Artificial intelligence assistant method and system based on PAD emotion driving

Abstract

The invention discloses an artificial intelligence assistant method and system based on PAD emotion driving, and the method comprises the steps: collecting multi-modal data of a user, and carrying out the preprocessing of the multi-modal data; identifying the current emotion and communication intention of the user according to the preprocessed data; personality parameters and internal emotional states of the AI assistant are dynamically updated, and personification and consistency are kept in long-term interaction; long-term interactive memory of the AI is managed, and personality consistency is kept; generating a personalized recommendation and guide strategy for the user based on the personality and emotional state; generating an anthropomorphic natural language; verifying the key facts by using a knowledge graph; and performing state updating and guiding feedback. And the recommendation result not only is matched with the user behavior, but also better conforms to the evolution style and the guide target of the AI assistant, so that the acceptance and the exploration intention of the user to the recommendation are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, and in particular to a PAD emotion-driven artificial intelligence assistant method and system. BACKGROUND

[0002] The existing AI assistant usually adopts the following solutions: Personality setting: usually adopts static, preset prompt words or fixed role setting based on rule set switching.

[0003] Emotion perception: using a two-dimensional emotion model of Valence Arousal (V, A) to simply classify user emotions.

[0004] Memory management: mainly relying on vector database retrieval (RAG) or fixed-length context window for short-term dialogue memory management, and some using simple fixed label priority for memory classification.

[0005] Recommendation and guidance: mainly based on collaborative filtering, user portrait fixed weight or historical behavior data for commodity / content recommendation, usually presented in the form of pop-up windows or hard prompts.

[0006] The existing technology has the following technical problems: Personality rigidity and lack of empathy: the personality of the existing AI is fixed, and it lacks the ability to continuously and smoothly evolve according to long-term interaction and emotional feedback of the user. The V, A emotion model cannot distinguish complex emotions with different dominance, such as anxiety (low control) and anger (high control), resulting in a lack of deep empathy and real long-term companionship in AI interaction.

[0007] Memory explosion and personality drift: long-term interaction data accumulates, and the fixed-length context window causes key historical information to be submerged (memory explosion). There is a lack of mechanism for dynamically evaluating the long-term value of memory (such as the contribution to the evolution of AI personality), resulting in the AI's behavior and style gradually deviating (personality drift) in long-term use.

[0008] Guidance is mechanical and weight is fixed: the weighting function of the recommendation / guidance system is fixed and cannot be coupled with the personality preference evolved by the AI itself. The timing and method of recommendation is harsh, and there is a lack of natural and subtle guidance according to the user's fine emotional state and AI personality style, affecting user experience and acceptance. SUMMARY

[0009] In view of the above problems, the present application is proposed to provide a PAD emotion-driven artificial intelligence assistant method and system to overcome the above problems or at least partially solve the above problems.

[0010] According to one aspect of the present application, a PAD emotion-driven artificial intelligence assistant method is provided, which comprises: Collecting multi-modal data of the user and preprocessing the same; Identifying the current emotion and communication intention of the user according to the preprocessed data; Dynamically updating the personality parameters and internal emotional state of the AI assistant to maintain personification and consistency in long-term interaction; Managing the long-term interaction memory of the AI and maintaining personality consistency; Generating personalized recommendation and guidance strategies for the user based on the personality and emotional state; Generating personified natural language; Verifying key facts using a knowledge graph; Performing state updating and guidance feedback.

[0011] Optionally, the collecting multi-modal data of the user and preprocessing the same specifically comprises: Obtaining multi-modal input of the user, and the multi-modal data includes text, voice, image, and sensor data; Data preprocessing, including data cleaning and data standardization; Using a multi-modal encoder to uniformly convert the input into a semantic vector for subsequent emotion and intention analysis.

[0012] Optionally, the identifying the current emotion and communication intention of the user according to the preprocessed data specifically comprises: Using a large language model (LLM) or a lightweight special-purpose model to analyze ; a user emotion vector ; a user intention vector , including asking questions, seeking help, and comforting needs.

[0013] Optionally, the dynamically updating the personality parameters and internal emotional state of the AI assistant specifically comprises: Collecting current personality parameters , AI old emotions , user emotion vectors , and user communication intentions ; Executing an evolution function to output new personality parameters and new emotions .

[0014] Using a smoothing algorithm to suppress sharp fluctuations; Emotion-personality evolution function new personality parameters The personality parameter increment calculation is configured to adjust based on the PAD emotion and the AI's own emotion, to realize continuous, smooth and adaptive evolution of the AI personality; The personality parameter is converted into style description words and tone modifiers through dynamic mapping, and is integrated into the prompt words through a dynamic description injection mechanism to affect the interactive style of the LLM output.

[0015] Optionally, the management of the long-term interaction memory of the AI and the maintenance of personality consistency specifically includes: Hierarchical structure: recent memory IM, semantic memory STM, emotional memory EM, meta-memory Meta, and core memory bank; Memory distillation: assign a dynamic importance score to each memory When the threshold is reached, low-score memories are deleted with a probability, and high-score memories are abstracted as personality anchors; Retrieval: retrieve relevant historical information for dialogue generation according to the current context and ; Each memory is assigned a dynamic importance score DIS(t), which is determined by an initial weight and a forgetting function based on time decay; The memory distillation engine is configured to periodically perform a random cleaning task, and the deletion probability is a nonlinear inverse function of ; Where γ is the forgetting factor; The distillation engine will be configured to extract long-term high-DIS patterns and stable components of the AI personality parameters, abstract them as personality anchors and store them in the core memory bank, to support the long-term consistency of the AI personality.

[0016] Optionally, the generation of personalized recommendation and guidance strategies for users based on personality and emotional state specifically includes: Final score of recommended items A multi-dimensional weighting function is used, where the weight of each dimension is a function of the personality parameters evolved by the AI teacher ; Achieving dynamic coupling of AI personality to recommendation strategy; The novelty / diversity regularization term is introduced into the scoring function ; λ is the penalty coefficient; is the novelty score of the candidate object; The guidance strategy is triggered by combining and Trigger the optimal guidance mode, and output through natural dialogue embedding.

[0017] Optionally, the generation of natural language personification specifically includes:​ Prompt factory: will Convert to style description words, combine memory abstracts and recommendation strategies, and build structured prompts.

[0018] Call large model to generate response text Response.

[0019] The application also provides a PAD emotion-driven artificial intelligence assistant system based on the PAD emotion-driven artificial intelligence assistant method described above, and the intelligent assistant system specifically includes: An input and vectorization module for collecting multi-modal data of a user and pre-processing; An emotion and intention analysis module for identifying the current emotion and communication intention of the user according to the pre-processed data; A personality adaptive evolution module for dynamically updating the personality parameters and internal emotional state of the AI assistant, and maintaining personification and consistency in long-term interaction; A memory layering and retrieval module for managing the long-term interaction memory of the AI and maintaining personality consistency; A guidance and recommendation decision module for generating personalized recommendations and guidance strategies for the user based on personality and emotional state; A generation and personality style injection module for generating personified natural language; A fact verification and output module for verifying key facts using a knowledge graph; A state update and guidance feedback module for state update and guidance feedback.

[0020] The PAD emotion-driven artificial intelligence assistant method and system provided by the application, the artificial intelligence assistant method includes: collecting multi-modal data of a user and pre-processing; identifying the current emotion and communication intention of the user according to the pre-processed data; dynamically updating the personality parameters and internal emotional state of the AI assistant, and maintaining personification and consistency in long-term interaction; managing the long-term interaction memory of the AI and maintaining personality consistency; generating personalized recommendations and guidance strategies for the user based on personality and emotional state; generating personified natural language; verifying key facts using a knowledge graph; and performing state update and guidance feedback. The personalized guidance realizes that the recommendation result not only matches the user behavior, but also conforms to the evolution style and guidance target of the AI assistant, significantly improving the acceptance and exploration willingness of the user to the recommendation.

[0021] The above description is only a summary of the technical solutions of the application, in order to more clearly understand the technical means of the application, the application can be implemented according to the content of the specification, and in order to make the above and other purposes, features and advantages of the application more obvious and easy to understand, the following specific embodiments of the application are described. BRIEF DESCRIPTION OF DRAWINGS

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0023] Figure 1 A flowchart illustrating a PAD-based emotion-driven AI assistant method provided in an embodiment of the present invention; Figure 2 A block diagram illustrating the composition of a PAD-based emotion-driven artificial intelligence assistant system provided in an embodiment of the present invention; Figure 3 A block diagram of the memory distillation system based on dynamic importance scoring and decay curves provided in this embodiment of the invention; Figure 4 A flowchart for guidance and recommendation based on dynamic weights of personality parameters provided in an embodiment of the present invention; Figure 5 A detailed flowchart of guidance and recommendation based on dynamic weights of personality parameters provided for embodiments of the present invention; Figure 6 A detailed flowchart of a PAD-based emotion-driven AI assistant method is provided for an embodiment of the present invention. Detailed Implementation

[0024] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0025] The terms "comprising" and "having," and any variations thereof, in the specification, embodiments, claims, and drawings of this invention are intended to cover non-exclusive inclusion, such as including a series of steps or units.

[0026] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0027] like Figure 1 and Figure 6 As shown, a method and system for a human-like, long-term companion AI assistant has the following workflow: 1. Input and Vectorization Module Function: Receive and preprocess (clean, standardize) user's multi-modal input (text, speech, image, sensor data).

[0028] Implementation Method: Use multi-modal encoders (e.g. CLIP, Wav2Vec2, BERT) to convert inputs into semantic vectors for subsequent sentiment and intent analysis.

[0029] 2. Sentiment and Intent Analysis Module Function: Identify user's current emotions and communication intent.

[0030] Implementation Method: Use large language models (LLM) or lightweight specialized models for analysis .

[0031] Output: User emotion vector User intent vector (e.g. asking questions, seeking help, comfort needs, etc.).

[0032] Personality Adaptive Evolution Module (Innovation Point One) Function: Dynamically update AI assistant's personality parameters and internal emotional state to maintain personification and consistency in long-term interactions.

[0033] Implementation Method: Input: (Current personality parameters), (AI old emotions), (User emotion vector), (User communication intent) Execute evolution function , output new personality parameters and new emotions .

[0034] Use smoothing algorithms (Kalman filter or moving average) to suppress sharp fluctuations.

[0035] Memory Layering and Retrieval Module (Innovation Point Two) Function: Manage AI's long-term interaction memory, avoid memory explosion, and maintain personality consistency.

[0036] Implementation Method: Hierarchical structure: Recent memory (IM), semantic memory (STM), emotional memory (EM), meta-memory (Meta), core memory bank (personality anchor).

[0037] Memory distillation: Assign a dynamic importance score to each memory .

[0038] When threshold reached, low-score memories are deleted by probability, high-score memories are abstracted as personality anchors.

[0039] Retrieval: relevant historical information is retrieved for dialogue generation based on current context and .

[0040] Guidance and recommendation module Function: generate personalized recommendation and guidance strategies for users based on personality and emotional states.

[0041] Implementation method: User profile: integrate cognitive, emotional, and habit dimensions.

[0042] Dynamic weighting: use personality parameters to calculate recommendation dimension weights .

[0043] Calculate candidate content score: Formula item-by-item explanation: Represents a certain candidate object (learning materials, books, courses, recommended targets).

[0044] The goal of this function is to calculate the recommendation score of the candidate object.

[0045] Represents the score of the candidate object in a certain dimension k.

[0046] Dimensions include: difficulty adaptation, content relevance, interest fit, emotional matching.

[0047] Represents the corresponding weight, which depends on the user's personality parameter P.

[0048] Different user personality parameters will result in different weight distribution: For example, "introverted personality" pays more attention to difficulty adaptation and comfort; "extroverted personality" pays more attention to sociality and novelty.

[0049] Effect: make the calculation results of the recommendation system vary due to individual differences of users.

[0050] The weighted sum of each dimension score and weight.

[0051] Meaning: consider multiple dimensions comprehensively, embody "weighted total score".

[0052] Sum of weights, used as normalization factor.

[0053] Significance: Ensures that the result is a "weighted average score" and does not become skewed due to overall weight increase.

[0054] Example: If the user's preference weights sum to 10 or 100, the final result can be standardized to a comparable range. Nature: Weighted average performance of the candidate object across multiple dimensions.

[0055] Embodiment: The recommendation system measures the overall suitability of the candidate object based on the preferences assigned by the personality parameters. Novelty : Novelty score of the candidate object (the higher the value, the more novel it is).

[0056] λ: Penalty coefficient, controls the influence of novelty in the recommendation score.

[0057] Logic: Objects with too high novelty may not be suitable for the current user, so they need to be "penalized" to avoid excessive recommendation of unfamiliar, unstable or deviating from learning goals.

[0058] Strategy selection: Combine with Trigger the optimal guidance mode (challenge, comfort, recommended resources, etc.).

[0059] Generate and inject personality style module Function: Generate natural language output with personification.

[0060] Implementation method: Prompt word factory: Convert into style description words, combine with memory summary and recommendation strategy to build structured prompts.

[0061] Agent execution: Call large model to generate response text Response.

[0062] Fact verification and output module Function: Ensure the reliability and personality consistency of the output.

[0063] Implementation method: Use knowledge graph (such as Neo4j) to verify key facts.

[0064] Virtual character engine gives tone, speed, expression modification according to and .

[0065] State update and guidance feedback module Function: Maintain system internal state and user profile.

[0066] Implementation method: Write the current interaction result into the hierarchical memory system.

[0067] According to the user state and emotion, automatically trigger the next round of guidance strategy.

[0068] As Figure 2 shown, the AI personality adaptive closed-loop evolution method based on PAD emotion driving includes: Through a continuous and regulated closed-loop mechanism, the personality parameters P of the AI assistant can respond to the fine emotions of the user, realizing the evolution of the style of personification.

[0069] Personality modeling and extension: Adopt a six-dimensional continuous parameter space P=[OCEAN, Pempathy]. Among them, the OCEAN dimension (Openness experience openness, conscientiousness, extraversion, agreeableness, neuroticism) is used to describe the basic personality tendency; The Pempathy empathy dimension added outside OCEAN is used to define the unique guidance style of AI as a long-term assistant.

[0070] Fine emotional feedback: Adopt a three-dimensional PAD emotion model [Pleasure, Arousal, Dominance]. The Dominance (dominance) dimension is used to capture the user's sense of control and self-confidence in the interaction. For example, identify whether the user is anxious (low D) or arguing (high D), providing fine input for subsequent personality evolution.

[0071] Nonlinear evolution function: Emotion-personality evolution function Receive user emotions , user intent and the previous emotion of AI itself as input.

[0072] Function variable description New personality parameter vector (PersonalityParameters, Updated) Old personality parameter vector (PersonalityParameters, Previous) α Learning rate / regulation coefficient (LearningRate) f(⋅) UpdateFunction UserPADState UserIntentRepresentation Environment / EpisodeState,Previous Function implemented with Multilayer Perceptron or small Transformer, where has a regulatory effect on the increment ΔP, so that the personality evolution of AI not only depends on the user, but also depends on the processing state of AI itself.

[0073] Dynamic Style Injection: Updated Continuous Personality Parameters Not as simple string concatenation. Instead, it is transformed by a dynamic mapping module into a set of style descriptors and tone modifiers (such as "affectionate, patient, slightly excited"), which are injected into the system instructions and examples of LLM through a dynamic description injection mechanism, thereby realizing the migration of underlying styles and generating highly personalized replies.

[0074] The AI assistant has the ability of "emotion-driven personality continuous evolution", avoiding personality rigidity and realizing long-term, emotionally deep interaction.

[0075] As shown in Figure 3 , based on the dynamic importance score (DynamicImportanceScore, hereinafter abbreviated as DIS) and the memory distillation method of decay curve, including: A mechanism for long-term memory distillation to solve the problems of Memory Explosion and Personality Drift in long-term interaction. This mechanism consists of three key steps:

[0076] Dynamic Importance Score (DIS): Each memory Mi is assigned a dynamic score DIS(t) instead of a fixed label. DIS(t) is determined by the following factors: Initial Importance Score (IIS): The IIS weight not only contains the emotional intensity (user PAD fluctuation) and the repetition frequency, but also contains the AI personality evolution contribution (|| norm), that is, the events that have a greater impact on AI itself have a higher IIS. − || norm), that is, the events that have a greater impact on AI itself have a higher IIS.

[0077] Decay function: Following a time-based non-linear decay model (Ebbinghaus Simplified Curve): where is the memory retention constant and is positively correlated.

[0078] : The i-th memory.

[0079] t: Current time.

[0080] : The probability of a memory being stored at a certain timestamp.

[0081] (InitialImportanceScore): The initial importance score is calculated by weighting the following factors: User emotional intensity (based on PAD model fluctuation); The repetition frequency of interaction events; The contribution to the evolution of AI personality (|| of − ||'s norm).

[0082] : The memory retention constant is positively correlated with , indicating that important memories decay more slowly.

[0083] Non-linear random pruning: When the total amount of memory reaches a threshold, a distillation task is started. The probability of memory deletion is a non-linear inverse function of DIS, calibrated by the forgetting factor γ: : The probability of a memory being deleted.

[0084] : The current importance score of a memory.

[0085] : The highest importance score in the current memory set. C: Normalization coefficient, ensuring 0 ≤

[0086] ≤ 1.

[0087] γ > 1: Forgetting factor, determines the accelerated elimination rate of low-importance memories.

[0088] ​Where γ>1, ensures that the lower the DIS, the faster the probability of the memory being deleted increases, while high DIS memories still have a very small probability of random forgetting, highly simulating the randomness of human forgetting.

[0089] Memory abstraction and personality anchor: The memory distillation engine periodically abstracts the long-term co-occurrence patterns of high DIS and the stable components of AI personality parameters into semantic vectors as "personality anchors" and stores them in the core memory pool (CoreMemoryPool). Core memory is used to stabilize the long-term personality characteristics and behavior prediction of AI.

[0090] Personality anchors ensure that the core personality characteristics of AI remain stable in long-term interaction. The core memory pool is decoupled from the temporary memory pool, avoiding short-term fluctuations that may cause deviations in long-term personality modeling. In the process of new interaction, the system prioritizes response generation based on personality anchors, improving behavior consistency and predictability.

[0091] As shown in Figure 4 and Figure 5 , the guidance and recommendation method based on dynamic weight of personality parameters specifically includes: Integrate the personality preferences and fine emotional states of the evolved AI into the recommendation weighting function to achieve personalized and stealthy guidance.

[0092] Dynamic weight function: The final score of the recommended item is calculated using a multi-dimensional weighting function, where the weight of each dimension is a function of the personality parameters P of the current evolved AI assistant The formula is explained item by item: represents a candidate object (learning material, book, course, recommended target).

[0093] The goal of the function is to calculate the recommendation score of the candidate object.

[0094] represents the score of the candidate object in dimension k.

[0095] The dimensions include: difficulty adaptation, content relevance, interest fit, and emotional matching.

[0096] represents the corresponding weight, which depends on the personality parameters P of the user.

[0097] Different user personality parameters will result in different weight distribution: For example, "introverted personality" places more emphasis on difficulty adaptation and comfort. ​"Extraversion" values sociality and novelty more.

[0098] Effect: The calculation result of the recommendation system varies due to individual differences in users.

[0099] The weighted sum of scores of each dimension and weights.

[0100] Significance: Consider multiple dimensions comprehensively, embody "weighted total score".

[0101] The sum of weights, used as a normalization factor.

[0102] Significance: Ensure that the result is a "weighted average score", and that the result will not be biased high due to the overall increase in weights.

[0103] Example: If the sum of the user's preference weights is 10 or 100, the final result can be standardized to a comparable range. Nature: The weighted average performance of the candidate object in multiple dimensions.

[0104] Embodiment: The recommendation system measures the overall adaptability of the candidate object according to the preferences assigned by the personality parameters. Novelty : Novelty score of the candidate object (the larger the value, the more novel).

[0105] λ: Penalty coefficient, controls the influence of novelty in the recommendation score.

[0106] Logic: Objects with too high novelty may not be suitable for the current user, so they need to be "penalized" to avoid excessive recommendation of unfamiliar, unstable, or deviated content from the learning goal.

[0107] Dynamic coupling: For example, if the AI's (diligence) is high, it will increase the weight WM(P) of the matching degree SM with the user's long-term goal (from distilled memory), making the recommendation more planned.

[0108] Novelty regularization term: Introduce a novelty / diversity regularization term in the scoring function . This term ensures that the system encourages exploration when recommending efficient tools or expanding resources, consistent with the AI assistant's role as a guide.

[0109] PAD state triggers and natural embedding: the recommendation timing is triggered by fine PAD emotional state, such as: triggering "challenge motivation" in high D (high control), triggering "empathy pacification + recommended simplified tool" in low P / low D (frustration, helplessness). All recommended content is output through personalized language in natural dialogue embedding mode, avoiding harsh interference on users.

[0110] Technical effects: realize "personalized guidance", the recommended results not only match the user behavior, but also meet the evolution style and guidance target of AI assistant, significantly improve the acceptance and exploration willingness of users to the recommendations.

[0111] Beneficial effects: Realize high personalization and emotional depth connection across scenes: Reason / mechanism: adopt PAD emotion model, introduce Dominance (control) dimension, combined with six-dimensional personality parameters (including PEmpathy). AI can distinguish the control tendency of user emotion and realize fine empathy.

[0112] AI assistant has cross-scene deep empathy ability, significantly enhances the emotional trust and dependence of users, can better pacify or motivate users in the work scene, and improves the authenticity and long-term stickiness of companionship.

[0113] Solve "memory explosion" and "personality drift" in long-term companionship: Reason / mechanism: introduce dynamic importance score , calculate the value of memory based on personality evolution contribution and time decay curve.

[0114] Adopt nonlinear random cleaning , simulate human forgetting mechanism.

[0115] Avoid context explosion and reduce computational overhead while ensuring that key emotions, core values and experiences are refined as personality anchors. Ensure that the personality characteristics and behavior style of AI remain highly consistent in long-term interaction of months or even years.

[0116] Realize dynamic and subtle guidance and recommendation based on AI personality: Reason / mechanism: the calculation of recommendation weight dynamically couples the personality parameters after AI evolution , and introduces novelty regularization term. The recommendation strategy is no longer fixed, but has the unique personality preference of AI assistant.

[0117] The recommended content is more in line with the user's immediate emotional state and the guidance target of AI assistant (such as work efficiency, emotional health). Realize the subtle personalized guidance, significantly reduce the cognitive load of users, and improve the willingness of autonomous exploration and improvement.

[0118] The above detailed description of the specific implementation is further detailed for the purpose, technical solution and beneficial effect of the present application, and it should be understood that the above is only a specific implementation of the present application and is not used to limit the protection scope of the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A PAD emotion driven artificial intelligence assistant method based on, characterized by, The artificial intelligence assistant method comprises: Collecting multi-modal data of the user and preprocessing; Identifying the current emotion and communication intention of the user according to the preprocessed data; Dynamically updating the personality parameters and internal emotional state of the AI assistant, and maintaining personification and consistency in long-term interaction; Managing the long-term interaction memory of the AI and maintaining personality consistency; Generating personalized recommendations and guidance strategies for the user based on personality and emotional state; Generating personified natural language; Verifying key facts using a knowledge graph; State updating and guidance feedback.

2. The PAD emotion-driven artificial intelligence assistant method according to claim 1, wherein, The collecting multi-modal data of the user and preprocessing specifically comprises: Obtaining multi-modal input of the user, and the multi-modal data includes text, voice, image and sensor data; Data preprocessing, including data cleaning and data standardization; The input is uniformly converted into a semantic vector by using a multi-modal encoder for subsequent sentiment and intent analysis.

3. The PAD emotion-driven artificial intelligence assistant method according to claim 1, wherein, The identifying the current emotion and communication intention of the user according to the preprocessed data specifically comprises: Using large language models (LLMs) or lightweight specialized models for parsing ; User emotion vector ; User intent vector including asking questions, seeking help, and comforting needs.

4. The PAD emotion-driven artificial intelligence assistant method of claim 1, wherein, The dynamically updating the personality parameters and internal emotional state of the AI assistant specifically comprises: Collecting current personality parameters , ai old emotions , user emotion vector , user communication intention ; executing an evolution function , outputting new personality parameters and new emotions ; Using a smoothing algorithm to suppress violent fluctuations; Emotion-personality evolution function New personality parameters The personality parameter increment calculation configured to be adjusted based on the PAD emotion and the AI self emotion, realizes the continuous, smooth and adaptive evolution of the AI personality. The personality parameters are converted into style description words and tone modifiers through dynamic mapping, and are integrated into prompt words through a dynamic description injection mechanism, affecting the interaction style of LLM output.

5. The PAD emotion-driven artificial intelligence assistant method based on PAD according to claim 1, wherein, The managing the long-term interaction memory of the AI and maintaining personality consistency specifically comprises: Hierarchical structure: recent memory IM, semantic memory STM, emotional memory EM, meta memory Meta, and core memory library; Memory distillation: assign dynamic importance scores to each memory ; low-score memories are deleted probabilistically and high-score memories are abstracted as personality anchors when threshold is reached Retrieving: retrieving relevant history information for dialogue generation according to the current context and , Each memory is assigned a dynamic importance score DIS(t), which is determined by an initial weight and a forgetting function based on time decay; The engine of memory distillation is configured to periodically perform a random clean-up task with a deletion probability of a non-linear inverse function where γ is a forgetting factor; The distillation engine will be configured to extract stable components of long-term high DIS mode and AI personality parameters, abstract them as personality anchor points and store them in the core memory library, to support the long-term consistency of AI personality.

6. The PAD emotion-driven artificial intelligence assistant method based on PAD according to claim 1, wherein, The generating personalized recommendations and guidance strategies for the user based on personality and emotional state specifically comprises: Final score of the recommended item Adopt a multi-dimensional weighting function, where the weight of each dimension is a function of the personality parameters currently evolved by the AI teacher The dynamic coupling of AI personality to the recommendation strategy is realized ​ introducing a novelty / diversity regularization term in the scoring function ； λ is a penalty coefficient; is the novelty score of the candidate object; The guiding strategy is combined with With The optimal guiding mode is triggered and output through natural dialogue embedding mode.

7. The PAD emotion-driven artificial intelligence assistant method of claim 1, wherein, The generating personified natural language specifically comprises: Prompt word factory: Will Convert to style description words, combine memory abstracts and recommendation strategies, build structured prompts; Calling a large model to generate a response text Response.

8. A PAD emotion-driven artificial intelligence assistant system based on the PAD emotion-driven artificial intelligence assistant method of any one of claims 1-7. The intelligent assistant system specifically comprises: An input and vectorization module for collecting multi-modal data of the user and preprocessing; An emotion and intention analysis module for identifying the current emotion and communication intention of the user according to the preprocessed data; A personality adaptive evolution module for dynamically updating the personality parameters and internal emotional state of the AI assistant, and maintaining personification and consistency in long-term interaction; A memory hierarchical and retrieval module for managing the long-term interaction memory of the AI and maintaining personality consistency; A guidance and recommendation decision module for generating personalized recommendations and guidance strategies for the user based on personality and emotional state; A generation and personality style injection module for generating personified natural language; A fact verification and output module for verifying key facts using a knowledge graph; A state updating and guidance feedback module for state updating and guidance feedback.

Images