Feature extraction method, device and equipment based on lightweight neural network, medium and product

Abstract

The invention provides a feature extraction method, device and equipment based on a lightweight neural network, a medium and a product, and relates to the field of big data. The method comprises the following steps: preprocessing a natural language text input by a user to generate standardized tensor data; performing feature extraction on the standardized tensor data by adopting depth separable convolution to generate a first feature tensor; the first feature tensor represents a user intention feature; the number of input channels and the number of output channels of the depth separable convolution are the same; performing channel shuffling operation on the first feature tensor to obtain a second feature tensor rich in context association; generating a third feature tensor of a fixed dimension according to the second feature tensor; the feature tensor of the fixed dimension is used for classification and question and answer generation. According to the method, the understanding ability and the response speed of the mobile banking intelligent question-answering system are improved, and the user experience and the service efficiency are improved.

Description

Technical Field

[0001] This application relates to the field of big data, and in particular to a feature extraction method, apparatus, device, medium and product based on a lightweight neural network. Background Technology

[0002] Mobile banking intelligent Q&A is a virtual assistant system based on artificial intelligence technology (especially natural language processing and deep learning). It is integrated into the bank's mobile app and allows users to query information, obtain service guidance, or complete simple transactions through natural language dialogue (text or voice) without having to manually navigate menus or fill out complex forms.

[0003] Traditional mobile banking intelligent question-answering systems rely on rule engines or static knowledge bases, which have weak comprehension capabilities and slow response speeds, resulting in poor user experience and low service efficiency. Summary of the Invention

[0004] This application provides a feature extraction method, apparatus, device, medium, and product based on a lightweight neural network to solve the technical problems of weak understanding ability and slow response speed in mobile banking intelligent question-answering systems, resulting in poor user experience and low service efficiency.

[0005] In a first aspect, this application provides a feature extraction method based on a lightweight neural network, comprising: preprocessing natural language text input by a user to generate standardized tensor data;

[0006] The standardized tensor data is subjected to feature extraction using depthwise separable convolution to generate a first feature tensor; the first feature tensor represents user intent features; the number of input channels and the number of output channels of the depthwise separable convolution are the same.

[0007] Perform a channel shuffling operation on the first feature tensor to obtain a second feature tensor rich in contextual relationships;

[0008] Based on the second feature tensor, a third feature tensor with fixed dimensions is generated; the fixed-dimensional feature tensor is used for classification and question answering generation.

[0009] Secondly, this application provides a feature extraction device based on a lightweight neural network, comprising:

[0010] The preprocessing module is used to preprocess the natural language text input by the user and generate standardized tensor data;

[0011] The feature extraction module is used to extract features from the standardized tensor data using depthwise separable convolution to generate a first feature tensor; the first feature tensor represents user intent features; the number of input channels and the number of output channels of the depthwise separable convolution are the same.

[0012] The feature extraction module is also used to perform a channel shuffling operation on the first feature tensor to obtain a second feature tensor rich in contextual association;

[0013] The output module is used to generate a fixed-dimensional third feature tensor based on the second feature tensor; the fixed-dimensional feature tensor is used for classification and question answering generation.

[0014] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0015] The memory stores computer-executed instructions;

[0016] The processor executes computer execution instructions stored in the memory to implement the method as described in any of the first aspects.

[0017] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any of the first aspects.

[0018] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the first aspects.

[0019] This application provides a feature extraction method, apparatus, device, medium, and product based on lightweight neural networks. It preprocesses user-input natural language text to generate standardized tensor data for subsequent feature extraction. Through depthwise separable convolution, it separates inter-channel feature extraction from intra-channel feature propagation, reducing redundant parameters, achieving model lightweighting, lowering model complexity, and improving response speed. Subsequently, a channel shuffling operation enhances the information propagation capability between different branches through channel reorganization, improving the understanding of complex user intentions in financial scenarios. This makes mobile banking intelligent question answering more context-sensitive, more human-like, and improves answer accuracy when facing different questions. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] Figure 1 This is a schematic diagram illustrating an application scenario involved in an embodiment of this application;

[0022] Figure 2 A flowchart illustrating a feature extraction method based on a lightweight neural network, provided for an embodiment of this application;

[0023] Figure 3 This is a schematic diagram of a lightweight neural network structure provided in an embodiment of this application;

[0024] Figure 4 This is a schematic diagram of the structure of a basic feature extraction module provided in an embodiment of this application;

[0025] Figure 5 A schematic diagram of another basic feature extraction module provided in an embodiment of this application;

[0026] Figure 6 A schematic diagram of a channel shuffling operation provided in an embodiment of this application;

[0027] Figure 7 A schematic diagram of a feature extraction device based on a lightweight neural network provided in an embodiment of this application;

[0028] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0029] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0030] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0031] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of the relevant data all comply with the relevant laws, regulations, and standards of the relevant countries and regions, have taken necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation access points for users to choose to authorize or refuse.

[0032] Furthermore, the technical solution involved in this application, which involves big data analysis of user information (including but not limited to personal biometrics, identity data, consumption data, asset data, electronic terminal operation data, etc.) and the use of artificial intelligence technology for automated decision-making, and makes decisions that have a significant impact on personal rights based on the results of automated decision-making, provides users with corresponding operation entry points for users to choose to agree to or reject the results of automated decision-making; if the user chooses to reject, the process will proceed to the expert decision-making process.

[0033] It should be noted that the feature extraction method, apparatus, device, medium and product based on lightweight neural networks provided in this application can be used in the field of big data, or in any field other than big data. The application fields of the feature extraction method, apparatus, device, medium and product based on lightweight neural networks in this application are not limited.

[0034] First, let me explain the terms used in this application:

[0035] Depthwise separable convolution consists of channel-wise convolution and pointwise convolution. Channel-wise convolution performs convolution operations on each channel of the input separately, which can greatly reduce the number of parameters while ensuring feature extraction, but lacks information exchange between channels. Pointwise convolution, on the other hand, can learn information between different channels at the same spatial location and recombine them into a new feature map, enhancing the correlation between channels.

[0036] The specific application scenario of this application is natural language processing (NLP) applications for the financial industry, which are suitable for intelligent question answering systems in mobile banking apps, and improve the user's interactive experience in mobile banking apps through artificial intelligence technology.

[0037] Figure 1 This is a schematic diagram illustrating an application scenario involved in an embodiment of this application, such as... Figure 1 As shown,

[0038] In practical applications, users input natural language commands (such as "check balance", "transfer 500 yuan to Zhang San" or "explain this month's deduction details") in the question input box of a mobile app via voice or text. The question-and-answer system needs to analyze the user's intent in real time, call the relevant financial service modules, and generate a response that conforms to the financial scenario specifications, which is then displayed in the answer box.

[0039] This scenario places extremely high demands on the system's accuracy, real-time performance, and security: On the one hand, financial transactions involve user asset security, requiring the system to have a very low fault tolerance rate (e.g., erroneous transfers must be immediately intercepted); on the other hand, user behavior varies greatly (e.g., dialects, non-standard expressions), requiring the system to understand complex issues through contextual understanding (e.g., "Which expense in last month's spending record is abnormal?"). Furthermore, financial industry terminology is complex (e.g., "structured deposits," "foreign exchange rates"), requiring the system to accurately identify professional terms and combine them with user historical behavior (e.g., transaction records, account types) to provide personalized services.

[0040] Traditional systems rely on rule engines or static knowledge bases, which are difficult to adapt to dynamic scenario requirements, resulting in poor user experience and low service efficiency. There is an urgent need to optimize model performance and interaction logic through deep learning technology.

[0041] In existing technologies, intelligent question answering in mobile banking mainly relies on the following two types of methods:

[0042] 1. Feature engineering solutions based on machine learning: This involves manually extracting text features (such as word frequency and sentence structure) from user input and building a classification model. However, this method relies heavily on prior knowledge to define feature rules and has poor generalization ability for unstructured data (such as colloquial expressions). Machine learning suffers from problems such as high requirements for prior knowledge, greater sensitivity to parameters, insufficient feature extraction, and high computational complexity, which are detrimental to the final implementation results.

[0043] 2. Solution based on traditional deep neural networks (DNN): Use fully connected layers or convolutional neural networks (CNN) to extract semantic features. However, such models have a large number of parameters (e.g., ResNet-50 has more than 23 million parameters), resulting in high training costs and long inference time, making it difficult to meet the real-time response requirements of mobile devices (e.g., user waiting time needs to be controlled within 500ms).

[0044] Therefore, improving the system's comprehension capabilities and response speed, thereby enhancing user experience and service efficiency, has become an urgent technical problem to be solved.

[0045] This application provides a feature extraction method, apparatus, device, medium, and product based on lightweight neural networks, aiming to solve the aforementioned technical problems of existing technologies. By introducing depthwise separable convolution, inter-channel feature extraction is separated from intra-channel feature propagation, reducing redundant parameters, achieving model lightweighting, reducing model complexity, and improving response speed. Furthermore, a channel shuffling operation is introduced to enhance the information propagation capability between different branches through channel recombination, thereby enhancing the understanding of complex user intentions in financial scenarios and improving the accuracy, response speed, and scenario adaptability of mobile banking intelligent question answering.

[0046] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0047] The execution subject of this application embodiment can be, for example, a feature extraction device, or an electronic device (such as a mobile phone) equipped with a feature extraction device. This application embodiment uses a feature extraction device as an example for illustration.

[0048] Figure 2 This application provides a flowchart illustrating a feature extraction method based on a lightweight neural network, as shown in the embodiments below. Figure 2 As shown, the method may include, for example, the following steps:

[0049] S201. Preprocess the natural language text input by the user to generate standardized tensor data;

[0050] The natural language text entered by the user contains financial industry terminology and general semantics;

[0051] For example, users can input natural language text (such as "check balance" or "explain the reason for deduction") through a mobile app. The electronic device receives the natural language text input by the user and performs standardization processing on the input text, such as word segmentation, word vector encoding, and channel adjustment, to generate standardized two-dimensional tensor data, i.e., standardized tensor data, which is convenient for subsequent feature extraction modules to process.

[0052] In one example, preprocessing user-input natural language text can specifically include:

[0053] Data cleaning is performed on the natural language text input by the user to obtain a structured text sequence;

[0054] The structured text sequence is encoded with word vectors to obtain the initial feature tensor;

[0055] The initial feature tensor is channel-adjusted and standardized to obtain standardized tensor data.

[0056] For example, data cleaning of user-input natural language text can remove special characters, standardize capitalization, etc., to obtain a structured text sequence. The word sequence after segmentation can be, for example, [w1, w2, w3, ..., wL], where L is the text length. When a user inputs "How much is my credit card bill?", the system first segments it into "[my][credit card][bill][is][how much][?]".

[0057] By employing pre-trained word vector models or dynamic encoding, each word in the structured text sequence is mapped to a fixed-dimensional word vector, generating an initial feature tensor [L, d]. Here, L is the sequence length, and d is the word vector dimension.

[0058] Word vector encoding maps financial terms (such as "structured deposits") to high-dimensional vectors, preserving their semantic information. 1×1 convolutions adjust the channel width to ensure that vectors of different dimensions adapt to the feature extraction needs of lightweight neural networks. For example, when a user inputs "explain the risks of structured deposits," word vector encoding maps "structured deposits" to [0.3, -0.5, 0.8], and the 1×1 convolution adjusts its channel count to match the model's input layer, avoiding feature loss due to channel mismatch. This technique enables the model to accurately recognize financial terms and generate personalized responses by combining user historical behavior (such as transaction records), significantly reducing the semantic parsing error rate. Furthermore, standardized input formats improve the model's adaptability to non-standard expressions (such as dialect accents), enhancing its generalization ability.

[0059] The initial feature tensor is channel-adjusted using c 1×1 convolutional kernels, reducing the channel dimension from d to c, resulting in a channel-adjusted feature tensor [L, c]. Then, the channel-adjusted feature tensor is standardized, for example, by scaling the feature values ​​to a distribution with a mean of 0 and a variance of 1, ultimately generating standardized tensor data for subsequent feature extraction.

[0060] Raw natural language text is a discrete sequence of symbols that cannot be directly processed by neural networks. Through preprocessing, the text is converted into standardized tensor data, providing a suitable input format for subsequent feature extraction.

[0061] S202. Use depthwise separable convolution to extract features from standardized tensor data to generate a first feature tensor; the first feature tensor represents the user intent feature; the number of input channels and the number of output channels of the depthwise separable convolution are the same.

[0062] Depthwise separable convolution significantly reduces the number of parameters by splitting standard convolution into depthwise convolution and pointwise convolution. Depthwise convolution processes each channel independently, avoiding redundant computation across channels, while pointwise convolution recombines channel features using a 1×1 kernel, allowing the model to reduce computational complexity while retaining accurate recognition of financial terms. The design of identical input and output channel widths further reduces memory access costs and avoids increased computational overhead due to channel expansion. This significantly improves inference speed on mobile devices, reducing user wait time to an acceptable range and meeting the stringent real-time response requirements of financial scenarios. Furthermore, the reduced number of parameters also decreases model training costs and improves deployment efficiency.

[0063] S203. Perform a channel shuffling operation on the first feature tensor to obtain a second feature tensor rich in contextual relationships;

[0064] Channel shuffling operations can include, for example, the transposition and reorganization of channel dimensions to enhance information exchange between different branches.

[0065] By employing multi-level deep separable convolutional blocks and channel shuffling operations, user intent features are extracted hierarchically, and information propagation between channels is enhanced.

[0066] The input data “[My][Credit Card][Bill][Is][How Much][?]” is downsampled in a deep separable block to extract key semantic features (such as the correlation between “Bill” and “Credit Card”); then the feature channels of different branches are reorganized through a channel shuffling operation to ensure the cross-channel propagation of information such as “Bill Amount” and “Billing Period”.

[0067] Perform a channel shuffling operation on the first feature tensor to obtain a second feature tensor rich in contextual relationships, specifically including:

[0068] The channel dimension of the first feature tensor is split into the grouping dimension and the intra-group channel dimension to obtain the seventh intermediate tensor;

[0069] Transpose the grouping dimension and the intra-group channel dimension of the seventh intermediate tensor to obtain the eighth intermediate tensor.

[0070] The transposed eighth intermediate tensor is reshaped into a single-channel dimension, forming the recombined second feature tensor.

[0071] For example, after processing by a depthwise separable convolutional unit, a channel shuffling operation is performed on the output data of each unit. First, the input channels are grouped according to a preset number of groups (g), splitting into a (g, n) dimension; then (g, n) is transposed to (n, g), and finally reshaped into a (g, n) channel dimension. This operation breaks the fixed channel order through channel reorganization, enhancing the information propagation capability between different branches. For example, in a multi-turn dialogue scenario, the channel information of "user's historical transaction records" and "current question intent" can be transferred across channels through the shuffling operation, providing richer contextual features for subsequent semantic parsing.

[0072] Building upon the improved feature extraction efficiency and semantic relevance achieved through deep separable convolutional units, channel shuffling further enhances the information exchange capabilities between channels. Channel reorganization avoids the feature silos problem caused by fixed channel order, enabling the model to dynamically integrate contextual information from multi-turn dialogues (such as "What is the account balance mentioned earlier?"). For example, when a user asks consecutively "What is my credit card bill?" and "Does this bill include interest?", channel shuffling ensures the cross-channel propagation of information regarding "billing amount" and "interest composition," significantly improving the accuracy of resolving complex questions.

[0073] S204. Generate a fixed-dimensional third feature vector based on the second feature tensor; the fixed-dimensional feature vector is used for classification and question answering generation.

[0074] For example, the output layer of a lightweight neural network can include pooling layers and fully connected layers, with the pooling layers employing global average pooling. The output layer performs average pooling on the spatial dimensions of the second feature tensor, compressing the feature values ​​of each channel into a scalar to generate a fixed-dimensional feature vector. For instance, an input of a 32×32×64 feature map would generate a 1×1×64 feature vector after global average pooling.

[0075] The feature extraction method based on lightweight neural networks provided in this application preprocesses the natural language text input by the user to generate standardized tensor data for subsequent feature extraction. Through depthwise separable convolution, inter-channel feature extraction is separated from intra-channel feature propagation, reducing redundant parameters, achieving model lightweighting, lowering model complexity, and improving response speed. Subsequently, a channel shuffling operation is performed, enhancing the information propagation capability between different branches through channel reorganization. This improves the understanding of complex user intentions in financial scenarios, making mobile banking intelligent question answering more context-sensitive, more human-like, and improving answer accuracy when facing different questions.

[0076] The following explains how to use depthwise separable convolution to extract features from standardized tensor data and generate the first feature tensor.

[0077] Example 1: Perform channel copying on normalized tensor data to obtain the first tensor data and the second tensor data;

[0078] Perform a first depthwise separable convolution operation on the first tensor data to obtain a first right branch feature tensor; and perform a second depthwise separable convolution operation on the second tensor data to obtain a first left branch feature tensor.

[0079] The first feature tensor is obtained by feature fusion of the first right branch feature tensor and the first left branch feature tensor.

[0080] Both the first and second depthwise separable convolution operations include, for example, depthwise convolution and pointwise convolution. The first depthwise separable convolution is divided into depthwise convolution and pointwise convolution. By setting different strides for the depthwise convolution of the two, the spatial size of the normalized tensor can be reduced, thus improving the response speed. For example, the stride of the first depthwise separable convolution can be set to 2, compressing the output size to half of the input size.

[0081] For example, the number of channels in both the first and second tensor data is the same as the number of input channels, denoted as C. The first depthwise separable convolution is divided into depthwise convolution and pointwise convolution, with the channel number change mainly determined by the pointwise convolution. For the first tensor data, the pointwise convolution uses C1 1×1 convolution kernels to adjust the channels of the depthwise convolution output, and the number of output channels of the first right branch feature tensor becomes C1. Similarly, for the second tensor data, the pointwise convolution uses C2 1×1 convolution kernels to adjust the channels of the depthwise convolution output, and the number of output channels of the first left branch feature tensor becomes C2. To make C1 = C2 = C / 2, we can directly concatenate the first right branch feature tensor and the first left branch feature tensor, obtaining C1 + C2 = C; ultimately, the number of channels in the first feature tensor is C, thus making the number of input channels the same as the number of output channels. If C1=C2=C, we can first concatenate the first right branch feature tensor and the first left branch feature tensor to obtain C1+C2=2C. Then, we use C 1×1 convolutional kernels to compress the channels to C, resulting in the first feature tensor having the same number of channels as the output tensor. If C1≠C2, we can first adjust the number of channels in both branches to C using 1×1 convolutional kernels, and then fuse them using addition / multiplication or concatenation with a 1×1 convolutional kernel. Finally, the first feature tensor has the same number of channels as the output tensor, resulting in the same number of channels as the input tensor.

[0082] By fusing the first right branch feature tensor and the first left branch feature tensor to make their input and output channels the same, memory usage can be reduced.

[0083] After channel duplication, two tensors with completely identical structures are generated. By setting the parameters of the two depthwise separable convolution operations differently, the spatial size of the standardized tensor data can be compressed, accelerating feature learning and associating contextual semantics, while reducing the impact of redundant parameters on processing speed. The fusion operation can adjust the input and output channel widths to be equal, reducing memory access costs and improving response speed.

[0084] Furthermore, the first right-branch feature tensor can be obtained by performing a first depthwise separable convolution operation on the first tensor data in the following manner:

[0085] Perform pointwise convolution on the first tensor data to obtain the first intermediate tensor;

[0086] Perform a regularization operation on the first intermediate tensor to obtain a regularized second intermediate tensor;

[0087] Applying the first activation function to the regularized second intermediate tensor enhances the nonlinear expressive power of the features, resulting in the third intermediate tensor;

[0088] The first right branch feature tensor is obtained by performing a depthwise separable convolution operation on the third intermediate tensor with a first preset stride.

[0089] For example, the number of channels in the first tensor data can be C_in. A 1×1 convolution kernel with a number of channels C_out can be used to perform a convolution operation on the first tensor data, converting the number of input channels from C_in to C_out, and outputting the first intermediate tensor. A regularization operation is applied to the first intermediate tensor output by pointwise convolution, calculating the mean and variance of each channel and standardizing it to eliminate numerical biases between different samples or feature dimensions, preventing the model from over-relying on certain noisy features, and suppressing overfitting.

[0090] The first activation function can be any function that can introduce non-linear expressive power into the model, such as Rectified Linear Unit 6 (ReLU6). Applying the ReLU6 activation function element-wise to the regularized tensor enhances the non-linear expressive power of the features, resulting in a third intermediate tensor that can capture the complex semantic relationships in the original text.

[0091] The first preset stride can be, for example, 2 or 3. By performing a depthwise separable convolution operation on the third intermediate tensor with the first preset stride, the spatial dimension of the tensor can be reduced to half of its original size (when the first preset stride is 2). By downsampling, redundant local details can be extracted, allowing feature extraction to focus more on key semantic units and improving response speed.

[0092] This branch allows for flexible adjustment of the number of channels through 1×1 convolutions. The computational cost of subsequent depthwise separable convolutions decreases as the number of channels decreases, resulting in a more lightweight overall process.

[0093] Furthermore, the first left-branch feature tensor can be obtained by performing a second depthwise separable convolution operation on the second tensor data in the following manner:

[0094] The first left branch feature tensor is obtained by performing a depthwise separable convolution operation on the second tensor data with a second preset stride.

[0095] The second preset step size can be the same as the first preset step size, or the second preset step size can be different from the first preset step size. The specific setting can be made according to actual conditions, and this application embodiment does not limit this.

[0096] The channel dimension of this branch is consistent with that of the input. It directly extracts global features through depthwise separable convolution, preserving the complete dimension of the original channel. This allows it to cover more potential correlations and avoids missing key information due to channel compression.

[0097] Example 2: Channel-segment the standardized tensor data to obtain a third tensor data and a fourth tensor data; wherein the number of feature channels of the third tensor data and the fourth tensor data is equal to half the number of feature channels of the standardized tensor data;

[0098] Perform a third depthwise separable convolution operation on the third tensor data to obtain the second right branch feature tensor;

[0099] The second right branch feature tensor and the fourth tensor data are fused to obtain the first feature tensor.

[0100] By fusing features from the second right-branch feature tensor and the fourth tensor data, the number of channels can be kept constant, ensuring that the number of input channels is the same as the number of output channels. The specific method for achieving this consistency between the number of input and output channels can be found in the previous sections and will not be elaborated upon here.

[0101] By dividing the channels of the standardized tensor data into two equal parts, the dimensionality of the features is decoupled. The semantic information carried by different channels (such as the "terminology feature channel" and "contextual association channel" in financial text) is split into two independent tensors. Depth-separable convolution is performed only on the third tensor data, allowing for the extraction of deeper features from this part of the channels and further reducing the computational cost of single-branch depth extraction. The fourth tensor data is not convolutionally processed, thus preserving the basic features of the original channels and avoiding information loss caused by excessive convolution, providing an original reference for fusion.

[0102] Furthermore, the second right-branch feature tensor can be obtained by performing a third depthwise separable convolution operation on the third tensor data in the following manner:

[0103] Perform pointwise convolution on the third tensor data to obtain the fourth intermediate tensor;

[0104] Regularize the fourth intermediate tensor to obtain the regularized fifth intermediate tensor;

[0105] The second activation function is applied to the regularized fifth intermediate tensor to enhance the nonlinear expressive power of the features, resulting in the sixth intermediate tensor.

[0106] The second right branch feature tensor is obtained by performing a depthwise separable convolution operation on the sixth intermediate tensor with a third preset stride.

[0107] By using 1×1 convolutional kernels to specifically adjust the channel dimension of the third tensor data, redundant channels can be compressed to reduce subsequent computational costs, or the number of channels can be expanded to accommodate richer sub-channel semantics. Regularization operations are used to standardize the sub-channel feature distribution of the fourth intermediate tensor, eliminating numerical bias in features within sub-channels and preventing the model from becoming overly dependent on certain types of noisy features.

[0108] The second activation function can be any function that can introduce non-linear expressive power into the model, such as a Smooth Maximum Unit (SMU). Smoothing avoids the vanishing gradient problem while preserving the ability to enhance key features.

[0109] The third preset stride can be 1. Combining the depthwise separable convolution with the third preset stride, spatial information of sub-channel features is extracted specifically, accurately matching it with the semantic characteristics of the sub-channels. Depthwise separable convolution can extract deep features from sub-channels while avoiding a surge in computation, making it suitable for small-scale feature processing scenarios involving sub-channels.

[0110] This branch refines the split sub-channels, leveraging their semantic specificity and improving feature quality through regularization and activation. The resulting second right-branch feature tensor efficiently complements the unconvolved fourth tensor, providing a high-quality feature source for subsequent fusion.

[0111] In summary, the feature extraction method based on lightweight neural networks provided in this application preprocesses the natural language text input by the user to generate standardized tensor data for subsequent feature extraction. Through depthwise separable convolution, inter-channel feature extraction is separated from intra-channel feature propagation, reducing redundant parameters, achieving model lightweighting, lowering model complexity, and improving response speed. Subsequently, a channel shuffling operation is performed, enhancing the information propagation capability between different branches through channel reorganization. This improves the understanding of complex user intentions in financial scenarios, making mobile banking intelligent question answering more context-sensitive, more human-like, and improving answer accuracy when facing different questions.

[0112] Furthermore, the mobile banking intelligent question answering system can perform the following intelligent operations for each question and can continuously improve itself, for example:

[0113] (1) For a question raised by a user for the first time, provide a professional and easy-to-understand answer. Subsequent answers should be contextually related to the previous question and more closely resemble human answers in order to address the issue of low-quality answers.

[0114] (2) The application of lightweight networks improves the response speed of intelligent question answering and can quickly transfer to human agents for questions that cannot be understood, thus solving the problem of excessively long response time.

[0115] (3) Provide early warnings or interventions when users' operations may or have already caused risks, protect users' property safety, and improve security early warning.

[0116] (4) After the user finishes asking a question, present the questions and actions that the user may still be interested in, guide the user to ask questions, enrich the knowledge base, and thus achieve personalized recommendations.

[0117] Figure 3 This is a schematic diagram of a lightweight neural network structure provided in an embodiment of this application, as shown below. Figure 3 As shown, the lightweight neural network includes an input preprocessing module, N basic feature extraction modules, and an output layer, where N is a positive integer.

[0118] Input data flows sequentially through these modules, ultimately achieving classification and localization. The data preprocessing module unifies the input data into a two-dimensional input mode, which is beneficial for subsequent feature extraction and processing. The basic feature extraction module enriches the features of the input data, mining the potential information within user features and strengthening feature propagation, thus improving the accuracy of user classification and localization. The output layer is responsible for normalizing the final classification results.

[0119] For example, after the input preprocessing module inputs data to the basic feature extraction module, the output layer outputs the data results; the input preprocessing module includes data input and ordinary convolution with a kernel size of 1×1; the basic feature extraction module includes depthwise separable blocks with a stride of 2, depthwise separable blocks with a stride of 1, channel shuffling, and channel segmentation; the output layer includes pooling layers and fully connected layers, and the data passes through the pooling layers and fully connected layers once, with the pooling layers using global average pooling.

[0120] Training methods for lightweight application networks may include the following steps:

[0121] S1. Obtain user usage data;

[0122] For example, when users interact with financial institutions through a mobile app, the financial institution's servers can store these data interaction records. For instance, when a user checks their balance, after receiving the query request, the server not only returns the balance data but also records the complete information of the request in the log. By accessing the database, the server can obtain user usage data, such as "check balance," "transfer xx yuan to xx," or "why was xx yuan deducted from my account this month, and what is the source of the deduction?"

[0123] S2. Randomly select a portion of the data as the test set, randomly select a portion of the data as the validation set, and use the remaining data as the training set.

[0124] The training set is used for learning model parameters to ensure that the model can fit the core patterns of the data (e.g., user behavior patterns, data feature distribution, etc.); the validation set is used for hyperparameter tuning during training (e.g., learning rate, network depth, etc.) to avoid the model overfitting the details of the training set; the test set serves as an independent sample that did not participate in training to objectively evaluate the model's generalization ability and ensure that the evaluation results are true and reliable.

[0125] S3. Input the user data from the training set into the lightweight neural network for training;

[0126] The lightweight neural network design reduces training computation costs, making it suitable for efficient processing of user data. Targeted training on the training set allows the model to focus on learning key features of user data (such as user behavior preferences, core patterns in the data, etc.), avoiding "memory noise" caused by redundant parameters, improving training efficiency while ensuring the model's ability to capture core patterns.

[0127] S4. Input the test set data into the trained lightweight neural network, calculate the feature information obtained by the user data in each layer, and then classify it.

[0128] By tracking the feature outputs of each layer of the neural network, we can analyze the model's understanding of user data and improve the model's interpretability. Classification based on hierarchical features allows the model to select the optimal feature level according to task requirements, improving classification accuracy and providing a clear direction for subsequent optimization.

[0129] The information features of intelligent responses in mobile banking mainly include financial industry terminology, industry jargon, deep semantic understanding, user behavior analysis, and scenario prediction. Due to the huge differences in user usage and behavior, it is necessary to train multiple dialogue scenarios, which leads to a huge number of parameters when training neural networks.

[0130] This application employs depthwise separable convolutions for feature learning, propagation, and classification, which accelerates rapid feature learning, connects to contextual semantics, and significantly reduces the impact of redundant parameters on training costs. The input and output channel widths of the basic feature extraction blocks are equal, minimizing memory access costs. Channel shuffling operations are used to improve communication between channels, further extracting and enriching user language features, semantic conversion, and improving the understanding of preceding semantics by following semantics. Global average pooling is used in the output layer instead of the nonlinear connections of fully connected layers, reducing the training burden of the model and further improving the lightweight nature of the network.

[0131] Figure 4 This is a schematic diagram of the structure of a basic feature extraction module provided in an embodiment of this application, as shown below. Figure 4 As shown, taking a downsampling feature extraction unit with a stride of 2 as an example, at the beginning of each unit, the input is copied to generate two parallel branches, left and right. The left branch consists of a depthwise separable convolution with a stride of 2, and the right branch consists of a convolution kernel with a size of... It consists of pointwise convolution with stride 1 and depthwise separable convolution with stride 2. To reduce memory usage, the left and right branches are merged so that their input and output channels are the same.

[0132] Figure 5 A schematic diagram of another basic feature extraction module provided in an embodiment of this application is shown below. Figure 5 As shown, taking a feature extraction unit with a stride of 1 as an example, at the beginning of each unit, channel segmentation divides the input with c feature channels into two equal parts, with each branch having c / 2 channels. Since fewer branches result in a faster model, the right branch consists of a convolutional kernel with a size of... It consists of pointwise convolution with stride 1 and depthwise separable convolution with stride 1. The left branch is not processed, and the left and right branches are merged to achieve the principle of keeping the number of channels unchanged.

[0133] Figure 6 This is a schematic diagram of a channel shuffling operation provided in an embodiment of this application, as shown below. Figure 6As shown, channel shuffling enables feature recombination, allowing information to flow between different channels. Assume the input layer is divided into g groups, with a total number of channels. First, the channel dimension is split into two dimensions (g, n). Then, these two dimensions are transposed to become (n, g), and finally reshaped into a single dimension (g, n). After fusion, the two feature extraction units with different time lengths undergo channel shuffling to ensure that the channels between the two branches can exchange information.

[0134] The feature extraction method based on lightweight neural networks provided in this application first utilizes the advantages of depthwise separable convolution—small parameter count and strong feature extraction capability—to build a basic feature extraction module with the same input and output channel width, reducing model memory and making the neural network model lightweight. Then, channel shuffling is used to redistribute features and channels, strengthening the feature transfer and propagation capabilities between different channels, which is beneficial to improving the network's ability to learn dialogues in a given scenario and its semantic conversion. This invention uses a lighter neural network model to learn the dialogue features of mobile banking users, improving the quality and efficiency of intelligent question answering.

[0135] Figure 7 A schematic diagram of a feature extraction device based on a lightweight neural network provided in this application embodiment is shown below. Figure 7 As shown, the device may include, for example, a preprocessing module 701, a feature extraction module 702, and an output module 703. The preprocessing module 701 may be, for example, a... Figure 3 The data preprocessing module shown; the feature extraction module 702 can be, for example, a data preprocessing module; Figure 3 The basic feature extraction module shown; the output module 703 can be, for example, the basic feature extraction module; Figure 3 The output layer shown.

[0136] The preprocessing module 701 is used to preprocess the natural language text input by the user to generate standardized tensor data;

[0137] The feature extraction module 702 is used to extract features from the standardized tensor data using depthwise separable convolution to generate a first feature tensor; the first feature tensor represents the user intent features; the number of input channels and the number of output channels of the depthwise separable convolution are the same.

[0138] The feature extraction module 702 is also used to perform a channel shuffling operation on the first feature tensor to obtain a second feature tensor rich in contextual association;

[0139] Output module 703 is used to generate a fixed-dimensional third feature tensor based on the second feature tensor; the fixed-dimensional feature tensor is used for classification and question answering generation.

[0140] One possible implementation is a feature extraction module 702, used to extract features from standardized tensor data using depthwise separable convolution to generate a first feature tensor, including:

[0141] Channel copying is performed on the standardized tensor data to obtain the first tensor data and the second tensor data;

[0142] Perform a first depthwise separable convolution operation on the first tensor data to obtain a first right branch feature tensor; and perform a second depthwise separable convolution operation on the second tensor data to obtain a first left branch feature tensor.

[0143] The first feature tensor is obtained by feature fusion of the first right branch feature tensor and the first left branch feature tensor.

[0144] One possible implementation is a feature extraction module 702, which performs a first depthwise separable convolution operation on the first tensor data to obtain a first right-branch feature tensor, including:

[0145] Perform pointwise convolution on the first tensor data to obtain the first intermediate tensor;

[0146] Perform a regularization operation on the first intermediate tensor to obtain a regularized second intermediate tensor;

[0147] Applying the first activation function to the regularized second intermediate tensor enhances the nonlinear expressive power of the features, resulting in the third intermediate tensor;

[0148] The first right branch feature tensor is obtained by performing a depthwise separable convolution operation on the third intermediate tensor with a first preset stride.

[0149] One possible implementation is a feature extraction module 702, which performs a second depthwise separable convolution operation on the second tensor data to obtain a first left-branch feature tensor, including:

[0150] The first left branch feature tensor is obtained by performing a depthwise separable convolution operation on the second tensor data with a second preset stride.

[0151] One possible implementation is a feature extraction module 702, which is used to extract features from the normalized tensor data using depthwise separable convolution to generate a first feature tensor, specifically including:

[0152] Channel-segmenting is performed on the standardized tensor data to obtain third and fourth tensor data; the number of feature channels of the third and fourth tensor data is equal to half the number of feature channels of the standardized tensor data.

[0153] Perform a third depthwise separable convolution operation on the third tensor data to obtain the second right branch feature tensor;

[0154] The second right branch feature tensor and the fourth tensor data are fused to obtain the first feature tensor.

[0155] One possible implementation is a feature extraction module 702, which performs a third depthwise separable convolution operation on the third tensor data to obtain a second right-branch feature tensor, including:

[0156] Perform pointwise convolution on the third tensor data to obtain the fourth intermediate tensor;

[0157] Regularize the fourth intermediate tensor to obtain the regularized fifth intermediate tensor;

[0158] The second activation function is applied to the regularized fifth intermediate tensor to enhance the nonlinear expressive power of the features, resulting in the sixth intermediate tensor.

[0159] The second right branch feature tensor is obtained by performing a depthwise separable convolution operation on the sixth intermediate tensor with a third preset stride.

[0160] One possible implementation is a feature extraction module 702, which performs a channel shuffling operation on the first feature tensor to obtain a second feature tensor rich in contextual relationships, including:

[0161] The channel dimension of the first feature tensor is split into the grouping dimension and the intra-group channel dimension to obtain the seventh intermediate tensor;

[0162] Transpose the grouping dimension and the intra-group channel dimension of the seventh intermediate tensor to obtain the eighth intermediate tensor.

[0163] The transposed eighth intermediate tensor is reshaped into a single-channel dimension, forming the recombined second feature tensor.

[0164] One possible implementation is a preprocessing module 701, used to preprocess the user-input natural language text to generate standardized tensor data, including:

[0165] Data cleaning is performed on the natural language text input by the user to obtain a structured text sequence;

[0166] The structured text sequence is encoded with word vectors to obtain the initial feature tensor;

[0167] The initial feature tensor is channel-adjusted and standardized to obtain standardized tensor data.

[0168] The feature extraction device based on a lightweight neural network provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0169] Figure 8This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 8 As shown, the electronic device may include at least one processor 801 and a memory 802.

[0170] The memory 802 is used to store programs. Specifically, the program may include program code, which includes computer operation instructions.

[0171] The memory 802 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0172] The processor 801 is used to execute computer execution instructions stored in the memory 802 to implement the actions in the foregoing method embodiments. The processor 801 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0173] Optionally, the electronic device may also include a communication interface 803 for communication and interaction with external devices. In specific implementations, if the communication interface 803, memory 802, and processor 801 are implemented independently, the communication interface 803, memory 802, and processor 801 can be interconnected via a bus to complete communication between them.

[0174] Optionally, in a specific implementation, if the communication interface 803, memory 802, and processor 801 are integrated on a single chip, then the communication interface 803, memory 802, and processor 801 can communicate through an internal interface.

[0175] This application also provides a computer-readable storage medium, which may include various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), and a random access memory (RAM). Specifically, the computer-readable storage medium stores program instructions, which are used to implement the actions of the above-described method implementation.

[0176] This application also provides a computer program product including executable instructions stored in a readable storage medium. At least one processor of an electronic device can read the executable instructions from the readable storage medium, and the at least one processor executes the executable instructions to cause the electronic device to perform the actions described in the method embodiments.

[0177] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A feature extraction method based on a lightweight neural network, characterized in that, include: The user-input natural language text is preprocessed to generate standardized tensor data meaning; The standardized tensor data is subjected to feature extraction using depthwise separable convolution to generate a first feature tensor; the first feature tensor represents user intent features; the number of input channels and the number of output channels of the depthwise separable convolution are the same. Perform a channel shuffling operation on the first feature tensor to obtain a second feature tensor rich in contextual relationships; Based on the second feature tensor, a third feature tensor with fixed dimensions is generated; the fixed-dimensional feature tensor is used for classification and question answering generation.

2. The method according to claim 1, characterized in that, The step of extracting features from the standardized tensor data using depthwise separable convolution to generate a first feature tensor includes: Channel copying is performed on the standardized tensor data to obtain first tensor data and second tensor data; Perform a first depthwise separable convolution operation on the first tensor data to obtain a first right branch feature tensor; and perform a second depthwise separable convolution operation on the second tensor data to obtain a first left branch feature tensor. The first feature tensor is obtained by feature fusion of the first right branch feature tensor and the first left branch feature tensor.

3. The method according to claim 2, characterized in that, The step of performing a first depthwise separable convolution operation on the first tensor data to obtain a first right-branch feature tensor includes: Perform pointwise convolution on the first tensor data to obtain the first intermediate tensor; Perform a regularization operation on the first intermediate tensor to obtain a regularized second intermediate tensor; Applying the first activation function to the regularized second intermediate tensor enhances the nonlinear expressive power of the features, resulting in a third intermediate tensor; The first right branch feature tensor is obtained by performing a depthwise separable convolution operation on the third intermediate tensor with a first preset stride.

4. The method according to claim 3, characterized in that, The step of performing a second depthwise separable convolution operation on the second tensor data to obtain the first left-branch feature tensor includes: The first left branch feature tensor is obtained by performing a depthwise separable convolution operation on the second tensor data with a second preset stride.

5. The method according to claim 1, characterized in that, The step of extracting features from the standardized tensor data using depthwise separable convolution to generate a first feature tensor specifically includes: The normalized tensor data is channel-segmented to obtain a third tensor data and a fourth tensor data; wherein the number of feature channels of the third tensor data and the fourth tensor data is equal to half the number of feature channels of the normalized tensor data; Perform a third depthwise separable convolution operation on the third tensor data to obtain the second right branch feature tensor; The second right branch feature tensor and the fourth tensor data are fused to obtain the first feature tensor.

6. The method according to claim 5, characterized in that, The step of performing a third depthwise separable convolution operation on the third tensor data to obtain the second right-branch feature tensor includes: Perform pointwise convolution on the third tensor data to obtain the fourth intermediate tensor; The fourth intermediate tensor is regularized to obtain the fifth intermediate tensor. The second activation function is applied to the regularized fifth intermediate tensor to enhance the nonlinear expressive power of the features, resulting in the sixth intermediate tensor; The second right branch feature tensor is obtained by performing a depthwise separable convolution operation on the sixth intermediate tensor with a third preset step size.

7. The method according to any one of claims 1 to 6, characterized in that, The step of performing a channel shuffling operation on the first feature tensor to obtain a second feature tensor rich in contextual relationships includes: The channel dimension of the first feature tensor is split into the grouping dimension and the intra-group channel dimension to obtain the seventh intermediate tensor; The grouping dimension and the intra-group channel dimension of the seventh intermediate tensor are transposed to obtain the eighth intermediate tensor. The transposed eighth intermediate tensor is reshaped into a single-channel dimension, forming the recombined second feature tensor.

8. The method according to any one of claims 1 to 6, characterized in that, The preprocessing of user-input natural language text to generate standardized tensor data includes: Data cleaning is performed on the natural language text input by the user to obtain a structured text sequence; The structured text sequence is encoded using word vectors to obtain an initial feature tensor; The initial feature tensor is channel-adjusted and standardized to obtain the standardized tensor data.

9. A feature extraction device based on a lightweight neural network, characterized in that, include: The preprocessing module is used to preprocess the natural language text input by the user and generate standardized tensor data; The feature extraction module is used to extract features from the standardized tensor data using depthwise separable convolution to generate a first feature tensor; the first feature tensor represents user intent features; the number of input channels and the number of output channels of the depthwise separable convolution are the same. The feature extraction module is also used to perform a channel shuffling operation on the first feature tensor to obtain a second feature tensor rich in contextual association; The output module is used to generate a fixed-dimensional third feature tensor based on the second feature tensor; the fixed-dimensional feature tensor is used for classification and question answering generation.

10. 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 to 8.

11. 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 to 8.

12. 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 to 8.

Images