Intelligent chat robot implementation method and system based on deep learning

Abstract

The invention relates to the technical field of man-machine interaction, and discloses an intelligent chat robot implementation method and system based on deep learning. The system comprises a knowledge reasoning module, a deep learning generation module, a dynamic fusion module, a strategy optimization module and an adaptive adjustment module. The knowledge reasoning module processes user input data and outputs an intention constraint value. The deep learning generation module processes the same user input data and outputs a response predicted value. And the dynamic fusion module receives the intention constraint value and the response prediction value, and generates a fusion response value after comprehensive processing. The policy optimization module formulates a dialog policy instruction based on the fused response value. And after executing the dialogue, the adaptive adjustment module analyzes the deviation between the feedback data and the fusion response value, and if the deviation exceeds a preset threshold value, the parameters of the deep learning generation module are updated online. The system can more accurately understand the intention of the user, generates a natural and smooth dialogue response conforming to knowledge constraints, and continuously improves the interaction quality through real-time learning.

Description

Technical Field

[0001] This invention relates to the field of human-computer interaction technology, specifically to a method and system for implementing intelligent chatbots based on deep learning. Background Technology

[0002] Intelligent chatbots, as an important interface for human-computer interaction, have been widely used in various fields such as customer service, information inquiry, health consultation, and smart home control. Traditional chatbot systems are usually based on rule bases or simple pattern matching, with limited dialogue capabilities, poor flexibility, and difficulty in handling open-domain or complex user queries.

[0003] With the development of deep learning technology, especially generative dialogue systems based on large language models (LLMs), significant progress has been made. These systems can generate more fluent and natural text responses. However, existing deep learning-based chatbots still face many challenges. First, while the content generated by the model is fluent, it may contain factual errors or be inconsistent with specific domain knowledge. The lack of reliable knowledge guidance and constraints leads to inaccurate answers or even misleading users. Second, the model's generation process is often a "black box," lacking accurate reasoning and understanding of the user's potential intentions, making it difficult to maintain logical consistency and goal orientation in complex multi-turn dialogues. Third, most systems use static models, whose knowledge is limited to the time point of the training data, lacking the ability to dynamically learn and update parameters based on user feedback in real-world interaction environments. Once deployed, their performance may degrade over time or when facing new scenarios, making it impossible to continuously optimize the user experience. Fourth, existing systems typically couple intent recognition, response generation, and strategy decision-making together, or use a simple chain approach, making it difficult to achieve an effective balance between intent constraints and generation freedom, and also unable to finely adjust interaction strategies according to the dialogue state.

[0004] While existing research has attempted to combine knowledge graphs or rules to guide generation, or to separate intent recognition and response generation, efficient and robust solutions are still lacking in areas such as how to deeply integrate the advantages of knowledge reasoning and generative models, achieve synergistic optimization of their outputs, and efficiently fine-tune the model based on real-time feedback during dialogue. This limitation is particularly evident in applications requiring high reliability, high accuracy, or continuous adaptation to new knowledge (such as health consultation, medication guidance, and elderly care), highlighting the urgent need for an intelligent dialogue system architecture that can more accurately understand intent, generate reliable responses, and possess continuous evolution capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based intelligent chatbot system to solve the problems mentioned in the background section.

[0006] To achieve the above objectives, the present invention provides a deep learning-based intelligent chatbot implementation system, the system comprising: The module includes a knowledge reasoning module, a deep learning generation module, a dynamic fusion module, a strategy optimization module, and an adaptive adjustment module. The knowledge reasoning module receives user input data and outputs intent constraint values; The deep learning generation module receives the user input data and outputs the response prediction value; The dynamic fusion module receives the intent constraint value and the response prediction value and generates a fused response value; The strategy optimization module generates dialogue strategy instructions based on the fused response value; When the adaptive adjustment module detects that the deviation between the feedback data after execution and the fused response value exceeds a preset threshold, it updates the parameters of the deep learning generation module online.

[0007] Preferably, the knowledge reasoning module includes a knowledge base unit and a rule engine unit; the knowledge base unit stores a domain knowledge graph; the rule engine unit sets intent boundary conditions based on the knowledge graph, and solves the intent constraint value through logical reasoning equations according to the intent boundary conditions.

[0008] Preferably, the deep learning generation module includes a temporal processing unit and a neural network unit; the temporal processing unit extracts contextual features from the user input data; and the neural network unit, based on the contextual features, outputs the response prediction value with the same dimension as the intent constraint value through a fully connected layer.

[0009] Preferably, the dynamic fusion module includes a weight calculation unit and a weighted fusion unit; the weight calculation unit calculates the weight coefficient γ of the knowledge reasoning module and the weight coefficient δ of the deep learning generation module through an attention mechanism based on the real-time dialogue error, where γ+δ=1; the weighted fusion unit performs a weighted summation of the intent constraint value and the response prediction value based on the weight coefficients γ and δ to generate the fused response value.

[0010] Preferably, the strategy optimization module includes a decision modeling unit and a parameter solving unit; the decision modeling unit constructs a state space, an action space, and a reward function based on a Markov decision process; the parameter solving unit updates the objective function based on the fused response value within the rolling time window, and solves for the current optimal dialogue action through the objective function to generate the dialogue strategy instruction.

[0011] Preferably, the adaptive adjustment module includes a deviation detection unit and a parameter update unit; the deviation detection unit calculates the actual response value based on the feedback data collected after executing the dialogue strategy instruction, using a quality assessment equation; the parameter update unit calculates the absolute value of the relative deviation between the actual response value and the fused response value, and triggers an online update when the absolute value exceeds a preset threshold.

[0012] Preferably, the parameter update unit freezes the parameters of the knowledge reasoning module and the hidden layer parameters of the deep learning generation module; calculates the gradient of the output layer parameters of the deep learning generation module based on the feedback data; and updates the output layer parameters according to the gradient until the absolute value of the relative deviation is lower than the preset threshold.

[0013] Preferably, the decision modeling unit includes a Critic evaluation module and an Actor execution module; the Critic evaluation module uses a neural network to model the current dialogue state and actions and outputs a decision value estimate; the Actor execution module uses a neural network to model the current dialogue state and outputs an action distribution, and generates the dialogue actions based on sampling and iterative updates of the action distribution.

[0014] Preferably, it further includes a state intersection detection module and a priority processing module; the state intersection detection module receives multi-user input data and detects the intent intersection parameters of each selected user combination; when the intent intersection parameters are detected, the priority processing module determines the priority corresponding to the intent intersection parameters and generates an environmental adjustment instruction based on the priority and parameter range.

[0015] The state intersection detection module traverses all user combinations in multiple detections; the priority processing module determines the priority based on the total historical interaction cost of each user combination; when the parameter value of the user input data is outside the parameter range corresponding to the high-priority intent intersection parameter, the environment adjustment instruction is output to the strategy optimization module.

[0016] Preferably, the present invention also includes a method for implementing a deep learning-based intelligent chatbot, comprising all the modules and method flows of the aforementioned deep learning-based intelligent chatbot system.

[0017] Compared with the prior art, the beneficial effects of the present invention are: This invention, by introducing a knowledge reasoning module, enables in-depth analysis of user input, outputting intent constraint values. This provides crucial semantic and logical boundaries for subsequent response generation, enhancing the system's depth and accuracy in understanding the user's true intent and avoiding irrelevant or off-topic responses that may occur with traditional generative models. The deep learning generation module, based on powerful language model capabilities, outputs fluent and diverse response predictions, maintaining the naturalness and flexibility of the dialogue. The dynamic fusion module is a key innovation. It effectively receives and integrates intent constraint values ​​from the knowledge reasoning module and response prediction values ​​from the generation module, generating a final fused response value through a specific fusion mechanism. This fusion mechanism ensures that the response conforms to knowledge specifications and user intent (guaranteed by the intent constraint values) while maximizing the preservation of the natural language characteristics of the generated response (provided by the response prediction values). This resolves the contradiction between accuracy and fluency in generated content, resulting in a final output that is both reliable and natural. The strategy optimization module further generates dialogue strategy instructions based on the fused response values, which can guide the system to perform interactive behaviors that are more in line with the context and user state, such as guiding the direction of the dialogue, taking the initiative to ask questions, or performing specific tasks, thereby improving the coherence and goal orientation of the dialogue.

[0018] The most prominent feature is the design of the adaptive adjustment module. During operation, the system continuously collects user feedback data on system responses. When a significant deviation (exceeding a preset threshold) is detected between the actual feedback and the fused response value generated internally by the system, this module triggers an online update process for the parameters of the deep learning generation module. This mechanism endows the system with powerful self-evolution capabilities. It does not require complete retraining of a massive model; instead, driven by real-world interactions, it selectively adjusts the parameters of the generation module, enabling it to quickly adapt to new expressions, new knowledge requirements, or changes in user preferences. This significantly overcomes the shortcomings of traditional static models, such as easy performance degradation and difficulty in adapting to new scenarios, allowing the chatbot to continuously learn from interactions and constantly improve the relevance of responses and user satisfaction. The entire system architecture is clear, and the modular design ensures that the responsibilities of each functional component are clearly defined, working collaboratively to achieve a comprehensive effect of more accurate intent understanding, more reliable response content, more intelligent dialogue processes, and sustainable improvement of system capabilities. This significantly optimizes the human-computer dialogue experience and is particularly suitable for application environments with high requirements for accuracy, reliability, and continuous learning capabilities. Attached Figure Description

[0019] Figure 1 This is a timing diagram of the deep learning-based intelligent chatbot implementation system described in this invention. Figure 2 This is a schematic diagram illustrating the working principle of the knowledge reasoning module. Figure 3 This is a schematic diagram illustrating the working principle of the strategy optimization module. Figure 4 This is a schematic diagram illustrating the working principle of the adaptive adjustment module. Detailed Implementation

[0020] 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.

[0021] Please see Figure 1 This invention provides a method and system for implementing an intelligent chatbot based on deep learning. Its core lies in achieving intent understanding, response generation, and strategy optimization through multi-module collaboration. The system includes a knowledge reasoning module, a deep learning generation module, a dynamic fusion module, a strategy optimization module, and an adaptive adjustment module.

[0022] User input data is first fed into the knowledge reasoning module and the deep learning generation module, which output intent constraint values ​​and response prediction values, respectively. The dynamic fusion module calculates weight coefficients based on real-time dialogue errors and generates a fused response value by weighted summation of the two. The strategy optimization module constructs a Markov decision model based on the fused response value to solve for the optimal dialogue action command. The adaptive adjustment module detects deviations in the fused response value through feedback data and triggers online parameter updates to optimize response quality. The implementation details of each module are described below through examples.

[0023] Example 1: See Figure 2The knowledge reasoning module loads the domain knowledge graph into the knowledge base unit upon system startup. This graph uses a resource description framework to store entity relationships, integrating heterogeneous data such as medication safety rules, equipment control logic, and health monitoring indicators. Graph nodes cover entities such as drug chemical properties, environmental equipment parameters, and physiological indicator thresholds. Edge relationships define contraindication combinations, control linkage rules, and health risk levels. The rule engine unit pre-sets intent boundary conditions, including a set of Boolean expressions for drug contraindications, threshold range definitions for environmental control instructions, and warning value determination logic for health monitoring. When user input data is passed to the module, the rule engine unit performs multi-stage logical reasoning: First, key entities are extracted through natural language parsing. For example, in the input statement "Can I drink grape juice after taking antihypertensive drugs?", "antihypertensive drugs" and "grape juice" are identified as drug entities and food entities, respectively. Then, the SPARQL query engine is launched to search the graph for contraindication edges such as "increases blood drug concentration" or "induces hypotension." Finally, a first-order predicate logic solver is used to assess the risk level and output a quantified intent constraint value. The constraint value is a multi-dimensional vector. For example, when a taboo relationship is detected, the output is [Taboo probability: 0.92, Risk level: 3], and when there is no taboo, the output is [Safety probability: 0.85]. This process strictly limits the response boundaries. For example, when a high taboo probability is detected, a safety warning response is forcibly locked.

[0024] The temporal processing unit of the deep learning generation module employs a bidirectional long short-term memory network architecture to process user input data. This unit maintains a dynamic dialogue state matrix, recording entity sequences, sentiment polarity labels, and time decay factors from historical dialogues. For the current input sentence, the network first performs word segmentation and embedding, mapping discrete words to a continuous semantic space through a word vector layer. Subsequently, a bidirectional LSTM layer extracts contextual features from both forward and backward dimensions. The forward layer captures the grammatical structure of the current sentence, while the backward layer fuses entity association information from the previous three rounds of dialogue. The feature fusion layer outputs a fixed-dimensional contextual feature vector. For example, when the input is "worsening cough at night," the feature vector includes [time period label: night, symptom entity: cough, historical association: history of chronic bronchitis, sentiment value: 0.7 anxiety]. The neural network unit receives this feature vector and performs a nonlinear transformation through a three-layer fully connected network: the first layer uses the ReLU activation function to expand the feature representation capability, the second layer performs dimensionality compression, and the last layer outputs a response prediction value that is strictly aligned with the dimension of the intent constraint value through a linear transformation. The predicted value is also a multi-dimensional vector. For example, for cough symptoms, it might output [Suggestion type: Medication reminder, Associated drug: Cough syrup, Confidence: 0.88]. The module design mandates that the output dimension be consistent with the intent constraint value to ensure mathematical operability in the subsequent fusion stage.

[0025] The two modules establish a parameter coordination mechanism during runtime. The output channel of the knowledge reasoning module shares a dimension verification interface with the feature output layer of the deep learning generation module. When a mismatch is detected between the dimension of the intent constraint value and the dimension of the response prediction value, a dimension realignment procedure is automatically triggered. This procedure maintains dimensional consistency by filling in default values ​​or truncating redundant dimensions. For example, if the knowledge module outputs a three-dimensional vector [0.6, 0.2, 0.9] while the generation module outputs a five-dimensional vector, the system automatically pads the knowledge module vector with zeros to expand it to five dimensions. An asynchronous execution thread pool is established between the modules, and user input data is simultaneously distributed to both modules for parallel processing. A timeout circuit breaker mechanism is set to control the maximum processing latency. When the knowledge reasoning module times out due to complex graph query, the system automatically downgrades to using cached historical constraint values; if the generation module times out, a lightweight text matching model is switched to generate the basic response. The output buffers of the two modules adopt a dual verification mechanism: first, numerical validity verification is performed (NaN values ​​are removed), and then normalization is performed to compress the values ​​of each dimension of the vector to the [0,1] range. Finally, the two modules output data packets containing intent constraint values ​​and response prediction values, along with processing timestamps and confidence metadata, which are then transmitted to the dynamic fusion module via a message queue.

[0026] The knowledge base unit implements a proactive update strategy. The system's backend service periodically crawls update logs from authoritative medical databases. When a change in drug contraindication data is detected, an incremental update of the knowledge graph is automatically triggered. The update process employs a two-phase commit protocol: first, the incremental data package is downloaded to a temporary storage area; then, after the rule engine unit verifies its logical consistency, it is atomically committed to the main graph. For rule defects reported by users, the system generates an anomaly report that triggers a manual review process. The rule engine unit maintains a version snapshot function, automatically saving historical version snapshots before each rule update and supporting rollback operations for erroneous rules. The deep learning generation module's model parameters undergo periodic incremental training. The training dataset comes from anonymized dialogue logs, and model version management uses a blue-green deployment strategy to ensure service continuity.

[0027] At the anomaly handling level, the knowledge reasoning module sets a graph query circuit breaker threshold. When a single query involves more than 50 entities or the relationship path depth exceeds 5 hops, an approximate query algorithm is automatically activated to compress query complexity through entity summarization technology. The deep learning generation module deploys an input filtering layer to clean or segment inputs containing malicious injection characters or excessively long statements (more than 256 characters). Both modules share an anomaly monitoring dashboard, displaying real-time resource utilization and processing success rate metrics. When the processing failure rate exceeds the threshold three times consecutively, the module restart process is automatically triggered.

[0028] Example 2: See Figure 3The weight calculation unit of the dynamic fusion module continuously monitors real-time dialogue error metrics, including the number of times the user actively interrupts the response, the frequency of manual correction operations, and semantic understanding failure logs. The error collector aggregates data in rounds, generating a normalized error value at the end of each round of dialogue. This value is mapped to the [0,1] interval through a linear transformation function. The weight coefficient generation mechanism adopts the attention mechanism principle, designing a trainable parameter matrix to project the error feature vector onto the weight space. In the specific execution process, the system initializes two weight coefficient generation channels: the knowledge reasoning module channel receives the error vector and outputs the initial weight value after transformation through a three-layer neural network; the deep learning generation module channel uses the same structure for parallel processing. Finally, the weight coefficient γ of the knowledge reasoning module is generated by constraining the Sigmoid function, while the weight coefficient δ of the deep learning generation module is automatically calculated by subtracting γ from 1, strictly maintaining the constraint condition γ+δ=1. The weight coefficients are globally updated every five minutes, but if the error mutation in a single round exceeds twice the historical average, an immediate recalculation is triggered.

[0029] The weighted fusion unit performs vector space alignment. After receiving the intent constraint value vector output by the knowledge reasoning module and the response prediction value vector output by the deep learning generation module, the unit first checks the dimensionality consistency between the two: if the dimensions do not match, a padding algorithm is called to pad the smaller dimension vector with zeros; if the dimension overflows the preset maximum value, PCA dimensionality reduction is initiated. During the fusion execution phase, an element-wise weighted summation algorithm is used to independently calculate the fusion value for each dimension of the two vectors. For example, if the first dimension of the intent constraint value is 0.8 and the corresponding dimension of the response prediction value is 0.5, when γ=0.6, the fusion value for that dimension is 0.6×0.8 + 0.4×0.5=0.68. The fusion result generates a multi-dimensional vector form of the fused response value, with source weight labels for each dimension. This module has a circuit breaker protection mechanism; when γ>0.9 for three consecutive rounds, it automatically switches to dual-channel independent output mode to prevent system stagnation due to a single module failure.

[0030] The decision modeling unit of the strategy optimization module constructs a Markov decision process framework. The state space is defined as a set of vectors of fused response values, with each state corresponding to a 256-dimensional feature vector. The action space pre-defines three types of basic operations: follow-up actions include 12 information completion templates, confirmation actions support 8 interactive verification forms, and execution actions are associated with 32 service interfaces such as medication reminders, environmental adjustment, and health monitoring. The reward function adopts a two-parameter weighted structure, with task completion quantified by the service call result code (e.g., 1 point for successful medication reminder delivery), and user satisfaction converted based on the sentiment analysis score of the voice evaluation (0.8 points for positive evaluation). The parameter solving unit maintains a rolling time window memory, which by default stores the fused response value sequence of the most recent five rounds of dialogue. The objective function updater calculates the state transition probability based on the time window data and uses a time difference algorithm to iteratively update the action value estimate. The solver uses an ε-greedy strategy to balance exploration and utilization: 90% probability of selecting the current optimal action, and 10% probability of randomly selecting a new action.

[0031] A three-tiered feedback loop is established during the action execution phase. Primary feedback collects system logs after the action is triggered, including API call latency and success status codes. Intermediate feedback obtains the user's voice evaluation text for that round of dialogue, outputting a satisfaction score through a sentiment analysis model. Advanced feedback monitors whether the user repeats the same request in the subsequent three rounds of dialogue. Feedback data is fed back to the reward function calculator in real time, dynamically adjusting the weighting of task completion and user satisfaction. For example, if the system detects that the user makes the same request again after the environmental adjustment command has been executed, the task completion weight for that action is automatically reduced.

[0032] In health management scenarios, when the fused response value indicates a user is at risk of falling (e.g., abnormal TOF radar data and postural feature vector exceeding a threshold), the strategy optimization module generates tiered response actions: First, a fall risk confirmation action is executed (outputting a voice prompt asking, "Are you feeling dizzy?"). If the user answers affirmatively, a second-level action is triggered (automatically dialing emergency contacts), and a third-level action is initiated (sending vital sign monitoring commands to the smart mattress). When multiple actions occur concurrently, an execution sequence manager is established, prioritizing actions and setting a minimum execution interval of 500 milliseconds.

[0033] The module is deployed with a disaster recovery mechanism. The decision modeling unit periodically snapshots policy network parameters to distributed storage. When three consecutive action selection conflicts are detected, it automatically rolls back to the most recent stable version. The parameter solving unit sets resource consumption thresholds. When a single solution takes more than 800 milliseconds or memory usage exceeds the limit, it automatically switches to a lightweight decision tree model to temporarily take over.

[0034] Example 3: See Figure 4The adaptive adjustment module's deviation detection unit establishes a multi-source feedback acquisition pipeline. User feedback data to dialogue strategy commands is acquired through the voice interaction interface. The raw audio stream, after noise reduction processing, is input into the speech recognition engine to transcribe it into text. The feedback classifier performs three levels of annotation on the text: semantic matching degree annotation identifies the relevance of the response content to the user's request (e.g., if the user says "the answer is irrelevant," the matching degree is 0.2); task resolution rate annotation judges whether the core requirement has been met (e.g., if the user says "the problem has been solved," the resolution rate is 1.0). The quality assessment equation is defined as: in Indicates the actual response value. This represents the normalized semantic matching score (0.0-1.0). This represents a quantitative value for the task completion rate (0.0 or 1.0). and These are dynamic weighting coefficients (initial values ​​0.6 / 0.4). The coefficient adjuster automatically adjusts the weights based on the feedback type: increasing the weight when the feedback contains emotional words. Weighting increases when task keywords are included. Weights.

[0035] The parameter update unit implements a deviation threshold determination mechanism. The system maintains real-time updated fusion response values. A cache queue stores the fusion values ​​from the last ten rounds of dialogue. The relative deviation calculator performs dual-channel verification: the main channel calculates the deviation for the current round. The auxiliary channel calculates the three-wheel moving average deviation. Update the trigger in Or three consecutive rounds The system activates the parameter update process. Once the trigger condition is met, the system generates a hierarchical optimization instruction: freezes all parameters of the knowledge reasoning module and the hidden layer parameters of the bidirectional LSTM layer in the deep learning generation module; and only exposes the weight matrix of the fully connected output layer of the generation module. Perform gradient calculation.

[0036] The gradient calculation engine employs a mini-batch sample processing mode. Each update randomly selects 50 sets of sample data from the historical dialogue database, each set containing user input. Target output and current model output The loss function constructor generates the objective equation: in For the generation module in the current parameters The output value is shown below. The gradient calculator solves the problem using automatic differentiation techniques. The actuator is updated using a momentum optimization algorithm to adjust parameters: in The adaptive learning rate is set to 0.01 initially. The momentum factor is a fixed value of 0.9. Record the last update direction. Each iteration processes a new batch, and when a validation set bias is detected... It may terminate when the upper limit of five iterations is reached.

[0037] A dual-safety mechanism is established during the update verification phase. The shadow model container copies the output layer parameters of the generation module at update startup, and all online requests during the update process are still handled by the original parameters. After the update, the verifier injects ten sets of standard test cases to compare the cosine similarity of the model output before and after the update. When the similarity is below 0.85 or a critical test case output is abnormal, the system automatically rolls back to the shadow parameters; if verification passes, the new parameters are switched to the production environment, and the shadow copy is deleted. For scenarios with consecutive update failures (three rollbacks), the system switches to a safe mode: the generation module degrades to using rule templates for response, and a manual diagnostic alert is triggered.

[0038] In a health management scenario, when a user reports an error in a medication reminder (e.g., "the aspirin dosage is stated as 50mg instead of 100mg"), the semantic matching score is labeled as 0.3, and the task resolution rate is labeled as 0.0. Assume... If it is 0.9, then The update process is triggered. The extracted historical samples contain drug dosage-related dialogues, and gradient calculation focuses on the dosage prediction neurons in the output layer. After the update, the model's output for the new input "aspirin dosage" is corrected from 100mg to 55mg, after two iterations. Updated to version 0.18.

[0039] Module deployment employs resource isolation strategies. Parameter update tasks are assigned to dedicated computing containers with a memory limit of 2GB and a single execution time not exceeding 15 seconds. Monitoring probes collect CPU utilization and memory usage data in real time; when resource limits are exceeded, updates are automatically paused and an optimization report is generated. The historical sample library implements a rolling update mechanism, retaining data from the most recent three months and cleaning up samples older than their expiration date monthly. The feedback data anonymization processor removes user identity information before storage, retaining only semantic features and annotation results.

[0040] Example 4: The decision modeling unit of the policy optimization module deploys a dual-network architecture consisting of a Critic evaluation module and an Actor execution module. The Critic evaluation module constructs a three-layer fully connected value network. The input layer receives the state feature vector transformed from the fused response values, and the output layer generates a scalar-form decision value estimate. This network is trained using a temporal difference learning mechanism, with a reward signal injected after each round of dialogue to adjust the network connection weights. The Actor execution module constructs a policy network. Its input layer shares the state feature vector with the Critic network, and its output layer generates a probability distribution of preset actions using a Softmax function. During network updates, a policy gradient algorithm is used to adjust the action probability weights based on the value estimate output by the Critic network.

[0041] The specific implementation process for the health management scenario is as follows: When the fused response value indicates that the user is at risk of hypertension (the feature vector contains [blood pressure reading: 182 / 110, time period: night, medication history: missed dose]), the Critic network receives this 256-dimensional feature vector. After the hidden layer of the network performs a nonlinear transformation, the output layer calculates the value score of each candidate action: Table 1: Data table of action value assessment under hypertension risk status. The Actor network receives the same status characteristics, and the probability distribution of the output actions is as follows: emergency medical reminder (42%), vital sign monitoring (28%), contacting the contact person (18%), guiding audio (7%), and appointment follow-up (5%). Based on the probability distribution, the system selects to execute the "send emergency medical reminder" action, while preparing "start vital sign monitoring equipment" as an alternative action in parallel.

[0042] A three-tiered collaborative mechanism is established during the action execution phase: the main execution thread drives the selected core action (calling the community emergency center), the auxiliary thread prepares secondary action resources (such as pre-loading blood pressure monitor control instructions), and the monitoring thread tracks user feedback in real time. When the user refuses the medical advice (feedback "no need to go to the hospital"), the system switches to the second-ranked action, "start vital signs monitoring device," within 300 milliseconds, activating the smart mattress's continuous blood pressure monitoring mode via the IoT interface.

[0043] The network update process employs an offline training mechanism. After accumulating 100 rounds of dialogue logs, the system initiates a dual-network collaborative update: the Critic network adjusts its value assessment model based on actual reward values ​​(e.g., high scores for user acceptance of suggestions); the Actor network corrects the action probability distribution based on the value estimates provided by the Critic. The training dataset undergoes time window slicing, with data from the most recent three days having three times the weight of the old data. After the update, the policy network optimizes the action distribution for the "refusal to seek medical treatment" scenario as follows: vital sign monitoring (51%), contacting the contact person (29%), scheduling a follow-up appointment (15%), and guiding audio (5%).

[0044] The fault tolerance mechanism includes an action conflict detector. When two actions output by the Actor network are mutually exclusive (such as "turn off all lights" and "turn on reading light"), the conflict detector calls a predefined rule base to identify conflict patterns and automatically suppresses the execution probability of low-value actions. Network parameter version management adopts an A / B testing model: new strategies are deployed to 10% of the user group, and are only fully launched after key indicators (such as user secondary request rate) compared with the original network meet the standards. The system retains snapshots of the three most recent stable parameters, and automatically reverts to the previous version when a new strategy causes five consecutive action execution failures.

[0045] In an age-friendly home control scenario, when the user intent analyzed by the fusion response value is "environment adjustment" (feature vector includes [voice command: "too bright", ambient light sensor data: 1500 lux, historical preference: preference 300 lux]), the Critic network evaluates "dimming the main light" as having the highest value (8.5 points), and the Actor network outputs this action with a probability of 67%. During execution, the brightness of the living room main light is adjusted from 70% to 30% via the smart home protocol, and a satisfaction feedback signal (the user says "this is much more comfortable") is injected after the action is completed to reinforce the strategy.

[0046] Example 5: The state intersection detection module deploys a distributed instruction acquisition architecture. The system synchronously receives input data from multiple users through multi-channel audio separation technology, with each user allocated an independent audio stream channel and speech recognition engine. The recognition result generates a structured intent description object, containing a triplet of core verb, target entity, and modifying parameters. The module initiates a combination traversal algorithm: the first detection selects user A and user B as a combination, extracts their intent vectors, and calculates similarity; the second detection cycles to the combination of user A and user C, until all possible combinations are traversed. The intent vector converter maps the triplet to a 128-dimensional semantic space, and the similarity calculation uses an improved cosine algorithm. When the similarity between combinations exceeds a preset threshold of 0.65, intent intersection parameters are generated. This parameter is recorded as a quadruple structure. For example, when user A's request for "temperature 26℃" and user B's request for "lower room temperature" form an intersection, it is marked as [Intersection ID: TC001, Entity type: temperature, Operation direction: cooling, Parameter range: 24-26℃].

[0047] The priority processing module constructs an interaction cost assessment model. The model maintains a historical interaction record table for each user combination, with the data structure containing five core metrics: average response time (in milliseconds), action error count, user correction frequency, second request rate, and session abandonment rate. The total historical interaction cost is calculated using a weighted summation algorithm: Combination cost = Response time × 0.3 + Error count × 0.25 + Correction frequency × 0.2 + Second request rate × 0.15 + Abandonment rate × 0.1. The priority sorter arranges combinations in ascending order of cost value, assigning the lowest cost combination the highest priority. When a new intent intersection parameter is detected, the system retrieves the priority tag for that combination and associates it with the parameter range.

[0048] The environmental adjustment instruction generator implements a triple verification mechanism. Upon receiving real-time user input data: First, it parses the current parameter value; for example, user C's new instruction "set 28℃" extracts the target value 28. Second, it matches the intent intersection parameter library, searching for temperature-related intersection parameters (such as TC001) for this combination. Finally, it determines the inclusion relationship between the current parameter value and the intersection parameter range. When it detects that the target value exceeds the range of high-priority parameters (e.g., the priority 1 TC001 range is 24-26℃, and the input value 28 exceeds the upper limit), it triggers the generation of an environmental adjustment instruction. The instruction encoding uses JSON-LD format and includes a conflict resolution field, for example, the generated instruction is {"action":"set_temperature", "value":26, "reason":"priority_override", "source_intersection":"TC001"}.

[0049] Typical process in multi-user health management scenario: User A (a 72-year-old patient with hypertension) made a voice request to "keep the bedroom warm"; User B (family member) also instructed, "The elderly person is afraid of feeling stuffy, so we need to ventilate the room." The state detection module calculates the intent vector similarity to 0.78 and generates the intersection parameters [ID:AC001, Entity: Bedroom environment, Operation: Balance temperature and humidity, Range: Temperature 22-24℃ / Humidity 40%~60%].

[0050] Priority module retrieval history: AB combination average response time 1800ms (cost value 54), AC combination 3500ms (cost value 105), giving AB higher priority.

[0051] User C's (caregiver's) new instruction "dehumidify to 35%" was detected: The humidity value of 35% is below the lower limit of the intersection parameter of 40%. Trigger the environment adjustment command {"device":"bedroom_humidifier", "action":"adjust", "value":40, "constraint_range":"40-60%"}.

[0052] The system deploys an intent conflict early warning mechanism. When the same intersection parameter is frequently violated (more than 5 times) for three consecutive minutes, a coordination dialogue mode is automatically initiated: a voice explanation is sent to low-priority users (e.g., "According to health management requirements, maintaining the current temperature at 26℃ is the safest"), while a confirmation request is pushed to high-priority users (e.g., "Do you agree to adjust to 25℃?"). The coordination result updates the original intersection parameter range, forming a dynamically adaptable parameter boundary.

[0053] The data storage layer adopts a time-series database architecture. Interaction metrics for all user combinations are stored at a minute-level granularity, supporting sliding window aggregation queries. The intent intersection parameter library implements versioned management, generating a new version record with each range update while retaining historical versions. Environmental adjustment instruction logs are stored in association with original audio clips, establishing a traceable decision-making chain.

[0054] The fault tolerance mechanism includes priority failure protection. When a high-priority combination is detected to have consecutive offline timeouts (e.g., user A's device disconnects), the system automatically downgrades to the parameter range of the next lower priority combination. In extreme conflict scenarios (where all parameter ranges are mutually exclusive), a safety threshold enforcement mode is activated: the environmental standard values ​​specified in the medical protocol (e.g., room temperature 24℃, humidity 45%) are used as the default adjustment benchmark.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0056] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based intelligent chat robot system, characterized by, The application relates to a dialogue management method and system. The application comprises a knowledge reasoning module, a deep learning generation module, a dynamic fusion module, a strategy optimization module and an adaptive adjustment module. The knowledge reasoning module receives user input data and outputs an intent constraint value. The deep learning generation module receives the user input data and outputs a response prediction value. The dynamic fusion module receives the intent constraint value and the response prediction value and generates a fusion response value. The strategy optimization module generates a dialogue strategy instruction based on the fusion response value. The adaptive adjustment module updates parameters of the deep learning generation module online when a deviation between feedback data after execution and the fusion response value exceeds a preset threshold. 2.The deep learning-based intelligent chat robot system according to claim 1, wherein, The knowledge reasoning module comprises a knowledge base unit and a rule engine unit; the knowledge base unit stores a field knowledge graph; the rule engine unit sets an intent boundary condition based on the knowledge graph and solves the intent constraint value through a logical reasoning equation according to the intent boundary condition. 3.The deep learning-based intelligent chat robot system according to claim 1, wherein, The deep learning generation module comprises a time sequence processing unit and a neural network unit; the time sequence processing unit extracts context features of the user input data; the neural network unit outputs the response prediction value with the same dimension as the intent constraint value through a full connection layer based on the context features. 4.The deep learning-based intelligent chat robot system according to claim 1, wherein, The dynamic fusion module comprises a weight calculation unit and a weighted fusion unit; the weight calculation unit calculates a weight coefficient gamma of the knowledge reasoning module and a weight coefficient delta of the deep learning generation module through an attention mechanism according to real-time dialogue errors, wherein gamma + delta = 1; the weighted fusion unit performs weighted summation on the intent constraint value and the response prediction value based on the weight coefficients gamma and delta to generate the fusion response value. 5.The deep learning based intelligent chat robot system according to claim 1, wherein, The strategy optimization module comprises a decision modeling unit and a parameter solving unit; the decision modeling unit constructs a state space, an action space and a reward function based on a Markov decision process; the parameter solving unit updates an objective function based on the fusion response value in a rolling time window and solves a current optimal dialogue action through the objective function to generate the dialogue strategy instruction. 6.The deep learning-based intelligent chat robot system according to claim 1, wherein, The adaptive adjustment module comprises a deviation detection unit and a parameter updating unit; the deviation detection unit calculates a response actual value through a quality evaluation equation based on feedback data collected after the dialogue strategy instruction is executed; the parameter updating unit calculates a relative deviation absolute value of the response actual value and the fusion response value and triggers online updating when the absolute value exceeds a preset threshold. 7.The deep learning based intelligent chat robot system according to claim 6, wherein, The parameter updating unit freezes parameters of the knowledge reasoning module and hidden layer parameters of the deep learning generation module; calculates a gradient of output layer parameters of the deep learning generation module based on the feedback data; updates the output layer parameters according to the gradient until the relative deviation absolute value is lower than the preset threshold. 8.The deep learning based intelligent chat robot system according to claim 5, wherein, The decision modeling unit comprises a Critic evaluation module and an Actor execution module; the Critic evaluation module models the input of the current dialogue state and action through a neural network and outputs a decision value estimate; the Actor execution module models the input of the current dialogue state through a neural network and outputs an action distribution, samples and iteratively updates the dialogue action based on the action distribution. 9.The deep learning based intelligent chat robot system according to claim 1, wherein, Further comprising a state intersection detection module and a priority processing module; the state intersection detection module receives multi-user input data, detects the intention intersection parameter of each selected user combination; the priority processing module determines the priority corresponding to the intention intersection parameter when detecting the intention intersection parameter, and generates an environment adjustment instruction based on the priority and the parameter range.

10. The state intersection detection module traverses all user combinations in multiple detections; the priority processing module determines the priority according to the total historical interaction cost of each user combination; when the parameter value of the user input data is outside the parameter range corresponding to the intention intersection parameter of the high priority, the environment adjustment instruction is output to the policy optimization module.

11. A deep learning-based intelligent chat robot implementation method, characterized in that, All modules and method processes of the intelligent chat robot system based on deep learning according to any one of claims 1 to 9.

Images