A Customer Attitude Recognition and Operational Intervention Method Based on AI Semantic Analysis

Abstract

This invention proposes a customer attitude recognition and operational intervention method based on AI semantic analysis, belonging to the field of data processing technology. The method includes: acquiring real-time voice streams of customer interactions and multi-source heterogeneous business context data, and generating interaction feature vectors and business context data sets respectively; performing a feature-level affine transformation on the current interaction feature vector based on the business context vector of the previous moment to generate a modulated feature vector; using a gated recurrent unit to perform temporal modeling on the modulated feature vector to output the customer attitude vector of the current moment; driving an attention mechanism based on the difference between the attitude vectors of the current moment and the previous moment to dynamically calculate the attention weight of each data item in the business context data set, and generating an updated context vector; feeding the updated context vector back to the feature modulation step, and simultaneously comparing the current attitude vector with a preset threshold vector to trigger corresponding operational intervention instructions.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for customer attitude recognition and operational intervention based on AI semantic analysis. Background Technology

[0002] Customer attitude recognition, as a crucial supporting technology for intelligent customer service, is widely used in interaction quality control during enterprise digital transformation. Among related technologies, a multimodal processing system, from acoustic feature extraction to semantic embedding generation, is constructed through the collaborative operation of speech recognition, natural language processing, and deep learning models. Specifically, this technology covers key stages from speech stream preprocessing and text semantic modeling to feature fusion, including Mel-frequency cepstral coefficient extraction, semantic embedding generation from pre-trained models such as BERT, and core operations such as feature-level affine transformation. As the complexity of customer service scenarios increases, existing technologies are gradually introducing temporal modeling methods such as bidirectional gated recurrent units and attention mechanisms; however, their technical systems are still primarily based on linear feedforward architectures, lacking a systematic design for dynamic modulation of business contexts.

[0003] However, existing customer attitude recognition methods directly fuse business background information (such as historical orders and membership levels) with real-time interaction features using a static splicing approach, without establishing a dynamic context modulation mechanism. This may lead to the system failing to recognize shifts in customer focus during long interactions (such as a shift from logistics issues to product quality issues), or misjudging atypical expressions. In practice, existing technologies typically classify single-point semantic information using pre-trained models, but this process severs the intrinsic connection between attitude evolution and contextual understanding, leaving the system's grasp of the interaction process discrete. This one-way processing mode not only limits the accuracy of contextual information focusing but also causes operational intervention to rely on post-event analysis, failing to respond immediately to real-time attitude changes. Among these, the dynamic context modulation module ( (Generation) and Adaptive Context Refocusing Module ( The lack of coordination in updating has become a key technical bottleneck restricting the improvement of the system's intelligence level. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, the first objective of this invention is to propose a customer attitude identification and operational intervention method based on AI semantic analysis.

[0006] The second objective of this invention is to propose a customer attitude recognition and operational intervention device based on AI semantic analysis.

[0007] The third objective of this invention is to provide an electronic device.

[0008] The fourth objective of this invention is to provide a computer-readable storage medium.

[0009] The fifth objective of this invention is to provide a computer program product.

[0010] To achieve the above objectives, a first aspect of the present invention proposes a method for customer attitude identification and operational intervention based on AI semantic analysis, comprising: S1, acquire real-time voice streams of customer interactions and multi-source heterogeneous business context data, and generate interaction feature vectors and business context data sets respectively; S2, based on the business context vector of the previous moment, performs a feature-level affine transformation on the current interaction feature vector to generate a modulated feature vector; S3 uses a gated loop unit to perform time-series modeling on the modulated feature vector and outputs the customer attitude vector at the current moment. S4. Based on the difference between the attitude vectors at the current time and the previous time, the attention mechanism is driven to dynamically calculate the attention weight of each data item in the business context data set and generate an updated context vector by weighting. S5 feeds the updated context vector back to the feature modulation step, and compares the current attitude vector with the preset threshold vector to trigger the corresponding operational intervention command.

[0011] In one embodiment of the present invention, S1 includes: S11 extracts acoustic features such as Mel-frequency cepstral coefficients, pitch, and energy from the real-time speech stream, generating an acoustic feature vector. ; S12, after converting the speech stream into text, inputs it into a pre-trained AI semantic analysis model to generate text semantic embedding vectors. And through vector concatenation operations and Combined into interactive feature vectors .

[0012] In one embodiment of the present invention, S2 includes: S21, the scaling factor vector is calculated through two parallel fully connected neural network layers. and offset factor vector ; S22, using the formula Interactive feature vectors Perform a feature-level affine transformation to generate the modulated feature vector. .

[0013] In one embodiment of the present invention, S3 includes: S31, through the update and reset gate mechanisms of the GRU unit, the current modulated feature vector is... Hidden state from the previous moment Merge and calculate the current hidden state. ; S32, hide state Input a fully connected layer with either a Softmax or Sigmoid activation function, decode to generate a customer attitude vector containing N quantized intensity values. .

[0014] In one embodiment of the present invention, S4 includes: S41, Calculate the current attitude vector attitude vector from the previous moment The difference As the query input for the attention mechanism; S42, applying attention weights to the business context data set Each data point is weighted and summed to generate a context vector that focuses on the relevant background information.

[0015] In one embodiment of the present invention, it further includes: S6 deploys the system as a streaming service, using a Kafka+Flink architecture to process voice stream segments in real time, and uses the MLOps system to achieve continuous integration, continuous deployment and performance monitoring of models.

[0016] To achieve the above objectives, a second aspect of the present invention provides a customer attitude recognition and operational intervention device based on AI semantic analysis, comprising: The data acquisition module is used to acquire real-time voice streams of customer interactions and multi-source heterogeneous business context data, and generate interaction feature vectors and business context data sets respectively. The feature modulation module is used to perform a feature-level affine transformation on the current interaction feature vector based on the business context vector of the previous time step, and generate a modulated feature vector. The temporal modeling module is used to perform temporal modeling on the modulated feature vector using a gated loop unit and output the customer attitude vector at the current moment. The context update module is used to drive the attention mechanism based on the difference between the attitude vector at the current time and the previous time, dynamically calculate the attention weight of each data item in the business context data set, and generate the updated context vector by weighting. The feedback intervention module is used to feed the updated context vector back to the feature modulation step, and at the same time compare the current attitude vector with the preset threshold vector to trigger the corresponding operational intervention instruction.

[0017] To achieve the above objectives, a third aspect of the present invention provides an electronic device, comprising: a processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of the first aspects.

[0018] To achieve the above objectives, a fourth aspect of the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, are used to implement the method as described in any one of the first aspects.

[0019] To achieve the above objectives, a fifth aspect of the present invention provides a computer program product that, when executed by a processor, implements the method described in any one of the first aspects.

[0020] The technical solutions provided by the embodiments of the present invention bring at least the following beneficial effects: This invention enables dynamic and continuous tracking of customer attitudes and adaptive focusing on business scenarios, and constructs a closed-loop feedback mechanism for identification and intervention, significantly improving the accuracy and intelligence of real-time operational response.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a customer attitude identification and operational intervention method based on AI semantic analysis, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a customer attitude recognition and operational intervention device based on AI semantic analysis, provided in an embodiment of the present invention. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0024] Figure 1This is a flowchart illustrating a customer attitude identification and operational intervention method based on AI semantic analysis, provided as an embodiment of the present invention. Figure 1 As shown, the method includes the following steps: S1 acquires real-time voice streams of customer interactions and multi-source heterogeneous business context data, and generates interaction feature vectors and business context data sets respectively.

[0025] Specifically, in the implementation of this invention, the step of "acquiring real-time voice streams of customer interactions and multi-source heterogeneous business context data, and generating interaction feature vectors and business context data sets respectively" is the basic data processing link for the entire system operation. Its technical implementation involves multiple technical aspects such as voice signal processing, text semantic embedding, multi-source data encoding and structured integration.

[0026] At the technical implementation level, this step first acquires the customer's real-time voice interaction data through a voice stream acquisition interface. The voice signal is processed in time steps... The signal is fed into an acoustic feature extractor, which extracts key acoustic features including Mel-frequency cepstral coefficients, pitch, and energy, generating an acoustic feature vector. Meanwhile, the speech signal is converted into text by an automatic speech recognition module, and the text is then input into a pre-trained semantic model (such as BERT) to generate semantic embedding vectors. These two vectors are combined into a comprehensive interactive feature vector for the current time step through vector concatenation or weighted fusion. Its dimensions are usually ,in Dimensions representing acoustic features This indicates the dimension of semantic embedding.

[0027] In terms of business context data processing, the system receives heterogeneous data from multiple sources, including CRM, order systems, and membership systems. This data includes static customer profiles (such as membership level and region), historical behavior (such as the time of the most recent service request and order amount), and textual data (such as historical work order summaries). For different data types, the system employs corresponding encoding strategies: categorical data uses one-hot encoding, numerical data undergoes normalization (such as Min-Max or Z-score), and textual data is converted into vector representations using a text encoder (such as BERT or TF-IDF). Ultimately, all business context features are organized into a unified business context data set. Its structure is usually a list or tensor containing multiple feature vectors, each vector corresponding to a specific business dimension.

[0028] At the parameter level, the sampling rate of an acoustic feature extractor is typically set to... Frame length is Frame shift is This is to ensure the temporal continuity of the speech signal. The output dimension of a semantic embedding model is generally [missing information]. or Specifically, it depends on the chosen pre-trained model. (Business context dataset) The dimensions are dynamically configured based on actual business needs, typically in... to The dimensions are designed to balance information richness and computational efficiency.

[0029] At the application level, this step is widely applicable to real-time interactive scenarios such as customer service, marketing operations, and complaint alerts. For example, in customer service hotlines, the system can analyze customer voice emotions in real time and, combined with their historical service records, determine whether they are currently in a high-risk state, thereby triggering a transfer to a live agent or sending reassurance strategies. In e-commerce customer service, the system can dynamically adjust its interpretation of customer attitudes by combining business data such as customer purchase records and return history, improving the accuracy of service responses.

[0030] Its purpose is to provide high-quality, computable input for subsequent dynamic context modulation, attitude evolution modeling, and closed-loop intervention by structuring real-time voice interaction and multi-source business context data separately. Its innovation lies in the dynamic modeling of business contexts and the fusion processing of multimodal features, laying a solid foundation for accurate identification of customer attitudes and adaptive operational intervention.

[0031] Furthermore, S1 includes: S11 extracts acoustic features such as Mel-frequency cepstral coefficients, pitch, and energy from the real-time speech stream, generating an acoustic feature vector. .

[0032] Specifically, in reality, acoustic features such as Mel-frequency cepstral coefficients (MFCC), pitch, and energy are extracted from real-time speech streams to generate acoustic feature vectors. This is one of the core operations of the system of this invention. The purpose of this step is to extract low-dimensional temporal features with discriminative capabilities from the original speech signal, providing structured input for subsequent semantic analysis and attitude recognition.

[0033] At the technical implementation level, the process first preprocesses the real-time speech stream, including pre-emphasis, framing, and windowing. The speech signal is framed in fixed time windows (e.g., 25ms), with frame shifts typically set to 10ms to ensure temporal continuity. Next, a Fast Fourier Transform (FFT) is performed on each frame to convert the time-domain signal into a frequency-domain representation. In the frequency domain, a Mel filter bank is used to nonlinearly map the spectrum, simulating the human ear's perception of different frequencies. Typically, a Mel filter bank contains 20-40 filters, covering the speech frequency band from 20Hz to 4kHz. The logarithmic energy of each filter output is taken, followed by a Discrete Cosine Transform (DCT) to extract 12-13 dimensional MFCC features, removing spectral correlations and enhancing the robustness of the speech.

[0034] In addition, pitch and energy features are extracted simultaneously. Pitch features are typically calculated using an autocorrelation function or the YIN algorithm. At a sampling frequency of 16kHz, the pitch range is usually set between 85Hz and 255Hz to cover the fundamental frequency range of adult speech. Energy features are obtained by summing the squared amplitudes of each frame of the signal and then taking the logarithm to reflect the intensity changes of the speech.

[0035] At the parameter level, MFCC feature extraction typically involves 12-13 coefficients (including energy coefficients), with a time step of 10ms, a frame length of 25ms, and a Hamming window function. The sampling rate for pitch features is consistent with the speech stream, typically 16kHz, and the fundamental frequency detection accuracy is required to be within ±5Hz. The calculation of energy features needs to consider logarithmic compression to avoid an excessively large dynamic range that could affect model stability.

[0036] In application scenarios, this step provides crucial acoustic input for subsequent semantic analysis and attitude recognition. By extracting features such as MFCC and pitch, the system can capture the prosody and timbre variations of speech, thereby assisting in the identification of customer emotional fluctuations, such as anger, anxiety, or satisfaction. This acoustic feature vector... With text semantic embedding vectors After concatenation, a comprehensive interactive feature vector is formed. As input to the dynamic context modulation module, it enables continuous modeling and real-time intervention of customer attitudes.

[0037] S12, after converting the speech stream into text, inputs it into a pre-trained AI semantic analysis model to generate text semantic embedding vectors. And through vector concatenation operations and Combined into interactive feature vectors .

[0038] Specifically, in the customer attitude recognition and operational intervention system of the present invention, the speech stream is converted into text and then input into a pre-trained AI semantic analysis model to generate a text semantic embedding vector. And through vector concatenation operations and Combined into interactive feature vectors This step is a key link in achieving semantic understanding of customer interaction and fusion of multimodal features.

[0039] At the technical implementation level, the speech stream is first transcribed in real time by an Automatic Speech Recognition (ASR) module to generate a corresponding text sequence. This text sequence is then input into a pre-trained AI semantic analysis model, such as BERT, RoBERTa, or their functional equivalents. These models are based on the Transformer architecture and possess powerful context-aware capabilities, enabling them to map the input text into fixed-dimensional semantic embedding vectors. Its dimensions are typically 768 or 1024, depending on the model version used. Semantic embedding vectors It captures the deep semantic information of the text, including emotional inclination, intention, and contextual relevance.

[0040] Meanwhile, the acoustic features in the speech stream are extracted by the acoustic feature extractor at time step. Generate acoustic feature vectors Its dimensions are typically 128 to 256, depending on the configuration of the feature extraction algorithm. This vector reflects the non-semantic properties of speech, such as emotion-related speech rhythm and pitch variations.

[0041] In the feature fusion stage, the system uses vector concatenation operations to... and Combined into interactive feature vectors Specifically, if and Then the spliced This operation maintains feature continuity over time, providing a unified input representation for subsequent modules (such as the dynamic context modulation module and the attitude evolution tracking module).

[0042] This step has a dual role in the system: on the one hand, it uses semantic embedding vectors... Improve the accuracy of recognizing customer intentions and emotions; on the other hand, by combining acoustic features The splicing of features enables complementary fusion of multimodal characteristics, enhancing the model's robustness to complex interactive scenarios. This fusion mechanism provides a high-quality input foundation for subsequent dynamic context modulation and attitude evolution modeling, and is a prerequisite for the system to achieve closed-loop feedback and real-time intervention.

[0043] S2, based on the business context vector of the previous moment, performs a feature-level affine transformation on the current interaction feature vector to generate a modulated feature vector.

[0044] Specifically, in some implementations, performing a feature-level affine transformation on the current interaction feature vector based on the business context vector from the previous time step is the core processing step of the dynamic context modulation module in this invention. This step aims to enhance the semantic understanding of the current interaction features by introducing the dynamic influence of the business context, thereby improving the accuracy and context relevance of customer attitude recognition.

[0045] From a technical implementation perspective, this step first receives two inputs: the interaction feature vector at the current time step t. and the previous time step Business Context Vector .in, It is a high-dimensional vector generated by the data processing unit 101 after fusing acoustic features and semantic embedding, typically with dimensions of 100. ,and It is the context vector output by the adaptive context refocusing module at the previous time step, and its dimension is... Consistent, denoted as The dynamic context modulation module contains two parallel fully connected neural network layers, one for generating the scaling factor vector. and offset factor vector The dimensions of these two vectors are... Same, used for Perform element-wise affine transformations.

[0046] Specifically, the mathematical expression for affine transformation is:

[0047] in, It represents the Hadamah accumulation. and Depend on The transformation is calculated through a fully connected layer. Essentially, it injects business context information into the current interaction features in a non-linear manner, enabling the system to dynamically adjust the expression strength and offset of the current features based on historical context, thereby achieving a more accurate model of customer attitudes.

[0048] In terms of parameter settings, fully connected layers typically use ReLU or Tanh activation functions to enhance nonlinear expressiveness. Scaling factor The output range is typically limited to [0.5, 1.5] to avoid excessive perturbation of the original features. Offset factor The output range is controlled between [-0.2, 0.2], which is used to fine-tune the baseline expression of features.

[0049] This step is particularly suitable for multi-round customer service interaction scenarios in practical applications. For example, when a customer moves from inquiring about product features to complaining about logistics issues, the system can... The dynamic adjustment makes the current interaction features more focused on business background information related to logistics, thereby improving the sensitivity to changes in customer attitudes.

[0050] In terms of technical effectiveness, this step achieves dynamic coupling between business context and interaction features, overcoming the limitations of statically splicing contextual information in traditional methods. Through element-wise affine transformation, the system can perform refined semantic modulation of interaction features, significantly improving the attitude recognition model's adaptability to contextual changes and providing higher-quality input features for subsequent adaptive context refocusing and operational intervention.

[0051] Furthermore, S2 includes: S21, the scaling factor vector is calculated through two parallel fully connected neural network layers. and offset factor vector .

[0052] Specifically, in the dynamic context modulation module of this invention, the scaling factor vector is calculated through two parallel fully connected neural network layers. and offset factor vector It is to realize interactive feature vectors A key step in dynamic modulation. This step is technically implemented based on the affine transformation mechanism of deep neural networks, aiming to integrate customer business context information into the current interaction feature processing flow in a non-linear and adaptive manner, thereby improving the contextual sensitivity and dynamic response capability of attitude recognition.

[0053] In some implementations, these two fully connected layers each receive the updated context vector from the previous time step. As input, the scaling factor vector. The output of the first fully connected layer has the same dimension as the current interaction feature vector. Consistency, used for Element-wise scaling is performed on each feature dimension; offset factor vector The output from the second fully connected layer is also the same as... Dimension alignment, used for Perform element-wise offset adjustments. These two vectors are adjusted using the following formula: Perform an affine transformation to generate the modulated feature vector. :

[0054] in, This represents the Hadamard product, which is an element-wise multiplication operation. The essence of this transformation is to inject business context information into the processing of interaction features in a parameterized manner, enabling the system to dynamically adjust its sensitivity and offset to current interaction features based on the customer's historical background, thereby achieving a more accurate model of customer attitudes.

[0055] In terms of parameter settings, fully connected layers typically use ReLU or Sigmoid activation functions to ensure the non-linear expressiveness of the output. Scaling factor vector The output range is typically limited to the [0, 1] interval by the Sigmoid function to avoid excessive perturbation of the original features; the offset factor vector They are usually not restricted and are allowed to vary freely in the real number field to provide more flexible offset adjustment capabilities.

[0056] In practical applications, particularly in customer service and marketing automation scenarios, this step significantly improves the system's responsiveness and depth of understanding of changes in customer emotions. Through dynamic modulation, the system can more effectively identify emotional turning points during customer interactions and provide high-quality input features for subsequent attitude evolution modeling and operational intervention. Its technological value lies in achieving a deep integration of business context and interaction semantics, providing crucial support for building an adaptive intelligent customer service system.

[0057] S22, using the formula Interactive feature vectors Perform a feature-level affine transformation to generate the modulated feature vector. .

[0058] Specifically, this step involves processing the interaction feature vector. Perform a feature-level affine transformation to generate the modulated feature vector. Its technical implementation is based on the formula:

[0059] in, It represents the Hadamah accumulation. For scaling factor vectors, Both are offset factor vectors, generated by two parallel fully connected neural network layers in the dynamic context modulation module based on the updated context vector from the previous time step. This formula is generated at each time step. The above pairs of interactive feature vectors Element-wise scaling and offset operations are performed to achieve dynamic modulation of features.

[0060] At the technical implementation level, this affine transformation is a key mechanism for deeply integrating business context information with real-time interaction features. Unlike traditional feature concatenation methods, this solution uses nonlinear transformation to... Each dimension is adjusted independently, allowing the feature representation to evolve dynamically according to the current context. Specifically, and Dimensions and Consistent, usually VI (e.g.) (corresponding to the BERT embedding dimension), its output is determined by the weight matrix of the fully connected layer. and The activation function can be either ReLU or Tanh to enhance nonlinear expressive power.

[0061] Regarding parameter settings, the input to the fully connected layer is the context vector from the previous time step. Its dimensions are usually or Specifically, it depends on how the business context data is encoded. The parameter size of the output layer and Directly related, for example in In this case, the number of parameters for each fully connected layer is approximately The overall modulation process has good scalability and computational efficiency.

[0062] This step plays a crucial role in the system, bridging the gap between preceding and subsequent steps, and its output... This will serve as input to the attitude evolution tracking module, driving the GRU unit to perform temporal modeling of customer attitudes. Through this affine transformation, the system can achieve dynamic contextual awareness of interaction features, thereby improving the accuracy and contextual relevance of attitude recognition and providing high-quality feature input for subsequent attention mechanisms and operational interventions.

[0063] S3 uses a gated loop unit to perform time-series modeling on the modulated feature vector and outputs the customer attitude vector at the current moment.

[0064] Specifically, in an embodiment of the present invention, the attitude evolution tracking module uses a gated recurrent unit (GRU) to perform temporal modeling on the modulated feature vector to output the customer attitude vector at the current moment. GRU, as a lightweight variant of recurrent neural network (RNN), has fewer parameters and higher computational efficiency compared to the traditional LSTM structure, making it suitable for dynamic modeling needs in real-time voice interaction scenarios.

[0065] In its implementation, the GRU unit operates at time steps. It receives two inputs: one is the modulated feature vector from the dynamic context modulation module. Secondly, the module itself was in the previous time step Hidden state GRU updates through its internal update gate. and reset door This mechanism enables selective fusion of the current input and historical states. The update gate controls the degree to which historical information is retained, while the reset gate determines the weight of the current input's influence on the hidden state update. The calculation formula is as follows:

[0066]

[0067]

[0068]

[0069] in, This represents the Sigmoid activation function. It is the hyperbolic tangent function. , , These are the weight matrices for the update gate, reset gate, and candidate state, respectively. This represents a vector concatenation operation. This is element-wise multiplication.

[0070] Obtain the hidden state The state is then fed into a fully connected output layer, typically configured as a neural network layer with a Softmax or Sigmoid activation function, to decode and generate the N-dimensional attitude vector for the current time step. .in, The number of categories for customer attitudes, such as satisfaction, anger, neutrality, etc., for each dimension. This indicates the customer's current attitude towards the first... The quantitative strength value of a class attitude typically ranges from [0, 1], representing probability or confidence level.

[0071] This step plays a crucial role in the entire system, and its output It is not only used to determine customers' emotional state in real time, but also serves as input to the adaptive context refocusing module, driving the attention mechanism to dynamically focus on business context data. Furthermore, It is also fed into a closed-loop operational intervention engine to trigger corresponding operational actions. Through GRU's temporal modeling capabilities, the system can capture the evolution trend of customer attitudes, achieve continuous perception and adaptive response to the interaction process, thereby significantly improving customer experience and operational efficiency.

[0072] Furthermore, S3 includes: S31, through the update and reset gate mechanisms of the GRU unit, the current modulated feature vector is... Hidden state from the previous moment Merge and calculate the current hidden state. .

[0073] Specifically, in the attitude evolution tracking module of this invention, the GRU unit, through its unique update gate and reset gate mechanism, realizes the updating of the current modulated feature vector. Hidden state from the previous moment The fusion of these elements is used to calculate the hidden state at the current time step. This step is the core component of the system's continuous modeling and evolution tracking of customer attitudes. Its technical implementation is based on a nonlinear state update mechanism using gated loop units, which can effectively capture the temporal dependencies of customer attitudes during the interaction process.

[0074] In the specific implementation, the update gate of the GRU unit and reset door All are vectors calculated using a fully connected layer and an activation function (usually the sigmoid function), and their dimensions and hidden states are... Consistent. Update gate Determine the current input information With historical status The fusion ratio, and the reset door Controlling historical states in calculating the current candidate state The degree of participation at any given time. The calculation formula is as follows:

[0075]

[0076]

[0077]

[0078] in, This represents the Sigmoid activation function. This represents the hyperbolic tangent activation function. , , This is the corresponding weight matrix. , , For bias vectors, This represents a vector concatenation operation. This indicates element-wise multiplication.

[0079] In practical applications, this GRU unit is typically deployed within a deep learning framework to support frame-by-frame processing of real-time speech streams. It incorporates input features... The dimension is usually set to Hidden state The dimension can be set as The general range of values ​​is The specific range depends on the complexity of the model and the limitations of computing resources. The training process of GRU units uses the standard backpropagation algorithm, and the loss function can be either cross-entropy loss or mean squared error (MSE) to adapt to the needs of different attitude classification tasks.

[0080] This step plays a crucial role in the overall system, acting as a bridge between preceding and subsequent steps. On one hand, it receives features modulated by dynamic context. This ensures that the current input information incorporates contextual information relevant to the customer's business background; on the other hand, it outputs a hidden state. Used to generate the attitude vector at the current moment. This information serves as input to the GRU unit in the next moment, enabling continuous state evolution. Through the GRU gating mechanism, the system can adaptively retain or forget historical information, thereby maintaining sensitivity and stability to changes in customer attitudes during long interactions and providing high-quality temporal state representations for subsequent attention refocusing and operational intervention.

[0081] S32, hide state Input a fully connected layer with either a Softmax or Sigmoid activation function, decode to generate a customer attitude vector containing N quantized intensity values. .

[0082] Specifically, in the attitude evolution tracking module, the hidden states are... The input is a fully connected layer with a Softmax or Sigmoid activation function, which is used to implement customer attitude vectors. The key step in decoding. This step is based on the sequence modeling mechanism in deep learning, aiming to decode the hidden states generated by the GRU unit at time step t. It is mapped to the quantitative intensity value of the customer's attitude towards N preset attitude categories at the current moment.

[0083] At the technical implementation level, this fully connected layer typically consists of a linear transformation layer, whose weight matrix... With bias vector Together they act in the hidden state ,in Let be the hidden state dimension of the GRU. The output of the linear transformation is a real vector of length N, representing the customer's original predicted score for each attitude at the current time step. This vector is then non-linearly activated using either a Softmax or Sigmoid function to generate the final customer attitude vector. .

[0084] If the Softmax activation function is used, the output satisfies the probability distribution characteristics, making it suitable for multi-class classification tasks, where a customer may exhibit one of several attitudes at a given time. Its calculation method is as follows:

[0085] in This represents the i-th output value after linear transformation. If the Sigmoid activation function is used, each attitude dimension independently outputs an intensity value between 0 and 1, suitable for multi-label classification tasks, where customers may simultaneously exhibit multiple attitudes. Its calculation method is as follows:

[0086] Regarding parameter settings, the weight matrix of the fully connected layer Training is typically performed using the backpropagation algorithm, with loss functions such as cross-entropy loss or mean squared error (MSE) used, depending on the classification nature of the task. The output dimension N can be set according to actual business needs, for example, including attitude categories such as "satisfied," "dissatisfied," "angry," and "neutral."

[0087] This step plays a central role in the entire system, and its output... It is not only used to determine customers' emotional state in real time, but also serves as input to the adaptive context refocusing module, driving the system to dynamically focus on business context information. By decoding hidden states into interpretable attitude intensity vectors, the system achieves a closed loop from semantic modeling to business decision-making, providing a precise basis for subsequent operational interventions.

[0088] S4: Driven by the difference between the attitude vectors at the current time and the previous time, the attention mechanism is dynamically calculated to determine the attention weight of each data item in the business context data set, and the updated context vector is generated by weighting.

[0089] Specifically, the step of “dynamically calculating the attention weight of each data item in the business context data set based on the difference between the attitude vector at the current time and the previous time, and generating an updated context vector by weighting” is the core link in realizing “adaptive context refocusing” in this invention. Its technical implementation principle is based on the attention mechanism driven by temporal difference, which aims to realize the dynamic coupling of customer attitude evolution and business context information.

[0090] At the technical implementation level, this step first uses the current attitude vector output by the attitude evolution tracking module. attitude vector from the previous moment Element-wise difference operations are performed to obtain the attitude evolution signal. This signal reflects the changing trend of customer attitudes between adjacent time steps and serves as the external driving signal for the attention mechanism. Subsequently, this signal is input as a query into the attention mechanism of the adaptive context refocusing module, while the business context data set... Each data item serves as both a key and a value in the calculation of attention weights. Attention mechanisms typically employ dot-product attention or learnable linear transformations to calculate the correlation between each piece of business contextual data and the current attitude change, thereby generating an attention weight vector. .

[0091] At the level of parameter indicators, attitude vector Dimensions Typically set to 5 or 7, corresponding to basic attitude categories such as satisfaction, dissatisfaction, anger, neutrality, and expectation, respectively. Attention weight. The calculation must satisfy the normalization condition, that is... This is to ensure the reasonableness of the weighted summation. Business context data set. Each piece of data needs to undergo standardized encoding processing. For example, categorical data uses One-Hot encoding, numerical data uses Min-Max normalization, and text data is embedded and encoded using a pre-trained model (such as BERT) to unify the vector space dimension.

[0092] At the application level, this step is suitable for scenarios that require real-time awareness of customer status and precise intervention, such as customer service, marketing operations, and complaint alerts. For example, in a voice interaction between a customer and customer service, if the customer's attitude changes from "satisfied" to "dissatisfied," the system will automatically increase the weight given to business context data such as the customer's historical complaint records and service response time, thereby more accurately judging the potential risks of the current interaction and triggering corresponding operational actions (such as transferring to a live agent or sending reassuring messages).

[0093] The technical advantage of this step lies in its ability to dynamically correlate attitude change signals with business context data, enabling the system to adaptively focus on customer background information and significantly improve the accuracy and real-time performance of contextual understanding. Simultaneously, this mechanism constructs a feedback loop from attitude recognition to contextual modulation, enhancing the system's adaptability and business response efficiency, and serving as a key technical support for achieving intelligent operational intervention.

[0094] Furthermore, S4 includes: S41, Calculate the current attitude vector attitude vector from the previous moment The difference As the query input for the attention mechanism.

[0095] Specifically, in the adaptive context refocusing module of the present invention, the current attitude vector is calculated. attitude vector from the previous moment The difference This is one of the key input steps in the attention mechanism. The technical implementation of this step is based on the idea of ​​time series modeling, which aims to capture the dynamic evolution trend of customer attitudes during the interaction process, thereby providing the attention mechanism with time-sensitive query signals.

[0096] In practical implementation, attitude vector The attitude evolution tracking module tracks the time steps. The modulated feature vector is processed through a GRU network. The output after encoding has the following dimensions. This indicates that the customer's current attitude towards... The quantitative intensity of a pre-defined attitude category. For example, It can be an emotion category such as "satisfied," "dissatisfied," "angry," or "neutral." At each time step, the system records the current attitude vector. attitude vector from the previous moment And calculate the difference. This difference vector serves as a quantitative representation of attitude change. Each element Indicates that the customer is at the The magnitude of change across attitude dimensions reflects the shifting trend of emotions or intentions.

[0097] This step serves as a query input in the attention mechanism, driving the processing of the business context data set. Dynamic focusing. By... As the basis for calculating attention weights, the system can identify which business background information (such as historical orders, service records, customer profiles, etc.) is most relevant to the current attitude change, thereby achieving adaptive modulation of contextual information. For example, if If customers show a stronger attitude towards "complaints," then the attention mechanism will focus more on... Features related to historical complaint records or service response times.

[0098] From a technical perspective, the calculation frequency of this step depends on the system's time window setting, typically updating once per second or per dialogue sentence. Difference vector The numerical range is typically between [-1, 1], depending on the specific value. The normalization method (such as using the Sigmoid or Softmax activation function) is crucial. This step relies on high-precision temporal modeling capabilities and a dynamic perception mechanism for multi-dimensional attitude vectors, and is one of the core components for achieving closed-loop feedback and real-time intervention in the system.

[0099] S42, applying attention weights to the business context data set Each data point is weighted and summed to generate a context vector that focuses on the relevant background information.

[0100] Specifically, in the adaptive context refocusing module, attention weights are applied to the business context data set. The process of weighted summation of each data point to generate a context vector focusing on relevant current background information is a key step in achieving closed-loop feedback for customer attitude identification and operational intervention. The core technical principle of this step is to utilize the difference between the attitude vectors at the current and previous moments. As an attention mechanism, the query (Query) The multidimensional business background information is dynamically weighted to extract the contextual features most relevant to the current attitude change.

[0101] In the specific implementation, It includes structured information such as static customer profiles (e.g., membership level, region), historical behavior (e.g., order records, service requests), and textual business data (e.g., historical work order summaries). Each data item is encoded as a fixed-dimensional vector representation. For example, categorical data uses one-hot encoding, numerical data undergoes normalization, and textual data is generated into semantic embeddings using a pre-trained text encoder (e.g., BERT). These vectors form keys and values, which are then input to the attention module.

[0102] Attention weights are typically calculated using dot product attention or learnable linear transformations. In this invention, and The similarity of each key vector in the dataset is calculated to obtain an attention score. ,in Indicates the first The key vector of each business context data item. Then, through a weighted summation operation, the context vector for the current time step is generated. ,in This is the corresponding value vector.

[0103] The parameter settings for this step include dimension matching for the attention mechanism, the Softmax normalization method, and the weighted summation vector space structure. For example, if The dimension is ,but and It should also be To ensure the feasibility of attention computation, the update frequency of attention weights is synchronized with the time step. Synchronization is typically calculated on a per-turn dialogue or voice segment basis.

[0104] In practical applications, this step significantly enhances the system's ability to interpret changes in customer attitudes. For example, when a customer's attitude shifts from "satisfied" to "dissatisfied," the attention mechanism can automatically focus on business data related to service quality and historical complaint records, thereby providing a more accurate basis for subsequent operational interventions. By dynamically coupling attitude evolution signals with business contextual information, this step achieves adaptive modulation of customer background information, effectively compensating for the shortcomings of static fusion methods in existing technologies and enhancing the system's contextual awareness and responsiveness.

[0105] S5 feeds the updated context vector back to the feature modulation step, and compares the current attitude vector with the preset threshold vector to trigger the corresponding operational intervention command.

[0106] Specifically, in the closed-loop operation intervention engine of this invention, feeding back the updated context vector to the feature modulation step, and comparing the current attitude vector with a preset threshold vector to trigger the corresponding operation intervention instruction, is a key step in achieving adaptive feedback and real-time response of the system. This step employs a joint decision-making architecture based on an attention mechanism and dynamic threshold comparison at the technical implementation level.

[0107] Specifically, the adaptive context refocusing module at time step Generate updated context vectors This vector, obtained through weighted summation, is derived from a multidimensional business context data set. Extract the most relevant background information. The dimension is usually related to the interaction feature vector. To maintain consistency and ensure element-wise affine transformations within the dynamic context modulation module. In some implementations, The update mechanism can adopt GRU units or linear combination, and its update frequency is consistent with the sampling frequency of the interaction features, usually 1 to 5 time steps per second, depending on the framing strategy of the speech stream and the system response requirements.

[0108] At the same time, the attitude evolution tracking module outputs the N-dimensional attitude vector at the current time step. Each element Indicates the customer at the time step For the The quantitative strength of the attitude category. The closed-loop operation intervention engine receives this attitude vector and compares it with a preset N-dimensional threshold vector. Perform an element-by-element comparison. If any... Then it is determined that the first one is triggered. Intervention conditions corresponding to class attitudes. Threshold vector. It can be dynamically configured according to business scenarios, such as in customer service scenarios. It can be set to 0.75 (normalized strength value) to ensure that intervention is triggered only when there is a significant change in customer attitude.

[0109] This step is widely used in practical applications such as customer service, marketing automation, and complaint alerts. For example, when a customer's mood deteriorates sharply, the system can automatically trigger a "transfer to a live agent" instruction, or push personalized promotional strategies when a customer shows strong purchase intent. This is achieved through contextual vectorization. Feedback is sent to the dynamic context modulation module, enabling the system to continuously modulate interaction features, thereby improving the contextual sensitivity of attitude recognition in subsequent time steps. This feedback mechanism significantly enhances the system's ability to continuously perceive customer status, providing closed-loop support for precise and real-time operational intervention.

[0110] S6 deploys the system as a streaming service, using a Kafka+Flink architecture to process voice stream segments in real time, and uses the MLOps system to achieve continuous integration, continuous deployment and performance monitoring of models.

[0111] Specifically, this step involves deploying the system as a streaming processing service, using a Kafka + Flink architecture to achieve real-time processing of audio stream segments, and using an MLOps system to complete continuous model integration, continuous deployment, and performance monitoring. This step is a crucial link in achieving closed-loop feedback and real-time response capabilities for the entire system, and its technical implementation must balance high throughput, low latency, scalability, and automated model management.

[0112] At the technical implementation level, Kafka, as a distributed message queue, is responsible for receiving and buffering real-time audio stream segments from the voice acquisition end. Each segment is identified by a timestamp to ensure the orderliness and timeliness of the data. The audio stream data in Kafka is divided into several time windows (e.g., 500ms or 1s) and encapsulated in JSON or Protobuf format, containing metadata such as voice encoding, session ID, and timestamp. Flink, as a streaming computing engine, subscribes to the audio stream data through Kafka consumers and performs real-time feature extraction, semantic analysis, and attitude recognition tasks within its streaming processing topology. Flink's event-time processing mechanism, combined with a watermarking strategy, ensures that state consistency is maintained even when there are delays or out-of-order delivery in the audio stream.

[0113] At the parameter level, the number of Kafka partitions is typically set to match the parallelism of Flink to achieve load balancing. The recommended checkpoint interval for Flink is 5 to 10 seconds to balance state recovery and processing latency. The model inference service is deployed as a Flink UDF (User-Defined Function), supporting asynchronous calls to reduce the impact of inference latency on overall processing performance. In the MLOps system, model version management uses MLflow or Kubeflow, and model deployment is containerized via Kubernetes, supporting rolling updates and canary releases. Performance monitoring metrics include inference latency (P99 < 200ms), throughput (≥ 1000 audio segments / second), and model accuracy (F1-score ≥ 0.85), visualized using Prometheus + Grafana.

[0114] At the application level, this deployment method is suitable for high-concurrency, low-latency customer service scenarios, such as online customer service and intelligent voice response systems. Voice stream segments can come from VoIP, call centers, or real-time conferencing systems. The system needs to complete feature extraction, semantic analysis, and attitude recognition within milliseconds, and trigger operational intervention actions (such as transferring to human agents, sending coupons, etc.) based on the recognition results.

[0115] The technical advantage of this step lies in the fact that, through a streaming processing architecture built with Kafka and Flink, real-time processing and analysis of voice data are achieved, ensuring the system's immediate response to changes in customer attitudes. Simultaneously, the introduction of the MLOps framework automates model iteration, version control, and performance monitoring, significantly improving the system's maintainability and adaptability, and providing a stable and efficient operational foundation for subsequent adaptive context refocusing and closed-loop intervention.

[0116] This invention discloses a customer attitude recognition and operational intervention method based on AI semantic analysis. By deploying the system as a streaming service based on Kafka+Flink, it achieves low-latency real-time processing of voice stream segments. Combined with the MLOps system to ensure continuous iteration and performance optimization of the model, it further improves the timeliness of customer attitude recognition and the stability and adaptability of the system in complex business scenarios.

[0117] To achieve the above embodiments, the present invention also proposes a customer attitude recognition and operational intervention device based on AI semantic analysis. Figure 2 This is a schematic diagram of a customer attitude recognition and operational intervention device based on AI semantic analysis, provided as an embodiment of the present invention. Figure 2 As shown, the device includes: The data acquisition module is used to acquire real-time voice streams of customer interactions and multi-source heterogeneous business context data, and generate interaction feature vectors and business context data sets respectively. The feature modulation module is used to perform a feature-level affine transformation on the current interaction feature vector based on the business context vector of the previous time step, and generate a modulated feature vector. The temporal modeling module is used to perform temporal modeling on the modulated feature vector using a gated loop unit and output the customer attitude vector at the current moment. The context update module is used to drive the attention mechanism based on the difference between the attitude vector at the current time and the previous time, dynamically calculate the attention weight of each data item in the business context data set, and generate the updated context vector by weighting. The feedback intervention module is used to feed the updated context vector back to the feature modulation step, and at the same time compare the current attitude vector with the preset threshold vector to trigger the corresponding operational intervention instruction.

[0118] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0119] To implement the above embodiments, the present invention also proposes an electronic device, comprising: a processor, and a memory communicatively connected to the processor; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory to implement the method provided in the foregoing embodiments.

[0120] To implement the above embodiments, the present invention also proposes a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the methods provided in the foregoing embodiments.

[0121] To implement the above embodiments, the present invention also proposes a computer program product, including a computer program that, when executed by a processor, implements the methods provided in the foregoing embodiments.

[0122] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0123] It should be noted that personal information collected from users should be used for legitimate and reasonable purposes and should not be shared or sold outside of these legitimate uses. Furthermore, such collection / sharing should only be conducted after receiving the user's informed consent, including but not limited to notifying the user to read the user agreement / user notice and sign an agreement / authorization that includes authorization of relevant user information before the user uses the function. In addition, any necessary steps must be taken to protect and safeguard access to such personal information data and ensure that others with access to personal information data comply with their privacy policies and procedures.

[0124] This invention is intended to provide implementation schemes for users to selectively prevent the use or access to personal information data. That is, this disclosure is intended to provide hardware and / or software to prevent or block access to such personal information data. Once personal information data is no longer needed, risks can be minimized by restricting data collection and deleting data. Furthermore, where applicable, such personal information can be de-identified to protect user privacy.

[0125] In the foregoing descriptions of the embodiments, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0126] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0127] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of the invention pertain.

[0128] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0129] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0130] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0131] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0132] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

[0133] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0134] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for customer attitude identification and operational intervention based on AI semantic analysis, characterized in that, Includes the following steps: S1, acquire real-time voice streams of customer interactions and multi-source heterogeneous business context data, and generate interaction feature vectors and business context data sets respectively; S2, based on the business context vector of the previous moment, performs a feature-level affine transformation on the current interaction feature vector to generate a modulated feature vector; S3 uses a gated loop unit to perform time-series modeling on the modulated feature vector and outputs the customer attitude vector at the current moment. S4. Based on the difference between the attitude vectors at the current time and the previous time, the attention mechanism is driven to dynamically calculate the attention weight of each data item in the business context data set and generate an updated context vector by weighting. S5 feeds the updated context vector back to the feature modulation step, and compares the current attitude vector with the preset threshold vector to trigger the corresponding operational intervention command.

2. The method according to claim 1, characterized in that, S1 includes: S11 extracts acoustic features such as Mel-frequency cepstral coefficients, pitch, and energy from the real-time speech stream, generating an acoustic feature vector. ; S12, after converting the speech stream into text, inputs it into a pre-trained AI semantic analysis model to generate text semantic embedding vectors. And through vector concatenation operations and Combined into interactive feature vectors .

3. The method according to claim 1, characterized in that, S2 includes: S21, the scaling factor vector is calculated through two parallel fully connected neural network layers. and offset factor vector ; S22, using the formula Interactive feature vectors Perform a feature-level affine transformation to generate the modulated feature vector. .

4. The method according to claim 1, characterized in that, The S3 includes: S31, through the update and reset gate mechanisms of the GRU unit, the current modulated feature vector is... Hidden state from the previous moment Merge and calculate the current hidden state. ; S32, hide state Input a fully connected layer with either a Softmax or Sigmoid activation function, decode to generate a customer attitude vector containing N quantized intensity values. .

5. The method according to claim 1, characterized in that, The S4 includes: S41, Calculate the current attitude vector attitude vector from the previous moment The difference As the query input for the attention mechanism; S42, applying attention weights to the business context data set Each data point is weighted and summed to generate a context vector that focuses on the relevant background information.

6. The method according to claim 1, characterized in that, Also includes: S6 deploys the system as a streaming service, using a Kafka+Flink architecture to process voice stream segments in real time, and uses the MLOps system to achieve continuous integration, continuous deployment and performance monitoring of models.

7. A customer attitude recognition and operational intervention device based on AI semantic analysis, characterized in that, include: The data acquisition module is used to acquire real-time voice streams of customer interactions and multi-source heterogeneous business context data, and generate interaction feature vectors and business context data sets respectively. The feature modulation module is used to perform a feature-level affine transformation on the current interaction feature vector based on the business context vector of the previous time step, and generate a modulated feature vector. The temporal modeling module is used to perform temporal modeling on the modulated feature vector using a gated loop unit and output the customer attitude vector at the current moment. The context update module is used to drive the attention mechanism based on the difference between the attitude vector at the current time and the previous time, dynamically calculate the attention weight of each data item in the business context data set, and generate the updated context vector by weighting. The feedback intervention module is used to feed the updated context vector back to the feature modulation step, and at the same time compare the current attitude vector with the preset threshold vector to trigger the corresponding operational intervention instruction.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory to implement the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method of any one of claims 1-6.

Images