Intention recognition method and equipment applied to intelligent customer service of power self-service terminal

Abstract

The invention provides an intention recognition method and intention recognition equipment applied to intelligent customer service of an electric power self-service terminal. According to the implementation scheme, based on a bidirectional long and short memory neural network, memory learning is carried out on a voice query text vector of the power self-service terminal, and a first time sequence text feature used for representing a query intention is obtained; dividing the first time sequence text feature to obtain a plurality of second time sequence text features; on the basis of the query vector, the key vector and the value vector of each second time sequence text feature, respectively determining a query related feature vector of each second time sequence text feature; performing feature fusion on the query related feature vector of each second time sequence text feature to obtain a third time sequence text feature; and determining a target query intention of the voice query text vector based on a residual connection result of the first time sequence text feature and the third time sequence text feature. According to the invention, the intention recognition accuracy can be improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an intent recognition method and device for intelligent customer service in power self-service terminals. Background Technology

[0002] With the intelligent development of power services, self-service terminals are being used more and more widely in power service halls. A self-service terminal is a device that integrates multiple functions, allowing users to pay electricity bills, inquire about services, and report faults. Currently, most self-service terminals use a fixed menu-based interactive system, requiring users to click through layers of the touchscreen to complete the operation.

[0003] In recent years, some terminals have introduced voice recognition functionality; however, their technical solutions still have significant limitations. In some technologies, the terminal only identifies the user's business intent through keyword matching, unable to provide dynamic answers based on the user's ambiguous intent. For example, when a user inquires about procedures outside of business processing, the system cannot recognize the user's specific intent and can only provide a fixed response, unable to offer personalized answers based on the user's voice question.

[0004] In some techniques, Long Short-Term Memory (LSTM) networks are used to construct intent classifiers. While the gating mechanism in LSTM can mitigate the effects of vanishing gradients, for extremely long sequences, crucial information may still be prematurely discarded by subsequent gating operations. Furthermore, its ability to extract local word-level or phrase-level features is relatively weak, resulting in unsatisfactory intent recognition performance for shorter sequence files. Summary of the Invention

[0005] This invention provides an intent recognition method, apparatus, device, and storage medium for intelligent customer service in power self-service terminals, which can solve at least one of the above-mentioned technical problems.

[0006] According to one aspect of the present invention, an intent recognition method for intelligent customer service in power self-service terminals is provided, comprising: Based on a bidirectional long short-term memory neural network, the voice query text vector of the power self-service terminal is learned to obtain the first temporal text feature used to represent the query intent. The first temporal text features are divided to obtain multiple second temporal text features; Based on the query vector, key vector, and value vector of each of the second time-series text features, the query-related feature vector of each of the second time-series text features is determined respectively; The query-related feature vectors of each of the second time-series text features are fused to obtain the third time-series text features; based on the residual connection results of the first time-series text features and the third time-series text features, the target query intent of the voice query text vector is determined.

[0007] According to another aspect of the present invention, an intent recognition device for intelligent customer service in a power self-service terminal is provided, comprising: The first intent determination module is used to learn the voice query text vector of the power self-service terminal based on a bidirectional long short-term memory neural network to obtain the first temporal text features that represent the query intent. The temporal feature decomposition module is used to decompose the first temporal text feature to obtain multiple second temporal text features; the query feature determination module is used to determine the query-related feature vector of each second temporal text feature based on its respective query vector, key vector and value vector. The feature fusion module is used to fuse the query-related feature vectors of each of the second time-series text features to obtain the third time-series text features; The second intent determination module is used to determine the target query intent of the voice query text vector based on the residual connection result between the first temporal text feature and the third temporal text feature.

[0008] According to one aspect of the present invention, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the processor, and the processor is configured to retrieve the instructions from the memory and execute the instructions to enable the processor to execute the intent recognition method for intelligent customer service in a power self-service terminal as described in any embodiment of the present invention.

[0009] According to one aspect of the present invention, a non-transitory computer-readable storage medium storing computer instructions is provided, the computer instructions being provided to a computer to instruct the computer to execute the intent recognition method for intelligent customer service in a power self-service terminal as described in any one of the embodiments of the present invention.

[0010] This invention employs a bidirectional long short-term memory (LSSM) neural network to learn and memorize the voice query text vectors from a power self-service terminal, obtaining first temporal text features representing the query intent. Thus, the LSSM neural network performs preliminary query intent recognition on the voice query text vectors. Next, the first temporal text features are divided to obtain multiple second temporal text features. Based on the query vector, key vector, and value vector of each second temporal text feature, query-related feature vectors are determined for each second temporal text feature. This yields some strongly relevant information related to the current query. Feature fusion is performed on the query-related feature vectors of each second temporal text feature to obtain third temporal text features. Based on the residual connection results between the first and third temporal text features, the target query intent of the voice query text vector is determined. By fusing all strongly relevant information related to the current query and comparing it with the initially determined query intent, the final query intent can be determined, improving the accuracy of intent recognition.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0012] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of the invention. Wherein: Figure 1 This is a flowchart of an intent recognition method applied to intelligent customer service in a power self-service terminal according to an embodiment of the present invention; Figure 2 This is a voice interaction flowchart of an intelligent customer service system for a power self-service terminal according to an embodiment of the present invention. Figure 3 This is a connection diagram of the vector conversion module and the intent recognition module according to an embodiment of the present invention; Figure 4 This is a network diagram of an intent recognition module according to an embodiment of the present invention; Figure 5 This is a network diagram of an intent recognition module according to another embodiment of the present invention; Figure 6 This is a flowchart of an intent recognition device applied to an intelligent customer service system for power self-service terminals according to an embodiment of the present invention; Figure 7 This is a block diagram of an electronic device used to implement the methods of embodiments of the present invention. Detailed Implementation

[0013] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0014] Figure 1 This is a flowchart of an intent recognition method applied to intelligent customer service in a power self-service terminal, according to an embodiment of the present invention.

[0015] like Figure 1 As shown, the intent recognition method includes: S110, based on a bidirectional long short-term memory neural network, performs memory learning on the voice query text vector of the power self-service terminal to obtain the first temporal text feature used to represent the query intent; S120, the first time-series text features are divided to obtain multiple second time-series text features; S130, based on the query vector, key vector and value vector of each second time-series text feature, determine the query-related feature vector of each second time-series text feature respectively; S140, perform feature fusion on the query-related feature vectors of each second time-series text feature to obtain the third time-series text feature; S150, determine the target query intent of the voice query text vector based on the residual connection results of the first time-series text feature and the third time-series text feature.

[0016] Understandably, the intent recognition method provided in this application can be applied to intelligent customer service in power self-service terminals to recognize the intent of user-provided voice. Of course, it can also be applied to other terminals to recognize the intent of user voice in other scenarios.

[0017] For example, such as Figure 2 As shown, the self-service terminal in the power business hall, using the method of this application, can achieve intelligent voice interaction capabilities, providing users with a more convenient and efficient business processing experience. Specifically, when a user activates the intelligent customer service system on the self-service terminal, the terminal begins initialization and inquires about the user's intended operation. The microphone of the self-service terminal collects the user's voice information and sends it to the speech recognition module to convert it into text. The intent recognition method provided in this application is used to understand the converted speech text. If the intent understanding result is a process instruction, the corresponding process of the business guidance module is executed to guide the user to complete the business processing. If the intent understanding result is a question, the intelligent question-and-answer module is invoked to generate a text answer based on the question, which is then input into the speech synthesis module and converted into speech to answer the user through the speaker.

[0018] For example, a smart Q&A customer service system can be set up in a self-service terminal. This smart Q&A customer service system may include a speech recognition module, an intent understanding module, a business guidance module, a smart Q&A module, and a speech synthesis module.

[0019] The speech recognition module processes the input speech and converts it into text. It employs a Conformer end-to-end model, and the training data includes recordings of real-world noise environments in power company service centers. The intent understanding module analyzes the text input by the user to accurately determine the user's intent, which is the foundation for the intelligent customer service system to achieve efficient guidance and question answering. This intent understanding module can execute the intent recognition method provided in the embodiments of this application.

[0020] The business guidance module is used to assist users in completing operations smoothly through voice and visual guidance when the user clearly understands the intention of the business operation.

[0021] The intelligent question-answering module is used to process user questions and generate answers based on a domain-adapted large language model, achieving accurate semantic understanding and related question handling in the power customer service scenario.

[0022] The speech synthesis module is used to generate human-like speech based on a lightweight speech synthesis method with dynamic emotion embedding, and the tone can be dynamically adjusted to match the business scenario.

[0023] For example, a voice recognition module is used to perform voice recognition on the user's query voice at the power self-service terminal to obtain the voice query text.

[0024] For example, for a speech recognition model, the user's speech signal is first acquired through a high-sensitivity microphone array, and the acquired speech signal is then subjected to noise reduction processing. An adaptive filtering algorithm is used to remove background noise and improve speech clarity.

[0025] For example, a speech recognition model can employ a Conformer hybrid architecture as its core algorithm. This model combines the local feature extraction capabilities of Convolutional Neural Networks (CNNs) with the global context modeling advantages of Transformer models, forming an end-to-end speech-to-text solution. In terms of model structure, the input layer directly receives the pre-processed audio signal, extracts 80-dimensional Mel-spectral features through frame-by-frame processing, and then feeds it into a 12-layer Conformer Block for deep feature learning. Each Conformer Block contains a multi-head self-attention module, a convolutional enhancement module, and a feedforward neural network module. The multi-head attention mechanism captures long-distance speech dependencies, and convolutional operations enhance the extraction capability of local acoustic patterns. The decoder employs a joint CTC / Attention mechanism. The CTC loss function is used to force alignment between acoustic features and text, while the Attention mechanism is responsible for generating the final text sequence. Joint training of these two mechanisms significantly improves the model's robustness in noisy environments.

[0026] Understandably, to adapt to the high-noise environment of power business halls, multi-dimensional data augmentation strategies can be introduced during the model training phase of speech recognition and intent recognition models.

[0027] For example, the basic training data can include a voice library in the power sector, covering typical business scenarios such as electricity bill payment, business processing, and policy consultation. At the same time, it can collect environmental noise from the business hall (such as printer noise, equipment humming, and human voice interference) and mix the data to simulate a complex acoustic environment with a signal-to-noise ratio of 0-20dB.

[0028] For example, speech recognition results, i.e., speech-recognized text, generally cannot be directly applied to intent recognition models. For instance... Figure 3 As shown, the speech recognition text also needs to be vectorized to obtain a speech recognition text vector before it can be input into the intent recognition module for intent recognition.

[0029] For example, a word segmentation algorithm based on Conditional Random Fields (CRF), combined with a vocabulary dictionary for electricity business processing, is used to segment the speech recognition text, ensuring that the segmented words are meaningful. Word frequency statistics are then performed on the segmented text to count the number of times each word appears in the text. Based on the word frequency statistics results, a word frequency threshold of 5 is set to construct a vocabulary, and an index is assigned to each word in the vocabulary. Subsequently, a sliding window is used to move across the original corpus, generating word pairs, which are then used to train the word2vec model. The trained module then feeds the speech recognition text into the trained word2vec model for processing, obtaining the speech recognition text vector.

[0030] Understandably, in step S110, the voice query text vector is input into a bidirectional long short-term memory neural network, which performs intent learning on the voice query text vector to obtain and output the first temporal text feature used to represent the query intent.

[0031] However, when processing long short-term memory (LSTM) text vectors, bidirectional long short-term memory (BSSM) neural networks may prematurely discard important information due to gating operations, and their ability to extract local word-level or phrase-level features is weak, resulting in poor performance in some fine-grained intent recognition. Therefore, steps S120 to S150 are used to enhance intent recognition and obtain accurate query intent.

[0032] For example, the first temporal text feature can be in the form of a three-dimensional matrix. This three-dimensional matrix can be divided, for example, by average division, to obtain multiple sub-matrices, which are the second temporal text features. A sub-matrix can also be understood as a subspace feature.

[0033] For example, query vector, key vector and value vector analysis are performed on each time-series feature in the second time-series text features to obtain triple vectors of multiple time-series features. Attention analysis is performed on the query vector and key vector in the triple vectors of these multiple time-series features to determine the value probability distribution of the corresponding value vector. Then, by using each value vector and its value probability distribution, the query-related feature vector of the second time-series text features can be determined.

[0034] Understandably, feature fusion is performed on the query-related feature vectors of each second time-series text feature to obtain a third time-series text feature used to characterize enhanced query intent.

[0035] For example, by comparing the first temporal text features representing the initial intent with the third temporal text features representing the enhanced query intent, i.e. the residual connection result of the two, the target query intent of the voice query text vector can be accurately determined.

[0036] According to the above implementation method, based on a bidirectional long short-term memory (LSTM) neural network, the voice query text vector of the power self-service terminal is learned through memory learning to obtain a first temporal text feature representing the query intent. Thus, a bidirectional LSTM neural network is used to perform preliminary query intent recognition on the voice query text vector. Then, the first temporal text feature is divided to obtain multiple second temporal text features; based on the query vector, key vector, and value vector of each second temporal text feature, the query-related feature vector of each second temporal text feature is determined. In this way, some strongly relevant information related to the current query can be obtained. The query-related feature vectors of each second temporal text feature are fused to obtain a third temporal text feature; based on the residual connection results of the first and third temporal text features, the target query intent of the voice query text vector is determined. Thus, by fusing all the strongly relevant information related to the current query and comparing it with the initially determined query intent, the final query intent can be determined, improving the accuracy of intent recognition.

[0037] In one implementation, based on the query vector, key vector, and value vector of each second time-series text feature, the query-related feature vector of each second time-series text feature is determined, including: using parallel first fully connected layers, second fully connected layers, and third fully connected layers to process the second time-series text features respectively, obtaining multiple triple vectors of the second time-series text features composed of query vectors, key vectors, and value vectors; calculating the dot product between the query vector and the key vector in each triple vector; scaling the dot product of each triple vector; using a first normalization layer to process the scaled dot product of each triple vector, obtaining the value probability distribution of the value vector in each triple vector; and performing a weighted summation of the value vectors in each triple vector based on the value probability distribution of the value vectors in each triple vector to obtain the query-related feature vector of the second time-series text feature.

[0038] Understandably, the second temporal text features can be a subspace matrix, which undergoes multiple attention processing to obtain the output vector of each subspace, i.e., the query-related feature vector.

[0039] For example, each subspace matrix is ​​processed through three fully connected layers to generate query vectors, key vectors, and value vectors, which together form a triplet vector, creating a differentiated semantic analysis perspective. The query vector represents the information or feature to be focused on, the key vector represents the identifier of the feature to be focused on, and the value vector represents the actual feature value of the feature to be focused on.

[0040] like Figure 4As shown, the temporal text features provided by the bidirectional LSTM are divided into subspaces. Then, in each subspace, the dot product of each query vector and each key vector is calculated to obtain an attention score matrix S = Query * Key. This score matrix represents the similarity between the query vector and each key vector.

[0041] To prevent the dot product from becoming too large and causing gradient vanishing or exploding problems, the dot product result is divided by a scaling factor sqrt(d_k) after calculation. Here, sqrt() represents the square root, and d_k represents the dimension of the key vector, resulting in the scaled fractional matrix SS = S / sqrt(d_k).

[0042] The scaled score matrix is ​​normalized using the Softmax function, ensuring that each score value is between 0 and 1, and that the sum of all scores is 1. This yields a probability distribution representing the relevance weight of each value vector to the current query vector.

[0043] Finally, the probability distributions of each value are weighted and summed with the corresponding value vectors to obtain the output vector of each subspace matrix, which is the query related feature vector.

[0044] In this example, the output vector of the subspace matrix can be seen as the result of a weighted combination of the value vectors at different positions in the input subspace sequence based on their relevance to the query vector. Thus, the output vector contains information related to the current query in the input subspace sequence. Using this output vector to enhance the query intent can improve the accuracy of intent recognition.

[0045] In one implementation, a first temporal text feature representing the query intent is obtained by learning the voice query text vector of a power self-service terminal based on a bidirectional long short-term memory neural network. This includes: using multiple parallel convolutional networks to convolve the voice query text vector to obtain convolutional text features output by each convolutional network; using channel attention modules connected to each convolutional network to perform attention weighting on the convolutional text features output by each convolutional network to obtain attention-weighted text features output by each channel attention module; fusing the attention-weighted text features output by each channel attention module to obtain fused text features; and using a bidirectional long short-term memory neural network to learn the fused text features to obtain the first temporal text feature.

[0046] In one implementation, multiple convolutional networks are used in parallel to convolve the voice query text vectors respectively, and the convolutional text features output by each convolutional network are obtained as follows: a first-scale convolutional network is used to convolve the voice query text vectors to obtain first-scale convolutional text features; a second-scale convolutional network is used to convolve the voice query text vectors to obtain second-scale convolutional text features; and a third-scale convolutional network is used to convolve the voice query text vectors to obtain third-scale convolutional text features; wherein the kernel size of the first-scale convolutional network is smaller than the kernel size of the second-scale convolutional network, and the kernel size of the second-scale convolutional network is smaller than the kernel size of the third-scale convolutional network.

[0047] For example, the kernel size of the first-scale convolutional network can be 2*2, the kernel size of the second-scale convolutional network can be 3*3, and the kernel size of the third-scale convolutional network can be 5*5.

[0048] like Figure 5 As shown, a parallel multi-scale convolutional neural network is employed, with each network using convolutional kernels of 2, 3, and 5 sizes to capture short-range phrases, mid-range collocations, and long-range entity associations, respectively. Each convolutional branch is followed by a channel attention module, which uses a global average pooling layer to compress features on each channel to reflect the importance of each channel. Next, a fully connected layer learns the weights for each channel, and then a sigmoid function is used to generate the final channel attention weights. These weights are multiplied by the input features to achieve feature weighting, resulting in fused text features. Finally, the fused text features are input into a bidirectional long short-term memory neural network (Bidirectional LSTM), which learns and memorizes the fused text features to obtain the first temporal text features.

[0049] In this example, by using multi-scale fusion and channel attention mechanisms, the model can adaptively learn important features in the text, thereby suppressing redundant features and strengthening the discrimination of local features.

[0050] In one implementation, the first temporal text features are divided to obtain multiple second temporal text features, including: splitting the third dimension of the three-dimensional matrix of the first temporal text features by average to determine multiple splitting positions; and based on each splitting position, dividing the three-dimensional matrix of the first temporal text features along the third dimension to obtain multiple second temporal text features.

[0051] For example, the three-dimensional matrix of the first temporal text feature is a (128, 1024, 64) dimension matrix, and the third dimension is 64 dimensions. Then, this dimension is divided into 8 sub-space matrices, and each sub-space matrix has a dimension of (128, 1024, 8).

[0052] In this example, by taking the last dimension of the three-dimensional matrix as a baseline and splitting the three-dimensional matrix corresponding to the first time-series text features on an average basis, multiple second time-series text features can be obtained.

[0053] In one implementation, feature fusion is performed on the query-related feature vectors of each second time-series text feature to obtain a third time-series text feature, including: concatenating the query-related feature vectors of each second time-series text feature to obtain a concatenated feature vector; and using a fourth fully connected layer to perform feature fusion on the concatenated feature vector to obtain the third time-series text feature.

[0054] like Figure 4 As shown, the output vectors of all subspaces are concatenated to form a higher-dimensional vector, which is then fed into a fully connected layer to obtain the final multi-attention output. In this way, information from multiple attention heads can be fused, enabling the model to capture complex relationships in the input sequence from different subspaces.

[0055] In one implementation, determining the query intent of a voice query text vector based on the residual connection result of the first temporal text features and the third temporal text features includes: using a second normalization layer to normalize the residual connection result of the first temporal text features and the third temporal text features to obtain the query intent of the voice query text vector.

[0056] like Figure 4 As shown, the multi-attention output vector is added to the temporal features of the bidirectional LSTM output through residual connections, and then processed by a normalization layer. The normalization layer uses the Softmax function. The combination of residual connections and normalization enables more stable model training, effectively solving the gradient vanishing problem in deep networks while preserving the original information of the input. Ultimately, it generates hybrid features that consider both local details and global semantics, which are then input into the Softmax classifier to determine intent.

[0057] Figure 6 This is a structural block diagram of an intent recognition device applied to intelligent customer service in a power self-service terminal, according to an embodiment of the present invention.

[0058] like Figure 6 As shown, the intent recognition device applied to intelligent customer service in power self-service terminals includes: The first intent determination module 610 is used to learn the voice query text vector of the power self-service terminal based on a bidirectional long short-term memory neural network to obtain the first temporal text features that represent the query intent. The temporal feature segmentation module 620 is used to segment the first temporal text features to obtain multiple second temporal text features; The query feature determination module 630 is used to determine the query-related feature vectors of each of the second time-series text features based on their respective query vectors, key vectors, and value vectors. The feature fusion module 640 is used to fuse the query-related feature vectors of each of the second time-series text features to obtain the third time-series text features; The second intent determination module 650 is used to determine the target query intent of the voice query text vector based on the residual connection result between the first temporal text feature and the third temporal text feature.

[0059] In one implementation, the query feature determination module includes: The triple vector determination unit is used to process the second time-series text features by using a parallel first fully connected layer, a second fully connected layer and a third fully connected layer, respectively, to obtain multiple triple vectors of the second time-series text features, which are composed of query vectors, key vectors and value vectors. The dot product calculation unit is used to calculate the dot product between the query vector and the key vector in each of the triplet vectors; A scaling processing unit is used to scale the dot product of each of the triplet vectors; The normalization processing unit is used to process the scaled dot product of each triplet vector using the first normalization layer to obtain the value probability distribution of the value vector in each triplet vector. The weighted summation unit is used to perform weighted summation on the value vectors in each of the triplet vectors based on the value probability distribution of the value vectors in each of the triplet vectors, so as to obtain the query-related feature vector of the second time-series text feature.

[0060] In one implementation, the first intent-determining module includes: A multi-scale convolutional unit is used to convolve the speech query text vector using multiple convolutional networks in parallel, so as to obtain the convolutional text features output by each of the convolutional networks. The attention processing unit is used to apply attention weighting to the convolutional text features output by each of the convolutional networks using the channel attention modules connected to each of the convolutional networks, so as to obtain the attention-weighted text features output by each of the channel attention modules. The feature fusion unit is used to fuse the attention-weighted text features output by each of the channel attention modules to obtain fused text features; The intent prediction unit is used to learn the fused text features by employing the bidirectional long short-term memory neural network to obtain the first temporal text features.

[0061] In one implementation, the multi-scale convolutional unit is specifically used for: A first-scale convolutional network is used to convolve the voice query text vector to obtain first-scale convolutional text features; a second-scale convolutional network is used to convolve the voice query text vector to obtain second-scale convolutional text features; a third-scale convolutional network is used to convolve the voice query text vector to obtain third-scale convolutional text features; wherein, the kernel size of the first-scale convolutional network is smaller than the kernel size of the second-scale convolutional network, and the kernel size of the second-scale convolutional network is smaller than the kernel size of the third-scale convolutional network.

[0062] In one embodiment, the time-series feature splitting module includes: The splitting position determination unit is used to perform an average splitting of the third dimension of the three-dimensional matrix of the first temporal text features to determine multiple splitting positions; The matrix splitting unit is used to segment the three-dimensional matrix of the first temporal text features along the third dimension based on each of the splitting positions, to obtain multiple second temporal text features.

[0063] In one embodiment, the feature fusion module includes: The feature vector concatenation unit is used to concatenate the query-related feature vectors of each of the second time-series text features to obtain a concatenated feature vector; The feature fusion unit is used to perform feature fusion on the concatenated feature vector using the fourth fully connected layer to obtain the third temporal text features.

[0064] In one implementation, the second intent-determining module is specifically used for: A second normalization layer is used to normalize the residual connection results of the first temporal text features and the third temporal text features to obtain the query intent of the voice query text vector.

[0065] The specific functions and examples of each module and submodule of the system in this embodiment of the invention can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0066] The acquisition, storage, and application of user personal information involved in the technical solution of this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0067] According to embodiments of the present invention, the present invention also provides a system and a readable storage medium.

[0068] Figure 7A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0069] like Figure 7 As shown, the electronic device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. The RAM 803 may also store various programs and data required for the operation of the electronic device 800. The computing unit 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0070] Multiple components in electronic device 800 are connected to I / O interface 805, including: input unit 806, such as keyboard, mouse, etc.; output unit 807, such as various types of displays, speakers, etc.; storage unit 808, such as disk, optical disk, etc.; and communication unit 809, such as network card, modem, wireless transceiver, etc. Communication unit 809 allows electronic device 800 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0071] The computing unit 801 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 801 performs the various methods and processes described above, such as the intent recognition method applied to intelligent customer service of a power self-service terminal. For example, in some embodiments, the intent recognition method applied to intelligent customer service of a power self-service terminal can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 800 via ROM 802 and / or communication unit 809. When the computer program is loaded into RAM 803 and executed by the computing unit 801, one or more steps of the intent recognition method applied to intelligent customer service of a power self-service terminal described above can be performed. Alternatively, in other embodiments, the computing unit 801 may be configured by any other suitable means (e.g., by means of firmware) to perform an intent recognition method applied to intelligent customer service for power self-service terminals.

[0072] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0073] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0074] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0075] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0076] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0077] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0078] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. 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 disclosed in this invention can be achieved, and this is not limited herein.

[0079] 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 principles of this invention should be included within the scope of protection of this invention.

Claims

1. An intent recognition method applied to intelligent customer service in power self-service terminals, characterized in that, include: Based on a bidirectional long short-term memory neural network, the voice query text vector of the power self-service terminal is learned to obtain the first temporal text feature used to represent the query intent. The first temporal text features are divided to obtain multiple second temporal text features; Based on the query vector, key vector, and value vector of each of the second time-series text features, the query-related feature vector of each of the second time-series text features is determined respectively; The query-related feature vectors of each of the second time-series text features are fused to obtain the third time-series text features; based on the residual connection results of the first time-series text features and the third time-series text features, the target query intent of the voice query text vector is determined.

2. The method according to claim 1, characterized in that, The step of determining the query-related feature vectors of each of the second time-series text features based on their respective query vectors, key vectors, and value vectors includes: using parallel first fully connected layers, second fully connected layers, and third fully connected layers to process the second time-series text features respectively, obtaining multiple triple vectors of the second time-series text features composed of query vectors, key vectors, and value vectors; and calculating the dot product between the query vector and the key vector in each triple vector. Scale the dot product of each of the triplet vectors; The first normalization layer is used to process the scaled dot product of each triplet vector to obtain the value probability distribution of the value vector in each triplet vector. Based on the value probability distribution of the value vectors in each of the triplet vectors, the value vectors in each of the triplet vectors are weighted and summed to obtain the query-related feature vector of the second time-series text feature.

3. The method according to claim 1, characterized in that, The method of learning the voice query text vector of the power self-service terminal based on the bidirectional long short-term memory neural network to obtain the first temporal text feature used to represent the query intent includes: using multiple parallel convolutional networks to convolve the voice query text vector respectively to obtain the convolutional text feature output by each of the convolutional networks. By using the channel attention modules connected to each of the convolutional networks, attention weighting is applied to the convolutional text features output by each of the convolutional networks to obtain the attention-weighted text features output by each of the channel attention modules. The attention-weighted text features output by each channel attention module are fused to obtain fused text features; the bidirectional long short-term memory neural network is used to learn the fused text features to obtain the first temporal text features.

4. The method according to claim 3, characterized in that, Multiple convolutional networks are used in parallel to convolve the speech query text vector, respectively, to obtain the convolutional text features output by each of the convolutional networks, including: A first-scale convolutional network is used to convolve the voice query text vector to obtain first-scale convolutional text features; a second-scale convolutional network is used to convolve the voice query text vector to obtain second-scale convolutional text features; a third-scale convolutional network is used to convolve the voice query text vector to obtain third-scale convolutional text features; wherein, the kernel size of the first-scale convolutional network is smaller than the kernel size of the second-scale convolutional network, and the kernel size of the second-scale convolutional network is smaller than the kernel size of the third-scale convolutional network.

5. The method according to any one of claims 1-4, characterized in that, The first temporal text feature is divided to obtain multiple second temporal text features, including: The third dimension of the three-dimensional matrix of the first temporal text features is averaged to determine multiple splitting positions; based on each of the splitting positions, the three-dimensional matrix of the first temporal text features is segmented along the third dimension to obtain multiple second temporal text features.

6. The method according to any one of claims 1-4, characterized in that, The feature fusion of the query-related feature vectors of each of the second time-series text features to obtain the third time-series text features includes: The query-related feature vectors of each of the second time-series text features are concatenated to obtain a concatenated feature vector; A fourth fully connected layer is used to perform feature fusion on the concatenated feature vectors to obtain the third temporal text features.

7. The method according to any one of claims 1-4, characterized in that, The step of determining the query intent of the voice query text vector based on the residual connection result of the first temporal text feature and the third temporal text feature includes: A second normalization layer is used to normalize the residual connection results of the first temporal text features and the third temporal text features to obtain the query intent of the voice query text vector.

8. An intent recognition device applied to intelligent customer service in power self-service terminals, characterized in that, include: The first intent determination module is used to learn the voice query text vector of the power self-service terminal based on a bidirectional long short-term memory neural network to obtain the first temporal text features that represent the query intent. The temporal feature decomposition module is used to decompose the first temporal text features to obtain multiple second temporal text features; The query feature determination module is used to determine the query-related feature vectors of each of the second time-series text features based on their respective query vectors, key vectors, and value vectors. The feature fusion module is used to fuse the query-related feature vectors of each of the second time-series text features to obtain the third time-series text features; The second intent determination module is used to determine the target query intent of the voice query text vector based on the residual connection result between the first temporal text feature and the third temporal text feature.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to said at least one processor; The memory stores instructions that can be executed by the processor, which retrieves the instructions from the memory and executes them to enable the processor to perform the intent recognition method for intelligent customer service in power self-service terminals as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are provided to the computer to instruct the computer to perform the intent recognition method for intelligent customer service in power self-service terminals according to any one of claims 1-7.

Images