Personalized semantic understanding learning method, medium and system under AI platform

Abstract

The invention provides a personalized semantic understanding learning method, medium and system under an AI platform, and belongs to the technical field of semantic understanding. According to the technical scheme, the personalized semantic understanding learning method comprises the steps that a personalized semantic file recording user question and answer habits and personal terms is constructed, and dual time decay values are set; a semantic understanding optimization model is established based on an attention mechanism, personalized feature vectors and input semantic vectors are fused, a concept drift detection algorithm adopting bipartite graph maximum matching and a Hungary algorithm is designed to monitor semantic habit changes in real time, and an incremental learning mechanism based on a variable sliding window is established to calculate a semantic difference matrix. A forgetting function based on a Gaussian kernel is applied to adjust the historical semantic feature weight according to the time distance and the use frequency, and a reinforcement learning feedback module is constructed to collect user satisfaction evaluation and generate reward signal optimization model parameters; and executing a self-iterative optimization process to periodically update semantic archives and models so as to continuously adapt to personalized semantic requirements of users.

Description

Technical Field

[0001] This invention belongs to the field of semantic understanding technology, and more specifically, relates to a personalized semantic understanding learning method, medium, and system under an AI platform. Background Technology

[0002] In the field of semantic understanding technology on artificial intelligence platforms, traditional natural language processing methods mainly employ static semantic models and fixed feature extraction mechanisms to achieve user intent recognition and semantic parsing. Existing semantic understanding systems are typically based on pre-trained large-scale language models, combined with attention mechanisms and deep neural network architectures, to achieve text understanding and intent recognition functions in application scenarios such as intelligent customer service, voice assistants, and question-answering systems. These systems can handle standardized language expressions and perform well in general semantic understanding tasks. However, traditional semantic understanding technologies have significant drawbacks: First, static semantic models cannot capture users' personalized language habits and terminology preferences, resulting in low accuracy in understanding personalized expressions; second, existing systems lack the ability to dynamically perceive changes in users' semantic habits and cannot adjust semantic understanding strategies in a timely manner; third, traditional methods perform poorly when dealing with semantic concept drift during long-term user interactions, making it difficult to maintain consistent understanding accuracy. Traditional technologies struggle to address the problem of continuously declining semantic understanding accuracy caused by the dynamic changes in users' personalized semantic features over time, preventing AI platforms from providing users with a continuously optimized personalized service experience. Summary of the Invention

[0003] In view of this, the present invention provides a personalized semantic understanding learning method, medium and system under an AI platform, which can solve the technical problem in the prior art that the AI ​​platform cannot effectively adapt to the personalized semantic changes of users, resulting in a continuous decline in the accuracy of semantic understanding.

[0004] This invention is implemented as follows: The first aspect of this invention provides a personalized semantic understanding learning method under an AI platform. This method constructs a personalized semantic profile of the user, recording user question-and-answer habit patterns, a personal terminology dictionary, and personalized wake words, and assigns a first time decay value and a second time decay value to each semantic feature. Based on an attention mechanism, a semantic understanding optimization model is established to fuse the user's personalized feature vector with the semantic vector of the input statement, adjusting the weight of historical information through a time decay balance value. A concept drift detection algorithm is designed to monitor changes in user semantic habits in real time, constructing a semantic drift matrix to record the degree of semantic feature shift. A bipartite graph maximum matching problem is used for semantic feature association matching, and the optimal semantic feature allocation scheme is solved using the Hungarian algorithm. The process involves: 1) Optimal matching of current semantic features with historical semantic features to determine the direction and extent of semantic concept drift; 2) Establishing an incremental learning mechanism based on a personalized semantic graph optimization model, using a variable sliding window to collect user interaction data and calculate the semantic difference matrix between current and historical features; 3) Applying a Gaussian kernel-based forgetting function to adjust the weights of historical semantic features, dynamically updating the user's personalized semantic profile based on time distance and usage frequency; 4) Constructing a reinforcement learning feedback module to collect user satisfaction evaluations of tool usage results and generate reward signals to optimize semantic understanding model parameters; and 5) Performing a self-iterative optimization process to periodically update the user's personalized semantic profile and semantic understanding optimization model to achieve continuous adaptation to the user's personalized semantic needs.

[0005] Specifically, the steps for setting the first time decay value and the second time decay value are as follows: the first time decay value is used to control the decay rate of the user's semantic features over time. The smaller the value, the stronger the stability of the semantic features. It mainly targets the long-term retention assessment of the user's core language habits. The second time decay value is used to adjust the decay difference of different types of semantic features. Different decay rates are set for personal terms and personalized wake words to reflect the changing trend of the user's frequency of word use.

[0006] Specifically, the time decay balance value is a weighted average of the first time decay value and the second time decay value, used to balance the importance of semantic features at different time scales, ensuring that the system achieves an optimal balance between sensitivity and stability to changes in user semantics.

[0007] Specifically, the semantic drift matrix is ​​a multi-dimensional matrix that records the trajectory of changes in user semantic habits. Rows represent semantic feature dimensions, columns represent time series, and matrix element values ​​reflect the degree of shift of each semantic feature at a given time point, which is used to quantify the evolution process of semantic concepts.

[0008] Specifically, the concept drift detection algorithm is a semantic feature matching algorithm based on bipartite graph theory. It takes the semantic feature set at the current moment and the semantic feature set at the historical moment as two vertex sets of the bipartite graph, respectively. It constructs edge weights by calculating the similarity between semantic features, and uses the Hungarian algorithm to solve for the maximum weight matching to determine the drift direction and drift magnitude of the semantic concept.

[0009] Specifically, the semantic difference matrix is ​​a symmetric matrix that quantifies the degree of difference between current semantic features and historical semantic features. It is constructed by calculating the similarity between feature vectors and is used to determine whether semantic concepts have drifted significantly, providing a decision-making basis for incremental learning.

[0010] Specifically, the variable sliding window is a data collection mechanism that adaptively adjusts the window size based on the frequency of user interaction and the speed of semantic changes. The window length changes dynamically within a preset range to ensure that the latest user semantic behavior patterns are captured, while avoiding oversensitivity or sluggish response.

[0011] Specifically, the Gaussian kernel-based forgetting function is used to smoothly assign weights to historical semantic features. The inputs include time distance, frequency of use of semantic features, and importance score. The output is the corresponding weight coefficient, which is used to adjust the importance of semantic features in the user's personalized semantic profile.

[0012] The personalized semantic graph optimization model is specifically a deep neural network based on the Transformer architecture, which includes a multi-head attention mechanism layer, a feedforward neural network layer, and a residual connection structure. The attention mask parameter is used to control the correlation strength between different semantic features, the hidden state dimension determines the representational power of the semantic feature vector, and the sequence length limits the length of the user interaction history in a single processing session.

[0013] Specifically, the steps for establishing the training dataset for the personalized semantic graph optimization model include collecting historical interaction records from multiple users as the original data source, performing word segmentation and semantic annotation on user input text, extracting user-specific terminology dictionaries and question-and-answer habit patterns, constructing user semantic feature vector sequences and corresponding tool call tags, dividing the data into training and validation sets according to time order, and performing data augmentation on the training samples to improve the model's generalization ability.

[0014] Specifically, the personalized semantic graph optimization model training steps involve using the Adam optimizer to update parameters and set a learning rate decay strategy to ensure training stability; employing the cross-entropy loss function to calculate the semantic understanding accuracy loss and combining it with the mean squared error loss function to evaluate the success rate prediction accuracy using tools; using an early stopping strategy to prevent model overfitting; and periodically evaluating model performance on a validation set during the training process.

[0015] Specifically, the gating weight function is used to adjust the gating mechanism parameters of the neural network. It calculates a balance value based on three data points: the user's semantic change rate, interaction frequency, and satisfaction score. When the balance value is in the range of 0 to 0.3, a linear weight adjustment function is used to enhance gating sensitivity. When the balance value is in the range of 0.3 to 0.7, a sigmoid weight adjustment function is used to maintain gating stability. When the balance value is in the range of 0.7 to 1.0, an exponentially decaying weight adjustment function is used to reduce the gating response strength.

[0016] Before executing the self-iterative optimization process, the process includes: time distance calculated by the difference between the current time and the historical interaction time; usage frequency obtained by counting the number of times semantic features appear in the historical interaction data; user semantic change rate calculated by the change rate of the semantic difference matrix; interaction frequency obtained by counting the number of user interactions per unit time; and satisfaction score derived from user evaluation data from the reinforcement learning feedback module.

[0017] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the aforementioned personalized semantic understanding learning method under an AI platform.

[0018] A third aspect of the present invention provides a personalized semantic understanding learning system under an AI platform, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0019] This invention proposes establishing a personalized semantic profile to record users' question-and-answer habits and a personal terminology dictionary, and combining this with a time decay mechanism to adjust the weights of historical semantic features. By designing a concept drift detection algorithm and an incremental learning mechanism, this invention effectively addresses the shortcomings of traditional technologies in dynamically adapting to changes in user semantics. The system employs bipartite graph maximum matching and the Hungarian algorithm to achieve the optimal association between current and historical semantic features, accurately identifying the direction and degree of semantic concept drift. Through dynamic semantic profile updates and real-time concept drift detection, the AI ​​platform can continuously adapt to users' personalized semantic needs. Simultaneously, it uses a variable sliding window and a Gaussian kernel-based forgetting function to accurately capture and smoothly adjust changes in user semantics. The introduction of a reinforcement learning feedback module further enhances the system's adaptive optimization capabilities, solving the core problem of decreased understanding accuracy caused by personalized semantic changes, and maintaining high accuracy and stability in semantic understanding. Attached Figure Description

[0020] Figure 1This is a flowchart of the method of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0022] like Figure 1 The diagram shown is a flowchart of a personalized semantic understanding learning method under an AI platform provided by the first aspect of this invention. This method includes the following steps:

[0023] S01. Construct a personalized semantic profile for each user, record the user's question-and-answer habit patterns, personal terminology dictionary and personalized wake words, and assign a first time decay value and a second time decay value to each semantic feature;

[0024] S02. Establish a semantic understanding optimization model based on the attention mechanism, fuse the user's personalized feature vector with the semantic vector of the input sentence, and adjust the weight of historical information through the time decay balance value, wherein the time decay balance value is calculated based on the first time decay value and the second time decay value;

[0025] S03. Design a concept drift detection algorithm to monitor changes in user semantic habits in real time, construct a semantic drift matrix to record the degree of semantic feature shift. The concept drift detection algorithm uses the bipartite graph maximum matching problem to perform semantic feature association matching, solves the optimal semantic feature allocation scheme through the Hungarian algorithm, and performs the best match between the current semantic feature and the historical semantic feature to determine the direction and degree of semantic concept drift. The current semantic feature comes from the user's current interaction data, and the historical semantic feature comes from the user's personalized semantic profile.

[0026] S04. An incremental learning mechanism is established based on the personalized semantic graph optimization model. User interaction data is collected using a variable sliding window, and the semantic difference matrix between the current semantic features and historical features is calculated. The input of the personalized semantic graph optimization model is the user interaction data and the user's personalized semantic profile. The output semantic understanding results are used for tool call decisions.

[0027] S05. Apply a Gaussian kernel-based forgetting function to adjust the weights of historical semantic features, and dynamically update the user's personalized semantic profile according to the time distance and usage frequency. The historical semantic features are derived from the user's personalized semantic profile. The time distance is calculated by the difference between the current time and the historical interaction time. The usage frequency is obtained by counting the number of times the semantic features appear in the historical interaction data.

[0028] S06. Construct a reinforcement learning feedback module to collect user satisfaction evaluations of tool call results and generate reward signals to optimize semantic understanding model parameters. The tool call results are derived from the semantic understanding results in step S04, and the reward signals are used to update the parameters of the personalized semantic graph optimization model.

[0029] S07. Perform a self-iterative optimization process, periodically update the user's personalized semantic profile and the semantic understanding optimization model to achieve continuous adaptation to the user's personalized semantic needs. The update is based on the reward signal and the semantic difference matrix.

[0030] First-time decay value: Used to control the decay rate of user semantic features over time. The smaller the value, the stronger the stability of the semantic features. It is mainly used to evaluate the long-term retention of users' core language habits.

[0031] The second time decay value is used to adjust the decay difference of different types of semantic features. Different decay rates are set for personal terms and personalized wake words to reflect the changing trend of users' frequency of word use.

[0032] Time decay balance value: The weighted average of the first time decay value and the second time decay value is used to balance the importance of semantic features at different time scales, ensuring that the system’s sensitivity and stability to changes in user semantics reach the optimal balance.

[0033] Semantic drift matrix: A multidimensional matrix that records the trajectory of changes in users' semantic habits. Rows represent semantic feature dimensions, columns represent time series, and matrix element values ​​reflect the degree of shift of each semantic feature at a given time point. It is used to quantify the evolution process of semantic concepts.

[0034] Semantic difference matrix: A symmetric matrix that quantifies the degree of difference between the current semantic features and the historical semantic features. It is constructed by calculating the similarity between feature vectors and is used to determine whether the semantic concept has drifted significantly, providing a decision basis for incremental learning.

[0035] Variable sliding window: A data collection mechanism that adaptively adjusts the window size based on user interaction frequency and semantic change speed. The window length changes dynamically within a preset range to ensure the capture of the latest user semantic behavior patterns, while avoiding oversensitivity or sluggish response.

[0036] The Gaussian kernel-based forgetting function is used to smoothly assign weights to the historical semantic features. The inputs include the time distance, the frequency of use of the semantic features, and the importance score. The output is the corresponding weight coefficient, which is used to adjust the importance of the semantic features in the user's personalized semantic profile.

[0037] Personalized Semantic Graph Optimization Model: The personalized semantic graph optimization model is structured as a deep neural network based on the Transformer architecture, including a multi-head attention mechanism layer, a feedforward neural network layer, and a residual connection structure. The attention mask parameter controls the correlation strength between different semantic features, the hidden state dimension determines the representational power of the semantic feature vector, the sequence length limits the length of user interaction history in a single processing iteration, and the information flow intensity between layers is dynamically adjusted through a gating weight function. The steps for establishing the training dataset for the personalized semantic graph optimization model include collecting historical interaction records from multiple users as the original data source, performing word segmentation and semantic annotation on user input text, extracting user-personalized terminology dictionaries and question-and-answer habit patterns, and constructing user semantic feature vector sequences and corresponding tools. The data is divided into training and validation sets according to time order by calling tags. Data augmentation is performed on the training samples to improve the model's generalization ability, ultimately forming a structured training dataset containing user identifiers, semantic feature sequences, timestamps, and tool call success rates. The personalized semantic graph optimization model training steps include updating parameters using the Adam optimizer, setting a learning rate decay strategy to ensure training stability, calculating the semantic understanding accuracy loss using the cross-entropy loss function, evaluating the tool call success rate prediction accuracy using the mean squared error loss function, preventing model overfitting through an early stopping strategy, periodically evaluating model performance on the validation set during training, adjusting the attention mask parameters and the hidden state dimension based on the validation results, and finally saving the model weights that perform best on the validation set as the deployment version.

[0038] Concept drift detection algorithm: A semantic feature matching algorithm based on bipartite graph theory. It takes the semantic feature set at the current moment and the semantic feature set at the historical moment as two vertex sets of the bipartite graph, respectively. It constructs edge weights by calculating the similarity between semantic features, and uses the Hungarian algorithm to solve the maximum weight matching to determine the drift direction and drift magnitude of semantic concepts.

[0039] Gating weight function: The gating weight function is used to adjust the gating mechanism parameters of the neural network. It calculates a balance value based on three data points: the rate of change in user semantics, the interaction frequency, and the satisfaction score. When the balance value is between 0 and 0.3, a linear weight adjustment function is used to enhance gating sensitivity; when the balance value is between 0.3 and 0.7, a sigmoid weight adjustment function is used to maintain gating stability; and when the balance value is between 0.7 and 1.0, an exponentially decaying weight adjustment function is used to reduce the gating response intensity. The rate of change in user semantics is calculated using the rate of change of the semantic difference matrix. The interaction frequency is obtained by counting the number of user interactions per unit time. The satisfaction score is derived from the user evaluation data of the reinforcement learning feedback module.

[0040] The specific implementation methods of the above steps are described in detail below.

[0041] Step S01 is implemented by continuously collecting the user's interaction history through a data acquisition module, including text input content, speech-to-text results, and operation sequence. This step first establishes a mapping between the user's identity identifier and interaction timestamps. Then, natural language processing technology is used to segment and semantically annotate the user input, extracting high-frequency words as candidates for the personal terminology dictionary. The system uses a word frequency statistics algorithm to calculate the frequency of each word; when a word's frequency exceeds a threshold of 5 times, it is included in the personal terminology dictionary. Personalized wake words are identified through a pattern matching algorithm, detecting recurring fixed phrases or expressions in the user's input. When the same expression pattern appears more than 3 times within 30 days, it is identified as a personalized wake word. The initial setting of the first time decay value is 0.95 to control the long-term stability of core language habits, and the initial setting of the second time decay value is 0.8 to adjust the dynamic changes of terms and wake words. The purpose of this step is to establish a semantic foundation profile for the user, providing data support for subsequent personalized understanding.

[0042] Step S02 is implemented by establishing a semantic fusion model based on the Transformer architecture's attention mechanism. This step performs multi-head attention calculations on the user's personalized feature vector and the semantic vector of the current input sentence, capturing the correlation between semantic features through a self-attention mechanism. The time decay balance value is calculated using a weighted average algorithm, with weight coefficients set to 0.6 for the first time decay value and 0.4 for the second. When the balance value is below 0.7, the weight of historical information is increased; when the balance value is above 0.9, the influence of historical information is reduced. Eight attention heads are used, each with a dimension of 64. Residual connections and layer normalization techniques ensure the stability of model training. The model uses a positional encoding mechanism to process sequence information, with a maximum sequence length of 512 tags. The purpose of this step is to achieve effective fusion of personalized features and input semantics, improving the accuracy and personalization of semantic understanding.

[0043] The specific implementation of step S03 involves constructing a concept drift detection mechanism based on bipartite graph theory. This step uses the current semantic feature set as the left vertex set of the bipartite graph and the historical semantic feature set as the right vertex set. The similarity between feature vectors is calculated using a cosine similarity algorithm and used as the edge weight. When the similarity exceeds a threshold of 0.75, an edge is established between the corresponding vertices, forming a weighted bipartite graph structure. The Hungarian algorithm is used to solve the maximum weight matching problem, determining the optimal correspondence between the current and historical features. The semantic drift matrix is ​​stored using a dynamic array structure, with the number of rows equal to the number of semantic feature dimensions and the number of columns equal to the time window length. A significant drift is identified when the rate of change of matrix element values ​​exceeds a threshold of 0.3. The direction of concept drift is determined by vector difference calculation, and the degree of drift is measured using Euclidean distance. The purpose of this step is to monitor the changing trends of user semantic habits in real time, providing a basis for model updates.

[0044] The specific implementation of step S04 involves establishing an incremental learning mechanism based on a variable sliding window. This step dynamically adjusts the window size according to the user interaction frequency. When the interaction frequency is higher than 5 times per hour, the window length is set to 100 interaction records; when the interaction frequency is lower than 1 time per hour, the window length expands to 500 interaction records. The semantic difference matrix is ​​constructed using the Pearson correlation coefficient, with the matrix dimension being the current number of features × the historical number of features. A correlation coefficient lower than 0.6 is considered to indicate a significant difference between features. The personalized semantic graph optimization model receives interaction data and user profile information within the window, and calculates the semantic understanding probability distribution through forward propagation. This probability distribution is used for tool invocation decisions. The hidden layer dimension of the model is set to 512, and the dropout ratio is set to 0.1 to prevent overfitting. The purpose of this step is to achieve rapid adaptation to changes in user semantics, ensuring the system's real-time responsiveness.

[0045] Step S05 is implemented by applying a Gaussian kernel function to perform weight decay processing on historical semantic features. This step uses the difference between the current time and the historical interaction time as the time distance input, and the number of times the semantic feature appears in the historical record as the usage frequency input. The Gaussian kernel function is used to calculate the retention weight of each historical feature. The bandwidth parameter of the Gaussian kernel is set to 7 days; when the time distance exceeds 30 days, the weight decays to below 0.1. Normalization of usage frequency is achieved through max-min scaling, and the ratio of frequency weight to time weight is 3:7. The importance score is calculated based on the feature's contribution to the tool's call success rate, with a score range of 0 to 1. The weight coefficients are updated using an exponential moving average algorithm, with a smoothing coefficient set to 0.9. The purpose of this step is to dynamically adjust the importance of historical information, ensuring the system's sensitivity to new semantic patterns while maintaining stability.

[0046] The specific implementation of step S06 involves constructing a user feedback processing module based on reinforcement learning. This step collects satisfaction ratings for tool usage results through the user interface. The ratings are categorized into three levels: satisfied, neutral, and dissatisfied, corresponding to reward values ​​of 1.0, 0.5, and -0.5, respectively. The reward signal is generated using a temporal difference learning algorithm, with a discount factor set to 0.9 and a learning rate set to 0.01. Delayed processing of user feedback is achieved through an experience replay mechanism, with a replay buffer size of 1000 samples. Model parameters are updated using a policy gradient algorithm, adjusting the weight parameters of the personalized semantic graph optimization model through backpropagation. The reward signal is accumulated using a moving average method, with a window size of 50 feedback samples. The purpose of this step is to establish a closed-loop feedback mechanism between user satisfaction and model performance, continuously optimizing the semantic understanding effect.

[0047] Step S07 is implemented by performing a regular model update and profile maintenance process. This step is set to update every 24 hours, triggering an immediate update when the accumulated reward signal change exceeds a threshold of 0.2 or the average value of the semantic difference matrix exceeds 0.4. Updates to the user's personalized semantic profile employ incremental learning; new features are added based on statistical significance testing with a significance level set to 0.05. Outdated features are removed based on a weight threshold; features with weights below 0.1 for seven consecutive days are removed from the profile. The parameters of the semantic understanding optimization model are updated using gradient descent, with a cosine annealing scheduling strategy for the learning rate, initially set to 0.001. Model performance is evaluated through cross-validation, with the validation set comprising 20% ​​of the total data. The purpose of this step is to achieve system self-optimization and continuous improvement, ensuring long-term adaptability to user-personalized needs.

[0048] The personalized semantic graph optimization model employs an improved Transformer architecture as its core structure, comprising two main components: an encoder and a decoder. The encoder consists of six identical layers, each containing a multi-head self-attention mechanism, a positional feedforward network, and a residual connection structure. The multi-head attention mechanism uses eight attention heads, each with a dimension of 64, resulting in a total model dimension of 512. The attention mask parameters are dynamically adjusted through a gating mechanism, assigning different mask weights based on the importance of the user's semantic features. Positional encoding uses a combination of sine and cosine functions, supporting a maximum sequence length of 1024 tags. The hidden layer dimension of the feedforward network is set to 2048, and the GELU activation function is used to provide better gradient flow. Layer normalization is applied to the input and output of each sub-layer, and the dropout rate is set to 0.1 to prevent overfitting.

[0049] The decoder also consists of six layers, including all the components of the encoder, plus an encoder-decoder attention layer to focus on relevant information in the input sequence. The output layer uses a linear transformation and a softmax activation function to generate the probability distribution of tool calls. The model has approximately 110 million parameters, and its size is optimized through weight sharing and parameter pruning techniques.

[0050] The training dataset was first established by collecting historical interaction records from multiple user platforms using a distributed web crawler system. Data sources included intelligent assistant dialogue logs, search query records, and application operation sequences. The raw data underwent privacy anonymization, removing personally identifiable information while retaining semantic features. In the data preprocessing stage, regular expressions and natural language processing tools were used for text cleaning, removing invalid characters and noise. Word segmentation employed a character-based word segmenter, supporting unified processing of multilingual text.

[0051] Semantic annotation is achieved through a pre-trained language model, assigning semantic category and importance weights to each word. A personalized terminology dictionary is extracted using the TF-IDF algorithm, selecting the top 1000 words by weight as the candidate set. Question-answering habit patterns are identified using a sequence pattern mining algorithm, with a minimum support of 0.05 and a minimum confidence of 0.7. User semantic feature vectors are represented using a fusion of Word2Vec and BERT models, with a vector dimension of 768.

[0052] Tool call labels were generated through a combination of expert and automatic annotation. Expert annotations covered 10% of the data samples, while automatic annotation was implemented using a rule engine and machine learning classifier. The dataset was divided into training, validation, and test sets in an 8:1:1 ratio to ensure consistency in temporal order. Data augmentation techniques included synonym replacement, sentence rearrangement, and noise injection, augmenting the number of samples to 1.5 times that of the original data. The final dataset includes user identifiers, semantic feature sequences, timestamps, tool call success rates, and user satisfaction ratings, with a total of 5 million records.

[0053] It should be noted that the concept drift detection algorithm of this invention uses bipartite graph theory to model the current semantic feature set and the historical semantic feature set as two vertex sets of a bipartite graph. The Hungarian algorithm is used to solve for the maximum weight matching to determine the drift direction and magnitude of the semantic concept. Compared to traditional drift detection methods based on threshold judgment or statistical analysis, this graph theory-based matching algorithm can achieve optimal association between semantic features from a global optimization perspective, avoiding misjudgments caused by local matching. Traditional methods often use the rate of change of a single feature or a similarity threshold to judge concept drift, which is easily affected by noisy data, resulting in false positives or false negatives. This invention, by constructing a complete bipartite graph matching model, can comprehensively consider the association relationships between all semantic features, achieving more accurate and stable concept drift identification.

[0054] Furthermore, this invention designs a dual decay mechanism with first and second time decay values, setting different decay rates for core user language habits and personalized terms, respectively. This, combined with a Gaussian kernel-based forgetting function, achieves smooth weight adjustment of historical semantic features. Traditional semantic understanding systems typically use fixed time windows or simple linear decay functions to process historical information, failing to distinguish the stability differences between different types of semantic features. This invention's dual decay mechanism adjusts its retention strategy based on the essential attributes of semantic features, ensuring the long-term stability of core language habits while allowing for flexible changes in personalized terms. The Gaussian kernel-based forgetting function, by comprehensively considering time distance, usage frequency, and importance scores, achieves a weight allocation pattern that better aligns with human memory patterns than traditional linear decay.

[0055] Furthermore, this invention employs a variable sliding window mechanism that adaptively adjusts the window size based on user interaction frequency and the speed of semantic change, combined with a semantic difference matrix to construct an incremental learning system. Traditional fixed window mechanisms cannot adapt to the differences in interaction patterns among different users, often resulting in data windows that are too small for active users and too large for inactive users, thus affecting learning effectiveness. The variable window mechanism of this invention can dynamically adjust the data collection range according to the actual user interaction, ensuring timely capture of semantic changes while avoiding oversensitivity. The semantic difference matrix provides precise decision-making basis for incremental learning by quantifying the degree of difference between current features and historical features, enabling the system to update relevant semantic features in a targeted manner without affecting stable semantic knowledge.

[0056] These three key technological approaches form a complete dynamic semantic adaptation closed-loop system, and their synergistic effect generates comprehensive advantages that surpass the simple sum of individual technologies. The concept drift detection algorithm provides the system with accurate semantic change perception capabilities, the dual time decay mechanism ensures the rationality and stability of semantic feature updates, and the variable sliding window mechanism guarantees the timeliness and relevance of data collection. The collaborative work of these three technologies enables the system to rapidly adapt to users' personalized semantic changes while maintaining the stability of semantic understanding, solving the fundamental problem that traditional static semantic models cannot handle dynamic personalized needs. This synergistic mechanism allows the AI ​​platform to continuously learn and optimize users' personalized semantic features, achieving dynamic improvement and long-term maintenance of semantic understanding accuracy.

[0057] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the aforementioned personalized semantic understanding learning method under an AI platform.

[0058] A third aspect of the present invention provides a personalized semantic understanding learning system under an AI platform, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0059] Specifically, the principle of this invention is:

[0060] The fundamental reason why this invention can solve the problem of decreased comprehension accuracy caused by personalized semantic changes lies in its establishment of a complete dynamic semantic adaptation mechanism. First, the construction of personalized semantic profiles provides the system with the basic data structure of user semantic features. By recording users' question-and-answer habits, personal terminology dictionaries, and personalized wake words, the system can accurately capture users' semantic preferences. The design of dual time decay values ​​enables the system to distinguish the stability differences of different types of semantic features, ensuring the long-term maintenance of core language habits and the flexible adjustment of personalized terminology.

[0061] The concept drift detection algorithm achieves precise quantification of semantic feature changes through the bipartite graph maximum matching theory. By modeling the current and historical semantic features as two vertex sets of a bipartite graph and using the Hungarian algorithm to solve for the optimal matching scheme, it can accurately identify the direction and degree of semantic concept drift. The introduction of a semantic drift matrix further enhances the system's ability to record and analyze semantic change trajectories, providing a reliable data foundation for subsequent model optimization.

[0062] The core of the incremental learning mechanism lies in the collaborative work of the variable sliding window and the semantic difference matrix. The variable sliding window adaptively adjusts the data collection range based on user interaction frequency and the speed of semantic change, ensuring the system can capture the latest semantic behavior patterns in a timely manner. The semantic difference matrix provides precise decision-making basis for incremental learning by quantifying the degree of difference between current and historical features. A Gaussian kernel-based forgetting function achieves smooth weight adjustment of historical semantic features, ensuring the system maintains stability and adapts to change through a comprehensive consideration of time distance and usage frequency.

[0063] The personalized semantic graph optimization model based on the Transformer architecture achieves association modeling between different semantic features through a multi-head attention mechanism. The gating weight function dynamically adjusts the information flow intensity based on the user's semantic change rate, interaction frequency, and satisfaction rating, enabling the model to achieve an optimal balance between stability and adaptability. The reinforcement learning feedback module generates reward signals by collecting user satisfaction evaluations, driving continuous optimization of model parameters and forming a complete adaptive learning closed loop.

[0064] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0065] The specific implementation of step S01 involves continuously collecting the user's interaction history through a data acquisition module to establish a personalized semantic profile for the user. First time decay value. Second time decay value The calculation method is as follows:

[0066] ;

[0067] ;

[0068] In the formula, This is the initial decay value, used to control the decay rate of core language habits; This is the second time decay value, used to adjust the decay difference between the term and the wake word; The time interval is in days; This is the first attenuation coefficient, with a default value of 0.05; This is the second attenuation coefficient, with a default value of 0.12; The first basic retention factor ranges from 0.15 to 0.25; This is the second basic retention coefficient, ranging from 0.08 to 0.18. The parameter is obtained as follows: The stability of user language habits is obtained by statistical analysis, including step 1: collecting user interaction data for 90 consecutive days; step 2: calculating the variance of the frequency of core vocabulary usage; and step 3: determining the attenuation coefficient based on the variance. The method involves obtaining terminology change frequency statistics, including step 1: extracting the user's personal terminology dictionary; step 2: counting the terminology update frequency; and step 3: calculating the average update cycle to determine the decay coefficient.

[0069] The specific implementation of step S02 is based on establishing a semantic understanding optimization model using an attention mechanism. (Time decay balance value) The calculation expression is:

[0070] ;

[0071] In the formula, This is the time decay balance value; This is the weighting coefficient for the first attenuation value, with a default value of 0.6; This is the weighting coefficient for the second attenuation value, with a default value of 0.4; To balance the error term, the range is -0.05 to 0.05. Semantic feature fusion vector. The calculation method is as follows:

[0072] ;

[0073] In the formula, This is the fused semantic feature vector; For the first Each user's personalized feature vector; For the first A semantic vector of each input statement; For the first Weight coefficients for each user feature; For the first The weight coefficients of each input feature; Number of user feature dimensions; The number of input feature dimensions.

[0074] The specific implementation of step S03 involves designing a concept drift detection algorithm to monitor changes in user semantic habits in real time. Semantic drift matrix. The expression is:

[0075] ;

[0076] In the formula, This is the semantic drift matrix; For the first A semantic feature at time 1 The degree of offset; The number of semantic feature dimensions; Length of the time series. Offset level. The calculation method is as follows:

[0077] ;

[0078] In the formula, For the current number One semantic feature vector; For the first time in history One semantic feature vector; This is the time decay factor, with a default value of 0.1; For reference time points; Let be the Euclidean distance norm of the vector.

[0079] The specific implementation of step S04 is to establish an incremental learning mechanism based on a personalized semantic graph optimization model. Semantic difference matrix. The calculation expression is:

[0080] ;

[0081] In the formula, This is the semantic difference matrix; For the first The current feature and the first The similarity of historical features ranges from -1 to 1; This represents the current number of features; This represents the number of historical features. Similarity. Calculated using cosine similarity:

[0082] ;

[0083] Resize sliding window The adaptive adjustment formula is:

[0084] ;

[0085] In the formula, This is the base window size, with a default value of 200. This is the interaction frequency impact coefficient, with a default value of 15; User interaction frequency, measured in times per hour; This is the coefficient for the rate of change, with a default value of 50. The semantic change rate is calculated by the magnitude of change of the elements in the semantic difference matrix.

[0086] The specific implementation of step S05 involves applying a Gaussian kernel-based forgetting function to adjust the weights of historical semantic features. The Gaussian kernel-based forgetting function... The expression is:

[0087] ;

[0088] In the formula, These are the weighting coefficients; This is a time interval, expressed in days. For usage frequency; Assess importance; This represents the mean of the time distribution; the default value is 3.5. This represents the standard deviation of the time distribution; the default value is 7. Maximum usage frequency; This is the importance index, with a default value of 1.2. Time distance. The calculation method is as follows:

[0089] ;

[0090] In the formula, This is the current timestamp; This is a historical interaction timestamp; 86400 represents the number of seconds in a day. Usage frequency. This is obtained by counting the number of times semantic features appear in historical interaction data.

[0091] Among them, the forgetting function of the Gaussian kernel The core principle is to dynamically adjust the retention weights of historical semantic features through a comprehensive evaluation across three dimensions. Among these, the Gaussian kernel function... Smooth weight decay is achieved based on time distance; the weight decreases as the historical interaction time is further away from the current time, but the decay process exhibits a bell-shaped distribution, avoiding the abrupt problem of linear decay. Frequency terms are used. Ensure that frequently used semantic features receive higher retention weights; Importance scoring items Through index By amplifying the impact of key features, features that significantly contribute to the success rate of tool invocation are prioritized for retention. This design allows the forgetting function to achieve a balance between time decay, usage frequency, and feature importance, ensuring the system's sensitivity to new semantic patterns while maintaining stable memory of users' core language habits, thus enabling intelligent and dynamic updates to personalized semantic profiles.

[0092] The specific implementation method of step S06 is the same as described above, and will not be repeated in detail here.

[0093] The specific implementation of step S07 is to perform a self-iterative optimization process. Gating weight function. The expression is:

[0094] when hour: ;

[0095] when hour: ;

[0096] when hour: ;

[0097] In the formula, This refers to the gating weight coefficient; This is the equilibrium value; This is the linear adjustment coefficient, with a default value of 2.5. This is the sigmoid adjustment factor, with a default value of 8; This is the exponential decay coefficient, with a default value of 3; These represent the error terms for each stage, ranging from -0.02 to 0.02. Equilibrium value. The calculation method is as follows:

[0098] ;

[0099] In the formula, For the speed of semantic changes in users; This refers to the interaction frequency, measured in times per hour. Satisfaction rating, ranging from 0 to 1; These are weighting coefficients, with default values ​​of 0.4, 0.3, and 0.3 respectively. User semantic change rate. Calculated using the rate of change of the semantic difference matrix:

[0100] ;

[0101] In the formula, For a moment The semantic difference matrix; For a moment The semantic difference matrix; The time interval is in hours; This represents the current number of features; Quantity of historical features. Interaction frequency. Satisfaction ratings are obtained by counting the number of user interactions per unit of time. User review data is derived from the reinforcement learning feedback module and uses a normalized rating of 0 to 1, where 1 represents very satisfied and 0 represents very dissatisfied.

[0102] It should be noted that the first time decay value Second time decay value An exponential decay model is used to simulate the natural forgetting process of user semantic features over time. The main focus is on controlling the long-term stability of users' core language habits through a small decay coefficient. Ensure that the basic language model can be maintained in the long term; Then set a faster decay rate for personal terms and wake words. This reflects the dynamic changes in these characteristics. Base retention coefficient. and This ensures that important features retain a minimum weight even after long time intervals, preventing the complete loss of valuable information. This differentiated decay mechanism enables the system to distinguish between semantic features with different levels of stability and dynamism, improving the accuracy of personalized understanding.

[0103] It should be noted that the time decay balance value By combining two time decay values ​​using a weighted average, the importance of semantic features at different time scales is balanced and adjusted. Weighting coefficients and The setting reflects the importance of core language habits over dynamic terminology, ensuring that the system relies more on the user's stable language patterns in personalized understanding. Error Term The introduction of this feature provides fine-tuning space, allowing the system to make subtle adjustments based on specific application scenarios. When the balance value is low, the system increases its reliance on historical information, while when the balance value is high, it pays more attention to the current input. This dynamic adjustment mechanism enables the system to find the optimal balance between the stability and adaptability of semantic understanding.

[0104] It should be noted that the semantic feature fusion vector This function organically integrates user-personalized features with the current input semantics through linear combination. The core idea is to inject the user's historical language habits as contextual information into the current semantic understanding process, with weighting coefficients... and The contribution levels of personalized features and input features are controlled separately. This fusion mechanism allows the system to fully consider the user's personal language style, terminology preferences, and expression habits when understanding the user's current input, thereby providing semantic parsing results that are more in line with the user's needs and significantly improving the naturalness and accuracy of human-computer interaction.

[0105] It should be noted that the semantic drift matrix The temporal changes in user semantic habits are recorded using a multi-dimensional matrix structure, where the degree of shift... The method quantifies the magnitude of change of each semantic feature at different times. It combines Euclidean distance measurement and a time decay mechanism, accurately capturing the intensity of semantic feature changes while considering the impact of time on the importance of those changes. (Time decay factor) This design ensures that recent semantic changes receive higher attention, while earlier changes are given less weight. This allows the system to build a complete profile of user semantic evolution, providing a precise data foundation for concept drift detection and model adaptive adjustments, and helping the system to promptly identify and adapt to the evolving trends of user language habits.

[0106] It should be noted that the semantic difference matrix Using cosine similarity Construct a similarity measurement matrix between current features and historical features. The choice of cosine similarity ensures that the calculation result is not affected by vector length and can purely reflect the consistency of semantic direction; its value range is... This intuitively represents the degree of correlation between features. A similarity close to 1 indicates highly consistent features, close to -1 indicates completely opposite features, and close to 0 indicates no correlation. This matrix-based difference representation method provides the system with a global view of semantic changes, enabling the incremental learning mechanism to accurately identify which features need to be updated and which need to be retained, thereby achieving a refined response to user semantic changes and efficient adaptive adjustment of the model.

[0107] It should be noted that the size of the sliding window can be changed. The size of the data collection window is dynamically adjusted based on the frequency of user interaction and the speed of semantic change. Basic window Provides a stable data foundation and a high frequency of interaction. This enables users who interact frequently to receive more sensitive responses and a faster rate of change. This ensures the system can promptly capture new patterns even when semantics change rapidly. This adaptive mechanism addresses the issue of fixed window sizes failing to adapt to varying user behavior patterns. Smaller windows are used for active users to improve response speed, while larger windows are used for infrequent users to ensure data sufficiency. This design allows the system to provide personalized data processing strategies for each user, optimizing system response performance while ensuring learning effectiveness.

[0108] It should be noted that the gating weight function A piecewise function design is adopted, based on the equilibrium value. Choose appropriate adjustment strategies for different intervals: Using linear functions in intervals enhances gating sensitivity. The sigmoid function is used to maintain gating stability in the interval. An exponential decay function is used to reduce the gating response intensity within the interval. This piecewise design fully considers the optimization requirements of the neural network gating mechanism under different states: rapid response to changes is needed at low equilibrium values, stable operation is needed at medium equilibrium values, and over-adjustment needs to be avoided at high equilibrium values. By dynamically adjusting the gating parameters, this function enables the personalized semantic graph optimization model to adaptively adjust the information flow intensity according to user behavior characteristics, improving the model's response accuracy and stability.

[0109] It should be noted that the balance value The decision-making basis for the gating weight function is provided by comprehensively considering three key indicators: the speed of user semantic change, interaction frequency, and satisfaction rating. Weight coefficients The allocation of this value reflects the dominant role of the speed of semantic change in system regulation, while also balancing user activity and satisfaction feedback. This multi-dimensional comprehensive evaluation avoids the bias that may arise from a single indicator, enabling the system to comprehensively and accurately judge the current state and make corresponding adjustments. When user semantics change rapidly, interactions are frequent, and satisfaction is low, the balance value will drive the system to make more proactive adaptive adjustments; conversely, it will maintain a relatively stable operating state, ensuring the continuity and consistency of personalized services.

[0110] It should be noted that the speed of user semantic change The rate of change of the semantic difference matrix over consecutive time intervals is used to quantify the speed of evolution of user semantic patterns. This metric uses the mean of the absolute differences between matrix elements as a measure of the intensity of change, with time intervals... Normalization ensures comparability across different time scales. This computational method can capture subtle changes in users' language habits, terminology usage, and expression preferences, providing the system with accurate user state awareness. When the semantic change rate is high, the system will correspondingly increase the learning rate and adaptive adjustment frequency; when the change rate is low, it will rely more on existing user models to ensure the stability and consistency of personalized services.

[0111] It should be noted that the concept drift detection algorithm, based on bipartite graph theory, constructs the current semantic feature set and the historical feature set as two vertex sets of a bipartite graph. Edge weights are constructed by calculating the similarity between features, and the Hungarian algorithm is used to solve for the maximum weight matching to determine the optimal feature correspondence. The core advantage of this algorithm lies in its ability to handle cases with unequal numbers of features. Through global optimization, it finds the best matching scheme between current and historical features, thereby accurately identifying the drift direction and magnitude of semantic concepts. When the similarity exceeds a threshold of 0.75, an edge is established to ensure that only truly relevant features are matched, avoiding erroneous associations. This graph theory-based detection mechanism has a strong theoretical foundation and high computational efficiency, and can monitor the changing trends of user semantic habits in real time, providing reliable decision support for the system's adaptive adjustment.

[0112] It should be noted that the reinforcement learning feedback mechanism generates reward signals by collecting user satisfaction ratings for tool usage results, and continuously optimizes model parameters using a temporal difference learning algorithm and a policy gradient method. This mechanism transforms user satisfaction ratings (Satisfied 1.0, Neutral 0.5, Dissatisfied -0.5) into numerical reward signals, balancing immediate rewards and long-term gains through a discount factor of 0.9, and ensuring parameter update stability with a learning rate of 0.01. The introduction of an experience replay buffer addresses user feedback latency and sample correlation issues, while a moving average method smooths fluctuations in the reward signal. This closed-loop feedback mechanism enables the system to learn from actual user experiences, continuously optimizing semantic understanding and tool usage accuracy, achieving continuous improvement driven by user satisfaction and a spiral enhancement of personalized service quality.

[0113] To better understand and implement this invention, a specific application scenario of the invention is provided below as Example 2: Researchers constructed an intelligent assistant system based on a personalized semantic understanding learning method to verify the effectiveness of the method in practical applications. Wang, a software engineer, was selected as the test user. He has a habit of using software-related terminology and a unique interaction pattern. The system first continuously collected Wang's interaction history for 30 days through a data acquisition module, including 265 text inputs, 73 speech-to-text results, and 152 operation sequences.

[0114] During step S01, the system performs word segmentation and semantic annotation on Wang's interaction data. Using a word frequency statistics algorithm, the system identified 236 high-frequency words, of which 89 words exceeding the threshold by 5 times were included in the personal terminology dictionary, including professional terms such as "API interface," "code refactoring," and "unit testing." Personalized wake-up words were identified using a pattern matching algorithm. The system detected that Wang repeatedly used fixed expressions such as "help me check" and "can you assist me" in the interaction. These expressions appeared 12 and 8 times respectively within 30 days, both exceeding the threshold of 3 times and were therefore identified as personalized wake-up words. The system then set an initial decay value for Wang. The initial value is 0.95, and the second time decay value is... The initial value is 0.8, which is used to control the decay characteristics of core language habits and terminology changes, respectively.

[0115] In the specific implementation of step S02, the system establishes an 8-head attention mechanism semantic fusion model based on the Transformer architecture. For Wang's query on day 25, "Please help me find the latest React framework performance optimization solutions," the system calculates the time decay balance value. The balance value is higher than 0.7 but lower than 0.9. The system adjusts the historical information weight to 0.65 and the current input weight to 0.35 accordingly. During semantic feature fusion, the user-personalized feature vector has a dimension of 768, and the input sentence semantic vector has a dimension of 512. The fused semantic feature vector... The dimension was expanded to 1024, providing rich contextual information for subsequent semantic understanding.

[0116] The concept drift detection algorithm in step S03 detected a significant change in Wang's semantic habits on day 20. As shown in Table 1, the semantic drift matrix constructed by the system records the degree of shift of the main semantic features:

[0117] Table 1. Statistical Table of Semantic Feature Deviation of Wang

[0118]

[0119] Using the bipartite graph maximum matching algorithm, the system calculates the cosine similarity between the current feature and historical features. It finds that the similarity between Wang's semantic features in React-related queries and his historical Java development features is only 0.68, below the threshold of 0.75, triggering the concept drift detection mechanism. The Hungarian algorithm solution shows that Wang's technology stack preference is migrating from Java to front-end technologies, with a semantic concept drift direction shift angle of [missing value]. The drift degree is calculated to be 1.34 using Euclidean distance.

[0120] In the incremental learning mechanism implementation in step S04, the system dynamically adjusts the sliding window size based on Wang's interaction frequency. Wang's average interaction frequency during days 20-25 was 3.2 times per hour, and the system calculates the variable sliding window size accordingly. The semantic difference matrix was constructed using Pearson correlation coefficients, with a dimension of 89×127 (current feature count × historical feature count). Feature pairs with correlation coefficients below 0.6 accounted for 34%, indicating significant semantic differences. The personalized semantic graph optimization model received 262 interaction records within its receiving window. After forward propagation, the probability distribution of correct tool invocation was 0.847, a 12% improvement compared to the traditional fixed-window method.

[0121] In step S05, the forgetting function based on a Gaussian kernel is used to adjust the weights of Wang's historical semantic features. For the semantic feature of "Maven dependency management" used by Wang 15 days ago, the time distance... Day, frequency of use Second, importance score The system calculates the weight coefficient of this feature. In contrast, the "React Hook" feature weight coefficient used 3 days ago was 0.576, demonstrating the significant impact of time distance on feature preservation. Table 2 shows the system's weight adjustment results for Wang's main historical features:

[0122] Table 2. Results of Weight Adjustment for Wang's Historical Semantic Features

[0123]

[0124] Step S06, the reinforcement learning feedback module, collects Wang's satisfaction ratings for the tool's usage results. During the 5-day test, Wang provided 47 feedback ratings, including 28 satisfactory ratings (reward value 1.0), 15 neutral ratings (reward value 0.5), and 4 unsatisfactory ratings (reward value -0.5). The cumulative reward signal generated by the system through the temporal difference learning algorithm is 32.5, and the expected long-term reward is 29.8 under a discount factor of 0.9. The experience replay mechanism randomly samples 128 samples from a buffer of 1000 samples for parameter updates, with a learning rate of 0.01 ensuring stable adjustment of model parameters. The average satisfaction calculated using the moving average window is 0.766, an improvement of 20.8% compared to the initial stage of 0.634.

[0125] During the self-iterative optimization process in step S07, the system calculates the semantic change rate of Wang. Interaction frequency Satisfaction rating per hour Equilibrium value After normalization, the value is 0.612. Since the equilibrium value lies in the interval (0.3, 0.7), the system uses the sigmoid weight adjustment function to calculate the gating weight coefficient. The weighting coefficients guide the personalized semantic graph optimization model to adjust the intensity of information flow between layers, achieving precise adaptation to changes in Wang's semantics.

[0126] After seven days of continuous optimization, the system created a personalized semantic profile for Mr. Wang containing 156 effective semantic features, 89 words in his personal terminology dictionary, and 15 personalized wake words. The model achieved a semantic understanding accuracy of 89.3%, a tool invocation success rate of 86.7%, and a user satisfaction rate of 76.6%. Compared to traditional rule-based semantic understanding methods, this system improved understanding accuracy by 16.8% and personalization adaptability by 18.2%. Traditional methods rely on predefined semantic rules and fixed feature weights, failing to effectively handle the dynamic changes in users' personalized needs. When faced with professional terminology and personal expression habits, the accuracy rate is only 72.5%, and the tool invocation success rate is 69.4%. This invention, by introducing a time decay mechanism, concept drift detection, and reinforcement learning feedback, enables the system to continuously learn and adapt to users' semantic changes, achieving a technological leap from static rule matching to dynamic personalized understanding, significantly improving the naturalness and effectiveness of human-computer interaction.

[0127] It should be noted that the variables involved in this invention are explained in detail in Table 3.

[0128] Table 3. Variable Explanation Table

[0129]

[0130] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A personalized semantic understanding learning method under an AI platform, characterized in that, This system constructs a personalized semantic profile to record user question-and-answer habits, a personal terminology dictionary, and personalized wake words, assigning a first and second time decay value to each semantic feature. A semantic understanding optimization model based on an attention mechanism is established, fusing the user's personalized feature vector with the semantic vector of the input statement, and adjusting the weight of historical information through a time decay balance value. A concept drift detection algorithm is designed to monitor changes in user semantic habits in real time, constructing a semantic drift matrix to record the degree of semantic feature shift. A bipartite graph maximum matching problem is used for semantic feature association matching, and the optimal semantic feature allocation scheme is solved using the Hungarian algorithm to determine the direction and degree of semantic concept drift by best matching the current semantic features with historical semantic features. An incremental learning mechanism is established based on the personalized semantic graph optimization model, using a variable sliding window to collect user interaction data and calculate the semantic difference matrix between the current and historical semantic features. A Gaussian kernel-based forgetting function is applied to adjust the weights of historical semantic features, dynamically updating the user's personalized semantic profile based on time distance and usage frequency. A reinforcement learning feedback module is constructed to collect user satisfaction evaluations of tool usage results, generating reward signals to optimize the semantic understanding model parameters. A self-iterative optimization process is executed to periodically update the user's personalized semantic profile and the semantic understanding optimization model, achieving continuous adaptation to the user's personalized semantic needs.

2. The personalized semantic understanding learning method under the AI ​​platform according to claim 1, characterized in that, The steps for setting the first and second time decay values ​​are as follows: the first time decay value is used to control the decay rate of user semantic features over time. The smaller the value, the stronger the stability of the semantic features. It is mainly used to evaluate the long-term retention of the user's core language habits. The second time decay value is used to adjust the decay difference of different types of semantic features. Different decay rates are set for personal terms and personalized wake words to reflect the changing trend of the user's frequency of word use.

3. The personalized semantic understanding learning method under the AI ​​platform according to claim 2, characterized in that, The aforementioned time decay balance value is specifically a weighted average of the first and second time decay values, used to balance the importance of semantic features at different time scales, ensuring that the system achieves an optimal balance between sensitivity and stability to changes in user semantics.

4. The personalized semantic understanding learning method under the AI ​​platform according to claim 3, characterized in that, The semantic drift matrix is ​​a multi-dimensional matrix that records the trajectory of changes in user semantic habits. Rows represent semantic feature dimensions, columns represent time series, and matrix element values ​​reflect the degree of shift of each semantic feature at a given time point, which is used to quantify the evolution process of semantic concepts.

5. The personalized semantic understanding learning method under the AI ​​platform according to claim 4, characterized in that, The concept drift detection algorithm is specifically a semantic feature matching algorithm based on bipartite graph theory. It takes the semantic feature set at the current moment and the semantic feature set at the historical moment as two vertex sets of the bipartite graph, respectively. It constructs edge weights by calculating the similarity between semantic features, and uses the Hungarian algorithm to solve for the maximum weight matching to determine the drift direction and drift magnitude of the semantic concept.

6. The personalized semantic understanding learning method under the AI ​​platform according to claim 5, characterized in that, The semantic difference matrix is ​​a symmetric matrix that quantifies the degree of difference between current semantic features and historical semantic features. It is constructed by calculating the similarity between feature vectors and is used to determine whether semantic concepts have drifted significantly, providing a basis for decision-making in incremental learning.

7. The personalized semantic understanding learning method under the AI ​​platform according to claim 6, characterized in that, The variable sliding window is a data collection mechanism that adaptively adjusts the window size based on the frequency of user interaction and the speed of semantic changes. The window length changes dynamically within a preset range to ensure that the latest user semantic behavior patterns are captured, while avoiding oversensitivity or sluggish response.

8. The personalized semantic understanding learning method under the AI ​​platform according to claim 7, characterized in that, The Gaussian kernel-based forgetting function is specifically used to smoothly assign weights to historical semantic features. The inputs include time distance, frequency of use of semantic features, and importance score. The output is the corresponding weight coefficient, which is used to adjust the importance of semantic features in the user's personalized semantic profile.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform a personalized semantic understanding learning method under an AI platform as described in any one of claims 1-8.

10. A personalized semantic understanding learning system based on an AI platform, characterized in that, The system includes the computer-readable storage medium of claim 9, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

Images