Data processing method and device, nonvolatile storage medium and electronic equipment

Abstract

The invention discloses a data processing method and device, a nonvolatile storage medium and electronic equipment. The method comprises the following steps: receiving input data; searching a preliminary candidate result set corresponding to the input data; performing vectorization processing on the preliminary candidate results in the preliminary candidate result set; sorting vectorization processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base; determining a task type of the first output result; inputting the first output result of which the task type is the simple task into a first preset large language model to obtain a target output result of the first preset large language model, and inputting the first output result of which the task type is the complex task into a second preset large language model to obtain a target output result of the second preset large language model; and displaying the target output result. According to the method and the device, the technical problem of relatively high response delay caused by limited capability of processing a plurality of requests by a large language model at the same time is solved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular, to a data processing method and device, a non-volatile storage medium, and an electronic device. Background Art

[0002] With the development of financial technology, financial institutions are undergoing unprecedented changes. In order to adapt to this change, many financial institutions are actively introducing artificial intelligence technology to improve operational efficiency and service quality. As a product of this trend, the artificial intelligence application platform of financial institutions has become the focus of the industry. The artificial intelligence application platform of financial institutions usually integrates a variety of advanced technologies such as machine learning, natural language processing, and big data analysis. The application of these technologies can not only help financial institutions to automate daily business, but also conduct in-depth mining of massive data through intelligent algorithms, thereby providing customers with a more personalized service experience.

[0003] Taking the application of big models in customer service scenarios as an example, big models can understand and process complex semantic structures, and more accurately parse and understand users' natural language input. Big models can make customer service more empathetic and humane, and effectively manage user emotions. Big models can intelligently generate knowledge questions and answers based on the chat content between agents and customers, improving the efficiency of knowledge base maintenance. The application of big models in the customer service field, with its powerful data processing capabilities and intelligent service model, improves service efficiency and customer satisfaction by answering customer questions 24 hours a day.

[0004] At the same time, the application of big models in the field of customer service still faces the following challenges: big models have limited ability to process multiple requests simultaneously, which can easily cause response delays.

[0005] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0006] The present application provides a data processing method and device, a non-volatile storage medium, and an electronic device to at least solve the technical problem of high response delay caused by the limited ability of a large language model to process multiple requests simultaneously.

[0007] According to one aspect of the present application, a data processing method is provided, including: receiving input data; searching for a preliminary candidate result set corresponding to the input data in a preset professional knowledge base; vectorizing the preliminary candidate results in the preliminary candidate result set; sorting the vectorized processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base; determining a task type of the first output result, wherein the task types include: simple tasks and complex tasks; inputting the first output result of the task type of the simple task into a first preset large language model to obtain a target output result of the first preset large language model, and inputting the first output result of the task type of the complex task into a second preset large language model to obtain a target output result of the second preset large language model, wherein the number of training samples of the first preset large language model is smaller than that of the second preset large language model, and the response speed of the first preset large language model is greater than that of the second preset large language model; and displaying the target output result.

[0008] Optionally, searching for a preliminary candidate result set corresponding to the input data in a preset professional knowledge base includes: extracting target keywords from the input data, searching for a preliminary candidate result set corresponding to the target keywords in the index information in the preset professional knowledge base, wherein the preliminary candidate result set includes: documents and / or paragraphs; vectorizing the preliminary candidate results in the preliminary candidate result set, including: converting the preliminary candidate results in the preliminary candidate result set into a second vector representation to obtain multiple second vector representations; sorting the vectorized processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base, including: converting the input data into a first vector representation, calculating the similarity between the first vector representation and each second vector representation to obtain multiple similarities; determining the comprehensive score corresponding to the candidate result according to whether the source of the candidate result is a preset database, the update time of the candidate result in the preset professional knowledge base, the text length of the candidate result and the similarity; sorting the paragraphs in the target candidate result with the highest comprehensive score in order of the degree of matching with the input data from high to low, to obtain the first output result of the preset professional knowledge base.

[0009] Optionally, determining the task type of the first output result includes: obtaining a keyword library in the financial field, wherein the keyword library includes: financial terms, financial product names, financial operation names, and query keywords related to finance; obtaining a rule library, wherein the rule library includes target rules for judging that the input data is a simple task, and the target rules include: the input data only includes a single financial product name and / or preset query keywords; if the first output result does not conform to the target rules in the rule library, using natural language processing technology to perform semantic analysis on the first output result to determine the semantic intent of the first output result; if the semantic intent points to a single financial product name and / or preset query keywords, determining that the first output result is a simple task; if the semantic intent includes multiple clauses and / or logical connectives, and the semantic intent points to the processing flow of multiple aspects of financial information, determining that the first output result is a complex task.

[0010] Optionally, the first preset large language model is trained by the following method: collecting financial field data and preprocessing the financial field data to generate a first training set and a first validation set; selecting a neural network architecture with a parameter value in the range of M to N as the first model architecture, wherein M ranges from 1 million to 100 million, and N ranges from 500 million to 1 billion; using a random parameter initialization method to assign initial values ​​to the neural network parameters of the first model architecture; inputting the first training set into the first model architecture for training, and the training algorithms used include: stochastic gradient descent method, wherein the initial learning rate of the stochastic gradient descent method is L 1 to L2, L1 ranges from 0.001 to 0.01, and L2 ranges from 0.1 to 0.5. During the training process, the learning rate is adjusted according to the loss function value of the first validation set. When the loss of the first validation set has not decreased for K consecutive training cycles, the learning rate is reduced, and the value range of K is 3 to 10. After each training cycle, the first validation set is used to evaluate the first performance indicator parameter of the first model architecture. If the first performance indicator parameter is less than the first preset threshold, the model architecture parameter is adjusted and the training is performed again. When the first performance indicator parameter is greater than or equal to the first preset threshold, the first preset large language model is obtained.

[0011] Optionally, the second preset large language model is trained by the following method: collecting financial field data and preprocessing the financial field data to generate a second training set and a second validation set; selecting a neural network architecture with a parameter value in the range of P to Q as the first model architecture, wherein M has a value range of 1 million to 100 million, and N has a value range of 500 million to 1 billion; using a random parameter initialization method to assign initial values ​​to the neural network parameters of the second model architecture; inputting the second training set into the second model architecture for training, and the training algorithms used include: stochastic gradient descent method, wherein the initial learning rate of the stochastic gradient descent method is between L3 and L4, L3 ranges from 0.0001 to 0.001, L4 ranges from 0.01 to 0.1, and the learning rate is adjusted according to the loss function value of the second validation set during training. When the loss of the second validation set does not decrease for J consecutive training cycles, the learning rate is reduced, and the value range of J is 5 to 15. After each training cycle, the second validation set is used to evaluate the second performance indicator parameter of the second model architecture. If the second performance indicator parameter is less than the second preset threshold, the model architecture parameter is adjusted and the training is performed again. When the second performance indicator parameter is greater than or equal to the second preset threshold, the second preset large language model is obtained.

[0012] Optionally, the financial field data is preprocessed, including: cleaning the financial field data to remove noise data, duplicate data, erroneous format data and irrelevant non-financial field specific data in the financial field data; and performing text standardization on the cleaned financial field data, wherein the text standardization includes: unifying text format, capitalization specification, punctuation processing and standardized conversion of preset financial professional terms.

[0013] Optionally, the financial field data includes: official document data of financial institutions, customer interaction data and financial market data, among which the official document data includes: business policy documents, product manuals, operating procedure manuals; the customer interaction data includes: customer consultation records, complaint records, suggestion feedback; the financial market data includes: interest rate change information, industry statistical reports.

[0014] According to another aspect of the present application, a data processing device is also provided, including: a receiving module, used to receive input data; a processing module, used to search for a preliminary candidate result set corresponding to the input data in a preset professional knowledge base; vectorizing the preliminary candidate results in the preliminary candidate result set; sorting the vectorized processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base; a determining module, used to determine the task type of the first output result, wherein the task type includes: a simple task and a complex task; an input module, used to input the first output result of the task type being a simple task into a first preset large language model to obtain a target output result of the first preset large language model, and input the first output result of the task type being a complex task into a second preset large language model to obtain a target output result of the second preset large language model, wherein the number of training samples of the first preset large language model is smaller than that of the second preset large language model, and the response speed of the first preset large language model is greater than that of the second preset large language model; and a display module, used to display the target output result.

[0015] According to another aspect of the present application, a non-volatile storage medium is provided, the storage medium includes a stored program, wherein when the program is executed, the device where the storage medium is located is controlled to execute the above data processing method.

[0016] According to another aspect of the present application, an electronic device is provided, including: a memory and a processor, the processor being used to run a program stored in the memory, wherein the above data processing method is executed when the program is run.

[0017] According to yet another aspect of the present application, a computer program is provided, wherein the above data processing method is implemented when the computer program is executed by a processor.

[0018] According to another aspect of the present application, a computer program product is provided, which includes a non-volatile computer-readable storage medium, wherein the non-volatile computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the above data processing method is implemented.

[0019] In the present application, input data is received; a preliminary candidate result set corresponding to the input data is searched in a preset professional knowledge base; the preliminary candidate results in the preliminary candidate result set are vectorized; the vectorized processing results of the preliminary candidate results in the preliminary candidate result set are sorted to obtain a first output result of the preset professional knowledge base; the task type of the first output result is determined, wherein the task types include: simple tasks and complex tasks; the first output result of the task type of the simple task is input into a first preset large language model to obtain a target output result of the first preset large language model, and the first output result of the task type of the complex task is input into a second preset large language model to obtain a target output result of the second preset large language model, wherein the number of training samples of the first preset large language model is less than that of the second preset large language model, and the response speed of the first preset large language model is greater than that of the second preset large language model; the method of displaying the target output result achieves the purpose of improving the ability of the large language model to process multiple requests simultaneously, thereby achieving the technical effect of reducing the response delay, and further solving the technical problem of high response delay caused by the limited ability of the large language model to process multiple requests simultaneously. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0021] Figure 1 is a flow chart of a data processing method according to an embodiment of the present application;

[0022] Figure 2 is a structural diagram of a data processing device according to an embodiment of the present application;

[0023] Figure 3 It is a hardware structure block diagram of a computer terminal according to a data processing method of an embodiment of the present application. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0025] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0026] According to an embodiment of the present application, a method embodiment of a data processing method is provided. It should be noted that the steps shown in the flowchart of 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 can be executed in an order different from that shown here.

[0027] Figure 1 is a flow chart of a data processing method according to an embodiment of the present application, such as Figure 1 As shown, the method comprises the following steps:

[0028] Step S101, receiving input data.

[0029] Users can submit questions or requests through the text input box of the front-end user interface, the voice recognition module, or by uploading files. When a user enters a question, such as "I want to apply for a loan, what documents do I need to prepare?", this input is received as data to be processed.

[0030] Step S102, searching for a preliminary candidate result set corresponding to the input data in a preset professional knowledge base; performing vectorization processing on the preliminary candidate results in the preliminary candidate result set; sorting the vectorization processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base.

[0031] Once the input data is received, the natural language processing (NLP) component will be called to pre-process the input data, including word segmentation, stop word removal, stemming, etc., to extract keywords and semantic information in the question. After the pre-processing is completed, these keywords are matched with the preset professional knowledge base to find the preliminary candidate result set related to the input data. For example, for the above question, search for related terms such as "loan" and "application materials".

[0032] After finding the preliminary candidate result set, use the pre-trained word vector model (such as Word2Vec, GloVe, etc.) to convert these text information into vector representations for more accurate similarity calculations. By calculating the cosine similarity or Euclidean distance between the preliminary candidate results and the input data vector, the preliminary candidate results are sorted, and the results that best match the input data are placed in front to form the first output result.

[0033] Step S103: determining the task type of the first output result, wherein the task type includes: a simple task and a complex task.

[0034] Analyze the first output to determine whether it is a simple task or a complex task. Simple tasks involve answers to frequently asked questions or standardized process guidance, such as checking account balances or understanding basic information about specific products. Complex tasks require deeper semantic understanding, logical reasoning, or cross-domain knowledge integration, such as developing personalized credit plans, analyzing market trends, etc.

[0035] Step S104: input the first output result of the simple task into the first preset large language model to obtain the target output result of the first preset large language model; input the first output result of the complex task into the second preset large language model to obtain the target output result of the second preset large language model, wherein the number of training samples of the first preset large language model is smaller than that of the second preset large language model, and the response speed of the first preset large language model is greater than that of the second preset large language model.

[0036] According to the determined task type, the corresponding large language model is selected for processing. Simple tasks will be input into the first preset large language model, which is relatively small and has less training data, but has a fast response speed and can give answers quickly. Complex tasks will be sent to the second preset large language model, which has more parameters and richer training data, and can generate more sophisticated response results through more complex reasoning and understanding.

[0037] For example, for the simple task of "loan application information", a standardized list of information is directly obtained from the first preset large language model; and for the complex task of "formulating a personalized credit plan for a loan", the second preset large language model will generate more detailed and personalized suggestions based on the user's specific situation and market analysis.

[0038] Step S105, displaying the target output result.

[0039] After the target output result is processed and generated, it is presented to the user through the user interface. For text results, they can be directly displayed in the chat window; for voice results, the text will be converted into voice for playback through text-to-speech technology. In addition, further interactive options such as "Learn More" and "Apply Now" can be provided to help users obtain the required information or complete the business process more conveniently.

[0040] According to the above steps, by adopting the technical solution of integrating the hierarchical model processing with the professional knowledge base, simple and direct inquiries are assigned to smaller and faster-processing models, ensuring instant answers to basic questions and greatly shortening the user waiting time. For complex questions involving multiple considerations and requiring in-depth reasoning, they are processed by models with larger parameter quantities and stronger understanding capabilities. Although the response time may be slightly longer, it can ensure the depth and personalization of the solution, meeting the consultation needs of users in complex scenarios, improving the instant response ability of the service, solving the contradiction between the response speed and service quality when handling banking business consultations, and achieving the technical effects of improving service efficiency, enhancing the user experience, and strengthening decision-making support capabilities.

[0041] The following Figure 1 gives an exemplary illustration and explanation of the steps shown.

[0042] According to some alternative embodiments of the present application, to find the preliminary candidate result set corresponding to the input data in the preset professional knowledge base, it can be achieved by the following method: extract the target keywords in the input data, and search for the preliminary candidate result set corresponding to the target keywords in the index information of the preset professional knowledge base, where the preliminary candidate result set includes: documents and / or paragraphs.

[0043] Use natural language processing technology to preprocess the input, including but not limited to: word segmentation, splitting the text into words or phrases; removing stop words, such as common but meaningless words like "of" and "is"; lemmatization, converting words to their basic forms; and context understanding, parsing the entities and relationships in the question. For example, for the question "How can I query the balance of my personal pension account?", the target keywords "personal pension", "query", and "account balance" will be extracted.

[0044] Based on the extracted target keywords, search the index information in the preset professional knowledge base to find the preliminary candidate result set that matches the keywords. The professional knowledge base stores a large number of documents and paragraphs related to banking business, including product descriptions, operation guides, policies and regulations, and FAQs. First, keyword matching is used to quickly locate relevant documents; then further screening is performed to ensure the accuracy and relevance of the results. For example, for the above target keywords, documents or paragraphs containing the keywords "personal pension", "account balance" and "query" are searched to form a preliminary candidate result set.

[0045] The vectorization processing of the preliminary candidate results in the preliminary candidate result set may be implemented by the following method: converting the preliminary candidate results in the preliminary candidate result set into a second vector representation to obtain a plurality of second vector representations.

[0046] Load pre-trained word vector models, such as Word2Vec, GloVe, or pre-trained large language models based on the Transformer architecture, which can convert words or sentences into numerical vectors for vectorization. The pre-trained models have been learned on a large amount of text data and can capture the semantic relationship and contextual information between words. Then, each word in the preliminary candidate results is input into the pre-trained word vector model to generate the corresponding word vector. The word vector captures the semantic features of the word and provides a basis for subsequent vector fusion. Finally, since the preliminary candidate results usually contain multiple words, these word vectors need to be fused to obtain the vector representation of the entire text. The fusion method can be to average all word vectors, weighted sum, or use more complex neural network models (such as LSTM, GRU, or self-attention mechanism) to comprehensively consider the order and contextual information between words.

[0047] Through the above steps, each document or paragraph in the preliminary candidate results is converted into a vector representation, that is, multiple second vector representations are generated. These second vector representations integrate the semantic features and contextual information of the text, providing a numerical comparison basis for subsequent similarity calculations.

[0048] Sorting the vectorized processing results of the preliminary candidate results in the preliminary candidate result set to obtain the first output result of the preset professional knowledge base can be achieved by the following method: converting the input data into a first vector representation, calculating the similarity between the first vector representation and each second vector representation, and obtaining multiple similarities; determining the comprehensive score corresponding to the candidate result according to whether the source of the candidate result is a preset database, the update time of the candidate result in the preset professional knowledge base, the text length of the candidate result and the similarity; sorting the paragraphs in the target candidate results with the highest comprehensive score in order of the degree of matching with the input data from high to low, and obtaining the first output result of the preset professional knowledge base.

[0049] The similarity calculation may be performed using a variety of methods, including but not limited to cosine similarity, Euclidean distance, or Manhattan distance, etc. A similarity score is calculated for each second vector representation, thereby obtaining multiple similarity scores, which reflect the semantic matching degree between the preliminary candidate results and the user input.

[0050] Then, the comprehensive score of each candidate result is determined based on the source, update time and text length of the candidate results, combined with the similarity score calculated above. The evaluation criteria include: if the candidate results come from a preset database, the authority and accuracy of this information are usually higher, and a higher score will be given; newer information will score higher to ensure that the information received by the user is the latest; the text length is standardized to avoid bias in the score caused by too long or too short text, thereby ensuring the fairness of the comprehensive score.

[0051] Finally, after calculating the comprehensive scores of all preliminary candidate results, these scores will be sorted according to the comprehensive scores to ensure that the most relevant and latest information is at the top. The paragraphs in the target candidate results will be sorted in the order of the comprehensive scores to obtain the first output result of the preset professional knowledge base. This output result will be passed to subsequent processing links, such as in-depth analysis of large models or directly presented to users, to ensure that users can obtain the most accurate and fastest consulting service experience.

[0052] According to some other optional embodiments of the present application, determining the task type of the first output result can be achieved by the following method: obtaining a keyword library in the financial field, wherein the keyword library includes: financial terms, financial product names, financial operation names, and query keywords related to finance; obtaining a rule library, wherein the rule library includes target rules for judging whether the input data is a simple task, and the target rules include: the input data only includes a single financial product name and / or preset query keywords; if the first output result does not conform to the target rules in the rule library, using natural language processing technology to perform semantic analysis on the first output result to determine the semantic intent of the first output result; if the semantic intent points to a single financial product name and / or preset query keywords, determining that the first output result is a simple task; if the semantic intent includes multiple clauses and / or logical connectives, and the semantic intent points to the processing flow of multiple aspects of financial information, determining that the first output result is a complex task.

[0053] A keyword library in the financial field is loaded, including but not limited to financial terms (such as "fixed deposit", "loan interest rate"), financial product names, financial operation names (such as "report loss"), and financial-related query keywords (such as "balance", "interest rate"). Subsequently, a preset rule library is loaded, including a target rule for judging that the input data is a simple task. For example, if the input data only contains "loan" or "check balance", the task is preliminarily judged to be a simple task. The first output result is analyzed to check whether it meets the target rules in the rule library. If the content involved in the first output result meets the pattern of a single financial product name or a preset query keyword, it is preliminarily judged to be a simple task; otherwise, its task type needs to be further analyzed. For the first output result that does not meet the preliminary simple task judgment, the natural language processing technology will be used to perform in-depth semantic analysis on the output result to identify the semantic intent, and then check whether the output result contains multiple clauses, logical conjunctions (such as "and", "or"), and processing flows pointing to multiple aspects of financial information. These features are usually associated with complex tasks. Based on the results of the semantic analysis, the task type of the first output result is further determined. If the semantic intent points to a single financial product name and / or preset query keywords, that is, the output result contains only one clear query or operation requirement, it will be determined as a simple task; on the contrary, if the semantic intent contains multiple clauses, logical connectives and processing flows pointing to multiple aspects of financial information, that is, the output result involves complex requirements such as compound queries, product comparisons or process guidance, it will be judged as a complex task.

[0054] In some optional embodiments of the present application, the first preset large language model is trained by the following method: collecting financial field data and preprocessing the financial field data to generate a first training set and a first validation set; selecting a neural network architecture with a parameter value in the range of M to N as the first model architecture, wherein M ranges from 1 million to 100 million, and N ranges from 500 million to 1 billion; using a random parameter initialization method to assign initial values ​​to the neural network parameters of the first model architecture; inputting the first training set into the first model architecture for training, and the training algorithms used include: a stochastic gradient descent method, wherein the initialization of the stochastic gradient descent method The learning rate is between L1 and L2, the L1 value range is 0.001 to 0.01, and the L2 value range is 0.1 to 0.5. During the training process, the learning rate is adjusted according to the loss function value of the first validation set. When the loss of the first validation set does not decrease for K consecutive training cycles, the learning rate is reduced, and the value range of K is 3 to 10. After each training cycle, the first validation set is used to evaluate the first performance indicator parameter of the first model architecture. If the first performance indicator parameter is less than the first preset threshold, the model architecture parameter is adjusted and the training is performed again. When the first performance indicator parameter is greater than or equal to the first preset threshold, the first preset large language model is obtained.

[0055] Financial data is collected from multiple sources, including but not limited to official bank policy documents, product descriptions, customer consultation records, transaction data, financial market dynamics, etc. The collected data must cover a wide range of financial business scenarios to ensure the comprehensiveness and effectiveness of model training. Next, the collected data is preprocessed, including text cleaning, standardization (such as unified capitalization, removal of special symbols), word segmentation, removal of stop words, etc., to remove noise, improve data quality and model training efficiency. The preprocessed data is further divided into the first training set and the first validation set. The proportion distribution must follow statistical principles to ensure the representativeness of the data. The neural network architecture is selected as the first model architecture, and the number of parameters must be in the range of M to N, with M ranging from 1 million to 100 million and N ranging from 500 million to 1 billion. The selection of this range is intended to balance the complexity of the model with the training efficiency, ensuring that the model has sufficient representation capabilities to understand the complex semantics in the financial field while avoiding unnecessary waste of computing resources. The architecture adopts a multi-layer Transformer structure, including encoders and decoders, to process sequence data and capture long-distance dependencies. In order to avoid the model training from falling into the local optimum, ensure that the model can learn from different starting points, and improve the generalization ability, the random parameter initialization method is used to assign initial values ​​to the neural network parameters of the selected first model architecture. The preprocessed first training set data is input into the first model architecture for training, and the stochastic gradient descent method (SGD) is used as the optimization algorithm. The initial learning rate is set between L1 and L2, with L1 values ​​ranging from 0.001 to 0.01 and L2 values ​​ranging from 0.1 to 0.5. During the training process, the first validation set data is regularly used to evaluate the model performance. The performance indicators include accuracy, recall rate, F1 value, etc. At the same time, the loss function value of the first validation set is monitored. If the validation set loss does not decrease within K consecutive training cycles (K values ​​range from 3 to 10), the learning rate will be automatically reduced to avoid model overfitting and promote training convergence. At the end of each training cycle, the first performance indicator parameter of the model is calculated using the first validation set data and compared with the preset first performance threshold: If the first performance indicator parameter is less than the first preset threshold, the model architecture parameters will be adjusted, such as increasing or decreasing the number of neural network layers, adjusting the number of hidden layer nodes, optimizing the loss function, etc., and then restarting the training. This process will continue to iterate until the model performance meets the preset standards. When the first performance indicator parameter reaches or exceeds the first preset threshold, the training is stopped, and the model at this time is the first preset large language model. This model has efficient understanding and generation capabilities when dealing with complex problems in the financial field, and can provide accurate and comprehensive answers and suggestions, effectively improving the service quality and user experience of bank intelligent customer service.

[0056] As some optional embodiments of the present application, the second preset large language model is trained by the following method: collect financial field data, preprocess the financial field data, generate a second training set and a second validation set; select a neural network architecture with a parameter value in the range of P to Q as the first model architecture, wherein M ranges from 1 million to 100 million, and N ranges from 500 million to 1 billion; use a random parameter initialization method to assign initial values ​​to the neural network parameters of the second model architecture; input the second training set into the second model architecture for training, and the training algorithms used include: stochastic gradient descent method, wherein the initial learning of the stochastic gradient descent method The learning rate is between L3 and L4, the L3 value range is 0.0001 to 0.001, and the L4 value range is 0.01 to 0.1. During the training process, the learning rate is adjusted according to the loss function value of the second validation set. When the loss of the second validation set does not decrease for J consecutive training cycles, the learning rate is reduced, and the value range of J is 5 to 15. After each training cycle, the second validation set is used to evaluate the second performance indicator parameter of the second model architecture. If the second performance indicator parameter is less than the second preset threshold, the model architecture parameter is adjusted and the training is performed again. When the second performance indicator parameter is greater than or equal to the second preset threshold, the second preset large language model is obtained.

[0057] A neural network architecture with a parameter range of P to Q is selected as the second model architecture, where P ranges from 1 million to 100 million and Q ranges from 500 million to 1 billion. This range is selected to build a large language model that can understand complex semantics in the financial field and can be trained within a reasonable time frame. The architecture can be based on the Transformer structure, with multiple layers of encoders and decoders, which can effectively process sequence data and capture long-distance dependencies. A random parameter initialization method, including Gaussian distribution, uniform distribution or other random distribution, is used to ensure the randomness of the parameters and the initial diversity of the model, and to assign initial values ​​to the neural network parameters of the second model architecture. This breaks the symmetry of model training, promotes parameter diversification, and avoids training from falling into local optimality. The preprocessed second training set data is input into the second model architecture for training. The training algorithm uses the stochastic gradient descent method, and the initial learning rate is set between L3 and L4, with L3 taking a value of 0.0001 to 0.001 and L4 taking a value of 0.01 to 0.1. During the training process, the learning rate is dynamically adjusted according to the loss function value of the second validation set. When the validation set loss does not decrease within J consecutive training cycles (the value range of J is 5 to 15), the learning rate is automatically reduced to prevent model overfitting and promote the convergence of the training process. At the end of each training cycle, the second validation set data is used to evaluate the second performance indicator parameters of the second model architecture. The performance indicators include accuracy, recall, F1 value, etc. If the second performance indicator parameter is less than the second preset threshold (that is, the model performance does not meet the expected standard), the model architecture parameters will be adjusted, such as increasing the number of neural network layers, adjusting the number of hidden layer nodes, optimizing the loss function, etc., and then the training process will be restarted until the model performance is improved. When the second performance indicator parameter reaches or exceeds the second preset threshold, the model performance meets the demand standards of intelligent customer service in the financial field, and the model at this time is the second preset large language model. This indicates that the model training is completed and can handle complex and changeable financial consulting problems and provide high-quality automatic services. The second preset large language model will play a core role in subsequent intelligent customer service applications and provide accurate and fast consulting answers.

[0058] In some optional embodiments of the present application, preprocessing of financial field data can be achieved by the following methods: performing a cleaning operation on the financial field data to remove noise data, duplicate data, erroneous format data, and irrelevant non-financial field-specific data in the financial field data; performing text standardization processing on the cleaned financial field data, wherein the text standardization processing includes: unifying text format, capitalization specification, punctuation processing, and standardized conversion of preset financial professional terms.

[0059] Preferably, the financial field data includes: official document data of financial institutions, customer interaction data and financial market data, among which the official document data includes: business policy documents, product manuals, and operating procedure manuals; the customer interaction data includes: customer consultation records, complaint records, and suggestion feedback; the financial market data includes: interest rate change information and industry statistical reports.

[0060] Collect financial data from multiple channels, including but not limited to official documents of financial institutions (such as business policy documents, product manuals, and operating procedure manuals), customer interaction records (such as customer consultation records, complaint records, and suggestion feedback), and financial market dynamics data (such as interest rate changes and industry statistical reports), covering a wide range of financial business scenarios and user needs, providing a rich source of information for model training. Perform preliminary cleaning on the collected financial data to remove noise data (such as advertisements, irrelevant links, etc.), duplicate data (to prevent bias in the training data set), and erroneous data in the format (such as damaged PDF files, incorrectly encoded text, etc.). At the same time, identify and exclude non-financial field-specific data to ensure the purity and relevance of the data set.

[0061] The cleaned financial data is subjected to text standardization. The specific steps include: ensuring that all text data is in a consistent format, such as unifying it into a plain text format, removing complex typesetting and style information, so that the model can understand and process it easily; converting the text into a unified case format, such as converting all to lowercase, to avoid information comprehension bias caused by case differences; standardizing the use of punctuation marks, such as standardizing punctuation marks such as quotation marks and commas in different language environments, to ensure that the model can correctly parse the text structure; establishing a standardized list of financial professional terms, and uniformly converting the terms that appear in the data to ensure that the model learns consistent terminology representations and improves comprehension capabilities. For example, "personal retirement funds" and "personal pensions" are uniformly converted to the standardized term "personal pension".

[0062] For official document data, such as business policy documents, special attention is paid to format standardization and terminology consistency. Automated tools are used to identify and extract key information in the documents, such as policy details, product features, and operating procedures, and convert them into a unified text format and terminology representation to facilitate model learning and understanding. For customer interaction data, such as complaint records, emotional color and problem type are analyzed, and different problems and feedback are classified and labeled to facilitate model learning of processing methods in different situations. Sentiment analysis technology can also be used to label the emotional tendency of the data to help the model understand user emotions and provide more humane services. For financial market data, such as industry statistical reports, focus on the timeliness and accuracy of the data. Market data is regularly updated through automated data crawling tools, and real-time checks are performed to ensure that the information learned by the model is the latest. In addition, text summarization technology can be used to extract key information from the report to reduce the computational burden of model training.

[0063] Figure 2 is a structural diagram of a data processing device according to an embodiment of the present application, such as Figure 2 As shown, the device comprises:

[0064] The receiving module 21 is used to receive input data.

[0065] The processing module 22 is used to search for a preliminary candidate result set corresponding to the input data in a preset professional knowledge base; perform vectorization processing on the preliminary candidate results in the preliminary candidate result set; sort the vectorization processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base.

[0066] The determination module 23 is used to determine the task type of the first output result, wherein the task type includes: a simple task and a complex task.

[0067] The input module 24 is used to input the first output result of the task type being a simple task into the first preset large language model to obtain the target output result of the first preset large language model, and input the first output result of the task type being a complex task into the second preset large language model to obtain the target output result of the second preset large language model, wherein the number of training samples of the first preset large language model is smaller than that of the second preset large language model, and the response speed of the first preset large language model is greater than that of the second preset large language model.

[0068] The display module 25 is used to display the target output result.

[0069] Optionally, the processing module 22 is also used to perform the following steps: extract the target keyword in the input data, and search for a preliminary candidate result set corresponding to the target keyword in the index information in the preset professional knowledge base, wherein the preliminary candidate result set includes: documents and / or paragraphs; vectorize the preliminary candidate results in the preliminary candidate result set, including: converting the preliminary candidate results in the preliminary candidate result set into a second vector representation to obtain multiple second vector representations; sorting the vectorized processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base, including: converting the input data into a first vector representation, calculating the similarity between the first vector representation and each second vector representation to obtain multiple similarities; determining the comprehensive score corresponding to the candidate result according to whether the source of the candidate result is a preset database, the update time of the candidate result in the preset professional knowledge base, the text length of the candidate result and the similarity; sorting the paragraphs in the target candidate result with the highest comprehensive score in order of the degree of matching with the input data from high to low, to obtain the first output result of the preset professional knowledge base.

[0070] Optionally, the determination module 23 is also used to perform the following steps: obtaining a keyword library in the financial field, wherein the keyword library includes: financial terms, financial product names, financial operation names, and query keywords related to finance; obtaining a rule library, wherein the rule library includes target rules for determining that the input data is a simple task, and the target rules include: the input data only includes a single financial product name and / or preset query keywords; if the first output result does not conform to the target rules in the rule library, using natural language processing technology to perform semantic analysis on the first output result to determine the semantic intent of the first output result; if the semantic intent points to a single financial product name and / or preset query keywords, determining that the first output result is a simple task; if the semantic intent includes multiple clauses and / or logical connectives, and the semantic intent points to a processing flow of multiple aspects of financial information, determining that the first output result is a complex task.

[0071] Optionally, the first preset large language model is trained by the following method: collecting financial field data and preprocessing the financial field data to generate a first training set and a first validation set; selecting a neural network architecture with a parameter value in the range of M to N as the first model architecture, wherein M ranges from 1 million to 100 million, and N ranges from 500 million to 1 billion; using a random parameter initialization method to assign initial values ​​to the neural network parameters of the first model architecture; inputting the first training set into the first model architecture for training, and the training algorithms used include: stochastic gradient descent method, wherein the initial learning rate of the stochastic gradient descent method is L 1 to L2, L1 ranges from 0.001 to 0.01, and L2 ranges from 0.1 to 0.5. During the training process, the learning rate is adjusted according to the loss function value of the first validation set. When the loss of the first validation set has not decreased for K consecutive training cycles, the learning rate is reduced, and the value range of K is 3 to 10. After each training cycle, the first validation set is used to evaluate the first performance indicator parameter of the first model architecture. If the first performance indicator parameter is less than the first preset threshold, the model architecture parameter is adjusted and the training is performed again. When the first performance indicator parameter is greater than or equal to the first preset threshold, the first preset large language model is obtained.

[0072] Optionally, the second preset large language model is trained by the following method: collecting financial field data and preprocessing the financial field data to generate a second training set and a second validation set; selecting a neural network architecture with a parameter value in the range of P to Q as the first model architecture, wherein M has a value range of 1 million to 100 million, and N has a value range of 500 million to 1 billion; using a random parameter initialization method to assign initial values ​​to the neural network parameters of the second model architecture; inputting the second training set into the second model architecture for training, and the training algorithms used include: stochastic gradient descent method, wherein the initial learning rate of the stochastic gradient descent method is between L3 and L4, L3 ranges from 0.0001 to 0.001, L4 ranges from 0.01 to 0.1, and the learning rate is adjusted according to the loss function value of the second validation set during training. When the loss of the second validation set does not decrease for J consecutive training cycles, the learning rate is reduced, and the value range of J is 5 to 15. After each training cycle, the second validation set is used to evaluate the second performance indicator parameter of the second model architecture. If the second performance indicator parameter is less than the second preset threshold, the model architecture parameter is adjusted and the training is performed again. When the second performance indicator parameter is greater than or equal to the second preset threshold, the second preset large language model is obtained.

[0073] Optionally, the data processing device is also used to execute the following steps: perform cleaning operations on financial field data to remove noise data, duplicate data, format error data and irrelevant non-financial field specific data in the financial field data; perform text standardization processing on the cleaned financial field data, wherein the text standardization processing includes: unified text format, capitalization standardization, punctuation processing and standardized conversion of preset financial professional terms.

[0074] Optionally, the financial field data includes: official document data of financial institutions, customer interaction data and financial market data, among which the official document data includes: business policy documents, product manuals, operating procedure manuals; the customer interaction data includes: customer consultation records, complaint records, suggestion feedback; the financial market data includes: interest rate change information, industry statistical reports.

[0075] It should be noted that the above Figure 2 Each module in the can be a program module (for example, a set of program instructions that implement a specific function) or a hardware module. For the latter, it can be expressed in the following forms, but is not limited to this: the expression form of each of the above modules is a processor, or the functions of each of the above modules are implemented by a processor.

[0076] It should be noted that Figure 2 The preferred implementation of the illustrated embodiment can be found in Figure 1 The relevant description of the illustrated embodiment will not be repeated here.

[0077] Figure 3 FIG. 1 shows a hardware structure block diagram of a computer terminal for implementing a data processing method. Figure 3 As shown, the computer terminal 30 may include one or more (302a, 302b, ..., 302n are used to illustrate) processors 302 (the processor 302 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 304 for storing data, and a transmission module 306 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the BUS bus), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Figure 3 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 3 More or fewer components as shown, or with Figure 3 Different configurations shown.

[0078] It should be noted that the one or more processors 302 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 30. As involved in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0079] The memory 304 can be used to store software programs and modules of application software, such as program instructions / data storage devices corresponding to the data processing method in the embodiment of the present application. The processor 302 executes various functional applications and data processing by running the software programs and modules stored in the memory 304, that is, realizing the above-mentioned data processing method. The memory 304 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 304 may further include a memory remotely arranged relative to the processor 302, and these remote memories may be connected to the computer terminal 30 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0080] The transmission module 306 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer terminal 30. In one example, the transmission module 306 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission module 306 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0081] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 30 .

[0082] It should be noted that, in some optional embodiments, the above Figure 3 The computer terminal shown may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of both hardware elements and software elements. Figure 3 This is merely one example of a particular embodiment and is intended to illustrate the types of components that may be present in the computer terminal described above.

[0083] It should be noted that Figure 3 The computer terminal shown is used to execute Figure 1 The data processing method shown, therefore the relevant explanations in the execution method of the above command are also applicable to the electronic device and will not be repeated here.

[0084] An embodiment of the present application further provides a non-volatile storage medium, which includes a stored program, wherein when the program is executed, the device where the storage medium is located is controlled to execute the above data processing method.

[0085] A program for a non-volatile storage medium to perform the following functions: receiving input data; searching for a preliminary candidate result set corresponding to the input data in a preset professional knowledge base; vectorizing the preliminary candidate results in the preliminary candidate result set; sorting the vectorized processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base; determining a task type of the first output result, wherein the task types include: simple tasks and complex tasks; inputting the first output result of the task type being a simple task into a first preset large language model to obtain a target output result of the first preset large language model, and inputting the first output result of the task type being a complex task into a second preset large language model to obtain a target output result of the second preset large language model, wherein the number of training samples of the first preset large language model is less than that of the second preset large language model, and the response speed of the first preset large language model is greater than that of the second preset large language model; and displaying the target output result.

[0086] An embodiment of the present application further provides an electronic device, including: a memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the above data processing method is executed when the program is run.

[0087] The processor is used to run a program that performs the following functions: receiving input data; searching for a preliminary candidate result set corresponding to the input data in a preset professional knowledge base; performing vectorization processing on the preliminary candidate results in the preliminary candidate result set; sorting the vectorization processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base; determining a task type of the first output result, wherein the task types include: simple tasks and complex tasks; inputting the first output result of the task type being a simple task into a first preset large language model to obtain a target output result of the first preset large language model, and inputting the first output result of the task type being a complex task into a second preset large language model to obtain a target output result of the second preset large language model, wherein the number of training samples of the first preset large language model is less than that of the second preset large language model, and the response speed of the first preset large language model is greater than that of the second preset large language model; and displaying the target output result.

[0088] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0089] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0090] In the above-mentioned embodiments of the present application, the collected information is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with relevant laws, regulations and standards, take necessary protection measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0091] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0092] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0093] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0094] If the integrated unit is implemented in the form of 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 the present application, or the part that contributes to the relevant technology or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, server or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk, etc. Various media that can store program codes.

[0095] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A data processing method, characterized in that: include: Receive input data; Searching for a preliminary candidate result set corresponding to the input data in a preset professional knowledge base; Performing vectorization processing on the preliminary candidate results in the preliminary candidate result set; Sorting the vectorized processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base; Determine a task type of the first output result, wherein the task type includes: a simple task and a complex task; Inputting the first output result of the task type being a simple task into a first preset large language model to obtain a target output result of the first preset large language model, and inputting the first output result of the task type being a complex task into a second preset large language model to obtain a target output result of the second preset large language model, wherein the number of training samples of the first preset large language model is smaller than that of the second preset large language model, and the response speed of the first preset large language model is greater than that of the second preset large language model; Display the target output result.

2. The method according to claim 1, characterized in that: Searching a preset professional knowledge base for a preliminary candidate result set corresponding to the input data includes: Extracting a target keyword from the input data, and searching for a preliminary candidate result set corresponding to the target keyword in the index information in the preset professional knowledge base, wherein the preliminary candidate result set includes: documents and / or paragraphs; The preliminary candidate results in the preliminary candidate result set are vectorized, including: Converting the preliminary candidate results in the preliminary candidate result set into a second vector representation to obtain a plurality of second vector representations; Sorting the vectorized processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base includes: Convert the input data into a first vector representation, and calculate the similarity between the first vector representation and each of the second vector representations to obtain multiple similarities; Determine a comprehensive score corresponding to the candidate result according to whether the source of the candidate result is a preset database, the update time of the candidate result in the preset professional knowledge base, the text length of the candidate result, and the similarity; The paragraphs in the target candidate results with the highest comprehensive scores are sorted in descending order of the degree of matching with the input data to obtain a first output result of the preset professional knowledge base.

3. The method according to claim 1, characterized in that Determining the task type of the first output result includes: Acquire a keyword library in the financial field, wherein the keyword library includes: financial terms, financial product names, financial operation names, and query keywords related to finance; Obtaining a rule base, wherein the rule base includes a target rule for determining that the input data is a simple task, wherein the target rule includes: the input data only includes a single financial product name and / or a preset query keyword; If the first output result does not conform to the target rule in the rule base, performing semantic analysis on the first output result using natural language processing technology to determine the semantic intent of the first output result; If the semantic intent points to a single financial product name and / or a preset query keyword, determining that the first output result is a simple task; If the semantic intent includes multiple clauses and / or logical connectives, and the semantic intent points to a processing flow of multiple aspects of financial information, it is determined that the first output result is a complex task.

4. The method according to claim 1, characterized in that The first preset large language model is trained by the following method: Collect financial field data, and pre-process the financial field data to generate a first training set and a first validation set; A neural network architecture with a parameter value in the range of M to N is selected as the first model architecture, wherein M ranges from 1 million to 100 million, and N ranges from 500 million to 1 billion; Using random parameter initialization to assign initial values ​​to the neural network parameters of the first model architecture; Inputting the first training set into the first model architecture for training, the adopted training algorithm includes: stochastic gradient descent method, wherein the initial learning rate of the stochastic gradient descent method is between L1 and L2, L1 has a value range of 0.001 to 0.01, and L2 has a value range of 0.1 to 0.

5. During the training process, the learning rate is adjusted according to the loss function value of the first validation set. When the loss of the first validation set does not decrease for K consecutive training cycles, the learning rate is reduced, and the value range of K is 3 to 10; After each training cycle, the first validation set is used to evaluate a first performance indicator parameter of the first model architecture, and if the first performance indicator parameter is less than a first preset threshold, the model architecture parameter is adjusted and training is performed again; When the first performance indicator parameter is greater than or equal to the first preset threshold, the first preset large language model is obtained.

5. The method according to claim 1, characterized in that The second preset large language model is trained by the following method: Collecting financial field data, and preprocessing the financial field data to generate a second training set and a second validation set; A neural network architecture with a parameter value in the range of P to Q is selected as the first model architecture, wherein M ranges from 1 million to 100 million, and N ranges from 500 million to 1 billion; Using random parameter initialization method, the neural network parameters of the second model architecture are given initial values; The second training set is input into the second model architecture for training, and the adopted training algorithm includes: Stochastic gradient descent method, wherein the initial learning rate of the stochastic gradient descent method is between L3 and L4, the value range of L3 is 0.0001 to 0.001, and the value range of L4 is 0.01 to 0.

1. During the training process, the learning rate is adjusted according to the loss function value of the second validation set. When the loss of the second validation set does not decrease for J consecutive training cycles, the learning rate is reduced, and the value range of J is 5 to 15; After each training cycle, the second performance indicator parameter of the second model architecture is evaluated using the second validation set, and if the second performance indicator parameter is less than a second preset threshold, the model architecture parameter is adjusted and training is performed again; When the second performance indicator parameter is greater than or equal to the second preset threshold, the second preset large language model is obtained.

6. The method according to claim 4, characterized in that Preprocessing the financial data includes: Performing a cleaning operation on the financial field data to remove noise data, duplicate data, data with incorrect formats, and irrelevant non-financial field specific data in the financial field data; The cleaned financial data is subjected to text standardization processing, wherein the text standardization processing includes: unifying text format, capitalization standardization, punctuation processing, and standardized conversion of preset financial professional terms.

7. The method according to claim 6, characterized in that The financial field data includes: official document data of financial institutions, customer interaction data and financial market data, among which the official document data includes: business policy documents, product manuals, and operating procedure manuals; the customer interaction data includes: customer consultation records, complaint records, and suggestion feedback; the financial market data includes: interest rate change information and industry statistical reports.

8. A data processing device, characterized in that: include: A receiving module, used for receiving input data; A processing module, used for searching a preset professional knowledge base for a preliminary candidate result set corresponding to the input data; Performing vectorization processing on the preliminary candidate results in the preliminary candidate result set; Sorting the vectorized processing results of the preliminary candidate results in the preliminary candidate result set to obtain a first output result of the preset professional knowledge base; A determination module, used to determine the task type of the first output result, wherein the task type includes: a simple task and a complex task; an input module, configured to input the first output result of the task type being a simple task into a first preset large language model to obtain a target output result of the first preset large language model, and input the first output result of the task type being a complex task into a second preset large language model to obtain a target output result of the second preset large language model, wherein the number of training samples of the first preset large language model is smaller than that of the second preset large language model, and the response speed of the first preset large language model is greater than that of the second preset large language model; A display module is used to display the target output result.

9. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored program, wherein when the program is executed, the device where the non-volatile storage medium is located is controlled to execute the data processing method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: include: A memory and a processor, wherein the processor is used to run a program stored in the memory, wherein the program executes the data processing method according to any one of claims 1 to 7 when running.

11. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the data processing method according to any one of claims 1 to 7 is implemented.

Images