Data processing method, electronic device, storage medium and computer program product

Abstract

The invention discloses a data processing method, electronic equipment, a storage medium and a computer program product, and relates to the technical field of large model technology and data query. The method comprises the following steps: acquiring seed data; pre-labeling the seed data in a label propagation mode to obtain labeled data; performing data quality inspection on the labeled data to obtain training data; and training the initial intention classification model by adopting the training data to obtain a target intention classification model. The technical problems that the training cost of a model with a tool selection function is high, the efficiency is low, and the tool selection accuracy and the query accuracy of the model obtained through training are poor in the prior art are solved.

Description

Technical Field

[0001] This application relates to the fields of large model technology and data query technology, and more specifically, to a data processing method, electronic device, storage medium, and computer program product. Background Technology

[0002] In the field of large-scale model technology, agents are capable of understanding user input and responding accordingly. In particular, tool selection is a crucial function of agents; they select appropriate tools to respond to user needs by understanding the intent behind the user's input. The accuracy of understanding the user's intent in this process affects the accuracy of the agent's tool selection, and consequently, the agent's output performance.

[0003] In related technologies, intelligent agents typically include an intent recognition model. User input is mapped to the intent recognition model for intent recognition, and a language model (i.e., a student model) corresponding to the intelligent agent is trained using conventional model distillation training methods. Specifically, unlabeled data is randomly sampled, and the sampling results are used as prompts for the large language model (i.e., the teacher model) to produce output labels corresponding to the unlabeled data (e.g., if the unlabeled data is query data, the output label is the tool label corresponding to the query data). This output label is used as a supervision signal to guide the training of the student model. However, the above-mentioned solutions provided by related technologies have the following drawbacks: user queries that require tool selection account for a low proportion of all queries, and using the large language model to predict for each user query would lead to a large waste of resources, i.e., low model training efficiency and high training cost; the accuracy of directly prompting the large language model to select tools (i.e., output labels) is difficult to meet the accuracy requirements of intent recognition model training, resulting in poor model training performance.

[0004] Therefore, how to train intelligent agents with low cost and high efficiency, improve their intent recognition accuracy, and enhance tool selection accuracy has become one of the important technical problems in related fields. Currently, no effective solution has been proposed to address these problems. Summary of the Invention

[0005] This application provides a data processing method, electronic device, storage medium, and computer program product to at least solve the technical problems of high training cost, low efficiency, poor tool selection accuracy, and poor query accuracy of the trained model for models with tool selection functions.

[0006] According to one aspect of the embodiments of this application, a data processing method is provided, comprising: acquiring seed data, wherein the seed data includes sample data pairs, the sample data pairs including sample query data and sample intent categories; pre-labeling the seed data through a label propagation method to obtain labeled data; performing data quality checks on the labeled data to obtain training data; and training an initial intent classification model using the training data to obtain a target intent classification model, wherein the target intent classification model is used to classify the intent of target query data to select the target tool corresponding to the target query data, and the target intent classification model is also used to call the target tool to obtain the target query result.

[0007] According to another aspect of the embodiments of this application, a data processing method is also provided, including: acquiring target query data; classifying the target query data by intent using a target intent classification model to select the target tool corresponding to the target query data; and calling the target tool to obtain the target query result; wherein, the target intent classification model is obtained by training an initial intent classification model with training data, the training data is obtained by performing data quality checks on labeled data, the labeled data is obtained by pre-labeling seed data through label propagation, and the seed data includes sample data pairs, the sample data pairs including sample query data and sample intent categories.

[0008] According to another aspect of the embodiments of this application, a data processing method is also provided, including: acquiring weather query data; classifying the weather query data into intents using a target intent classification model to select a weather forecast tool corresponding to the weather query data; and calling the weather forecast tool to obtain weather query results; wherein the target intent classification model is obtained by training an initial intent classification model with training data, the training data is obtained by performing data quality checks on labeled data, the labeled data is obtained by pre-labeling seed data through label propagation, and the seed data includes sample data pairs, each sample data pair including sample query data and sample intent category.

[0009] According to another aspect of the embodiments of this application, a data processing method is also provided, comprising: obtaining a data processing request through a first application programming interface, wherein the request data carried in the data processing request includes: target query data; returning a data processing response through a second application programming interface, wherein the response data carried in the data processing response includes: target query results; wherein the target query results are obtained by classifying the target query data according to the intent of the target query data using a target intent classification model to select the target tool corresponding to the target query data and by calling the target tool, the target intent classification model is obtained by training an initial intent classification model using training data, the training data is obtained by performing data quality checks on labeled data, the labeled data is obtained by pre-labeling seed data using a label propagation method, the seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories.

[0010] According to another aspect of the embodiments of this application, a data processing method is also provided, comprising: acquiring a currently input data processing dialogue request, wherein the request data carried in the data processing dialogue request includes: target query data; responding to the data processing dialogue request, returning a data processing dialogue response, wherein the information carried in the data processing dialogue response includes: target query results; displaying the target query results in a graphical user interface; wherein the target query results are obtained by classifying the target query data using a target intent classification model to select the target tool corresponding to the target query data and by calling the target tool, the target intent classification model is obtained by training an initial intent classification model using training data, the training data is obtained by performing data quality checks on labeled data, the labeled data is obtained by pre-labeling seed data using a label propagation method, the seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories.

[0011] According to another aspect of the embodiments of this application, an electronic device is also provided, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the data processing method of any one of the above-mentioned methods when it runs.

[0012] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform any of the above-mentioned data processing methods.

[0013] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the data processing method described above.

[0014] In this embodiment, seed data is obtained, which includes sample data pairs, each including sample query data and sample intent category; the seed data is pre-labeled using a label propagation method to obtain labeled data; the labeled data undergoes a data quality check to obtain training data; the initial intent classification model is trained using the training data to obtain a target intent classification model, wherein the target intent classification model is used to classify the intent of the target query data to select the target tool corresponding to the target query data; the target intent classification model is also used to call the target tool to obtain the target query results.

[0015] It is noteworthy that this application embodiment combines a label propagation scheme and a data quality check scheme to pre-label sample data pairs and further construct training data for training the target intent classification model through data quality checks. This training data is of high quality. Based on this, the target intent classification model can more accurately identify the intent of target query data and select a more suitable target tool. Calling this target tool can obtain more accurate target query results. Furthermore, the method provided in this application embodiment avoids the large amount of redundant prediction and resource waste caused by using conventional model distillation training schemes. Therefore, this application embodiment achieves the goal of training a target intent classification model based on label propagation and data quality checks, thereby realizing the technical effects of accurately identifying query intent, selecting appropriate target tools, improving the accuracy of model query results, reducing model training costs, and improving training efficiency. This solves the technical problems of high training costs, low efficiency, and poor tool selection and query accuracy of related technologies for models with tool selection functions.

[0016] It is worth noting that the general description above and the detailed description that follow are merely for illustrative purposes and do not constitute a limitation on this application. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0018] Figure 1 This is a schematic diagram illustrating an application scenario of a data processing method according to an embodiment of this application;

[0019] Figure 2 This is a flowchart of a data processing method according to an embodiment of this application;

[0020] Figure 3 This is a flowchart of a data processing method according to an embodiment of this application;

[0021] Figure 4This is a flowchart of a data processing method according to an embodiment of this application;

[0022] Figure 5 This is a flowchart of a data processing method according to an embodiment of this application;

[0023] Figure 6 This is a flowchart of a data processing method according to an embodiment of this application;

[0024] Figure 7 This is a schematic diagram of the structure of a data processing apparatus according to an embodiment of this application;

[0025] Figure 8 This is a schematic diagram of the structure of a data processing apparatus according to an embodiment of this application;

[0026] Figure 9 This is a schematic diagram of the structure of a data processing apparatus according to an embodiment of this application;

[0027] Figure 10 This is a schematic diagram of the structure of a data processing apparatus according to an embodiment of this application;

[0028] Figure 11 This is a schematic diagram of the structure of a data processing apparatus according to an embodiment of this application;

[0029] Figure 12 This is a structural block diagram of a computer terminal according to an embodiment of this application. Detailed Implementation

[0030] 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 should fall within the scope of protection of the present application.

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

[0032] The technical solution provided in this application is mainly implemented using large-scale model technology. Here, "large-scale model" refers to a deep learning model with a massive number of parameters, typically containing hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of parameters. Large-scale models are also known as foundation models. They are pre-trained using large-scale unlabeled corpora to produce pre-trained models with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability. Examples include Large Language Models (LLMs) and multi-modal pre-training models.

[0033] It should be noted that, in practical applications, large models can be fine-tuned using a small number of samples to adapt them to different tasks. For example, large models can be widely used in Natural Language Processing (NLP), computer vision, and speech processing. Specifically, they can be applied to computer vision tasks such as Visual Question Answering (VQA), Image Captioning (IC), and Image Generation, as well as NLP tasks such as text-based sentiment classification, text summarization, and machine translation. Therefore, the main application scenarios for large models include, but are not limited to, digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design. In this embodiment, data processing using a large language model in a data query scenario is used as an example for explanation.

[0034] First, some nouns or terms that appear in the description of the embodiments of this application shall be interpreted as follows.

[0035] Large Language Model (LLM): refers to an artificial intelligence model that possesses a certain degree of general-purpose capabilities.

[0036] An intelligent agent is an entity in computer science, particularly artificial intelligence, capable of perceiving its environment and taking actions to influence it. An intelligent agent can be a software program, a robot, or other object with autonomous behavior. The design purpose of an intelligent agent is to enable it to perform tasks in a specific environment (such as playing games, navigating, or using recommendation systems) and to possess a certain degree of learning and adaptability. In this embodiment, the intelligent agent can be implemented based on a language model; for example, an intelligent agent that implements functions such as task planning and tool usage based on a language model.

[0037] Tool use function: refers to the ability of an intelligent agent to select the appropriate tool to use based on the specific task intent.

[0038] Intent-driven: refers to the process of determining the processing flow based on intent. For example, an agent selects to use a weather forecast tool based on the user's intent to query the weather, thereby determining the weather query process.

[0039] Clustering is an unsupervised learning method. It divides data points in a dataset into multiple groups, making data points within the same group similar to each other, while data points in different groups are significantly different.

[0040] Label propagation is a semi-supervised learning technique that uses a small amount of labeled data and a large amount of unlabeled data to improve classifier performance. In label propagation, known label information from the labeled data is "propagated" to the unlabeled data through the node connections in a neural network, inferring the possible clusters to which data points in the unlabeled data belong.

[0041] Model distillation is a machine learning technique involving a smaller model (also called a student model) and a larger model (also called a teacher model). The student model learns from the teacher model. In model distillation, the soft output (i.e., probability distribution rather than hard classification results) of the teacher model is used as additional training signals to train the student model. This allows the student model to capture the knowledge of the teacher model, reducing computational complexity and improving training efficiency. Model distillation is suitable for creating lightweight and efficient models in mobile devices or other resource-constrained scenarios.

[0042] According to an embodiment of this application, a data processing method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0043] Considering the large number of model parameters in large models and the limited computing resources of mobile terminals, the data processing method provided in this application embodiment can be applied to, for example, Figure 1 The application scenarios shown are not limited to these. In, for example... Figure 1 In the application scenario shown, the large model is deployed on server 10. Server 10 can connect to one or more client devices 20 via a local area network (LAN), wide area network (WAN), internet connection, or other types of data network. These client devices 20 may include, but are not limited to, smartphones, tablets, laptops, PDAs, personal computers, smart home devices, and in-vehicle devices. Client devices 20 can interact with users through a graphical user interface to access the large model, thereby implementing the method provided in this embodiment.

[0044] In this embodiment, the system consisting of a client device and a server can perform the following steps: the client device transmits seed data to the server, the seed data including sample data pairs, each sample data pair including sample query data and sample intent category; the server pre-labels the seed data using a label propagation method to obtain labeled data, performs data quality checks on the labeled data to obtain training data; the initial intent classification model is trained using the training data to obtain a target intent classification model, wherein the target intent classification model is used to classify the intent of the target query data to select the target tool corresponding to the target query data, and the target intent classification model is also used to call the target tool to obtain the target query result; further, the server returns the target intent classification model to the client device.

[0045] In another application scenario, the server provides the client device with an interface to use the aforementioned target intent classification model. The client device uses the interface to transmit the target query data to the target intent classification model on the server, and the server returns the target query results obtained by calling the target tool using the target intent classification model to the client device.

[0046] Under the aforementioned operating environment, this application provides the following: Figure 2 The data processing method shown. Figure 2 This is a flowchart of a data processing method according to an embodiment of this application, such as... Figure 2 As shown, the data processing method includes the following steps S21 to S24.

[0047] Step S21: Obtain seed data, wherein the seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories.

[0048] The data processing method provided in this application embodiment can be used to construct training data in a preset application scenario by combining tag propagation technology and data quality inspection technology, and then train a target intent classification model. The target intent classification model can be used to drive an intelligent agent (such as a chatbot). The intelligent agent has strong performance in data query and data processing tasks in the preset application scenario. Specifically, the intelligent agent can accurately understand the query intent corresponding to the target query data, and then select a suitable target tool and call the target tool to obtain accurate target query results.

[0049] The aforementioned pre-defined application scenarios may include, but are not limited to, scenarios involving the use of intelligent agents for data querying and processing in fields such as e-commerce, education, healthcare, conferencing, social networks, financial products, logistics, and navigation. For example, these pre-defined application scenarios may include, but are not limited to: intelligent dialogue scenarios (such as intelligent customer service, intelligent voice assistants, and intelligent enterprise services), intelligent recommendation scenarios (such as intelligent search, personalized recommendations, shopping recommendations, and travel recommendations), content generation scenarios (such as intelligent writing and intelligent reading comprehension), online teaching scenarios (such as online learning assistants, course searches, and homework assistance), and healthcare scenarios (such as medical consultations and drug searches).

[0050] The seed data mentioned above can be a set of pre-selected sample data pairs, each pair including sample query data (also called seed query) and sample intent category. The sample query data can be query questions or requests that users might raise while using the agent driven by the target intent classification model. The sample intent category is the intent category corresponding to the sample query data. The sample intent categories in the sample data pairs can be obtained by technicians or other annotation tools through high-precision annotation of the sample query data. This seed data serves as a crucial foundation for the training process of the target intent classification model and can guide the correct direction of subsequent data annotation.

[0051] In one exemplary application scenario, technicians, based on a deep understanding of user needs, can select and label a series of representative query data. This query data and its corresponding intent category constitute seed data. For example, the agent development team might collect and label query data such as "What's the weather like today?" or "Will it rain tomorrow?" The query intent corresponding to this query data is labeled as "weather query".

[0052] Step S22: Pre-label the seed data using a label propagation method to obtain labeled data.

[0053] The aforementioned label propagation method can be a semi-supervised learning technique. Specifically, in this method, known label information is propagated among similar data in the dataset to add labels to unlabeled data (i.e., pre-labeling), resulting in labeled data. This labeled data can be a query dataset with intent categories obtained after pre-labeling the seed data using the label propagation method.

[0054] For example, in the scenario of intelligent question-answering assistants, the implementation of pre-labeling seed data to obtain labeled data through label propagation can include: using a clustering algorithm to calculate the similarity between the seed data and a large-scale unlabeled historical query data, forming multiple clusters. Then, the historical query data in these clusters will be automatically labeled with the same intent category as the seed data, resulting in the aforementioned labeled data. For instance, if "Beijing weather" is labeled with the intent "weather query," then similar historical query data such as "Beijing weather forecast" and "Is it hot in Beijing today?" will also be pre-labeled with the intent "weather query."

[0055] By pre-labeling seed data through label propagation, we can obtain labeled data, which can expand the scale of seed data and provide richer training data for subsequent training of target intent recognition models.

[0056] Step S23: Perform a data quality check on the labeled data to obtain the training data.

[0057] To ensure the accuracy and reliability of the labeled data, a pre-set high-precision mechanism is used to perform data quality checks on the pre-labeled data, identifying and correcting erroneous labels to obtain training data. For example, this pre-set high-precision mechanism could be a mechanism for semantic analysis and intent checking using a pre-trained large language model. The training data can be labeled data that has been confirmed to be accurate after data quality checks.

[0058] For example, in the scenario of intelligent question-answering assistants, a large language model is used to check the data quality of labeled data. This large language model is regarded as a "data quality inspector". The large language model will determine whether each query in the labeled data accurately expresses the pre-labeled intent. For example, during the data quality check, if a query is pre-labeled as "weather query" but the large language model believes that it actually expresses the intent of "news query", then the label of this query will be corrected to "news query".

[0059] By using labeled data to perform data quality checks to obtain training data, the accuracy of the training data can be ensured, the data quality can be improved, and more reliable data support can be provided for the training of the target intent classification model.

[0060] Step S24: Train the initial intent classification model using training data to obtain the target intent classification model. The target intent classification model is used to classify the intent of the target query data to select the target tool corresponding to the target query data. The target intent classification model is also used to call the target tool to obtain the target query results.

[0061] Before training begins, the initial intent classification model mentioned above is a pre-designed, untuned small model with a small number of parameters; that is, a lightweight model. For example, the initial intent classification model could be a Bidirectional Encoder Representations from Transformers (BERT) model.

[0062] During the iterative training of the initial intent classification model using training data, the initial intent classification model will gradually become more intelligent, able to accurately identify the user's query intent and select appropriate tools accordingly, until the output of the model in the current training meets the preset conditions, thus obtaining the target intent classification model.

[0063] The training data mentioned above was obtained by checking the amount of data in the large language model (which can be used as the large model in model distillation, i.e., the teacher model). In the process of training the initial intent classification model (which can be used as the small model in model distillation, i.e., the student model) with this training data, the initial intent classification model can learn the knowledge of the large language model and gradually become the target intent classification model that can quickly and accurately classify user queries.

[0064] Furthermore, the accurate intent classification capability of the target intent classification model, combined with the predefined "intent-tool" correspondence, can determine the target tool that matches the target query data, call the target tool to perform query processing, and obtain the target query result corresponding to the target query data.

[0065] Through steps S21 to S24, the data processing method described above efficiently constructs an accurate target intent classification model (a lightweight small model), significantly reducing the cost and latency of model training. The seed data selection and label propagation pre-labeling steps effectively reduce the dependence of the small model training process on the large model, avoiding resource waste. The iterative processes of data quality checking and model distillation training ensure the accuracy and reliability of the model, improving the final model's performance in practical applications. Especially in the application scenario of agent tool selection, the embodiments of this application can help the agent more accurately understand the user's real needs and select the most suitable tool to respond, thereby significantly improving user satisfaction and agent efficiency.

[0066] In this embodiment, seed data is obtained, including sample data pairs, each containing sample query data and sample intent categories. The seed data is pre-labeled using a label propagation method to obtain labeled data. The labeled data undergoes a data quality check to obtain training data. The initial intent classification model is trained using the training data to obtain a target intent classification model. This target intent classification model is used to classify the intent of the target query data to select the corresponding target tool. It is also used to invoke the target tool to obtain the target query results. This embodiment combines a label propagation scheme and a data quality check scheme. Pre-labeling the sample data pairs and further using data quality checks constructs training data for training the target intent classification model. This training data is of high quality, enabling the target intent classification model to more accurately identify the intent of the target query data and select a more suitable target tool. Invoking the target tool yields more accurate target query results. Furthermore, the method provided in this embodiment avoids the significant redundant predictions and resource waste associated with conventional model distillation training schemes. As described above, the embodiments of this application achieve the goal of training a target intent classification model based on label propagation and data quality inspection, thereby realizing the technical effects of selecting appropriate target tools by accurately identifying query intent, improving the accuracy of model query results, reducing model training costs and improving training efficiency. This solves the technical problems of high training costs, low efficiency and poor tool selection accuracy and query accuracy of the trained models in related technologies.

[0067] In an optional embodiment, in step S22, the seed data is pre-labeled using a label propagation method to obtain labeled data, including the following method steps:

[0068] Step S221: Obtain historical query data, wherein the historical query data is unlabeled real query data generated within a historical time period;

[0069] Step S222: Perform cluster analysis on the historical query data and sample query data to obtain the clustering results;

[0070] Step S223: Perform label propagation on the clustering results to obtain labeled data.

[0071] The aforementioned historical query data refers to real, unlabeled user query data generated within a historical time period in the context of intelligent agent applications. This historical query data can record actual interactions between the tea garden, users, and intelligent agents, such as user searches, questions, or requests.

[0072] In one exemplary application scenario, during the operation of the intelligent agent, the system records the interaction data between the user and the intelligent agent (such as the user's entered query statements). Furthermore, historical query data is extracted from this interaction data as the basis for subsequent clustering and label propagation. For example, in an intelligent customer service system, a user might query "I want to know my order status" or "How do I apply for a refund?" These user queries, which are not manually labeled, can be extracted as historical query data.

[0073] The sample query data mentioned above includes tagged query data, meaning that the sample query data can include query text and corresponding intent categories.

[0074] The clustering analysis described above can be an unsupervised learning technique used to divide a dataset (here, including historical query data and sample query data) into multiple clusters. Data within the same cluster are similar to each other, while data between different clusters differ significantly. Based on this, the clustering results described above represent the data grouping results obtained after clustering analysis, with each group (i.e., cluster) representing a class of similar query data.

[0075] The clustering results described above will be used for label propagation. Specifically, the intent category label carried by the sample query data within each cluster will be "propagated" to the historical query data in the same cluster, thereby adding predicted intent category labels to these historical query data and forming labeled data.

[0076] For example, if a cluster contains multiple sample queries corresponding to the intent category "query order status", then the historical query data in that cluster will be labeled with the intent category "query order status". Thus, influenced by the seed data (sample query data), a large amount of unlabeled historical query data is initially labeled with intent categories.

[0077] Through steps S221 to S223 described above, the technical solution provided in this application embodiment can obtain intent-labeled data from massive historical query data at low cost, and then use it to train an intent classification model. This method effectively improves the efficiency of data labeling, reduces the dependence on large pre-trained models for direct intent classification, and thus saves computing resources. Simultaneously, through cluster analysis, data distribution can be naturally controlled, ensuring that the final training dataset covers various intent categories, which helps improve the model's generalization ability. Label propagation, while ensuring data volume, maintains a high labeling accuracy, providing a high-quality data foundation for subsequent model training and distillation. Furthermore, the cluster-based label propagation process in this application embodiment helps to build an efficient and accurate intent classification model, significantly enhancing the agent's ability in tool selection and task response, and optimizing the user experience.

[0078] In an optional embodiment, in step 222, cluster analysis is performed on the historical query data and the sample query data to obtain the clustering results, including the following method steps:

[0079] Step S2221: Vectorize the historical query data to obtain the first vector, and vectorize the sample query data to obtain the second vector;

[0080] Step S2222: Calculate the similarity between the first vector and the second vector based on a preset threshold condition to obtain the target calculation result;

[0081] Step S2223: Use a preset clustering algorithm to perform cluster analysis on the target calculation results to obtain the clustering results.

[0082] The vectorization process described above can be the process of converting text-based query data into numerical vectors. Vectorization facilitates machine understanding and calculation of text similarity. The first vector above is a numerical vector representation of historical query data obtained through vectorization. The second vector above is a numerical vector representation of sample query data obtained through vectorization.

[0083] In an exemplary application scenario, a text vectorization model is used to convert collected historical query data into numerical vectors, resulting in a first vector; a text vectorization model is also used to convert sample query data labeled with intent categories into numerical vectors, resulting in a second vector. The aforementioned text vectorization model can be a pre-trained text embedding model, such as the BERT model or a Generative Pre-trained Transformer (GPT) model. This text vectorization model can capture the semantic features of the query text in both historical and sample query data and map these semantic features into a high-dimensional numerical space.

[0084] The aforementioned similarity calculation refers to the method of evaluating the degree of similarity between two vectors by calculating the distance between them in a numerical vector space (such as cosine similarity, Euclidean distance, etc.).

[0085] The aforementioned preset threshold condition is used to determine whether two vectors are sufficiently similar after similarity calculation. The preset threshold condition is usually determined based on a threshold; if the similarity between two vectors is higher than this threshold, the two vectors are considered similar.

[0086] Furthermore, based on the aforementioned preset threshold conditions, the similarity between the first vector and the second vector is calculated to obtain the target calculation result. This target calculation result is the similarity calculation result between the historical query data vector and the sample query data vector, and can serve as an important basis for cluster analysis. The target calculation result can include the similarity values ​​between all historical query data and the sample query data.

[0087] In an exemplary application scenario, the system calculates the similarity between the first vector corresponding to each user query in the historical query data and the second vector corresponding to the sample query data, based on a preset similarity threshold, to obtain the target calculation result. For example, if the first vector corresponding to the historical query "What will the weather be like in Beijing tomorrow?" and the second vector corresponding to the sample query "What will the weather be like in Beijing tomorrow?" have a similarity in the numerical vector space higher than the similarity threshold, then the historical query is considered to express an intent similar to the sample query.

[0088] The aforementioned pre-defined clustering algorithms may include, but are not limited to, K-means clustering, density-based clustering (DBSCAN), hierarchical clustering, etc. The pre-defined clustering algorithms are used to perform cluster analysis on the target calculation results. Specifically, this involves a deep analysis of the target calculation results, using clustering algorithms to divide data points (which may include each query data point in historical query data and sample query data) into multiple clusters, thus obtaining the clustering results. The clustering results not only visually demonstrate the distribution of historical query data in the intent space but also provide a clear basis for tag propagation.

[0089] Through steps S2221 to S2223, the technical method provided in this application embodiment can effectively mine data with similar intent to sample query data from massive unlabeled historical query data and classify them through cluster analysis. This process greatly reduces the reliance on direct labeling of large pre-trained models, lowers costs, and ensures high data quality by setting reasonable similarity thresholds and selecting appropriate clustering algorithms. The clustering results provide a foundation for further label propagation and model training, enabling the agent to more accurately select and invoke tools based on the user's query intent, ultimately improving the user service experience and the agent's efficiency. The method provided in this application embodiment can be applied not only to intelligent question-answering assistants but also to various scenarios that require understanding and responding to users' natural language queries, such as intelligent customer service, voice assistants, and search recommendation engines.

[0090] In one optional embodiment, the clustering category corresponding to the clustering result is determined based on a preset number of intentions.

[0091] In the above optional embodiments, the selection of the number of cluster centers is also an important factor in the clustering analysis of the target calculation results. In this embodiment, the number of cluster categories (i.e., the number of clusters) in the clustering results can be determined according to the preset intention number. For example, according to the application scenario requirements, the preset intention number is determined to be N, then the number of cluster categories in the clustering analysis process can be set to 2N+1.

[0092] In the scenario of an intelligent question-answering assistant, suppose the system presets the number of intents to be identified as N, which may include common user query intents such as "weather query," "news retrieval," and "movie recommendation." During cluster analysis, the system will determine the number of cluster categories in the clustering results based on this preset number of intents. To ensure that each intent category has sufficient representativeness and the clusters are distinguishable, the system typically sets the number of cluster categories to 2N+1, which includes: N cluster categories that match each of the N intent categories, N cluster categories that do not match each of the N intent categories (but may be similar), and one "other" category (such as abnormal or unidentified queries).

[0093] For example, if the preset number of intents N is 5, the number of cluster categories could be set to 11: 5 intent categories, 5 additional categories, and 1 other category. This setting ensures that each intent category has at least one corresponding cluster (i.e., intent category). Meanwhile, the additional clusters (i.e., additional categories) can be used to distinguish queries with similar but different intent categories. For instance, if the preset intents include "weather query," the query data "Beijing weather forecast" might be categorized into the "weather query" cluster (one of the intent categories), while the query data "Beijing air quality" might be categorized into the "environmental information query" cluster (one of the additional categories), thereby improving the accuracy of subsequent label propagation and the quality of model training.

[0094] By introducing a strategy that determines the cluster categories corresponding to the clustering results based on a preset number of intents, this embodiment of the application can effectively classify historical query data, ensuring that each intent category has sufficient representative data to enter the model distillation process. This strategy helps balance the distribution of training data, avoids intent recognition bias caused by improper clustering numbers, and improves the model's generalization ability and accuracy.

[0095] Specifically, in the scenario of agent tool selection, introducing the above strategies can ensure that the model can cover a wide range of diverse user query intents, improving the accuracy and response speed of the agent in selecting tools when faced with complex queries. Simultaneously, by setting an appropriate number of cluster categories, the system can better control the quality and quantity of training data, thereby reducing the computational cost of model training and achieving efficient, low-latency intent classification model construction, providing users with a smoother and more satisfactory agent service experience.

[0096] In an optional embodiment, the preset threshold condition includes: a multi-level similarity threshold. In step S2222, the similarity between the first vector and the second vector is calculated based on the preset threshold condition to obtain the target calculation result, including the following method steps:

[0097] Step S2224: Select the target similarity threshold from the multi-level similarity thresholds;

[0098] Step S2225: Calculate the similarity between the first vector and the second vector using a preset vector similarity calculation method to obtain the initial calculation result;

[0099] Step S2226: Compare the initial calculation result with the target similarity threshold to obtain the comparison result;

[0100] Step S2227: Based on the comparison results, the initial calculation results are filtered to obtain the target calculation results.

[0101] The aforementioned multi-level similarity thresholds include multiple thresholds used to measure the similarity between the first vector and the second vector at different levels, thereby adapting to the query data matching needs of different intent categories. During the similarity calculation process, the target similarity threshold to be used is first selected from the multi-level similarity thresholds.

[0102] For example, in the scenario of an intelligent question-answering assistant, to achieve fine-grained control over user query intent, the system can pre-set multiple similarity thresholds: 0.85, 0.9, and 0.95, each corresponding to different similarity matching levels. When processing a user query "Beijing's weather for the next week" (corresponding to the first vector), the system may choose a higher target similarity threshold (such as 0.95) to ensure a sufficiently high match with the sample query "query a week's weather forecast" (corresponding to the second vector), thereby avoiding misclassifying queries with sufficiently similar but not completely matching intents.

[0103] The aforementioned preset vector similarity calculation methods may include, but are not limited to, cosine similarity calculation methods and Euclidean distance calculation methods. Using these preset vector similarity calculation methods, the similarity between historical query data (corresponding to the first vector) and sample query data (corresponding to the second vector) is evaluated, generating initial calculation results. Taking cosine similarity as an example, the system calculates the cosine value of the angle between the two vectors; the closer the value is to 1, the more similar the two vectors are, meaning the more similar the query intentions.

[0104] Furthermore, taking the scenario of an intelligent question-answering assistant as an example, when comparing the initial calculation result with the target similarity threshold, if the similarity between the first vector and the second vector calculated using cosine similarity is 0.92, and the selected target similarity threshold is 0.9, then the comparison result shows that the similarity between the first vector and the second vector exceeds the target similarity threshold, which means that the intentions expressed by the historical query data and the sample query data are sufficiently similar.

[0105] Furthermore, based on the comparison results, data pairs with similarity values ​​higher than or equal to the target similarity threshold are selected from the initial calculation results to form the target calculation results. For example, for the similarity calculation results of all historical query data and sample query data, the system will only retain data pairs with similarity values ​​higher than or equal to 0.9, and these data pairs will serve as candidates for subsequent label propagation.

[0106] By introducing multi-level similarity thresholds and filtering mechanisms, the embodiments of this application can more precisely control the similarity matching process between historical query data and sample query data, thereby ensuring the accuracy and efficiency of label propagation. This process can avoid misclassifying semantically dissimilar queries into the same intent category, reducing unnecessary data noise and improving the data quality for model training.

[0107] In particular, in the scenario of agent tool selection, the solution provided in this application can significantly improve the accuracy of the agent's understanding of user intent, thereby more accurately selecting and calling tools to meet user needs and optimize user experience. Simultaneously, by effectively filtering data, it reduces the computational resources required for subsequent model training, lowering development and operating costs. This filtering and matching strategy based on multi-level similarity thresholds has significant technical value in improving the agent's tool selection capabilities, especially when handling complex and diverse user queries.

[0108] In an optional embodiment, in step S223, label propagation is performed on the clustering results to obtain labeled data, including the following method steps:

[0109] Step S2231: Determine the intent category of the multiple clusters contained in the clustering results;

[0110] Step S2232: Label the data categories within the same cluster in multiple clusters as intent categories to obtain labeled data.

[0111] The aforementioned clusters refer to multiple data groups obtained by analyzing historical and sample query data using clustering algorithms. The data within each cluster are similar to each other in the data vector space.

[0112] Taking an intelligent question-answering assistant scenario as an example, after dividing historical query data into multiple clusters using a clustering algorithm, the intent category corresponding to each cluster is determined. Furthermore, all historical query data within each cluster is uniformly labeled with the intent category corresponding to that cluster. Thus, the previously unlabeled historical query data now obtains preliminary intent classification labels, becoming labeled data that can be used for model training. For example, in the "weather query" cluster of the intelligent question-answering assistant, all unlabeled queries within this cluster, such as "Beijing tomorrow's weather" and "Shanghai weekly weather forecast," will be labeled with the "weather query" intent category.

[0113] Through steps S2231 and S2232, the technical method provided in this application embodiment can significantly improve the efficiency of intent category labeling for historical query data while maintaining high labeling accuracy. This process utilizes the grouping capability of clustering algorithms and the known intent categories of sample query data to achieve preliminary intent classification of large-scale data with low manual cost. Uniform intent category labeling for data within each cluster helps expand the training dataset, enabling the final distilled intent model to cover a wider range of query intents, thereby improving the model's generalization ability and response speed.

[0114] In particular, in the scenario of agent tool selection, the method described in the embodiments of this application can help the agent more accurately understand the user's intent, thereby selecting and calling the most suitable tool to meet the user's needs, improving user satisfaction and overall service quality. Meanwhile, since label propagation is based on clustering results, this method can effectively avoid data skewness problems, ensuring the diversity and balance of training data, which is of great significance for building high-precision, low-latency intent models.

[0115] In an optional embodiment, step S2231, determining the intended category of multiple clusters included in the clustering results, includes the following method steps:

[0116] Step S2232: Based on the voting results of the sample query data located in the same cluster in multiple clusters, determine the intent category of multiple clusters contained in the clustering result.

[0117] The above voting results are based on a "voting" mechanism for clusters, in which the intent category of the majority of sample query data within a cluster is determined as the intent category of that cluster. For example, if most sample query data in a cluster is labeled with the intent of "weather query", then this cluster will be classified as the "weather query" cluster.

[0118] Through step S2232, the technical solution provided in this application embodiment can effectively add intent category labels to historical query data without requiring individual manual annotation of each data entry, significantly reducing annotation costs and workload. Determining the intent category of a cluster based on voting results avoids the risk of the entire cluster being misclassified due to the intent category of a few sample query data.

[0119] Specifically, in scenarios involving agent tool selection, the aforementioned technical solutions help the agent more accurately understand user intent and select and invoke the most suitable tool to respond to user queries. For example, in the intent cluster of "weather query," the agent will more efficiently invoke weather query tools (such as weather forecast service interfaces) without mistakenly invoking other irrelevant tools. Furthermore, because the voting mechanism considers the majority intent of the data, this helps the model focus more on recognizing mainstream intents during training, enhancing the model's usability and response speed, and providing users with a more accurate and faster service experience.

[0120] In other words, the embodiments of this application determine the intent category of clusters through voting results, which not only enables the effective use of a large amount of unlabeled data, but also ensures the accuracy of data labeling and the efficiency of model training. This makes an important contribution to building a high-precision, low-latency agent tool intent selection model.

[0121] In an optional embodiment, step S23 involves performing a data quality check on the labeled data to obtain training data, including the following method steps:

[0122] Step S231: Set the sample query data as the reference example and the labeled data as the judgment object, and generate prompt text;

[0123] Step S232: Based on the prompt text, control the target language model to perform intent judgment on the labeled data and obtain the judgment result, wherein the judgment result is used to determine whether the labeled data matches the sample intent category;

[0124] Step S233: In response to the judgment result indicating that the labeled data does not match the sample intent category, the target data pair is updated to the training data so as to continue iterative training of the initial intent classification model using the updated training data. Here, the target data pair includes the non-matching data pair between the labeled data and the sample intent category.

[0125] The aforementioned prompt text can include a reference example and a judgment object, used to guide the target language model in determining intent. For example, sample query data can be set as the reference example; for instance, "What's the weather like in Beijing tomorrow?" is labeled as the intent category "weather query". Simultaneously, labeled data obtained through clustering label propagation is used as the judgment object, such as historical queries like "Beijing weather for today and tomorrow" being initially labeled as the intent category "weather query". Based on this, a prompt text is generated to guide the target language model (such as a pre-trained large language model) in determining the intent of the judgment object. The constructed prompt text could be, "Please determine whether the following query expresses the intent of a weather query: Beijing weather for today and tomorrow."

[0126] Furthermore, the target language model, based on the prompt text and its own comprehension capabilities, determines whether the intent category of the labeled data matches the sample intent category corresponding to the reference example. For example, for the judgment object "Beijing weather for today and tomorrow" in the prompt text, the large language model might determine that the judgment object does indeed express the intent of "weather query" (i.e., the labeled data matches the sample intent category), or it might determine that the judgment object actually expresses the intent of "travel consultation" (i.e., the labeled data does not match the sample intent category), thus generating the corresponding judgment result.

[0127] Furthermore, when the target language model's judgment result shows that the intent category of the labeled data does not match the intent category of the sample, this set of non-matching data pairs (labeled data and sample intent categories) will be labeled and updated in the training dataset.

[0128] For example, if the large language model determines that "Beijing weather today and tomorrow" should actually belong to the "travel consultation" intent category rather than "weather query", then this misclassified data pair will be recorded and used for subsequent model iteration training.

[0129] Through steps S231 to S233, the technical solution provided in this application embodiment can effectively utilize the capabilities of a large model to verify and adjust the intent category of the initially labeled data, ensuring the quality of the training data. In particular, by adding the target data pairs (labeled data and sample intent category) identified as having pre-labeling errors during the data quality check to the training data, the seed data is expanded, resulting in richer training data, thereby improving the model's accuracy and generalization ability.

[0130] In one optional embodiment, the data processing method further includes the following method steps:

[0131] Step S25: In response to the judgment result indicating that the labeled data matches the sample intent category, stop iteratively training the initial intent classification model using the labeled data and sample intent categories.

[0132] If the target language model determines that the labeled data matches the sample intent category, it assumes that the model has learned how to classify the intent of this data and will not use this labeled data and sample intent category as training data.

[0133] For example, in the scenario of an intelligent question-answering assistant, the system uses the clustering label propagation method to initially label the historical query data "Beijing Weather Forecast" with the intent category of "weather query". Further, if the large language model determines that the intent category of this labeled data is indeed "weather query" (that is, the judgment result shows that the labeled data matches the sample intent category), the system will stop using this set of data (i.e., "Beijing Weather Forecast" and its "weather query" intent category) to continue iteratively training the initial intent classification model.

[0134] By employing step 25 above, the model avoids repeatedly learning from correctly classified data during training, thereby saving computational resources and improving training efficiency. In particular, in the scenario of agent tool selection, the above-described solution in this application helps to build a fast and accurate intent classification model, enabling the agent to understand and respond to user queries more quickly, accurately select and invoke tools, and improve user satisfaction and experience.

[0135] Furthermore, since the model training process filters out correctly classified data, the training dataset will focus more on query samples that are not classified accurately or have complex intents. This helps the model to focus more on improving its ability to identify edge cases and complex intents during training, thereby enhancing the model's robustness and generalization ability and enabling it to provide more accurate services when faced with diverse user queries.

[0136] The data processing method provided in this application embodiment enables a data production and intent model distillation scheme based on clustering label propagation and large model-assisted quality inspection, significantly improving the efficiency, cost-effectiveness, and data quality of intent recognition model training. Specifically, the technical innovations and beneficial effects of this application embodiment are as follows.

[0137] First, by implementing a clustering-based label propagation strategy, this application can quickly identify and pre-label queries related to tool usage intent from massive amounts of historical query data. This effectively filters out a large number of queries unrelated to tool usage, significantly reducing the number of subsequent large model calls and lowering overall computational costs. Simultaneously, the clustering process facilitates manual data filtering, ensuring a balanced distribution of intent categories in the training data and avoiding category bias during model training.

[0138] Secondly, unlike traditional model distillation methods, this application's embodiments do not directly rely on the output of a large language model for data annotation. Instead, they utilize the large language model as a judgment tool to perform secondary quality checks on the pre-labeled data. By constructing a simple judgment task, the parsing difficulties of complex task outputs are avoided, improving the accuracy and success rate of data processing. The large language model performs excellently on simple intent judgment tasks, efficiently identifying and correcting errors in pre-labeling, thereby producing more accurate and high-quality training data and ensuring that the distilled smaller model achieves better performance in intent recognition.

[0139] Third, this application's embodiments introduce an iterative training mechanism, which continuously improves the accuracy and generalization ability of the intent model through a cyclic distillation process. In each iteration, the system not only expands the training dataset but also identifies and adds "difficult" examples that are semantically similar but have different intents (i.e., data pairs where the labeled data and sample intent categories do not match). These examples are crucial for improving the model's ability to handle complex queries. With the assistance of a large model, the training data of the small model is continuously optimized and updated, ensuring a steady improvement in model performance and ultimately achieving high-precision, low-latency intent recognition, providing strong support for the selection of intelligent agent tools.

[0140] In summary, this application's embodiments achieve a highly efficient, low-cost, and high-quality intent model distillation method through cluster label propagation and large-scale model quality inspection strategies. This method not only rapidly mines effective intent samples from large amounts of historical data but also continuously optimizes model performance through precise diagnosis and control of the large-scale model, significantly improving the accuracy and response speed of the agent in tool selection and providing users with a more intelligent and user-friendly service experience.

[0141] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0142] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0143] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, they can also be implemented by hardware. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM), random access memory (RAM), magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0144] In the aforementioned operating environment, this application also provides, as follows: Figure 3 This illustrates a data processing method. Figure 3 This is a flowchart of a data processing method according to an embodiment of this application, such as... Figure 3 As shown, the data processing method includes:

[0145] Step S31: Obtain the target query data;

[0146] Step S32: Use a target intent classification model to classify the target query data to select the target tool corresponding to the target query data;

[0147] Step S33: Call the target tool to obtain the target query results;

[0148] The target intent classification model is obtained by training the initial intent classification model with training data. The training data is obtained by checking the quality of the labeled data. The labeled data is obtained by pre-labeling the seed data through label propagation. The seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories.

[0149] The aforementioned target query data is the query data input by the user. A target intent classification model is used to classify the intent of the target query data. After accurately identifying the query intent category reflected in the target query data, an appropriate target tool is selected based on this query intent category. This target tool is used to process the query data and obtain the target query results. The aforementioned target intent classification model can be trained using the methods and steps described in the embodiments of this application.

[0150] In this embodiment, target query data is acquired; a target intent classification model is used to classify the target query data to select the target tool corresponding to the target query data; and the target tool is invoked to obtain the target query results. The target intent classification model is obtained by training an initial intent classification model with training data. The training data is obtained by performing data quality checks on labeled data. The labeled data is obtained by pre-labeling seed data using a label propagation method. The seed data includes sample data pairs, each including sample query data and sample intent category. This embodiment combines a label propagation scheme and a data quality check scheme to pre-label sample data pairs and further construct training data for training the target intent classification model through data quality checks. This training data is of high quality. Based on this, the target intent classification model can more accurately identify the intent of the target query data and select a more suitable target tool. Invoking the target tool can obtain more accurate target query results. Furthermore, the method provided in this embodiment avoids the large amount of redundant prediction and resource waste caused by using conventional model distillation training schemes. As described above, the embodiments of this application achieve the goal of training a target intent classification model based on label propagation and data quality inspection, thereby realizing the technical effects of selecting appropriate target tools by accurately identifying query intent, improving the accuracy of model query results, reducing model training costs and improving training efficiency. This solves the technical problems of high training costs, low efficiency and poor tool selection accuracy and query accuracy of the trained models in related technologies.

[0151] It should be noted that the preferred embodiments of steps S31 to S34 described above can be found in the foregoing description, and will not be repeated here.

[0152] In the aforementioned operating environment, this application provides the following: Figure 4 This illustrates a data processing method. Figure 4 This is a flowchart of a data processing method according to an embodiment of this application, such as... Figure 4 As shown, the data processing method includes:

[0153] Step S41: Obtain weather query data;

[0154] Step S42: Use a target intent classification model to classify the intent of the weather query data in order to select the weather forecast tool corresponding to the weather query data;

[0155] Step S43: Call the weather forecast tool to obtain the weather query results;

[0156] The target intent classification model is obtained by training the initial intent classification model with training data. The training data is obtained by checking the quality of the labeled data. The labeled data is obtained by pre-labeling the seed data through label propagation. The seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories.

[0157] The data processing method provided in this application can be used to provide weather query services to users in weather query application scenarios, for example, to provide intelligent weather query functions for intelligent chatbots. The aforementioned weather query data refers to query data input by the user for querying the weather (such as query text, query voice, or query video). A target intent classification model is used to classify the intent of the weather query data. After accurately identifying the category of weather query intent reflected in the weather query data, an appropriate weather forecast tool is selected based on the category of weather query intent. This weather forecast tool is used to perform targeted weather query processing based on the weather query data to obtain weather query results. The aforementioned target intent classification model can be trained using the methods and steps described in the embodiments of this application.

[0158] In this embodiment, weather query data is acquired; a target intent classification model is used to classify the intent of the weather query data to select the corresponding weather forecast tool; and the weather forecast tool is invoked to obtain the weather query results. The target intent classification model is obtained by training an initial intent classification model using training data. The training data is obtained by performing data quality checks on labeled data. The labeled data is obtained by pre-labeling seed data using a label propagation method. The seed data includes sample data pairs, each including sample query data and sample intent category. This embodiment combines a label propagation scheme and a data quality check scheme. The sample data pairs are pre-labeled and further tested for data quality to construct training data for training the target intent classification model. This training data is of high quality. Based on this, the target intent classification model can more accurately identify the intent of the target query data and select a more suitable target tool. Invoking the target tool can obtain more accurate target query results. Furthermore, the method provided in this embodiment avoids the large amount of redundant prediction and resource waste caused by using conventional model distillation training schemes. As described above, the embodiments of this application achieve the goal of training a target intent classification model based on label propagation and data quality inspection, thereby realizing the technical effects of selecting appropriate target tools by accurately identifying query intent, improving the accuracy of model query results, reducing model training costs and improving training efficiency. This solves the technical problems of high training costs, low efficiency and poor tool selection accuracy and query accuracy of the trained models in related technologies.

[0159] It should be noted that the preferred embodiments of steps S41 to S43 described above can be found in the relevant descriptions in the foregoing embodiments, and will not be repeated here.

[0160] In the aforementioned operating environment, this application provides the following: Figure 5 This illustrates a data processing method. Figure 5 This is a flowchart of a data processing method according to an embodiment of this application, such as... Figure 5 As shown, the data processing method includes:

[0161] Step S51: Obtain a data processing request through the first application programming interface, wherein the request data carried in the data processing request includes: target query data;

[0162] Step S52: Return a data processing response through the second application programming interface, wherein the response data carried in the data processing response includes: the target query result;

[0163] Specifically, the target query results are obtained by classifying the target query data using a target intent classification model to select the target tool corresponding to the target query data and by calling the target tool. The target intent classification model is obtained by training the initial intent classification model using training data. The training data is obtained by checking the data quality of the labeled data. The labeled data is obtained by pre-labeling the seed data through label propagation. The seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories.

[0164] The first and second application programming interfaces (APIs) mentioned above can be the same or different APIs. In one optional embodiment, the interface parameters in the first and second APIs may include, but are not limited to: a global interface identifier, an interface signing key, an interface timestamp, an interface request identifier, and a system call credential identifier. The first API can use GET or POST as the interface request method to obtain the file processing request. The second API can use JSON format to return the file processing response.

[0165] In this embodiment, a data processing request is obtained through a first application programming interface (API), wherein the request data carried in the data processing request includes target query data; a data processing response is returned through a second API, wherein the response data carried in the data processing response includes target query results; wherein the target query results are obtained by classifying the target query data using a target intent classification model to select the target tool corresponding to the target query data and by calling the target tool, the target intent classification model is obtained by training an initial intent classification model using training data, the training data is obtained by performing data quality checks on labeled data, the labeled data is obtained by pre-labeling seed data using a label propagation method, and the seed data includes sample data pairs, each of which includes sample query data and sample intent categories. This application combines a label propagation scheme and a data quality check scheme. Sample data pairs are pre-labeled, and further data quality checks are performed to construct training data for training a target intent classification model. This training data is of high quality. Based on this, the target intent classification model can more accurately identify the intent of target query data and select a more suitable target tool. Calling this target tool can obtain more accurate target query results. Furthermore, the method provided in this application avoids the large amount of redundant predictions and resource waste caused by using conventional model distillation training schemes. Therefore, this application achieves the goal of training a target intent classification model based on label propagation and data quality checks. This achieves the technical effects of accurately identifying query intent, selecting appropriate target tools, improving the accuracy of model query results, reducing model training costs, and improving training efficiency. It also solves the technical problems of high training costs, low efficiency, and poor tool selection and query accuracy of related technologies for models with tool selection functions.

[0166] It should be noted that the preferred embodiments of steps S51 to S52 described above can be found in the relevant descriptions in the foregoing embodiments, and will not be repeated here.

[0167] In the aforementioned operating environment, this application provides the following: Figure 6 This illustrates a data processing method. Figure 6 This is a flowchart of a data processing method according to an embodiment of this application, such as... Figure 6 As shown, the data processing method includes:

[0168] Step S61: Obtain the currently input data processing dialogue request, wherein the request data carried in the data processing dialogue request includes: target query data;

[0169] Step S62: In response to the data processing dialogue request, return a data processing dialogue response, wherein the information carried in the data processing dialogue response includes: the target query result;

[0170] Step S63: Display the target query results within the graphical user interface;

[0171] Specifically, the target query results are obtained by classifying the target query data using a target intent classification model to select the target tool corresponding to the target query data and by calling the target tool. The target intent classification model is obtained by training the initial intent classification model using training data. The training data is obtained by checking the data quality of the labeled data. The labeled data is obtained by pre-labeling the seed data through label propagation. The seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories.

[0172] Based on the above method steps, a visualization scheme for data processing functions is provided. The terminal device provides a graphical user interface (GUI), which displays at least one data query scenario. The GUI's display content also includes input components (such as text input boxes, voice input controls, etc.) and display components (such as text display windows, image display windows, etc.). The user inputs a data processing dialog request through the input components to specify the target query data in the data processing task. After detecting the user's input, a data processing flow is executed based on the target query data to obtain the target query result. Furthermore, the target query result is displayed through the display components within the GUI.

[0173] In this embodiment, a data processing dialogue request is obtained from the current input. The request data carried in the data processing dialogue request includes target query data. In response to the data processing dialogue request, a data processing dialogue reply is returned. The information carried in the data processing dialogue reply includes target query results. The target query results are displayed in the graphical user interface. The target query results are obtained by classifying the target query data using a target intent classification model to select the target tool corresponding to the target query data and by calling the target tool. The target intent classification model is obtained by training an initial intent classification model using training data. The training data is obtained by performing data quality checks on the labeled data. The labeled data is obtained by pre-labeling the seed data using a label propagation method. The seed data includes sample data pairs, and each sample data pair includes sample query data and sample intent category. This application combines a label propagation scheme and a data quality check scheme. Sample data pairs are pre-labeled, and further data quality checks are performed to construct training data for training a target intent classification model. This training data is of high quality. Based on this, the target intent classification model can more accurately identify the intent of target query data and select a more suitable target tool. Calling this target tool can obtain more accurate target query results. Furthermore, the method provided in this application avoids the large amount of redundant predictions and resource waste caused by using conventional model distillation training schemes. Therefore, this application achieves the goal of training a target intent classification model based on label propagation and data quality checks. This achieves the technical effects of accurately identifying query intent, selecting appropriate target tools, improving the accuracy of model query results, reducing model training costs, and improving training efficiency. It also solves the technical problems of high training costs, low efficiency, and poor tool selection and query accuracy of related technologies for models with tool selection functions.

[0174] It should be noted that the preferred embodiments of steps S61 to S63 described above can be found in the relevant descriptions in the foregoing embodiments, and will not be repeated here.

[0175] According to an embodiment of this application, an apparatus embodiment for implementing the above-described data processing method is also provided. Figure 7 This is a schematic diagram of the structure of a data processing apparatus according to an embodiment of this application, such as... Figure 7 As shown, the device includes:

[0176] The acquisition module 701 is used to acquire seed data, wherein the seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories;

[0177] The annotation module 702 is used to pre-annotate the seed data through label propagation to obtain labeled data;

[0178] The inspection module 703 is used to perform data quality checks on the labeled data to obtain training data.

[0179] Training module 704 is used to train the initial intent classification model using training data to obtain the target intent classification model. The target intent classification model is used to classify the intent of the target query data to select the target tool corresponding to the target query data. The target intent classification model is also used to call the target tool to obtain the target query results.

[0180] Optionally, the labeling module 702 is further configured to: acquire historical query data, wherein the historical query data is unlabeled real query data generated within a historical time period; perform cluster analysis on the historical query data and sample query data to obtain clustering results; and perform label propagation on the clustering results to obtain labeled data.

[0181] Optionally, the annotation module 702 is further configured to: vectorize historical query data to obtain a first vector, and vectorize sample query data to obtain a second vector; calculate the similarity between the first vector and the second vector based on a preset threshold condition to obtain a target calculation result; and perform cluster analysis on the target calculation result using a preset clustering algorithm to obtain a clustering result.

[0182] Optionally, the preset threshold conditions include: multi-level similarity thresholds. The annotation module 702 is further used to: select a target similarity threshold from the multi-level similarity thresholds; calculate the similarity between the first vector and the second vector using a preset vector similarity calculation method to obtain an initial calculation result; compare the initial calculation result with the target similarity threshold to obtain a comparison result; and filter the initial calculation result based on the comparison result to obtain the target calculation result.

[0183] Optionally, in the above data processing device, the clustering categories corresponding to the clustering results are determined according to the preset intention quantity.

[0184] Optionally, the annotation module 702 is further configured to: determine the intent category of multiple clusters contained in the clustering result; and annotate the data categories located in the same cluster within the multiple clusters as intent categories to obtain annotated data.

[0185] Optionally, the annotation module 702 is further configured to: determine the intent category of multiple clusters contained in the clustering result based on the voting results of sample query data located in the same cluster among multiple clusters.

[0186] Optionally, the aforementioned inspection module 703 is further configured to: set the sample query data as a reference example and the labeled data as the judgment object, and generate prompt text; control the target language model to perform intent judgment on the labeled data based on the prompt text, and obtain a judgment result, wherein the judgment result is used to determine whether the labeled data matches the sample intent category; in response to the judgment result indicating that the labeled data does not match the sample intent category, update the target data pair to the training data, so as to continue iteratively training the initial intent classification model using the updated training data, wherein the target data pair includes non-matching data pairs of labeled data and sample intent categories.

[0187] Optionally, in addition to all the modules mentioned above, the data processing device also includes a stop module (not shown in the figure), which is used to stop iteratively training the initial intent classification model using the labeled data and the sample intent category in response to the judgment result indicating that the labeled data matches the sample intent category.

[0188] It should be noted that the above-mentioned acquisition module 701, annotation module 702, inspection module 703 and training module 704 correspond to steps S21 to S24 in the embodiments. The four modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the foregoing embodiments.

[0189] According to an embodiment of this application, an apparatus embodiment for implementing the data processing method in the above embodiments is also provided. Figure 8 This is a schematic diagram of the structure of a data processing apparatus according to an embodiment of this application, such as... Figure 8 As shown, the device includes:

[0190] Module 801 is used to obtain the target query data;

[0191] Selection module 802 is used to classify the target query data using a target intent classification model in order to select the target tool corresponding to the target query data.

[0192] Module 803 is invoked to call the target tool to obtain the target query results;

[0193] The target intent classification model is obtained by training the initial intent classification model with training data. The training data is obtained by checking the quality of the labeled data. The labeled data is obtained by pre-labeling the seed data through label propagation. The seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories.

[0194] It should be noted that the above-mentioned acquisition module 801, selection module 802 and calling module 803 correspond to the aforementioned steps S31 to S33. The three modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments.

[0195] According to an embodiment of this application, an apparatus embodiment for implementing the data processing method in the above embodiments is also provided. Figure 9 This is a schematic diagram of the structure of a data processing apparatus according to an embodiment of this application, such as... Figure 9 As shown, the device includes:

[0196] Module 901 is used to obtain weather query data;

[0197] Select module 902 is used to classify the intent of weather query data using a target intent classification model in order to select the weather forecast tool corresponding to the weather query data.

[0198] Module 903 is used to invoke the weather forecast tool to obtain weather query results;

[0199] The target intent classification model is obtained by training the initial intent classification model with training data. The training data is obtained by checking the quality of the labeled data. The labeled data is obtained by pre-labeling the seed data through label propagation. The seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories.

[0200] It should be noted that the above-mentioned acquisition module 901, selection module 902 and calling module 903 correspond to the aforementioned steps S41 to S43. The three modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments.

[0201] According to an embodiment of this application, an apparatus embodiment for implementing the data processing method in the above embodiments is also provided. Figure 10 This is a schematic diagram of the structure of a data processing apparatus according to an embodiment of this application, such as... Figure 10 As shown, the device includes:

[0202] The request module 1001 is used to obtain a data processing request through a first application programming interface, wherein the request data carried in the data processing request includes: target query data;

[0203] The response module 1002 is used to return a data processing response through a second application programming interface, wherein the response data carried in the data processing response includes: the target query result;

[0204] Specifically, the target query results are obtained by classifying the target query data using a target intent classification model to select the target tool corresponding to the target query data and by calling the target tool. The target intent classification model is obtained by training the initial intent classification model using training data. The training data is obtained by checking the data quality of the labeled data. The labeled data is obtained by pre-labeling the seed data through label propagation. The seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories.

[0205] It should be noted that the above-mentioned request module 1001 and response module 1002 correspond to the aforementioned steps S51 to S52. The two modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments.

[0206] According to an embodiment of this application, an apparatus embodiment for implementing the data processing method in the above embodiments is also provided. Figure 11 This is a schematic diagram of the structure of a data processing apparatus according to an embodiment of this application, such as... Figure 11 As shown, the device includes:

[0207] The acquisition module 1101 is used to acquire the currently input data processing dialogue request, wherein the request data carried in the data processing dialogue request includes: target query data;

[0208] Return module 1102 is used to respond to a data processing dialog request and return a data processing dialog response, wherein the information carried in the data processing dialog response includes: the target query result;

[0209] Display module 1103 is used to display the target query results within a graphical user interface;

[0210] Specifically, the target query results are obtained by classifying the target query data using a target intent classification model to select the target tool corresponding to the target query data and by calling the target tool. The target intent classification model is obtained by training the initial intent classification model using training data. The training data is obtained by checking the data quality of the labeled data. The labeled data is obtained by pre-labeling the seed data through label propagation. The seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories.

[0211] It should be noted that the above-mentioned acquisition module 1101, return module 1102 and display module 1103 correspond to the aforementioned steps S61 to S63. The three modules and the corresponding steps implement the same instances and application scenarios, but are not limited to the content disclosed in the above embodiments.

[0212] It should be noted that the above-mentioned modules or units may be hardware or software components stored in memory and processed by one or more processors. The above-mentioned modules may also be part of a device and run in a computer terminal.

[0213] It should be noted that the preferred implementation of this embodiment can be found in the relevant descriptions of the foregoing embodiments, and will not be repeated here.

[0214] According to an embodiment of this application, an electronic device is also provided, which can be any terminal device in a group of computer terminals. Optionally, in this embodiment, the above-mentioned electronic device can also be replaced by a mobile terminal or other terminal device.

[0215] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0216] In this embodiment, the above-mentioned electronic device can execute the program code of the following steps in the data processing method: obtaining seed data, wherein the seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories; pre-labeling the seed data through label propagation to obtain labeled data; performing data quality checks on the labeled data to obtain training data; and training the initial intent classification model using the training data to obtain a target intent classification model, wherein the target intent classification model is used to classify the intent of the target query data to select the target tool corresponding to the target query data. The target intent classification model is also used to call the target tool to obtain the target query results.

[0217] Optionally, Figure 12 This is a structural block diagram of an electronic device according to an embodiment of this application, such as... Figure 12 As shown, the electronic device 120 may include: one or more (only one is shown in the figure) processors 1202, memory 1204, memory controller 1206, and peripheral interface 1208, wherein the peripheral interface 1208 is connected to a radio frequency module, an audio module, and a display.

[0218] The memory 1204 can be used to store software programs and modules, such as the program instructions / modules corresponding to the data processing method and apparatus in this embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby implementing the aforementioned data processing method. The memory 1204 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1204 may further include memory remotely located relative to the processor, and these remote memories can be connected to the electronic device 120 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0219] The processor 1202 can invoke the information and application program stored in the memory through the transmission device to perform the following steps: acquiring seed data, wherein the seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories; pre-labeling the seed data through label propagation to obtain labeled data; performing data quality checks on the labeled data to obtain training data; and training the initial intent classification model using the training data to obtain a target intent classification model, wherein the target intent classification model is used to classify the intent of the target query data to select the target tool corresponding to the target query data. The target intent classification model is also used to invoke the target tool to obtain the target query results.

[0220] This application provides an electronic device scheme for implementing the above-described data processing method. The process involves: acquiring seed data, which includes sample data pairs, each pair comprising sample query data and sample intent categories; pre-labeling the seed data using a label propagation method to obtain labeled data; performing data quality checks on the labeled data to obtain training data; and training an initial intent classification model using the training data to obtain a target intent classification model. This target intent classification model is used to classify the intent of target query data to select the corresponding target tool. It is also used to invoke the target tool to obtain the target query results. This application combines a label propagation scheme and a data quality check scheme. Pre-labeling the sample data pairs and further performing data quality checks constructs training data for training the target intent classification model. This training data is of high quality, enabling the target intent classification model to more accurately identify the intent of the target query data and select a more suitable target tool. Invoking the target tool yields more accurate target query results. Furthermore, the method provided in this application avoids the large amount of redundant prediction and resource waste associated with conventional model distillation training schemes. As described above, the embodiments of this application achieve the goal of training a target intent classification model based on label propagation and data quality inspection, thereby realizing the technical effects of selecting appropriate target tools by accurately identifying query intent, improving the accuracy of model query results, reducing model training costs and improving training efficiency. This solves the technical problems of high training costs, low efficiency and poor tool selection accuracy and query accuracy of the trained models in related technologies.

[0221] Those skilled in the art will understand that Figure 12 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones (such as Android phones, iOS phones, etc.), tablets, PDAs, and mobile internet devices (MIDs) and other terminal devices. Figure 12 This does not limit the structure of the aforementioned electronic device. For example, electronic device 120 may also include components that are more... Figure 12 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 12 The different configurations shown.

[0222] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, ROM, RAM, disk or optical disk, etc.

[0223] According to an embodiment of this application, a computer-readable storage medium is also provided. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the data processing method provided in the above embodiments.

[0224] Optionally, in this embodiment, the storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0225] Optionally, in this embodiment, the computer-readable storage medium is configured to store program code for performing the following steps: obtaining seed data, wherein the seed data includes sample data pairs, the sample data pairs including sample query data and sample intent categories; pre-labeling the seed data through label propagation to obtain labeled data; performing data quality checks on the labeled data to obtain training data; training the initial intent classification model using the training data to obtain a target intent classification model, wherein the target intent classification model is used to classify the intent of the target query data to select the target tool corresponding to the target query data; the target intent classification model is also used to call the target tool to obtain the target query results.

[0226] Embodiments of this application also provide a computer program product. Optionally, in this embodiment, the computer program product may include a computer program that, when executed by a processor, implements the methods provided in the embodiments described above.

[0227] Embodiments of this application also provide a computer program product. Optionally, the computer program product may include a non-volatile computer-readable storage medium, which can be used to store a computer program that, when executed by a processor, implements the method provided in the above embodiments.

[0228] Embodiments of this application also provide a computer program. Optionally, in this embodiment, when the computer program is executed by a processor, it implements the method provided in the above embodiments.

[0229] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0230] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0231] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0232] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0233] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, ROM, RAM, portable hard drives, magnetic disks, or optical disks.

[0234] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A data processing method, characterized in that, include: Acquire seed data, wherein the seed data includes sample data pairs, and the sample data pairs include sample query data and sample intent categories; The seed data is pre-labeled using a tag propagation method to obtain labeled data; The labeled data is subjected to data quality checks to obtain training data; The initial intent classification model is trained using the training data to obtain a target intent classification model. The target intent classification model is used to classify the intent of the target query data to select the target tool corresponding to the target query data. The target intent classification model is also used to call the target tool to obtain the target query results.

2. The data processing method according to claim 1, characterized in that, The seed data is pre-labeled using the aforementioned tag propagation method, resulting in labeled data including: Obtain historical query data, wherein the historical query data is unlabeled real query data generated within a historical time period; Cluster analysis is performed on the historical query data and the sample query data to obtain clustering results; The clustering results are then propagated through labeling to obtain the labeled data.

3. The data processing method according to claim 2, characterized in that, Cluster analysis is performed on the historical query data and the sample query data to obtain the clustering results, including: The historical query data is vectorized to obtain a first vector, and the sample query data is vectorized to obtain a second vector; The similarity between the first vector and the second vector is calculated based on a preset threshold condition to obtain the target calculation result; The target calculation results are clustered using a preset clustering algorithm to obtain the clustering results.

4. The data processing method according to claim 3, characterized in that, The preset threshold conditions include: multi-level similarity thresholds. Based on the preset threshold conditions, similarity calculations are performed on the first vector and the second vector to obtain the target calculation result, including: Select the target similarity threshold from the multi-level similarity thresholds; The first vector and the second vector are similar to each other using a preset vector similarity calculation method to obtain an initial calculation result; The initial calculation result is compared with the target similarity threshold to obtain the comparison result; The initial calculation results are filtered based on the comparison results to obtain the target calculation results.

5. The data processing method according to claim 3, characterized in that, The clustering categories corresponding to the clustering results are determined based on the preset number of intentions.

6. The data processing method according to claim 2, characterized in that, The clustering results are then propagated through labels to obtain the labeled data, which includes: Determine the intent category of the multiple clusters contained in the clustering results; The data categories located within the same cluster in the multiple clusters are labeled as the intent categories to obtain the labeled data.

7. The data processing method according to claim 6, characterized in that, Determining the intent category of the plurality of clusters included in the clustering results includes: Based on the voting results of sample query data located within the same cluster in the multiple clusters, the intent category of the multiple clusters included in the clustering result is determined.

8. The data processing method according to any one of claims 1 to 7, characterized in that, The labeled data is subjected to a data quality check to obtain the training data, which includes: The sample query data is set as a reference example, and the labeled data is set as the judgment object to generate prompt text; Based on the prompt text, the target language model is used to determine the intent of the labeled data and obtain a determination result, wherein the determination result is used to determine whether the labeled data matches the sample intent category; In response to the judgment result indicating that the labeled data does not match the sample intent category, the target data pair is updated to the training data so that the initial intent classification model can be iteratively trained using the updated training data. The target data pair includes the non-matching data pair between the labeled data and the sample intent category.

9. The data processing method according to claim 8, characterized in that, The data processing method further includes: In response to the judgment result indicating that the labeled data matches the sample intent category, the iterative training of the initial intent classification model using the labeled data and the sample intent category is stopped.

10. A data processing method, characterized in that, include: Retrieve the target query data; A target intent classification model is used to classify the target query data to select the target tool corresponding to the target query data; Invoke the target tool to obtain the target query results; The target intent classification model is obtained by training the initial intent classification model with training data. The training data is obtained by performing data quality checks on the labeled data. The labeled data is obtained by pre-labeling the seed data through label propagation. The seed data includes sample data pairs, and each sample data pair includes sample query data and sample intent category.

11. A data processing method, characterized in that, include: Get weather query data; A target intent classification model is used to classify the intent of the weather query data in order to select the weather forecast tool corresponding to the weather query data. Call the weather forecast tool to obtain weather query results; The target intent classification model is obtained by training the initial intent classification model with training data. The training data is obtained by performing data quality checks on the labeled data. The labeled data is obtained by pre-labeling the seed data through label propagation. The seed data includes sample data pairs, and each sample data pair includes sample query data and sample intent category.

12. A data processing method, characterized in that, include: A data processing request is obtained through a first application programming interface, wherein the request data carried in the data processing request includes: target query data; A data processing response is returned through a second application programming interface, wherein the response data carried in the data processing response includes: the target query result; The target query result is obtained by classifying the target query data using a target intent classification model to select the target tool corresponding to the target query data and by calling the target tool. The target intent classification model is obtained by training an initial intent classification model using training data. The training data is obtained by checking the quality of the labeled data. The labeled data is obtained by pre-labeling the seed data using a label propagation method. The seed data includes sample data pairs, and each sample data pair includes sample query data and sample intent category.

13. A data processing method, characterized in that, include: Obtain the currently input data processing dialogue request, wherein the request data carried in the data processing dialogue request includes: target query data; In response to the data processing dialogue request, a data processing dialogue response is returned, wherein the information carried in the data processing dialogue response includes: the target query result; The target query results are displayed within a graphical user interface; The target query result is obtained by classifying the target query data using a target intent classification model to select the target tool corresponding to the target query data and by calling the target tool. The target intent classification model is obtained by training an initial intent classification model using training data. The training data is obtained by checking the quality of the labeled data. The labeled data is obtained by pre-labeling the seed data using a label propagation method. The seed data includes sample data pairs, and each sample data pair includes sample query data and sample intent category.

14. An electronic device, characterized in that, include: Memory, which stores executable programs; A processor for running the program, wherein the program, when running, performs the data processing method according to any one of claims 1 to 13.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored executable program, wherein, when the executable program is executed, it controls the device on which the computer-readable storage medium is located to perform the data processing method according to any one of claims 1 to 13.

16. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the data processing method according to any one of claims 1 to 13.

Images