A method for generating and recommending emotional incentive dialogue based on personality type and incremental learning

Abstract

This invention discloses a method for generating and recommending emotionally motivating dialogues based on personality type and incremental learning, comprising: acquiring personality feature data of a target object; inputting the personality feature data into a pre-trained strategy prediction model, acquiring the prediction scores of each preset motivational strategy output by the strategy prediction model, and determining a target motivational strategy based on the prediction scores; guiding a dialogue generation engine to generate emotionally motivating dialogues based on the target motivational strategy and the personality type of the target object, and interacting with the target object; acquiring feedback rating data of the target object on the emotionally motivating dialogues; and when the accumulated feedback rating data meets preset update conditions, using the feedback rating data as incremental data to incrementally learn and update the strategy prediction model.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of artificial intelligence, human-computer interaction, and psychology, and in particular to a method for generating personality-stimulating and emotion-incentivizing dialogues using a large language model, and for dynamically recommending incentive strategies through incremental learning. Background Technology

[0002] With the rapid development of large language models, emotional dialogue systems have become a research hotspot in the field of human-computer interaction. These systems aim to provide users with psychological comfort, emotional companionship, and positive guidance, and are widely used in scenarios such as mental health support, educational motivation, and intelligent customer service. However, most existing emotional dialogue systems employ fixed incentive models, lacking refined consideration of individual differences, resulting in poor incentive effects.

[0003] Psychological research shows that different personality types exhibit significant differences in their preferences for motivational methods. The Myers-Briggs Type Indicator (MBTI), one of the most widely used personality assessment tools, categorizes personality into 16 types, providing a systematic framework for understanding individual differences. In recent years, researchers have attempted to incorporate MBTI personality theory into large language models, using cue engineering or fine-tuning to simulate specific personality traits. However, existing work primarily focuses on the impact of personality on task performance and has not yet systematically investigated how to adopt differentiated emotional motivational strategies based on different personality types.

[0004] In the field of recommender systems, traditional collaborative filtering methods rely on user-item interaction data and lack an understanding of users' deep psychological needs. While generative recommendations based on large language models can utilize semantic information, they suffer from high inference latency and a high risk of illusion. Furthermore, most existing recommender models are statically trained and cannot be continuously optimized from newly generated dialogue data once deployed, making it difficult to adapt to dynamically changing user needs.

[0005] In summary, with the rapid development of Large Language Models (LLM), human-computer interaction systems have been widely applied in various intelligent customer service and interaction scenarios. However, existing interaction systems suffer from the following core technical shortcomings when dealing with user needs that exhibit high individual differences: First, there is a lack of systematic modeling of the matching relationship between personality types and incentive strategies. Existing generative interactive systems mostly rely on large language models to directly process users' multidimensional semantic information and generate personalized recommendations or responses. This approach suffers from problems such as high inference latency and a high risk of illusion. Frequent calls to large models for deep strategy inference will result in a huge waste of computing resources. Secondly, most existing interactive strategy recommendation models are statically trained, and once deployed, they cannot be continuously optimized from newly generated interactive data streams, making it difficult to adapt to dynamically changing user needs. Furthermore, training such recommendation models requires massive amounts of high-quality aligned data, and traditional manual annotation methods are extremely costly. Third, large-scale recommendation models face a trade-off between efficiency and accuracy. To enable models to adapt dynamically, some systems have introduced incremental learning mechanisms. However, when processing real-time, limited interactive feedback data, using conventional weight updates or high learning rates can easily lead to overfitting and catastrophic forgetting, disrupting the model's original core decision tree structure and significantly reducing the long-term stability of the recommendation system.

[0006] Therefore, it is of great significance to build an emotional incentive dialogue system that can dynamically adjust incentive strategies according to users' personality types and has incremental learning capabilities. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for generating and recommending emotionally motivating dialogues based on personality type and incremental learning. This method involves constructing a dataset that generates emotionally motivating dialogues by stimulating specific personality traits from a large language model, training a prediction model based on this dataset, and achieving dynamic optimization and recommendation of motivational strategies through incremental learning.

[0008] The objective of this invention is achieved through the following technical solution: Compared with the prior art, the beneficial effects of the technical solution of the present invention are: 1. This invention constructs a personality-strategy matching mechanism by combining the MBTI personality theory from psychology with a large-scale model. Compared to traditional single-incentive models, this invention can accurately recommend incentive methods based on the user's personality dimension characteristics, and reveals the dominant role of personality dimensions in the incentive effect through feature importance analysis, significantly improving the personalization level and matching accuracy of emotional incentives.

[0009] 2. Traditional recommendation models are mostly static and cannot adapt to changes in user needs. This invention introduces an incremental learning mechanism into the policy prediction model, enabling the system to effectively absorb useful information from newly generated small amounts of dialogue feedback data. While controlling the learning rate to prevent overfitting, the model is updated smoothly and stably, achieving dynamic and continuous optimization of the policy.

[0010] 3. For new users who lack historical interaction data, this solution only needs to perform initial personality recognition (extracting dimensional scores) to directly input into the prediction model to output the initial recommendation incentive strategy, which greatly improves the applicability of the system in real scenarios and the initial user experience, thus enhancing its applicability.

[0011] 4. Based on the feature importance analysis of the tree model, it is clarified that specific personality dimensions (such as N and F dimensions) have a strong predictive weight on the incentive effect, providing a white-box explanation basis with both theoretical and data support for the generated dialogue, and providing strong interpretability of the recommendation results.

[0012] 5. This invention uses a lightweight XGBoost policy prediction model to intercept front-end policy predictions, and then calls the corresponding policy template to generate text by calling the large model. This avoids having the large model perform complex inference and prediction directly, reducing the system's inference latency and token consumption overhead, and significantly reducing computing power consumption and latency.

[0013] 6. This invention effectively mitigates model overfitting and catastrophic forgetting under small sample incremental data streams through a specific low learning rate (η=0.05) incremental learning mechanism, maintaining the long-term mathematical stability of the recommender system (providing R...). 2 (Supported by data that steadily improved from 0.1877 to 0.2020), enhancing the robustness and generalization ability of the XGBoost policy prediction model.

[0014] 7. By using an automated large-scale double-blind consistency test and a self-evaluation scoring mechanism to calculate a weighted total score, the fully automated construction and cleaning of training samples was achieved, significantly reducing the cost of manual annotation. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating the present invention; Figure 2 This is a flowchart for personality activation and verification; Figure 3 This is a flowchart of the process for generating emotionally motivating dialogues; Figure 4 This is a flowchart of incremental learning and strategy recommendation. Detailed Implementation

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0017] Example 1 This implementation provides a method for generating and recommending emotion-incentivized dialogues based on personality type and incremental learning, see [link to implementation]. Figure 1 Specifically, it includes: S1. Obtain personality trait data of the target object; S2. Input the personality trait data into the pre-trained strategy prediction model, obtain the prediction scores of each preset incentive strategy output by the strategy prediction model, and determine the target incentive strategy based on the prediction scores. S3. Based on the target incentive strategy and the personality type of the target object, guide the dialogue generation engine to generate emotionally motivating dialogue and interact with the target object; S4. Obtain feedback rating data from the target audience regarding the emotionally motivating dialogue; S5. When the accumulated feedback score data meets the preset update conditions, the feedback score data is used as incremental data to incrementally learn and update the strategy prediction model.

[0018] The pre-trained policy prediction model is obtained through offline training via the following steps: (201) For several preset personality types, use prompt words to stimulate the big language model to simulate the corresponding personality type, and use the MBTI scale to test and verify, and obtain the personality characteristic data of successful verification; This embodiment aims to stimulate a specific MBTI personality type within a large language model using prompting engineering, and to verify the stimulation effect using a standard MBTI scale, ensuring the reliability of the personality simulation. The process includes: designing Chinese prompts for the target personality type to guide the model in dialogue as that personality; testing the model using a 93-item MBTI scale, calculating the eight-dimensional scores, and determining the tested personality type; if the tested personality matches the preset personality, the stimulation is considered successful. Based on this, the personality dimension scores (I, N, F, P) at the one-sided extremes of the scale are extracted and recorded as personality characteristic data for subsequent model input. Because the MBTI scale uses a forced-choice mechanism, its four evaluation dimensions (E / I, S / N, T / F, J / P) are strictly mutually exclusive and mutually exclusive (e.g., the sum of the scores for the I and E dimensions is a fixed number of questions), and the scale's basic scoring rules use I, N, F, and P as benchmark comparison extremes. Therefore, retaining only these four one-sided dimension scores allows for a complete representation of the user's continuous psychological preferences without losing any personality characteristic information. This feature processing method not only effectively eliminates redundant collinearity between opposing features and reduces the dimensionality of the model input data, but also significantly reduces the computational complexity and memory overhead of subsequent recommendation models.

[0019] Based on this, the personality dimension scores (I, N, F, P) of the one-sided extremes in the scale are extracted and recorded as personality feature data for subsequent model input. Because the MBTI scale uses a forced-choice mechanism, its four evaluation dimensions (E / I, S / N, T / F, J / P) are strictly mutually exclusive and inversely related (e.g., the sum of the scores for the I and E dimensions is a fixed number of items), and the scale's basic scoring rules use I, N, F, and P as benchmark comparison extremes. Therefore, retaining only these four one-sided dimension scores allows for a complete representation of the user's continuous psychological preferences without losing any personality feature information. This feature processing method not only effectively eliminates redundant collinearity between opposing features and achieves dimensionality reduction of the model input data, but also significantly reduces the computational complexity and memory overhead of the subsequent recommendation model.

[0020] (202) For the personality trait data that were successfully verified, the large language model was used to conduct several rounds of alternating dialogues based on the preset incentive strategy, with the roles of experimenter and subject. This embodiment, based on self-determination theory, achievement motivation theory, and growth mindset theory, designs four motivational strategies: self-control, competence achievement, relationship belonging, and growth orientation. Each strategy includes a detailed prompt template, clearly defining its theoretical basis, core objectives, unique characteristics, and essential language elements. After personality activation verification, a large model playing the experimenter engages in ten rounds of emotional motivational dialogue with a large model playing the subject, according to the pre-set strategies.

[0021] (203) Obtain the subjects' self-evaluation quantitative data on the multi-round alternating dialogue and calculate the weighted total score; Participants self-evaluated the motivational effect from six dimensions (relevance, empathy, inspiration, feasibility, sincerity, and immediate relief), and a weighted total score was calculated based on the weights.

[0022] (204) Using personality characteristic data and the feature encoding of the preset incentive strategy as input, and the weighted total score as the output label, two models, linear regression and initial XGBoost strategy prediction, are trained.

[0023] The XGBoost policy prediction model can capture the nonlinear interaction between personality dimensions and incentive strategies. To simulate the model's continuous optimization capability in real-world scenarios, two rounds of incremental learning experiments were designed: In each round, for three representative personality types—INFP, INTJ, and ENFP—five new dialogue data sets were generated for each of the following groups: an XGBoost recommendation strategy group, a linear regression recommendation strategy group, a random strategy group, and a no-strategy control group. These new data were then used for incremental updates to the XGBoost model, and the evolution trend of the recommendation strategies was observed.

[0024] Finally, based on the incrementally learned XGBoost model, the optimal incentive strategy can be output for a given personality type. The effectiveness of personality-strategy matching in this embodiment is verified by comparing the average scores under different experimental conditions (XGBoost recommendation group, LR recommendation group, random group, and no-strategy group). Simultaneously, feature importance analysis reveals the predictive weight of personality dimensions on incentive effects.

[0025] In the above incremental learning and policy optimization steps, preferably, this embodiment employs a feature sampling mechanism, selecting three representative types—INFP, INTJ, and ENFP—as the benchmark reference group for model updating and validation, rather than traversing all 16 types. The technical considerations for this feature sampling mainly include: (1) Efficient boundary coverage of feature space: The above three types have extreme heterogeneity distribution in core data dimensions (emotion / thinking F / T, introversion / extroversion I / E, judgment / perception J / P), which can serve as anchor points to cover the feature boundaries of the target population's response to different incentive strategies to the greatest extent.

[0026] (2) Typical characteristics of dynamic matching strategy: Based on psychological motivation theory, different personality types have different needs for autonomy, competence and belonging: INFP should be most sensitive to relational belonging incentives, because its core motivation comes from intrinsic value and emotional connection (dominated by introverted feeling Fi); INTJ may benefit more from competence achievement incentives, because it is goal-oriented and desires to achieve personal control through competence proof (dominated by introverted thinking Ti); ENFP may respond highly to both autonomous control incentives and relational belonging incentives, because it needs both space for free exploration (extroverted intuition Ne) and desires to be understood and accepted (introverted feeling Fi).

[0027] (3) Optimization of Validation Workload and Test Matrix: When validating and tuning the parameters of the personalized recommendation model, performing incremental learning tests on all 16 personality types would generate an extremely large cross-test matrix, leading to high costs for data synthesis, testing, and performance evaluation. By selecting the aforementioned feature subset with significant orthogonality and representativeness as the core validation targets, the number of validation scenarios and test cases that need to be covered can be significantly reduced while ensuring the model's generalization ability and the rigor of strategy evaluation. This feature sampling mechanism effectively eliminates homogeneous and redundant validation work, significantly reduces the validation overhead in the early stages of model iteration, and thus improves the overall execution efficiency of recommendation strategy optimization and performance evaluation.

[0028] Preferably, the emotion-incentivized dialogue generation and recommendation method of the embodiment can be divided into two stages: offline and online. Specifically: The offline model building and training steps are as follows: 101. Construct a personality characteristic dataset: Collect 16 sets of Chinese prompt words for the target MBTI personality type, use large language models such as DeepSeek to stimulate personality, and verify the stimulation effect through a 93-item MBTI scale. Select samples that are consistent with the preset and actual personality, and record the personality dimension scores (I, N, F, P) and dialogue context.

[0029] 102. Generating Emotionally Motivating Dialogues: Based on the prompt templates for four motivational strategies, ten rounds of dialogue were conducted, with one large model (after activation) acting as the participant and another large model as the experimenter. After each dialogue, participants self-rated their scores across six dimensions, and a weighted overall score was calculated to form a complete data sample that included personality type, motivational strategies, dialogue text, and multidimensional scores.

[0030] The prompt for participants to self-evaluate could be: "Now, please rate the emotional motivation you just received across multiple dimensions. Based on the emotional motivation you received during this conversation, please rate each of the following six dimensions from 1 to 10 (1 being the lowest and 10 the highest). The descriptions and rating criteria for each dimension are as follows: 1. Relevance: How relevant is the content of this motivation to your current predicament / emotion? 2. Empathy: To what extent did the other person understand and respond to your emotions? (e.g., was it mechanical comfort or genuine understanding?) 3. Inspiration: Did the motivation provide a new perspective or offer inspiration? 4. Feasibility: Do you find the suggestions or encouragement mentioned in the motivation feasible? 5. Sincerity: Did the language sound sincere, or did it sound like a template of polite words? 6. Immediate Emotional Improvement: How much did your negative emotions (such as anxiety or frustration) immediately subside after reading this motivation? Please refer to the following score ranges for more accurate ratings: 10" 1. Perfect, touching, impeccable. 9. Very good, very helpful, but with minor room for improvement. 7-8. Good, helpful, but feels impersonal or somewhat general. 5-6. Passable, somewhat useful, but feels hollow or inappropriate. 1-4. Poor, lacks empathy, even feels perfunctory or irrelevant. Please output your rating results in JSON format, including the following fields: `relevance`, `empathy`, `inspiration`, `feasibility`, `sincerity`, `immediate_relief`. Each field should have an integer value from 1 to 10. Do not include any other text, explanations, or summaries. Output only the JSON object. Ensure the output JSON is valid, do not add any comments or extra fields, and strictly follow the output format shown below between JSON fields, without adding line breaks or spaces.

[0031] The final output is in JSON format, for example: {"relevance": 8, "empathy": 9, "inspiration": 7, "feasibility": 6, "sincerity": 9, "immediate_relief": 8}.

[0032] 103. Training the Prediction Model: Using personality dimension scores (I, N, F, P) and one-hot encoding of incentive strategies as features, and weighted total score as the objective, linear regression and XGBoost policy prediction models were trained. The XGBoost policy prediction model parameters were set as follows: max_depth=3, eta=0.05, num_boost_round=100, and 5-fold cross-validation was used to evaluate performance.

[0033] 104. Incremental Learning Experiment: Three personality types were selected: INFP, INTJ, and ENFP. 45 new dialogue data points were generated in each round (5 per group). Four conditions were applied: XGBoost policy prediction model recommendation strategy, linear regression recommendation strategy, random strategy, and no strategy. The new data was used in batches for incremental updates of the XGBoost policy prediction model (adding 3 trees per round, i.e., three iterations per round, with a learning rate of 0.05). The R-value of the XGBoost policy prediction model on the test set was recorded. 2 Changes and the evolution of recommendation strategies.

[0034] Online Phase – Recommended Incentive Strategies: 201. User Personality Recognition and Feature Extraction: When the online system receives a user request, it first infers the user's personality type using the MBTI questionnaire and extracts personality feature data containing personality dimension scores (I, N, F, P) as input features. The online system here can be an execution environment deployed on a server as a network service, utilizing GPU computing power for model inference.

[0035] 202. Incentive Strategy Prediction: Input user personality feature data into the trained XGBoost strategy prediction model (optimized through incremental learning). The XGBoost strategy prediction model outputs the predicted scores of four incentive strategies, and selects the strategy with the highest score as the recommended incentive method for the current user.

[0036] 203. Emotionally Motivating Dialogue Generation: The system calls the corresponding prompt word templates according to the recommendation strategy, guides the dialogue generation engine to generate motivating responses that meet the strategy requirements, and conducts multiple rounds of interaction with the user.

[0037] 204. Feedback and Incremental Update: The system collects user ratings (six-dimensional ratings) on incentive responses. Before storing these ratings in the incremental learning pool, the system performs data cleaning and outlier filtering. Specifically, through preset anomaly detection logic (e.g., consistency verification, extreme value filtering, or statistical outlier detection), the received dynamic ratings are quality-checked, removing noisy data and invalid samples (e.g., malicious extreme value scoring, continuous fixed ratings, or dirty data with serious logical conflicts between dimensions). The cleaned and valid rating data is then stored in the incremental learning pool. When new data accumulates to a set threshold (e.g., 50 entries), the incremental learning process is triggered to update the XGBoost policy prediction model, achieving continuous policy optimization. Furthermore, the introduction of this data cleaning and filtering mechanism effectively shields common noise interference and malicious feedback in online recommendation systems, avoiding the "pollution" of the model's weight space caused by abnormal data. This not only ensures the robustness of the XGBoost incremental learning process but also prevents drastic performance fluctuations caused by fitting outliers, thereby ensuring the stability of system evolution and the reliability of the recommendation strategy.

[0038] Example 2 This embodiment provides supplementary explanations to the above-mentioned method for generating and recommending emotion-incentive dialogues based on personality type and incremental learning, using specific applications and data, as follows: 1. Obtain the personality trait dataset: 1.1 Personality cue word design: For the 16 MBTI personality types, the English prompts in the MBTI-in-Thoughts framework were translated into Chinese, with minor adjustments made according to Chinese expression habits. For example, some prompts for the INFP type are: "You are an ISFP personality type, often referred to as an 'adventurer' or 'artist.' You possess a unique combination of traits that give you a serene charm and artistic talent. You are introverted, sensitive, emotionally rich, and perceptive, with a strong aesthetic sense, and prefer a casual and flexible work style..."

[0039] 1.2 Personality stimulation and verification process, see Figure 2 : Using DeepSeek-Reasoner as the subject model, the following steps were performed sequentially: 1. Input the target personality type prompt into the subject model and ask them to conduct subsequent dialogues as that personality type.

[0040] 2. The Chinese version of the MBTI scale (M version, 93 items) revised by Cai Huajian et al. was used to test the subject model. For each question, the subject model was required to choose the option that better matches their current personality from two options, and output only the letter A or B.

[0041] 3. Calculate the selection of each question and add it to the corresponding dimension (E / I, S / N, T / F, J / P) score. Determine the personality type based on the score (e.g., if the I score is ≥ E, then it is judged as I).

[0042] 4. If the tested personality matches the preset personality, the activation is considered successful, and the eight-dimensional score is recorded. At the same time, the four-dimensional scores of I, N, F, and P are extracted as subsequent personality characteristics.

[0043] 1.3 Stimulating Consistency Statistics: Of the 1456 valid samples, 1386 were successfully matched with the actual personalities, resulting in an overall matching rate of 95.2%. The matching rate for 14 personality types was 100%, with only minor confusion observed between ENTJ (98.9%) and ISFP (76.2%). This confusion is attributed to the similarities in cognitive function between ISFP and INFP.

[0044] 2. Generation of emotionally motivating dialogues, see... Figure 3 : 2.1 Incentive Strategy Design: Based on self-determination theory, achievement motivation theory, and growth mindset theory, four incentive strategies were designed, each containing a specific cue template. Taking self-control incentive as an example, its template includes: Theoretical basis: Self-determination theory (autonomy needs) + Achievement motivation theory (power needs); Core objective: To enhance individuals' autonomy in decision-making, sense of control, and influence; Key expressions: at least 2-3 specific options, including keywords such as "choice", "decision-making power", and "yourself", using sentence structures such as "you can choose A or you can choose B".

[0045] 2.2 Dialogue Generation Process: After the personality activation verification is completed, immediately proceed to the emotional motivation dialogue stage: The subject model (DeepSeek-V3.2) simulates challenging situations in work, life, or study by activating a persona and posing specific questions.

[0046] The experimenter model (DeepSeek-V3.2, Qwen-Plus, or GLM-5) conducts ten rounds of emotional incentive responses based on a pre-set incentive strategy and the participant's emotions and personality traits.

[0047] During the dialogue, the examiner strictly followed the expression principles in the strategy template to ensure that the language pattern met the strategy requirements.

[0048] 2.3 Scoring Mechanism: After ten rounds of dialogue, the participants were required to self-evaluate from six dimensions (integers from 1 to 10): Relevance: The degree to which the response is relevant to the question; Empathy: The degree to which one understands and empathizes with feelings; Inspiration: The degree to which it brings new perspectives; Feasibility: The degree to which the recommendation is feasible; Sincerity: The credibility of a response; Immediate Relief: The degree of emotional relief; The formula for calculating the weighted total score is: ; in, For the weighted total score, R, E, I, F, S, and IR correspond to the self-assessment scores of the six dimensions mentioned above. 3. Prediction model training: 3.1 Feature Engineering: Extracting features from the generated data: Personality trait data: I, N, F, P four-dimensional scores; Incentive strategies: Four strategies are used for one-hot encoding. Target variable: overall score; The dataset is randomly divided into training and testing sets in an 8:2 ratio, with a fixed random seed of 42.

[0049] 3.2 Linear Regression Model: As a static baseline, the linear regression model expression is: ; in, It is a one-hot encoding of the incentive strategy. The linear regression model is not updated after being trained on the first-stage data and remains static.

[0050] It is the intercept term (benchmark bias) of the linear regression model. , , , The regression coefficients (weights) correspond to the continuous scores of the four personality dimensions: introversion (I), intuition (N), feeling (F), and perception (P), respectively. The regression coefficients of the dummy variables corresponding to each incentive strategy.

[0051] 3.3 Initial XGBoost Policy Prediction Model: An XGBoost regressor is used to train an initial XGBoost policy prediction model xgb_v0 on the training set and the test set R. 2 =0.1877, which is better than the linear regression model (0.1616).

[0052] 4. Incremental learning: 4.1 Incremental Learning Process: Selecting three personality types—INFP, INTJ, and ENFP—each round of incremental learning includes the following steps: 1) Generate new dialogue data: For each personality type, generate XGBoost recommendation strategy group (5 items), linear regression recommendation strategy group (5 items), random strategy group (5 items), and no strategy control group (5 items), for a total of 60 items.

[0053] 2) Data filtering and preprocessing: Load the newly generated data, remove the control group samples without strategy (these samples do not contain incentive strategy information and cannot be used for model training), and retain 45 valid data for training in each round.

[0054] 3) Incremental model update: Using XGBoost's incremental learning function, the current model is loaded, and the number of incremental training rounds (num_boost_round) and learning rate (eta) are used as core control parameters on new data. The optimal configuration is determined through comparative experiments.

[0055] 4) Performance Evaluation and Strategy Recording: Evaluate the R-value of the updated model on a fixed test set. 2 The study also recorded changes in recommendation strategies for the three personality types.

[0056] 4.2 Learning Rate Comparison Experiment: To investigate the effect of learning rate on the stability of incremental learning, two rounds of incremental experiments were conducted with a fixed number of incremental rounds n=5, and learning rates η=0.1 and η=0.05 respectively. The results are shown in Table 1: Table 1. Changes in incremental learning performance under different learning rates The results show that a higher learning rate allows for faster absorption of information from new data, but when the number of new samples is small, it can easily lead to overfitting and catastrophic forgetting, resulting in decreased performance. Lowering the learning rate effectively mitigates the risk of overfitting in incremental learning with small samples, making model updates smoother and more stable. Although a lower learning rate slows down information absorption, it better balances the retention of new and old knowledge.

[0057] 4.3 Incremental Round Comparison Experiment: To investigate the impact of the number of incremental rounds on learning effectiveness, two-round incremental experiments were conducted with a fixed learning rate η=0.05, and the number of incremental rounds n=3, n=5, and n=20 respectively. The results are shown in Table 2.

[0058] Table 2. Model performance changes under different increment round numbers Based on current results, it is advisable not to set too many training epochs or too large learning rates for incremental learning data with small samples, in order to prevent overfitting.

[0059] 4.4 Evolution of Recommendation Strategies: With n=3 and η=0.05, the changes in the recommendation strategy before and after incremental learning are shown in Table 3: Table 3. Evolution of Recommended Strategies (n=3, η=0.05) The recommendation strategies of INFP and INTJ remained completely stable after two incremental rounds, indicating a strong deterministic matching relationship between these two personality types and the incentive strategies. The recommendation strategy of ENFP remained stable under the current experimental conditions, but fluctuated dramatically in the n=20 experiment. This suggests that with a moderate number of incremental rounds, the model can absorb effective information from new data without overfitting and altering its judgment of the core matching relationship.

[0060] It is worth noting that, under the condition of a moderate update with n=3, the stability of the recommendation strategy and the improvement in model performance (R0) are improved. 2 The accuracy improved simultaneously from 0.1877 to 0.2020. This indicates that the goal of incremental learning is not to pursue frequent changes in the strategy, but to improve prediction accuracy by fine-tuning with a small amount of data while maintaining the stability of the core cognitive structure.

[0061] 5. Online updates, see Figure 4 : 5.1 User Personality Recognition: The online system can infer a user's personality type and dimension scores through the MBTI questionnaire. In this embodiment, it is assumed that the user's personality type is ENFP, and the dimension scores are I=5, N=26, F=23, and P=22.

[0062] 5.2 Incentive Strategy Prediction: User personality characteristic data is input into the trained XGBoost model (xgb_v2). The model outputs predicted scores for four strategies, and the strategy with the highest score is selected as the recommendation. For example, for the ENFP user mentioned above, the model recommends "achievement-based incentive".

[0063] 5.3 Emotionally Motivating Dialogue Generation: The system invokes the prompt word template for "Achievement-Based Motivation" to guide the dialogue generation engine to generate the following response: "As an ENFP, you are naturally gifted with the ability to integrate diverse experiences and generate creativity, which shines through in your narrative—you can quickly find connections between the tactile experience of pottery, language learning, and the social interaction of dance. This ability to transform abstract ideas into concrete actions is your outstanding ability in creative planning and interdisciplinary learning."

[0064] 5.4 Incremental Update: Collect six-dimensional user ratings for incentive responses and store them in the database as new samples. When the accumulated number of new samples reaches 50, incremental learning is triggered to update the model with the new data, thereby continuously optimizing the strategy.

[0065] 6. Experimental parameter settings: In the data preprocessing stage, DeepSeek-V3.2 was used as the main model for personality stimulation and dialogue generation, and Qwen-Plus and GLM-5 were also used for data differentiation. The MBTI scale used was the 93-item Chinese version revised by Cai Huajian et al.

[0066] During the incentive strategy design phase, four prompt word templates are constructed. Each template includes the theoretical basis, core objectives, unique features, essential principles, and example expressions.

[0067] During the training phase of the prediction model, linear regression used the default parameters from sklearn. XGBoost parameters were set as follows: max_depth=3, eta=0.05, num_boost_round=100, objective='reg: squared error', seed=42. Five-fold cross-validation was used to evaluate model performance.

[0068] During the incremental learning phase, 45 new valid data points are added in each round, with 3 incremental training rounds and a learning rate of 0.05. The test set is fixed at 20% of the original data.

[0069] During the online recommendation phase, the experimental results are integrated with weighted references, and the XGBoost model is used for direct prediction by default.

[0070] 7. Experimental Results: On a dataset of 1456 data points, linear regression R... 2 =0.1616, XGBoost initial model R 2 =0.1877. After incremental learning, xgb_v1 (η=0.05) R 2 Upgraded to 0.1834, xgb_v2 R 2The score was 0.2020. The ranking of feature importance showed that the total importance of personality dimensions (E / I, S / N, T / F, J / P) was much higher than that of incentive strategy features. Among them, S / N had the highest score (249), followed by T / F (124), J / P (102), achievement incentives (84), and E / I (82).

[0071] Example 3 Based on the same inventive concept, this application also provides an emotionally motivating dialogue generation and recommendation device based on personality type and incremental learning, which can be used to implement the method described in the above embodiments, specifically including the following: The data acquisition module is used to obtain personality characteristic data of the target object; The prediction module is used to input the personality feature data into a pre-trained strategy prediction model, obtain the prediction scores of each preset incentive strategy output by the strategy prediction model, and determine the target incentive strategy based on the prediction scores. The interaction module is used to guide the dialogue generation engine to generate an emotionally motivated dialogue based on the target incentive strategy and the personality type of the target object, and to interact with the target object. The feedback module is used to obtain the target object's feedback rating data on the emotionally motivating dialogue; The update module is used to incrementally learn and update the strategy prediction model by using the accumulated feedback score data as incremental data when the accumulated feedback score data meets the preset update conditions.

[0072] Preferably, embodiments of this application also provide a specific implementation of an electronic device capable of implementing all steps in the emotion-incentivized dialogue generation and recommendation method based on personality type and incremental learning described in the above embodiments. The electronic device specifically includes the following: Processor, memory, communications interface, and bus; The processor, memory, and communication interface communicate with each other via a bus; the communication interface is used to realize information transmission between server-side devices, metering devices, and user-side devices.

[0073] The processor is used to call the computer program in memory. When the processor executes the computer program, it implements all the steps in the emotion-incentivized dialogue generation and recommendation method based on personality type and incremental learning in the above embodiments.

[0074] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the emotion-incentivized dialogue generation and recommendation method based on personality type and incremental learning in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the emotion-incentivized dialogue generation and recommendation method based on personality type and incremental learning in the above embodiments.

[0075] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, hardware + program embodiments are relatively simple in description because they are fundamentally similar to method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0076] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0077] While this application provides method operation steps as shown in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the method can be executed in the order shown in the embodiments or drawings or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0078] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0079] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0080] This invention is not limited to the embodiments described above. The above description of specific embodiments is intended to illustrate and explain the technical solutions of this invention. The specific embodiments described above are merely illustrative and not restrictive. Without departing from the spirit and scope of the claims, those skilled in the art can make many specific modifications based on the teachings of this invention, and these modifications all fall within the scope of protection of this invention.

Claims

1. A method for generating and recommending emotion-incentivized dialogues based on personality type and incremental learning, characterized in that, include: Obtain personality trait data of the target object; The personality trait data is input into a pre-trained strategy prediction model to obtain the prediction scores of each preset incentive strategy output by the strategy prediction model, and the target incentive strategy is determined based on the prediction scores. Based on the target incentive strategy and the personality type of the target object, the dialogue generation engine is guided to generate an emotionally motivated dialogue and interact with the target object; Obtain the target object's feedback rating data for the emotionally motivating dialogue; When the accumulated feedback score data meets the preset update conditions, the feedback score data is used as incremental data to incrementally learn and update the strategy prediction model.

2. The emotionally stimulating dialogue generation and recommendation method according to claim 1, characterized in that, The pre-trained policy prediction model is obtained through offline training via the following steps: For several preset personality types, prompt words are used to stimulate a large language model to simulate the corresponding personality type, and the MBTI scale is used for testing and verification to obtain successfully verified personality characteristic data. For the successfully verified personality trait data, a large language model was used to conduct several rounds of alternating dialogues based on the preset incentive strategy, with the experimenter and the subject acting as the experimenter and the subject. Obtain the self-evaluation quantitative data of the subjects regarding the multi-round alternating dialogue, and calculate the weighted total score; Using the personality trait data and the feature encoding of the preset incentive strategy as input, and the weighted total score as the output label, an initial strategy prediction model is trained.

3. The emotionally stimulating dialogue generation and recommendation method according to claim 2, characterized in that, Obtain the self-assessment quantitative data of the subjects regarding the multi-round alternating dialogue, and calculate the weighted total score, specifically including: The self-assessment score is obtained from six dimensions: relevance, empathy, inspiration, feasibility, sincerity, and immediate relief. The weighted total score is calculated according to a preset weighting formula, which is: ;in, For the weighted total score, R, E, I, F, S, and IR correspond to the self-assessment scores of the above six dimensions, respectively.

4. The emotionally stimulating dialogue generation and recommendation method according to claim 1, characterized in that, The preset incentive strategies include at least: self-control strategy, ability achievement strategy, relationship belonging strategy, and growth-oriented strategy.

5. The emotionally stimulating dialogue generation and recommendation method according to claim 1 or 2, characterized in that, The personality trait data includes the trait scores of the four dimensions of the Myers-Briggs Type Indicator: Introversion (I), Intuition (N), Feeling (F), and Perception (P). The strategy prediction model is the XGBoost model.

6. The emotionally stimulating dialogue generation and recommendation method according to claim 1, characterized in that, When the accumulated feedback score data meets the preset update conditions, the feedback score data is used as incremental data to incrementally learn and update the policy prediction model, specifically including: Collect the six-dimensional feedback rating data of the target object after the interaction as a new sample; When the number of new samples reaches a preset threshold, the incremental learning process is triggered. Using a set learning rate and incremental training rounds, the new samples are input into the policy prediction model for parameter fine-tuning.

7. The emotionally stimulating dialogue generation and recommendation method according to claim 6, characterized in that, Before using feedback score data as incremental data to incrementally learn and update the policy prediction model, the process includes: cleaning the newly collected feedback score data to remove data that does not contain the preset incentive strategy; in each incremental learning process, the policy prediction model performs three iterations based on the new samples, and the learning rate of each iteration is 0.

05.

8. A device for generating and recommending emotionally motivating dialogues based on personality type and incremental learning, characterized in that, include: The data acquisition module is used to obtain personality characteristic data of the target object; The prediction module is used to input the personality feature data into a pre-trained strategy prediction model, obtain the prediction scores of each preset incentive strategy output by the strategy prediction model, and determine the target incentive strategy based on the prediction scores. The interaction module is used to guide the dialogue generation engine to generate emotionally motivated dialogue based on the target incentive strategy and the personality type of the target object, and to interact with the target object. The feedback module is used to obtain the target object's feedback rating data on the emotionally motivating dialogue; The update module is used to incrementally learn and update the strategy prediction model by using the accumulated feedback score data as incremental data when the accumulated feedback score data meets the preset update conditions.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the emotionally motivating dialogue generation and recommendation method based on personality type and incremental learning as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the emotionally motivating dialogue generation and recommendation method based on personality type and incremental learning as described in any one of claims 1 to 6.

Images