Event list classification method and device, program product and electronic equipment

Abstract

The invention discloses an event list classification method and device, a program product and electronic equipment, and relates to the field of financial science and technology or other related fields, and the classification method comprises the steps: obtaining an event list, inputting an event description text into a preset classification model, outputting a target category label, and classifying the event list based on the target category label. And determining a target category to which the event list belongs. According to the method and the device, the technical problem of relatively low efficiency of classifying the event lists in related technologies is solved.

Description

Technical Field

[0001] This invention relates to the field of financial technology, and more specifically, to a method, apparatus, program product, and electronic device for classifying event tickets. Background Technology

[0002] An incident ticket is a manually generated form by branch staff based on customer feedback regarding issues encountered while using banking applications. Incident tickets typically circulate between different branches of the bank via email. Email handlers forward emails based on the incident description in the ticket, ultimately sending the email to the designated functional team responsible for that function. Incident tickets need to be categorized according to the application function they pertain to during this circulation process. Currently, incident tickets are categorized manually. The person handling the email at the time reads the incident description and, based on their subjective judgment, forwards the email to the next person in line, and so on, until it is finally assigned to a developer.

[0003] With the rapid development of fintech, banks are accelerating their technological and digital transformation. Consequently, an increasing number of banking applications are entering the market, and their iteration speed is accelerating. As the number of bank customers grows, the number of event requests related to these applications will also increase. Relying solely on manual methods for event request classification would be extremely resource-intensive. Furthermore, the limitations of human labor and the constraints of human effort can lead to inaccurate and slow event request classification, thus reducing the efficiency of event request processing.

[0004] There is currently no effective solution to the above problems. Summary of the Invention

[0005] This invention provides a method, apparatus, program product, and electronic device for classifying event tickets, in order to at least solve the technical problem of low efficiency in classifying event tickets in related technologies.

[0006] According to one aspect of the present invention, a method for classifying event slips is provided, comprising: acquiring an event slip, wherein the event slip includes a description of a problem reported by a user when using a financial application, and the problem description is recorded as event description text; inputting the event description text into a preset classification model and outputting a target category label, wherein the preset classification model includes at least: a text representation structure, a semantic extraction structure, and a classification structure, wherein the text representation structure is used to extract features from the event description text to obtain a text vector representation, the semantic extraction structure is used to process the text vector representation to obtain temporal fusion features, and the classification structure is used to process the temporal fusion features to obtain a category label; and determining the target category to which the event slip belongs based on the target category label.

[0007] Furthermore, before inputting the event description text into the preset classification model and outputting the target category label, the process includes: constructing a text representation structure based on a language representation model; constructing a semantic extraction structure based on a bidirectional long short-term memory network; and constructing an initial classification model based on the text representation structure, semantic extraction structure, and classification structure. The classification structure includes at least a linear layer and an output layer. The linear layer is used to transform the feature dimension of the temporal fusion features, and the output layer is used to output the category label corresponding to the temporal fusion features after the feature dimension transformation.

[0008] Furthermore, after constructing the initial classification model, the process includes: collecting historical event slips and labeling them based on the event description text contained in each historical event slip to obtain the corresponding label; training the initial classification model using the historical event slips and their corresponding labels; obtaining the category label of each historical event slip output by the initial classification model after each training iteration, and determining the current training fit rate based on the category label and the label; determining that the initial classification model training is complete when the fit rate reaches its maximum, and using the trained model parameters indicated by the maximum fit rate as the target model parameters; and adjusting the initial classification model using the target model parameters to obtain the preset classification model.

[0009] Furthermore, the steps of inputting the event description text into a preset classification model and outputting the target category label include: extracting features from the event description text using a text representation structure to obtain a text vector representation; processing the text vector representation using a semantic extraction structure to obtain temporal fusion features; and processing the temporal fusion features using a classification structure to obtain the target category label.

[0010] Furthermore, the step of extracting features from the event description text using a text representation structure to obtain a text vector representation includes: segmenting the event description text into multiple words; extracting features from each word using a text representation structure to obtain a character vector corresponding to each word; and generating a text vector representation based on all character vectors.

[0011] Furthermore, the semantic extraction structure includes at least a forward long short-term memory (LSTM) unit and a backward long short-term memory (LSTM) unit. The step of processing the text vector representation using the semantic extraction structure to obtain temporal fusion features includes: inputting the text vector representation into the semantic extraction structure; performing forward propagation processing on the text vector representation using the forward LSTM unit to obtain the forward temporal features at the final time step, wherein the input of the forward LSTM unit at the initial time step is the text vector representation, and the output at the initial time step is the input at the next time step; the initial time step is the moment when the semantic extraction structure receives the text vector representation, and the final time step is the moment when the forward LSTM unit completes processing the text vector representation; performing backward propagation processing on the text vector representation using the backward LSTM unit to obtain the backward temporal features at the final time step; and concatenating the forward and backward temporal features to obtain the temporal fusion features.

[0012] Furthermore, the step of processing the temporal fusion features using a classification structure to obtain the target category label includes: transforming the feature dimension of the linear layer temporal fusion features in the classification structure to obtain the temporal fusion features after feature dimension transformation; and normalizing the temporal fusion features after feature dimension transformation using a preset activation function to obtain the target category label.

[0013] According to another aspect of the present invention, an event order classification apparatus is also provided, comprising: an acquisition unit for acquiring an event order, wherein the event order includes a problem description reported by a user when using a financial application, and the problem description is recorded as event description text; an input unit for inputting the event description text into a preset classification model and outputting a target category label, wherein the preset classification model includes at least: a text representation structure, a semantic extraction structure, and a classification structure, wherein the text representation structure is used to extract features from the event description text to obtain a text vector representation, the semantic extraction structure is used to process the text vector representation to obtain temporal fusion features, and the classification structure is used to process the temporal fusion features to obtain a category label; and a determination unit for determining the target category to which the event order belongs based on the target category label.

[0014] Furthermore, the classification device also includes: a first construction module, used to construct a text representation structure based on a language representation model before inputting the event description text into a preset classification model and outputting the target category label; a second construction module, used to construct a semantic extraction structure based on a bidirectional long short-term memory network; and a third construction module, used to construct an initial classification model based on the text representation structure, the semantic extraction structure, and the classification structure, wherein the classification structure includes at least: a linear layer and an output layer, the linear layer is used to transform the feature dimension of the temporal fusion features, and the output layer is used to output the category label corresponding to the temporal fusion features after the feature dimension transformation.

[0015] Furthermore, the classification device also includes: a first acquisition module, used to acquire historical event slips after constructing the initial classification model, and to annotate the historical event slips based on the event description text contained in each historical event slip to obtain the annotation labels corresponding to the historical event slips; a first training module, used to train the initial classification model using the historical event slips and the annotation labels corresponding to each historical event slip; a first acquisition module, used to acquire the category label of each historical event slip output by the initial classification model after each training is completed, and to determine the fitting rate of the current training based on the category label and the annotation label; a first determination module, used to determine that the initial classification model training is complete when the fitting rate reaches its maximum, and to use the trained model parameters indicated by the maximum fitting rate as the target model parameters; and a first adjustment module, used to adjust the initial classification model using the target model parameters to obtain a preset classification model.

[0016] Furthermore, the input unit includes: a first extraction module, used to extract features from the event description text using a text representation structure to obtain a text vector representation; a first processing module, used to process the text vector representation using a semantic extraction structure to obtain temporal fusion features; and a second processing module, used to process the temporal fusion features using a classification structure to obtain a target category label.

[0017] Furthermore, the first extraction module includes: a first processing submodule, used to perform word segmentation on the event description text to obtain multiple word segments; a first extraction submodule, used to extract features from each word segment using a text representation structure to obtain the character vector corresponding to each word segment; and a first generation submodule, used to generate a text vector representation based on all character vectors.

[0018] Furthermore, the semantic extraction structure includes at least: a forward long short-term memory unit and a backward long short-term memory unit. The first processing module includes: a first input submodule, used to input the text vector representation into the semantic extraction structure; a second processing submodule, used to perform forward propagation processing on the text vector representation using the forward long short-term memory unit to obtain the forward temporal features at the last moment, wherein the input of the forward long short-term memory unit at the initial moment is the text vector representation, and the output at the initial moment is the input at the next moment; the initial moment is the moment when the semantic extraction structure receives the text vector representation, and the last moment is the moment when the forward long short-term memory unit completes processing on the text vector representation; a third processing submodule, used to perform backward propagation processing on the text vector representation using the backward long short-term memory unit to obtain the backward temporal features at the last moment; and a first concatenation submodule, used to concatenate the forward temporal features and the backward temporal features to obtain temporal fusion features.

[0019] Furthermore, the second processing module includes: a first transformation submodule, used to perform feature dimension transformation using the linear layer temporal fusion features in the classification structure to obtain the temporal fusion features after feature dimension transformation; and a fourth processing submodule, used to normalize the temporal fusion features after feature dimension transformation using a preset activation function to obtain the target category label.

[0020] According to another aspect of the present invention, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the classification method of any of the above-mentioned event sheets.

[0021] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement any of the above-described event list classification methods.

[0022] In this invention, an event order is obtained, the event description text is input into a preset classification model, a target category label is output, and the target category to which the event order belongs is determined based on the target category label, thereby solving the technical problem of low efficiency in classifying event orders in related technologies.

[0023] In this invention, by inputting the event description text provided by users into a pre-defined classification model, the aim is to automatically extract text features and perform accurate classification, thereby achieving the technical effect of automated event classification and solving the technical problems of slow speed, low efficiency, and inaccurate classification caused by manual classification. Specifically, the text representation structure in the pre-defined classification model uses Roberta-wwm (Robustly Optimized BERT approach with Whole Word Masking, a pre-trained language representation model) to obtain the semantic features of the event description text and transform them into dynamic vector representations; the semantic extraction structure further mines the inherent temporal dependencies of the text through BiLSTM (Bidirectional Long Short-Term Memory network) to form temporal fusion features; the classification structure outputs target category labels based on the above features, thereby automatically determining the financial application subcategory to which the event belongs, such as account services, payment settlement, etc., improving processing efficiency and classification accuracy. Attached Figure Description

[0024] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0025] Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing a classification method for event tickets is shown.

[0026] Figure 2 This is a flowchart of the event form classification method according to Embodiment 1 of the present invention;

[0027] Figure 3 This is a schematic diagram of the classification model architecture according to Embodiment 1 of the present invention;

[0028] Figure 4 This is a schematic diagram of an optional event form classification device according to an embodiment of the present invention;

[0029] Figure 5 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0032] To facilitate understanding of the present invention by those skilled in the art, some terms or nouns involved in the various embodiments of the present invention are explained below:

[0033] The RoBerta-wwm pre-trained model (RoBerta-wwm, Robustly Optimized BERT approach with Whole Word Masking) is a pre-trained language representation model optimized based on the BERT (Bidirectional Encoder Representations from Transformers) model. This model is specifically designed for processing Chinese text. "WWM" is an abbreviation for "Whole Word Masking," referring to the use of a whole word masking strategy during pre-training.

[0034] Text classification: Text classification is a fundamental task in natural language processing, automatically assigning text data to predefined categories. It can be used in scenarios such as sentiment analysis and topic recognition. Text classification can be implemented using various techniques, ranging from simple rule-based and statistical methods to complex machine learning and deep learning models. These models learn how to map text to corresponding categories based on text features through training data.

[0035] It should be noted that all relevant information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data) collected and involved in this invention are 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 this data comply with the relevant laws, regulations, and standards of the relevant regions, necessary confidentiality measures have been taken, and it does not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse. For example, this system has an interface with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface. The relevant information is obtained only after receiving consent from the aforementioned user or organization. If the user chooses to refuse, the process proceeds to an expert decision-making process.

[0036] In this invention, an event-based text classification model is constructed using the Roberta-wwm pre-trained model and a BiLSTM deep learning classifier. This Roberta-wwm-BiLSTM model can automatically classify the sub-application to which an event belongs, based on the specific circumstances of the application. For example, sub-applications of a banking application may include account services, payment settlement, channel management, fund management, and financial markets.

[0037] This invention solves the problems of slow speed, low efficiency and inaccuracy caused by manual classification of event tickets, and achieves the goal of automated classification of event tickets through artificial intelligence.

[0038] The present invention will now be described in detail with reference to various embodiments.

[0039] Example 1

[0040] According to an embodiment of this application, an embodiment of a method for classifying event tickets is also provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0041] The method embodiment provided in Embodiment 1 of this application can be executed on a mobile terminal, computer terminal, or similar computing device. Figure 1 A hardware block diagram of a computer terminal (or mobile device) for implementing a classification method for event tickets is shown. Figure 1 As shown, computer terminal 10 (or mobile device) may include one or more ( Figure 1 The processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions may also be included. In addition, it may include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera, wherein the network interface can be connected to wired and / or wireless networks. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0042] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0043] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the event ticket classification method in this embodiment. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby implementing the aforementioned event ticket classification method. The memory 104 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0044] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0045] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10 (or mobile device).

[0046] Under the aforementioned operating environment, this application provides the following: Figure 2 The event form classification method shown. Figure 2 This is a flowchart of the event form classification method according to Embodiment 1 of the present invention, as shown below. Figure 2 As shown, the method includes the following steps:

[0047] Step S201: Obtain an event form, wherein the event form includes a description of a problem reported by the user when using the financial application, and the problem description is recorded as an event description text.

[0048] In this embodiment of the invention, event forms submitted by bank customers when using financial applications (such as mobile banking applications) can be collected and processed. The event forms include descriptions of problems reported by users when using the financial applications (such as inability to make payments), recording the specific problems encountered by users during operation. These problem descriptions are recorded as event description text, which serves as the basis for subsequent analysis and classification.

[0049] Step S202: Input the event description text into the preset classification model and output the target category label. The preset classification model includes at least: a text representation structure, a semantic extraction structure and a classification structure. The text representation structure is used to extract features from the event description text to obtain a text vector representation. The semantic extraction structure is used to process the text vector representation to obtain temporal fusion features. The classification structure is used to process the temporal fusion features to obtain the category label.

[0050] In this embodiment of the invention, the event description text is input into a pre-trained classification model to obtain the corresponding target category label. This model consists of a text representation structure, a semantic extraction structure, and a classification structure. The text representation structure uses a Roberta-wwm pre-trained model to process the text data. Roberta-wwm is a deep learning model optimized for Chinese, capable of extracting the semantic features of each character in the text through its internal self-attention mechanism and whole-word masking strategy, forming a dynamic vector representation containing contextual information. The semantic extraction structure is a BiLSTM (Bidirectional Long Short-Term Memory) network that receives the vector information output by the text representation structure as input. BiLSTM can process sequence data in both forward and backward directions, thus preserving the temporal information of the text when extracting features from the event description text, fully exploring the hidden temporal correlations and sequence dependencies in the event description, and outputting temporal fusion features. The temporal fusion features, after semantic extraction, are input into the classification structure, which is a fully connected layer. The softmax function (an activation function used to transform the model's raw output into probabilistic class predictions) is then used to calculate the probability distribution, ultimately outputting the target class label. The class label is a predefined label representing the application subcategory to which the event belongs, such as 0, 1, 2, etc.

[0051] Step S203: Determine the target category to which the event order belongs based on the target category label.

[0052] In this embodiment of the invention, the sub-application category to which an event ticket should belong is automatically determined based on the category labels output by the model. For example, category label 0 represents account services, 1 represents payment settlement, and 2 represents channel management, etc. This process automates the classification of event tickets, ensuring that event tickets can be quickly and accurately forwarded to the corresponding functional teams, improving the efficiency and accuracy of event processing, while also reducing the burden of manual classification.

[0053] In summary, by inputting user-reported event description text into a pre-defined classification model, the goal of automatically extracting text features and performing accurate classification is achieved. This realizes the technical effect of automated event classification, thus solving the technical problems of slow speed, low efficiency, and inaccurate classification caused by manual classification. Specifically, the text representation structure in the pre-defined classification model uses Roberta-wwm (Robustly Optimized BERT approach with WholeWord Masking, a pre-trained language representation model) to obtain semantic features of the event description text and transform them into dynamic vector representations. The semantic extraction structure further mines the inherent temporal dependencies of the text through BiLSTM (Bidirectional Long Short-Term Memory network) to form temporal fusion features. The classification structure outputs target category labels based on the above features, thereby automatically determining the financial application subcategory to which the event belongs, such as account services or payment settlement, improving processing efficiency and classification accuracy.

[0054] To improve the accuracy of constructing the initial classification model, in the event list classification method provided in Embodiment 1 of this application, before inputting the event description text into the preset classification model and outputting the target category label, a text representation structure is constructed based on a language representation model; a semantic extraction structure is constructed based on a bidirectional long short-term memory network; and an initial classification model is constructed based on the text representation structure, the semantic extraction structure, and the classification structure. The classification structure includes at least a linear layer and an output layer. The linear layer is used to transform the feature dimension of the temporal fusion features, and the output layer is used to output the category label corresponding to the temporal fusion features after the feature dimension transformation.

[0055] In this embodiment of the invention, addressing the problem that current language models cannot dynamically extract semantic information from text context, this embodiment uses RoBerta-wwm to fully acquire semantic feature information in the text. Since the self-attention mechanism of the RoBerta-wwm model weakens the positional information in the original sequence to some extent, considering the temporal semantic information of the text, the semantic information output by RoBerta-wwm is used as a feature input into the bidirectional LSTM (Long Short-Term Memory) model to mine the temporal association information hidden in the text, obtain more sequence dependency information, and thus complete the construction of the RoBerta-wwm-BiLSTM event single text classification model.

[0056] Specifically, firstly, the text representation structure is constructed using the RoBerta-wwm pre-trained model. RoBerta-wwm is a deep learning model optimized for Chinese, employing a "Whole Word Masking" (WWM) strategy during pre-training to more effectively process the semantic information of Chinese text. In practical applications, after the input event description text is processed by the RoBerta-wwm model, a dynamic vector representation of each character is obtained. These vectors contain rich semantic features and contextual information, forming the text vector representation X={x1, x2, x3, ..., xn}, where each xi (i=1, 2, 3, ..., n) represents the semantic feature vector of a character in the text, and the dimension of X is b. s e, b are the batch size length, s is the length of each text (each line in the event description text is represented as one text), and e is the dimension of the semantic feature vector.

[0057] Then, a semantic extraction structure is constructed using BiLSTM (Bidirectional Long Short-Term Memory). BiLSTM consists of two LSTM networks operating in opposite directions, capable of simultaneously processing both forward and reverse information from sequence data, thus obtaining more comprehensive temporal semantic features. The text vector representation X output from RoBerta-wwm is fed into the BiLSTM as a whole, with xi as the input at each time step t. The BiLSTM processes the semantic features of each event description text through its internal self-loop mechanism, extracting temporal correlation information from the text. Finally, the hidden layer output h1 (temporal feature) of the last time step of the forward LSTM unit and the hidden layer output h1' (temporal feature) of the last time step of the backward LSTM unit are concatenated to obtain the temporal fusion feature. In this way, by considering the temporal order in the text, the semantic representation of event descriptions is further enriched, the dependencies between text sequences are explored, and deeper feature information is provided for classification.

[0058] Next, based on the text representation structure, semantic extraction structure, and classification structure, an initial classification model is constructed. Here, the classification structure consists of a linear layer and an output layer. The linear layer receives the temporal fusion features output by the BiLSTM and transforms the features into a suitable feature representation by setting specific weights and biases. The output layer uses the softmax function to convert the transformed feature representation into a probability distribution of the corresponding category label, ultimately outputting the sub-application category label to which the event belongs. In this way, the complex semantic features of the event description are transformed into specific classification results, achieving automated classification of event orders.

[0059] Figure 3 This is a schematic diagram of the classification model architecture according to Embodiment 1 of the present invention, as shown below. Figure 3 As shown, the classification model architecture includes: RoBERTa-wwm Layer, Forward LSTM (composed of multiple LSTM units), Backward LSTM (composed of multiple LSTM units), Linear Layer, and Output Layer. The input text of this model is the event description text in an event log. For example, in multi-task deep learning research, RoBERTa-wwm is used to obtain word vectors containing semantic information, such as x1, x2, x3, x4, x5, ..., xn, ultimately obtaining the dynamic vector representation of the text X = {x1, x2, x3, ..., xn}. X (Input Vector) is used as the input to both the Forward and Backward LSTMs. The vector information propagates forward within the Forward LSTM, and at each time step, the input is propagated forward within the LSTM to obtain the corresponding output. Simultaneously, the vector information propagates backward within the Backward LSTM, and at each time step, the input is propagated backward within the LSTM to obtain the corresponding output. Then, the hidden layer h1 at the last time step of the forward LSTM unit and the first layer h1' at the last time step of the backward LSTM unit are concatenated to output BiLSTMOutput (including: P1, P2, P3, P4, P5, ..., Pn). Next, the output of the bidirectional LSTM is used as the input to a linear layer to transform the feature dimensions. Finally, the softmax function is applied to output the classification result, thus achieving multi-class classification for multiple events.

[0060] In this embodiment, by combining the semantic representation capabilities of the Roberta-wwm pre-trained model and the temporal information mining capabilities of BiLSTM, as well as the classification capabilities of the linear and output layers, a highly efficient and accurate automated event single classification model is constructed. This model can significantly improve the efficiency of event single processing, reduce the workload of manual classification, and improve the accuracy of classification through deep learning technology, avoiding subjective errors that may occur in manual classification.

[0061] To accurately train the preset classification model, in the event ticket classification method provided in Embodiment 1 of this application, after constructing the initial classification model, historical event tickets are collected, and based on the event description text contained in each historical event ticket, the historical event tickets are labeled to obtain the corresponding label tags; the initial classification model is trained using the historical event tickets and the corresponding label tags; after each training is completed, the category label of each historical event ticket output by the initial classification model is obtained, and the current training fit rate is determined based on the category label and the label tags; when the fit rate reaches the maximum, it is determined that the initial classification model training is complete, and the trained model parameters indicated by the maximum fit rate are used as the target model parameters; the initial classification model is adjusted using the target model parameters to obtain the preset classification model.

[0062] In this embodiment of the invention, firstly, a large number of processed historical event slips are collected. These event slips contain detailed event description text, and are labeled according to their content. Each event slip is assigned one or more sub-application categories as its true label, such as "account services," "payment settlement," and "channel management." Then, the collected historical event slips and their corresponding labels are used to train an initial classification model built on Roberta-wwm and BiLSTM. During training, the model attempts to understand and learn the relationship between the event description text and the classification labels, continuously updating its parameters to improve classification performance. After each training cycle, the model predicts the historical event slips used for training based on the current parameters, outputting the predicted category label. By comparing the predicted label with the true label, the model's prediction accuracy, i.e., the fit rate, can be calculated. The fit rate measures the model's performance on the training data, reflecting the model's classification ability under the current parameter settings. The goal of model training is to find a set of parameter combinations that maximizes the fit rate (i.e., classification accuracy). When the model's classification accuracy on the training data no longer significantly improves, or reaches a predetermined threshold, the model training is considered complete. At this point, record this set of model parameters as the target model parameters. Then, use the determined target model parameters to adjust the initial classification model to obtain the final preset classification model. This model has been sufficiently trained and possesses good event single-item classification capabilities, which can be used for automated processing and classification of new event singles.

[0063] In this embodiment, by training and optimizing an existing historical event ticket dataset, the initial model based on the Roberta-wwm pre-trained model and the BiLSTM deep learning classifier is transformed into a high-precision, high-performance automatic event ticket classification model. Through continuous training and parameter optimization, the model can gradually master the characteristics of event description text, accurately determine the category of the event ticket, thereby improving the efficiency and accuracy of event processing in practical applications, reducing manual intervention, and enhancing the automation level of internal bank processes.

[0064] In order to accurately output the target category label, the event form classification method provided in Embodiment 1 of this application uses a text representation structure to extract features from the event description text to obtain a text vector representation; uses a semantic extraction structure to process the text vector representation to obtain temporal fusion features; and uses a classification structure to process the temporal fusion features to obtain the target category label.

[0065] In this embodiment of the invention, for each event to be classified, its event description text is first input into a text representation structure based on the RoBerta-wwm pre-trained model. The RoBerta-wwm model performs deep semantic analysis on each character or word in the event description text based on its pre-trained weight matrix, generating a dynamic vector representation X={x1, x2, x3, ..., xn} containing rich semantic information. Here, the "text vector representation" is X. By converting the text into a numerical vector, the contextual relevance and semantic depth of the text are preserved, which is an important foundation for subsequent semantic extraction and classification.

[0066] The obtained text vector representation X is then fed into a BiLSTM-based semantic extraction structure. BiLSTM processes X simultaneously from both forward and backward directions, capturing the temporal relationships and potential dependencies between different parts of the text, thus obtaining more comprehensive temporal fusion features. This allows for further extraction of time-series information from the text vector representation, enabling the model to understand the semantic evolution process in event description texts and improve classification accuracy.

[0067] The acquired temporal fusion features are fed into a classification structure, which includes at least a linear layer and an output layer. The linear layer transforms the feature dimensions of the temporal fusion features, adjusting the feature space for easier classification. Subsequently, the output layer uses a softmax function to convert the transformed feature information into a probability distribution, outputting the most likely category label for the event ticket. This is the final step in the model application, directly determining the result of the automated event ticket classification. Through the powerful classification capabilities of deep learning models, the sub-application category to which the event ticket belongs can be quickly and accurately located, such as "account services" or "payment settlement."

[0068] In this embodiment, deep analysis and intelligent classification of event description text are achieved through the combined application of text representation structure, semantic extraction structure, and classification structure. This process fully leverages the advantages of deep learning technology, effectively capturing the semantic features and temporal relationships of text, automating and accurately completing single-classification tasks for events, improving the efficiency and accuracy of internal event processing within the bank, while also reducing reliance on manual operation and enhancing the system's intelligence level.

[0069] In order to accurately obtain the text vector table, in the event form classification method provided in Embodiment 1 of this application, the event description text is segmented into multiple words; the text representation structure is used to extract features from each word to obtain the character vector corresponding to each word; and a text vector representation is generated based on all character vectors.

[0070] In this embodiment of the invention, the event description text in each event form can first be preprocessed using Chinese word segmentation technology, splitting it into a series of meaningful words or symbols. Word segmentation is a fundamental step in natural language processing, capable of dividing continuous text into individual words, facilitating subsequent semantic analysis and feature extraction. In Chinese text processing, since there are no natural separators between Chinese characters and words, word segmentation is particularly important, directly affecting the model's understanding of the text content. Then, the text representation structure of the RoBerta-wwm pre-trained model is used to extract features from each word segment. The RoBerta-wwm model can generate a fixed-length vector representation, i.e., a "character vector," for each word segment. These character vectors not only contain the surface form of the word segment but also its semantic information in the context. Through RoBerta-wwm's Whole Word Masking (WWM) mechanism, the model further enhances its understanding and representation of Chinese words, generating more comprehensive and accurate character vectors. Next, the word vectors of each segment are integrated to generate the vector representation X={x1, x2, x3, ..., xn} of the entire event description text. This process typically employs methods such as weighted averaging, sequence concatenation, or sequence modeling using recurrent networks like LSTM to ensure that the text vector representation X comprehensively reflects the semantic features and contextual information of the event description text, providing sufficient information support for subsequent classification tasks.

[0071] In this embodiment, by converting the event description text into a series of word vectors with rich semantic information, the model can more effectively understand and classify event slips, improving the accuracy and efficiency of automated classification. This process fully leverages the advantages of deep learning technology in the field of natural language processing, effectively solving problems such as low efficiency and high subjectivity in manual classification, and providing a solid foundation for the automation and intelligence of banking technology services.

[0072] Optionally, the semantic extraction structure includes at least a forward long short-term memory (LSTM) unit and a backward long short-term memory (LSTM) unit. To accurately obtain the temporal fusion features, in the event list classification method provided in Embodiment 1 of this application, the text vector representation is input to the semantic extraction structure; the forward LSTM unit is used to perform forward propagation processing on the text vector representation to obtain the forward temporal features at the last moment. The input of the forward LSTM unit at the initial moment is the text vector representation, and the output at the initial moment is the input at the next moment. The initial moment is the moment when the semantic extraction structure receives the text vector representation, and the last moment is the moment when the forward LSTM unit finishes processing the text vector representation. The backward LSTM unit is used to perform backward propagation processing on the text vector representation to obtain the backward temporal features at the last moment. The forward temporal features and the backward temporal features are concatenated to obtain the temporal fusion features.

[0073] In this embodiment of the invention, after the text vector representation is input into the semantic extraction structure, the forward LSTM unit in the BiLSTM first receives the first vector x1 in X as the initial input, and then processes the word vectors of each word segment in the text one by one. At each time point t, the forward LSTM unit generates a hidden state ht, which contains the semantic and temporal information of all words segmented before time t. Finally, after all word vectors have been processed, the forward LSTM unit outputs the hidden state at the last time point, i.e., the forward temporal feature at the last time point, which contains the semantic flow and temporal dependencies of the entire text from left to right. Furthermore, the backward LSTM unit in the BiLSTM starts from the last word segment xn in the text and processes the word vectors of each word segmented in the text progressively backward. Similarly, at each time point t, the backward LSTM unit generates a hidden state ht', which contains the semantic and temporal information of all words segmented after time t. After processing the complete text, the latent state of the first time step output by the backward LSTM unit is considered as the backward temporal feature of the last time step, reflecting the semantic flow and temporal dependencies of the entire text from right to left. After obtaining the forward and backward temporal features, these two features are concatenated to form a larger vector, namely the temporal fusion feature. The temporal fusion feature integrates information from the beginning to the end and from the end to the beginning of the text, enabling the model to consider both the context and temporal relationships of the text simultaneously, thus gaining a more comprehensive understanding of the meaning and details of the event description text.

[0074] In this embodiment, by performing bidirectional propagation processing on the event description text vector representation using forward and backward LSTM units, richer temporal dependencies and semantic associations can be captured from the text. The resulting temporal fusion features will be used as the basis for the final classification decision. This feature extraction method improves the model's ability to understand text complexity, enabling the classification model to more accurately identify the sub-application category to which the event belongs, reducing the possibility of misclassification, and improving the overall classification efficiency and accuracy.

[0075] In order to accurately obtain the target category label, in the event list classification method provided in Embodiment 1 of this application, the linear layer temporal fusion features in the classification structure are used to transform the feature dimension to obtain the temporal fusion features after feature dimension transformation; a preset activation function is used to normalize the temporal fusion features after feature dimension transformation to obtain the target category label.

[0076] In this embodiment of the invention, after the semantic extraction module generates temporal fusion features, these features are fed into a linear layer in the classification structure for feature dimension transformation. The linear layer maps the high-dimensional temporal fusion features to a lower-dimensional feature space through a linear combination of weight matrices and bias vectors, thus simplifying the model's output and making it more suitable for the classifier's input requirements, while also facilitating subsequent activation function processing. After the feature dimension transformation is complete, a preset activation function, such as the Softmax function, normalizes the transformed features. The activation function maps the feature values ​​output by the linear layer to a probability distribution. The Softmax function converts these feature values ​​into probability values ​​for each category, with the sum of all category probabilities being 1. This probability distribution represents the model's prediction of the category to which the event belongs, and the category with the highest probability value is the target category label predicted by the model. The purpose of normalization is to enable the model's output to be directly mapped to category probabilities, facilitating model decision-making and result interpretation.

[0077] In this embodiment, a linear layer and a softmax activation function are used to transform temporal fusion features into single-class event prediction, ensuring that the output class labels have clear probabilistic interpretations and enhancing the model's classification decision-making ability. Feature dimension transformation uses a linear layer to convert complex temporal fusion features into a form easily understood by the classifier, while the normalization process of the activation function gives the classification results probabilistic meaning, making the model output more intuitive and interpretable.

[0078] The event order classification method provided in this application achieves the goal of automatically extracting text features and performing accurate classification by inputting the event description text provided by the user into a preset classification model. This realizes the technical effect of automated event order classification and solves the technical problems of slow speed, low efficiency and inaccurate classification caused by manual classification. Specifically, the text representation structure in the preset classification model uses Roberta-wwm (RobustlyOptimized BERT approach with Whole Word Masking, a pre-trained language representation model) to obtain the semantic features of the event description text and transform them into dynamic vector representations. The semantic extraction structure further mines the inherent temporal dependencies of the text through BiLSTM (Bidirectional Long Short-Term Memory network) to form temporal fusion features. The classification structure outputs target category labels based on the above features, thereby automatically determining the financial application subcategory to which the event order belongs, such as account services, payment settlement, etc., improving processing efficiency and classification accuracy.

[0079] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0080] Example 2

[0081] This application also provides an event ticket classification device. It should be noted that the event ticket classification device of this application can be used to execute the event ticket classification method provided in this application. The event ticket classification device provided in this application will be described below.

[0082] According to an embodiment of this application, an apparatus for implementing the above-described event form classification method is also provided. Figure 4 This is a schematic diagram of an optional event ticket classification device according to an embodiment of the present invention, such as... Figure 4 As shown, the classification device may include: an acquisition unit 40, an input unit 41, and a determination unit 42.

[0083] The acquisition unit 40 is used to acquire event slips, wherein the event slips include a description of a problem reported by the user when using the financial application, and the problem description is recorded as event description text.

[0084] Input unit 41 is used to input event description text into a preset classification model and output target category label. The preset classification model includes at least: text representation structure, semantic extraction structure and classification structure. The text representation structure is used to extract features from the event description text to obtain text vector representation. The semantic extraction structure is used to process the text vector representation to obtain temporal fusion features. The classification structure is used to process the temporal fusion features to obtain category label.

[0085] The determination unit 42 is used to determine the target category to which the event form belongs based on the target category label.

[0086] The event order classification device provided in this application achieves the goal of automatically extracting text features and performing accurate classification by inputting the event description text fed by the user into a preset classification model. This realizes the technical effect of automated event order classification and solves the technical problems of slow speed, low efficiency and inaccurate classification caused by manual classification. Specifically, the text representation structure in the preset classification model uses Roberta-wwm (RobustlyOptimized BERT approach with Whole Word Masking, a pre-trained language representation model) to obtain the semantic features of the event description text and transform them into dynamic vector representations. The semantic extraction structure further mines the inherent temporal dependencies of the text through BiLSTM (Bidirectional Long Short-Term Memory network) to form temporal fusion features. The classification structure outputs target category labels based on the above features, thereby automatically determining the financial application subcategory to which the event order belongs, such as account services, payment settlement, etc., improving processing efficiency and classification accuracy.

[0087] Optionally, the classification device further includes: a first construction module, used to construct a text representation structure based on a language representation model before inputting the event description text into a preset classification model and outputting the target category label; a second construction module, used to construct a semantic extraction structure based on a bidirectional long short-term memory network; and a third construction module, used to construct an initial classification model based on the text representation structure, the semantic extraction structure, and the classification structure, wherein the classification structure includes at least: a linear layer and an output layer, the linear layer being used to transform the feature dimensions of the temporal fusion features, and the output layer being used to output the category label corresponding to the temporal fusion features after the feature dimension transformation.

[0088] Optionally, the classification device further includes: a first acquisition module, used to acquire historical event slips after constructing the initial classification model, and to annotate the historical event slips based on the event description text contained in each historical event slip to obtain the annotation labels corresponding to the historical event slips; a first training module, used to train the initial classification model using the historical event slips and the annotation labels corresponding to each historical event slip; a first acquisition module, used to acquire the category label of each historical event slip output by the initial classification model after each training is completed, and to determine the fitting rate of the current training based on the category label and the annotation label; a first determination module, used to determine that the initial classification model training is complete when the fitting rate reaches its maximum, and to use the trained model parameters indicated by the maximum fitting rate as the target model parameters; and a first adjustment module, used to adjust the initial classification model using the target model parameters to obtain a preset classification model.

[0089] Optionally, the input unit includes: a first extraction module, used to extract features from the event description text using a text representation structure to obtain a text vector representation; a first processing module, used to process the text vector representation using a semantic extraction structure to obtain temporal fusion features; and a second processing module, used to process the temporal fusion features using a classification structure to obtain a target category label.

[0090] Optionally, the first extraction module includes: a first processing submodule, used to perform word segmentation on the event description text to obtain multiple word segments; a first extraction submodule, used to extract features from each word segment using a text representation structure to obtain a character vector corresponding to each word segment; and a first generation submodule, used to generate a text vector representation based on all character vectors.

[0091] Optionally, the semantic extraction structure includes at least: a forward long short-term memory unit and a backward long short-term memory unit. The first processing module includes: a first input submodule, used to input the text vector representation into the semantic extraction structure; a second processing submodule, used to perform forward propagation processing on the text vector representation using the forward long short-term memory unit to obtain the forward temporal features at the last moment, wherein the input of the forward long short-term memory unit at the initial moment is the text vector representation, and the output at the initial moment is the input at the next moment; the initial moment is the moment when the semantic extraction structure receives the text vector representation, and the last moment is the moment when the forward long short-term memory unit completes processing on the text vector representation; a third processing submodule, used to perform backward propagation processing on the text vector representation using the backward long short-term memory unit to obtain the backward temporal features at the last moment; and a first concatenation submodule, used to concatenate the forward temporal features and the backward temporal features to obtain temporal fusion features.

[0092] Optionally, the second processing module includes: a first transformation submodule, used to perform feature dimension transformation using the linear layer temporal fusion features in the classification structure to obtain the temporal fusion features after feature dimension transformation; and a fourth processing submodule, used to normalize the temporal fusion features after feature dimension transformation using a preset activation function to obtain the target category label.

[0093] The classification device described above may also include a processor and a memory. The acquisition unit 40, input unit 41, determination unit 42, etc., are all stored in the memory as program units, and the processor executes the program units stored in the memory to realize the corresponding functions.

[0094] The processor described above contains a kernel, which retrieves the corresponding program unit from memory. One or more kernels can be configured, and by adjusting kernel parameters, the target category to which the event belongs can be determined based on the target category label.

[0095] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0096] It should be noted that the acquisition unit 40, input unit 41, and determination unit 42 mentioned above correspond to steps S201 to S203 in Embodiment 1. The instances and application scenarios implemented by the above units and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above units can be hardware or software components stored in memory (e.g., memory 104) and processed by one or more processors (e.g., processors 102a, 102b, ..., 102n). The above units can also be part of a device and run in the computer terminal 10 provided in Embodiment 1.

[0097] Example 3

[0098] Embodiments of this application may provide an electronic device. Figure 5 This is a structural block diagram of an electronic device according to an embodiment of the present invention. Figure 5 As shown, the electronic device may include: one or more ( Figure 5 (Only one is shown) processor 502, memory 504, memory controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0099] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the event ticket classification method and apparatus in this application embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the aforementioned event ticket classification method. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0100] The processor can access information and applications stored in memory via a transmission device to perform the following steps: acquiring an event ticket, which includes a description of a problem reported by a user while using a financial application, recorded as event description text; inputting the event description text into a preset classification model and outputting a target category label, wherein the preset classification model includes at least: a text representation structure, a semantic extraction structure, and a classification structure; the text representation structure is used to extract features from the event description text to obtain a text vector representation; the semantic extraction structure is used to process the text vector representation to obtain temporal fusion features; and the classification structure is used to process the temporal fusion features to obtain a category label; and determining the target category to which the event ticket belongs based on the target category label.

[0101] The processor can access the information and application programs stored in the memory via the transmission device to perform the following steps: constructing a text representation structure based on a language representation model; constructing a semantic extraction structure based on a bidirectional long short-term memory network; and constructing an initial classification model based on the text representation structure, the semantic extraction structure, and the classification structure. The classification structure includes at least a linear layer and an output layer. The linear layer is used to transform the feature dimensions of the temporal fusion features, and the output layer is used to output the category labels corresponding to the temporal fusion features after the feature dimension transformation.

[0102] The processor can access information and applications stored in the memory via a transmission device to perform the following steps: collect historical event logs and annotate them based on the event description text contained in each historical event log to obtain the corresponding annotation labels; train the initial classification model using the historical event logs and their corresponding annotation labels; after each training iteration, obtain the category label of each historical event log output by the initial classification model, and determine the current training fit rate based on the category label and annotation labels; when the fit rate reaches its maximum, determine that the initial classification model training is complete, and use the trained model parameters indicated by the maximum fit rate as the target model parameters; adjust the initial classification model using the target model parameters to obtain the preset classification model.

[0103] The processor can access the information and application programs stored in the memory via the transmission device to perform the following steps: extracting features from the event description text using a text representation structure to obtain a text vector representation; processing the text vector representation using a semantic extraction structure to obtain temporal fusion features; and processing the temporal fusion features using a classification structure to obtain a target category label.

[0104] The processor can access the information and application programs stored in the memory via the transmission device to perform the following steps: segment the event description text to obtain multiple segments; extract features from each segment using a text representation structure to obtain the character vector corresponding to each segment; and generate a text vector representation based on all character vectors.

[0105] The processor can access information and applications stored in memory via a transmission device to perform the following steps: inputting the text vector representation into the semantic extraction structure; performing forward propagation processing on the text vector representation using a forward long short-term memory unit to obtain the forward temporal features at the final time step, wherein the input of the forward long short-term memory unit at the initial time step is the text vector representation, and the output of the forward long short-term memory unit at the initial time step is the input of the next time step; the initial time step is the time when the semantic extraction structure receives the text vector representation, and the final time step is the time when the forward long short-term memory unit completes processing of the text vector representation; performing backward propagation processing on the text vector representation using a backward long short-term memory unit to obtain the backward temporal features at the final time step; concatenating the forward temporal features and the backward temporal features to obtain the temporal fusion features.

[0106] The processor can access the information and application programs stored in the memory via the transmission device to perform the following steps: transform the feature dimension using the linear layer temporal fusion features in the classification structure to obtain the temporal fusion features after feature dimension transformation; normalize the temporal fusion features after feature dimension transformation using a preset activation function to obtain the target category label.

[0107] This application provides a scheme for classifying event slips. By inputting the event description text provided by the user into a preset classification model, the scheme achieves the goal of automatically extracting text features and performing accurate classification, thereby realizing the technical effect of automated event slip classification and solving the technical problems of slow speed, low efficiency, and inaccurate classification caused by manual classification. Specifically, the text representation structure in the preset classification model uses Roberta-wwm (Robustly Optimized BERT approach with Whole Word Masking, a pre-trained language representation model) to obtain the semantic features of the event description text and transform them into dynamic vector representations; the semantic extraction structure further mines the inherent temporal dependencies of the text through BiLSTM (Bidirectional Long Short-Term Memory network) to form temporal fusion features; the classification structure outputs target category labels based on the above features, thereby automatically determining the financial application subcategory to which the event slip belongs, such as account services, payment settlement, etc., improving processing efficiency and classification accuracy.

[0108] Those skilled in the art will understand that Figure 5 The structure shown is for illustrative purposes only. Electronic devices can also be terminal devices such as smartphones, tablets, PDAs, and mobile internet devices (MIDs). Figure 5 This does not limit the structure of the aforementioned electronic device. For example, electronic devices may also include components that are more... Figure 5 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 5 The different configurations shown.

[0109] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0110] Example 4

[0111] Embodiments of this application also provide a storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the event ticket classification method provided in Embodiment 1.

[0112] Optionally, in this embodiment, the storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0113] This application also provides a computer program product that, when executed on a data processing device, is suitable for performing the steps of a classification method for event sheets.

[0114] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0115] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0116] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0117] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0118] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0119] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0120] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for classifying event tickets, characterized in that, include: Obtain an event ticket, wherein the event ticket includes a description of a problem reported by a user when using a financial application, and the problem description is recorded as an event description text; The event description text is input into a preset classification model, and the target category label is output. The preset classification model includes at least a text representation structure, a semantic extraction structure, and a classification structure. The text representation structure is used to extract features from the event description text to obtain a text vector representation. The semantic extraction structure is used to process the text vector representation to obtain temporal fusion features. The classification structure is used to process the temporal fusion features to obtain a category label. Based on the target category label, the target category to which the event order belongs is determined.

2. The classification method according to claim 1, characterized in that, Before inputting the event description text into the preset classification model and outputting the target category label, the process also includes: The text representation structure is constructed based on the language representation model; The semantic extraction structure is constructed based on a bidirectional long short-term memory network; Based on the text representation structure, the semantic extraction structure, and the classification structure, an initial classification model is constructed. The classification structure includes at least a linear layer and an output layer. The linear layer is used to transform the feature dimension of the temporal fusion features, and the output layer is used to output the category label corresponding to the temporal fusion features after the feature dimension transformation.

3. The classification method according to claim 2, characterized in that, After building the initial classification model, the following is also included: Collect historical event records, and annotate the historical event records based on the event description text contained in each historical event record to obtain the annotation tags corresponding to the historical event records; The initial classification model is trained using the historical event list and the corresponding annotation label for each historical event list; Upon completion of each training iteration, the category label of each historical event output by the initial classification model is obtained, and the fitting rate of the current training is determined based on the category label and the annotation label. When the fitting rate reaches its maximum, it is determined that the initial classification model training is complete, and the trained model parameters indicated by the maximum fitting rate are used as the target model parameters. The initial classification model is adjusted using the target model parameters to obtain the preset classification model.

4. The classification method according to claim 1, characterized in that, The steps of inputting the event description text into a preset classification model and outputting the target category label include: The event description text is used to extract features using the text representation structure to obtain the text vector representation; The semantic extraction structure is used to process the text vector representation to obtain the temporal fusion feature; The temporal fusion features are processed using the classification structure to obtain the target category label.

5. The classification method according to claim 4, characterized in that, The step of extracting features from the event description text using the aforementioned text representation structure to obtain the text vector representation includes: The event description text is segmented into multiple words. The text representation structure is used to extract features from each word segment to obtain the character vector corresponding to each word segment; The text vector representation is generated based on all the aforementioned character vectors.

6. The classification method according to claim 4, characterized in that, The semantic extraction structure includes at least a forward long short-term memory unit and a backward long short-term memory unit. The step of processing the text vector representation using the semantic extraction structure to obtain the temporal fusion features includes: The text vector representation is input into the semantic extraction structure; The text vector representation is processed by the forward propagation unit to obtain the forward temporal features at the last moment. The input of the forward long short-term memory unit at the initial moment is the text vector representation, and the output at the initial moment is the input at the next moment. The initial moment is the moment when the semantic extraction structure receives the text vector representation, and the last moment is the moment when the forward long short-term memory unit completes the processing of the text vector representation. The text vector representation is backpropagated using the backpropagation long short-term memory unit to obtain the backpropagation temporal features at the last moment; The forward temporal features and the backward temporal features are concatenated to obtain the temporal fusion features.

7. The classification method according to claim 4, characterized in that, The step of processing the temporal fusion features using the classification structure to obtain the target category label includes: The temporal fusion features of the linear layer in the classification structure are transformed by feature dimension transformation to obtain the temporal fusion features after feature dimension transformation. The temporal fusion features after feature dimension transformation are normalized using a preset activation function to obtain the target category label.

8. A classification device for event tickets, characterized in that, include: The acquisition unit is used to acquire event tickets, wherein the event tickets include a description of a problem reported by a user when using a financial application, and the problem description is recorded as an event description text; An input unit is used to input the event description text into a preset classification model and output a target category label. The preset classification model includes at least a text representation structure, a semantic extraction structure, and a classification structure. The text representation structure is used to extract features from the event description text to obtain a text vector representation. The semantic extraction structure is used to process the text vector representation to obtain temporal fusion features. The classification structure is used to process the temporal fusion features to obtain a category label. The determining unit is used to determine the target category to which the event order belongs based on the target category label.

9. A computer program product, characterized in that, The method includes a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the event sheet classification method according to any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes one or more processors and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the event list classification method according to any one of claims 1 to 7.

Images