Intention recognition method based on large language model and customer portrait classification model

Abstract

An intent recognition method based on a large language model and a customer portrait classification model, the method comprising: obtaining a customer's current relevant order and background information according to customer ID information; a customer portrait classification model obtains a first score vector according to the relevant order and the background information; the first score vector is used to indicate a current customer portrait classification score; a fine-tuned large model performs semantic recognition on the context of an input customer conversation to obtain a second score vector; the second score vector is used to indicate a customer current emotional feedback score; the fine-tuned large model is obtained by fine-tuning a large language model; a comprehensive score vector is obtained by fusing the first score vector and the second score vector; the comprehensive score vector is used to indicate different tendency weights in different business scenarios; an inference large model performs inference according to the comprehensive score vector, and outputs a customer intent; a corresponding business logic is driven to be executed according to the customer intent, and the corresponding business logic comprises providing customized suggestions or triggering a specific service process.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to an intent recognition method based on a large language model and a customer portrait classification model. Background Art

[0002] Intent recognition is a critical component in natural language processing. With the continuous development of the internet and artificial intelligence (AI), intelligent conversational systems, such as intelligent customer service and intelligent assistants, are becoming increasingly popular in various application scenarios. However, traditional intent recognition methods often rely on extensive manual rule-based knowledge and traditional NLP recognition methods, such as conversation intent slot filling. These methods require significant manual effort and suffer from low accuracy.

[0003] Advances in deep learning technology have enabled large language models (such as GPT and LLAMA) to achieve significant results in natural language processing tasks. At the same time, customer profile classification models are gaining increasing attention due to their ability to better understand customer needs and preferences. By mining customer attributes and behavioral data, customer profile classification models can provide rich contextual information. Therefore, combining large language models with customer profile classification models for intent recognition has significant research significance and application value.

[0004] The rapid development of big data and machine learning technologies has promoted the widespread application of intelligent customer service systems and intelligent personal assistants. Intent recognition, as the core technology of these applications, has become a research hotspot.

[0005] Current intent recognition methods primarily rely on predefined rules or traditional machine learning models, which suffer from poor robustness and low adaptability. Large language models (such as GPT and LLAMA) excel in natural language processing tasks due to their powerful language generation and understanding capabilities. However, current large language models still have certain limitations when dealing with personalized needs and complex scenarios. Customer intent recognition implemented by large language models is generally based on two approaches: configuring customer conversation scenarios through prompt word engineering, and training large amounts of customer conversation data through continuous pre-training or fine-tuning methods (such as SFT and LORA) to solve the problem of understanding and recognizing customer intent.

[0006] With the development of internet technology, businesses and organizations need to process large amounts of customer interaction data to better understand customer needs and intentions. However, existing intent recognition systems typically rely on predefined rules, keyword matching, rule engines, or simple statistical models. These methods have significant limitations when dealing with complex and changing customer intent.

[0007] Although large language models (such as BERT and GPT) have demonstrated potential in natural language processing, existing technologies, while capable of understanding contextual information in natural language, lack the ability to provide personalized services. Customer profiling, a method for analyzing and describing customer characteristics and behaviors, can help product and operations personnel better understand customer needs and preferences, thereby providing higher-quality, more personalized products and services. Customer profiling technology also has considerable advantages in understanding customer information.

[0008] While customer profile classification models can provide customer characteristics, they still need improvement in real-time intent recognition. Existing technologies have yet to fully explore and leverage the advantages of combining these two approaches. Currently, some existing use cases that integrate large language models with customer profiling primarily use them to extract richer and deeper features, such as customer emotions, attitudes, intentions, preferences, and values, from raw customer data, such as conversation text, customer behavior information, and work order details. This is because conversation text may reflect a customer's true needs and personality, and it can increase the dimensionality and complexity of the data, thereby increasing the amount of information and discriminability. Therefore, the natural language processing capabilities of large language models are leveraged to extract multiple levels of customer features, including semantics, pragmatics, tone, style, emotion, expression, voice, and gestures, from various data formats, including text, images, audio, and video. This allows for the construction of a more comprehensive and detailed customer profile. However, these integrations of large language models and customer profiling primarily use the large language model as a feature extraction or data source for customer profiling, serving a one-sided purpose, rather than integrating the two to meet the demands of specific, diverse, and diverse business scenarios.

[0009] Especially changing business scenarios involve highly personalized and contextually diverse customer interactions. Traditional intent recognition systems often struggle to accurately capture customers' true intentions. Therefore, a more intelligent technology is needed that can deeply understand customer characteristics and effectively predict their needs, thereby improving the accuracy and robustness of intent recognition. This new intent recognition technology should combine the strengths of large language models and customer profile classification models to cope with complex language environments and changing customer needs, thereby achieving more intelligent customer intent recognition and more precise subsequent intelligent action execution. Summary of the Invention

[0010] The embodiments of this application provide an intent recognition method based on a large language model and a customer profile classification model. This method combines deep learning technology with big data analysis, integrating natural language understanding with customer behavior analysis to achieve high-precision recognition and response processing of customer intent. This embodiment of the application can improve the accuracy and efficiency of intent recognition and more accurately understand and predict customer intent, thereby providing more personalized and efficient customer service.

[0011] In the first aspect, an embodiment of the present application proposes an intent recognition method based on a large language model and a customer portrait classification model, the method comprising: obtaining the customer's current relevant orders and background information based on the customer ID information; the customer portrait classification model obtains a first scoring vector based on the relevant orders and background information; the first scoring vector is used to indicate the current customer portrait classification score; the fine-tuning large model performs semantic recognition on the context of the input customer conversation to obtain a second scoring vector; the second scoring vector is used to indicate the customer's current emotional feedback score; the fine-tuning large model is obtained by fine-tuning the large language model; the first scoring vector and the second scoring vector are fused to obtain a comprehensive scoring vector; the comprehensive scoring vector is used to indicate different tendency weights in different business scenarios; the inference large model performs inference based on the comprehensive scoring vector and outputs the customer's intent; the corresponding business logic is driven to execute according to the customer's intent, and the corresponding business logic includes providing customized suggestions or triggering specific service processes.

[0012] In some embodiments, a linear weighted algorithm is used to concatenate the first score vector and the second score vector to form a comprehensive score vector.

[0013] In some embodiments, fusing the first rating vector and the second rating vector to obtain a comprehensive rating vector includes: using a multi-layer perceptron (MLP) to perform weighted fusion of the first rating vector and the second rating vector by learning a set of fixed weights to obtain the comprehensive rating vector.

[0014] In some embodiments, the large inference model performs inference based on the comprehensive scoring vector and outputs the customer's intention classification, including: the number of customer intention classifications is M, and the top-K intentions with the highest scores are selected based on the M intention classifications; or low-confidence intentions among the M intentions are filtered out by setting a threshold to obtain K intentions that meet the requirements; or M intentions are sorted by priority; or M intentions are executed in parallel.

[0015] In some embodiments, the first score vector and the second score vector are fused to obtain a comprehensive score vector, including: arranging the output weights of the first score vector and the second score vector through agent arrangement and prompt word engineering.

[0016] In some embodiments, the output weights of the first scoring vector and the second scoring vector are arranged through intelligent agent orchestration and prompt word engineering, including: determining the key nodes of the task process of the inference large model, the key nodes including: the customer's original intention input node, the inference call node of the customer portrait classification model, the inference call node of the fine-tuning large model, the weight merging decision node, the behavior execution node and the multi-module combination process node; converting the output of the customer portrait classification model inference call node into a descriptive statement based on semantic understanding of the fine-tuning large model; configuring the prompt word engineering for the first scoring vector and the second scoring vector, and adjusting different tendency weights in different business scenarios.

[0017] In some embodiments, the execution of corresponding business logic is driven according to the customer intent classification, including: determining an agent behavior library, the agent behavior library pre-setting tools and / or atomic capability services for registering business operation behaviors; selecting a corresponding agent combination in the agent behavior library according to the customer intent classification, and the agent combination is used to orchestrate and execute behaviors.

[0018] In some embodiments, the method further includes: obtaining the results of the interactive behavior of customer feedback based on the contextual text of the customer interaction; using the results of the interactive behavior of customer feedback as training corpus for retraining the feature weights of the customer portrait classification model; the customer portrait classification model retrains the feature weights based on the training corpus; including data integration and feature expansion, feature quantization of feedback annotation, automatic feature engineering, AutoML automatic machine learning training, incremental learning and online updates.

[0019] In some embodiments, the method further includes: fine-tuning the large model to automatically generate new annotations based on the results of the interactive behavior fed back by the customer; and continuously fine-tuning the large model using the conversation interaction results and the new annotations as input-output pairs to align the conversation and intent recognition.

[0020] In a second aspect, an embodiment of the present application proposes an intent recognition system based on a large language model and a customer profile classification model, based on the method provided in the first aspect, comprising: an information acquisition module for acquiring the customer's current relevant orders and background information based on customer ID information; a customer profile classification model for obtaining a first scoring vector based on the relevant orders and background information; the first scoring vector is used to indicate the current customer profile classification score; a fine-tuning large model for semantically recognizing the context of the input customer conversation to obtain a second scoring vector; a fusion module for the second scoring vector to indicate the customer's current emotional feedback score; the fine-tuning large model is obtained by fine-tuning the large language model; the first scoring vector and the second scoring vector are fused to obtain a comprehensive scoring vector; the comprehensive scoring vector is used to indicate different tendency weights in different business scenarios; an inference large model for inferring based on the comprehensive scoring vector and outputting the customer's intent; an intent driving module for driving the execution of corresponding business logic based on the customer's intent, the corresponding business logic including providing customized suggestions or triggering specific service processes. In a third aspect, an embodiment of the present application provides an electronic device, comprising: at least one memory for storing programs; at least one processor for executing the programs stored in the memory, and when the programs stored in the memory are executed, the processor is used to execute the method provided in the first aspect.

[0021] In a fourth aspect, an embodiment of the present application provides a computer storage medium, in which instructions are stored. When the instructions are executed on a computer, the computer executes the method provided in the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the multiple embodiments disclosed in this specification, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings described below are only the multiple embodiments disclosed in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0023] The following is a brief introduction to the drawings required for describing the embodiments or prior art.

[0024] Figure 1 A schematic diagram of a process engine framework for implementing agent orchestration provided in Example 1 of the present application;

[0025] Figure 2 A flowchart of the training phase of the customer profile classification model provided in an embodiment of the present application;

[0026] Figure 3 Flowchart of the large language model fine-tuning phase provided in the embodiment of this application;

[0027] Figure 4 A flowchart for building and training a large inference model provided in the embodiments of this application;

[0028] Figure 5 Schematic diagram of the dual-model fusion training process;

[0029] Figure 6 A schematic diagram of implementing the inference model C by arranging prompt word engineering through the agent orchestration tool;

[0030] Figure 7 A schematic diagram for result feedback and continuous optimization;

[0031] Figure 8 Schematic diagram of the process engine framework for implementing agent orchestration provided in Example 1 of the present application. DETAILED DESCRIPTION

[0032] In the description of the embodiments of the present application, words such as "exemplary," "for example," or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary," "for example," or "for example" in the embodiments of the present application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "exemplary," "for example," or "for example" is intended to present the relevant concepts in a concrete manner.

[0033] In the description of the embodiments of this application, the term "and / or" is simply a description of the association relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent the following three situations: A exists alone, B exists alone, and A and B exist at the same time. In addition, unless otherwise specified, the term "plurality" means two or more. For example, "multiple systems" refers to two or more systems, and "multiple terminals" refers to two or more terminals.

[0034] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly identifying the technical features being referred to. Thus, features specified as "first" or "second" may explicitly or implicitly include one or more of such features. The terms "include," "comprising," "having," and their variations all mean "including but not limited to," unless otherwise specifically emphasized.

[0035] In the description of the embodiments of the present application, reference is made to “some embodiments”, which describe a subset of all possible embodiments, but it can be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments, and may be combined with each other without conflict.

[0036] In the description of the embodiments of the present application, the terms "first\second\third, etc." or module A, module B, module C, etc. are only used to distinguish similar objects and do not represent a specific ordering of the objects. It can be understood that the specific order or sequence can be interchanged where permitted so that the embodiments of the present application described here can be implemented in an order other than that illustrated or described here.

[0037] In the description of the embodiments of the present application, the numbers representing the steps, such as S110, S120, etc., do not necessarily mean that the steps must be executed in this manner. If permitted, the order of the previous and next steps can be interchanged, or they can be executed simultaneously.

[0038] Relevant scientific terms used in this application:

[0039] Pre-training, first train a deep learning model on a large-scale dataset, such as using self-supervised learning or unsupervised learning.

[0040] Fine-tuning: Fine-tuning the pre-trained model using the training set for the target task. Typically, only a subset of the layers in the pre-trained model are fine-tuned, such as the last few layers or some intermediate layers. During fine-tuning, the model is optimized using the backpropagation algorithm to achieve better performance on the target task.

[0041] Evaluation: Use the test set of the target task to evaluate the fine-tuned model and obtain the performance indicators of the model on the target task.

[0042] LoRA is a method for fine-tuning large pre-trained language models by inserting low-rank matrices at specific layers to reduce the number of parameters and improve efficiency and performance.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which this application pertains. The terms used herein are for the purpose of describing the embodiments of this application only and are not intended to limit this application.

[0044] The embodiment of the present application proposes a method based on a large language model and a customer portrait classification model, which fuses the output results of the customer portrait classification model with the fine-tuning model of the large language model based on historical customer behavior, so that the large language model can combine the feature data of the customer portrait and the current real-time customer conversation context as targeted reasoning, thereby completing a more accurate understanding and judgment of customer intentions. Respond to subsequent customer behaviors and call relevant capabilities in the form of intelligent agents to meet the customer's intention needs. After obtaining positive or negative feedback from customers, the results of this round of interactive behavior are used as training corpus for further fine-tuning and alignment of the large model, and the feature weights in the original customer portrait classification model are adjusted.

[0045] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings.

[0046] Figure 1 This is a flow chart of the intent recognition method based on a large language model and a customer profile classification model provided in the embodiment of the present application. Figure 1 As shown, the method includes the following steps:

[0047] S1, obtain the customer's current related orders and background information based on the customer ID information.

[0048] S2, the customer profile classification model obtains a first scoring vector based on relevant orders and background information; the first scoring vector is used to indicate the current customer business classification score.

[0049] S3, fine-tuning the large model to perform semantic recognition on the context of the input customer conversation to obtain a second scoring vector; the second scoring vector is used to indicate the customer's current emotional feedback score; the fine-tuning large model is obtained by fine-tuning the language large model.

[0050] S4: The inference model fuses the first and second scoring vectors to obtain a comprehensive scoring vector. The comprehensive scoring vector is used to indicate different tendency weights in different business scenarios. Inference is performed based on the comprehensive scoring vector to output the customer's intention.

[0051] S5 drives the execution of corresponding business logic based on the customer's intention, which includes providing customized suggestions or triggering specific service processes.

[0052] Based on the above application process, the technical solutions in the embodiments of this application are explained one by one through multiple aspects such as customer portrait classification model, fine-tuning large model, reasoning large model, intent-driven execution, etc.

[0053] First, let’s introduce the customer portrait classification model. Figure 2 This is a flowchart of the construction and training phase of the customer portrait classification model provided in the embodiment of this application. Figure 2 As shown, the following steps are included:

[0054] S21, obtain customer data and preprocess the customer data; the customer data includes historical customer order records and / or work order records extracted from the database based on customer ID information, and the preprocessing includes data cleaning, data standardization, missing value processing and outlier processing.

[0055] S22, obtaining a sample set of customer portrait classification based on the customer data, and dividing the sample set into a training set sample and a test set sample of the customer portrait classification.

[0056] In some implementations, basic customer information features can be constructed based on customer data, customer tags can be generated, current resource information can be extracted, historical behavior features can be calculated, and a customer portrait classification sample set can be constructed based on basic customer information features, customer tags, current resource information, and historical behavior features.

[0057] Among them, basic customer information features include age, gender, region, etc.; customer tags include high value, potential churn, etc.; current resource information includes the type of package that has been ordered, contract period, activated resource status, etc.; historical behavior characteristics include the number of complaints, diagnosis frequency, and the probability of daily subsequent operation classification.

[0058] S23, build a customer portrait classification model based on the existing machine learning classification model, and train the customer portrait classification model.

[0059] The customer profile classification model can be built and trained by following the steps below:

[0060] S231, determine the algorithm of customer portrait classification model A.

[0061] The algorithm of the customer portrait classification model A can select a classification algorithm, where the classification algorithm includes random forest, XGBoost, and AdaBoost.

[0062] S232, determining the training strategy, including using the grid search method to tune the hyperparameters, and using the customer portrait classification training set samples to use the grid search method to tune the hyperparameters to train the customer portrait classification model.

[0063] S24, evaluate and iterate the customer portrait classification model A.

[0064] Model A is evaluated using the test set for customer portrait classification to obtain the performance indicators of model A on the customer portrait classification task.

[0065] Indicators such as precision, recall, and F1 value can be used as performance indicators of customer profile classification model A. Based on the evaluation results of these performance indicators, erroneous samples can be identified, and the features in the samples can be corrected or the parameters of the model can be adjusted.

[0066] In some implementations, the system may also use feedback data from the context of customer conversations to periodically retrain the customer profile classification model A to adapt to new customer data.

[0067] In an actual customer conversation, the customer's current order and background information are inferred using the trained customer profile classification model A to obtain a first prediction result. This first prediction result is then normalized and converted into a unified score vector, denoted as the first score vector. The first score vector indicates the score of the current customer profile classification.

[0068] For example, the first scoring vector is such as [upgrade cloud computer package: 0.8, solve network problem: 0.7, ...].

[0069] Next, we will introduce the fine-tuning large model based on the large language model.

[0070] Fine-tuning involves first training a large language model using self-supervised or unsupervised learning on a large dataset as a pre-trained model. The pre-trained model is then fine-tuned using the fine-tuning corpus for the target task. Typically, only a subset of the layers in the pre-trained model are fine-tuned, such as the last few layers or certain intermediate layers. During fine-tuning, the model is optimized using a backpropagation algorithm, resulting in improved performance on the target task.

[0071] If a pre-trained and fine-tuned baseline large model for a specific vertical domain class already exists, this baseline large model can be used directly as a pre-trained large model.

[0072] In some implementations, various open-source or closed-source customer service models can be used directly as pre-trained models. The customer service models are used to reason about the customer interaction context and output scores for customer intent classification or customer behavior prediction classification.

[0073] For example, Figure 3 This is a flow chart of the fine-tuning phase of the large language model provided in the embodiment of this application. Figure 3 As shown, the following steps are included:

[0074] S31, obtain fine-tuning corpus on a large-scale dataset and determine input-output pairs, where the input is the context of the customer interaction conversation and the output is the corresponding reasoning description of the pre-trained large model.

[0075] In some implementations, the behavioral dialogues of historical customer conversations can be used as fine-tuning information to construct a large amount of fine-tuning corpus for inference training.

[0076] For example, daily conversations are input into a pre-trained large model, and the corresponding reasoning description is output.

[0077] Based on the input customer description, the inference output is performed through the generation task of the pre-trained large model. The input customer description and the output inference response constitute the questions and answers of the training dialogue of the large model, thereby obtaining fine-tuning corpus.

[0078] For example, the pre-trained large model generates JSONL fine-tuning data in the following format based on the input customer description:

[0079]

[0080] Based on the fine-tuning corpus, we construct input-output pairs: the input is the customer description, and the output is the corresponding reasoning explanation. The reasoning explanation includes the intent classification score or the behavior prediction classification score.

[0081] S32, determining a fine-tuning method.

[0082] The first fine-tuning method uses supervised fine-tuning (SFT). SFT includes the following steps: S321, determining a pre-trained large model, which can be Qwen or Llama; S322, determining a fine-tuning task, which includes an intent classification task or a behavior prediction classification task; S323, using the prepared fine-tuning corpus to fine-tune all parameters of the pre-trained large model.

[0083] The second fine-tuning method uses a low-rank adapter. The fine-tuning steps of the low-rank adapter include: S324, selecting a basic large language model such as Qwen or LLaMA; S325, defining the rank of the adaptation layer, training only the newly added low-rank matrix, and keeping the basic pre-trained large model unchanged.

[0084] S33 uses the two fine-tuning methods provided by S32 to fine-tune the pre-trained large model respectively. During the fine-tuning process, parameters such as learning rate and batch size are set respectively, and appropriate optimizers such as AdamW are used to monitor training loss and verification indicators to obtain the first fine-tuned large model and the second fine-tuned large model.

[0085] S34, evaluate the performance of the first fine-tuned large model and the second fine-tuned large model, select the fine-tuned large model with the best performance, and deploy the best fine-tuned large model to the production environment.

[0086] In some embodiments, a test set can be extracted based on historical customer context interaction results, and the performance of the first fine-tuned large model obtained under the SFT method and the second fine-tuned large model obtained under the LoRA method can be evaluated based on the test set. The effects of SFT and LoRA are compared, the best solution is selected, and the best fine-tuned large model is deployed to the production environment.

[0087] It should be understood that the fine-tuned large model obtained in S34 in the embodiment of the present application performs semantic recognition on the customer context conversation, determines the customer's current emotional feedback, and feeds back a score.

[0088] In the production environment, the customer interaction context is inferred by fine-tuning the large model B to obtain a second prediction result. The second prediction result is normalized and converted into a unified score vector, which is recorded as the second score vector.

[0089] For example, the second rating vector is such as [consulting package: 0.4, complaint: 0.5, ...].

[0090] Next, we introduce the inference model C.

[0091] For example, Figure 4 This is a flowchart of building and training the inference model C provided in the embodiment of this application. Figure 4 As shown, the following steps are included:

[0092] S31, calling the second scoring vector output by the inference of the fine-tuned large model B in the production environment and the first scoring vector output by the inference of the customer portrait classification model A.

[0093] S32 can select step 3A dual-model fusion training or step 3B intelligent agent orchestration real-time reasoning as one of the two routes for the actual implementation of the inference large model C based on the customer's computing resources and actual development scenarios.

[0094] If the computing power is sufficient and the implementation time period is sufficient, step 3A can be selected to perform dual-model fusion training to determine the inference large model C.

[0095] For example, Figure 5 The flowchart of dual-model fusion training and inference of large model C provided in step 3A is shown in the figure. Figure 5 As shown in Figure 2, the process of merging the dual-model inference results is as follows:

[0096] Sa1, performing vector fusion on the first scoring vector and the second scoring vector, determining the weight of each scoring vector, and forming a comprehensive scoring vector c.

[0097] In some implementations, a linear weighting algorithm can be used to concatenate the first rating vector a and the second rating vector b to form a composite rating vector c. The linear weighting algorithm is as follows: c = w1·a + w2·b. w1 and w2 are linear weights, and their values ​​can be preset or adjusted based on experience. The first rating vector a is derived from the current customer profile classification score inferred by the customer profile model; the second rating vector b is derived from the current customer sentiment feedback score inferred by the fine-tuned large model based on conversational reasoning.

[0098] Exemplarily, a linear weighted algorithm is used to merge the first score vector a and the second score vector b, which is achieved by the following steps:

[0099] Sa11, assuming a = [0.2, 0.4, 0.5], where the example vector a represents the customer profile evaluation values ​​based on the three rating dimensions generated by the customer profile model, such as purchase intention score 0.2, budget margin score 0.4, and intention to replace existing orders score 0.5;

[0100] b = [0.3, 0.2, 0.4], where the example vector b represents the real-time scores for the three customer rating dimensions generated by the fine-tuned large model based on the current contextual conversation. For example, a score of 0.3 for purchase intent, a score of 0.2 for budget margin, and a score of 0.4 for intent to replace an existing order.

[0101] Sa12, determine the weights. Based on a pre-set experience, select the values ​​of weights w1 and w2. The weights satisfy w1 + w2 = 1, ensuring that the combined score vector c remains within the same range. If normalization is not required, this condition does not need to be met.

[0102] Sa13, calculate the comprehensive score vector c: perform weighted summation on each corresponding element of vectors a and b to obtain the comprehensive score vector c.

[0103] The formula for the linear weighted algorithm is as follows: c = w1·a + w2·b. The specific operation is as follows: Assume w1 = 0.6 and w2 = 0.4, which means that the customer profile weights 60% and the customer dialogue fine-tuning model weights 40%. Then: ci = 0.6·ai + 0.4·bi.

[0104] The specific calculation is as follows:

[0105] c1=0.6·0.2+0.4·0.3=0.12+0.12=0.24

[0106] c2=0.6·0.4+0.4·0.2=0.24+0.08=0.32

[0107] c3=0.6·0.5+0.4·0.4=0.30+0.16=0.46

[0108] Therefore, the combined vector c is: c = [0.24, 0.32, 0.46]. This process can be summarized as a weighted linear combination of the two vectors. Each element ci of vector c is the weighted sum of the rating vectors ai and bi generated by the two original models. Finally, the final vector c, derived from offline weighting, represents the customer's real-time scores for the three rating dimensions, such as a purchase intent score of 0.24, a budget margin score of 0.32, and an intent to replace an existing order score of 0.46.

[0109] In some implementations, a linear weighted average may be performed to combine the first scoring vector a and the second scoring vector b to determine the comprehensive scoring vector c.

[0110] In some embodiments, a multi-layer perceptron (MLP) algorithm can be used to assign weights to each score vector. An MLP is a feedforward neural network typically used to process fixed-size input vectors. The MLP learns the complex relationships between input features through a series of linear transformations and nonlinear activation functions. The MLP performs weighted fusion by learning a set of fixed weights.

[0111] The specific steps of using MLP to assign weights to each rating vector are as follows:

[0112] Sa31, concatenates multiple input rating vectors into a larger rating vector.

[0113] Sa32, processes the concatenated score vector through a multi-layer perceptron to learn the complex relationship of the input features.

[0114] Sa33, outputs the fused comprehensive score vector c through the last layer of MLP.

[0115] The scoring vector dimension is calculated at the input layer of the MLP; ReLU activation and Dropout regularization are performed at the hidden layer of the MLP; Softmax classification is performed at the output layer of the MLP to output the probability corresponding to each intent.

[0116] Sa4, train the inference model C based on the comprehensive score vector.

[0117] The comprehensive rating vector is used as input, and the inference model performs inference based on the comprehensive rating vector and outputs the customer's intention classification. The true intention of the historical data is used as the label, and the cross entropy loss is minimized through iterative calculation to obtain the trained inference model C.

[0118] In some embodiments, the large inference model performs inference based on the comprehensive score vector and outputs M customer intent categories. An intent decision strategy can be formulated. The intent decision strategy includes one or a combination of the following:

[0119] Select the top-K intents based on the M intent classifications output by the inference model C; filter out low-confidence intents from the top-K intents by setting a threshold (such as 0.6) to obtain multiple intents that meet the requirements; or process multiple intents in a set manner, such as sorting M intents by priority or executing M intents in parallel.

[0120] In step 3A, the reasoning capabilities of customer profile classification model A and fine-tuned large model B are integrated. This dual-model reasoning and prediction capabilities enhance the inference performance of customer intent. While step 3A incurs additional training costs for model fusion, it improves the inference efficiency of large inference model C in actual production deployments.

[0121] When computing power is limited and rapid implementation is required, you can choose step 3B to understand the intention based on the dual model and determine the inference large model C through the intelligent agent orchestration prompt word engineering.

[0122] For example, Figure 6 The schematic diagram of step 3B is provided for arranging prompt word engineering through the intelligent agent arrangement tool to realize the reasoning large model C. Step 3B includes adjusting the output of customer profile classification model A and fine-tuning large model B through intelligent agent arrangement and prompt word engineering in the actual reasoning scenario, and arranging the output weights of the intent classification of customer profile classification model A and fine-tuning large model B. Among them, prompt word engineering can be implemented through intelligent agent arrangement tools or large language model application development framework. Figure 6 As shown, step 3B includes the following steps:

[0123] Sb1, use BPMN to design a flowchart and define the key nodes of the task process of the reasoning model C.

[0124] In some implementations, an agent orchestration tool may be used to determine key nodes in the reasoning process of the large model C.

[0125] Exemplarily, the agent orchestration tool can be a self-developed process orchestration tool or a large-model application development framework, including Langchain, semantic-kernel, etc.

[0126] First, determine the key nodes of the inference task process of the large model C, including the customer's original intention input node, the inference call node of the customer portrait classification model A, the inference call node of the fine-tuning large model B, the weight merging decision node (based on the prompt word engineering configuration weight), the behavior execution node (intelligent agent tool call) and the multi-module combination process node.

[0127] It can be understood that the weighted merging decision node is also the prompt word engineering node, which describes the task requirements of the large inference model C through natural language description and performs merged reasoning. The weighted merging decision node is determined by the configuration weight based on the prompt word engineering.

[0128] A behavior execution node is an inference model agent node that can make external Assistant / Tool API calls. For example, "Query the weather forecast" is an action execution node for an agent tool. Because inference model C itself cannot generate the current real-time Beijing weather, the agent tool node relies on the inference model C to access weatherdt.com to obtain weather information, and then continue inference. For customer ordering scenarios, "Query customer business data" or "Initiate a customer order" are also external API capability calls.

[0129] Sb2, in order to merge the subsequent weighted decision points, converts the output results of the inference call node of the customer portrait classification model A into a descriptive statement that can be semantically understood by the fine-tuned large model B.

[0130] Through the prompt word engineering + intelligent agent node, the inference model C is based on the prompt words and is set according to the user's text requirements. It makes decision inferences and executes subsequent operations.

[0131] For example, the customer's purchase intention score is 7 (0 is the lowest and 10 is the highest); the customer's current experience feedback score is 4 (0 is very dissatisfied and 10 is very satisfied).

[0132] The multiple prompt word engineering nodes of the process orchestration constitute a process of weighted merging of the first scoring vector output by the customer portrait classification model A and the second scoring vector output by the fine-tuned large model B.

[0133] Sb3, determine the weight merging decision node and configure the prompt word project.

[0134] To handle complex and implicit intent, we configure prompt word engineering based on the output of customer portrait classification model A and the output of fine-tuned large model B, and adjust different preference weights for different business scenarios.

[0135] In some implementations, advanced prompt word engineering techniques such as thought chaining may be used to define the following: configuring a first prompt word based on an inference call node of a customer profile classification model, the first prompt word being used to guide the system to call the customer profile classification model to obtain a first score vector; the first score vector including a score result predicted for the customer profile classification;

[0136] A second prompt word is configured on the inference call node based on the fine-tuned large model. The second prompt word is used to guide the system to call the fine-tuned large model to obtain a second scoring vector. The second scoring vector includes a score result of sentiment reasoning based on the customer context.

[0137] A third prompt word is configured based on the weighted merging decision node. The third prompt word is used to guide the system to compile the output weights of the customer profile classification model and the intent classification of the fine-tuned large model based on the current different customer status information according to the first scoring vector and the second scoring vector, so as to determine the customer's intention tendency;

[0138] A fourth prompt word is configured based on the behavior execution node, where the fourth prompt word is used to guide the system to output a selectable operation function;

[0139] Based on the multi-module combination process node, the fifth prompt word is configured. This prompt word guides the system to combine the first, second, third, and fourth prompt words to compose the output weights of the customer profile classification model and fine-tune the intent classification of the large model. For example, a reference example of a prompt word project is as follows:

[0140]

[0141] Sb4, determine the intelligent agent behavior library, which is used to preset the tools and atomic capability services for registering business operation behaviors.

[0142] In some embodiments, the business operation behavior tools pre-registered in the agent behavior library include one or a combination of the following types:

[0143] Ordering, including calling the service activation API to create and change orders;

[0144] Business consulting, including intelligent agents in RAG scenarios, to search, enhance and generate business knowledge;

[0145] In some embodiments, the pre-registered atomic capability services in the agent behavior library include one or a combination of the following capabilities:

[0146] Graph question answering; and / or recommendation services, including calling external models of collaborative filtering algorithms (customer-based collaborative filtering) to perform customer preference recommendation tasks.

[0147] Sb5, based on the agent behavior library, selects the corresponding agent combination according to the intention, arranges the behavior weights and executes the behavior.

[0148] For example, customer portrait model A concludes that the customer is a person who is prone to complain, and the SCORE for "dissatisfaction" is high; fine-tuning large model B concludes that the probability SCORE of the customer complaining about "slow network performance" is high; then based on the weighted results of step 3A or the prompt word engineering combination results of step 3B, the conclusion is: [slow network 0.8, dissatisfied 0.9]; for customer portrait classification and the emotion "slow network + dissatisfied", the corresponding intelligent agent combination can be selected to solve the above results after customer evaluation. The intelligent agent combination includes a network diagnosis intelligent agent and a soothing intelligent agent. The network diagnosis intelligent agent is used for dedicated line configuration test calls, route detection calls, ARP detection interface calls, and link status detection calls; the soothing intelligent agent is used to generate analysis results + generate apology text + generate compensation suggestions.

[0149] Compared to the linear weighted or MLP neural network calculation used in step 3A, step 3B is relatively flexible. As long as the user adjusts the natural language text describing the requirements in the prompt word node of the process, it is easier to understand and operate for the end user.

[0150] Optionally, the large inference model C can also use integrated learning to vote on the results of decision reasoning of multiple inference chains and select the result with the most votes as the correct intention.

[0151] If the COT-SC (Self-Consistency with Chain of Thought) method is used in the prompt word engineering arrangement, it is equivalent to voting for the most correct results based on the results of decision reasoning using multiple reasoning chains. This method can effectively improve the accuracy of reasoning in the large reasoning model C.

[0152] For example, there are five large model reasoning routes, three of which lead to the conclusion that "the customer will buy Package A", and the other two are "buy Package B" and "not buy"; then the voting result is "the customer will buy Package A", and subsequent measures are taken.

[0153] The method provided in the embodiment of the present application also obtains customer feedback based on the behavioral results of the inference large model C, and continuously optimizes the customer portrait classification model A and fine-tunes the large model B.

[0154] Figure 7 Schematic diagram of result feedback and continuous optimization in the method provided in the embodiment of this application. Figure 7 As shown, the following steps are included:

[0155] S41, based on the contextual text of the customer interaction, obtain the results of the interactive behavior of the customer feedback, and use the results of the interactive behavior of the customer feedback as the feature weight retraining of the customer portrait classification model A in steps S23-S24 and the training corpus for fine-tuning and aligning the large model B in steps S31-S34, including feedback alignment corpus generation, customized fine-tuning or preference training corpus generation.

[0156] The customer profile classification model dynamically adjusts feature weights based on the training corpus to enhance or weaken the influence of the corresponding features.

[0157] Customer feedback includes both positive and negative feedback. The results of customer feedback interactions include: explicit action choices made on the front-end interface, such as clicking "satisfied" or "unsatisfied"; sentiment recognition of customer feedback responses through fine-tuning of the large model B; and context-guided feedback during the conversation, such as positive words ("very good") versus negative words ("no"); and feedback results obtained through monitoring front-end interaction patterns, such as rapid, continuous replies (likely indicating impatience) and prolonged silence (likely indicating confusion or dissatisfaction).

[0158] S42, fine-tuning the large model B to automatically generate new annotations based on the results of the customer feedback interaction behavior. This is achieved through the following steps:

[0159] S421, tracing back multiple rounds of contextual conversations;

[0160] S422, applying the SpanBERT algorithm of the fine-tuned large model B inference or NLP model to identify pronoun references and complete reference resolution;

[0161] S423 uses the ToD-BERT algorithm to track the conversation state. The ToD-BERT algorithm is a general BERT model for understanding user intent.

[0162] S424, automatically generating new annotations for the fine-tuned large model B;

[0163] Exemplarily, GPT4 or QWEN110B is used to capture the transition from “unsatisfied” -> “problem solved” -> “satisfied”.

[0164] S425, fine-tune the large model B. For feedback results that cannot be clearly labeled, mark them as "fuzzy" and record the feedback results.

[0165] Judgment records,For a small number of interaction result records marked as “fuzzy”, a manual intervention process is used to annotate feedback to improve the quality of feedback.

[0166] S43, retraining customer portrait classification model A, including data integration and feature expansion, feature quantification of feedback annotation, automatic feature engineering, AutoML automatic machine learning training, incremental learning and online updates.

[0167] S44: Fine-tune the pre-trained model using the conversation interaction results and the new annotations to obtain an updated fine-tuned model B, such as a language model or a customer service interaction model. This includes task-oriented fine-tuning and re-fine-tuning. Task-oriented fine-tuning includes the customer feedback dialogue after the feedback in this round has been annotated.

[0168] S45, retrain the inference model C to adapt to the updated customer portrait classification model A and fine-tune the large model B to improve the accuracy of inference.

[0169] The intent recognition method based on a large language model and a customer portrait classification model provided in the embodiment of the present application can be applied to a customer dialogue system based on a cloud platform. It receives customer input, selects an inference implementation route to identify customer intent based on the computing power and production environment conditions provided by the customer, and performs subsequent tasks; calls the large language model for semantic analysis, and at the same time obtains and updates the customer portrait from the database, calculates the final intent recognition result through a fusion algorithm, and then drives the execution of corresponding business logic, such as providing customized suggestions or triggering specific service processes.

[0170] Example 1

[0171] Embodiment 1 is an application of an intent recognition method based on a large language model and a customer portrait classification model in an operator cloud computer ordering customer service system.

[0172] During the chat with the AI customer service, the system inputs the customer's statement into the fine-tuned large model B based on the large language model through preprocessing for preliminary intent recognition, and combines the customer's historical purchase records, browsing habits and other information to build a customer portrait.

[0173] The reasoning large model C fuses these two parts of information to make the final intent determination, and the customer service system can accurately identify whether the user wants to consult products, feedback problems or perform after-sales service, etc. Intention, so as to provide accurate replies.

[0174] Figure 8 The flow engine framework for implementing the agent arrangement provided in Embodiment 1 of the present application is shown in the figure. Figure 8 As shown, through a flow engine framework for implementing the agent arrangement, the customer intent classification model A and the fine-tuned large model B perform real-time double-model reasoning on the customer's conversation, and obtain the customer's complete intent behavior, and then call the relevant follow-up business operation.

[0175] S81, the customer inputs the interaction information.

[0176] S82, query the customer's related orders and background information according to the customer ID, including whether the customer is due, new order, new account, customer type, etc.

[0177] S83, according to the customer's order status, account status and customer type, judge whether it is a new customer, new order or order expiration, etc.; perform customer portrait reasoning, (based on label reasoning), according to the customer's related information obtained by querying, perform customer portrait classification prediction, and perform ordering customer portrait classification prediction.

[0178] S84, call the customer portrait classification prediction API to obtain the customer portrait classification score.

[0179] In some embodiments, the customer portrait classification model can be called to generate a classification score (Score) of the customer's ordering intent according to the customer's existing package and customer background information.

[0180] Illustratively, an Adaboost classification model is called to perform customer portrait classification, and a classification score (Score) of the customer's ordering intent is generated according to the customer's existing package and customer background information.

[0181] S85, reason the customer's context interaction text.

[0182] A large model based on Qwen14B fine-tuned can be used for text reasoning service.

[0183] Here, the Qwen14B large model is selected for fine-tuning. The Qwen14B large model is a large model with fewer parameters (14B = 14 billion parameters), which has been evaluated to be sufficient to complete the efficient fine-tuning training task, and achieves the usability of intent classification under relatively less demand for computing power. The Qwen14B fine-tuning large model is selected to perform semantic recognition on the customer context dialogue, judge the current emotional feedback of the customer, and output the current emotional feedback score.

[0184] S86, based on the prompt word engineering application implemented on the inference large model C (Qwen70B model), to realize intent classification.

[0185] According to the output content of the two models, the prompt words are constructed and the weights are assigned: the classification score of the customer ordering intent generated by the customer portrait classification model and the current emotional feedback score output by the fine-tuned large language model are combined to arrange the prompt words and assign the weights.

[0186] It should be noted that the inference large model C selects the Qwen70B (70 billion parameters) large model with more parameters, because the large model with more parameters has better prompt word understanding and logical reasoning tasks, and can understand the thinking chain definition of the prompt engineering output by the double model, and perform high-quality inference.

[0187] In the embodiment 1 of the present application, the prompt word engineering helps the inference large model C to understand the judgment results generated by the previous double model through the prompt description of a thinking chain.

[0188] The inference large model C performs intent reasoning, according to the intent classification result of the prompt word engineering, and the customer portrait classification model, calls the corresponding tool capability API operation, realizes the follow-up task behavior of the classification result as follows:

[0189] Intent classification result A: customer service recommendation, such as customer service related API.

[0190] Intent classification result B: purchase recommendation, such as purchase recommendation API.

[0191] Intent classification result C: cloud computer business subscription, such as inference based on Owen70B large model.

[0192] Intent classification result D: other business recommendation, such as other recommendation API.

[0193] S87, based on the prompt word engineering, the operation result of different execution classification results of S86, feedback and interaction with the customer.

[0194] Based on the large model inference service C (based on Qwen70B), process customer interaction feedback.

[0195] S88, record customer feedback for evaluation.

[0196] Large models can be used for automatic evaluation and manual evaluation and recording of some results.

[0197] The embodiment of the present application provides an intent recognition method based on a large language model and a customer portrait classification model, and combines the multi-dimensional information processing capabilities of fine-tuning the large model and the customer portrait classification model to realize a real-time personalized intent recognition algorithm and model intent recognition reasoning. It has a self-learning and optimization feedback mechanism, provides optional dual-route reasoning implementation, and can be based on dual-model fusion training reasoning or real-time reasoning based on a process orchestration engine according to the customer's actual computing resources and production links.

[0198] The intent recognition method based on the large language model and the customer portrait classification model provided in the embodiment of the present application can greatly improve the accuracy and efficiency of intent recognition. By combining the two models for intent recognition, the fine-tuned large language model can better capture the sentiment and intent information of the semantic text in the customer context, thereby significantly improving the accuracy of intent recognition and reducing misjudgments and missed judgments. Through the pre-trained customer portrait classification model, the prediction and classification information based on the clear feature information in the customer context is improved. The combination of the two models improves personalized services and enhances customer experience, can better understand customer needs and preferences, realize personalized services, and thus improve customer experience.

[0199] The intent recognition method based on a large language model and a customer portrait classification model provided in the embodiments of the present application can provide more personalized services or recommendations that meet customer expectations, thereby enhancing customer satisfaction and loyalty.

[0200] An embodiment of the present application proposes an intention recognition system based on a large language model and a customer portrait classification model, based on the method provided in the above embodiment, including: an information acquisition module, used to obtain the customer's current relevant orders and background information based on the customer ID information; a customer portrait classification model, used to obtain a first scoring vector based on the relevant orders and background information; the first scoring vector is used to indicate the current customer portrait classification score; a fine-tuning large model, used to perform semantic recognition on the context of the input customer conversation to obtain a second scoring vector; a fusion module, used for the second scoring vector to indicate the customer's current emotional feedback score; the fine-tuning large model is obtained by fine-tuning the large language model; the first scoring vector and the second scoring vector are fused to obtain a comprehensive scoring vector; the comprehensive scoring vector is used to indicate different tendency weights in different business scenarios; an inference large model, used to perform inference based on the comprehensive scoring vector and output the customer's intention; an intention-driven module, used to drive the execution of corresponding business logic based on the customer's intention, and the corresponding business logic includes providing customized suggestions or triggering specific service processes.

[0201] The system's self-learning and optimization capabilities: The system has the ability to self-learn and optimize, and can adapt to changing customer needs, ensuring that the system can develop as customer behavior changes and market trends.

[0202] Multi-dimensional feedback capture uses implicit signals such as conversational features, sentiment vocabulary, and interaction patterns to fully understand customers' true feelings. This allows the system to capture subtle emotional changes even when customers are reluctant to express them explicitly.

[0203] Dynamic interaction adjustments: Based on real-time interpretation of customer sentiment and preferences, the system can dynamically adjust its inquiry style during the interaction, such as being concise and to the point for impatient customers and patiently guiding those who are confused. This adaptability greatly improves the customer experience.

[0204] The continuous evolution of the dual models, based on a continuous closed-loop semi-supervised learning process, completes the majority of feedback annotation through continuous feedback collection and automatic annotation by a large language model. For ambiguous samples that are difficult to judge, manual intervention is combined to improve annotation quality, and the training set is continuously expanded.

[0205] The customer portrait classification model dynamically adjusts feature weights based on each interaction. Through technologies such as gradient boosting trees and SHAP values, the system can accurately enhance or weaken the influence of certain features, making the portrait closer to the customer's actual status.

[0206] The continuous evolution of the fine-tuned large model is not just a one-time fine-tuning, but continuous optimization based on the results of conversation interactions and feedback annotations that occur continuously in production, so that the fine-tuned large model is not only optimized at the technical level, but also aligned with the behavioral standards of professional staff.

[0207] An embodiment of the present application provides an electronic device, comprising: at least one memory for storing programs; and at least one processor for executing the programs stored in the memory. When the programs stored in the memory are executed, the processor is used to execute the above-mentioned method for predicting ecological indicators.

[0208] An embodiment of the present application provides a computer storage medium, in which instructions are stored. When the instructions are executed on a computer, the computer executes the above-mentioned method for predicting ecological indicators.

[0209] It is understood that the processor in the embodiments of the present application may be a central processing unit (CPU), or may be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. The general-purpose processor may be a microprocessor or any conventional processor.

[0210] The method steps in the embodiments of the present application can be implemented by hardware or by a processor executing software instructions. The software instructions can be composed of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, mobile hard disks, CD-ROMs or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can be located in an ASIC.

[0211] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When software is used for implementation, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted via a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via a wired (e.g., coaxial cable, optical fiber, digital customer line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrations. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium (e.g., a solid state drive (SSD)).

[0212] It will be understood that the various numerical numbers involved in the embodiments of the present application are merely distinctions for the convenience of description and are not intended to limit the scope of the embodiments of the present application.

Claims

1. An intent recognition method based on a large language model and a customer profile classification model, characterized in that: The method comprises: Obtain the customer's current related orders and background information based on the customer ID information; The customer profile classification model obtains a first scoring vector based on the relevant orders and background information; the first scoring vector is used to indicate the current customer profile classification score; The fine-tuned large model performs semantic recognition on the context of the input customer conversation to obtain a second scoring vector; the second scoring vector is used to indicate the customer's current emotional feedback score; the fine-tuned large model is obtained by fine-tuning the large language model; fusing the first scoring vector and the second scoring vector to obtain a comprehensive scoring vector; the comprehensive scoring vector is used to indicate different tendency weights in different business scenarios; The inference model performs inference based on the comprehensive score vector and outputs the customer's intention; The corresponding business logic is driven to execute according to the customer's intention, and the corresponding business logic includes providing customized suggestions or triggering specific service processes.

2. The method according to claim 1, characterized in that The fusing the first scoring vector and the second scoring vector to obtain a comprehensive scoring vector includes: The first scoring vector and the second scoring vector are concatenated using a linear weighted algorithm to form the comprehensive scoring vector.

3. The method according to claim 1, characterized in that The fusing the first scoring vector and the second scoring vector to obtain a comprehensive scoring vector includes: The comprehensive scoring vector is obtained by weighted fusion of the first scoring vector and the second scoring vector by learning a set of fixed weights using a multilayer perceptron.

4. The method according to any one of claims 1 to 3, characterized in that The large inference model performs inference based on the comprehensive score vector and outputs the customer's intention, including: The customer has M intentions, and the top-K intentions with the highest scores are selected from the M intentions; or Filtering the low-confidence intents among the M intents by setting a threshold to obtain K intents that meet the requirements; or sorting the M intentions by priority; Or execute the M intentions in parallel.

5. The method according to claim 1, wherein The fusing the first scoring vector and the second scoring vector to obtain a comprehensive scoring vector includes: The output weights of the first scoring vector and the second scoring vector are arranged through intelligent agent arrangement and prompt word engineering.

6. The method according to claim 5, characterized in that The step of arranging the output weights of the first and second scoring vectors through agent arrangement and prompt word engineering includes: Determine the key nodes of the task flow of the large inference model, including: the customer's original intention input node, the inference call node of the customer portrait classification model, the inference call node of the fine-tuning large model, the weight merging decision node, the behavior execution node, and the multi-module combination process node; Convert the output of the inference call node of the customer portrait classification model into a description statement based on semantic understanding of the fine-tuned large model; The prompt word engineering is configured for the first scoring vector and the second scoring vector, and different tendency weights are adjusted in different business scenarios.

7. The method according to claim 1 or 5, characterized in that The driving and executing of corresponding business logic according to the customer intention includes: Determine an agent behavior library, wherein the agent behavior library pre-sets tools and / or atomic capability services for registering business operation behaviors; According to the customer intention, a corresponding agent combination is selected from the agent behavior library, and the agent combination is used to orchestrate and execute behaviors.

8. The method according to claim 1, characterized in that The method further comprises: Obtaining a result of the customer's feedback interaction behavior based on the context text of the customer interaction; Using the results of the interactive behavior of the customer feedback as training corpus for retraining the feature weights of the customer portrait classification model; The customer portrait classification model retrains the feature weights based on the training corpus; including data integration and feature expansion, feature quantification of feedback annotations, automatic feature engineering, AutoML automatic machine learning training, incremental learning and online updates.

9. The method according to claim 8, characterized in that The method further comprises: The fine-tuned large model automatically generates new annotations based on the results of the interactive behavior fed back by the customer; The conversation interaction results and new annotations are used as input-output pairs to continuously fine-tune the fine-tuned large model to align conversation and intent recognition.

10. An electronic device comprising: at least one memory for storing a program; At least one processor is used to execute the program stored in the memory, and when the program stored in the memory is executed, the processor is used to execute the method according to any one of claims 1 to 8.

Images