An AI follow-up robot intelligent interaction method and system

Abstract

The application relates to the field of artificial intelligence and natural language processing technology, and discloses an AI follow-up robot intelligent interaction method and system, wherein the method comprises the following steps: extracting a structured entity; constructing a semantic context vector based on the structured entity; performing emotion source classification according to the emotion polarity; outputting an individualized response expected time; judging whether the response is overdue and triggering a compensation reminder strategy. Compared with the follow-up interaction response mode in the prior art which depends on a rule template or a keyword trigger, especially in the unstructured scene where a patient expresses the drug experience, subjective feeling or non-timely response in free language, the technical problem that accurate recognition, emotion understanding and individualized compensation are difficult to realize is solved, and since the application introduces a structured label extraction mechanism, fuses an emotion discrimination method, and combines a response prediction model of historical behavior and emotion state, sensitive information recognition in doctor-patient interaction is realized, and the understanding ability and follow-up interaction efficiency of the AI follow-up robot are improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, and in particular to an AI follow-up robot intelligent interaction method and system. Background Technology

[0002] Currently, with the development of telemedicine and intelligent health management, AI-based follow-up methods using social platforms (such as WeChat groups and app chat rooms) are widely used in scenarios such as chronic disease monitoring, postoperative recovery assessment, and mental health intervention. Doctors or nurses typically guide patients to regularly report their physical condition and medication status using pre-set scripts or questionnaire templates, hoping to improve the efficiency and coverage of follow-ups through AI. However, most existing follow-up systems have the following shortcomings:

[0003] (1) Patients have a weak ability to understand unstructured natural language information. In actual communication, they often describe their condition in a subjective, colloquial, or even emotional way, such as: "I still feel chest tightness after taking the medicine for three days" or "I feel very bad, I don't know if it's a side effect." Such descriptions combine symptom characteristics, time characteristics, and subjective emotional expression. They contain a lot of information but are expressed in a highly free manner. Traditional methods based on keyword dictionaries or rule matching are difficult to accurately identify the symptom-time-drug association structure and cannot generate standardized medical structure data (such as entity labels: symptom = chest tightness; drug = antihypertensive drug; time = three days).

[0004] (2) Lack or poor ability to perceive emotional state. In doctor-patient interactions, patients' expression of negative emotions (such as anxiety, depression, anger) often accompanies changes in their condition or subjective assessment of treatment efficacy, which has important clinical early warning value. However, most existing follow-up systems have not built an emotion recognition mechanism. Even if some systems have simple emotion classification models, it is difficult to effectively distinguish the boundary between medical symptom descriptions and subjective emotion expressions, which leads to "heart discomfort" and "bad mood" being confused as similar risk warnings, misleading doctors' judgment.

[0005] (3) The lack of a personalized response expectation modeling mechanism makes it common for patients not to reply to follow-up tasks in WeChat groups in a timely manner. The current system usually reminds patients based on a fixed time window (e.g., if no reply is received within 24 hours, the system will push the message again), without considering the patient's historical response behavior, the context of the time of sending the message, and the content of the message type (whether it is a sensitive topic). Therefore, it is impossible to determine whether "silence" is a clinical signal that requires active intervention, such as intentional neglect, network abnormality, or low mood, which may lead to misjudgment or missed intervention opportunities.

[0006] Therefore, there is an urgent need for an AI follow-up robot intelligent interaction method that integrates medical knowledge graphs, emotion semantic dual-channel discrimination mechanism and behavior prediction model, so as to realize the structured label extraction of unstructured messages, real-time identification of emotional state, identification of trigger source type and intelligent modeling of individual response behavior. Summary of the Invention

[0007] To address the aforementioned technical shortcomings, the present invention aims to propose an intelligent interaction method for AI follow-up robots. This method addresses the technical challenges of existing follow-up interaction response modes that rely on rule templates or keyword triggers, particularly in unstructured scenarios where patients freely express their medication experiences, subjective feelings, or fail to respond promptly. These challenges make it difficult to achieve accurate identification, emotional understanding, and personalized compensation.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides an intelligent interaction method for an AI follow-up robot.

[0009] The AI ​​follow-up robot's intelligent interaction method includes:

[0010] Step S10: Obtain the original text message P of the target patient in the preset chat platform follow-up group, construct a time window ΔT, and perform structured label extraction processing on the original text message P based on the knowledge-enhanced semantic parsing mechanism to output a structured entity set E.

[0011] Step S20: Construct a semantic context vector S based on the structured entity set E, input the semantic context vector S into a preset emotion recognition network, and output emotion polarity labels. And confidence level of emotional polarity ;

[0012] Step S30: Based on emotion polarity labels And confidence level of emotional polarity Perform a secondary classification task for emotion sources, determine and output the type of emotion trigger source. Among them, the types of emotion triggers Including subjective emotion expression types and medical symptom description types ;

[0013] Step S40: Output the type of emotion trigger source Then, obtain the message sending time of the original text message P. and patient response status ;Message sending time and patient response status Input into a pre-defined patient response behavior prediction model, and output individualized expected response time. ;

[0014] Step S50: Based on individualized expected response time A multi-channel linkage closed-loop control mechanism is adopted to perform response timeout judgment and compensation reminder triggering.

[0015] Preferably, in step S10, the structured entity set E includes three entities. ,in, For the i-th identified entity item; , represents the entity type label corresponding to the i-th identified entity item; is the starting character position in the original text message P; n is the total number of entities.

[0016] Preferably, step S10, which involves performing structured tag extraction processing on the original text message P based on a knowledge-enhanced semantic parsing mechanism to output a structured entity set E, specifically includes:

[0017] First, part-of-speech tagging is performed: the original text message P is initially scanned using a pre-set part-of-speech tagging library. When the original text message P contains a sentence structure that combines adverbs, verbs and nouns, it is determined that there are symptom-expressing features or medication experience features. When the original text message P contains a sentence structure that combines time adverbs and verbs, it is determined that there are time-related behavioral event features.

[0018] Then, synonym phrase clustering and recognition are performed: based on the pre-constructed medical knowledge graph, combined with the characteristics of expressed symptoms, medication experience, and time behavior events, a dual discrimination method based on fuzzy editing distance matching and word vector cosine similarity fusion is used to perform fuzzy matching and word vector semantic similarity supplementary judgment, and output structured entity candidates and entity category labels;

[0019] Then, a medical-sensitive starting position factor is introduced: a medical-sensitive starting position factor is constructed, and the position attention factor includes an initiation template phrase factor, a transition structure cue factor, and an interaction history enhancement factor; based on the medical-sensitive starting position factor, the structured entity candidates and entity category labels are weighted, and finally the candidate semantic window set W is output;

[0020] Finally, based on the candidate semantic window set W, a dual-channel semantic discrimination mechanism is further used to classify labels and confirm structured entities, outputting a structured entity set E.

[0021] Preferably, in step S10, the final step of further using a dual-channel semantic discrimination mechanism to classify labels and confirm structured entities based on the candidate semantic window set W, and outputting a structured entity set E, specifically includes:

[0022] Static grammar channel stage: Extract syntactic dependency relations and contextual part-of-speech flow from the candidate semantic window set W. Based on the syntactic dependency relations and contextual part-of-speech flow, call the sequence model BiLSTM-CRF to determine and output the first entity category. and first confidence level;

[0023] Dynamic Context Passage Stage: The pre-trained medical language model MedBERT is invoked to perform enhanced classification of the candidate semantic window set W based on the overall context, and the second entity category is output. Second confidence level;

[0024] Based on the first entity category Second Entity Category The consistency score is obtained by using the semantic vector cosine similarity method: when the consistency score is greater than or equal to the preset consistency score threshold, the label is directly classified; when the consistency score is less than the preset consistency score threshold, the label is classified by confidence fusion method based on the first confidence score and the second confidence score.

[0025] Preferably, in step S20, the preset emotion recognition network adopts a dual-channel semantic emotion fusion mechanism driven by medical entity perception; the emotion recognition network includes an input layer for receiving a set of structured entities E; a medical entity perception layer for using an entity attention mechanism to achieve semantic alignment between structured entities and context; a semantic context fusion layer for using a context modeling method with a bidirectional gating mechanism to fuse the dependency and emotion propagation relationships between contexts; a multi-label emotion classification layer for using a preset output mechanism combining a multi-label softmax function and a sigmoid function to map the dependency and emotion propagation relationships between the fused contexts to multiple emotion category spaces, including anxiety category space, anger category space, complaint category space, low mood category space, neutral statement category space, and positive gratitude category space; an emotion consistency discrimination layer for performing a consistency check between emotion expression and medical entity description before the final output; and an output layer for outputting emotion polarity labels. And confidence level of emotional polarity .

[0026] Preferably, in step S40, the patient's response status The result is obtained after sequentially determining the message passing status and the interaction response status, specifically including:

[0027] Get ,based on Message passing status determination:

[0028]

[0029] in, This is the message passing status;

[0030] The patient's response status is continuously monitored and output within a time window ΔT:

[0031]

[0032] in, The patient's response status.

[0033] Preferably, in step S40, the preset patient response behavior prediction model is constructed using a hybrid deep prediction architecture based on temporal behavior embedding and joint modeling of emotional state. The patient response behavior prediction model specifically includes:

[0034] The behavior embedding layer is used to receive and encode the historical interaction behavior sequence of the preset input, and uses a multi-channel embedding mechanism to process the historical interaction behavior sequence and output the historical behavior embedding matrix.

[0035] The emotion state injection layer is used to receive the output historical behavior embedding matrix and emotion trigger source type. A position-aligned encoding mechanism is used to embed the output historical behavior matrix and the emotion trigger source type. Process the data to output a joint emotion-behavior vector sequence;

[0036] The temporal modeling layer receives the joint emotion-behavior vector sequence, processes the joint emotion-behavior vector sequence using a bidirectional gated recurrent unit network (Bi-GRU), and outputs a deep semantic representation of the patient's behavioral trends.

[0037] The doctor-patient interaction feature fusion layer is used to receive deep semantic expressions of patient behavior trends, obtain task interaction cues, and use an attention-weighted fusion mechanism to fuse the deep semantic expressions and task interaction cues, outputting a fused representation vector; among them, task interaction cues include "whether it includes a specifier" and "whether it is a secondary reminder";

[0038] The response duration regression prediction layer receives the fused representation vector, employs a multilayer perceptron, and outputs an individualized expected response time. Response confidence score and response time rating label.

[0039] This invention also provides an AI follow-up robot intelligent interaction system comprising:

[0040] The semantic parsing and entity extraction module is used to obtain the original text message P of the target patient in the preset chat platform follow-up group, construct a time window ΔT, and perform structured label extraction processing on the original text message P based on the knowledge-enhanced semantic parsing mechanism to output a structured entity set E.

[0041] The emotion recognition and context construction module is used to construct a semantic context vector S based on a structured entity set E. The semantic context vector S is then input into a pre-defined emotion recognition network, and the module outputs emotion polarity labels. And confidence level of emotional polarity ;

[0042] The emotion trigger source identification module is used to identify emotions based on emotion polarity labels. And confidence level of emotional polarity Perform a secondary classification task for emotion sources, determine and output the type of emotion trigger source. Among them, the types of emotion triggers Including subjective emotion expression types and medical symptom description types ;

[0043] The response behavior modeling module is used to output the type of emotion trigger source. Then, obtain the message sending time of the original text message P. and patient response status ;Message sending time and patient response status Input into a pre-defined patient response behavior prediction model, and output individualized expected response time. ;

[0044] The closed-loop judgment and compensation reminder module is used to determine the individualized expected response time. A multi-channel linkage closed-loop control mechanism is adopted to perform response timeout judgment and compensation reminder triggering.

[0045] The present invention also provides an AI follow-up robot intelligent interaction device, comprising: a memory, a processor, and an AI follow-up robot intelligent interaction program stored in the memory and executable on the processor. When the AI ​​follow-up robot intelligent interaction program is executed by the processor, it implements an AI follow-up robot intelligent interaction method.

[0046] The present invention also provides a computer program product, including an AI follow-up robot intelligent interaction program, which, when executed by a processor, implements the AI ​​follow-up robot intelligent interaction method.

[0047] The beneficial effects of this invention are as follows: This invention can accurately extract standardized medical entity tags from unstructured natural language text of patients in follow-up groups, and combine knowledge graphs and dual-channel semantic discrimination mechanisms to achieve highly accurate analysis of symptom descriptions, medication experiences and time-based behavioral events, effectively improving the follow-up robot's ability to understand medical information and its ability to express it in a structured manner.

[0048] This invention establishes an individualized response expectation time model based on emotional state and behavioral timing through the deep integration of emotion recognition and patient response behavior prediction, and introduces a multi-channel linkage response timeout compensation mechanism, thereby enhancing the personalized response efficiency and humanistic care capability of follow-up interaction, and significantly reducing the risk of message neglect and missed response in passive follow-up. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart illustrating the first embodiment of an AI follow-up robot intelligent interaction method according to the present invention.

[0051] Figure 2 This is a schematic diagram comparing the number of structured tags extracted in the first embodiment of the AI ​​follow-up robot intelligent interaction method of the present invention.

[0052] Figure 3 This is a schematic diagram comparing the accuracy of an AI follow-up robot's intelligent interaction method in an emotion recognition task, according to the first embodiment of the present invention.

[0053] Figure 4 This is a schematic diagram comparing the confidence distribution in the emotion trigger source identification task of the first embodiment of the AI ​​follow-up robot intelligent interaction method of the present invention.

[0054] Figure 5 This is a schematic diagram illustrating service activation guidance for a first embodiment of an AI follow-up robot intelligent interaction method according to the present invention.

[0055] Figure 6 This is a schematic diagram illustrating the synchronous guidance of analysis results in the first embodiment of an AI follow-up robot intelligent interaction method of the present invention.

[0056] Figure 7 This is a schematic diagram of an AI follow-up robot intelligent interaction method according to the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the AI ​​follow-up robot intelligent interaction method of the present invention, which presents the first embodiment of the AI ​​follow-up robot intelligent interaction method of the present invention.

[0059] In the first embodiment, the AI ​​follow-up robot intelligent interaction method includes:

[0060] Step S10: Obtain the original text message P of the target patient in the preset chat platform follow-up group, construct a time window ΔT, and perform structured label extraction processing on the original text message P based on the knowledge-enhanced semantic parsing mechanism to output a structured entity set E.

[0061] It should be noted that the "knowledge-enhanced semantic parsing mechanism" refers to a semantic understanding technology framework that integrates medical knowledge graphs and natural language processing algorithms. This mechanism includes: a context-embedded word vector generation module (such as a pre-trained medical BERT model), a medical terminology entity recognition and disambiguation module (supporting mapping between generic and brand names), a rule-enhanced time normalization module (e.g., converting "these two days" and "the night before last" into standard time segments), and a multi-dimensional label classifier for symptoms, medications, and emotions (for structured annotation); enabling detailed extraction of potential medical-related entities (such as symptoms, medication, timeliness, and efficacy feedback) from the original text message P.

[0062] Understandably, this step involves introducing a time window ΔT to transform the isolated understanding of a single message into a dynamic semantic aggregation process within a continuous temporal context, thereby improving the contextual consistency and accuracy of tag extraction. It identifies information with stronger behavioral intent or implicit emotional tendencies within continuous messages. For example, a patient's expressions "My stomach is upset today" and "I took medication but it didn't help" belong to different messages, but can be uniformly parsed based on temporal aggregation to obtain structural tags such as [symptom: stomach discomfort], [medication: currently taking medication], and [efficacy feedback: not relieved], providing stable semantic input for subsequent emotion recognition and interaction judgment.

[0063] It should be understood that, compared to traditional medical question-answering robots that use rule trees or keyword matching strategies to only recognize explicit terms (such as "pain" or "fever"), this invention, based on a mechanism that integrates semantic parsing and knowledge enhancement, can identify implicit expressions or omitted information and automatically complete structural labels. For example, in the phrase "This medicine is really uncomfortable," the symptom name does not appear directly, but by combining the drug name and emotion words, the label "drug side effects" can be inferred. Furthermore, the introduction of time normalization and drug disambiguation modules effectively reduces the technical shortcomings of traditional methods, such as ambiguous terminology, high ambiguity recognition rates, and contextual breaks.

[0064] For example, such as Figure 2As shown, with the same message sample size, the number of structured labels extracted by the method of this invention is significantly higher than that of the traditional CRF model method, indicating that it has stronger robustness and extraction depth in semantic understanding, medical knowledge fusion, and context recognition. Especially when the number of samples is large (e.g., more than 800), the label growth curve of the method of this invention continues to rise steadily, while the traditional method gradually approaches a plateau, reflecting that the invention has better generalization ability in dealing with long texts and multi-turn dialogues.

[0065] Step S20: Construct a semantic context vector S based on the structured entity set E, input the semantic context vector S into a preset emotion recognition network, and output emotion polarity labels. And confidence level of emotional polarity ;

[0066] It should be noted that the "semantic context vector S" refers to a multi-dimensional vector representation constructed based on medical tags such as symptoms, drugs, and time extracted from the structured entity set E, combined with contextual semantic features, behavioral history, and typical expression methods, through a pre-defined vector generation mechanism (such as fusing word vector representations from medical BERT with node embeddings from the symptom-drug co-occurrence map). This semantic context vector not only includes the literal features of the current message (such as average pooling of word vectors in the text content), but also encodes background factors such as the trend of emotional changes in the previous n rounds of interaction, the stage of the treatment cycle, and the status of medication feedback, thus providing a multimodal joint input for the emotion recognition model.

[0067] Understandably, this step effectively solves the problems of "context loss" and "emotional semantic ambiguity" in medical dialogues by integrating structured semantics with contextual historical behavior into a unified semantic context vector S. Compared to model designs that directly judge the emotion of a single message, this invention adopts a context-aware mechanism that can capture the implicit anxiety caused by "continuous medication without therapeutic effect" or the potential risk of depression caused by "continuous expression of helplessness," thereby achieving more accurate emotional polarity label output (such as negative, neutral, positive) and its corresponding confidence score (usually a probability value between 0 and 1).

[0068] It should be understood that, compared to the simple LSTM and emotion dictionary strategies used in traditional medical dialogue analysis systems, this invention introduces a hybrid perception architecture in the emotion recognition network that integrates structural labels, medical knowledge graph priors, and temporal embedding mechanisms. This effectively improves the recognition ability in complex scenarios such as implicit expressions, elliptical emotional sentences, and negative reinforcement expressions. For example, for sentences expressing "feeling useless" that have negative semantics but no explicit emotion words, traditional methods may judge them as neutral or fail to recognize them. However, this invention, after identifying the drug entity and the historical feedback "persistently unrelieved," matches them with a negative emotion template and finally outputs a "negative" emotion label, significantly improving the accuracy of recognition and the credibility of emotional confidence.

[0069] For example, such as Figure 3 As shown, the proposed "semantic context vector and hybrid emotion network" scheme significantly outperforms traditional methods (such as BiLSTM and emotion dictionaries) across all dimensions of emotion recognition: the overall accuracy is improved by 19.3 percentage points, demonstrating stronger generalization ability; in the challenging category of "negative emotion recognition," the accuracy increases from 52.1% to 83.4%, an improvement of over 60%, significantly enhancing the ability to recognize the emotional fluctuations of sensitive patients; and it also improves by approximately 9-17 percentage points in "positive" and "neutral emotions," respectively, improving the accuracy of perceiving multiple emotional states. This advantage is mainly due to the introduction of "structured entity-driven semantic context modeling" in this step, which not only focuses on the emotion words themselves but also considers their behavioral context, temporal state, and symptom information (e.g., "I took the medicine but still feel unwell"), thereby avoiding the misjudgment of omitted expressions and semantically ambiguous sentences by traditional methods.

[0070] Step S30: Based on emotion polarity labels And confidence level of emotional polarity Perform a secondary classification task for emotion sources, determine and output the type of emotion trigger source. Among them, the types of emotion triggers Including subjective emotion expression types and medical symptom description types ;

[0071] It should be noted that the "secondary classification task of emotion source" refers to, based on obtaining the emotion polarity label (such as positive, neutral, negative) and confidence level, further determining whether the emotion is triggered by subjective psychological state (such as anxiety, anger, helplessness, etc.) or by objective medical symptoms (such as pain, vomiting, fatigue, etc.).

[0072] Understandably, this step aims to improve the accuracy of emotion perception in complex doctor-patient dialogues, distinguishing between two types of emotional expressions: "objective feedback from patients about their own symptoms" and "subjective dissatisfaction from patients with treatment and the medical process." For example, "I took the medicine but it didn't work at all" includes objective feedback of symptom ineffectiveness and potential negative emotions; "I really feel like I can't go on anymore" leans more towards psychological and emotional imbalance. By differentiating the source of emotions, precise responses can be made: the former can prompt doctors to optimize medication strategies, while the latter can trigger psychological intervention mechanisms.

[0073] It should be understood that, compared to traditional medical AI systems that merely classify emotions as positive or negative or simply label them, this invention, by constructing a subjective-objective emotion discrimination mechanism, can accurately perceive the context triggering emotions, greatly reducing "emotion mismatches." Traditional systems easily misjudge "emotions caused by changes in the patient's condition" as "patient emotional instability," leading to unnecessary psychological interventions or misunderstandings between doctors and patients. This invention, however, identifies "medically driven emotions" through symptom entities and contextual reasoning, constructing a more semantically understanding follow-up interaction path.

[0074] For example, such as Figure 4 As shown, the method of this invention exhibits more concentrated confidence output in both subjective emotion expression and medical symptom description categories, with mean values ​​of 88% and 85% respectively, significantly higher than the 68% and 72% of traditional rule-based methods. Furthermore, it shows smaller fluctuations and stronger robustness. The figure reveals a strong overlap in the recognition of the two categories by traditional methods, making it prone to misidentifying the emotion source type. In contrast, the method of this invention, supported by structured entity guidance and semantic context enhancement mechanisms, can more clearly distinguish the semantic boundary between subjective emotions and pathological statements, significantly improving the responsiveness to sensitive emotional scenarios and the accuracy of automatic intervention.

[0075] Step S40: Output the type of emotion trigger source Then, obtain the message sending time of the original text message P. and patient response status ;Message sending time and patient response status Input into a pre-defined patient response behavior prediction model, and output individualized expected response time. ;

[0076] It should be noted that the "patient response behavior prediction model" refers to a prediction mechanism that integrates time series modeling and behavioral feature learning to individually estimate the timing of a patient's response in follow-up conversations. Specifically, it includes: constructing an individual behavioral feature vector based on a time-aware neural network (such as a time-gated recurrent network T-GRU) built from historical message response time series, combined with patient profile features (such as age group, disease type, message type sensitivity, and average past response delays); the model input consists of the current message sending time, emotion label, symptom label, and the k most recent response behaviors, and the output is the patient's expected response time interval in the current scenario.

[0077] Understandably, this step, through joint modeling of patients' historical behavioral data and contextual features, can dynamically predict the appropriate response time for patients to the current message. It considers not only the time factor but also the emotional urgency and symptom importance of the message content, allowing for customized response expectations to be set for different patients.

[0078] It should be understood that, compared to the traditional method of using a "uniform duration" as the response timeout for all patients (e.g., no response within 8 hours is considered ignored), the method of this invention uses neural networks to perform personalized modeling of historical behavioral data, which can predict the reasonable response window of the patient immediately after the message is sent, avoiding false alarms for slow-responding users or delayed intervention for fast-responding users.

[0079] Step S50: Based on individualized expected response time A multi-channel linkage closed-loop control mechanism is adopted to perform response timeout judgment and compensation reminder triggering.

[0080] It should be noted that the "multi-channel linkage closed-loop control mechanism" refers to the automatic activation of a compensation mechanism that includes multiple interactive channels such as text reminders, voice push notifications, outbound telephone calls, and doctor-side prompts when it is determined that the patient has not completed an effective response within the expected individualized response time. The mechanism also adjusts in a closed loop based on real-time feedback from the patient's response status. The control logic integrates a response delay assessment function, a risk level assessment module, and resource scheduling strategies to ensure that different response levels trigger different compensation paths. The entire process possesses a complete feedback chain of "delay determination—reminder triggering—status update—strategy adjustment."

[0081] Understandably, this step, by establishing a continuous monitoring mechanism for individual patient response behavior, can accurately identify "abnormal delayed response" situations and proactively intervene. It not only judges whether the timeout has occurred based on static time points, but also dynamically adjusts the compensation method by combining factors such as the emotional urgency of the message content and the severity of symptoms. This achieves intelligent linkage control strategies such as "key symptoms - priority outbound calls, minor delays - text reminders," thereby improving intervention efficiency and reducing manual workload.

[0082] For example, such as Figure 5 As shown, after logging in, users need to actively click the "WeChat Doctor Assistant" service entry and add the robot as a friend. During this process, the system does not force users to participate or automatically bind their accounts, demonstrating the user's clear autonomy in triggering the robot's service. Furthermore, the prompt "Click the button below to have the robot serve you" serves only as a guide and does not constitute a mandatory process, further demonstrating the effectiveness of the user-centric logic. This graphical interface visually illustrates the "user-driven" service activation strategy emphasized in this invention, effectively avoiding the incompatibility issues of forced system pushes and passive user acceptance in traditional health assistant systems. Figure 6 As shown, in the actual dialogue interface between the robot and the patient, the robot proactively uses inquiring emotional guidance language (such as "How are you feeling today?" or "Do you need me to make an appointment with a doctor?"), and automatically identifies the polarity of emotions and classifies the trigger source type based on the different types of responses from the patient. Furthermore, the system determines whether the emotion belongs to the category of subjective emotional expression or medical symptom description based on the identification results, and displays "Synchronization Successful" in the bottom status prompt bar, indicating that the system has synchronized the current user's emotional state to the cloud database. This process, combined with message interaction timestamps and response behavior, further triggers subsequent response expectation time prediction and compensation mechanisms, intuitively demonstrating the invention's closed-loop control capability of completing emotion perception, semantic understanding, state reasoning, and linked response in real-time interaction.

[0083] Example 2: Furthermore, the present invention provides an AI follow-up robot intelligent interaction system that employs an AI follow-up robot intelligent interaction method from the above embodiments, thereby solving a technical problem related to AI follow-up robot intelligent interaction. Compared with the prior art, the beneficial effects of the AI ​​follow-up robot intelligent interaction system provided by the present invention are the same as those of the AI ​​follow-up robot intelligent interaction method provided in the above embodiments, and other technical features of the AI ​​follow-up robot intelligent interaction system are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0084] Example 3: This invention provides an AI follow-up robot intelligent interactive device, please refer to... Figure 7An AI follow-up robot intelligent interaction device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to execute the AI ​​follow-up robot intelligent interaction method described in Embodiment 1 above. The AI ​​follow-up robot intelligent interaction device in this embodiment may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. This AI follow-up robot intelligent interaction device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this invention. An AI follow-up robot intelligent interaction device may include a processing device 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. Random access memory 1004 also stores various programs and data required for the operation of an AI follow-up robot intelligent interactive device. Processing device 1001, read-only memory 1002, and random access memory 1004 are interconnected via bus 1005. I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows an AI follow-up robot intelligent interactive device to communicate wirelessly or wiredly with other devices to exchange data. Although an AI follow-up robot intelligent interactive device with various systems is shown in the figure, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented or possessed alternatively.

[0085] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the AI ​​follow-up robot intelligent interaction method described above. The computer program product provided by this invention can solve the technical problem of intelligent interaction in an AI follow-up robot. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the AI ​​follow-up robot intelligent interaction method provided in the above embodiments, and will not be repeated here.

[0086] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0087] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0088] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An intelligent interaction method for an AI follow-up robot, characterized in that, The methods include: Step S10: Obtain the original text message P of the target patient in the preset chat platform follow-up group, construct a time window ΔT, and perform structured label extraction processing on the original text message P based on the knowledge-enhanced semantic parsing mechanism to output a structured entity set E; wherein, the step of performing structured label extraction processing on the original text message P based on the knowledge-enhanced semantic parsing mechanism to output a structured entity set E specifically includes: First, part-of-speech tagging is performed: the original text message P is initially scanned using a pre-set part-of-speech tagging library. When the original text message P contains a sentence structure that combines adverbs, verbs and nouns, it is determined that there are symptom-expressing features or medication experience features. When the original text message P contains a sentence structure that combines time adverbs and verbs, it is determined that there are time-related behavioral event features. Then, synonym phrase clustering and recognition are performed: based on the pre-constructed medical knowledge graph, combined with the characteristics of expressed symptoms, medication experience, and time behavior events, a dual discrimination method based on fuzzy editing distance matching and word vector cosine similarity fusion is used to perform fuzzy matching and word vector semantic similarity supplementary judgment, and output structured entity candidates and entity category labels; Then, a medical-sensitive starting position factor is introduced: a medical-sensitive starting position factor is constructed, and the position attention factor includes an initiation template phrase factor, a transition structure cue factor, and an interaction history enhancement factor; based on the medical-sensitive starting position factor, the structured entity candidates and entity category labels are weighted, and finally, a candidate semantic window set W is output; Finally, based on the candidate semantic window set W, the dual-channel semantic discrimination mechanism is further used to classify labels and confirm structured entities, and output the structured entity set E. Step S20: Construct a semantic context vector S based on the structured entity set E, input the semantic context vector S into a preset emotion recognition network, and output emotion polarity labels. and confidence level of emotional polarity ; Step S30: Based on emotion polarity labels and confidence level of emotional polarity Perform a secondary classification task for emotion sources, determine and output the type of emotion trigger source. Among them, the types of emotion triggers Including subjective emotion expression types and medical symptom description types ; Step S40: Output the type of emotion trigger source Then, obtain the message sending time of the original text message P. and patient response status ;Message sending time and patient response status Input into a pre-defined patient response behavior prediction model, and output individualized expected response time. ; Step S50: Based on individualized expected response time A multi-channel linkage closed-loop control mechanism is adopted to perform response timeout judgment and compensation reminder triggering.

2. The AI ​​follow-up robot intelligent interaction method as described in claim 1, characterized in that, In step S10, the structured entity set E includes three entities. ,in, For the i-th identified entity item; , represents the entity type label corresponding to the i-th identified entity item; is the starting character position in the original text message P; n is the total number of entities.

3. The AI ​​follow-up robot intelligent interaction method as described in claim 1, characterized in that, In step S10, the final step, based on the candidate semantic window set W, further utilizes a dual-channel semantic discrimination mechanism to perform label classification and structured entity verification, outputting a structured entity set E. This step specifically includes: Static grammar channel stage: Extract syntactic dependency relations and contextual part-of-speech flow from the candidate semantic window set W. Based on the syntactic dependency relations and contextual part-of-speech flow, call the sequence model BiLSTM-CRF to determine and output the first entity category. and first confidence level; Dynamic Context Passage Stage: The pre-trained medical language model MedBERT is invoked to perform enhanced classification of the candidate semantic window set W based on the overall context, and the second entity category is output. Second confidence level; Based on the first entity category Second Entity Category The consistency score is obtained by using the semantic vector cosine similarity method: when the consistency score is greater than or equal to the preset consistency score threshold, the label is directly classified; when the consistency score is less than the preset consistency score threshold, the label is classified by confidence fusion method based on the first confidence score and the second confidence score.

4. The AI ​​follow-up robot intelligent interaction method as described in claim 1, characterized in that, In step S20, the preset emotion recognition network adopts a dual-channel semantic emotion fusion mechanism driven by medical entity perception. The emotion recognition network includes an input layer for receiving a set of structured entities E; a medical entity perception layer for using an entity attention mechanism to achieve semantic alignment between structured entities and context; and a semantic context fusion layer for using a context modeling method with a bidirectional gating mechanism to fuse the dependency and emotion propagation relationships between contexts. The multi-label emotion classification layer is used to map the dependency and emotion propagation relationships between the fused contexts to multiple emotion category spaces using a pre-defined output mechanism that combines the multi-label softmax function and the sigmoid function. These include anxiety category space, anger category space, complaint category space, low mood category space, neutral statement category space, and positive gratitude category space. The emotion consistency discrimination layer is used to perform a round of consistency verification between emotion expression and medical entity description before the final output; Output layer, used to output emotion polarity labels. and confidence level of emotional polarity .

5. The AI ​​follow-up robot intelligent interaction method as described in claim 1, characterized in that, In step S40, the patient's response status The result is obtained after sequentially determining the message passing status and the interaction response status, specifically including: Get ,based on Message passing status determination: ； in, This is the message passing status; The patient's response status is continuously monitored and output within a time window ΔT: ； in, The patient's response status.

6. The AI ​​follow-up robot intelligent interaction method as described in claim 1, characterized in that, In step S40, the preset patient response behavior prediction model is constructed using a hybrid deep prediction architecture based on joint modeling of temporal behavior embedding and emotional state. The patient response behavior prediction model specifically includes: The behavior embedding layer is used to receive and encode the historical interaction behavior sequence of the preset input, and uses a multi-channel embedding mechanism to process the historical interaction behavior sequence and output the historical behavior embedding matrix. The emotion state injection layer is used to receive the output historical behavior embedding matrix and emotion trigger source type. A position-aligned encoding mechanism is used to embed the output historical behavior matrix and the emotion trigger source type. Process the data to output a joint emotion-behavior vector sequence; The temporal modeling layer receives the joint emotion-behavior vector sequence, processes the joint emotion-behavior vector sequence using a bidirectional gated recurrent unit network (Bi-GRU), and outputs a deep semantic representation of the patient's behavioral trends. The doctor-patient interaction feature fusion layer is used to receive deep semantic expressions of patient behavior trends, obtain task interaction cues, and use an attention-weighted fusion mechanism to fuse the deep semantic expressions and task interaction cues, outputting a fused representation vector; among which, task interaction cues include "whether it includes a specifier" and "whether it is a secondary reminder"; The response duration regression prediction layer receives the fused representation vector, employs a multilayer perceptron, and outputs an individualized expected response time. Response confidence score and response time rating label.

7. An AI follow-up robot intelligent interaction system, applied to the AI ​​follow-up robot intelligent interaction method according to any one of claims 1 to 6, characterized in that, The AI ​​follow-up robot intelligent interaction system includes: The semantic parsing and entity extraction module is used to obtain the original text message P of the target patient in the preset chat platform follow-up group, construct a time window ΔT, and perform structured label extraction processing on the original text message P based on the knowledge-enhanced semantic parsing mechanism to output a structured entity set E. The emotion recognition and context construction module is used to construct a semantic context vector S based on a structured entity set E. The semantic context vector S is then input into a pre-defined emotion recognition network, and the module outputs emotion polarity labels. and confidence level of emotional polarity ; The emotion trigger source identification module is used to identify emotions based on emotion polarity labels. and confidence level of emotional polarity Perform a secondary classification task for emotion sources, determine and output the type of emotion trigger source. Among them, the types of emotion triggers Including subjective emotion expression types and medical symptom description types ; The response behavior modeling module is used to output the type of emotion trigger source. Then, obtain the message sending time of the original text message P. and patient response status ;Message sending time and patient response status Input into a pre-defined patient response behavior prediction model, and output individualized expected response time. ; The closed-loop judgment and compensation reminder module is used to determine the individualized expected response time. A multi-channel linkage closed-loop control mechanism is adopted to perform response timeout judgment and compensation reminder triggering.

8. An AI-powered follow-up robot intelligent interactive device, characterized in that, The AI ​​follow-up robot intelligent interaction device includes: a memory, a processor, and an AI follow-up robot intelligent interaction program stored in the memory and executable on the processor. When the AI ​​follow-up robot intelligent interaction program is executed by the processor, it implements an AI follow-up robot intelligent interaction method according to any one of claims 1 to 6.

9. A computer program product, characterized in that, The computer program product includes an AI follow-up robot intelligent interaction program, which, when executed by a processor, implements an AI follow-up robot intelligent interaction method according to any one of claims 1 to 6.

Images