User intention recognition method and device, equipment and storage medium

Abstract

The invention discloses a user intention recognition method and device, equipment and a storage medium. According to the scheme, user multi-source data containing structured data and unstructured text data is obtained; and processing the structured data by using a gradient lifting decision tree model to obtain a first intention recognition result, and processing the unstructured text data by using a large language model to obtain a second intention recognition result. Determining a first weight and a second weight for fusion by combining the historical performance indexes of the two models and the business scene characteristics of the current user; and carrying out weighted fusion on the two recognition results based on the weights to obtain a final recognition result. According to the scheme, the value of the multi-source data is fully mined through the double-model parallel architecture, intelligent scene adaptation is realized through dynamic weight fusion, the accuracy of user intention recognition and scene adaptability are remarkably improved, and reliable technical support is provided for refined operation.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and data processing technology, and in particular to a user intent recognition method, apparatus, device and storage medium. Background Technology

[0002] With the continued growth in demand for intelligent services, accurately identifying user intent has become a core technological challenge for improving operational efficiency and user experience. Existing technological solutions, in addressing this challenge, typically rely on a single type of data or model architecture, which has inherent limitations.

[0003] One mainstream approach relies on structured user data to infer intent through machine learning models. While this method can handle explicit numerical and categorical information, it struggles to effectively interpret and utilize unstructured text generated during interactions such as conversations and emails. This results in the neglect of a significant amount of semantic information containing real-time intent, sentiment, and complex context, leaving the recognition results superficial and failing to deeply understand the user's true needs.

[0004] Another approach focuses on directly performing semantic analysis and intent extraction from unstructured text. While these methods have made continuous progress in parsing text content, their analysis process is often independent of structured background information such as the user's historical behavior and attribute characteristics. This isolated analysis results in a lack of consideration for the integration of long-term user preferences and stable profiles in intent judgment, making it difficult to guarantee the accuracy, stability, and personalization of the recognition results in complex and ever-changing real-world scenarios. Summary of the Invention

[0005] To address the aforementioned issues, this application provides a user intent recognition method, apparatus, device, and storage medium, with the aim of improving the accuracy and scenario adaptability of user intent recognition.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] A first aspect of this application provides a user intent recognition method, the method comprising:

[0008] Acquire multi-source data associated with users, including structured data and unstructured text data;

[0009] The structured data is processed using a gradient boosting decision tree model to obtain a first intent recognition result; and the unstructured text data is processed using a large language model to obtain a second intent recognition result.

[0010] Based on the historical performance metrics of the gradient boosting decision tree model and the large language model, and the current business scenario characteristics of the user, determine the first weight and the second weight for fusion.

[0011] Based on the first weight and the second weight, the first intent recognition result and the second intent recognition result are weighted and fused to obtain the final recognition result of the user intent.

[0012] In an optional implementation, before processing the structured data using a gradient boosting decision tree model to obtain a first intent recognition result, and processing the unstructured text data using a large language model to obtain a second intent recognition result, the user intent recognition method further includes:

[0013] The structured data is subjected to feature engineering to obtain structured features for inputting the gradient boosting decision tree model;

[0014] The unstructured text data is processed into text vectors to obtain text features for input into the large language model.

[0015] In an optional implementation, determining the first and second weights for fusion based on the historical performance metrics of the gradient boosting decision tree model and the large language model, and the current business scenario characteristics of the user, includes:

[0016] Based on the historical performance metrics of the gradient boosting decision tree model and the large language model, determine the first basic weight and the second basic weight;

[0017] Based on the characteristics of the business scenario, the first basic weight and the second basic weight are adjusted to obtain the first weight and the second weight.

[0018] In an optional implementation, the historical performance metrics include the AUC values ​​of the gradient boosting decision tree model and the large language model under historically similar tasks; determining the first basic weight and the second basic weight based on the historical performance metrics of the gradient boosting decision tree model and the large language model includes:

[0019] The AUC value is normalized using the Softmax function to obtain the first basic weight and the second basic weight.

[0020] In the optional implementation, the business scenario characteristics include one or more of customer value attributes, product lifecycle stages, and real-time user interaction behavior characteristics; the step of adjusting the first basic weight and the second basic weight according to the business scenario characteristics to obtain the first weight and the second weight includes:

[0021] If the business scenario characteristics indicate that the user is a high-value customer, a positive bias is added to the first basic weight;

[0022] If the business scenario characteristics indicate that the user is in the new product consultation stage or actively interacting, a positive bias is added to the second basic weight.

[0023] In an optional implementation, after obtaining the final recognition result of the user intent, the user intent recognition method further includes:

[0024] Based on the final identification result, a corresponding business response strategy is generated, which is used to indicate differentiated outreach to the user.

[0025] In an optional implementation, the method further includes:

[0026] Obtain intent recognition feedback data for the user;

[0027] Based on the feedback data, update the parameters of the gradient boosting decision tree model and the large language model, and / or optimize the determination process of the first weight and the second weight.

[0028] A second aspect of this application provides a user intent recognition device, the device comprising:

[0029] The data acquisition module is used to acquire multi-source data associated with the user, including structured data and unstructured text data.

[0030] The parallel processing module is used to process the structured data using a gradient boosting decision tree model to obtain a first intent recognition result; and to process the unstructured text data using a large language model to obtain a second intent recognition result.

[0031] The weight determination module is used to determine the first weight and the second weight for fusion based on the historical performance indicators of the gradient boosting decision tree model and the large language model and the current business scenario characteristics of the user.

[0032] The intent recognition module is used to perform weighted fusion of the first intent recognition result and the second intent recognition result based on the first weight and the second weight to obtain the final recognition result of the user intent.

[0033] A third aspect of this application provides a user intent recognition device, the device comprising: a processor and a memory.

[0034] The memory is used to store program code and transmit the program code to the processor;

[0035] The processor is configured to execute, according to instructions in the program code, the steps of the user intent recognition method described in any implementation of the first aspect.

[0036] A fourth aspect of this application provides a computer-readable storage medium for storing program code for performing the steps of the user intent recognition method described in any implementation of the first aspect.

[0037] Compared with the prior art, this application has the following beneficial effects:

[0038] In this application's technical solution, firstly, multi-source data associated with the user is acquired, including structured data and unstructured text data. Secondly, the structured data is processed using a gradient boosting decision tree model to obtain a first intent recognition result. Then, the unstructured text data is processed using a large language model to obtain a second intent recognition result. Next, based on the historical performance indicators of the gradient boosting decision tree model and the large language model, and the current business scenario characteristics of the user, a first weight and a second weight are determined for fusion. Finally, based on the first weight and the second weight, the first intent recognition result and the second intent recognition result are weighted and fused to obtain the final user intent recognition result. It is evident that this application, through a dual-model parallel processing architecture, simultaneously mines the quantitative value of structured data and the semantic value of unstructured text, solving the problem of insufficient data utilization in existing technologies. By dynamically adjusting the fusion weights based on the model's historical performance and business scenario characteristics, intelligent adaptation of the fusion strategy is achieved, significantly improving the accuracy and scenario adaptability of user intent recognition, providing reliable technical support for refined operations and precise business responses. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A flowchart of a user intent recognition method provided in an embodiment of this application;

[0041] Figure 2 This is a schematic diagram of the structure of a user intent recognition device provided in an embodiment of this application. Detailed Implementation

[0042] As described earlier, with the continued growth in demand for intelligent services, accurately identifying user intent has become a core technological challenge for improving operational efficiency and user experience. Existing technological solutions, in addressing this challenge, typically rely on a single type of data or model architecture, which has inherent limitations.

[0043] One mainstream approach relies on structured user data to infer intent through machine learning models. While this method can handle explicit numerical and categorical information, it struggles to effectively interpret and utilize unstructured text generated during interactions such as conversations and emails. This results in the neglect of a significant amount of semantic information containing real-time intent, sentiment, and complex context, leaving the recognition results superficial and failing to deeply understand the user's true needs.

[0044] Another approach focuses on directly performing semantic analysis and intent extraction from unstructured text. While these methods have made continuous progress in parsing text content, their analysis process is often independent of structured background information such as the user's historical behavior and attribute characteristics. This isolated analysis results in a lack of consideration for the integration of long-term user preferences and stable profiles in intent judgment, making it difficult to guarantee the accuracy, stability, and personalization of the recognition results in complex and ever-changing real-world scenarios.

[0045] To address the aforementioned problems, the inventors have proposed a user intent recognition method, apparatus, device, and storage medium after research.

[0046] First, multi-source data associated with the user is acquired, including structured data and unstructured text data. Second, the structured data is processed using a gradient boosting decision tree model to obtain a first intent recognition result. Then, the unstructured text data is processed using a large language model to obtain a second intent recognition result. Next, based on the historical performance indicators of the gradient boosting decision tree model and the large language model, and the current business scenario characteristics of the user, a first weight and a second weight are determined for fusion. Finally, based on the first weight and the second weight, the first intent recognition result and the second intent recognition result are weighted and fused to obtain the final user intent recognition result. It is evident that this application, through a dual-model parallel processing architecture, simultaneously mines the quantitative value of structured data and the semantic value of unstructured text, solving the problem of insufficient data utilization in existing technologies. By dynamically adjusting the fusion weights by combining historical model performance with business scenario characteristics, intelligent adaptation of the fusion strategy is achieved, significantly improving the accuracy and scenario adaptability of user intent recognition, and providing reliable technical support for refined operations and precise business responses.

[0047] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0048] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0049] See Figure 1 This figure is a flowchart of a user intent recognition method provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps:

[0050] S101. Obtain multi-source data associated with the user.

[0051] In this embodiment of the application, the multi-source data associated with the user includes structured data and unstructured text data.

[0052] Structured data refers to data with a fixed format that can be directly quantified and analyzed, including basic customer information (such as age, gender, occupation, and region), policy data (such as policy type, effective date, expiration date, and premium amount), transaction records (such as payment records and claims records), and behavioral data (such as platform login frequency, product browsing history, and operation behavior sequence).

[0053] Unstructured text data refers to data that exists in natural language without a fixed format, including customer service conversation records, email content, social media interaction information, customer feedback text, etc.

[0054] In one example implementation, structured user data is obtained from the insurance business system, including the user's age of 35, occupation as a company employee, critical illness insurance policy (effective date May 2020, annual premium of 8,000 yuan), 5 consecutive years of payment with no claims record, and logging into the insurance APP 2-3 times per month; unstructured text data is obtained from the customer service system, including the user's dialogue records with customer service such as "I want to know about the coverage of the new children's critical illness insurance" and "Can I consult about policy issues on WeChat at any time?"

[0055] This application embodiment comprehensively collects heterogeneous data from multiple sources, providing complete data support for subsequent parallel processing of dual models. This avoids the one-sidedness of intent recognition caused by single data, and ensures that user intent can be characterized from both quantitative and semantic features.

[0056] S102. The structured data is processed using a gradient boosting decision tree model to obtain the first intent recognition result; and the unstructured text data is processed using a large language model to obtain the second intent recognition result.

[0057] The gradient boosting decision tree model is an ensemble learning model built on the gradient boosting algorithm. It constructs multiple decision trees through multiple rounds of iterative training, using the prediction residuals of the previous model as the learning objective. It automatically selects important features and captures the nonlinear relationships between features, and has the advantage of efficiently processing structured data.

[0058] The large language model is a pre-trained language model based on the Transformer architecture. It has powerful natural language understanding and semantic parsing capabilities. Through domain-adaptive fine-tuning, it can be adapted to specific industry scenarios and can extract deep semantics, intent, and sentiment from unstructured text.

[0059] In this embodiment, the gradient boosting decision tree model is preferably the LightGBM model, which improves the efficiency and accuracy of structured data processing through optimization techniques such as histogram splitting and gradient one-sided sampling. The large language model adopts a pre-training combined with fine-tuning mode. After loading the general pre-trained model, it is fine-tuned using insurance industry professional terminology corpus and business scenario dialogue corpus to adapt it to the insurance customer intent recognition scenario.

[0060] In one example implementation, numerical features (such as age, premium amount, etc.) in structured data are standardized, and categorical features (such as policy type, occupation, etc.) are one-hot encoded. After being input into the LightGBM model, the model is trained through multiple rounds of iterations and outputs the first intent recognition result in the range of 0-1, such as 0.72. The higher the score, the stronger the user's intention to add the WeChat account.

[0061] Unstructured text data is cleaned, segmented, and embedded to obtain text feature vectors, which are then input into a fine-tuned large language model. Semantic features are extracted through multi-layer Transformer architecture encoding and self-attention mechanism. After processing by fully connected layers and Sigmoid activation function, the second intent recognition result in the 0-1 range is output, such as 0.86.

[0062] This application embodiment utilizes a dual-model parallel processing approach, leveraging the quantitative advantages of the gradient boosting decision tree model in processing structured data and the semantic advantages of the large language model in parsing unstructured text. This enables in-depth mining of the value of multi-source data, providing high-quality preliminary recognition results for subsequent accurate fusion.

[0063] S103. Based on the historical performance indicators of the gradient boosting decision tree model and the large language model and the current user's business scenario characteristics, determine the first weight and the second weight for fusion.

[0064] In this application embodiment, historical performance metrics refer to model evaluation metrics of the two models under similar historical business scenarios, such as AUC value, and may also include precision, recall, etc., to reflect the historical prediction performance of the model.

[0065] The current user's business scenario characteristics refer to the business scenario attributes related to the user, including customer value attributes (such as policy value, spending power), business stage characteristics (such as new product promotion period, regular operation period), and interaction behavior characteristics (such as online interaction frequency, whether new product keywords are mentioned), etc.

[0066] In one alternative implementation, the steps of determining the first weight and the second weight for fusion include:

[0067] Step 1: Determine the first and second basic weights based on the historical performance metrics of the gradient boosting decision tree model and the large language model.

[0068] Optionally, historical performance metrics include the AUC values ​​of the gradient boosting decision tree model and the large language model on historically similar tasks. The determination of the first and second basic weights specifically includes:

[0069] The AUC values ​​are normalized using the Softmax function to obtain the first and second basic weights.

[0070] In this embodiment of the application, the formula for calculating the first basic weight is:

[0071] ;

[0072] in, As the first basic weight, It is the natural constant, approximately equal to 2.71828. It is the AUC value of the gradient boosting decision tree model. It is the AUC value of the large language model.

[0073] The formula for calculating the second basic weight is:

[0074] ;

[0075] in, As the second basic weight, It is the natural constant, approximately equal to 2.71828. It is the AUC value of the gradient boosting decision tree model. It is the AUC value of the large language model.

[0076] In one example implementation, under historically similar tasks, if the AUC value of the gradient boosting decision tree model is 0.85 and that of the large language model is 0.82, the first basic weights are calculated. ≈0.507, second basic weight ≈0.493.

[0077] Step 2: Adjust the first basic weight and the second basic weight according to the characteristics of the business scenario to obtain the first weight and the second weight.

[0078] Optionally, the business scenario characteristics include one or more of the following: customer value attributes, product lifecycle stages, and real-time user interaction behavior characteristics. Adjustments to the first and second basic weights specifically include:

[0079] When business scenario characteristics indicate that users are high-value customers, structured data has higher predictive value, thus adding a positive bias to the first basic weight.

[0080] When business scenario characteristics indicate that users are in the new product consultation stage or have active interactions, the semantic information of unstructured text becomes more critical, thus adding a positive bias to the second basic weight.

[0081] In this embodiment, the bias adjustment amount is obtained through historical business data analysis, reflecting the actual business importance of each model under different scenarios.

[0082] In this embodiment of the application, the formula for calculating the adjusted first weight is as follows:

[0083] ;

[0084] in, As the first weight, As the first basic weight, As the second basic weight, The positive bias added to the first basic weight. The positive bias added to the second basic weight.

[0085] The adjusted formula for calculating the second weight is as follows:

[0086] ;

[0087] in, As the second weight, As the first basic weight, As the second basic weight, The positive bias added to the first basic weight. The positive bias added to the second basic weight.

[0088] In one example implementation, the first basic weight ≈0.507, second basic weight ≈0.493, for high-value customers with policy values ​​exceeding 500,000 yuan but with less than one interaction per month on average, the first basic weight is increased. A positive bias of 0.08 is applied, and the second base weight is increased. A positive bias of 0.01 is used to adjust the first weight. ≈0.54, second weight ≈0.46.

[0089] The embodiments of this application use a dynamic weight determination mechanism to ensure that the fusion weights match both the historical performance of the model and the current business scenario, thus ensuring that the fusion results are more in line with actual business needs.

[0090] S104. Based on the first weight and the second weight, the first intent recognition result and the second intent recognition result are weighted and fused to obtain the final recognition result of the user intent.

[0091] In this embodiment of the application, the final identification result of the user's intent refers to the user's intention score for adding WeChat, with a value range of 0-1. The higher the score, the stronger the user's intention to add WeChat.

[0092] In this embodiment, a weighted summation method is used for fusion calculation, and the calculation formula is as follows:

[0093]

[0094] in, The final result of user intent recognition. This is the result of the first intent recognition. For the result of second intent recognition, As the first weight, It is the second weight.

[0095] In one practical implementation, the first intent recognition result is obtained by combining the aforementioned example data. =0.72, second intent recognition result =0.86, first weight =0.54, second weight =0.46, the final recognition result is obtained: =0.54×0.72 + 0.46×0.86≈0.78, indicating that the user has a high intention to add WeChat.

[0096] This application embodiment organically combines the recognition results of the two models through weighted fusion, achieving complementary advantages, avoiding the limitations of a single model, and significantly improving the accuracy and reliability of user intent recognition.

[0097] This application's embodiments utilize a dual-model parallel processing architecture to simultaneously mine the quantitative value of structured data and the semantic value of unstructured text, solving the problem of insufficient data utilization in existing technologies. By dynamically adjusting the fusion weights based on the model's historical performance and business scenario characteristics, intelligent adaptation of the fusion strategy is achieved, significantly improving the accuracy of user intent recognition and scenario adaptability, and providing reliable technical support for refined operations and precise business responses.

[0098] In an optional implementation, to improve the quality of the model input data and ensure model prediction performance, the user intent recognition method further includes the following steps before step S102:

[0099] Step 1: Perform feature engineering on the structured data to obtain structured features for input into the gradient boosting decision tree model.

[0100] In this embodiment of the application, feature engineering processing includes two steps: data quality verification and feature transformation.

[0101] Data quality verification covers integrity verification (detecting missing values ​​in key fields and filling them with mean, median or business rules), consistency verification (verifying data logical relationships through the business rule engine, such as correcting logically conflicting data if the policy effective date is earlier than the expiration date), and reasonableness judgment (identifying abnormal values ​​based on statistical distribution, such as marking or removing values ​​that exceed the age range of 0-120 years).

[0102] Feature transformation includes standardization of numerical features (such as Z-score standardization) and encoding of categorical features (such as one-hot encoding).

[0103] In one practical implementation, Z-score standardization is performed on the user's age of 35 and annual premium of 8,000 yuan, and one-hot encoding is performed on the "critical illness insurance" policy type and the "company employee" occupation to generate a structured feature vector with a dimension of 12.

[0104] Step 2: Perform text vectorization on the unstructured text data to obtain text features for input into the large language model.

[0105] In this embodiment of the application, text vectorization processing includes multi-level cleaning and feature construction: first, basic text cleaning (removing special characters, HTML tags, and irrelevant punctuation) is performed, then deep semantic cleaning (identifying and processing duplicate content and irrelevant advertising information) is performed, followed by word segmentation and stop word removal, and finally, word embedding technology (such as Word2Vec and GloVe) is used to construct text feature vectors.

[0106] In one practical implementation, after cleaning the user's dialogue record "want to know the coverage of the new children's critical illness insurance", word segmentation is performed to obtain effective words such as "know", "new product", "children's critical illness insurance" and "coverage", which are then mapped into a 100-dimensional text feature vector through the Word2Vec model.

[0107] In an optional implementation, in order to convert the intent recognition result into an executable business action, after step S104, the user intent recognition method further includes:

[0108] Based on the final identification results, a corresponding business response strategy is generated, which is used to indicate differentiated outreach to users.

[0109] In this embodiment of the application, the final identification results are divided into three intention levels: high, medium and low by a preset threshold, such as thresholds of 0.8 and 0.5. A value greater than or equal to 0.8 is considered high intention, 0.5-0.8 is considered medium intention, and less than 0.5 is considered low intention. Different levels correspond to different outreach resource configurations, script templates and follow-up plans.

[0110] In one example implementation, a final recognition result of 0.80 indicates a high level of intent, leading to the generation of a business response strategy: prioritizing the allocation of senior consultant resources, and based on the key request extracted from the large language model, "inquiry about the coverage of the new children's critical illness insurance," generating a personalized message: "Hello! The coverage of the new children's critical illness insurance you previously inquired about has been compiled. We will add you via WeChat to provide you with one-on-one exclusive consultation. Is this convenient for you?" and suggesting direct contact via WeChat.

[0111] In another example implementation, if the final identification result is of moderate interest, the initial contact will be made via telephone or intelligent outbound calling, and subsequently incorporated into the content marketing nurturing system.

[0112] In another example implementation, if the final identification result is low interest, use general caring language and a targeted content push plan to gradually increase their interest.

[0113] In an alternative implementation, to enable continuous evolution of the model and weights and adapt to changes in the business environment, the user intent recognition method also includes:

[0114] Obtain user intent recognition feedback data. Based on the feedback data, update the parameters of the gradient boosting decision tree model and the large language model, and / or optimize the process of determining the first and second weights.

[0115] In this embodiment of the application, the feedback data refers to the actual WeChat addition results of the user (successful addition is a positive sample, and rejection is a negative sample). After accumulating a preset number (such as 1,000) of feedback data, the incremental learning process is started.

[0116] In one example implementation, after collecting 1000 user feedback data points, the decision tree structure parameters of the decision tree model are updated using the new samples. New customer dialogue records are used as fine-tuning corpus to incrementally fine-tune the large language model. The historical AUC values ​​of both models are recalculated, and the rules for adjusting the bias are calibrated to achieve continuous optimization of model performance and weighting strategies.

[0117] Based on the user intent recognition method provided in the foregoing embodiments, this application also provides a user intent recognition device. Figure 2 This is a schematic diagram of a user intent recognition device provided in an embodiment of this application. Figure 2 As shown, the user intent recognition device includes: a data acquisition module 201, a parallel processing module 202, a weight determination module 203, and an intent recognition module 204.

[0118] The data acquisition module 201 is used to acquire multi-source data associated with the user, including structured data and unstructured text data.

[0119] The parallel processing module 202 is used to process structured data using a gradient boosting decision tree model to obtain a first intent recognition result; and to process unstructured text data using a large language model to obtain a second intent recognition result.

[0120] The weight determination module 203 is used to determine the first weight and the second weight for fusion based on the historical performance indicators of the gradient boosting decision tree model and the large language model and the current user's business scenario characteristics.

[0121] The intent recognition module 204 is used to perform weighted fusion of the first intent recognition result and the second intent recognition result based on the first weight and the second weight to obtain the final recognition result of the user intent.

[0122] This embodiment of the application achieves comprehensive collection and in-depth utilization of multi-source data, efficient parallel processing of dual models, dynamic intelligent adaptation of fused weights, and accurate output of intent recognition results through the coordinated cooperation of data acquisition module 201, parallel processing module 202, weight determination module 203, and intent recognition module 204.

[0123] In an optional implementation, the user intent recognition device further includes a data processing module, used for:

[0124] Feature engineering is performed on structured data to obtain structured features that can be used as input to the gradient boosting decision tree model.

[0125] Unstructured text data is vectorized to obtain text features for input into a large language model.

[0126] In an optional implementation, the weight determination module 203 includes a determination unit and an adjustment unit.

[0127] The determination unit is used to determine the first and second basic weights based on the historical performance indicators of the gradient boosting decision tree model and the large language model.

[0128] The adjustment unit is used to adjust the first basic weight and the second basic weight according to the characteristics of the business scenario, so as to obtain the first weight and the second weight.

[0129] In the optional implementation, historical performance metrics include the AUC values ​​of the gradient boosting decision tree model and the large language model on historically similar tasks, and the determining unit is specifically used for:

[0130] The AUC values ​​are normalized using the Softmax function to obtain the first and second basic weights.

[0131] In the optional implementation, the business scenario characteristics include one or more of the following: customer value attributes, product lifecycle stages, and real-time user interaction behavior characteristics. The adjustment unit is specifically used for:

[0132] If the business scenario characteristics indicate that the user is a high-value customer, a positive bias is added to the first basic weight.

[0133] When the business scenario characteristics indicate that the user is in the new product consultation stage or has active interaction, a positive bias is added to the second basic weight.

[0134] In an optional implementation, the user intent recognition device further includes a response strategy generation module, which generates a corresponding business response strategy based on the final recognition result. The business response strategy is used to indicate differentiated outreach to the user.

[0135] In an optional implementation, the user intent recognition device further includes a feedback optimization module, used for:

[0136] Obtain user intent recognition feedback data;

[0137] Based on the feedback data, update the parameters of the gradient boosting decision tree model and the large language model, and / or optimize the process of determining the first and second weights.

[0138] Furthermore, embodiments of this application also provide a user intent recognition device, which includes a processor and a memory.

[0139] The memory is used to store program code and transmit the program code to the processor;

[0140] The processor is used to execute the steps of the user intent recognition method described in any of the above method embodiments according to the instructions in the program code.

[0141] Furthermore, embodiments of this application also provide a computer-readable storage medium for storing program code for performing the steps of the user intent recognition method described in any of the above method embodiments.

[0142] It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for the device and equipment embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the description of the method embodiments. The device and equipment embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components indicated as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the solution in this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0143] The above description is merely one specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for recognizing user intent, characterized in that, include: Acquire multi-source data associated with users, including structured data and unstructured text data; The structured data is processed using a gradient boosting decision tree model to obtain a first intent recognition result; and the unstructured text data is processed using a large language model to obtain a second intent recognition result. Based on the historical performance metrics of the gradient boosting decision tree model and the large language model, and the current business scenario characteristics of the user, determine the first weight and the second weight for fusion. Based on the first weight and the second weight, the first intent recognition result and the second intent recognition result are weighted and fused to obtain the final recognition result of the user intent.

2. The method according to claim 1, characterized in that, The structured data is processed using a gradient boosting decision tree model to obtain a first intent recognition result; Before processing the unstructured text data using a large language model to obtain the second intent recognition result, the method further includes: The structured data is subjected to feature engineering to obtain structured features for inputting the gradient boosting decision tree model; The unstructured text data is processed into text vectors to obtain text features for input into the large language model.

3. The method according to claim 1, characterized in that, The step of determining the first and second weights for fusion based on the historical performance metrics of the gradient boosting decision tree model and the large language model, and the current business scenario characteristics of the user, includes: Based on the historical performance metrics of the gradient boosting decision tree model and the large language model, determine the first basic weight and the second basic weight; Based on the characteristics of the business scenario, the first basic weight and the second basic weight are adjusted to obtain the first weight and the second weight.

4. The method according to claim 3, characterized in that, The historical performance metrics include the AUC values ​​of the gradient boosting decision tree model and the large language model on historically similar tasks; the determination of the first basic weight and the second basic weight based on the historical performance metrics of the gradient boosting decision tree model and the large language model includes: The AUC value is normalized using the Softmax function to obtain the first basic weight and the second basic weight.

5. The method according to claim 3, characterized in that, The business scenario characteristics include one or more of customer value attributes, product lifecycle stages, and real-time user interaction behavior characteristics; the adjustment of the first basic weight and the second basic weight according to the business scenario characteristics to obtain the first weight and the second weight includes: If the business scenario characteristics indicate that the user is a high-value customer, a positive bias is added to the first basic weight; If the business scenario characteristics indicate that the user is in the new product consultation stage or is actively interacting, a positive bias is added to the second basic weight.

6. The method according to claim 1, characterized in that, After obtaining the final recognition result of the user intent, the method further includes: Based on the final identification result, a corresponding business response strategy is generated, which is used to indicate differentiated outreach to the user.

7. The method according to claim 1, characterized in that, The method further includes: Obtain intent recognition feedback data for the user; Based on the feedback data, update the parameters of the gradient boosting decision tree model and the large language model, and / or optimize the determination process of the first weight and the second weight.

8. A user intent recognition device, characterized in that, include: The data acquisition module is used to acquire multi-source data associated with the user, including structured data and unstructured text data. The parallel processing module is used to process the structured data using a gradient boosting decision tree model to obtain a first intent recognition result; and to process the unstructured text data using a large language model to obtain a second intent recognition result. The weight determination module is used to determine the first weight and the second weight for fusion based on the historical performance indicators of the gradient boosting decision tree model and the large language model and the current business scenario characteristics of the user. The intent recognition module is used to perform weighted fusion of the first intent recognition result and the second intent recognition result based on the first weight and the second weight to obtain the final recognition result of the user intent.

9. A user intent recognition device, characterized in that, include: Processor and memory: The memory is used to store program code and transmit the program code to the processor; The processor is configured to execute the steps of the user intent recognition method according to any one of claims 1 to 7, based on the instructions in the program code.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program code for performing the steps of the user intent recognition method according to any one of claims 1 to 7.

Images