A flow data-oriented customer intention real-time identification intervention method and system

Abstract

The application discloses a kind of customer intention real-time identification intervention methods and systems for streaming data, method includes: receiving multimodal customer interaction stream data, constructs space-time behavior tensor, and tensor element is mapped as semantic-sentiment feature vector;The semantic-sentiment feature vector is carried out temporal logic analysis, constructs intention state conversion diagram, and the diagram structure is decomposed into several atomic intention units;Atomic intention unit is extracted using graph neural network feature, and atomic intention feature vector is generated;Atomic intention feature vector is input into pre-trained hybrid expert intention recognition model, and customer intention recognition result is generated;Intervention decision is made to customer intention recognition result based on reinforcement learning strategy optimizer, and finally the optimal intervention scheme containing intervention timing, mode and content is output. Using the embodiment of the application, customer intention recognition accuracy can be improved, and the intervention strategy of explainable and landable can be generated, and the real-time and individual level of customer service is improved.

Description

Technical Field

[0001] This invention belongs to the field of intent recognition technology, and in particular to a method and system for real-time recognition and intervention of customer intent for streaming data. Background Technology

[0002] With the widespread adoption of digital services, customer interaction scenarios are becoming increasingly complex. Real-time analysis and intent recognition of multimodal streaming data (such as voice, text, and behavioral logs) have become crucial for improving service experience. Traditional methods, typically based on static data modeling, struggle to capture the temporal dynamics and multimodal relationships in customer interactions, leading to lags or biases in intent recognition. Existing technologies lack fine-grained characterization of intent evolution in rule-based or single-modal recognition models, while deep learning methods, although capable of extracting features, fail to address the interpretability of intent primitives and the optimization of real-time intervention decisions. Furthermore, the coupling relationship between spatiotemporal behavioral patterns of customer intent and emotional semantics has not been effectively modeled, and intervention strategies often rely on human experience, lacking dynamic adjustment capabilities. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for real-time identification and intervention of customer intent based on streaming data, in order to overcome the shortcomings of the prior art, improve the accuracy of customer intent identification, generate interpretable and implementable intervention strategies, and improve the real-time and personalized level of customer service.

[0004] One embodiment of this application provides a method for real-time identification and intervention of customer intent based on streaming data, the method comprising:

[0005] Receive multimodal customer interaction stream data, construct a spatiotemporal behavior tensor, and map the tensor elements to semantic-sentiment feature vectors, wherein the semantic-sentiment feature vectors are associated with the temporal evolution dimension of customer behavior;

[0006] Temporal logic analysis is performed on the semantic-emotion feature vector to construct an intent state transition graph. The graph structure is decomposed into several atomic intent units, where each atomic intent unit represents an indivisible customer intent primitive.

[0007] The atomic intention unit is used to extract features using a graph neural network to generate an atomic intention feature vector, wherein the dimension of the atomic intention feature vector represents the semantic embedding space of the intention primitive.

[0008] The atomic intent feature vector is input into a pre-trained hybrid expert intent recognition model to generate a customer intent recognition result, wherein the intent recognition model dynamically selects an expert network suitable for the current input through a gating mechanism;

[0009] The intervention decision is made based on the customer intent recognition results using a reinforcement learning strategy optimizer, generating a Pareto optimal intervention strategy set. The feasibility is verified by combining the knowledge graph, and finally the optimal intervention plan containing the timing, method and content of the intervention is output.

[0010] Optionally, the step of receiving multimodal customer interaction stream data, constructing a spatiotemporal behavior tensor, and mapping tensor elements to semantic-sentiment feature vectors, wherein the semantic-sentiment feature vectors are associated with the temporal evolution dimension of customer behavior, including:

[0011] Receive multimodal customer interaction stream data, align different modal data by timestamp, filter out noisy data using rules, and output a spatiotemporally synchronized multimodal interaction dataset;

[0012] Temporal features are extracted from a multimodal interaction dataset. Specifically, text word frequency, speech emotion intensity, and click frequency are counted in a 5-second sliding window. The time difference sequence of behaviors within the window is calculated, and the temporal feature matrix is ​​output.

[0013] Using customer ID as the spatial dimension, time window as the temporal dimension, and temporal features as the attribute dimension, a three-dimensional spatiotemporal behavior tensor is constructed. The tensor element values ​​are the standardized strengths of the corresponding features, and the initial spatiotemporal behavior tensor is output.

[0014] Feature extraction is performed on the initial spatiotemporal behavior tensor. Text features are extracted using BERT to extract semantic vectors, speech features are extracted using a sentiment classification model to extract sentiment vectors, and behavior features are reduced in dimensionality using PCA. These are then concatenated into a 64-dimensional semantic-sentiment feature vector. The vector dimension corresponds to the temporal evolution of the behavior, generating a semantic-sentiment feature vector set with temporal correlation.

[0015] Optionally, the step of performing temporal logic analysis on the semantic-emotion feature vector to construct an intent state transition graph decomposes the graph structure into several atomic intent units, wherein each atomic intent unit represents an indivisible customer intent primitive, including:

[0016] The semantic-sentiment feature vector set is sorted by time window, and the temporal logical operator is used to label the temporal relationship between the vectors, and the temporal relationship labeling table is output.

[0017] Intent state nodes are constructed based on a temporal relationship annotation table. Each node contains a feature vector, sentiment label, and core behavior within a window. The node ID is incremented by time, and the output is a set of state nodes.

[0018] Calculate the transition probability between state nodes. When the transition probability is greater than 20%, establish a directed edge with the edge weight equal to the transition probability, and output the intention state transition graph.

[0019] Perform community detection on the intent state transition graph to identify strongly connected components. Each component corresponds to a composite intent, and the composite intent component set is output.

[0020] The composite intent component set is decomposed and split into indivisible primitives using the minimum description length criterion. Each primitive contains a single behavioral, emotional, and semantic feature, generating atomic intent units.

[0021] Optionally, the step of using a graph neural network to extract features from the atomic intention units and generate atomic intention feature vectors, wherein the dimension of the atomic intention feature vectors represents the semantic embedding space of the intention primitives, including:

[0022] Construct an association graph of atomic intention units, where nodes are atomic intention units, edges are co-occurrence relationships, and edge weights are co-occurrence frequencies, thus obtaining the atomic intention association graph;

[0023] Initialize the node features of the atomic intent association graph, using the semantic-sentiment feature vector of each atomic intent unit as the initial node feature, and supplement the unit's duration and trigger probability to obtain the supplementary node feature matrix;

[0024] Based on the supplementary node feature matrix, a graph attention network is used for feature extraction. The attention of a node to its neighbors is calculated through a multi-head attention mechanism. The neighbor features are aggregated to update the node's own features. After three layers of GAT iteration, the updated node features are output.

[0025] The updated node features are globally pooled and combined with the temporal position features of the units, then mapped to a 32-dimensional vector through a fully connected layer, which serves as the candidate atomic intent feature vector.

[0026] Calculate the cosine similarity between candidate atomic intention feature vectors. When the similarity is greater than 0.85, merge redundant vectors and normalize them to make the vector magnitude 1, thus generating the final atomic intention feature vector.

[0027] Optionally, the step of inputting the atomic intent feature vector into a pre-trained hybrid expert intent recognition model to generate a customer intent recognition result, wherein the intent recognition model dynamically selects an expert network suitable for the current input through a gating mechanism, including:

[0028] A hybrid expert intent recognition model is constructed, which includes 5 expert networks and 1 gating network. The expert networks adopt an MLP structure, and the gating network takes atomic intent feature vectors as input and outputs expert selection weights, thus obtaining the model architecture.

[0029] Input the atomic intent feature vector into the gating network, output the weight of each expert, include experts with weights > 0.3 into the candidate set, and output the expert candidate set and weights;

[0030] The feature vectors are inferred using an expert network in the candidate set, and the intent category and confidence level are output. The results are then fused by a gated network according to the weights, and the fused recognition result is output.

[0031] When the highest confidence level of the fusion recognition result is greater than 0.8, the customer intent is directly output; otherwise, the final customer intent recognition result is determined by majority voting.

[0032] Optionally, the step of using a reinforcement learning-based policy optimizer to make intervention decisions based on the customer intent recognition results generates a Pareto-optimal set of intervention strategies. Feasibility is then verified using a knowledge graph, ultimately outputting an optimal intervention plan that includes the timing, method, and content of the intervention, including:

[0033] The customer intent recognition result is used as the state input for reinforcement learning. The action space is defined as the intervention method, the reward function is the increase in customer satisfaction after the intervention, the strategy optimizer is trained, and the intervention strategy generation model is output.

[0034] The intervention strategy generation model outputs 10 candidate intervention strategies, calculates the effect index of each group, and uses Pareto optimization to screen non-dominated solutions to obtain the Pareto optimal intervention strategy set.

[0035] Input the optimal set of intervention strategies into the customer knowledge graph, verify the feasibility of the strategies, eliminate infeasible strategies, and output a subset of feasible strategies.

[0036] The feasible strategy subset is sorted by comprehensive score, and the strategy with the highest score that includes the timing, method and content of intervention is selected as the optimal intervention plan.

[0037] Another embodiment of this application provides a real-time customer intent recognition and intervention system for streaming data, the system comprising:

[0038] The mapping module is used to receive multimodal customer interaction stream data, construct a spatiotemporal behavior tensor, and map the tensor elements to semantic-sentiment feature vectors, wherein the semantic-sentiment feature vectors are associated with the temporal evolution dimension of customer behavior.

[0039] The decomposition module is used to perform temporal logic analysis on the semantic-emotion feature vector, construct an intent state transition graph, and decompose the graph structure into several atomic intent units, wherein each atomic intent unit represents an indivisible customer intent primitive.

[0040] The extraction module is used to extract features from the atomic intention units using a graph neural network to generate atomic intention feature vectors, wherein the dimension of the atomic intention feature vectors represents the semantic embedding space of the intention primitives.

[0041] The recognition module is used to input the atomic intent feature vector into a pre-trained hybrid expert intent recognition model to generate a customer intent recognition result, wherein the intent recognition model dynamically selects an expert network suitable for the current input through a gating mechanism;

[0042] The intervention module is used to make intervention decisions based on the customer intent recognition results using a reinforcement learning strategy optimizer, generate a Pareto optimal intervention strategy set, perform feasibility verification using a knowledge graph, and finally output the optimal intervention plan including the timing, method, and content of the intervention.

[0043] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0044] Another embodiment of this application provides an electronic device including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the method described in any of the preceding claims.

[0045] Compared with existing technologies, this invention provides a real-time customer intent recognition and intervention method for streaming data. It receives multimodal customer interaction stream data, constructs a spatiotemporal behavior tensor, and maps tensor elements to semantic-sentiment feature vectors. Temporal logic analysis is performed on the semantic-sentiment feature vectors to construct an intent state transition graph, which is then decomposed into several atomic intent units. A graph neural network is used to extract features from the atomic intent units, generating atomic intent feature vectors. These atomic intent feature vectors are input into a pre-trained hybrid expert intent recognition model to generate customer intent recognition results. An intervention decision is made based on the customer intent recognition results using a reinforcement learning policy optimizer, ultimately outputting an optimal intervention plan that includes the timing, method, and content of the intervention. This improves the accuracy of customer intent recognition, generates interpretable and implementable intervention strategies, and enhances the real-time and personalized level of customer service. Attached Figure Description

[0046] Figure 1 A hardware structure block diagram of a computer terminal for a real-time customer intent recognition and intervention method for streaming data provided in an embodiment of the present invention;

[0047] Figure 2 A flowchart illustrating a real-time customer intent identification and intervention method for streaming data, provided in an embodiment of the present invention;

[0048] Figure 3 A flowchart illustrating another real-time customer intent identification and intervention method for streaming data provided in an embodiment of the present invention;

[0049] Figure 4A flowchart illustrating another method for real-time identification and intervention of customer intent based on streaming data, provided in an embodiment of the present invention;

[0050] Figure 5 This is a schematic diagram of the structure of a real-time customer intent recognition and intervention system for streaming data provided in an embodiment of the present invention. Detailed Implementation

[0051] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0052] This invention first provides a method for real-time identification and intervention of customer intent based on streaming data. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0053] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a real-time customer intent recognition and intervention method based on streaming data, provided in an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0054] Non-volatile storage media can store operating systems and computer programs. These computer programs include program instructions that, when executed, cause the processor to perform any real-time client intent recognition and intervention method oriented towards streaming data.

[0055] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0056] Internal memory provides an environment for the execution of computer programs in non-volatile storage media. When executed by a processor, the computer program enables the processor to perform any real-time identification and intervention method for client intent oriented towards streaming data.

[0057] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0058] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0059] See Figures 2-4 The present invention provides a method for real-time identification and intervention of customer intent based on streaming data, which may include the following steps:

[0060] S201, Receive multimodal customer interaction stream data, construct a spatiotemporal behavior tensor, and map the tensor elements to semantic-sentiment feature vectors, wherein the semantic-sentiment feature vectors are associated with the temporal evolution dimension of customer behavior; specifically, this may include:

[0061] S2011 receives multimodal customer interaction stream data, aligns different modal data by timestamp, removes noisy data by rule filtering, and outputs a spatiotemporally synchronized multimodal interaction dataset;

[0062] Multimodal customer interaction flow data encompasses various types of data generated during customer-business interactions, typically including text data (such as online chat logs, email content, and entered query keywords), voice data (such as customer service call recordings and voice commands), and behavioral data (such as webpage click locations and times, app dwell time, and button operation records). For example, in the customer interaction data of an e-commerce platform, text data might be chat messages asking "When will this product ship?", voice data might be a customer's voice message saying "I want to return this item" when calling customer service, and behavioral data might be a customer's three clicks and two-minute dwell time record on the product details page.

[0063] Timestamp alignment is the core of achieving spatiotemporal synchronization of multimodal data. Each type of data carries a timestamp upon generation. Text messages have timestamps accurate to milliseconds (e.g., "2025-08-01 10:05:23.123"), voice data records time frame by frame (each frame is 20 milliseconds, timestamps such as "2025-08-01 10:06:00.000" and "2025-08-01 10:06:00.020"), and behavioral data has click timestamps accurate to milliseconds (e.g., "2025-08-01 10:05:30.456"). During alignment, the timestamps of different modalities are interpolated and matched based on the most precise millisecond-level timestamp. For example, the timestamp of a certain frame of voice data is "10:06:00.010", which needs to be bound to text or behavioral data within the same millisecond; if the timestamp of a certain modality is missing, it is supplemented by linear interpolation of the time difference between adjacent data (e.g., the previous line has a timestamp of "10:05:20.000", the next line has a timestamp of "10:05:25.000", and the missing timestamp in between is filled by incrementing by one second).

[0064] Rule-based filtering is used to remove noisy data. Specific rules include: (1) removing duplicate data, such as keeping only the first message if a customer sends the same "Are you there?" message three times in a row; (2) filtering invalid format data, such as text messages containing only meaningless symbols like "!!!", and silent segments in voice data with a signal-to-noise ratio below 20dB (judged as having no valid information); (3) removing abnormal time data, such as data with timestamps earlier than the system's launch time (e.g., "2020-01-01 00:00:00", the system was launched in 2023) or later than the current time. For example, in a customer's behavior data, a click record with a timestamp of "2025-08-02 15:30:00" (the current time is 2025-08-01) is judged as abnormal and filtered.

[0065] After timestamp alignment and rule filtering, the output spatiotemporally synchronized multimodal interaction dataset is arranged in chronological order. Each record contains a unified timestamp, modality type, data content, and associated customer ID. For example, one record is "Timestamp: 2025-08-01 10:05:23.123, Modality: Text, Content: 'When will this product be shipped?', Customer ID: C001"; another synchronized record is "Timestamp: 2025-08-01 10:05:23.123, Modality: Behavior, Content: Clicked the 'Logistics Inquiry' button, Customer ID: C001". This ensures that different modality data are associated on the same time dimension, laying the foundation for subsequent feature extraction.

[0066] S2012 extracts temporal features from a multimodal interaction dataset, including text word frequency, speech emotion intensity, and click frequency statistics in a 5-second sliding window, calculating the time difference sequence of behaviors within the window, and outputting a temporal feature matrix.

[0067] Temporal feature extraction uses a 5-second sliding window (the window size is set based on the instantaneous nature of customer interaction; too short a window will lead to feature fragmentation, while too long a window will mask short-term behavioral changes), with a window sliding step of 1 second (to ensure the feature continuity of consecutive windows). Feature quantization is performed on the multimodal data within each window.

[0068] Text frequency statistics, based on the text data within the window, first perform word segmentation (e.g., using the jieba word segmentation tool to segment "When will this product be shipped?" into "this", "product", "when", and "shipped"), filter out stop words (such as "this" and "when"), retain the core words ("product" and "shipped"), and then calculate the word frequency (number of times a word appears / total number of words in the window). For example, if a customer sends "When will the product be shipped? When will the product arrive?" within a 5-second window, after word segmentation, the core words are "product", "shipped", "time", and "arrived". "Product" appears 2 times, the total number of words is 4, and the word frequency is 2 / 4 = 0.5.

[0069] Voice emotion intensity is extracted using a pre-trained voice emotion classification model (such as a CNN-based emotion recognition model). The model input is a voice segment (5 seconds of audio) within a window, and the output is an emotion intensity value of 0-1 (0 represents negative, 1 represents positive). For example, the emotion intensity of a customer's voice "I am very dissatisfied with this service" is 0.2 (leaning towards negative), and the intensity of "This solution is good" is 0.8 (leaning towards positive).

[0070] Click frequency is the ratio of the number of clicks made by the customer within the window to the duration of the window (clicks per second). For example, if a customer clicks the "Add to Cart" button 3 times within a 5-second window, the click frequency = 3 / 5 = 0.6 clicks per second.

[0071] The time difference sequence of actions within a window is used to characterize the temporal intervals of actions. For actions that occur consecutively within a window (such as multiple clicks or consecutive message sending), the time difference (in seconds) between two adjacent actions is calculated. For example, if the customer clicks for the first time at 0 seconds, the second time at 2 seconds, and the third time at 4 seconds, the time difference sequence is [2-0=2, 4-2=2]. If there is only one action within the window, the time difference sequence is empty and filled with 0.

[0072] The temporal features extracted from each 5-second window include: text word frequency (the frequency of the top 10 high-frequency core words, forming a 10-dimensional feature), voice emotion intensity (1-dimensional), click frequency (1-dimensional), and time difference sequence (the first 5 time differences, padded with 0 if less than 5, forming a 5-dimensional feature), for a total of 17 dimensions. The features from all windows are arranged in chronological order to form a temporal feature matrix, with the number of rows equal to the number of windows and 17 columns (feature dimensions). For example, the feature matrix for a customer's first 3 windows is: the first row [0.5,0,0,...,0.6,2,2,0,0,0] ("product" word frequency 0.5, click frequency 0.6, time difference [2,2]), the second row [0,0.3,0,...,0.2,1,3,0,0,0] (voice emotion intensity 0.3, click frequency 0.2), clearly showing the trend of customer behavior over time.

[0073] S2013 constructs a three-dimensional spatiotemporal behavior tensor with customer ID as the spatial dimension, time window as the time dimension, and temporal features as the attribute dimension. The tensor element values ​​are the standardized strengths of the corresponding features, and the initial spatiotemporal behavior tensor is output.

[0074] The three dimensions of the three-dimensional spatiotemporal behavior tensor are spatial dimension, time dimension, and attribute dimension, which together characterize the spatiotemporal distribution of customer behavior.

[0075] The spatial dimension is identified by customer ID, with each ID representing an independent customer. For example, the spatial dimension containing customer IDs "C001", "C002", and "C003" has a size of 3 (i.e., 3 customers). If analyzing the intent of a group of customers, the spatial dimension can be expanded into customer group tags (such as "new customer" or "old customer"), but here we focus on individuals, so we use ID as the unit.

[0076] The time dimension uses 5-second sliding windows, numbered sequentially by time (e.g., window 1 corresponds to seconds 0-5, window 2 to seconds 1-6, and so on, until the last window). The size of the time dimension is equal to the total number of windows. For example, if a customer interaction lasts for 100 seconds, 96 windows are generated (100-5+1=96), and the time dimension size is 96.

[0077] The attributes are 17-dimensional temporal features (text word frequency, speech sentiment intensity, etc.), each feature represents an attribute, and the attribute dimension is 17.

[0078] Tensor element values ​​are the standardized strengths of the corresponding features. Min-max standardization is used to map feature values ​​to the [0,1] interval, with the formula: Standardized value = (Original value - Minimum feature value) / (Maximum feature value - Minimum feature value). For example, the original value range of the "click frequency" feature is 0-2 times / second, and the click frequency of a certain window is 0.6 times / second, so the standardized value = (0.6-0) / (2-0) = 0.3; the original value of "voice emotion intensity" is already in the [0,1] interval, and remains unchanged after standardization. For missing values ​​(such as a window with no text data, where the word frequency feature is empty), they are filled with 0 before standardization.

[0079] The initial spatiotemporal behavior tensor is formatted as (spatial dimension, temporal dimension, attribute dimension). For example, a tensor containing 3 customers, 96 windows for each customer, and 17 attributes would have the shape (3, 96, 17). An element (C001, 5, 3) in the tensor represents the standardized value of the third attribute (e.g., the frequency of the word "shipment") for customer "C001" in the fifth window. Assuming a value of 0.7, this reflects a high level of attention the customer pays to "shipment" in that window. This three-dimensional structure can distinguish behavioral differences among different customers (spatial dimension), capture the temporal changes in the behavior of the same customer (temporal dimension), and refine the characteristic composition of behavior (attribute dimension), providing a structured data foundation for subsequent feature mapping.

[0080] S2014 extracts features from the initial spatiotemporal behavior tensor. Text features are extracted using BERT to extract semantic vectors, speech features are extracted using a sentiment classification model to extract sentiment vectors, and behavior features are reduced in dimensionality using PCA. These are then concatenated into a 64-dimensional semantic-sentiment feature vector. The vector dimensions correspond to the temporal evolution of the behavior, generating a set of semantic-sentiment feature vectors with temporal correlation.

[0081] Feature extraction of the initial spatiotemporal behavior tensor needs to be processed by modality splitting and then fused into a unified semantic-sentiment feature vector, while preserving temporal evolution information.

[0082] Text feature extraction employs a BERT pre-trained model (such as bert-base-uncased). For text data within each time window (e.g., chat messages sent by customers), the text is converted into a token sequence (maximum length 512), which is then input into the [CLS] position of the BERT model to output a 768-dimensional semantic vector (containing deep semantic information of the text). For example, the BERT semantic vector for the text "When will this product be shipped?" is 768-dimensional, with higher values ​​for dimensions related to "logistics" and "time," reflecting the customer's focus on shipping time. To reduce dimensionality, the 768-dimensional vector is L2 normalized, and the top 20 dimensions are used as the text semantic features (retaining more than 90% of the semantic information).

[0083] Speech feature extraction uses a fine-tuned speech emotion classification model (such as a model based on Wav2Vec2). It takes a speech segment within an input window as input and outputs a 20-dimensional emotion vector (more detailed than a 1-dimensional intensity vector, including probabilities for sub-emotions such as joy, anger, and confusion). For example, in the emotion vector of a customer's voice saying "I haven't received the goods yet, I'm very anxious," the "anger" dimension value is 0.6, and the "confusion" dimension value is 0.3, clearly distinguishing the emotion type and intensity.

[0084] Behavioral features include non-text / speech features such as click frequency and time difference sequences, totaling 6 dimensions (split from 17 dimensions, excluding 10 dimensions of text word frequency and 1 dimension of speech intensity). Principal component analysis (PCA) is used to reduce the dimensionality of these 6-dimensional behavioral features, retaining principal components that contribute up to 95% of the cumulative variance, typically 4 dimensions (the 6-dimensional features have high correlation, while 4 dimensions retain the main information). For example, the correlation between click frequency and time difference sequences is 0.8. After PCA, the first principal component comprehensively reflects behavioral activity, while the second principal component reflects the regularity of behavioral intervals.

[0085] After concatenating the text semantic features (20-dimensional), speech sentiment vector (20-dimensional), and behavioral PCA features (4-dimensional), the total dimensions are 44. These are then mapped to 64 dimensions through a fully connected layer (using ReLU activation function) to supplement the dimensions and enhance expressiveness, forming a semantic-sentiment feature vector. The vector's dimensions are related to the temporal evolution of behavior: the first 32 dimensions focus on short-term behavior (features from the most recent 3 windows), while the latter 32 dimensions focus on long-term trends (cumulative features from the past 10 windows). For example, the 5th dimension corresponds to the semantic strength of "shipping" in the most recent window, and the 40th dimension corresponds to the average click frequency of the past 10 windows, allowing the vector to reflect both immediate intent and long-term behavioral patterns.

[0086] The semantic-sentiment feature vector set with temporal association is arranged in chronological order according to time windows. Each window corresponds to a 64-dimensional vector, and the vectors are associated with each other through timestamps (vectors of adjacent windows are continuous in the temporal dimension). For example, the 5th window vector for customer C001 is 64-dimensional, with the high values ​​of the first 32 dimensions concentrated in the "logistics" related dimensions, and the last 32 dimensions showing an upward trend in recent click frequency. The overall vector reflects that the customer's attention to logistics is increasing, providing high-dimensional feature support for subsequent intent state analysis.

[0087] S202, perform temporal logic analysis on the semantic-emotion feature vector, construct an intent state transition graph, and decompose the graph structure into several atomic intent units, wherein each atomic intent unit represents an indivisible customer intent primitive;

[0088] Specifically, the semantic-sentiment feature vector set can be sorted by time window, and temporal logical operators can be used to label the temporal relationship between the vectors to output a temporal relationship labeling table;

[0089] Data sorting:

[0090] The semantic-sentiment feature vector set is generated as described above, with each vector corresponding to customer interaction features (such as a fusion representation of text, voice, and behavior) within a 5-second sliding window. First, the vectors are arranged in ascending order of timestamps to form a time series (e.g., window 1 corresponds to t0-t5, window 2 corresponds to t1-t6, and so on). For example, the window sequence for a certain e-commerce customer:

[0091] Window 1 (t0-t5): Sends the text "When will the goods be shipped?", with semantic vectors containing "logistics" and "time", and sentiment neutral (no voice, no obvious emotion in the text).

[0092] Window 2 (t1-t6): Sends the text "Has it not been shipped yet?", with semantic vectors containing "logistics" and "urging", and sentiment leaning towards negative (textual anxiety).

[0093] Window 3 (t2-t7): Click the "Logistics Inquiry" button. The semantic vector is mainly based on behavioral features and is sentiment-neutral.

[0094] Temporal logic operator selection:

[0095] Temporal logical operators (such as "next" □, "until" U, "once" P, "parallel" ∧) are used to label the temporal relationships between vectors:

[0096] □ (Next): The direct successor window of the current window satisfies the intent association. For example, Window 1 → Window 2, the customer goes from "Inquire about delivery" to "Extend order", marked as Node1 □ Node2 (continuous evolution of intent).

[0097] U (Until): The current window continues until a new intent appears in a subsequent window. For example, if window 1 (consultation) continues until window 3 (logistics inquiry), it is marked as Node1 U Node3 (consultation behavior until the inquiry occurs).

[0098] P(Previous): The current window's history windows have related intents. For example, if window 3 (query) has a query in its history window 1, it is marked with P(Node1,Node3) (Previous query, current query).

[0099] ∧(Parallel): Multimodal behaviors coexist within the same window (such as sending text and clicking simultaneously), but since the window slides for 5 seconds, it is mainly used for the synchronization of cross-modal features (such as text consultation and clicking logistics buttons in the same window).

[0100] Annotation process:

[0101] Traverse all window pairs (adjacent or spaced, based on time association) and analyze intent evolution:

[0102] Window 1 → Window 2: The text changes from "consultation" to "order follow-up", and the sentiment changes from none to negative. Mark (Node1, Node2, □) (next relationship).

[0103] Window 2 → Window 3: After expediting the order, click Query and mark (Node2, Node3, □) (Next Relationship).

[0104] Window 1 and Window 3: Query after consultation, separated by 1 window, labeled (Node1, Node3, P) (formerly related).

[0105] Finally, a temporal relationship labeling table is formed, which records the operators and associated logic of each pair of windows.

[0106] Intent state nodes are constructed based on a temporal relationship annotation table. Each node contains a feature vector, sentiment label, and core behavior within a window. The node ID is incremented by time, and the output is a set of state nodes.

[0107] Node feature definition:

[0108] Feature vector: Directly uses the 64-dimensional semantic-sentiment feature vector (integrating text semantics, voice sentiment, and behavioral features) generated in the above steps. For example, in the vector of window 1, the "logistics" and "time" dimensions have higher values, reflecting the consultation intent.

[0109] Emotional labeling: Integrates speech emotion classification (e.g., using the 20-dimensional emotion probability output by the Wav2Vec2 model, taking the maximum value as the corresponding label, such as "anxious") and text sentiment analysis (e.g., the VADER algorithm, which gives -0.6 points to "Has it not been shipped yet?" and labels it as "anxious").

[0110] Core behavior: Determined through behavioral characteristics (click frequency, time difference sequence) and text / voice content:

[0111] Window 1: The text contains "shipping", and the core behavior is "inquire about shipping time".

[0112] Window 2: The text contains "not shipped yet", and the core action is "order expediting".

[0113] Window 3: Click "Logistics Inquiry", the core action is "Inquire about logistics status".

[0114] Node creation and ID assignment:

[0115] Assign unique IDs to nodes in time window order (e.g., Node1 corresponds to window 1, Node2 corresponds to window 2), and store the following for each node:

[0116] Node 1:

[0117] Feature vector: [0.8 (logistics semantics), 0.2 (time semantics), ...] (64 dimensions);

[0118] Emotional label: "Neutral" (the text shows no obvious emotion);

[0119] Core action: "Inquire about delivery time".

[0120] Node2:

[0121] Feature vector: [0.7 (logistics semantics), 0.6 (urging semantics), ...];

[0122] Emotional tag: "Anxiety" (textual anxiety);

[0123] Core behavior: "Order expediting".

[0124] All nodes form a set of state nodes, clearly depicting the intent "snapshot" for each time window.

[0125] Calculate the transition probability between state nodes. When the transition probability is greater than 20%, establish a directed edge with the edge weight equal to the transition probability, and output the intention state transition graph.

[0126] Transition probability calculation:

[0127] Count the number of state transitions for all customers (across multiple interaction sessions). Assuming we analyze interaction data from 100 customers and extract 1000 state nodes:

[0128] NodeA (consultation and shipment, corresponding to Node1) appears 200 times, of which 50 times it is transferred to NodeB (order expediting, corresponding to Node2) → conversion probability = 50 / 200 = 25% (>20%, condition met).

[0129] NodeB (order expediting) occurred 150 times, of which 40 times were transferred to NodeC (logistics inquiry, corresponding to Node3) → conversion probability = 40 / 150≈26.7% (>20%, condition met).

[0130] NodeC (logistics query) appeared 180 times, and only 20 times it was transferred to NodeA → conversion probability ≈ 11.1% (<20%, not satisfied).

[0131] Directed edge construction:

[0132] For node pairs with a transition probability greater than 20%, construct directed edges with edge weights equal to the probability values:

[0133] NodeA→NodeB: Weight 0.25 (25% transition probability);

[0134] NodeB→NodeC: weight 0.267 (26.7% transition probability).

[0135] The final result is an intent state transition graph, where nodes represent intent states and edges represent transition relationships and probabilities, visually demonstrating the high-frequency evolution path of "consultation → order reminder → inquiry".

[0136] Perform community detection on the intent state transition graph to identify strongly connected components. Each component corresponds to a composite intent, and the composite intent component set is output.

[0137] Community detection algorithm:

[0138] The Louvain algorithm (based on modularity optimization, measuring the difference between edge density within and between communities) is used to divide the intent state transition graph into multiple communities:

[0139] Logistics tracking community: NodeA (consultation), NodeB (order expediting), and NodeC (query) have high edge weights (frequent mutual transfers), forming a strongly connected component (any two nodes can reach each other, such as NodeC can return to NodeA through the P operator of the history window, which means that the customer may have "consulted repeatedly → expedited orders → queried repeatedly").

[0140] Product research community: Another set of nodes (such as Node4 "Consultation Specifications" and Node5 "Comparison Models") form an independent community, corresponding to the "product research" intent.

[0141] Composite Intent Mapping:

[0142] Each strongly connected component corresponds to a composite intent:

[0143] Logistics tracking component: includes NodeA, NodeB, NodeC, and edges A→B and B→C, representing the composite intent of "logistics tracking" (a continuous behavior from inquiry to order reminder to query).

[0144] Product research components: including Node4 and Node5, representing the combined intent of "product research" (parallel actions of inquiring about specifications and comparing models).

[0145] Output a composite intent component set, recording the node, edge, and intent type of each component.

[0146] The composite intent component set is decomposed and split into indivisible primitives using the minimum description length criterion. Each primitive contains a single behavioral, emotional, and semantic feature, generating atomic intent units.

[0147] Minimum Description Length (MDL) Criterion:

[0148] The MDL formula is that the description length = model complexity + data fitting error. Try to split the composite intention components:

[0149] Split of the logistics tracking component:

[0150] Primitive 1: NodeA (consult about shipment) → Behavior "consult about the shipment time", Emotion "neutral", Semantics "logistics, time";

[0151] Primitive 2: NodeB (urge for order) → Behavior "urge for order", Emotion "anxious", Semantics "logistics, urge";

[0152] Primitive 3: NodeC (query logistics) → Behavior "query logistics status", Emotion "neutral", Semantics "logistics, query".

[0153] Calculate the MDL after splitting. If it is smaller than the MDL of the composite state (more concise and better at fitting data), then the split is effective.

[0154] Judgment of non - decomposability:

[0155] If a primitive cannot be split into simpler intentions (for example, "urge for order" cannot be split into "urge" and "order" because their semantic association is tight and the intention integrity is lost after splitting), then it is determined as an atomic intention unit:

[0156] Atomic intention Figure 1 : {Behavior: consult about the shipment time, Emotion: neutral, Semantics: logistics, time};

[0157] Atomic intention Figure 2 : {Behavior: urge for order, Emotion: anxious, Semantics: logistics, urge};

[0158] Atomic intention Figure 3 : {Behavior: query logistics, Emotion: neutral, Semantics: logistics, query}.

[0159] Generate atomic intention units: All the split primitives form a set of atomic intention units. Each unit represents an indecomposable customer intention primitive, providing a basis for subsequent feature extraction by the graph neural network.

[0160] S203. Use the graph neural network to extract features from the atomic intention units to generate atomic intention feature vectors, where the dimension of the atomic intention feature vector represents the semantic embedding space of the intention primitive;

[0161] Specifically, an association graph of atomic intention units can be constructed, where the nodes are atomic intention units, the edges are co - occurrence relationships, and the edge weights are co - occurrence frequencies, to obtain an atomic intention association graph;

[0162] Atomic intent units are the smallest basic units of customer intent, such as "inquire about shipping time," "expedite order," "check logistics status," "inquire about product specifications," and "compare prices." Each unit corresponds to a single, indivisible intent. The core of constructing a relationship graph is to capture the co-occurrence relationships of these units in the customer interaction process—that is, the probability that different atomic intents appear simultaneously or successively in the same customer's continuous interactions.

[0163] Co-occurrence is determined based on customer interaction logs: if two atomic intent units appear in the same customer's interaction sequence within a time interval of ≤30 seconds (a reasonable threshold for the coherence of short-term customer intents), they are considered co-occurring. For example, if a customer first sends "When will this product be shipped?" (corresponding to the "Inquire about shipping time" unit), and then clicks the "Logistics Inquiry" button within 30 seconds (corresponding to the "Check Logistics Status" unit), then these two units constitute a co-occurrence relationship.

[0164] The formula for calculating edge weight (co-occurrence frequency) is: Co-occurrence frequency = Number of times the two units co-occur ÷ Sum of the total number of times the two units appear. Assuming we analyze 1000 customer interaction sequences, "Inquire about shipping time" appears 300 times, "Inquire about logistics status" appears 250 times, and the two co-occur 150 times. Therefore, the co-occurrence frequency = 150 ÷ ​​(300 + 250) = 150 ÷ ​​550 ≈ 0.27 (rounded to two decimal places), meaning the edge weight is 0.27. If the co-occurrence frequency = 0 (the two units never co-occur), then no edge is created.

[0165] In the final atomic intent association graph, nodes represent all atomic intent units (e.g., 5 nodes: A = inquire about shipping time, B = urge order, C = check logistics status, D = inquire about product specifications, E = compare prices). Edges and weights reflect co-occurrence strength: AC edge weight 0.27 (high-frequency co-occurrence), BC edge weight 0.22 (second-highest frequency), AB edge weight 0.15 (medium frequency), DE edge weight 0.3 (high-frequency co-occurrence of product research related units), while AD edge weight 0 (no co-occurrence, no connection). This graph clearly presents the association pattern of intent primitives, providing a structural foundation for subsequent feature extraction.

[0166] Initialize the node features of the atomic intent association graph, using the semantic-sentiment feature vector of each atomic intent unit as the initial node feature, and supplement the unit's duration and trigger probability to obtain the supplementary node feature matrix;

[0167] The initial node features of each atomic intent unit are based on the 64-dimensional semantic-sentiment feature vector generated in the above steps. This vector integrates the unit's textual semantics (such as keywords like "shipping" and "logistics"), sentiment (such as the anxious emotion of "urging order"), and behavioral features (such as click frequency). For example, in the initial feature vector of "inquiring about shipping time", the dimensions related to "logistics" and "time" have high values ​​(such as 0.8 and 0.7), the sentiment dimension is close to neutral (0.5), and the behavioral dimension reflects low-frequency clicks (0.2).

[0168] Supplementary features include "duration" and "trigger probability" to enhance the completeness of node features:

[0169] Duration: The average time (in seconds) for this atomic intent unit to appear and end during customer interaction. For example, "Inquire about shipping time" is usually a single text question, lasting an average of 5 seconds; "Check logistics status" includes clicks and page dwell, lasting an average of 15 seconds.

[0170] Trigger probability: The probability that this unit will be triggered in all customer interactions. The formula is: Trigger probability = Number of times the unit appears ÷ Total number of interactions. For example, if "Inquire about delivery time" appears 300 times in 1000 interactions, the trigger probability = 300 ÷ 1000 = 0.3.

[0171] The supplementary node feature matrix is ​​a concatenation of the initial feature vector and supplementary features: 64-dimensional semantic-sentiment vector + 1-dimensional duration + 1-dimensional trigger probability = 66-dimensional features. For example, the supplementary feature vector for "consultation and delivery time" is [0.8 (logistics semantics), 0.7 (time semantics), ..., 5 (duration), 0.3 (trigger probability)] (66 dimensions in total). The number of rows in the matrix equals the number of atomic intent units (e.g., 5 rows), and the number of columns is 66, fully characterizing the feature attributes of each unit.

[0172] Based on the supplementary node feature matrix, a graph attention network is used for feature extraction. The attention of a node to its neighbors is calculated through a multi-head attention mechanism. The neighbor features are aggregated to update the node's own features. After three layers of GAT iteration, the updated node features are output.

[0173] Graph Attention Networks (GATs) capture the strength of associations between nodes and their neighbors through an attention mechanism, making them suitable for extracting structured features from atomic intent graphs. Their core is "multi-head attention"—multiple independent attention heads compute attention in parallel, and then the results are concatenated to enhance feature representation.

[0174] Multi-head attention mechanism calculates attention: For each node, calculate its attention weight to all neighboring nodes, the formula is: attention α_ij=softmax(LeakyReLU(a^T[h_i||h_j])).

[0175] Where h_i and h_j are the feature vectors of node i and its neighbor j, || denotes concatenation, a is a learnable weight vector, LeakyReLU is the activation function, and softmax ensures that the sum of the attention of all neighbors is 1.

[0176] For example, the attention given to node A (inquiring about shipping time) to its neighbor C (inquiring about logistics status) is calculated as follows:

[0177] By concatenating the 66-dimensional feature vectors of A and C, a 132-dimensional vector is obtained.

[0178] By applying weight vector a and LeakyReLU activation, an intermediate value (such as 2.5) can be obtained;

[0179] Performing a softmax operation on the median values ​​of all neighbors (B, C) of A yields α_AC = 0.7 (C has higher attention) and α_AB = 0.3 (B has lower attention), because A and C co-occur more frequently.

[0180] Eight attention heads are used (balancing computational cost and feature richness), each head outputs 66-dimensional features, which are concatenated to obtain 8×66=528-dimensional features. These features are then compressed back to 66 dimensions through a linear layer to complete one attention update.

[0181] 3-layer GAT iterative update:

[0182] Layer 1: Focus on local relationships and capture the co-occurrence features of direct neighbors. For example, after aggregating features of A and C, the dimension value related to "logistics query" increases (e.g., from 0.8 to 0.85).

[0183] Layer 2: Extend to second-order neighbors to capture indirect relationships. For example, A aggregates the features of B (order reminder) through C, and the dimension value related to the emotion of "anxiety" is slightly increased (e.g., from 0.5 to 0.55).

[0184] Layer 3: Global association and integration, outputting the final node features. For example, A's features retain its core semantics of "consultation time" while also incorporating C's "query behavior" and B's "sentiment tendency" to form a more comprehensive 66-dimensional feature vector.

[0185] In the updated node feature matrix, the features of each atomic intent unit not only include its own attributes, but also incorporate key information from related units, which better reflects the semantic role of the intent primitive in the overall interaction.

[0186] The updated node features are globally pooled and combined with the temporal position features of the units, then mapped to a 32-dimensional vector through a fully connected layer, which serves as the candidate atomic intent feature vector.

[0187] The purpose of global pooling is to extract global statistical information from the node features output by GAT, thereby enhancing the robustness of the features. "Mean pooling" is employed—the mean of all nodes is calculated for each feature dimension. For example, in a 66-dimensional feature set, the mean of the first dimension (logistics semantics) is calculated as (0.85 of A + 0.7 of B + 0.9 of C + 0.1 of D + 0.2 of E) ÷ 5 = 0.55, which serves as the global logistics semantic strength. After pooling, a 66-dimensional global feature is obtained, which is then concatenated with the 66-dimensional features of each node to form a 132-dimensional fused feature set (local features + global features).

[0188] Temporal position features are used to characterize the temporal position of atomic intentions in the customer interaction sequence, and are divided into three categories: "early," "mid," and "late," using one-hot encoding: early = [1,0,0], mid = [0,1,0], late = [0,0,1]. For example, "inquiring about product specifications" often appears in the early stage of the interaction (when the customer has just come into contact with the product), with a temporal feature of [1,0,0]; "urging an order" often appears in the mid-stage (after the customer has waited for a period of time), with a temporal feature of [0,1,0]. The 3-dimensional temporal features are concatenated with the 132-dimensional fusion features to obtain 135-dimensional features.

[0189] The 135-dimensional features are mapped to a 32-dimensional vector, namely the candidate atomic intent feature vector, through a fully connected layer (64 hidden neurons, ReLU activation function). The mapping process learns a weight matrix to retain the most crucial information for intent recognition; for example, dimensions such as "logistics semantics," "anxiety emotion," and "query behavior" are strengthened, while irrelevant noise is suppressed. For instance, in the candidate vector for "inquiry about shipping time," the first 10 dimensions emphasize logistics-related semantics, the middle 10 dimensions reflect neutral emotion, and the last 12 dimensions relate to early temporal features. The overall 32-dimensional vector compactly expresses the core attributes of the intent primitive.

[0190] Calculate the cosine similarity between candidate atomic intention feature vectors. When the similarity is greater than 0.85, merge redundant vectors and normalize them to make the vector magnitude 1, thus generating the final atomic intention feature vector.

[0191] Cosine similarity measures the directional consistency between two vectors, and is calculated as: cosθ = (v1·v2) / (||v1||·||v2||). Here, v1 and v2 are candidate feature vectors, · represents the dot product, and ||·|| represents the magnitude. The closer the similarity is to 1, the more similar the vectors (the closer their intention primitives).

[0192] For example, the candidate vectors for "Consulting on delivery time" (v1) and "Consulting on logistics time" (v2) are: v1·v2 = the sum of the products of corresponding elements of the 32-dimensional vector = 28.5; ||v1|| = √(sum of squares of all elements in v1) = 5.8, ||v2|| = 5.9; cosθ = 28.5 ÷ (5.8 × 5.9) = 28.5 ÷ 34.2 ≈ 0.83 (< 0.85, not redundant).

[0193] Candidate vectors for “Query logistics status” (v3) and “View logistics information” (v4): v3·v4=30.2, ||v3||=5.7, ||v4||=5.8; cosθ=30.2÷(5.7×5.8)=30.2÷33.06≈0.91 (>0.85, deemed redundant).

[0194] Redundant vectors are merged using a "weighted average": the merged vector v = (v3×f3 + v4×f4) ÷ (f3 + f4), where f3 and f4 are the occurrence frequencies of the two units (e.g., f3 = 250, f4 = 150), ensuring that the intention primitives that occur more frequently have higher weights.

[0195] Normalization is achieved using the L2 norm: each element of the merged vector is divided by the vector's magnitude, resulting in a final vector magnitude of 1. For example, if the merged vector has a magnitude of 6.2, normalizing each element results in the original element divided by 6.2, yielding a 32-dimensional vector with a magnitude of 1.

[0196] The final atomic intent feature vector set contains all intent primitive vectors after redundancy removal. For example, "inquire about delivery time", "expedite order", "query logistics status (after merging)", "inquire about product specifications", and "compare prices" each correspond to a 32-dimensional vector. The vectors retain the distinctiveness of unique intents while facilitating subsequent model processing on the same scale, providing accurate feature representations for customer intent recognition.

[0197] S204, the atomic intent feature vector is input into a pre-trained hybrid expert intent recognition model to generate a customer intent recognition result, wherein the intent recognition model dynamically selects an expert network suitable for the current input through a gating mechanism;

[0198] Specifically, a hybrid expert intent recognition model can be constructed, which includes 5 expert networks and 1 gating network. The expert networks adopt an MLP structure, and the gating network takes atomic intent feature vectors as input and outputs expert selection weights, thus obtaining the model architecture.

[0199] The core of Hybrid Expert Model (MoE) is to process different types of intents by dividing the work among multiple "expert networks" and then using a "gating network" to dynamically select the most suitable expert, thereby improving the recognition accuracy of complex intents.

[0200] Expert Network Design: Five expert networks are designed to optimize customer intent for different scenarios. Each network employs a 3-layer MLP (Multilayer Perceptron) structure, with inputs consisting of 32-dimensional atomic intent feature vectors and outputs probability distributions for intent categories (e.g., 8 intent categories such as "logistics inquiry," "product inquiry," and "complaint"). The specific structure is as follows:

[0201] Input layer: 32 neurons (matching the dimension of the atomic intent feature vector);

[0202] Hidden layer 1: 64 neurons, activation function is ReLU (for handling non-linear features);

[0203] Hidden layer 2: 32 neurons, activation function is ReLU;

[0204] Output layer: 8 neurons (corresponding to 8 types of intent), activation function is Softmax (output probability distribution).

[0205] Each expert network focuses on different types of intent:

[0206] Expert 1: Optimize "logistics-related intents" (such as tracking logistics and expediting orders), with 70% of the training data being logistics-related samples;

[0207] Expert 2: Optimize "product inquiry intent" (such as specifications, price, material);

[0208] Expert 3: Optimize "after-sales intent" (such as returns, exchanges, and repairs);

[0209] Expert 4: Optimize the "complaint intent" (e.g., dissatisfaction with service, quality issues);

[0210] Expert 5: Optimize "generalization intent" (low-frequency, mixed-type intent).

[0211] Gated network design: The input is a 32-dimensional atomic intention feature vector, and the output is the selection weights of 5 experts (summing up to 1). The structure is a 2-layer MLP.

[0212] Input layer: 32 neurons;

[0213] Hidden layer: 16 neurons, activation function is Tanh;

[0214] Output layer: 5 neurons, activation function is Softmax (ensuring weights are non-negative and sum to 1).

[0215] The gating network assigns weights based on the semantic tendency of the input features. For example, when the values ​​of "logistics" and "time" are high in the input features, the weight of expert 1 will increase; when the values ​​of "product" and "specification" are high, the weight of expert 2 will increase. The model architecture is jointly trained through backpropagation, with the loss function being cross-entropy loss (which measures the difference between the predicted probability and the true intent), the learning rate set to 0.001, and 100 iterations until the loss converges.

[0216] Input the atomic intent feature vector into the gating network, output the weight of each expert, include experts with weights > 0.3 into the candidate set, and output the expert candidate set and weights;

[0217] After the atomic intent feature vector (32-dimensional) is input into the gating network, the weights of the five experts are output through Softmax, reflecting the degree of adaptation of each expert to the current intent.

[0218] Weight Calculation Example: Assume a certain atomic intent feature vector corresponds to "querying logistics status," with a "logistics semantics" dimension value of 0.8, a "time semantics" dimension value of 0.7, and a "behavior frequency" dimension value of 0.6 (consistent with logistics intent features). The gated network calculates the output weights as follows:

[0219] Expert 1 (Logistics): 0.45 (highest, due to its suitability for logistics); Expert 2 (Product): 0.1 (low, unrelated to the product); Expert 3 (After-sales): 0.25 (medium, logistics may be related to after-sales); Expert 4 (Complaints): 0.05 (low, no complaint characteristics); Expert 5 (Generalization): 0.15 (medium, for preventing uncovered issues).

[0220] Candidate set selection: Set a weight threshold of 0.3 (an empirical value to balance the number and accuracy of experts). Experts with a weight > 0.3 are included in the candidate set. In the example above, expert 1 (0.45 > 0.3) and expert 3 (0.25 < 0.3, not included) are both included. Is only expert 1 included in the candidate set? Correction: Expert 3's weight 0.25 < 0.3, so the candidate set only contains expert 1 with a weight of 0.45. Another example: If the feature vector corresponds to a mixed intent of "logistics + after-sales", and the gating output weights are expert 1 (0.35), expert 3 (0.32), and others < 0.3, then the candidate set is expert 1 (0.35) and expert 3 (0.32), with weights of 0.35 and 0.32 respectively.

[0221] The output expert candidate set must clearly specify the expert number and corresponding weight, such as "candidate set: expert 1 (0.45)" or "candidate set: expert 1 (0.35), expert 3 (0.32)", to provide a basis for subsequent reasoning.

[0222] The feature vectors are inferred using an expert network in the candidate set, and the intent category and confidence level are output. The results are then fused by a gated network according to the weights, and the fused recognition result is output.

[0223] The candidate expert network independently infers from the input atomic intent feature vectors, outputting their respective intent categories and confidence levels (probability values ​​of a certain type of intent), which are then fused by the gating network according to weights to enhance the influence of highly fit experts.

[0224] Example of expert reasoning:

[0225] Scenario 1: The candidate set consists only of expert 1 (logistics expert), and the input is the feature vector of "query logistics status".

[0226] After reasoning, Expert 1 outputs the intent category and confidence level: "Logistics Inquiry" (0.92), "Order Expediting" (0.05), Other (<0.03), with the highest confidence level being 0.92 ("Logistics Inquiry").

[0227] Scenario 2: The candidate set contains expert 1 (0.35) and expert 3 (0.32), and the input is a mixed feature vector of "logistics delay + application for return".

[0228] Expert 1's reasoning: "Logistics inquiry" (0.6), "Order expediting" (0.3), other (0.1);

[0229] Expert 3 reasoning: "Return application" (0.7), "Logistics inquiry" (0.2), Other (0.1).

[0230] Weighted fusion: The confidence scores for the same intent are weighted and summed according to expert weights. The formula is as follows:

[0231] Fusion confidence (intent i) = Σ(expert j weight × expert j's confidence in intent i)

[0232] In Scenario 2:

[0233] The confidence level of the "logistics inquiry" fusion is calculated as follows: 0.35 × 0.6 + 0.32 × 0.2 = 0.21 + 0.064 = 0.274.

[0234] The fusion confidence level for the "return request" is 0.35 × 0 + 0.32 × 0.7 = 0 + 0.224 = 0.224.

[0235] The fusion confidence level for "order reminder" is 0.35 × 0.3 + 0.32 × 0 = 0.105 + 0 = 0.105.

[0236] The highest fusion confidence level is 0.274 (“Logistics Inquiry”).

[0237] The fusion recognition result should include the fusion confidence of all intents and the intent category corresponding to the highest confidence. For example, scenario 1 outputs "intent: logistics query, confidence: 0.92"; scenario 2 outputs "intent: logistics query, confidence: 0.274".

[0238] When the highest confidence level of the fusion recognition result is greater than 0.8, the customer intent is directly output; otherwise, the final customer intent recognition result is determined by majority voting.

[0239] Set a confidence threshold of 0.8 (based on business needs, to ensure the reliability of high-confidence results), and handle two cases:

[0240] Case 1: Highest confidence level > 0.8

[0241] The intent is output directly because the experts' judgments on the intent are highly consistent. For example, in scenario 1, expert 1 outputs "logistics query" with a confidence level of 0.92 > 0.8, and directly outputs "customer intent: logistics query".

[0242] Case 2: Highest confidence level ≤ 0.8

[0243] The result is determined by majority voting (the number of experts supporting a certain intent among the candidate experts is the highest), thus resolving ambiguity under low confidence. For example, in scenario 2, candidate expert 1 supports "logistics query" and expert 3 supports "return application", with a vote ratio of 1:1 (a tie). At this time, expert 5 (a generalized expert, although its weight is 0.13 < 0.3 and it is not included in the candidate set, it participates in the voting) is introduced. Expert 5 infers that the "logistics + return mixed intent" tends to favor "return application", and the final vote ratio is 1 (logistics):2 (return), outputting "customer intent: return application".

[0244] Another example: Candidate expert 1 (weight 0.35) supports "product consultation", expert 2 (weight 0.31) supports "product consultation", and expert 3 (weight 0.29) supports "price comparison". The highest fusion confidence is 0.35×0.7+0.31×0.6=0.245+0.186=0.431≤0.8. The voting result is that "product consultation" gets 2 votes and "price comparison" gets 1 vote. The output is "Customer intent: product consultation".

[0245] The final customer intent identification result must be accompanied by a confidence level or voting basis, such as "Customer intent: Logistics inquiry (confidence level 0.92)" or "Customer intent: Return request (2:1 vote passed)" to ensure the interpretability of the result.

[0246] By employing a hybrid expert model with division of labor and dynamic selection, combined with weighted fusion and majority voting, this step can effectively handle single intents, mixed intents, and low-frequency intents, improving the accuracy and robustness of customer intent recognition and providing reliable input for subsequent intervention decisions.

[0247] S205, Based on the reinforcement learning policy optimizer, intervention decisions are made based on the customer intent recognition results, generating a Pareto-optimal intervention strategy set. Feasibility is then verified using a knowledge graph, and finally, the optimal intervention plan, including the timing, method, and content of the intervention, is output. Specifically, this may include:

[0248] S2051 uses the customer intent recognition result as the state input for reinforcement learning, defines the action space as the intervention method, the reward function as the increase in customer satisfaction after intervention, trains the strategy optimizer, and outputs the intervention strategy generation model.

[0249] The core of reinforcement learning is to enable an agent to learn the optimal decision-making strategy through a cyclical interaction of "state-action-reward". In customer intention intervention scenarios, three key elements need to be defined: state, action space, and reward function, and a model capable of generating effective intervention strategies should be trained using a policy optimizer.

[0250] State definition: Centered on the output customer intent recognition result, it includes intent category, customer sentiment tendency, and interaction history characteristics. For example:

[0251] Intent categories: "Logistics inquiry", "Product complaint", "Price inquiry", etc. (8 predefined intent categories);

[0252] Sentiment tendency: extracted from semantic-sentiment feature vectors, divided into "positive", "neutral", "negative" and "angry" (e.g., "complaint" intention is often accompanied by "angry" emotion);

[0253] Interaction history: Number of customer interactions in the past hour (e.g., 3 times), and whether intervention has been received (e.g., "No").

[0254] The state is represented in vector form. For example, "logistics query + neutral sentiment + 2 interactions + no intervention" can be encoded as [0,1,2,0] (0 represents logistics query, 1 represents neutral sentiment, 2 represents number of interactions, and 0 represents no intervention).

[0255] Action space definition: This refers to the possible intervention methods, encompassing multi-channel and multi-level intervention approaches, which must match the customer's intentions. For example:

[0256] Automated intervention: SMS push logistics links (A1), APP pop-up display of specifications (A2), intelligent voice broadcast of prices (A3);

[0257] Manual intervention: Transfer to online customer service (A4), arrange for a dedicated consultant to call back (A5), send a text message promising manual handling (A6);

[0258] Combined intervention: first, automatic information push + manual follow-up 10 minutes later (A7), etc.

[0259] The action space includes 10 intervention methods, each with a clear execution cost (e.g., A1 costs 0.1 yuan / time, A5 costs 5 yuan / time) and response time (e.g., A1 is immediate, A5 requires 5 minutes).

[0260] Reward function design: The core focus is on the increase in customer satisfaction after intervention, while also considering cost control. The calculation formula is: Reward = (Post-intervention satisfaction - Pre-intervention satisfaction) × 10 - Intervention cost × 0.5. Satisfaction is quantified through real-time customer feedback (e.g., clicking the "Satisfied" button = 5 points, "Dissatisfied" = 1 point) or semantic sentiment analysis (anger → 1 point, positive → 5 points). For example:

[0261] Customer satisfaction score was 3 before intervention and 4 after intervention. Using A1 (cost 0.1 yuan), the reward is (4-3)×10-0.1×0.5=10-0.05=9.95;

[0262] If the satisfaction score remains at 3 after intervention, the reward = 0 - 0.05 = -0.05 (punishment).

[0263] Policy optimizer training: The Deep Deterministic Policy Gradient (DDPG) algorithm is employed due to its continuous action space (allowing for fine-tuning of intervention intensity) and high state dimensionality. Training process:

[0264] Initialize the policy network (outputting action probabilities) and the value network (evaluating action values);

[0265] The agent performs actions in a simulated environment (built based on historical interaction data) to obtain rewards and new states;

[0266] The network is trained by storing "state-action-reward-new state" samples through an experience replay mechanism and randomly sampling them.

[0267] Iterate 10,000 times until the average reward of the action output by the policy network converges (fluctuation <1% for 100 consecutive rounds).

[0268] After training, the model that generates intervention strategies is output. This model receives the customer's intention state and outputs the execution probability of each intervention method (e.g., for the "logistics inquiry + neutral" state, the probability of A1 is 0.7 and the probability of A7 is 0.3).

[0269] S2052, using the intervention strategy generation model to output 10 sets of candidate intervention strategies, calculate the effect index of each set, and use Pareto optimization to screen non-dominated solutions to obtain the Pareto optimal intervention strategy set;

[0270] Candidate intervention strategies need to cover combinations and intensities of different intervention methods, evaluate their merits and demerits through effectiveness indicators, and then select "no disadvantage" strategies through Pareto optimization.

[0271] Generate 10 candidate strategies: The intervention strategy generation model outputs 10 feasible strategies based on the current customer intent state (e.g., "complaint + anger + 5 interactions"). Each strategy includes the intervention method and execution parameters (e.g., timing, content length). For example:

[0272] Strategy 1: A5 (dedicated consultant calls back), executed within 5 minutes, including apology and solution;

[0273] Strategy 2: A7 (Automatic push of apology SMS + customer service follow-up 10 minutes later);

[0274] Strategy 3: A4 (Immediately transfer to online customer service), priority "high"; ...

[0276] Strategy 10: A6 (manual processing of commitment SMS messages) + A3 (voice broadcast of processing progress).

[0277] Performance metrics calculation: Each strategy group needs to be evaluated based on 3 core metrics (balancing effectiveness, cost, and efficiency):

[0278] Satisfaction improvement value: Expected increase in satisfaction after intervention (based on historical data prediction, such as Strategy 1 is expected to improve by 2 points, and Strategy 2 is expected to improve by 1.5 points);

[0279] Execution costs include costs for manpower, communication, etc. (e.g., Strategy 1 costs 5 yuan, Strategy 2 costs 0.5 + 3 = 3.5 yuan).

[0280] Response time: The average time from strategy execution to customer feedback (e.g., 10 minutes for strategy 1, 15 minutes for strategy 2).

[0281] Pareto optimization for selecting nondominated solutions: Pareto optimality refers to a strategy where "no other strategy is better on all metrics." For example:

[0282] Strategy 1: Increase satisfaction by 2 points, cost 5 yuan, response time 10 minutes;

[0283] Strategy 2: Satisfaction score increased by 1.5 points, cost 3.5 yuan, response time 15 minutes;

[0284] Strategy 3: Increase satisfaction by 2 points, cost 6 yuan, response time 8 minutes;

[0285] Strategy 4: Satisfaction score increased by 1.8 points, cost 4 yuan, response time 12 minutes.

[0286] Analysis shows that: Strategy 3 has a higher cost than Strategy 1, but is better in other indicators; however, Strategy 1 is superior in cost, and neither is dominant. Strategy 2 is inferior to Strategy 1 in satisfaction and response time, and is a dominant solution (which can be eliminated). Strategy 4 has lower satisfaction than Strategy 1 and higher cost than Strategy 2, and is also a dominant solution (which can be eliminated). Ultimately, strategies 1, 3, and the other non-dominated solutions are retained, forming the Pareto optimal intervention strategy set (usually containing 3-5 strategies).

[0287] S2053: Input the optimal set of intervention strategies into the customer knowledge graph, verify the feasibility of the strategies, eliminate infeasible strategies, and output a subset of feasible strategies.

[0288] A customer knowledge graph is a semantic network that stores information such as customer attributes, preferences, and system resources. It is used to verify whether policies conform to actual constraints and avoid executing ineffective interventions.

[0289] Knowledge graph structure: includes entities (customers, intervention methods, resources) and relationships (customers - preferences, intervention methods - required resources, resources - available resources). For example:

[0290] Customer entity: C001, with attributes including "Reject voice calls", "Accept SMS only", and "VIP customer (high priority)";

[0291] Intervention method entity: A5 (dedicated consultant callback), relationship "required resources = consultant personnel" "applicable customers = non-refusal voice call customers";

[0292] Resource Entity: Consultants, with attributes "Current Availability = 3 people" and "Maximum Responses = 5 times / hour".

[0293] Feasibility verification rules:

[0294] Customer preference constraint: If the strategy includes intervention methods that result in customer rejection (such as C001 rejection voice, strategy 1 includes A5 callback), it is deemed infeasible;

[0295] Resource constraints: If the resources required by a strategy exceed the current available resources (e.g., strategy 3 requires 4 consultants, but only 3 are currently available), then it is not feasible.

[0296] Time constraint: If the strategy response time exceeds the customer's tolerance threshold (e.g., a 5-minute tolerance threshold for complaints and a 10-minute strategy response time), it is not feasible.

[0297] Example of verification process:

[0298] For strategy 1 (A5 call back) in the Pareto optimal set: query the knowledge graph, attribute C001 "reject voice call" → violates preference constraint → infeasible, remove;

[0299] Strategy 3 (A4 Online Customer Service): Query the knowledge graph, accept online interaction at C001, and the current available online customer service is ≥10 people → feasible;

[0300] Strategy 5 (A7 Composite Intervention): The required resources, "SMS channel + 1 customer service representative", are currently sufficient, and the customer accepts SMS messages → feasible.

[0301] After removing infeasible strategies, output a subset of feasible strategies (such as strategies 3, 5, and 7).

[0302] S2054. Sort the feasible strategy subsets by comprehensive score, and select the strategy with the highest score that includes the timing, method and content of intervention as the optimal intervention plan.

[0303] The overall score quantifies the merits of the strategy using weighted performance indicators, ultimately selecting the optimal solution that balances satisfaction, cost, and efficiency.

[0304] Overall score calculation: The three effectiveness indicators of feasible strategies (satisfaction improvement, cost, and response time) are weighted and summed. The weights are set according to business objectives (e.g., satisfaction weight 0.5, cost 0.2, response time 0.3): Overall score = satisfaction improvement × 0.5 + (1 / cost) × 0.2 × 10 + (1 / response time) × 0.3 × 10, where cost and response time are taken as reciprocals and scaled to make the indicator range uniformly 0-10.

[0305] For example, feasible strategies:

[0306] Strategy 3: Satisfaction improved by 1.8 points, cost 3 yuan, response time 5 minutes, score = 1.8×0.5+(1 / 3)×0.2×10+(1 / 5)×0.3×10=0.9+0.67+0.6=2.17;

[0307] Strategy 5: Satisfaction increased by 2 points, cost 4 yuan, response time 8 minutes, score = 2×0.5 + (1 / 4)×0.2×10 + (1 / 8)×0.3×10 = 1 + 0.5 + 0.375 = 1.875;

[0308] Strategy 7: Satisfaction increase by 1.5 points, cost 2 yuan, response time 6 minutes, score = 1.5×0.5+(1 / 2)×0.2×10+(1 / 6)×0.3×10=0.75+1+0.5=2.25.

[0309] Sorting and selection of the optimal solution: Sorting in descending order of comprehensive score, Strategy 7 (2.25) > Strategy 3 (2.17) > Strategy 5 (1.875), so Strategy 7 is selected as the optimal solution.

[0310] Optimal solution content:

[0311] Intervention timing: Execute an automatic SMS immediately (0 minutes), and initiate online customer service follow-up after 6 minutes;

[0312] Intervention method: A7 (automatic SMS + manual follow-up);

[0313] Intervention content: SMS: "Dear customer, your reported logistics delay issue has been expedited. The logistics status is expected to be updated within 1 hour, and a dedicated customer service representative will provide you with a detailed explanation in 6 minutes [link]"; Manual follow-up script: "Hello, regarding your logistics issue, we have contacted the warehouse to prioritize shipment. The current location is XX, and the estimated delivery time is XX. Do you have any other needs?"; The final solution is output to ensure accurate and efficient intervention and maximize customer satisfaction.

[0314] Another embodiment of the present invention provides a real-time customer intent recognition and intervention system for streaming data, see [link to relevant documentation]. Figure 5 The system may include:

[0315] The mapping module 501 is used to receive multimodal customer interaction stream data, construct a spatiotemporal behavior tensor, and map the tensor elements to semantic-sentiment feature vectors, wherein the semantic-sentiment feature vectors are associated with the temporal evolution dimension of customer behavior.

[0316] The decomposition module 502 is used to perform temporal logic analysis on the semantic-emotion feature vector, construct an intent state transition graph, and decompose the graph structure into several atomic intent units, wherein each atomic intent unit represents an indivisible customer intent primitive.

[0317] The extraction module 503 is used to extract features from the atomic intention unit using a graph neural network to generate an atomic intention feature vector, wherein the dimension of the atomic intention feature vector represents the semantic embedding space of the intention primitive.

[0318] The recognition module 504 is used to input the atomic intent feature vector into a pre-trained hybrid expert intent recognition model to generate a customer intent recognition result, wherein the intent recognition model dynamically selects an expert network suitable for the current input through a gating mechanism;

[0319] The intervention module 505 is used to make intervention decisions based on the customer intent recognition results using a reinforcement learning strategy optimizer, generate a Pareto optimal intervention strategy set, perform feasibility verification using a knowledge graph, and finally output the optimal intervention plan including the timing, method, and content of the intervention.

[0320] This invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0321] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0322] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for real-time identification and intervention of customer intent based on streaming data, characterized in that, The method includes: Receive multimodal customer interaction stream data, construct a spatiotemporal behavior tensor, and map the tensor elements to semantic-sentiment feature vectors, wherein the semantic-sentiment feature vectors are associated with the temporal evolution dimension of customer behavior; Temporal logic analysis is performed on the semantic-sentiment feature vectors to construct an intent state transition graph. The graph structure is decomposed into several atomic intent units, where each atomic intent unit represents an indivisible customer intent primitive. The semantic-sentiment feature vector set is sorted by time window, and temporal logic operators are used to label the temporal relationships between vectors, outputting a temporal relationship labeling table. Intent state nodes are constructed based on the temporal relationship labeling table, where each node contains the feature vector, sentiment tag, and core behavior within the window, with node IDs increasing over time, outputting a state node set. The transition probability between state nodes is calculated; when the transition probability is >20%, directed edges are established, with edge weights equal to the transition probability, outputting the intent state transition graph. Community detection is performed on the intent state transition graph to identify strongly connected components, each component corresponding to a composite intent, outputting a composite intent component set. The composite intent component set is decomposed into indivisible primitives using the minimum description length criterion, with each primitive containing a single behavior, sentiment, and semantic feature, generating atomic intent units. The atomic intention unit is used to extract features using a graph neural network to generate an atomic intention feature vector, wherein the dimension of the atomic intention feature vector represents the semantic embedding space of the intention primitive. The atomic intent feature vector is input into a pre-trained hybrid expert intent recognition model to generate a customer intent recognition result, wherein the intent recognition model dynamically selects an expert network suitable for the current input through a gating mechanism; The intervention decision is made based on the customer intent recognition results using a reinforcement learning strategy optimizer, generating a Pareto optimal intervention strategy set. The feasibility is verified by combining the knowledge graph, and finally the optimal intervention plan containing the timing, method and content of the intervention is output.

2. The method according to claim 1, characterized in that, The process involves receiving multimodal customer interaction stream data, constructing a spatiotemporal behavior tensor, and mapping tensor elements to semantic-sentiment feature vectors. These semantic-sentiment feature vectors are associated with the temporal evolution dimension of customer behavior, including: Receive multimodal customer interaction stream data, align different modal data by timestamp, filter out noisy data using rules, and output a spatiotemporally synchronized multimodal interaction dataset; Temporal features are extracted from a multimodal interaction dataset. Specifically, text word frequency, speech emotion intensity, and click frequency are counted in a 5-second sliding window. The time difference sequence of behaviors within the window is calculated, and the temporal feature matrix is ​​output. Using customer ID as the spatial dimension, time window as the temporal dimension, and temporal features as the attribute dimension, a three-dimensional spatiotemporal behavior tensor is constructed. The tensor element values ​​are the standardized strengths of the corresponding features, and the initial spatiotemporal behavior tensor is output. Feature extraction is performed on the initial spatiotemporal behavior tensor. Text features are extracted using BERT to extract semantic vectors, speech features are extracted using a sentiment classification model to extract sentiment vectors, and behavior features are reduced in dimensionality using PCA. These are then concatenated into a 64-dimensional semantic-sentiment feature vector. The vector dimension corresponds to the temporal evolution of the behavior, generating a semantic-sentiment feature vector set with temporal correlation.

3. The method according to claim 2, characterized in that, The graph neural network is used to extract features from the atomic intention units to generate atomic intention feature vectors, wherein the dimension of the atomic intention feature vectors represents the semantic embedding space of the intention primitives, including: Construct an association graph of atomic intention units, where nodes are atomic intention units, edges are co-occurrence relationships, and edge weights are co-occurrence frequencies, thus obtaining the atomic intention association graph; Initialize the node features of the atomic intent association graph, using the semantic-sentiment feature vector of each atomic intent unit as the initial node feature, and supplement the unit's duration and trigger probability to obtain the supplementary node feature matrix; Based on the supplementary node feature matrix, a graph attention network is used for feature extraction. The attention of a node to its neighbors is calculated through a multi-head attention mechanism. The neighbor features are aggregated to update the node's own features. After three layers of GAT iteration, the updated node features are output. The updated node features are globally pooled and combined with the temporal position features of the units, then mapped to a 32-dimensional vector through a fully connected layer, which serves as the candidate atomic intent feature vector. Calculate the cosine similarity between candidate atomic intention feature vectors. When the similarity is greater than 0.85, merge redundant vectors and normalize them to make the vector magnitude 1, thus generating the final atomic intention feature vector.

4. The method according to claim 3, characterized in that, The step of inputting the atomic intent feature vector into a pre-trained hybrid expert intent recognition model to generate a customer intent recognition result, wherein the intent recognition model dynamically selects an expert network suitable for the current input through a gating mechanism, including: A hybrid expert intent recognition model is constructed, which includes 5 expert networks and 1 gating network. The expert networks adopt an MLP structure, and the gating network takes atomic intent feature vectors as input and outputs expert selection weights, thus obtaining the model architecture. Input the atomic intent feature vector into the gating network, output the weight of each expert, include experts with weights > 0.3 into the candidate set, and output the expert candidate set and weights; The feature vectors are inferred using an expert network in the candidate set, and the intent category and confidence level are output. The results are then fused by a gated network according to the weights, and the fused recognition result is output. When the highest confidence level of the fusion recognition result is greater than 0.8, the customer intent is directly output; otherwise, the final customer intent recognition result is determined by majority voting.

5. The method according to claim 4, characterized in that, The reinforcement learning-based policy optimizer makes intervention decisions based on the customer intent recognition results, generates a Pareto-optimal intervention strategy set, performs feasibility verification using a knowledge graph, and finally outputs an optimal intervention plan that includes the timing, method, and content of the intervention, including: The customer intent recognition result is used as the state input for reinforcement learning. The action space is defined as the intervention method, the reward function is the increase in customer satisfaction after the intervention, the strategy optimizer is trained, and the intervention strategy generation model is output. The intervention strategy generation model outputs 10 candidate intervention strategies, calculates the effect index of each group, and uses Pareto optimization to screen non-dominated solutions to obtain the Pareto optimal intervention strategy set. Input the optimal set of intervention strategies into the customer knowledge graph, verify the feasibility of the strategies, eliminate infeasible strategies, and output a subset of feasible strategies. The feasible strategy subset is sorted by comprehensive score, and the strategy with the highest score that includes the timing, method and content of intervention is selected as the optimal intervention plan.

6. A real-time customer intent recognition and intervention system for streaming data, characterized in that, The system includes: The mapping module is used to receive multimodal customer interaction stream data, construct a spatiotemporal behavior tensor, and map the tensor elements to semantic-sentiment feature vectors, wherein the semantic-sentiment feature vectors are associated with the temporal evolution dimension of customer behavior. The decomposition module performs temporal logic analysis on the semantic-sentiment feature vectors, constructs an intent state transition graph, and decomposes the graph structure into several atomic intent units, where each atomic intent unit represents an indivisible customer intent primitive. Specifically, the semantic-sentiment feature vector set is sorted by time window, and temporal logic operators are used to label the temporal relationships between vectors, outputting a temporal relationship labeling table. Intent state nodes are constructed based on the temporal relationship labeling table, where each node contains a feature vector, sentiment tag, and core behavior within the window, with node IDs increasing over time, outputting a set of state nodes. The transition probability between state nodes is calculated; when the transition probability is greater than 20%, directed edges are established with edge weights equal to the transition probability, outputting the intent state transition graph. Community detection is performed on the intent state transition graph to identify strongly connected components; each component corresponds to a composite intent, outputting a set of composite intent components. The set of composite intent components is decomposed by splitting it into indivisible primitives using the minimum description length criterion; each primitive contains a single behavior, sentiment, and semantic feature, generating atomic intent units. The extraction module is used to extract features from the atomic intention units using a graph neural network to generate atomic intention feature vectors, wherein the dimension of the atomic intention feature vectors represents the semantic embedding space of the intention primitives. The recognition module is used to input the atomic intent feature vector into a pre-trained hybrid expert intent recognition model to generate a customer intent recognition result, wherein the intent recognition model dynamically selects an expert network suitable for the current input through a gating mechanism; The intervention module is used to make intervention decisions based on the customer intent recognition results using a reinforcement learning strategy optimizer, generate a Pareto optimal intervention strategy set, perform feasibility verification using a knowledge graph, and finally output the optimal intervention plan including the timing, method, and content of the intervention.

7. The system according to claim 6, characterized in that, The mapping module is specifically used for: Receive multimodal customer interaction stream data, align different modal data by timestamp, filter out noisy data using rules, and output a spatiotemporally synchronized multimodal interaction dataset; Temporal features are extracted from a multimodal interaction dataset. Specifically, text word frequency, speech emotion intensity, and click frequency are counted in a 5-second sliding window. The time difference sequence of behaviors within the window is calculated, and the temporal feature matrix is ​​output. Using customer ID as the spatial dimension, time window as the temporal dimension, and temporal features as the attribute dimension, a three-dimensional spatiotemporal behavior tensor is constructed. The tensor element values ​​are the standardized strengths of the corresponding features, and the initial spatiotemporal behavior tensor is output. Feature extraction is performed on the initial spatiotemporal behavior tensor. Text features are extracted using BERT to extract semantic vectors, speech features are extracted using a sentiment classification model to extract sentiment vectors, and behavior features are reduced in dimensionality using PCA. These are then concatenated into a 64-dimensional semantic-sentiment feature vector. The vector dimension corresponds to the temporal evolution of the behavior, generating a semantic-sentiment feature vector set with temporal correlation.

8. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images