An AI platform-based user question intent analysis method

Abstract

The application provides a user question intention analysis method under an AI platform, and belongs to the technical field of AI question answering.The application extracts semantic features and calculates a graph neural network structure index by constructing a hierarchical intention classifier, establishes a multi-task learning framework to simultaneously identify main and sub intentions and generate a micrograph structure index, models intention nodes by using a graph neural network and calculates a judgment accuracy, optimizes parameters by using a game theory model and solves an optimal combination by Nash equilibrium, solves a key path by using a maximum flow minimum cut algorithm and dynamically adjusts an identification strategy, establishes an intention vector space to calculate intention ranks of various types to realize intelligent distribution, and finally outputs an analysis result by an intention semantic generation model based on a Transformer, thereby solving the technical problems of low accuracy of complex multi-level user intention recognition and difficulty in processing dynamic association relationships between intentions in the prior art.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of AI question and answer, and in particular, relates to a user question intention analysis method under an AI platform. BACKGROUND

[0002] In the field of artificial intelligence platforms and intelligent customer service systems, user question intention analysis is a key technical basis for realizing intelligent interaction. Traditional intention recognition methods mainly use rule-based pattern matching, statistical machine learning, and deep learning classifiers, etc. technical means, by constructing an intention classification model to perform semantic understanding and intention classification on user input text. These methods are widely used in e-commerce platform intelligent customer service, voice assistants, chat robots, etc. application scenarios, and can handle relatively simple and single user intention recognition tasks. However, traditional technologies have significant defects when dealing with complex multi-level intentions. Existing methods usually use a flat classification strategy, which is difficult to effectively model the hierarchical relationship and dynamic association between user intentions, resulting in a significant decrease in recognition accuracy when faced with complex user inputs containing main intentions and multiple sub-intentions. At the same time, traditional methods lack the ability to dynamically model the intention evolution process, and cannot fully utilize user historical interaction information to optimize the current intention recognition result. Therefore, traditional technologies are difficult to solve the core technical problems of low accuracy of complex multi-level user intention recognition and difficulty in handling dynamic association between intentions, especially in application scenarios where user expression is ambiguous, intention hierarchy is complex, and comprehensive judgment needs to be combined with context. The limitations of existing technologies are more prominent. That is, there is a technical problem of low accuracy of complex multi-level user intention recognition and difficulty in handling dynamic association between intentions in existing technologies. SUMMARY

[0003] Therefore, the present application provides a user question intention analysis method under an AI platform, which can solve the technical problem of low accuracy of complex multi-level user intention recognition and difficulty in handling dynamic association between intentions in existing technologies.

[0004] The application is implemented in the following manner: the application provides a user problem intention analysis method under an AI platform, which comprises: constructing a hierarchical intention classifier to pre-process user input text, extracting a semantic feature vector and calculating a graph neural network structure index; establishing a multi-task learning framework to simultaneously identify a main intention and a sub-intention, generating a micrograph structure index for quantifying the correlation strength between intentions; modeling the intention nodes using a graph neural network and dividing them into multiple micrograph units to calculate intention determination accuracy; establishing a training data set after intention labeling of historical interaction data, coupling and adjusting the graph neural network parameters and the micrograph parameters using a game theory model, and solving the optimal parameter combination through Nash equilibrium; solving the key path in the intention graph using the maximum flow minimum cut algorithm, calculating the intention determination gain amount through an intention path optimization function, and dynamically adjusting the intention recognition strategy; establishing an intention vector space, calculating parallel intention rank, hierarchical intention rank and sequential intention rank, and transmitting the recognized intention result to the corresponding service module to complete intelligent distribution; and processing the recognized intention result through an intention semantic generation model to output the analysis result text of the user problem intention.

[0005] The graph neural network structure index is specifically a quantitative index for measuring the complexity of the intention graph network, which is obtained by calculating the comprehensive score of node degree distribution, clustering coefficient and path length.

[0006] The micrograph structure index is specifically the structural complexity quantitative value of each subgraph after the large-scale intention graph is decomposed into several small-scale subgraphs, reflecting the closeness of local intention relationships.

[0007] The intention determination accuracy is specifically the proportion of correct recognition of the user's real intention by the model, which is calculated by the matching degree of the predicted intention and the actual intention.

[0008] The intention path optimization function is used to calculate the optimal path from the user input to the final intention recognition result, the input including the current intention intensity, historical interaction weight, semantic similarity, time decay factor and user preference coefficient, and the output being the numerical result of the intention determination gain amount.

[0009] The parallel intention rank is specifically the rank of the vector group composed of multiple intention vectors existing in parallel in the same input text, representing the maximum number of linearly independent intention vectors, and reflecting the dimension complexity of the intention space.

[0010] The hierarchical intention rank is specifically the rank of the vector space composed of intention vectors at different abstraction levels in the intention classification system, used to measure the complexity of the intention hierarchical structure from abstraction to concreteness.

[0011] The order intention rank, in particular, is a rank of a vector group composed of an intention vector sequence arranged in time sequence in a user continuous interaction process, and reflects a degree of independence in an intention evolution process.

[0012] The structure of the intention semantic generation model is a text generation neural network based on a Transformer architecture, including 12 layers of encoder and decoder structures, wherein the number of attention heads is dynamically adjusted according to an exponential graph neural network structure, the number of message passing iterations is associated with a micrograph structure index, and a temperature coefficient is adaptively controlled by an intention determination accuracy.

[0013] The main intention, in particular, is a main and core demand expression in user input text, representing a task to be completed or a service to be obtained by the user, and is located at a top level in an intention hierarchy; the sub-intention, in particular, is a specific detail demand or auxiliary demand developed around the main intention, and is located at a lower level of the main intention in the intention hierarchy, and is used to supplement and perfect the overall demand expression of the user.

[0014] Before training the intention semantic generation model, the step of establishing a training data set is further included, specifically, collecting large-scale user question and corresponding intention annotation data, pairing the user input text in the historical interaction data with the artificially annotated intention category, constructing a mapping relationship from the intention recognition result to the natural language description, and finally forming a three-tuple training data set including the user question text, the intention recognition result vector and the corresponding natural language description.

[0015] The step of training the intention semantic generation model, in particular, is to use the AdamW optimizer for gradient descent training, set the learning rate to 0.0001 and use the cosine annealing strategy for adjustment, set the batch size to 16 samples, use the cross-entropy loss function to calculate the difference between the predicted text and the real text, set the training rounds to 50, and stop training when the validation set performance does not improve for 5 consecutive rounds.

[0016] The hierarchical intention classifier, in particular, is a classification model with a multi-level structure, including a coarse-grained intention classification layer and a fine-grained intention classification layer, the coarse-grained layer is responsible for identifying main intention categories, and the fine-grained layer further subdivides specific intention subcategories based on the main intention categories, and realizes step-by-step intention recognition from abstract to specific through hierarchical processing.

[0017] The semantic feature vector, in particular, is a high-dimensional numerical vector representation obtained by encoding the user input text through word embedding and language model, including lexical semantic information, syntactic structure information and context relationship information, and is used for subsequent intention recognition and classification processing.

[0018] The multi-task learning framework is specifically a machine learning architecture for simultaneously training multiple related tasks, and realizes parallel processing of main intent recognition, sub-intent recognition and intent relationship modeling through shared underlying feature representation and independent task output heads, thereby improving learning efficiency and generalization capability of the overall model.

[0019] The application proposes a hierarchical intent analysis method based on a graph neural network and multi-task learning, which realizes accurate recognition of complex intents by constructing an intent relationship graph and adopting micrograph decomposition technology. The method can simultaneously process the recognition tasks of main intent and sub-intent, and establish a dynamic association model between intents, thereby significantly improving the accuracy and robustness of multi-level intent recognition; the graph neural network is introduced to model the complex relationship between intent nodes, effectively solving the defect that the traditional planar classification method cannot process the intent hierarchy. The game theory model is used for parameter optimization and the maximum flow minimum cut algorithm is used to identify the key intent path, realizing adaptive adjustment of the intent recognition strategy and overcoming the problem of lack of dynamic optimization capability of the traditional method. At the same time, the intent vector space is established and various intent rank indexes are calculated, providing a quantitative analysis basis for complex intent relationships; the technical problems of low accuracy of complex multi-level user intent recognition and difficulty in processing dynamic association relationships between intents are solved, and the core principle is to convert intent recognition from traditional planar classification to graph structure modeling problem, capture the complex association between intents through the message passing mechanism of the graph neural network, realize the collaborative recognition of main and sub-intents in combination with the multi-task learning framework, and realize the dynamic adjustment of the recognition strategy by using the game theory optimization and path analysis technology, thereby realizing the breakthrough of the traditional technology in theory and practice. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The flowchart of the method of the application.

[0021] Figure 2 The structure diagram of the intent semantic generation model.

[0022] Figure 3 The process diagram for intent classification of user input text in embodiment 2.

[0023] Figure 4 The graph neural network structure and micrograph division diagram in embodiment 2.

[0024] Figure 5 The graph convolutional neural network processing layer diagram in embodiment 2. DETAILED DESCRIPTION

[0025] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application.

[0026] As Figure 1 shown, it is a flow chart of a user problem intent analysis method under an AI platform provided by the application, the method includes the following steps:

[0027] S01, a hierarchical intent classifier is constructed to preprocess the user input text, extract semantic feature vectors and calculate graph neural network structure indexes;

[0028] S02, a multi-task learning framework is established to simultaneously identify main intent and sub-intent, and generate micrograph structure indexes for quantifying the correlation strength between intents;

[0029] S03, the intent nodes are modeled using a graph neural network to obtain a complex intent graph, which is divided into multiple micrograph units, and the intent determination accuracy is calculated;

[0030] S04, after the historical interaction data is labeled with intent, a training data set is established, a game theory model is used to couple and adjust the graph neural network parameters and micrograph parameters, and the optimal parameter combination is solved through Nash equilibrium;

[0031] S05, the maximum flow minimum cut algorithm is used to solve the key path in the intent graph, the intent determination gain is calculated through the intent path optimization function, and the intent recognition strategy is dynamically adjusted;

[0032] S06, an intent vector space is established, and the parallel intent rank, hierarchical intent rank and sequential intent rank are calculated, and the identified intent result is transmitted to the corresponding service module to complete intelligent distribution;

[0033] S07, the identified intent result is processed through an intent semantic generation model, the analysis result text of the user problem intent is output, and user feedback data is collected to continuously optimize the overall system performance.

[0034] Among them, the graph neural network structure index is a quantitative index for measuring the complexity of the intent graph network, which is obtained by calculating the comprehensive score of node degree distribution, clustering coefficient and path length.

[0035] Among them, the micrograph structure index is the structure complexity quantitative value of each subgraph after the large-scale intent graph is decomposed into several small-scale subgraphs, reflecting the closeness of local intent relationship.

[0036] Among them, the intent determination accuracy is the proportion of the model correctly identifying the user's real intent, which is calculated by the matching degree of the predicted intent and the actual intent.

[0037] Among them, the intent path optimization function is used to calculate the optimal path from user input to the final intent recognition result, the input includes current intent intensity, historical interaction weight, semantic similarity, time decay factor and user preference coefficient, and the output is the numerical result of intent determination gain.

[0038] wherein, the parallel intention rank is specifically the rank of the vector group composed of multiple intention vectors existing in the same input text in parallel, represents the maximum number of linearly independent intention vectors, and reflects the dimension complexity of the intention space.

[0039] wherein, the hierarchical intention rank is specifically the rank of the vector space composed of intention vectors at different abstraction levels in the intention classification system, and is used to measure the complexity of the intention hierarchy from abstraction to concreteness.

[0040] wherein, the sequential intention rank is specifically the rank of the vector group composed of intention vectors arranged in time sequence in the process of continuous user interaction, and reflects the degree of independence in the intention evolution process.

[0041] As shown in Figure 2 The specific structure of the intention semantic generation model is a text generation neural network based on the Transformer architecture, including 12 layers of encoder and decoder structures, wherein the number of attention heads is dynamically adjusted according to the graph neural network structure index, the number of message passing iterations is associated with the micrograph structure index, the temperature coefficient is adaptively controlled by the intention determination accuracy, the encoder layer is responsible for understanding the identified intention feature vector, the decoder layer generates the corresponding text description according to the intention semantic relationship, the multi-head attention mechanism is used to capture the relevance between different levels of intentions, the residual connection and layer normalization ensure the stability of the model training, and the final output layer generates the probability distribution of the intention analysis result through the softmax function.

[0042] The steps of establishing the training data set of the intention semantic generation model specifically include collecting large-scale user questions and corresponding intention annotation data, pairing the user input text in the historical interaction data with the artificially annotated intention categories, constructing the mapping relationship from the intention recognition result to the natural language description, generating diversified text description templates for each intention category, including intention type, confidence, related context and suggestion operation information, expanding the number of training samples through data enhancement technology, ensuring that the model can handle various complex intention combination situations, and finally forming a three-tuple training data set containing user question text, intention recognition result vector and corresponding natural language description.

[0043] The step of training the intent semantic generation model specifically includes gradient descent training using an AdamW optimizer, a learning rate set to 0.0001 and adjusted using a cosine annealing strategy, a batch size set to 16 samples, a cross-entropy loss function used to calculate the difference between the predicted text and the real text during training, model parameters updated through a backpropagation algorithm, a training round set to 50 rounds, model performance evaluated on a validation set after each round of training, BLEU score and ROUGE score used as evaluation indicators, training stopped in advance when the validation set performance does not improve for 5 consecutive rounds, and the model weight with the best performance on the validation set saved as the final intent semantic generation model.

[0044] The hierarchical intent classifier is specifically a multi-level classification model that includes a coarse-grained intent classification layer and a fine-grained intent classification layer. The coarse-grained layer is responsible for identifying main intent categories, and the fine-grained layer further subdivides specific intent subcategories based on main intent categories, achieving gradual intent recognition from abstract to specific through hierarchical processing.

[0045] The semantic feature vector is specifically a high-dimensional numerical vector representation obtained by encoding user input text through word embedding and language model, containing lexical semantic information, syntactic structure information, and context relationship information, used for subsequent intent recognition and classification processing.

[0046] The multi-task learning framework is specifically a machine learning architecture that simultaneously trains multiple related tasks, achieving parallel processing of main intent recognition, sub-intent recognition, and intent relationship modeling through shared underlying feature representation and independent task output heads, improving the learning efficiency and generalization ability of the overall model.

[0047] The main intent is specifically the most important and core demand expression in the user input text, representing the task or service that the user most wants to complete, and is located at the top level in the intent hierarchy.

[0048] The sub-intent is specifically the specific detail requirement or auxiliary requirement around the main intent, and is located at the lower level of the main intent in the intent hierarchy, used to supplement and perfect the user's overall demand expression.

[0049] The intent node is specifically a node element representing a single intent in the intent graph structure, containing intent type, intent intensity, intent priority, and intent context attributes, and forming a complete intent relationship network through edges.

[0050] The Nash equilibrium is specifically a stable state in game theory where all participants cannot improve their benefits by changing their strategies alone, representing the optimal balance configuration state of the graph neural network parameters and micrograph parameters in the parameter optimization process.

[0051] The maximum flow minimum cut algorithm is a classical algorithm in graph theory for solving the maximum flow and minimum cut set of a network, which finds the maximum flow path from the source node to the sink node by constructing a residual network and finding an augmenting path, and is used to identify the key propagation path in the intent graph.

[0052] The current intent intensity is a weight value of the importance of the current user's expressed intent in the overall intent system, which is calculated by analyzing the semantic intensity, sentiment orientation and keyword density of the user input.

[0053] The historical interaction weight is a weight coefficient calculated based on the user's historical interaction behavior, reflecting the influence degree of the user's past intent preference on the current intent judgment, which is calculated by time decay and frequency statistics.

[0054] The semantic similarity is a numerical indicator that measures the semantic closeness between the current user input and the historical intent expression, which is calculated by word vector cosine similarity, semantic distance and context matching degree.

[0055] The time decay factor is a time-related coefficient for reducing the influence of historical information, which decreases in value with the increase of time interval, ensuring that recent interaction information has higher reference value than long-term information.

[0056] The user preference coefficient is a preference quantitative indicator calculated based on the user's personal behavior pattern and selection history, reflecting the user's individualized tendency in different intent categories, and is used for personalized intent recognition optimization.

[0057] The intent vector space is a mathematical space in which all possible intents are represented as high-dimensional vectors, and the distance and angle relationship between the intent vectors reflect the similarity and relevance between the intents, providing a geometric representation basis for intent analysis.

[0058] The specific implementation of the above steps is described in detail below.

[0059] The specific implementation of step S01 is to first use a lexical analysis algorithm to process the user input text, remove stop words and punctuation, and obtain a clean word sequence. Then use word vector embedding technology to map each word to a 300-dimensional dense vector representation, and initialize it using a pre-trained word vector model such as Word2Vec or GloVe. Then build a bidirectional long short-term memory network to encode the word vector sequence, extract context-related semantic features, and generate a 512-dimensional semantic feature vector. The hierarchical intent classifier uses a multi-layer perceptron structure, including an input layer, two hidden layers, and an output layer. The first hidden layer contains 256 neurons for coarse-grained intent classification, and the second hidden layer contains 128 neurons for fine-grained intent classification. The activation function uses a rectified linear unit function. The calculation of the graph neural network structure index is based on the complexity quantification method in graph theory. The degree centrality is calculated by counting the degree distribution of the nodes in the intent graph, the clustering degree of the network is calculated using the local clustering coefficient algorithm, and the average path length is calculated using the shortest path algorithm. The three indicators are weighted and summed according to the weights of 0.4, 0.3, and 0.3 to obtain the structure index, which is used to measure the complexity of the intent relationship network and provide quantitative basis for subsequent graph decomposition and parameter adjustment.

[0060] The specific implementation of step S02 is to establish a multi-task learning architecture based on a shared encoder. The bottom layer uses a transformer encoder structure to extract general semantic representations, and the upper layer sets a main intent classification head and a sub-intent classification head for parallel training. The main intent recognition uses a multi-classifier to map user input to 10 main intent categories, including information query, task execution, question help, entertainment interaction, learning guidance, life service, business consultation, technical support, emotional communication, and other categories. The sub-intent recognition uses a multi-label classifier to allow simultaneous recognition of multiple sub-intents, with 15 to 25 sub-intent categories set under each main intent. The micrograph structure index is generated based on graph partitioning algorithms, using spectral clustering methods to decompose large-scale intent graphs into several micrograph units containing 3 to 8 nodes. The cohesion and coupling degrees of each micrograph are calculated, the cohesion is calculated by the sum of the edge weights within the micrograph, and the coupling is calculated by the sum of the edge weights between the micrograph and other micrographs. The micrograph structure index is defined as the ratio of cohesion to coupling, which reflects the tightness of local intent relationships. The index value is usually between 0.2 and 2.5, and the larger the value, the tighter the internal relationship of the micrograph.

[0061] The specific implementation of step S03 is to use a graph convolutional neural network to learn and update the features of the intent nodes. Each intent node is initialized as a 768-dimensional feature vector containing intent semantic information, statistical information, and context information. The graph convolution operation updates the current node representation by aggregating neighbor node information, and uses a 3-layer graph convolution layer for message passing, each layer using different aggregation functions including sum aggregation, average aggregation, and attention aggregation. The division of micrograph units uses a hierarchical clustering algorithm, with a similarity threshold of 0.75, and aggregates intent nodes with similar semantics into micrograph units, each containing between 5 and 10 nodes. The calculation of intent determination accuracy uses a cross-validation method, dividing the test dataset into 5 subsets, calculating the accuracy, recall, and F1 score for each subset, and the final accuracy is the arithmetic mean of the 5 validation results. The accuracy calculation formula is the number of correctly predicted intents divided by the total number of predicted intents. This index is used to evaluate the performance of the model in the intent recognition task, and the accuracy target value is set to 85% or higher.

[0062] The specific implementation of step S04 is to manually annotate historical interaction data to establish a training dataset, including four dimensions of main intent category, sub-intent category, intent intensity, and intent priority. The training dataset contains 500,000 labeled samples, of which 70% are used for training, 15% for validation, and 15% for testing. The construction of the game theory model considers the graph neural network parameters and micrograph parameters as two game parties. The graph neural network parameters include weight matrices, bias vectors, and activation function parameters, and the micrograph parameters include micrograph division thresholds, aggregation function weights, and structure index calculation coefficients. The strategy space in game theory is defined as the feasible value range of the parameters, and the reward function is designed as the comprehensive performance index of the model on the validation set, including accuracy, computational efficiency, and generalization ability. The solution of Nash equilibrium uses an iterative optimization algorithm, which initializes the parameters and alternately optimizes the strategies of the two game parties. In each iteration, the parameters of one party are fixed to optimize the parameters of the other party, and the equilibrium state is considered to be reached when the reward change is less than 0.001 for 10 consecutive iterations. The optimal parameter combination ensures the balance between model performance and computational efficiency.

[0063] The specific implementation of step S05 is to construct the intent graph as a directed graph model, with nodes representing intent categories and edge weights representing transition probabilities or correlation strengths between intents. The implementation of the maximum flow minimum cut algorithm first sets the source point as the user input starting intent and the sink point as the final identified intent, and then finds the maximum flow path from the source point to the sink point through the Ford-Fulkerson algorithm or the Edmonds-Karp algorithm. During the iterative process of the algorithm, a residual network is constructed to find an augmented path until no new augmented path can be found. The path corresponding to the maximum flow is the key intent propagation path. The calculation of the intent path optimization function takes the current intent intensity, historical interaction weight, semantic similarity, time decay factor, and user preference coefficient as input parameters. The current intent intensity ranges from 0 to 1, the historical interaction weight is calculated based on interaction frequency and time interval and ranges from 0 to 0.8, the semantic similarity is calculated using cosine similarity and ranges from -1 to 1, the time decay factor is calculated using an exponential decay function and ranges from 0 to 1, and the user preference coefficient is based on user historical behavior statistics and ranges from 0 to 1. The optimization function outputs an intent determination gain, which is used to dynamically adjust the intent recognition strategy. When the gain is greater than 0.6, the recognition weight of the corresponding intent is increased, and when the gain is less than 0.3, the recognition weight is decreased.

[0064] The specific implementation of step S06 is to establish a 1024-dimensional intent vector space, with each intent category corresponding to a vector point in the space, and the Euclidean distance between vectors reflecting the similarity between intents. The calculation of the parallel intent rank uses the matrix rank calculation method in linear algebra, with multiple intent vectors recognized from the same input text forming a matrix. The rank of the matrix is calculated through Gaussian elimination or singular value decomposition, with the rank value representing the number of linearly independent intent vectors and reflecting the dimensional complexity of the intent space, typically ranging from 1 to 8. The calculation of the hierarchical intent rank targets the hierarchical intent structure, with intent vectors at different abstraction levels forming a block matrix. The rank value of the overall matrix is calculated, which measures the complexity of the intent hierarchy, with a range of 2 to 15. The calculation of the sequential intent rank is based on the intent sequence generated by continuous user interactions, with the time-ordered intent vectors forming a time series matrix. The matrix rank reflects the independence of intent evolution, with a range of 1 to 12. The intelligent distribution module calculates the matching degree with each service module based on the recognized intent results, and uses the weighted nearest neighbor algorithm to distribute the intent to the service module with the highest matching degree. The matching degree threshold is set to 0.7, and intents below the threshold will be distributed to the general processing module.

[0065] The specific implementation of step S07 is to use an intent semantic generation model to perform natural language processing on the recognition result and generate a user-friendly intent analysis report. The model receives an intent recognition result vector as input, extracts intent features through an encoder, and generates corresponding text descriptions through a decoder. The generated text includes four parts: intent category, confidence, related suggestions, and follow-up operations. The text length is controlled between 50 and 200 characters. The collection of user feedback data uses a combination of explicit feedback and implicit feedback. Explicit feedback is obtained through user ratings and satisfaction surveys, and implicit feedback is obtained through user follow-up behavior and dwell time analysis. The system performance optimization uses an online learning algorithm to adjust model parameters in real time based on newly collected feedback data. The learning rate is set to 0.00001, and parameter updating is performed once every 100 new feedback data collected to ensure that the system can adapt to changes and evolution in user needs.

[0066] The detailed structure of the intent semantic generation model is designed based on the transformer architecture, which includes 12 layers of encoder and 12 layers of decoder. Each layer contains multi-head self-attention mechanism and feedforward neural network. The encoder is responsible for understanding the input intent feature vector. The first to fourth layers mainly extract low-level semantic features, the fifth to eighth layers fuse middle-level semantic information, and the ninth to twelfth layers construct high-level abstract representation. The decoder is responsible for generating intent analysis text using the autoregressive method to generate words one by one. Each layer contains three sub-layers: masked self-attention, encoder-decoder attention, and feedforward network. The multi-head attention mechanism sets 16 attention heads, each with a dimension of 64. The pruning strategy of the attention head is dynamically adjusted according to the graph neural network structure index. When the structure index is greater than 1.5, all 16 heads are retained. When it is between 1.0 and 1.5, 12 heads are retained. When it is less than 1.0, 8 heads are retained. The number of message passing iterations is related to the micrograph structure index. When the micrograph structure index is greater than 1.8, the number of iterations is set to 6. When it is between 1.2 and 1.8, the number of iterations is set to 4. When it is less than 1.2, the number of iterations is set to 2. The temperature coefficient is adaptively controlled according to the intent determination accuracy. When the accuracy is greater than 90%, the temperature coefficient is set to 0.8. When it is between 80% and 90%, the temperature coefficient is set to 1.0. When it is less than 80%, the temperature coefficient is set to 1.2. The temperature coefficient is used to control the randomness and diversity of the generated text.

[0067] The establishment of the training data set includes four stages of data collection, preprocessing, labeling and enhancement. The data collection stage obtains user question texts from multiple channels, including customer service dialogue records, search query logs, question and answer community data and simulated user test data, totaling 2 million original text data. The preprocessing stage cleans the original text, removes personal privacy information, too short or too long text, repeated content and meaningless characters, and retains 1.5 million valid texts after preprocessing. The labeling stage invites 50 domain experts to label the text, each text is independently labeled by 3 experts, the labeling content includes intent category, intent intensity, context information and recommended reply, and finally the samples with consistency of more than 85% are included in the training set. The data enhancement stage uses synonym replacement, sentence transformation, back translation technology and adversarial generation method to expand the training samples, and the enhanced data set contains 3 million samples. The finally constructed triple data set contains user question text, intent recognition result vector and corresponding natural language description, wherein the intent recognition result vector is a 1024-dimensional dense vector, the average length of the natural language description is 120 characters, and the data set is divided into training set, validation set and test set according to the ratio of 8:1:1.

[0068] It should be noted that the technical idea of the present application is to model the intent node by using the graph neural network and to divide the complex intent graph into multiple micrograph units, which has a significant technical advantage over the traditional planar classification method. The traditional intent recognition method usually regards each intent as an independent classification label, ignoring the internal relationship between intents, resulting in semantic understanding deviation and classification conflict when dealing with complex multi-level intents. The graph neural network can effectively capture the complex dependency relationship and semantic association between intents by representing the intent as a node in the graph and the relationship between the intents as an edge, thereby making full use of the global intent structure information while maintaining the local intent features, and realizing the technical breakthrough from local optimization to global coordination.

[0069] The traditional intent recognition method usually adopts a single task learning mode, and trains the main intent recognizer and the sub-intent recognizer respectively. This independent processing method is easy to cause logical contradiction and semantic inconsistency between the main and sub-intent recognition results. The multi-task learning framework established by the present application realizes the parallel processing of main intent recognition, sub-intent recognition and intent relationship modeling by sharing the bottom layer feature representation and setting independent task output heads, so that the model can automatically discover and utilize the common information and complementary information between different tasks in the learning process, thereby significantly improving the consistency and accuracy of intent recognition, and avoiding the feature learning redundancy and model parameter waste problems caused by task independence in the traditional method.

[0070] The technical idea of the application is to couple and adjust the parameters of the graph neural network and the micrograph parameters by using the game theory model, and to solve the optimal parameter combination by Nash equilibrium. The traditional parameter optimization method solves the fundamental problem of local optimization leading to overall performance degradation. The traditional optimization strategy usually uses local search algorithms such as gradient descent, which is easy to fall into local optimal solution and difficult to balance the parameter configuration between different model components, resulting in limited overall performance of the system. The game theory model regards the parameters of the graph neural network and the micrograph as different game participants, builds a payoff function and a strategy space, and uses Nash equilibrium theory to find a stable state where all participants cannot change their strategies to obtain better benefits, ensuring the global optimality of parameter configuration and the coordination between system components.

[0071] The synergistic effect of the above three key technical ideas forms an organic and unified technical system, producing a comprehensive technical effect far exceeding the individual effect of each technical idea. The intention relationship modeling of the graph neural network provides rich structured information for the multi-task learning framework, enabling the collaborative recognition of main and sub-intentions based on more accurate intention relationship graphs. The collaborative features generated by the multi-task learning framework provide more accurate node representations for the message passing process of the graph neural network, forming a positive feedback loop. At the same time, the game theory-driven parameter optimization strategy coordinates the parameter configuration of the graph neural network and the multi-task learning framework, ensuring that the two core components can perform at their best under optimal parameters, avoiding the performance loss caused by parameter mismatch in traditional methods, and achieving a technical leap from simple classification to complex relationship modeling, from single-task processing to multi-task collaboration, and from local optimization to global equilibrium.

[0072] Specifically, the principle of the application is that the application can solve the technical problems of low accuracy of complex multi-level user intention recognition and difficulty in handling dynamic association between intentions. The root cause of the problem lies in the innovative combination of graph theory, game theory and deep learning technology, which constructs a multi-level and adaptive intention analysis framework.

[0073] Firstly, by designing hierarchical intention classifiers and multi-task learning frameworks, the application breaks through the limitations of traditional flat classification. The hierarchical classifier uses a step-by-step recognition strategy from coarse to fine granularity, which can effectively handle the hierarchical structure of intentions. The multi-task learning framework realizes the collaborative modeling of main and sub-intentions by sharing feature representations, avoiding information loss and conflict problems that may occur when processing each type of intention independently.

[0074] Secondly, the introduction of graph neural networks provides strong technical support for modeling the relationship between intents. By representing intents as nodes and relationships between intents as edges in a graph, graph neural networks can capture complex intent dependencies through a message passing mechanism and divide large-scale intent graphs into local structures that can be processed through micrograph decomposition techniques, ensuring both computational efficiency and accuracy of relationship modeling.

[0075] Thirdly, the application of game theory models achieves global optimization of system parameters. By treating graph neural network parameters and micrograph parameters as game participants, the optimal parameter combination is solved using Nash equilibrium, ensuring coordinated cooperation between different components and avoiding the problem of overall performance degradation caused by local optimization.

[0076] Finally, the design of the maximum flow minimum cut algorithm and the intent path optimization function realizes the dynamic adjustment of the recognition strategy. By identifying the key propagation path in the intent graph, the system can dynamically adjust the recognition strategy based on current intent intensity, historical interaction weights and other factors, achieving a technological leap from static classification to dynamic adaptation.

[0077] A specific embodiment 1 of the present application is provided below, and the specific implementation of each step in embodiment 1 is described in detail as follows.

[0078] The specific implementation of step S01 is to first use a lexical analysis algorithm to perform word segmentation processing on the user input text, remove stop words and punctuation, and obtain a clean word sequence. Then use word vector embedding technology to map each word to a 300-dimensional dense vector representation, and initialize using a pre-trained word vector model. Then build a bidirectional long short-term memory network to encode the word vector sequence, extract context-related semantic features, and generate a 512-dimensional semantic feature vector. The hierarchical intent classifier uses a multi-layer perceptron structure, including an input layer, two hidden layers, and an output layer. The first hidden layer contains 256 neurons for coarse-grained intent classification, and the second hidden layer contains 128 neurons for fine-grained intent classification. The calculation of the graph neural network structure index is based on the complexity quantification method in graph theory, and is specifically represented as follows:

[0079] I gnn = α1D c + α2C avg + α3L avg + δ1;

[0080] In the formula, I gnn is the graph neural network structure index; D c is the degree centrality index; C avg is the average clustering coefficient; L avgis the average path length; a1, a2, a3 are weight coefficients, respectively taking 0.4, 0.3, 0.3; d1 is a structure complexity adjustment factor.

[0081] wherein the parameter acquisition method is: D c By counting the degree of each node in the intent graph and calculating the centrality, the calculation formula is wherein d i is the degree of node i, and N is the total number of nodes in the graph. avg The local clustering coefficient algorithm is used to calculate the ratio of the number of edges between neighbor nodes to the possible number of edges for each node, and then the average value is calculated. avg The shortest path algorithm is used to calculate the shortest distance between all node pairs, and then the average value is calculated. The range of d1 is 0.1 to 0.3, which is used to adjust the complexity evaluation deviation of different graph structures.

[0082] The specific implementation of step S02 is to establish a multi-task learning architecture based on a shared encoder, and a transformer encoder structure is used at the bottom to extract a general semantic representation. A main intent classification head and a sub-intent classification head are set at the upper layer for parallel training. The main intent recognition adopts a multi-classifier to map the user input to 10 main intent categories. The sub-intent recognition adopts a multi-label classifier to allow simultaneous recognition of multiple sub-intents, and 15 to 25 sub-intent categories are set under each main intent. The generation of the micrograph structure index is based on a graph partitioning algorithm, and a spectral clustering method is used to decompose a large-scale intent graph into several micrograph units containing 3 to 8 nodes, which is specifically represented as follows:

[0083]

[0084] In the formula, I micro is the micrograph structure index; H inner is the micrograph cohesion degree; C outer is the micrograph coupling degree; m1, m2 are numerical stability adjustment parameters; and d1 is a micrograph partitioning error term.

[0085] wherein the parameter acquisition method is: H inner is obtained by calculating the sum of the edge weights within the micrograph, and is represented as wherein E inner is the micrograph internal edge set, w uv is the weight of edge (u, v), and u and v represent the starting node and the ending node of the edge, respectively. outer is obtained by calculating the sum of the edge weights between the micrograph and other micrographs, and is represented as wherein E outerThis represents the set of external edges of the micrograph. μ1 and μ2 are both 0.01 to avoid the case where the denominator is zero. ε1 ranges from 0.05 to 0.15, reflecting the uncertainty of the micrograph partitioning algorithm.

[0086] The specific implementation of step S03 involves using a graph convolutional neural network to learn features and update the representation of intent nodes. Each intent node is initialized as a 768-dimensional feature vector, containing semantic information, statistical information, and contextual information. The graph convolution operation updates the current node representation by aggregating neighbor node information, and a three-layer graph convolutional layer is used for message passing. The division of micro-graph units uses a hierarchical clustering algorithm, with a similarity threshold of 0.75. The calculation of intent determination accuracy uses a cross-validation method, specifically as follows:

[0087]

[0088] In the formula, A intent TP represents the accuracy of intent determination; K represents the cross-validation fold number; k Let be the number of true cases in the k-th fold verification; FP k FN is the number of false positives in the k-th fold verification. k β1 represents the number of false negatives in the k-th fold validation; β1 is the accuracy correction factor.

[0089] The parameter acquisition method is as follows: TP k The number of samples correctly predicted as positive in the k-th fold test set is obtained by counting the number of such samples. k The number of samples incorrectly predicted as positive in the k-th fold test set is obtained by counting the number of such samples. FN k The value was obtained by counting the number of samples mispredicted as negative in the k-th fold test set. K was set to 5, and 5-fold cross-validation was used. β1 ranged from -0.02 to 0.02 and was used to correct for evaluation bias between different datasets.

[0090] The specific implementation of step S04 involves manually annotating historical interaction data to create a training dataset. The annotations include four dimensions: main graph category, sub-intent category, intent strength, and intent priority. The training dataset contains 500,000 annotated samples, with 70% used for training, 15% for validation, and 15% for testing. The game theory model treats the graph neural network parameters and micrograph parameters as two players, employing an iterative optimization algorithm to solve for the Nash equilibrium. An equilibrium state is considered reached when the payoff change is less than 0.001 after 10 consecutive iterations.

[0091] The specific implementation of step S05 is to construct the intent graph as a directed graph model, with nodes representing intent categories and edge weights representing transition probabilities or correlation strengths between intents. The implementation of the maximum flow minimum cut algorithm uses the Ford-Fulkerson algorithm to find the maximum flow path from the source point to the sink point. The calculation of the intent path optimization function is based on multiple input parameters, which are specifically represented as follows:

[0092] G path = γ1S current + γ2W history + γ3R semantic + γ4D time + γ5P user + ζ1

[0093] where G path is the intent determination gain; S current is the current intent strength; W history is the historical interaction weight; R semantic is the semantic similarity; D time is the time decay factor; P us er is the user preference coefficient; γ1, γ2, γ3, γ4, γ5 are weight parameters; ζ1 is the path optimization bias term.

[0094] where the parameter acquisition method is: S current is calculated by analyzing the semantic strength, sentiment orientation and keyword density of user input, with a value range of 0 to 1. W history is calculated according to user historical interaction behavior, with a formula of where f i is the frequency of the i-th historical interaction, t i is the time interval, M is the total number of historical interactions, and λ is the time decay parameter with a value of 0.5 and a value range of 0 to 0.8. R semantic is calculated using cosine similarity, with a formula of where v1, v2 are the semantic vectors to be compared, with a value range of -1 to 1. D time is calculated using an exponential decay function, with a formula of D time = e -η·Δt , where η is the decay rate with a value of 0.1 and Δt is the time interval with a value range of 0 to 1. P user is based on user historical behavior statistics, with a value range of 0 to 1. γ1, γ2, γ3, γ4, γ5 take values of 0.3, 0.25, 0.2, 0.15 and 0.1 respectively. ζ1 ranges from -0.1 to 0.1.

[0095] The specific implementation of step S06 is to establish a 1024-dimensional intention vector space, and each intention category corresponds to a vector point in the space. The calculation of the parallel intention rank adopts the matrix rank calculation method in linear algebra, and is specifically represented as follows:

[0096] R parallel =rank(M parallel )+θ1;

[0097] In the formula, R parallel is the parallel intention rank; M parallel is the parallel intention matrix; rank(·) is a matrix rank function; and θ1 is a rank calculation correction term.

[0098] The calculation of the hierarchical intention rank is for the hierarchical intention structure, and is specifically represented as follows:

[0099] R hierarchy =rank(M hierarchy )+θ2;

[0100] In the formula, R hierarchy is the hierarchical intention rank; M hierarchy is the hierarchical intention matrix; and θ2 is a hierarchical correction term.

[0101] The calculation of the sequential intention rank is based on the intention sequence generated by the continuous interaction of the user, and is specifically represented as follows:

[0102] R sequence =rank(M sequence )+θ3;

[0103] In the formula, R sequence is the sequential intention rank; M sequence is the sequential intention matrix; and θ3 is a sequence correction term.

[0104] The parameter acquisition method is as follows: M parallel is obtained by arranging multiple intention vectors recognized from the same input text in rows to form a matrix. M hierarchy is obtained by arranging intention vectors of different abstraction levels to form a block matrix. M sequence is obtained by arranging intention vectors in time sequence to form a time sequence matrix. The ranges of θ1, θ2 and θ3 are all 0 to 0.5, which are used to correct the deviation of different calculation methods. The intelligent distribution module calculates the matching degree with each service module according to the recognized intention result, distributes the intention to the service module with the highest matching degree by using a weighted nearest neighbor algorithm, and the matching degree threshold is set to 0.7.

[0105] The specific implementation of step S07 is to use an intent semantic generation model to perform natural language processing on the recognition result to generate a user-friendly intent analysis report. The model receives an intent recognition result vector as input, extracts intent features through an encoder, and generates a corresponding text description through a decoder. The generated text includes intent category, confidence, related suggestions, and subsequent operations, with a text length controlled between 50 and 200 characters. User feedback data is collected using a combination of explicit feedback and implicit feedback, and system performance optimization uses an online learning algorithm with a learning rate set to 0.00001, with parameter updates performed every 100 new feedback data collected.

[0106] The detailed structure of the intent semantic generation model is designed based on a transformer architecture, including 12 layers of encoder and 12 layers of decoder, each layer containing multi-head self-attention mechanism and feedforward neural network. The encoder is responsible for understanding the input intent feature vector, and the decoder is responsible for generating the intent analysis text using an autoregressive method to generate words one by one. The multi-head attention mechanism sets 16 attention heads, each with a dimension of 64, and the pruning strategy of the attention heads is dynamically adjusted according to the graph neural network structure index. When the structure index is greater than 1.5, all 16 heads are retained, when it is between 1.0 and 1.5, 12 heads are retained, and when it is less than 1.0, 8 heads are retained. The number of message passing iterations is related to the micrograph structure index, when the micrograph structure index is greater than 1.8, the number of iterations is set to 6 times, when it is between 1.2 and 1.8, the number of iterations is set to 4 times, and when it is less than 1.2, the number of iterations is set to 2 times. The temperature coefficient is adaptively controlled according to the intent determination accuracy, when the accuracy is greater than 90%, the temperature coefficient is set to 0.8, when it is between 80% and 90%, the temperature coefficient is set to 1.0, and when it is less than 80%, the temperature coefficient is set to 1.2.

[0107] The establishment of the training data set includes four stages of data collection, preprocessing, labeling, and enhancement. In the data collection stage, user question texts are obtained from multiple channels, totaling 2 million original text data. In the preprocessing stage, the original text is cleaned, and after preprocessing, 1.5 million valid texts are retained. In the labeling stage, 50 domain experts are invited to label the text intent, each text is independently labeled by 3 experts, and finally the samples with consistency of more than 85% are included in the training set. In the data enhancement stage, synonym replacement, sentence transformation, back-translation technology, and adversarial generation methods are used to expand the training samples, and the enhanced data set contains 3 million samples. The finally constructed triple training data set contains user question text, intent recognition result vector, and corresponding natural language description for each sample, and the data set is divided into training set, validation set, and test set in the ratio of 8:1:1.

[0108] For better understanding and implementation of the present application, the following provides an embodiment 2 of a specific application scenario of the present application: researchers develop an AI intelligent customer service system, and use the method of the present application to analyze the user's intention. The researchers collected 10,000 pieces of user consultation text from an online shopping platform as test data, which contains various intention types such as return and exchange consultation, commodity inquiry, logistics inquiry, after-sales service, etc.

[0109] In the implementation process of step S01, as shown in Figure 3 , the researchers first preprocess the user input text "I bought a mobile phone that charges very slowly and overheats seriously, can I return it". After lexical analysis, the keyword sequence ["mobile phone", "charge", "slow", "overheat", "serious", "return"] is obtained, and each word is mapped to a 300-dimensional vector through word vector embedding technology. After bidirectional long short-term memory network coding, a 512-dimensional semantic feature vector is generated, of which the first 128 dimensions mainly contain product attribute information, the middle 256 dimensions contain problem description information, and the last 128 dimensions contain user demand information. The hierarchical intention classifier processes the semantic feature vector, and the first hidden layer output shows that the coarse-grained intention is the "after-sales service" category, with a confidence of 0.87, and the second hidden layer further subdivides it into the "return application" sub-intention, with a confidence of 0.82. According to the calculation formula I gnn = α1D c + α2C avg + α3L avg + δ1, the graph neural network structure index is 1.73, wherein the degree centrality D c is 0.45, the average clustering coefficient C avg is 0.68, the average path length L avg is 2.34, and the structure complexity adjustment factor δ1 takes the value 0.15.

[0110] In step S02, the multi-task learning framework simultaneously processes the main intention and sub-intention recognition tasks. The main intention recognition results are shown in Table 1:

[0111] Table 1 Main intention recognition results table

[0112] Main intent category Confidence Weight coefficient Information query 0.12 0.08 Task execution 0.08 0.05 Problem solving 0.15 0.12 After-sales service 0.87 0.65 Business consultation 0.09 0.06 Technical support 0.23 0.18

[0113] The micrograph structure index is calculated by the formula . The micrograph corresponding to the user's question contains 5 intention nodes, the cohesion H inner is 3.2, the coupling C outer is 1.8, the numerical stability adjustment parameters μ1 and μ2 are both 0.01, the micrograph division error term ε1 is 0.08, and the final micrograph structure index is 1.86. As Figure 4 Figure 5The figure neural network structure and micrograph division schematic diagram and the figure convolutional neural network processing layer schematic diagram are shown.

[0114] Step S03 uses a figure convolutional neural network to model the intent node. The initialized 768-dimensional intent node feature vector is processed by 3 layers of figure convolution. The first layer mainly extracts local features of the node, the second layer aggregates neighbor node information, and the third layer generates a global representation. The intent determination accuracy is calculated by cross-validation, and the formula is The calculation result is 0.89, wherein the accuracy of each fold of 5-fold cross-validation is 0.91, 0.88, 0.87, 0.90, and 0.89, and the accuracy correction factor β1 is 0.01.

[0115] The training data set established in step S04 contains 500,000 labeled samples, and the researchers select historical data similar to the current user's question for game theory model training. The figure neural network parameters include weight matrix, bias vector and activation function parameters, the micrograph parameters include division threshold, aggregation function weight and structure index calculation coefficient. After 16 iterations of optimization, the continuous income change is reduced to 0.0008, which is less than the set threshold 0.001, and reaches the Nash equilibrium state.

[0116] The parameter values of the intent path optimization function in step S05 are shown in Table 2:

[0117] Table 2 Intent path optimization parameter table

[0118] Parameter name Parameter value Calculation method Current intent strength S current ]]> 0.87 Semantic analysis Historical interaction weight W history ]]> 0.52 Frequency statistics Semantic similarity R semantic ]]> 0.74 Cosine similarity time decay factor D time ]]> 0.93 Exponential decay User preference coefficient P user ]]> 0.68 Behavior analysis

[0119] According to the formula G path = γ1S current + γ2W history + γ3R semantic + γ4D time + γ5P user + ζ1, each weight parameter is 0.3, 0.25, 0.2, 0.15, 0.1, and the path optimization bias term ζ1 is 0.02. The calculated intent determination gain is 0.72, which exceeds the set threshold 0.6, so the recognition weight of this intent path is increased.

[0120] In the 1024-dimensional intent vector space established in step S06, the current user's question identifies 3 parallel intent vectors, which form a parallel intent matrix M parallel is a 3x1024 matrix, and the parallel intent rank R parallel is 3. The hierarchical intent structure contains two levels of coarse granularity and fine granularity, involving 8 intent vectors. The hierarchical intent matrix M hierarchy is an 8x1024 matrix, and the hierarchical intent rank Rhierarchy The sequence of intentions generated by the user's 3 consecutive interactions constitutes an order intention matrix M sequence The sequence of intentions generated by the user's 3 consecutive interactions constitutes an order intention matrix M sequence The sequence of intentions generated by the user's 3 consecutive interactions constitutes an order intention matrix M

[0121] The intention semantic generation model of step S07 processes the intention recognition result based on the transformer architecture. The 12-layer structure of the encoder gradually extracts intention features, the first 4 layers extract basic semantic information, the middle 4 layers fuse context relationships, and the last 4 layers generate abstract representations. The decoder generates analysis text in an autoregressive manner, and the output result is "the user reflects that the mobile phone charging speed is slow and the heating problem, the main intention is to apply for return, and it is suggested to transfer to the after-sales return department for processing, the priority is high". According to the graph neural network structure index 1.73, the multi-head attention mechanism retains 12 attention heads. The micrograph structure index 1.86 corresponds to the message passing iteration number set to 4 times. The intention judgment accuracy 0.89 corresponds to the temperature coefficient set to 1.0.

[0122] The researchers further tested the performance of the system on different types of user problems, and the experimental results are shown in Table 3:

[0123] Table 3 Different intention type recognition performance table

[0124] Intent type Sample quantity Recognition accuracy Average response time (ms) Product consultation 1820 0.91 147 Logistics inquiry 1654 0.88 132 Return and exchange application 1432 0.89 156 After-sales service 1389 0.87 168 Technical support 1276 0.85 174 Complaint and suggestion 987 0.82 189 Account problem 934 0.86 143 Promotion activity 876 0.90 138

[0125] To verify the technical advantages of the present application, the researchers compared the traditional keyword matching and rule engine-based intention recognition method. The traditional method mainly relies on pre-set keyword dictionaries and manually written rules for intention classification, and cannot handle complex semantic relationships and context information. The comparison results on the same test data set are shown in Table 4:

[0126] Table 4 Technical method performance comparison table

[0127] Evaluation index Traditional keyword method Invention method Promotion range Overall recognition accuracy 0.74 0.87 17.6％ Complex intent processing rate 0.52 0.61 17.3％ Multi-intent recognition rate 0.43 0.51 18.6％ Average response time (ms) 89 152 -70.8％

[0128] The traditional technical means for solving user problem intent analysis mainly include keyword matching method and rule engine method. The keyword matching method determines the intent category by establishing an intent keyword dictionary and calculating the matching degree of user input text and keywords in the dictionary. This method is simple and direct, but cannot handle semantic changes and context relationships. The rule engine method identifies different intent patterns by manually writing if-then rules. Although it can handle certain logical relationships, the rule maintenance cost is high and the scalability is poor. The main technical progress of the present application relative to traditional methods is reflected in several aspects. First, the hierarchical intent classifier and graph neural network modeling can better capture the complex relationships between intents, and the overall recognition accuracy is improved from 74% to 87%, with an improvement of 17.6%. Second, the multi-task learning framework and micrograph structure analysis enable the system to simultaneously process main intent and sub-intent, and the complex intent processing rate is improved from 52% to 61%, with an improvement of 17.3%. Third, the intent vector space and multiple intent rank calculation methods enhance the system's ability to recognize parallel and hierarchical intents, and the multi-intent recognition rate is improved from 43% to 51%, with an improvement of 18.6%. Finally, the game theory model and intent path optimization ensure the optimality of parameter configuration and the adaptive adjustment of recognition strategy, significantly improving the robustness and generalization ability of the system. Although the response time of the present application method is increased compared with the traditional method, considering the significant improvement in recognition accuracy, this time cost is acceptable.

[0129] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1.A method for analyzing user problem intent under an AI platform, characterized in that, The application relates to an intelligent intent recognition method and device. The method comprises the following steps: preprocessing user input text, extracting semantic feature vectors, and calculating a graph neural network structure index; establishing a multi-task learning framework to simultaneously identify main and sub intents, generating a micrograph structure index to quantify the correlation strength between intents; modeling the intents by using a graph neural network and dividing the intents into multiple micrograph units to calculate intent determination accuracy; labeling historical interaction data to establish a training data set, coupling and adjusting the graph neural network parameters and the micrograph parameters by using a game theory model, and solving the optimal parameter combination by using a Nash equilibrium; solving the key path in the intent graph by using a maximum flow minimum cut algorithm, calculating the intent determination gain by using an intent path optimization function, and dynamically adjusting the intent recognition strategy; establishing an intent vector space, calculating parallel intent rank, hierarchical intent rank and sequential intent rank, and transmitting the recognized intent result to a corresponding service module to complete intelligent distribution; and processing the recognized intent result by using an intent semantic generation model to output an analysis result text of the user problem intent. The graph neural network structure index is a quantitative index for measuring the complexity of the intent graph network, and is obtained by calculating the comprehensive score of node degree distribution, clustering coefficient and path length. The micrograph structure index is a quantitative value of the structural complexity of each subgraph after the large-scale intent graph is decomposed into several small-scale subgraphs, and reflects the tightness of local intent relationships. The intent determination accuracy is the proportion of correctly recognized user real intents by the model, and is calculated by matching the predicted intent with the actual intent. The intent path optimization function is used to calculate the optimal path from user input to the final intent recognition result, and the input includes current intent intensity, historical interaction weight, semantic similarity, time decay factor and user preference coefficient, and the output is the numerical result of the intent determination gain. The parallel intent rank is the rank of a vector group composed of multiple intent vectors existing in parallel in the same input text, represents the maximum number of linearly independent intent vectors, and reflects the dimension complexity of the intent space. The hierarchical intent rank is the rank of a vector space composed of intent vectors at different abstraction levels in the intent classification system, and is used to measure the complexity of the intent hierarchy from abstraction to concreteness. The sequential intent rank is the rank of a vector group composed of an intent vector sequence arranged in time sequence in the continuous interaction process of the user, and reflects the independence degree in the intent evolution process. The structure of the intent semantic generation model is a text generation neural network based on a Transformer architecture, which comprises 12 layers of encoder and decoder structures, wherein the number of attention heads is dynamically adjusted according to the graph neural network structure index, and the number of message passing iterations is associated with the micrograph structure index. ​ 2.The AI platform-based user question intent analysis method of claim 1, wherein, The main intent, specifically the most important and core requirement expression in the user input text, represents the task to be completed or the service to be obtained by the user most, and is located at the top level in the intent hierarchy structure; the sub-intent, specifically the specific detail requirement or auxiliary requirement developed around the main intent, is located at the lower level of the main intent in the intent hierarchy structure, and is used to supplement and perfect the overall requirement expression of the user.

Images