User intention recognition method and device based on big and small model fusion and electronic equipment

Abstract

The invention provides a user intention recognition method and device based on large and small model fusion and electronic equipment. The method comprises the steps that user information is acquired, the user information is input into a small model set for user intention prediction, an intention prediction set is obtained, the intention recognition confidence coefficient of the small model set is determined according to the intention prediction set, and the user information comprises user questions and user portrait features; according to the intention recognition confidence coefficient, the customer service agent is adopted to analyze and process the user information to obtain a COT analysis table, and the COT analysis table is sent to the large model; and according to the COT analysis table and the user information, adopting the large model to perform preset processing on the user question to obtain a target user question, and inputting the target user question into the small model to obtain a user intention recognition result. The problem that in the prior art, user intention recognition independently adopting a large model or a small model is low in user satisfaction degree is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of user intention recognition, in particular to a user intention recognition method and device based on large and small model fusion, a computer readable storage medium and an electronic device. BACKGROUND

[0002] Large models are good at understanding, summarizing and providing high-level guidance, while small models are better at perceiving and executing specific tasks. However, due to the large number of model parameters of large models, the deployment cost is high, and the online inference response time is long. The traditional small model has a short inference response time, but its accuracy in non-professional fields is relatively lower.

[0003] In addition, due to the illusion of large models, in the financial field, it is acceptable to face internal employees, but in the field of intelligent customer service facing customers, the supervision is very strict, and without record, large models cannot directly face customers. According to Article 17 of the "Provisional Management Measures for Generative Artificial Intelligence Services", "any large model product facing the public in China, with public opinion attributes and social mobilization capabilities, needs to go through large model registration procedures", therefore, without registration, large models cannot be directly used to face customers.

[0004] And the prior art is based on small model intention recognition, or keyword-based intention recognition, or directly based on large model intention recognition. The prior art based on small model intention recognition has the problems of rigid intention understanding, only understanding the trained problems, and single round of intelligent understanding, poor semantic coherence, rigid answer, poor generalization ability, and inflexible problems. The prior art based on large model intention recognition has the problems of high deployment and inference cost of large models, long online service time delay, illusion, and difficulty in directly facing customers. SUMMARY

[0005] The main purpose of the present application is to provide a user intention recognition method and device based on large and small model fusion, a computer readable storage medium and an electronic device, to at least solve the problem of low user satisfaction of the prior art using large or small models alone.

[0006] In order to achieve the above object, according to one aspect of the present application, a user intention recognition method based on size model fusion is provided, comprising: acquiring user information and inputting the user information into a small model set for user intention prediction to obtain an intention prediction set, and determining an intention recognition confidence of the small model set according to the intention prediction set, wherein the small model set comprises a plurality of small models, the intention prediction set comprises intention prediction probabilities output by the small models, the intention recognition confidence is determined by a plurality of the intention prediction probabilities, and the user information comprises a user question and a user portrait feature; according to the intention recognition confidence, a customer service agent is used to analyze and process the user information to obtain a COT analysis table, and the COT analysis table is sent to a large model; according to the COT analysis table and the user information, the large model is used to perform preset processing on the user question to obtain a target user question, and the target user question is input into the small model to obtain a user intention recognition result.

[0007] Optionally, according to the intention recognition confidence, the customer service agent is used to analyze and process the user information, comprising: determining whether the intention recognition confidence is greater than a first threshold and less than a second threshold, wherein the second threshold is greater than the first threshold; in the case that the intention recognition confidence is less than or equal to the first threshold, it is represented that the user has no business intention, and in the case that the intention recognition confidence is greater than or equal to the second threshold, it is represented that the user has the business intention; in the case that the intention recognition confidence is greater than the first threshold and less than the second threshold, it is represented that the small model set cannot recognize whether the user has the business intention, and the customer service agent is used to analyze and process the user information.

[0008] Optionally, before the user question is input into the small model set for user intention prediction, the method further comprises: constructing a small model training set, wherein the small model training set comprises a plurality of training data groups, each training data group in the plurality of training data groups comprises historical user information and label data corresponding to the historical user information acquired in a historical time period, and the label data represents whether the user has a business intention; the small model training set is used to train each small model, and an AUC index of each small model is calculated.

[0009] Optionally, according to the intention prediction set, the intention recognition confidence of the small model set is determined, comprising: according to an AUC index corresponding to each small model, determining a weight of each small model; and according to the intention prediction probability corresponding to each small model and the weight, determining the intention recognition confidence of the small model set.

[0010] Optionally, the weight of each small model is determined according to the AUC index corresponding to each small model, including: using a first formula: determining the weight of each small model, wherein w i is the weight of the small model, auc i is the AUC index of the small model, is the sum of the AUC indexes of all the small models; the intent recognition confidence of the small model set is determined according to the intent prediction probability corresponding to each small model and the weight, including: using a second formula: determining the intent recognition confidence of the small model set, wherein P is the intent recognition confidence, p i is the intent prediction probability corresponding to the small model.

[0011] Optionally, after inputting the target user question into the small model to obtain a user intent recognition result, the method further includes: constructing a small model training corpus set based on the target user question and the user information; and iteratively optimizing the small model set using the small model training corpus set to obtain an iteratively optimized small model set.

[0012] Optionally, before inputting the user information into the small model set for user intent prediction, the method further includes: constructing a plurality of small models through a deep learning framework to obtain the small model set, wherein the deep learning framework includes tensorflow and pytorch, and the small model set includes a logistic regression model, a LightGBM model, a BERT model and an RNN model.

[0013] According to another aspect of the present application, a user intent recognition device based on large-small model fusion is provided, including: an intent prediction unit configured to obtain user information and input the user information into a small model set for user intent prediction to obtain an intent prediction set, and determine an intent recognition confidence of the small model set according to the intent prediction set, wherein the small model set includes a plurality of small models, the intent prediction set includes intent prediction probabilities output by each small model, the intent recognition confidence is determined by a plurality of intent prediction probabilities, and the user information includes a user question and a user portrait feature; an analysis processing unit configured to analyze and process the user information using a customer service intelligent agent according to the intent recognition confidence to obtain a COT analysis table, and send the COT analysis table to a large model; and a prediction processing unit configured to perform a preset processing on the user question using the large model according to the COT analysis table and the user information to obtain a target user question, input the target user question into the small model to obtain a user intent recognition result.

[0014] According to another aspect of the present application, a computer readable storage medium is provided, the computer readable storage medium comprising a stored program, wherein the computer readable storage medium controls the device where the computer readable storage medium is located to execute any one of the user intent recognition methods based on size model fusion when the program runs.

[0015] According to another aspect of the present application, an electronic device is provided, comprising one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs comprise a program for executing any one of the user intent recognition methods based on size model fusion.

[0016] According to the technical solution of the present application, user information is obtained, and the user information is input into a small model set for user intent prediction to obtain an intent prediction set. The intent recognition confidence of the small model set is determined according to the intent prediction set, wherein the small model set comprises a plurality of small models, the intent prediction set comprises intent prediction probabilities output by each small model, the intent recognition confidence is determined by a plurality of intent prediction probabilities, and the user information comprises a user question and a user portrait feature. According to the intent recognition confidence, a customer service agent is used to analyze and process the user information to obtain a COT analysis table, and the COT analysis table is sent to a large model. According to the COT analysis table and the user information, the large model is used to perform preset processing on the user question to obtain a target user question, and the target user question is input into the small model to obtain a user intent recognition result. According to the size of the intent recognition confidence of the small model set, it is determined whether to use the large model to polish the user question. The advantages of small models and large models are ingeniously combined, not only ensuring high accuracy of intent understanding, but also considering reasoning efficiency and deployment cost. The contradiction between accuracy and cost in the prior art is solved, and efficient and low-cost intent recognition is achieved. The problem of low user satisfaction in the user intent recognition of the prior art using a large model or a small model alone is solved. BRIEF DESCRIPTION OF DRAWINGS

[0017] The drawings accompanying the specification of the present application are used to provide further understanding of the present application, the illustrative embodiments of the present application and the description thereof are used to explain the present application, and do not constitute improper limitations on the present application. In the drawings:

[0018] Figure 1 A hardware structure block diagram of a mobile terminal for executing a user intent recognition method based on size model fusion is shown according to an embodiment of the present application;

[0019] Figure 2 A flowchart of a user intent recognition method based on size model fusion is shown according to an embodiment of the present application;

[0020] Figure 3 A flowchart of user intention recognition and dialogue flow library construction according to an embodiment of the present application is shown;

[0021] Figure 4 A flowchart of a user intention recognition method based on size model fusion according to an embodiment of the present application is shown;

[0022] Figure 5 A diagram of a prediction probability interval according to an embodiment of the present application is shown;

[0023] Figure 6 A flowchart of a user intention recognition method according to an embodiment of the present application is shown;

[0024] Figure 7 A structural block diagram of a user intention recognition device based on size model fusion according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0025] It should be noted that the embodiments and features of the embodiments in the present application can be combined with each other without conflict. The technical solutions in the embodiments of the present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0026] In order to enable persons skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should fall within the scope of protection of the present application.

[0027] 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 do not necessarily mean a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0028] As described in the background section, existing technologies that use either large or small models alone for user intent recognition result in low user satisfaction. To address this issue, embodiments of this application provide a user intent recognition method, apparatus, computer-readable storage medium, and electronic device based on the fusion of large and small models.

[0029] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0030] The methods and embodiments provided in this application can be executed on a mobile terminal, a computer terminal, or a similar computing device. Taking running on a mobile terminal as an example, Figure 1 This is a hardware structure block diagram of a mobile terminal based on a user intent recognition method using size model fusion, according to an embodiment of the present invention. Figure 1 As shown, a mobile terminal may include one or more ( Figure 1 Only one is shown in the diagram. A processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.) and a memory 104 for storing data are also shown. The mobile terminal may further include a transmission device 106 for communication functions and an input / output device 108. 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 mobile terminal described above. For example, the mobile terminal may also include components that are more... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0031] The memory 104 can be used to store computer programs, such as software programs of application software and modules, such as a computer program corresponding to the user intent recognition method based on size model fusion in the embodiments of the present application. The processor 102 can execute various functional applications and data processing, i.e., implement the above method, by running the computer program stored in the memory 104. The memory 104 can include a high-speed random access memory, and can further include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory 104 can further include a memory remotely arranged with respect to the processor 102, which can be connected to the mobile terminal through a network. Examples of the above network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof. The transmission device 106 is used to receive or send data via a network. The specific examples of the above network can include a wireless network provided by a communication provider of the mobile terminal. In one example, the transmission device 106 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 device 106 can be a radio frequency (Radio Frequency, RF) module which is used to communicate with the Internet in a wireless manner.

[0032] In the present embodiment, a user intent recognition method based on size model fusion running on a mobile terminal, a computer terminal or a similar computing device 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 herein can be executed in an order different from that shown.

[0033] Figure 2 is a flowchart of the user intent recognition method based on size model fusion according to the embodiments of the present application. As shown in Figure 2 , the method comprises the following steps:

[0034] In step S201, user information is obtained, and the user information is input into a small model set for user intent prediction to obtain an intent prediction set. An intent recognition confidence of the small model set is determined according to the intent prediction set. The small model set includes a plurality of small models. The intent prediction set includes intent prediction probabilities output by the small models. The intent recognition confidence is determined by a plurality of intent prediction probabilities. The user information includes a user question and a user portrait feature.

[0035] Among them, the user portrait features include the user's age, gender, product use activity, etc. These features can help the model understand the user's background more comprehensively, thereby improving the accuracy of intent recognition. Through the fusion of small model set, each small model focuses on processing specific types of problems or features, for example, the logistic regression model may be better at processing user portrait-based predictions, while the BERT model excels in understanding text semantics. This parallel prediction and confidence fusion mechanism of multiple models not only speeds up the processing speed, but also improves the accuracy of recognition through the complementarity between models, solving the limitations of single models, especially when dealing with complex and variable user intents. For example, the small model set can include but is not limited to logistic regression model, LightGBM model, BERT model and RNN model, the combined use of these models can cover a wide range from structured data processing to unstructured text understanding, thereby achieving more comprehensive user intent recognition.

[0036] Step S202, according to the above intent recognition confidence, the customer service agent analyzes and processes the above user information to obtain a COT analysis table, and sends the COT analysis table to the large model;

[0037] Specifically, the COT analysis table (Chain of Thought Analysis Table) is generated by the customer service agent based on the prediction results of the small model set and the user information after in-depth analysis, which contains multi-level understanding of user problems, such as keywords, potential intent, emotion analysis, etc. When the intent recognition confidence of the small model set is lower than the set threshold, the customer service agent intervenes to make up for the uncertainty of the small model through more detailed analysis, and the generated COT analysis table can provide more rich input for the large model to help the large model make more accurate judgments. This mechanism ensures fast response while maintaining high accuracy of recognition, solving the accuracy problem of user intent recognition in low confidence scenarios. For example, the COT analysis table can include but is not limited to keyword extraction, semantic understanding, emotion analysis, etc. The comprehensive use of these analysis results can help the system more accurately understand the user's real needs.

[0038] Step S203, according to the above COT analysis table and the above user information, using the above large model to pre-process the above user problem to obtain a target user problem, and inputting the target user problem into the above small model to obtain a user intent recognition result.

[0039] The preset processing can be question polishing processing, for example, the user respectively asks: "Help me check the details in August", "Last month", the small model set cannot understand the second question, so the large model needs to polish the second question, and the target user question is "Help me query the details in July".

[0040] Specifically, the large model, such as a large pre-training model based on a Transformer architecture, can process more complex and deeper semantic information. Based on the COT analysis table and user information, the large model can polish the user question to convert the original question into a clearer and more specific target user question, which helps to improve the accuracy of subsequent intent recognition. For example, the large model can convert a vague user question into an explicit business requirement description, thereby improving the recognition efficiency and accuracy of the small model set. Through the collaborative work of the large model and the small model, the present scheme can effectively handle complex scenarios in user intent recognition and solve the problem of decreased recognition accuracy when facing user questions with ambiguous semantics or incomplete information.

[0041] Through the above step S201, the above step S202 and the above step S203, it is determined whether to polish the user question by the large model according to the size of the intent recognition confidence of the small model set, which ingeniously combines the advantages of small models and large models, not only ensures high accuracy of intent understanding, but also considers reasoning efficiency and deployment cost, solves the contradiction between accuracy and cost in the prior art, and realizes efficient and low-cost intent recognition. The problem of low user satisfaction in the user intent recognition of the prior art by using a large model or a small model alone is solved.

[0042] In addition, considering that there is still a certain uncontrollable risk in the polishing process of the large model, the present embodiment introduces a security guarantee mechanism in the process design: after the large model completes the polishing of the user question, the small model performs intent recognition on the polished target question, and matches based on the knowledge base. Since the content in the knowledge base is a standard answer entered into the warehouse after manual audit, it has high compliance and accuracy, thereby effectively ensuring that the final output result is 100% safe and controllable, and realizing a solution to explore the large model to empower existing business scenarios under the requirement of strong supervision in the financial field. The unrecorded and unreported face-to-face financial scenarios.

[0043] In a specific implementation process, according to the intention recognition confidence, the customer service agent is used to analyze and process the user information in step S202, including: determining whether the intention recognition confidence is greater than a first threshold and less than a second threshold, wherein the second threshold is greater than the first threshold; in the case that the intention recognition confidence is less than or equal to the first threshold, it is represented that the user has no business intention, and in the case that the intention recognition confidence is greater than or equal to the second threshold, it is represented that the user has the business intention; in the case that the intention recognition confidence is greater than the first threshold and less than the second threshold, it is represented that the small model set cannot identify whether the user has the business intention, and the customer service agent is used to analyze and process the user information.

[0044] By setting the first threshold and the second threshold, the method can effectively screen out user problems that need further analysis. When the confidence is lower than the first threshold, the system can quickly determine that the user may not have a clear business demand, thereby avoiding unnecessary waste of resources; when the confidence is higher than the second threshold, the system can confirm that the user has a clear business intention, and directly enter the subsequent processing flow to improve efficiency. When the confidence is between the two thresholds, it indicates that the small model set has uncertainty in identifying the user's intention, and the customer service agent is introduced for in-depth analysis at this time, which can effectively improve the accuracy of identification, and solves the technical problem of how to balance fast response and accurate identification in user intention recognition. The specific size of the first threshold and the second threshold can be flexibly adjusted according to the accuracy of the small model prediction, for example, the first threshold can be set to 0.3, and the second threshold can be set to 0.8. Such threshold setting can ensure that the system neither misses users with business intentions nor over-analyzes users with no clear demands, thereby finding the best balance point between user experience and system efficiency.

[0045] In addition, the embodiment provides a specific determination method for the first threshold and the second threshold, wherein the candidate set of the first threshold is {0.01, 0.02, 0.03, …, 1}, the candidate set of the second threshold is {1, 9.99, 9.98, …, 0}, and the prediction accuracy threshold is set to 95% (the actual size can be set according to specific needs); first, a test sample is obtained, which includes a plurality of user problems and intention labels (with intention or without intention) corresponding to the user problems; the small model set is used to predict each user problem in the test sample to obtain the intention prediction probability (the intention prediction probability of each user problem in the test sample) output by the small model set; the intention prediction probability < the first threshold is regarded as a no-intention label output by the small model set, and the candidate set of the first threshold is traversed from small to large, until the accuracy of no-intention prediction < 95% under the condition, it is considered that the value in the current candidate set is the first threshold.

[0046] The calculation formula for calculating the accuracy of the non-intention prediction is: the accuracy of the non-intention prediction = (the number of small models predicted as non-intention and the real label is non-intention) / the number of real non-intention.

[0047] Similarly, the determination method of the second threshold is the same as that of the first threshold, and the candidate set of the second threshold is traversed from large to small, until the accuracy of the intention prediction is less than 95%, and it is considered that the value in the current candidate set is the second threshold.

[0048] The calculation formula for calculating the accuracy of the intention prediction is: the accuracy of the intention prediction = (the number of small models predicted as intention and the real label is intention) / the number of real intention.

[0049] Specifically, before inputting the above user question into the small model set to predict the user intention, the above method further comprises: constructing a small model training set, wherein the small model training set comprises a plurality of groups of training data, each group of training data in the plurality of groups of training data comprises historical user information and label data corresponding to the historical user information obtained in a historical time period, and the label data represents whether the user has business intention; training each small model using the small model training set, and calculating the AUC index of each small model.

[0050] In this method, the construction of the small model training set is completed by collecting historical user information and its corresponding label data, and this step is the basis for ensuring that the small model set can accurately predict user intention. Historical user information can be user questions, user portrait features, user behavior data, etc., and label data explicitly labels whether the user has business intention, for example, when the user asks about product price or purchase process, the label data is "has business intention", and when the user only expresses the preference or complaint for the product, the label data is "no business intention". By training the small model set, the system can learn the association between different features and user intention, so as to quickly and accurately make predictions when facing new user questions. Calculating the AUC index (Area Under the Curve, Area Under the Receiver Operating Characteristic Curve) can evaluate the prediction performance of the model, and the closer the AUC value is to 1, the stronger the classification ability of the model, which solves the problem of how to quantitatively evaluate the performance of the small model set. For example, the training of the small model set can use historical user data, including but not limited to user questions, user portrait features, etc., by continuously optimizing model parameters, the prediction accuracy of the small model set is improved.

[0051] More specifically, determining the intent recognition confidence of the small model set according to the above-mentioned intent prediction set comprises: determining the weight of each small model according to the AUC index corresponding to each small model; and determining the intent recognition confidence of the small model set according to the intent prediction probability corresponding to each small model and the weight.

[0052] The weight distribution of each small model in the method is based on its AUC index. Small models with higher AUC indexes will be given greater weights, which means that in the fusion of prediction results, the models that perform better will have greater influence. For example, if the AUC index of the logistic regression model is 0.85 and the AUC index of the BERT model is 0.95, then in determining the weight, the BERT model will obtain a higher weight value. This performance-based weight distribution strategy ensures that the prediction results of the small model set can reflect the judgment of the most accurate model, thereby improving the overall intent recognition confidence. In this way, the present scheme solves the technical problem of how to reasonably distribute the weights among multiple small models to improve the prediction accuracy of the overall model. For example, the weight distribution can use but is not limited to the above-mentioned first formula. This formula calculation can ensure that the distribution of weights takes into account the performance of the model and avoids excessive concentration of weights on a certain model, thereby achieving more balanced contributions among models.

[0053] Further, determining the weight of each small model according to the AUC index corresponding to each small model comprises: using a first formula: determining the weight of each small model, wherein w i is the weight of the small model, auc i is the AUC index of the small model, is the sum of the AUC indexes of all small models; and determining the intent recognition confidence of the small model set according to the intent prediction probability corresponding to each small model and the weight comprises: using a second formula: determining the intent recognition confidence of the small model set, wherein P is the intent recognition confidence, p i is the intent prediction probability corresponding to the small model.

[0054] The use of the first formula and the second formula in the method is the core of the weight distribution and confidence calculation in the present scheme. The first formula calculates the weight of each model by comparing the AUC index of each small model with the sum of the AUC indexes of all small models. This calculation ensures that the weight distribution reflects the performance of the model and maintains the sum of the weights as 1, thereby achieving a fair contribution assessment among models. The second formula calculates the comprehensive confidence of the small model set based on the prediction probability of the small model and the weight. This calculation can fully utilize the advantages of each small model and obtain more stable and accurate user intent recognition results through weighted averaging. Through the above formulas, the present scheme solves the technical problem of how to quantitatively evaluate and fuse the prediction results of multiple small models to improve the accuracy of user intent recognition. For example, by adjusting the model combination and weight distribution strategy in the small model set, the system can continuously optimize its intent recognition ability to adapt to changing user needs.

[0055] Specifically, after inputting the above target user question into the above small model and obtaining the user intent recognition result, the method further comprises: constructing a small model training corpus set based on the above target user question and the above user information; and iteratively optimizing the above small model set using the above small model training corpus set to obtain an iteratively optimized small model set.

[0056] The construction of the small model training corpus set corresponding to the method is completed by collecting target user questions and their corresponding user information. These data are used to further optimize the performance of the small model set. Through continuous iterative optimization, the small model set can learn the latest user behavior patterns and intent expression methods, thereby continuously improving its prediction accuracy. For example, if the system finds that the user intent recognition confidence of the small model set is difficult to accurately predict when processing user questions, the system can collect these biased cases and construct a new training corpus set to perform targeted optimization training on the small model set. This mechanism ensures that the small model set can continuously adapt to changes in user needs and solves the technical problem of how to continuously optimize model performance to cope with changing user behavior patterns in user intent recognition. For example, the iterative optimization process can use but is not limited to the above method to continuously collect new data, adjust model parameters, and improve the prediction ability of the small model set.

[0057] Further, before inputting the above user information into the small model set for user intent prediction, the method further comprises: constructing a plurality of the above small models through a deep learning framework to obtain the above small model set, wherein the deep learning framework comprises tensorflow and pytorch, and the small model set comprises a logistic regression model, a LightGBM model, a BERT model, and an RNN model.

[0058] The deep learning frameworks in this method, such as tensorflow and pytorch, provide tools and environments for building and training deep learning models. Through these frameworks, small models such as logistic regression models, LightGBM models, BERT models, and RNN models can be efficiently constructed, trained, and optimized. For example, the tensorflow framework provides rich APIs and tools to support the entire process from model definition to training and testing, while the pytorch framework is favored by many researchers due to its dynamic computation graph and ease of debugging. By using these deep learning frameworks, this solution can quickly build and optimize a small model set, solving the technical problem of how to efficiently build and train multiple small models to achieve rapid user intent recognition. For example, the construction of the small model set can use but is not limited to the above-mentioned deep learning frameworks, and by selecting appropriate frameworks and model types, the system can more flexibly and efficiently handle user intent recognition tasks.

[0059] In addition, the present embodiment also includes adaptive model selection and dynamic parameter tuning, which specifically includes the following contents:

[0060] When the system receives a new inquiry, it will first analyze the complexity of the problem and whether it involves a specific professional field. If it is a routine problem or a non-professional field problem, the system will preferentially call the small model for rapid response; while for complex problems or professional field problems, the system will automatically switch to the large model for in-depth analysis. In addition, the dynamic parameter tuning algorithm will adjust the model parameters according to real-time feedback to ensure that the model maintains the best state when facing changing user demands and technical environments.

[0061] Through the implementation of adaptive model selection and dynamic parameter tuning, the system can more intelligently allocate computing resources, significantly reducing response time and inference cost. At the same time, dynamic parameter tuning can optimize model performance for specific scenarios, improving the accuracy of intent recognition.

[0062] In addition, the present embodiment also includes multi-modal fusion and context awareness, which specifically includes the following contents:

[0063] Considering that the behavior patterns and preferences of users in different contexts may affect their intent expression, this embodiment further integrates multi-modal data (such as voice, image, text) and context information (such as user location, time, historical interaction records) for intent recognition. For example, the system can analyze the user's voice tone (emotion recognition), facial expression (during video calls), and historical transaction records to comprehensively judge the user's true intent. This multi-modal fusion and context awareness approach can capture more rich user information, thus more accurately identifying user intent.

[0064] After the addition of multi-modal fusion and context awareness, the system's understanding of user intent is more comprehensive and in-depth. In an experiment involving 1000 users, through multi-modal fusion, the system's accuracy in intent recognition improved by 15% when facing queries containing implicit emotions or complex information. Especially in customer service scenarios involving emotional fluctuations, this feature significantly improves service quality, reduces misunderstandings, and enhances user experience. At the same time, combined with context information, the system can provide more personalized services, such as prioritizing and providing quick responses to urgent fund query requests during the lunchtime peak on weekdays.

[0065] In order for those skilled in the art to more clearly understand the technical solutions of the present application, the implementation process of the user intent recognition method based on size model fusion of the present application will be described in detail below in conjunction with specific embodiments.

[0066] The present embodiment relates to a specific user intent recognition method based on size model fusion, as shown in Figure 3 The small model understands the intent according to the customer's question and the opening, and determines according to the confidence level of the small model output. In the case of confidence level in the fuzzy interval, the small model cannot understand the user's question, at which point the large model is needed to rewrite and polish the user's question, input the polished question into the small model, and the small model performs identification. According to the identification result, match through the dialogue flow library to obtain the business dialogue corresponding to the user's question, and output to the customer. The construction of dialogue flow tree and the generation of dialogue are constructed through the response of business audit to ensure that the business dialogue matched by the small model is in line with the business standard. The flowchart of the specific user intent recognition method based on size model fusion is shown in Figure 4 as follows:

[0067] First, the input question is 2 pieces of information for the user, asking "Help me check the details for August" and "Last month."

[0068] The small model does not have the ability to understand the context and cannot extract the effective slot for the customer's second sentence, so the confidence scores of the small model's intent understanding and slot extraction are relatively low, and the confidence score obtained by the small model result fusion is in the fuzzy interval.

[0069] Then, determine whether the number of times the large model is used to polish and rewrite the question exceeds N times. In the case of less than N times, through the semantic routing judgment of the large model intelligent hub, accurately route to the customer service intelligent agent, rather than other intelligent agents.

[0070] Customer service agent: Through the context history record of human-computer interaction, the intention inheritance is realized, the COT analysis table of the large model is combined, the customer's question is polished, and the customer's intention is clarified, wherein the size of N value can be set according to specific identification requirements, for example, it can be set to 2 times, 3 times, etc.

[0071] The large model sends the result of transcribing the user's question into the small model again, and re-performs intention understanding, and at this time the polished question is: "Help me query the details in July".

[0072] After the small model re-understands, the answer confirmed by the artificial in the knowledge base is matched and returned to the customer.

[0073] It is worth noting that if the answer polished by the large model is still not understood by the small model, the large model can continue to polish again, and the number of cycles can be configured and adjusted according to the actual situation of the system.

[0074] For multiple intention questions of the customer, the small model does not have the ability to disassemble the question, and the large model can disassemble the intention, and then let the small model understand one by one, and match the answer confirmed by the artificial in the knowledge base and return to the customer.

[0075] For the question that the small model does not understand the first time, after the large model polishes and rewrites the user's question, the small model can understand the question and deposit the knowledge as the accumulation of training corpus for training the small model.

[0076] In addition, when the number of times of polishing and rewriting the question by the large model exceeds N times, and the confidence of the small model output is still in the fuzzy interval, the small model outputs the dialogue with the user.

[0077] The training process of the small model includes the following contents:

[0078] 1) Data preparation; through the log system, N data is collected to form a data set D={(x1, y1),..., (x N , y N )}, x represents the characteristics, which are composed of user portrait characteristics (age, product use activity, etc.), scene characteristics (time, place, etc.), y represents business intention or no business intention, only two categories, 1 represents intention, and 0 represents no intention. Further divide the training set D train =(x1, y1),..., (x n , y n ), and the test set D test =(x1, y1),..., (x m , y m ), generally 10% of the data is selected as the test set, and the other data is the training set.

[0079] 2) Offline training of small models; construct k small models through tensorflow or pytorch, etc. The small models have fewer parameters, not exceeding 100,000 parameters. The selection of small models can be logistic regression models, LightGBM, BERT, RNN, etc. Use the training data in 1) to train the small models, and test the AUC index of each model on the test set. AUC = {auc i ,..., auc k}. The larger this index is, the higher the prediction accuracy of the model.

[0080] 3) Calculation of the fusion weights of small models. Since the probability of the predicted category is from 0 to 1, it is necessary to fuse the predicted values of each small model as the final output, and it is necessary to ensure that each weight is positively correlated with the accuracy. The higher the accuracy, the higher the weight. Use auc division sum as the weight:

[0081] 4) Confidence judgment. a is the lower threshold for unintentional prediction; b is the upper threshold for intentional prediction; if the prediction probability < a, it is unintentional; if the prediction probability > b, it is intentional; if the probability falls within the interval [a, b], it needs to enter the large model for polishing. However, it is necessary to determine how much is considered a high probability and how much is considered a low probability. As Figure 5 shown, it is necessary to determine the values of the interval [a, b].

[0082] Adopt the fusion parameters in 3), fuse the predicted probability values of k small models on the test set, and obtain the fusion probability as p = {p1,...., p m}, and the true label y = {y1,..., y m};

[0083] Determine the value of a. Assume the candidate set of a is a 候选 = {0.01, 0.02,...., 1}, a sequence from 0 to 1 with an interval of 0.01. For each candidate threshold a (from small to large), calculate: at this threshold, "unintentional prediction accuracy" = TP / (TP + FN); when the accuracy < the target threshold (such as 95%), stop and select the current a value as the final threshold, TP is the number of true unintentional, and (TP + FN) is the number of predicted unintentional and true label unintentional.

[0084] Traverse the candidate set starting from 0.01. The current traversed value is represented by a tmp . When the fusion probability value is less than this value, p m <a tmp : Then it is predicted as unintentional.

[0085] For the test set, the accuracy of the "unintentional" prediction can be calculated = (the number of predictions as unintentional and the real label as unintentional) / the number of real unintentional; if the "unintentional" prediction accuracy is less than 95%, exit the loop, and the value of b is the current value; otherwise, continue the loop.

[0086] Determine the value of b, determine the value of b, assume that the b candidate set is b 候选 ={0.01, 0.02,..., 1}, a sequence of 0 to 1 with a spacing of 0.01. For the candidate threshold b (from large to small), calculate: "intentional prediction accuracy" = TN / (TN + FP), when the accuracy < target threshold, stop and select the current b value as the final threshold; TN is the number of real intentions, (TN + FP) is the number of predictions as intentions and real labels as intentions.

[0087] From 1 to the end of the candidate set, the current traversal value b tmp Use to represent that the fusion probability value is less than the value: p m >b tmp Then predict as intentional.

[0088] For the test set, the accuracy of the "unintentional" prediction can be calculated = (the number of predictions as unintentional and the real label as unintentional) / the number of real unintentional; if the "unintentional" prediction accuracy is less than 95%, exit the loop, and the value of b is the current value; otherwise, continue the loop.

[0089] For values in the interval [a, b] of the prediction probability, it is considered to be a fuzzy intention, that is, the prediction certainty is not high. For this part of the data, the large model needs to be used for judgment. If the prediction probability is not in the interval, it is predicted as the corresponding unintentional or intentional label, and directly output.

[0090] Wherein, the value of a and b can be flexibly adjusted according to the accuracy of the small model prediction.

[0091] 5) Large model COT reasoning, for the data of the intention model, enter the large model reasoning thinking process, use the following prompt to activate the large model to reason and think about the intention, and get the large model analysis result.

[0092] Examples are as follows:

[0093]

Large model input part

[0094] System prompt word = "You are a financial field expert, please clarify and optimize the customer's current problem according to the customer's historical context, customer's current problem, customer's label and other related information.

[0095] Output requirements:

[0096] 1. Clarification and polishing: Accurately understand the true intention of the customer's current question and rewrite it as a clear, complete, and natural language question that conforms to the expression habits of the financial business scenario.

[0097] 2. Slot extraction and reasoning: Based on the polished question, reason and extract the specific business slots required for the query.

[0098] 3. Output format: Output the final result in structured JSON format, containing a key-value pair: {"query":"polished complete question"}.

[0099] Processing logic explanation:

[0100] * Carefully analyze the logical association between the customer's historical context and the current question (such as time, object, operation, etc.).

[0101] * Consider the query preferences or information needs implied by the customer's label (such as "young, low deposit") (such as more attention to recent, small transactions, etc.).

[0102] * Clarify ambiguous references (such as "last month", "that") to specific business terms (such as specific months).

[0103] * Ensure that the polished question can be directly used for subsequent business system queries.

[0104] "";

[0105] Business cues: {

[0106] * Historical context: "Help me query the bill for July";

[0107] * Current question: "What about last month?";

[0108] * Customer label: ["young", "low deposit"];

[0109]

Expected output example

[0110] Based on the above input, the reasonable output should be:

[0111] {

[0112] "query":"Help me query the bill for June"

[0113] }

[0114] Supplementary explanation:

[0115] In this embodiment, the "intention recognition" task is divided into two levels:

[0116] First layer: Intent Presence Detection:

[0117] Determine whether the user's current question contains an explicit business handling intention (for example: "Is there any preferential activity?" Yes, there is a business intention; "What is the weather today?" No business intention).

[0118] Second layer: Intent Classification: On the basis of confirming that the user has an explicit business intention, further identify the specific intention category (such as "query balance", "apply for loan", "transfer remittance", etc.).

[0119] The intent recognition confidence P mainly serves the first layer judgment. Only when the P value falls in the fuzzy interval [a, b], the large model will be triggered to clarify the context, optimize the semantics, and then be identified by the small model again to complete more detailed intent clarification and classification. The specific process logic is shown in Figure 6 .

[0120] 6) Further use the large model to extract the final prediction label from the above analysis results.

[0121] 7) Data backflow: Large model perfects small model knowledge system: based on the generalization ability of the large model and the reply of the large model, re-labeling is performed again, the training corpus of the small model is supplemented, and a knowledge closed loop is formed.

[0122] The embodiment of the present application also provides a user intent recognition device based on large and small model fusion. It should be noted that the user intent recognition device based on large and small model fusion of the embodiment of the present application can be used to execute the user intent recognition method based on large and small model fusion provided by the embodiment of the present application. The device is used to realize the above-mentioned embodiments and preferred embodiments, and will not be described here. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware, or a combination of software and hardware implementation is also possible and is conceived.

[0123] The user intent recognition device based on large and small model fusion provided by the embodiment of the present application is introduced below.

[0124] Figure 7 is a schematic diagram of the user intent recognition device based on large and small model fusion according to the embodiment of the present application. As Figure 7 shown, the device includes:

[0125] The intention prediction unit 71 is configured to obtain user information, input the user information into a small model set to predict user intention, obtain an intention prediction set, determine an intention recognition confidence of the small model set according to the intention prediction set, wherein the small model set includes a plurality of small models, the intention prediction set includes intention prediction probabilities output by the small models, the intention recognition confidence is determined by the intention prediction probabilities, and the user information includes a user question and a user portrait feature.

[0126] The analysis processing unit 72 is configured to analyze and process the user information by using a customer service agent according to the intention recognition confidence, obtain a COT analysis table, and send the COT analysis table to a large model.

[0127] The prediction processing unit 73 is configured to perform preset processing on the user question by using the large model according to the COT analysis table and the user information, obtain a target user question, input the target user question into the small model, and obtain a user intention recognition result.

[0128] In this embodiment, the intention prediction unit is configured to obtain user information, input the user information into a small model set to predict user intention, obtain an intention prediction set, determine an intention recognition confidence of the small model set according to the intention prediction set, wherein the small model set includes a plurality of small models, the intention prediction set includes intention prediction probabilities output by the small models, the intention recognition confidence is determined by the intention prediction probabilities, and the user information includes a user question and a user portrait feature. The analysis processing unit is configured to analyze and process the user information by using a customer service agent according to the intention recognition confidence, obtain a COT analysis table, and send the COT analysis table to a large model. The prediction processing unit is configured to perform preset processing on the user question by using the large model according to the COT analysis table and the user information, obtain a target user question, input the target user question into the small model, and obtain a user intention recognition result. It is verified by experiments that the polished user question is more consistent with the small model training corpus distribution, and the recognition accuracy is significantly improved. According to the size of the intention recognition confidence of the small model set, it is determined whether to use the large model to polish the user question. The advantages of small models and large models are ingeniously combined, which not only ensures high accuracy of intention understanding, but also considers reasoning efficiency and deployment cost, solves the contradiction between accuracy and cost in the prior art, and realizes efficient and low-cost intention recognition. The user intention recognition by using only the large model or the small model in the prior art has the problem of low user satisfaction.

[0129] As an optional solution, the analysis processing unit comprises a first determination module and an analysis processing module; the first determination module applies to determine whether the intention recognition confidence is greater than a first threshold and less than a second threshold, wherein the second threshold is greater than the first threshold; in the case that the intention recognition confidence is less than or equal to the first threshold, it is represented that the user has no service intention, and in the case that the intention recognition confidence is greater than or equal to the second threshold, it is represented that the user has the service intention; the analysis processing module is used to represent that the small model set cannot identify whether the user has the service intention in the case that the intention recognition confidence is greater than the first threshold and less than the second threshold, and the customer service intelligent agent is used to analyze and process the user information.

[0130] As an optional solution, the device further comprises a first construction unit and a training unit; the first construction unit is used to construct a small model training set before the user question is input into the small model set for user intention prediction, wherein the small model training set comprises a plurality of groups of training data, and each group of training data in the plurality of groups of training data comprises historical user information and label data corresponding to the historical user information acquired in a historical time period, and the label data represents whether the user has a service intention; the training unit is used to train each small model by using the small model training set and calculate the AUC index of each small model.

[0131] As an optional solution, the intention prediction unit comprises a second determination module and a third determination module; the second determination module is used to determine the weight of each small model according to the AUC index corresponding to each small model; and the third determination module is used to determine the intention recognition confidence of the small model set according to the intention prediction probability corresponding to each small model and the weight.

[0132] As an optional solution, the second determination module comprises a first determination submodule, which is used to determine the weight of each small model by using a first formula: wherein w i is the weight of the small model, auc i is the AUC index of the small model, is the sum of the AUC indexes of all the small models; and the third determination module comprises a second determination submodule, which is used to determine the intention recognition confidence of the small model set by using a second formula: wherein P is the intention recognition confidence, p i is the intention prediction probability corresponding to the small model.

[0133] An optional solution, the device further comprises a second construction unit and an iterative optimization unit; the second construction unit is used to construct a small model training corpus based on the target user question and the user information after inputting the target user question into the small model and obtaining the user intent recognition result; the iterative optimization unit is used to perform iterative optimization processing on the small model set by using the small model training corpus, and obtain the small model set after iterative optimization processing.

[0134] An optional solution, the device further comprises a third construction unit, before inputting the user information into the small model set for user intent prediction, a plurality of small models are constructed through a deep learning framework to obtain the small model set, wherein the deep learning framework includes tensorflow and pytorch, and the small model set includes a logistic regression model, a LightGBM model, a BERT model and an RNN model.

[0135] The user intent recognition device based on the size model fusion comprises a processor and a memory, the intent prediction unit, the analysis processing unit, the prediction processing unit and the like are stored in the memory as program units, and the corresponding functions are realized by the processor executing the program units stored in the memory. The modules are located in the same processor; or, the modules are located in different processors in any combination.

[0136] The processor comprises a core, and the core retrieves the corresponding program unit from the memory. The core can be set to one or more, and the problem of low user satisfaction in the prior art user intent recognition using a large model or a small model alone can be solved by adjusting the core parameters.

[0137] The memory can include non-permanent memory in a computer readable medium, random access memory (RAM) and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM), and the memory includes at least one memory chip.

[0138] The embodiment of the application provides a computer readable storage medium, and the computer readable storage medium comprises a stored program, wherein the program controls the device where the computer readable storage medium is located to execute the user intent recognition method based on the size model fusion when the program runs.

[0139] Specifically, the user intent recognition method based on the size model fusion comprises:

[0140] In step S201, user information is acquired, and the user information is input into a small model set to perform user intent prediction, to obtain an intent prediction set, and an intent recognition confidence of the small model set is determined according to the intent prediction set, wherein the small model set includes a plurality of small models, the intent prediction set includes intent prediction probabilities output by the small models, the intent recognition confidence is determined by the plurality of intent prediction probabilities, and the user information includes a user question and a user portrait feature.

[0141] In step S202, according to the intent recognition confidence, a customer service agent is used to analyze and process the user information, to obtain a COT analysis table, and the COT analysis table is sent to a large model.

[0142] In step S203, according to the COT analysis table and the user information, the large model is used to perform preset processing on the user question, to obtain a target user question, and the target user question is input into the small model, to obtain a user intent recognition result.

[0143] An embodiment of the present application provides a processor, which is used to run a program, wherein the processor executes the user intent recognition method based on large and small model fusion when the program runs.

[0144] Specifically, the user intent recognition method based on large and small model fusion includes:

[0145] In step S201, user information is acquired, and the user information is input into a small model set to perform user intent prediction, to obtain an intent prediction set, and an intent recognition confidence of the small model set is determined according to the intent prediction set, wherein the small model set includes a plurality of small models, the intent prediction set includes intent prediction probabilities output by the small models, the intent recognition confidence is determined by the plurality of intent prediction probabilities, and the user information includes a user question and a user portrait feature.

[0146] In step S202, according to the intent recognition confidence, a customer service agent is used to analyze and process the user information, to obtain a COT analysis table, and the COT analysis table is sent to a large model.

[0147] In step S203, according to the COT analysis table and the user information, the large model is used to perform preset processing on the user question, to obtain a target user question, and the target user question is input into the small model, to obtain a user intent recognition result.

[0148] An embodiment of the present application provides an electronic device, which includes a processor, a memory, and a program stored in the memory and capable of running on the processor, and the processor implements at least the following steps when executing the program:

[0149] Step S201, obtain user information, and input the user information into a small model set for user intent prediction to obtain an intent prediction set, determine an intent recognition confidence of the small model set according to the intent prediction set, wherein the small model set includes a plurality of small models, the intent prediction set includes intent prediction probabilities output by the small models, the intent recognition confidence is determined by the plurality of intent prediction probabilities, and the user information includes a user question and a user portrait feature;

[0150] Step S202, analyze and process the user information by using a customer service intelligent agent according to the intent recognition confidence, obtain a COT analysis table, and send the COT analysis table to a large model;

[0151] Step S203, perform preset processing on the user question by using the large model according to the COT analysis table and the user information, obtain a target user question, input the target user question into the small model, and obtain a user intent recognition result.

[0152] The device herein can be a server, a PC, a PAD, a mobile phone, or the like.

[0153] The application also provides a computer program product adapted to execute a program including at least the following method steps when executed on a data processing device:

[0154] Step S201, obtain user information, and input the user information into a small model set for user intent prediction to obtain an intent prediction set, determine an intent recognition confidence of the small model set according to the intent prediction set, wherein the small model set includes a plurality of small models, the intent prediction set includes intent prediction probabilities output by the small models, the intent recognition confidence is determined by the plurality of intent prediction probabilities, and the user information includes a user question and a user portrait feature;

[0155] Step S202, analyze and process the user information by using a customer service intelligent agent according to the intent recognition confidence, obtain a COT analysis table, and send the COT analysis table to a large model;

[0156] Step S203, perform preset processing on the user question by using the large model according to the COT analysis table and the user information, obtain a target user question, input the target user question into the small model, and obtain a user intent recognition result.

[0157] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0158] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0159] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0160] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0161] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process.Figure 1 one or more processes and / or functions specified in one or more blocks Figure 1 one or more processes and / or functions specified in one or more blocks

[0162] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0163] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) about which the processor can execute instructions. The memory can also include non-volatile memory, such as read only memory (ROM), electrically programmable read only memory (EPROM), electrically erasable programmable read only memory (EEPROM), flash memory, or other memory technologies. The memory is an example of computer readable media.

[0164] Computer readable media includes permanent and non-permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically programmable read only memory (EEPROM), flash memory or other memory technologies, compact disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer readable media does not include transitory media such as modulated data signals and carrier waves.

[0165] It should also be noted that the terms "comprising", "containing", or any other variant thereof are intended to encompass a non-exclusive inclusion, such that a process, method, article or apparatus that comprises a list of elements does not include only those elements recited, but can also include other elements not expressly listed or inherent to such process, method, article or apparatus. Without further limitation, an element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or apparatus that includes the element.

[0166] The above description is merely that of the preferred embodiments of the application and is not intended to limit the application. The application can be modified and varied greatly without departing from the spirit and scope of the application, wherein any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A user intention recognition method based on size model fusion, characterized in that, The method comprises the following steps: obtaining user information and inputting the user information into a small model set for user intent prediction to obtain an intent prediction set, and determining an intent recognition confidence of the small model set according to the intent prediction set, wherein the small model set comprises a plurality of small models, the intent prediction set comprises intent prediction probabilities output by the small models, the intent recognition confidence is determined by a plurality of the intent prediction probabilities, and the user information comprises a user question and a user portrait feature; according to the intent recognition confidence, analyzing and processing the user information by using a customer service intelligent agent to obtain a COT analysis table, and sending the COT analysis table to a large model; according to the COT analysis table and the user information, performing preset processing on the user question by using the large model to obtain a target user question, and inputting the target user question into the small model to obtain a user intent recognition result.

2. The method of claim 1, wherein, According to the intent recognition confidence, analyzing and processing the user information by using a customer service intelligent agent, comprising: determining whether the intent recognition confidence is greater than a first threshold and less than a second threshold, wherein the second threshold is greater than the first threshold; in the case that the intent recognition confidence is less than or equal to the first threshold, it is represented that the user has no business intention, and in the case that the intent recognition confidence is greater than or equal to the second threshold, it is represented that the user has the business intention; in the case that the intent recognition confidence is greater than the first threshold and less than the second threshold, it is represented that the small model set cannot identify whether the user has the business intention, and the customer service intelligent agent is used to analyze and process the user information.

3. The method of claim 1, wherein, Before inputting the user question into the small model set for user intent prediction, the method further comprises: constructing a small model training set, wherein the small model training set comprises a plurality of training data sets, each of the plurality of training data sets comprises historical user information and label data corresponding to the historical user information obtained in a historical time period, and the label data represents whether the user has a business intention; training each of the small models by using the small model training set, and calculating an AUC index of each of the small models.

4. The method of claim 3, wherein, According to the intent prediction set, determining the intent recognition confidence of the small model set, comprising: determining the weight of each of the small models according to the AUC index corresponding to each of the small models; determining the intent recognition confidence of the small model set according to the intent prediction probability corresponding to each of the small models and the weight.

5. The method of claim 4, wherein, after inputting the target user question into the small model to obtain a user intent recognition result, the method further comprises: According to the AUC indicators corresponding to each small model, weights of each small model are determined, including: The weights of each small model are determined, wherein w i is the weight of the small model, auc i is the AUC indicator of the small model, is the sum of the AUC indicators of all the small models; According to the intention prediction probability corresponding to each small model and the weight, the intention recognition confidence of the small model set is determined, including: adopting a second formula: The intention recognition confidence of the small model set is determined, wherein P is the intention recognition confidence, p i is the intention prediction probability corresponding to the small model.

6. The method of claim 1, wherein, constructing a small model training corpus set based on the target user question and the user information; iteratively optimizing the small model set by using the small model training corpus set to obtain an iteratively optimized small model set. Before inputting the user information into the small model set for user intent prediction, the method further comprises:

7. The method of claim 1, wherein, ​ A plurality of small models are constructed through a deep learning framework to obtain the small model set, wherein the deep learning framework includes tensorflow and pytorch, and the small model set includes a logistic regression model, a LightGBM model, a BERT model, and an RNN model. 8.A user intention recognition device based on size model fusion, characterized by, Comprise: An intent prediction unit is configured to obtain user information, input the user information into a small model set for user intent prediction, obtain an intent prediction set, and determine an intent recognition confidence of the small model set according to the intent prediction set, wherein the small model set includes a plurality of small models, the intent prediction set includes intent prediction probabilities output by the small models, the intent recognition confidence is determined by a plurality of the intent prediction probabilities, and the user information includes a user question and a user portrait feature; An analysis processing unit is configured to analyze and process the user information using a customer service agent according to the intent recognition confidence, obtain a COT analysis table, and send the COT analysis table to a large model; A prediction processing unit is configured to perform a preset processing on the user question using the large model according to the COT analysis table and the user information, obtain a target user question, input the target user question into the small model, and obtain a user intent recognition result.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium comprises a stored program, wherein the program controls the device where the computer-readable storage medium is located to execute the user intent recognition method based on large and small model fusion according to any one of claims 1 to 7 when the program is running.

10. An electronic device, comprising: Comprise: One or more processors, memories, and one or more programs, wherein the one or more programs are stored in the memories and configured to be executed by the one or more processors, and the one or more programs comprise a program for executing the user intent recognition method based on large and small model fusion according to any one of claims 1 to 7.

Images