data range consistent with the local calendar. When there is a clock deviation in the reference time data, it can be corrected according to a time calibration source to avoid overall offset caused by relative date derivation.

[0081] The dynamic calendar knowledge graph organizes time information according to nodes and relationships. Date nodes record specific dates, weekday attributes, work status, and the month they belong to; calendar event nodes record holidays, adjusted workdays, anniversaries, or other events with time boundaries; and time period nodes record morning, afternoon, evening, and other time intervals. Time relationship data is used to record date sequence, event start and end, date affiliation, time period inclusion, adjacent dates, and cross-cycle correspondences. When calendar arrangements change, the corresponding node attributes and relationships can be updated to ensure that the time information remains consistent with the current calendar state.

[0082] When reading date nodes, calendar event nodes, and time period nodes, the query range can be limited based on baseline time data. For example, the month of the current date and adjacent months can be used as the scope of the graph search to reduce time nodes that are irrelevant to the current semantics. The read results retain node identifiers, node names, start and end attributes, time zone attributes, and node relationships, providing structured data for subsequent time matching.

[0083] Time indicators are used to mark dates, days of the week, months, or time periods; event indicators are used to mark content related to calendar events; direction indicators are used to mark relative directions such as before, after, within, or outside; and range indicators are used to mark boundary information such as days before, days after, a whole week, or a specific time period. All types of indicators can be extracted using sequence labeling models, time dictionary matching, or a combination of both, and their corresponding text positions and mutual modification relationships can be recorded.

[0084] Fuzzy time entities are formed by combining indicator words with the same temporal meaning with adjacent modifiers. For content such as two days after a holiday, after next weekend, or a few days before the end of the month, it is necessary to preserve the dependencies between events, directions, and ranges, and avoid splitting each word into unrelated time segments. Time constraint data records relative direction, range length, boundary inclusion status, time period range, and periodicity attributes, which are used to limit the selection of subsequent candidate nodes.

[0085] The reference date node is determined from the date nodes based on the date and time zone in the reference time data. When regional calendar differences exist, date nodes with consistent regional markings can be selected first; when the current time crosses a date boundary, the reference date node can be determined according to the date after conversion to a unified time zone. The reference date node serves as a reference node for relative time derivation, used to determine the preceding and following positions and interval range between candidate nodes and the current date.

[0086] When performing semantic matching between fuzzy time entities and date nodes, calendar event nodes, and time period nodes, the semantic similarity between entity words and node names, the degree of consistency in event type, the degree of consistency in time attributes, and the degree of contextual matching can be compared. The matching results record the matching scores and corresponding relationships between the entity and each node. Based on the node matching data, nodes that satisfy relative direction and periodic relationships are searched from the baseline date node along the time relationship data, forming a candidate time node set. For example, when the entity includes "after a holiday," the event end date can be located from the matched calendar event node, and then candidate date nodes can be searched along the subsequent date relationships.

[0087] Time constraints are used to filter out candidate nodes that do not meet the direction, range, time period, or boundary conditions. Directional constraints exclude nodes located in the wrong time direction; range constraints limit the date interval and duration; boundary constraints determine whether the start and end nodes are included within the time range; and time period constraints retain nodes that meet the morning, afternoon, or evening criteria. After filtering, nodes with earlier times that meet the starting boundary conditions are identified as the target start time nodes, and nodes with later times that meet the ending boundary conditions are identified as the target end time nodes.

[0088] The start date and time data are extracted from the date attribute and time period start attribute of the target start time node, and the end date and time data are extracted from the date attribute and time period end attribute of the target end time node. Time zone markings can be obtained from node attributes, base time data, or regional calendar attributes. These three types of data are combined using a unified time format to form a standard time window containing clearly defined start and end boundaries and time zone information. When a specific time period is missing, the complete time range of the corresponding date can be used; when only a single time node is available, start and end boundaries can be generated according to the node's valid time period.

[0089] This embodiment determines the reference date and time zone for relative time through benchmark time data, organizes dates, calendar events, time periods and their time relationships using a dynamic calendar knowledge graph, and forms fuzzy time entities and time constraint data based on time indicators, event indicators, direction indicators and range indicators. It can filter start and end nodes that meet the semantic direction and range conditions from multiple candidate time nodes, and then generate a standard time window with clear time boundaries and time zone information, reducing date ambiguity and range deviation caused by fuzzy time content such as after holidays, after the next weekend, and a few days before the end of the month.

[0090] In one embodiment, such as Figure 6 As shown, step S40 above includes: S401, acquire task scheduling data, extract task deadline and task waiting time from the task scheduling data to generate task time limit data, extract resource availability period and resource load status from the task scheduling data to generate resource status data, and extract historical delay count and historical reach response from the task scheduling data to generate historical interaction data. S402, within the standard time window, set multiple initial wake-up times according to a preset time granularity to generate an initial wake-up time set; S403, Based on the task time limit data, filter the initial wake-up time set to generate a candidate wake-up time set; S404, Based on the resource status data, filter the candidate wake-up time set to generate an available wake-up time set; S405, Based on the user state vector and the historical interaction data, determine the delayed intention level, emotional state level and historical response level, and generate scheduling state data; S406, Based on the scheduling status data, the task time limit data, and each available wake-up time, generate a scheduling priority for each available wake-up time, and determine a target wake-up time from the set of available wake-up times based on the scheduling priority; S407, determine the reach method, reach content type and reach interval based on the scheduling status data and the historical interaction data, and generate reach strategy data based on the reach method, reach content type and reach interval.

[0091] In this embodiment, task scheduling data can be read from task records, resource status records, and interaction records, and associated according to task identifiers. The task deadline represents the latest time a task is allowed to complete, and the task waiting time represents the time a task has remained unprocessed since its creation, last occurrence, or last status change. Both are combined using a unified time unit to form task time limit data, used to determine whether each time point exceeds the deadline boundary and whether continuing to wait will cause the task to exceed the allowed time.

[0092] Resource availability time periods record the date and time range within which resources can handle tasks. Resource load status records the number of tasks assigned, remaining processing capacity, or occupancy level within each time period. Resource status data associates availability time periods with load status according to time location, ensuring that each candidate time period has both resource availability and resource capacity. Historical postponement counts are obtained by statistically analyzing the number of times users have requested postponement for the same or similar tasks. Historical response counts can include whether a response was received, the response time, the response channel, and the response result. The above data is arranged according to interaction time to form historical interaction data.

[0093] The standard time window is divided into multiple time positions according to a preset time granularity. The preset time granularity can be set to minutes, hours, time periods, or dates depending on the window length; a finer granularity is used for shorter windows, and a coarser granularity is used for longer windows. Date, time, and time zone information are written to each time position to form an initial wake-up time set. Wake-up indicates that the task has entered an accessible state at the corresponding time and does not involve waking a person from sleep.

[0094] Task deadline data is used to filter out initial wake-up times later than the task deadline and to determine whether the interval from the current time to each initial wake-up time exceeds the allowed waiting time. Time that meets the deadline boundary and waiting time requirements is retained to form a candidate wake-up time set. Resource status data is used to further check whether each candidate wake-up time falls within the resource availability period and to exclude times when resources are unavailable or the load exceeds the set range, forming an available wake-up time set.

[0095] The intent category, intent intensity, emotion category, and emotion intensity in the user state vector are used to determine the delayed intent level and emotion state level. The delayed intent level can be classified according to the delay category and its intensity, and the emotion state level can be classified according to the emotion category and its intensity. Historical delay counts, historical reach response counts, average response time, and effective response ratio can be converted into historical response levels. These three levels are combined using fixed fields to form scheduling state data, which describes the current user state and past interaction states.

[0096] Each available wake-up time is associated with the task deadline, allowed waiting time, and scheduling status data. When the task deadline is approaching and the waiting time is long, the scheduling priority of earlier times is increased; when the intention to delay is high or resistance is strong, the scheduling priority of earlier times is decreased; when historical response levels show that the user has a good response status during a specific period, the scheduling priority of that time is increased. The priority data can be combined according to preset weights, or a hierarchical sorting method can be used to compare task deadlines, user status, and time position sequentially. The target wake-up time is determined from the set of available wake-up times based on the sorting results.

[0097] The delivery method can include text messages, application notifications, voice notifications, or other information output formats. The delivery content type is used to distinguish between reminders, confirmations, prompts, or operation instructions. The delivery interval indicates the time distance between adjacent deliveries. When the delay intention level is high, a delivery method with lower interference can be selected and the delivery interval extended; when the emotional state level indicates strong resistance, a weak urging content type can be used; when the historical response level shows that a certain delivery method has a good response result, the selection priority of that delivery method can be increased. The delivery method, delivery content type, and delivery interval are combined according to a fixed data structure to generate delivery strategy data.

[0098] This embodiment limits the wake-up time range by setting task deadlines and task waiting times, excludes times when tasks cannot be handled by utilizing available resource periods and resource load status, and adjusts the scheduling priority of each available time based on user state vectors and historical interaction data. This enables the generation of a target wake-up time that simultaneously satisfies task deadlines, resource status, and user status. By delaying intent level, emotional state level, and historical response level to determine the reach method, reach content type, and reach interval, it can reduce time deviations and reach mismatches caused by reach at fixed times and in fixed forms.

[0099] In one embodiment, such as Figure 7 As shown, step S50 above includes: S501, extract task identifier, target data, task status data and task topic data from the task scheduling data to generate basic task data; S502, generate time index data according to the date level and time period level of the target wake-up time, and generate a wake-up task identifier based on the task identifier and the target wake-up time; S503, combine the reach strategy data and the task basic data to generate task content data; S504, generate wake-up task data based on the wake-up task identifier, the target wake-up time, the time index data, and the task content data; S505, the task storage area is divided into a task content area and a time index area. The wake-up task identifier and task content data in the wake-up task data are written into the task content area, and the wake-up task identifier and time index data in the wake-up task data are written into the time index area. S506, based on the wake-up task identifier written into the task content area and the time index area respectively, establish the correspondence between the task content data and the time index data.

[0100] In this embodiment, task scheduling data is read according to task identifiers. Task identifiers are used to distinguish different items to be scheduled and can consist of a task source marker, creation time, and sequence code. The target data records the receiving target and address; the task status data records states such as waiting, paused, completed, or canceled; and the task subject data records the item to be transmitted, its content, and the target action. All of the above data is organized according to fixed fields to form the basic task data, ensuring that task identity, receiving target, current status, and content subject are associated.

[0101] The target wake-up time is converted to a standard date and time format under a unified time zone and split into date level and time period level. The date level can include year, month, day, and calendar period, while the time period level can include intraday intervals, hourly intervals, and minute intervals. The time index data records the index value of each level and the complete target wake-up time, enabling tasks to be grouped according to a coarse time range while retaining accurate time positions.

[0102] The wake-up task identifier is generated jointly by the task identifier and the target wake-up time. The generation process can either concatenate the task identifier with a standardized time code, or perform digest encoding on the combined content. Different identifiers are generated for the same task with different target wake-up times. When the same task and the same target wake-up time are generated repeatedly, the identifier remains consistent, reducing the duplication of writing to the same wake-up task.

[0103] The outreach strategy data records the outreach method, content type, and interval, while the task basic data records the task's own content. During the combination process, both types of data are written into the same structured record according to field name and data type to avoid overlap between outreach control information and task content. Outreach object data, task status data, and task topic data are stored as task content fields, while outreach strategy data is stored as outreach control fields, forming the task content data.

[0104] The task data to be woken up combines the wake-up task identifier, target wake-up time, time index data, and task content data according to a fixed field structure. After combination, it verifies whether the wake-up task identifier is empty, whether the target wake-up time format is consistent, whether the time index data matches the target wake-up time, and whether the task content data contains the necessary fields. After passing the verification, a writable structured task record is formed.

[0105] The task content area and the time index area can be set as different logical areas within the same storage space, or as independent storage areas. The task content area stores task content data using the wake-up task identifier as the key, while the time index area stores time index data and the wake-up task identifier according to date and time period levels. The task content area focuses on storing complete task content, while the time index area focuses on storing data arranged by time, reducing the duplication of large task content in the time index.

[0106] The wake-up task identifier is written to both regions as a common associated field. A one-to-one correspondence is established between the wake-up task identifier in the time index region and the identifier with the same name in the task content region, enabling the time index data to locate the corresponding task content data. The two regions can be written in the same commit batch, and the identifier consistency is checked before committing to avoid data separation caused by writing only the task content or only the time index.

[0107] This embodiment extracts task identifiers, reachable objects, task status, and task topics to form structured task content, and generates hierarchical time indexes and unique wake-up task identifiers based on the target wake-up time. This maintains consistency between task content, reach control information, and time attributes. By partitioning and writing the task content area and the time index area and associating them with common identifiers, it reduces problems such as duplicate storage of task content, task identifier conflicts, and mismatches between the time index and task content, thereby improving the integrity of the data to be woken up and the efficiency of storage organization.

[0108] In one embodiment, such as Figure 8 As shown, step S60 above includes: S601, Obtain real-time date and time data, compare the time represented by the real-time date and time data with the target wake-up time, and generate a time comparison result; S602, when the time comparison result indicates that the time represented by the real-time date and time data is not earlier than the target wake-up time, the task data to be woken up is obtained from the task storage area according to the target wake-up time; S603, extract task content data from the task data to be woken up, and extract reach strategy data, reach object data, task status data and task topic data from the task content data; S604, when the task status data is in a waiting-to-reach state, extract the reach method and reach content type from the reach strategy data; S605, generate outreach content based on the task topic data, the task status data, and the outreach content type; S606, output the reach content to the reach object corresponding to the reach object data according to the reach method.

[0109] In this embodiment, the real-time date and time data is provided by the system clock, time synchronization service, or trusted time source, and includes the current date, current time, time zone marker, and time precision. After reading, the real-time date and time data and the target wake-up time are converted to the same time zone and date and time format, and second-level, millisecond-level, or minute-level differences are processed according to uniform precision to avoid the inconsistency of format and time zone affecting time judgment.

[0110] The two unified time values ​​are compared to generate a time comparison result indicating whether the time is earlier than, equal to, or later than the target wake-up time. If the real-time date and time data is equal to or later than the target wake-up time, the task to be woken up is determined to meet the triggering condition; if the real-time date and time data is earlier than the target wake-up time, the corresponding task content is not read.

[0111] Retrieve the corresponding record in the task storage area based on the target wake-up time. When a time index is set in the task storage area, the time index data can be located according to the target wake-up time. The wake-up task identifier can be obtained from the time index data, and then the task data to be woken up can be read according to the wake-up task identifier. When multiple tasks are associated with the same target wake-up time, they are read separately according to each wake-up task identifier to avoid missing tasks at the same time position.

[0112] From the task data to be woken up, locate the task content data, and extract the outreach strategy data, outreach object data, task status data, and task topic data according to the field structure of the task content data. The outreach object data includes the receiving target and its corresponding receiving address; the task status data records whether the task is currently allowed to be reached; the task topic data records the content to be delivered; and the outreach strategy data controls the output format and content category.

[0113] When the task status data is in the "waiting to be reached" state, continue reading the reach strategy data. When the task status is in the "completed," "cancelled," "paused," or other non-waiting state, stop reaching the corresponding task to reduce the duplicate output of expired tasks. The reach method is used to determine the information output interface, and the reach content type is used to determine the organization format and information fields of the output content.

[0114] When generating outreach content, the content composition fields are determined according to the outreach content type. The task object, task content, and target action are extracted from the task theme data, and the current status and pending tasks are extracted from the task status data. The corresponding data is then written into the content composition fields. Reminder-type content highlights the task and time information; confirmation-type content highlights the current status and pending confirmations; and operation guidance-type content highlights the target action and processing entry point, thus forming outreach content consistent with the task theme and task status.

[0115] Select the corresponding information output interface according to the reach method, convert the reach object data into a receiving address that the interface can recognize, and encode the reach content into the data format required by the interface. When the reach method is a text message, generate text message data; when it is an application notification, generate a notification title and notification body; when it is a voice notification, convert the reach content into voice output data and then output it to the corresponding receiving end of the reach object.

[0116] This embodiment reduces premature or delayed triggering caused by time format differences by unifying the time zone, format, and precision of real-time date and time data with the target wake-up time; it reduces task omissions and duplicate triggering of invalid tasks by locating the task data to be woken up by the target wake-up time and wake-up task identifier, and controls whether to reach it based on the task status; it ensures consistency in the timing, content, and output format of the reach by generating reach content through task topic data, task status data, and reach content type, and then outputting it to the corresponding reach object according to the reach method.

[0117] In one embodiment, the insurtech platform is used to handle follow-up tasks during the self-service auto insurance claims process, guiding customers to the repair shop. After the customer completes the claim and damage assessment, they can input "um, the car hasn't been repaired yet, I'm very busy with work these days, let's talk about it after I finish this, maybe after the holiday, please don't keep calling" via mobile terminal. The interaction interface forms an initial natural language sequence according to the message sending order, and retains the session identifier, case task identifier, input time, and source channel. The initial natural language sequence contains colloquial omissions, repetitive semantics, ambiguous time content, and conflicting tones, and cannot be directly used as a basis for subsequent scheduling.

[0118] The text processing algorithm segments the input content according to semantic pauses, punctuation marks, and time boundaries, resulting in initial text units such as "vehicle not yet repaired," "busy with work recently," "we'll talk after this period," "we'll go after the holiday," and "don't keep calling." The algorithm records the start and end positions and adjacency relationships of each text unit. From the initial text unit sequence, it identifies the omitted spoken language content corresponding to "we'll go after this period," identifies the repetitive semantics between "vehicle has been very busy these past two days" and "we'll talk after this period," identifies potential word ambiguities between "in-site repair" and "in-factory repair" in speech transcription, and identifies "vehicle has been very busy these past two days," "we'll talk after this period," and "after the holiday" as non-standard time expressions.

[0119] To address the lack of action objects, supplementary content is extracted from the previous sections on "vehicle not yet repaired" and "car insurance task," resulting in "re-entry for repair after the holiday." For repetitive content between "very busy" and "finished," valid information regarding customers' inability to process transactions recently and those requiring further processing after completing current tasks is retained and grouped under "recently busy." Candidate terms for "entry" and "re-entry to the workshop" are written into contextual text fragments, and "entry to the workshop" is determined as the target term based on the repair business context and semantic coherence of preceding and following words. For "these two days," "finished," and "after the holiday," only character and word forms are standardized, and vague time attributes are retained; specific dates are not specified during text processing. Each corrected text unit is written back to its original position according to the replacement correspondence, and then reorganized based on text boundary data to generate a standard text sequence: "vehicle not yet repaired, recently busy, scheduled for repair after the holiday, please reduce phone reminders."

[0120] The standard text sequence input shares a semantic encoding layer. Word encoding, position encoding, and context encoding jointly extract information about vehicle repair matters, tendency to postpone processing, post-holiday time semantics, and tone information to reduce phone reminders. The shared semantic feature sequence is used for intent recognition and emotion recognition respectively. Intent recognition, based on recent busy work, post-holiday arrangements, and the desire to reduce reminders, identifies the customer's intent as explicit postponement, and generates a higher intent intensity based on the number of postponement words, the explicitness of the tone, and the relationship of action delay. Emotion recognition extracts emotion word features, negation modification features, and degree modification features from content such as "busy" and "don't keep calling," and identifies the emotion category as mild anxiety accompanied by resistance. The intent category and emotion category are converted into fixed-dimensional category codes, and the intent intensity and emotion intensity are normalized to a unified numerical range. Then, they are combined according to preset positions to form a user state vector containing explicit postponement, high postponement intensity, mild anxiety, and moderate resistance intensity.

[0121] The time parsing section uses the time semantic parsing module to read the customer's input time as the base time data. For example, if a customer inputs "go after the holiday" at 14:20 on September 29th of the configured year, the base time data includes September 29th, 14:20, the business area, and the East 8 time zone. The dynamic calendar knowledge graph includes specific date nodes, statutory holiday nodes, weekend nodes, weekday nodes, morning time period nodes, afternoon time period nodes, and the start and end dates, inclusion, adjacency, and sequence relationships between nodes. It can also maintain local weather warnings as calendar event nodes with time ranges.

[0122] The temporal semantic parsing module extracts sections from the standard text sequence as event indicators, then uses these as directional indicators, and combines the sections after the holiday into fuzzy time entities. Based on September 29th as the baseline date node, the module semantically matches the sections after the holiday with the National Day holiday event nodes in the knowledge graph. Candidate date nodes are then searched along the subsequent date relationships from the holiday's end node. The graph shows that the holiday ends on October 7th, and October 8th is the first working day after the holiday. Since there are no regional weather warnings covering the entire day on October 8th, October 8th is selected as the target date. If the customer does not specify a specific time period, a standard time window from 9:00 AM to 6:00 PM on October 8th is generated based on the available repair service time periods and marked with the East 8 time zone. If the graph shows weather warnings affecting travel on October 8th, the affected time periods can be excluded from the candidate range, while retaining the available time range after the warning is lifted.

[0123] The task scheduling data consists of case task records, partner repair shop resource records, and historical interaction records. Case task records include the task deadline, current waiting time, and case processing status; resource records include the appointment capacity and load status of partner repair shops during different time periods on October 8th; historical interaction records include the number of times the customer has postponed, past phone call answering status, SMS viewing status, and mobile message response status. For example, if a current case requires repair arrangement confirmation by October 15th, has been waiting for six days, has a previous postponement record, has low response times via telephone, but fast response times via mobile text messages, and the partner repair shop has high load from 9:00 AM to 12:00 PM on October 8th, but has remaining appointment capacity from 3:00 PM to 6:00 PM.

[0124] The dynamic expedited repair state machine sets multiple initial wake-up times between 9:00 AM and 6:00 PM on October 8th, using a two-hour time granularity. Task deadlines and waiting times are used to exclude times that might cause further task backlogs. Repair shop availability and resource load status are used to exclude high-load periods such as 9:00 AM to 12:00 PM, resulting in a set of available wake-up times including 3:00 PM and 5:00 PM. User state vectors and historical interaction data are used to determine high levels of delay intent, moderate levels of resistance, and historically good response levels from mobile text channels. Based on this, the state machine lowers the priority of premature reach times, increases the priority of afternoon periods when repair shop load is lower and customers are more likely to respond, and, considering both case deadlines and current waiting times, determines 3:00 PM on October 8th as the target wake-up time.

[0125] The outreach strategy is generated based on scheduling status data and historical interaction data. If the customer explicitly requests fewer phone calls, the outreach method is selected as mobile text messaging, avoiding outbound calls. If the customer is experiencing their first or fewest delays and shows signs of resistance, the outreach content type is selected as a caring confirmation, minimizing urging language. If the customer has historically responded quickly to mobile messages, the outreach interval is set to a low frequency of interference. If the customer has already delayed multiple times and the case is close to its deadline, the dynamic reminder state machine can increase the task's urgency level, shorten the outreach interval, and adjust the outreach content type to a time-sensitive reminder, while still selecting an appropriate text intensity based on the customer's emotional state.

[0126] The task management records extract the case task identifier, customer mobile receiving address, waiting contact status, and in-factory repair arrangement confirmation topic to form the basic task data. The target wake-up time is generated by creating time index data corresponding to October 8th and the afternoon time slot according to date and time period levels. The case task identifier is then combined with 15:00 on October 8th to generate the wake-up task identifier. The contact method, care confirmation content, and contact interval are combined with the basic task data to form the task content data. The wake-up task identifier, target wake-up time, time index data, and task content data together constitute the task data to be woken up.

[0127] The task storage area is divided into a task content area and a time index area. The task content area stores customer receiving address, task status, in-service repair topic, and outreach strategy data, using the wake-up task identifier as the associated field. The time index area stores time index data and the same wake-up task identifier according to the time period of October 8th and the afternoon. The two storage areas are linked by the wake-up task identifier, enabling the scheduling service to locate task content based on the time index, while avoiding duplicate storage of complete customer information and outreach content in the time index area.

[0128] Upon arrival at 3 PM on October 8th, the dispatch service reads real-time date and time data from a trusted time source and converts it to the same time zone and time precision as the target wake-up time. If the time comparison result shows that the real-time time is equal to or later than the target wake-up time, the wake-up task identifier is obtained based on the afternoon time index of October 8th, and the corresponding wake-up task data is then read from the task content area. The dispatch service extracts the mobile text delivery method, care confirmation content type, customer receiving address, waiting to be reached status, and factory repair arrangement confirmation topic from the task content data.

[0129] When the task status is still "awaiting contact," contact content is generated based on the task topic, task status, and contact content type. For example, it might say, "Hello, you previously indicated that you would like to schedule vehicle repairs after the holiday. Our current partner repair shop has afternoon appointments available. If you wish to schedule, you can confirm in the current conversation. If it is still inconvenient to handle this in the near future, you can continue to inform us via text." The dispatch service sends the contact content to the customer's corresponding conversation address using the mobile text contact method. The entire process translates the customer's natural language expressions such as "after the holiday," "busy recently," and "reducing phone follow-ups" into specific time windows, user state vectors, and differentiated contact strategies, enabling unstructured dialogues to form storable, location-based, and timely triggered claims follow-up tasks.

[0130] In one embodiment, a reach device based on a dynamic calendar knowledge graph is provided, which corresponds one-to-one with the reach methods based on dynamic calendar knowledge graphs in the above embodiments. (Refer to...) Figure 9 , Figure 9 This is a schematic diagram of the functional modules of a preferred embodiment of the outreach device based on a dynamic calendar knowledge graph according to the present invention. The outreach device based on a dynamic calendar knowledge graph includes: a text preprocessing module 10, a user state recognition module 20, a fuzzy time parsing module 30, a dynamic scheduling decision module 40, a wake-up task storage module 50, and a timed outreach module 60. Detailed descriptions of each functional module are as follows: Text preprocessing module 10 is used to obtain an initial natural language sequence, perform text processing on the initial natural language sequence, and generate a standard text sequence; User state recognition module 20 is used to perform intent recognition and emotion recognition based on the standard text sequence and generate a user state vector. The fuzzy time parsing module 30 is used to acquire reference time data and dynamic calendar knowledge graph, extract fuzzy time entities from the standard text sequence, and perform time parsing on the fuzzy time entities based on the reference time data and the dynamic calendar knowledge graph to generate a standard time window; The dynamic scheduling decision module 40 is used to acquire task scheduling data, perform dynamic scheduling processing based on the user state vector, the standard time window and the task scheduling data, and generate target wake-up time and reach strategy data. The wake-up task storage module 50 is used to generate wake-up task data based on the target wake-up time, the reach strategy data and the task scheduling data, and write the wake-up task data into the task storage area. The timed outreach module 60 is used to acquire real-time date and time data. When the time represented by the real-time date and time data is not earlier than the target wake-up time, it acquires the task data to be woken up from the task storage area and performs outreach processing according to the outreach strategy data contained in the task data to be woken up.

[0131] In one embodiment, the text preprocessing module 10 is specifically used for: Obtain an initial natural language sequence, divide the initial natural language sequence according to semantic pause positions, punctuation positions, and time expression boundaries, and generate an initial text unit sequence and text boundary data; Identify spoken omission text units, repeated text units, ambiguous word text units, and non-standard time expression text units from the initial text unit sequence to generate a set of text units to be processed; Extract the context text units located before and after each text unit to be processed from the initial text unit sequence to generate a context text unit set; Based on the context text unit set, semantic completion is performed on the spoken omission text unit to generate a semantically completed text unit set and a semantic completion replacement relationship. Based on the context text unit set, content merging is performed on the repeated text unit to generate a content merged text unit set and a content merged replacement relationship. For each ambiguous text unit, a candidate word set is generated based on the context text unit set, and the candidate words in the candidate word set are used to replace the corresponding ambiguous text units to generate a candidate text fragment set. Based on the semantic coherence of the candidate text fragment set, the target words corresponding to each ambiguous text unit are determined from the candidate word set to generate a target word set. The target words in the target word set are used to replace the corresponding ambiguous text units to generate a word correction text unit set and a word correction replacement relationship. Word normalization is performed on spoken time words in the non-standard time expression text units to generate a time-normalized text unit set and time-normalized substitution relationship; The semantic completion text unit set, the content merging text unit set, the word correction text unit set, and the time normalization text unit set are summarized to generate a correction text unit set. The semantic completion replacement relationship, the content merging replacement relationship, the word correction replacement relationship, and the time normalization replacement relationship are summarized to generate a replacement correspondence relationship. According to the replacement correspondence, the text units to be processed in the initial text unit sequence are replaced by the corrected text unit set to generate the corrected text unit sequence; The corrected text unit sequence is reorganized according to the text boundary data, and the character format and punctuation are unified to generate a standard text sequence.

[0132] In one embodiment, the user state recognition module 20 is specifically used for: The standard text sequence is input into a shared semantic coding layer, and word semantic features, contextual semantic features, and tone expression features are extracted from the standard text sequence through the shared semantic coding layer to generate a shared semantic feature sequence. Intent recognition is performed based on the shared semantic feature sequence to generate intent recognition data including intent category and intent intensity; Emotion word features, negation modification features, and degree modification features are extracted from the shared semantic feature sequence to generate an emotion judgment feature sequence; Based on the emotion determination feature sequence, emotion recognition is performed to generate emotion recognition data including emotion category and emotion intensity; The intent category in the intent recognition data and the emotion category in the emotion recognition data are categorized and encoded to generate intent category codes and emotion category codes; The intent intensity in the intent recognition data and the emotion intensity in the emotion recognition data are normalized to generate normalized intent intensity and normalized emotion intensity. A user state vector is generated by combining the intent category code, the normalized intent intensity, the emotion category code, and the normalized emotion intensity according to preset feature positions.

[0133] In one embodiment, the fuzzy time resolution module 30 is specifically used for: Obtain baseline time data and dynamic calendar knowledge graph, and read date nodes, calendar event nodes, time period nodes and time relationship data between nodes from the dynamic calendar knowledge graph; Extract time indicators, event indicators, direction indicators, and range indicators from the standard text sequence, and generate fuzzy time entities and time constraint data based on the time indicators, event indicators, direction indicators, and range indicators; The reference date node is determined from the date node based on the reference time data; The fuzzy time entity is semantically matched with the date node, the calendar event node, and the time period node to generate node matching data, and a candidate time node set is generated based on the node matching data, the baseline date node, and the time relationship data. Based on the time constraint data, the candidate time node set is filtered to determine the target start time node and the target end time node; Extract start date and time data, end date and time data, and time zone marker from the target start time node and the target end time node, and generate a standard time window based on the start date and time data, the end date and time data, and the time zone marker.

[0134] In one embodiment, the dynamic scheduling decision module 40 is specifically used for: Obtain task scheduling data, extract task deadline and task waiting time from the task scheduling data to generate task time limit data, extract resource availability period and resource load status from the task scheduling data to generate resource status data, and extract historical delay count and historical reach response from the task scheduling data to generate historical interaction data. Within the standard time window, multiple initial wake-up times are set according to a preset time granularity to generate an initial wake-up time set. Based on the task time limit data, the initial wake-up time set is filtered to generate a candidate wake-up time set; Based on the resource status data, the candidate wake-up time set is filtered to generate an available wake-up time set; Based on the user state vector and the historical interaction data, the delayed intention level, emotional state level and historical response level are determined, and scheduling state data is generated. Based on the scheduling status data, the task time limit data, and each available wake-up time, a scheduling priority for each available wake-up time is generated, and a target wake-up time is determined from the set of available wake-up times based on the scheduling priority. Based on the scheduling status data and the historical interaction data, the reach method, reach content type and reach interval are determined, and reach strategy data is generated based on the reach method, reach content type and reach interval.

[0135] In one embodiment, waking up the task storage module 50 is specifically used for: Extract task identifier, target data, task status data, and task topic data from the task scheduling data to generate basic task data; Time index data is generated according to the date level and time period level of the target wake-up time, and a wake-up task identifier is generated based on the task identifier and the target wake-up time; Combine the reach strategy data and the task basic data to generate task content data; Data to be woken up is generated based on the wake-up task identifier, the target wake-up time, the time index data, and the task content data. The task storage area is divided into a task content area and a time index area. The wake-up task identifier and task content data in the wake-up task data are written into the task content area, and the wake-up task identifier and time index data in the wake-up task data are written into the time index area. Based on the wake-up task identifiers written into the task content area and the time index area respectively, a correspondence between the task content data and the time index data is established.

[0136] In one embodiment, the timing notification module 60 is specifically used for: Acquire real-time date and time data, compare the time represented by the real-time date and time data with the target wake-up time, and generate a time comparison result; When the time comparison result indicates that the time represented by the real-time date and time data is not earlier than the target wake-up time, the task data to be woken up is obtained from the task storage area according to the target wake-up time. Extract task content data from the task data to be woken up, and extract out outreach strategy data, outreach object data, task status data, and task topic data from the task content data; When the task status data is in a waiting-to-reach state, the reach method and reach content type are extracted from the reach strategy data; Generate outreach content based on the task topic data, the task status data, and the outreach content type; The reach content is output to the reach object corresponding to the reach object data according to the reach method.

[0137] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements server-side functions or steps of a dynamic calendar knowledge graph-based outreach method.

[0138] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 11 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides determination and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When executed by the processor, the computer program implements client-side functions or steps of a dynamic calendar knowledge graph-based outreach method.

[0139] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain an initial natural language sequence, perform text processing on the initial natural language sequence, and generate a standard text sequence; Intent recognition and emotion recognition are performed based on the standard text sequence to generate a user state vector; Acquire baseline time data and dynamic calendar knowledge graph, extract fuzzy time entities from the standard text sequence, and perform time parsing on the fuzzy time entities based on the baseline time data and the dynamic calendar knowledge graph to generate a standard time window; Obtain task scheduling data, perform dynamic scheduling processing based on the user state vector, the standard time window, and the task scheduling data, and generate target wake-up time and reach strategy data; Based on the target wake-up time, the reach strategy data, and the task scheduling data, wake-up task data is generated and written into the task storage area; Acquire real-time date and time data. If the time represented by the real-time date and time data is not earlier than the target wake-up time, retrieve the task data to be woken up from the task storage area and perform outreach processing according to the outreach strategy data contained in the task data to be woken up.

[0140] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, and a computer program is stored thereon, which, when executed by a processor, performs the following steps: Obtain an initial natural language sequence, perform text processing on the initial natural language sequence, and generate a standard text sequence; Intent recognition and emotion recognition are performed based on the standard text sequence to generate a user state vector; Acquire baseline time data and dynamic calendar knowledge graph, extract fuzzy time entities from the standard text sequence, and perform time parsing on the fuzzy time entities based on the baseline time data and the dynamic calendar knowledge graph to generate a standard time window; Obtain task scheduling data, perform dynamic scheduling processing based on the user state vector, the standard time window, and the task scheduling data, and generate target wake-up time and reach strategy data; Based on the target wake-up time, the reach strategy data, and the task scheduling data, wake-up task data is generated and written into the task storage area; Acquire real-time date and time data. If the time represented by the real-time date and time data is not earlier than the target wake-up time, retrieve the task data to be woken up from the task storage area and perform outreach processing according to the outreach strategy data contained in the task data to be woken up.

[0141] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0142] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0143] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0144] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0145] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for reaching users based on a dynamic calendar knowledge graph, characterized in that, Includes the following steps: Obtain an initial natural language sequence, perform text processing on the initial natural language sequence, and generate a standard text sequence; Intent recognition and emotion recognition are performed based on the standard text sequence to generate a user state vector; Acquire baseline time data and dynamic calendar knowledge graph, extract fuzzy time entities from the standard text sequence, and perform time parsing on the fuzzy time entities based on the baseline time data and the dynamic calendar knowledge graph to generate a standard time window; Obtain task scheduling data, perform dynamic scheduling processing based on the user state vector, the standard time window, and the task scheduling data, and generate target wake-up time and reach strategy data; Based on the target wake-up time, the reach strategy data, and the task scheduling data, wake-up task data is generated and written into the task storage area; Acquire real-time date and time data. If the time represented by the real-time date and time data is not earlier than the target wake-up time, retrieve the task data to be woken up from the task storage area and perform outreach processing according to the outreach strategy data contained in the task data to be woken up.

2. The outreach method based on a dynamic calendar knowledge graph as described in claim 1, characterized in that, Obtaining an initial natural language sequence, performing text processing on the initial natural language sequence to generate a standard text sequence, including: Obtain an initial natural language sequence, divide the initial natural language sequence according to semantic pause positions, punctuation positions, and time expression boundaries, and generate an initial text unit sequence and text boundary data; Identify spoken omission text units, repeated text units, ambiguous word text units, and non-standard time expression text units from the initial text unit sequence to generate a set of text units to be processed; Extract the context text units located before and after each text unit to be processed from the initial text unit sequence to generate a context text unit set; Based on the context text unit set, semantic completion is performed on the spoken omission text unit to generate a semantically completed text unit set and a semantic completion replacement relationship. Based on the context text unit set, content merging is performed on the repeated text unit to generate a content merged text unit set and a content merged replacement relationship. For each ambiguous text unit, a candidate word set is generated based on the context text unit set, and the candidate words in the candidate word set are used to replace the corresponding ambiguous text units to generate a candidate text fragment set. Based on the semantic coherence of the candidate text fragment set, the target words corresponding to each ambiguous text unit are determined from the candidate word set to generate a target word set. The target words in the target word set are used to replace the corresponding ambiguous text units to generate a word correction text unit set and a word correction replacement relationship. Word normalization is performed on spoken time words in the non-standard time expression text units to generate a time-normalized text unit set and time-normalized substitution relationship; The semantic completion text unit set, the content merging text unit set, the word correction text unit set, and the time normalization text unit set are summarized to generate a correction text unit set. The semantic completion replacement relationship, the content merging replacement relationship, the word correction replacement relationship, and the time normalization replacement relationship are summarized to generate a replacement correspondence relationship. According to the replacement correspondence, the text units to be processed in the initial text unit sequence are replaced by the corrected text unit set to generate the corrected text unit sequence; The corrected text unit sequence is reorganized according to the text boundary data, and the character format and punctuation are unified to generate a standard text sequence.

3. The outreach method based on a dynamic calendar knowledge graph as described in claim 1, characterized in that, Intent recognition and emotion recognition are performed based on the standard text sequence to generate a user state vector, including: The standard text sequence is input into a shared semantic coding layer, and word semantic features, contextual semantic features, and tone expression features are extracted from the standard text sequence through the shared semantic coding layer to generate a shared semantic feature sequence. Intent recognition is performed based on the shared semantic feature sequence to generate intent recognition data including intent category and intent intensity; Emotion word features, negation modification features, and degree modification features are extracted from the shared semantic feature sequence to generate an emotion judgment feature sequence; Based on the emotion determination feature sequence, emotion recognition is performed to generate emotion recognition data including emotion category and emotion intensity; The intent category in the intent recognition data and the emotion category in the emotion recognition data are categorized and encoded to generate intent category codes and emotion category codes; The intent intensity in the intent recognition data and the emotion intensity in the emotion recognition data are normalized to generate normalized intent intensity and normalized emotion intensity. A user state vector is generated by combining the intent category code, the normalized intent intensity, the emotion category code, and the normalized emotion intensity according to preset feature positions.

4. The outreach method based on a dynamic calendar knowledge graph as described in claim 1, characterized in that, Acquire baseline time data and a dynamic calendar knowledge graph; extract fuzzy time entities from the standard text sequence; and perform time parsing on the fuzzy time entities based on the baseline time data and the dynamic calendar knowledge graph to generate a standard time window, including: Obtain baseline time data and dynamic calendar knowledge graph, and read date nodes, calendar event nodes, time period nodes and time relationship data between nodes from the dynamic calendar knowledge graph; Extract time indicators, event indicators, direction indicators, and range indicators from the standard text sequence, and generate fuzzy time entities and time constraint data based on the time indicators, event indicators, direction indicators, and range indicators; The reference date node is determined from the date node based on the reference time data; The fuzzy time entity is semantically matched with the date node, the calendar event node, and the time period node to generate node matching data, and a candidate time node set is generated based on the node matching data, the baseline date node, and the time relationship data. Based on the time constraint data, the candidate time node set is filtered to determine the target start time node and the target end time node; Extract start date and time data, end date and time data, and time zone marker from the target start time node and the target end time node, and generate a standard time window based on the start date and time data, the end date and time data, and the time zone marker.

5. The outreach method based on a dynamic calendar knowledge graph as described in claim 1, characterized in that, Acquire task scheduling data, perform dynamic scheduling processing based on the user state vector, the standard time window, and the task scheduling data, and generate target wake-up time and reach strategy data, including: Obtain task scheduling data, extract task deadline and task waiting time from the task scheduling data to generate task time limit data, extract resource availability period and resource load status from the task scheduling data to generate resource status data, and extract historical delay count and historical reach response from the task scheduling data to generate historical interaction data. Within the standard time window, multiple initial wake-up times are set according to a preset time granularity to generate an initial wake-up time set. Based on the task time limit data, the initial wake-up time set is filtered to generate a candidate wake-up time set; Based on the resource status data, the candidate wake-up time set is filtered to generate an available wake-up time set; Based on the user state vector and the historical interaction data, the delayed intention level, emotional state level and historical response level are determined, and scheduling state data is generated. Based on the scheduling status data, the task time limit data, and each available wake-up time, a scheduling priority for each available wake-up time is generated, and a target wake-up time is determined from the set of available wake-up times based on the scheduling priority. Based on the scheduling status data and the historical interaction data, the reach method, reach content type and reach interval are determined, and reach strategy data is generated based on the reach method, reach content type and reach interval.

6. The outreach method based on a dynamic calendar knowledge graph as described in claim 1, characterized in that, Based on the target wake-up time, the reach strategy data, and the task scheduling data, wake-up task data is generated, and the wake-up task data is written to the task storage area, including: Extract task identifier, target data, task status data, and task topic data from the task scheduling data to generate basic task data; Time index data is generated according to the date level and time period level of the target wake-up time, and a wake-up task identifier is generated based on the task identifier and the target wake-up time; Combine the reach strategy data and the task basic data to generate task content data; Data to be woken up is generated based on the wake-up task identifier, the target wake-up time, the time index data, and the task content data. The task storage area is divided into a task content area and a time index area. The wake-up task identifier and task content data in the wake-up task data are written into the task content area, and the wake-up task identifier and time index data in the wake-up task data are written into the time index area. Based on the wake-up task identifiers written into the task content area and the time index area respectively, a correspondence between the task content data and the time index data is established.

7. The outreach method based on a dynamic calendar knowledge graph as described in claim 1, characterized in that, Acquire real-time date and time data. If the time represented by the real-time date and time data is not earlier than the target wake-up time, retrieve the task data to be woken up from the task storage area, and perform outreach processing according to the outreach strategy data contained in the task data to be woken up, including: Acquire real-time date and time data, compare the time represented by the real-time date and time data with the target wake-up time, and generate a time comparison result; When the time comparison result indicates that the time represented by the real-time date and time data is not earlier than the target wake-up time, the task data to be woken up is obtained from the task storage area according to the target wake-up time. Extract task content data from the task data to be woken up, and extract out outreach strategy data, outreach object data, task status data, and task topic data from the task content data; When the task status data is in a waiting-to-reach state, the reach method and reach content type are extracted from the reach strategy data; Generate outreach content based on the task topic data, the task status data, and the outreach content type; The reach content is output to the reach object corresponding to the reach object data according to the reach method.

8. A reach device based on a dynamic calendar knowledge graph, characterized in that, The reach device based on the dynamic calendar knowledge graph includes: The text preprocessing module is used to obtain an initial natural language sequence, perform text processing on the initial natural language sequence, and generate a standard text sequence. The user state recognition module is used to perform intent recognition and emotion recognition based on the standard text sequence and generate a user state vector. The fuzzy time parsing module is used to acquire reference time data and dynamic calendar knowledge graph, extract fuzzy time entities from the standard text sequence, and perform time parsing on the fuzzy time entities based on the reference time data and the dynamic calendar knowledge graph to generate a standard time window; The dynamic scheduling decision module is used to acquire task scheduling data, perform dynamic scheduling processing based on the user state vector, the standard time window and the task scheduling data, and generate target wake-up time and reach strategy data. The wake-up task storage module is used to generate wake-up task data based on the target wake-up time, the reach strategy data and the task scheduling data, and write the wake-up task data into the task storage area; The timed outreach module is used to acquire real-time date and time data. When the time represented by the real-time date and time data is not earlier than the target wake-up time, it acquires the task data to be woken up from the task storage area and performs outreach processing according to the outreach strategy data contained in the task data to be woken up.

9. A computer device, characterized in that, The computer device includes a memory, a processor, and a dynamic calendar knowledge graph-based outreach program stored in the memory and executable on the processor. When executed by the processor, the dynamic calendar knowledge graph-based outreach program implements the steps of the dynamic calendar knowledge graph-based outreach method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a reach program based on a dynamic calendar knowledge graph. When the reach program based on the dynamic calendar knowledge graph is executed by the processor, it implements the steps of the reach method based on the dynamic calendar knowledge graph as described in any one of claims 1-7.

Images