Government affair large model-based intention recognition method and device, medium and equipment

Through the intention recognition method of the government affairs big model, clean and vectorize user expression text, combined with multiple rounds of question templates and incremental training, the problem of accurate and efficient intention recognition in traditional government affairs services is solved, and efficient and accurate users are achieved Intent identification and government services reduce operating costs.

CN119088965APending Publication Date: 2024-12-06SHANDONG LANGCHAO YUNTOU INFORMATION TECH CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411195785.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-29
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

In traditional government services, the accuracy and efficiency of intention identification are inaccurate and efficient, making it difficult to meet the public's needs for convenient and efficient services, especially when facing complex and changeable user questions, traditional rules-based or template-based methods are unable to do so.

Method used

Using the intent recognition method based on the government affairs big model, by obtaining the text of the user's initial expression, cleaning and vectorization transformation, inputting a pre-trained intent recognition model to obtain the initial intent, and using multiple rounds of questioning templates that gradually narrow the scope of questioning for interaction , determine the final intention and continuously optimize the model through incremental training and performance evaluation.

Benefits of technology

Improve the accuracy of intention recognition, reduce misjudgment and misjudgment, improve the efficiency and user experience of government services, reduce dependence on manual customer services, reduce operational costs, and support the intelligent and efficient development of government service systems. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119088965A_ABST
    Figure CN119088965A_ABST
Patent Text Reader

Abstract

The invention provides an intention recognition method and device based on a large government affair model, a medium and equipment. The method comprises the steps of obtaining an initial expression text of a user; performing cleaning processing on the primary expression text; converting the cleaned primary expression text into a primary expression vector in a numerical form; inputting the initial expression vector into a pre-trained intention recognition model to obtain an initial intention of the user; determining a corresponding question template according to the initial intention; wherein multiple rounds of questioning information with the strategy of gradually narrowing the questioning range is designed in the questioning template; and performing multiple rounds of interaction with the user according to the multiple rounds of questioning information in the questioning template, and determining a final intention of the user according to an interaction result. According to the embodiment of the invention, the intention recognition accuracy can be improved, the government affair service efficiency is improved, the user experience is optimized, and the labor cost is reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of government service technology, and in particular to an intention recognition method and device, medium, and equipment based on a government big model. Background Art

[0002] In recent years, with the rapid development of information technology and the public's expectations for the quality of government services, traditional government service methods have become inadequate. The public expects to be able to easily access government services at any time and any place, and hopes that the services can respond to their needs quickly and accurately. In this context, the government industry urgently needs to introduce intelligent and automated technologies into the service process to improve service efficiency, optimize user experience, and meet the public's new expectations for government services. In the process of intelligent services, intent recognition is a very critical link. Summary of the invention

[0003] In response to at least one of the above technical problems, an embodiment of the present invention provides an intent recognition method, device, medium, and equipment based on a big government model.

[0004] According to the first aspect, the intent recognition method based on the government affairs big model provided by the embodiment of the present invention includes:

[0005] Get the user's first expression text;

[0006] Performing cleaning processing on the initial expression text;

[0007] Convert the cleaned initial expression text into an initial expression vector in numerical form;

[0008] Inputting the initial expression vector into a pre-trained intention recognition model to obtain the initial intention of the user;

[0009] Determine a corresponding question template according to the initial intention; wherein the question template is designed with multiple rounds of question information with the strategy of gradually narrowing the scope of the question;

[0010] Perform multiple rounds of interactions with the user according to the multiple rounds of question information in the question template, and determine the user's final intention according to the interaction results.

[0011] In one embodiment, the cleaning process of the first expression text includes:

[0012] Removing redundant information from the initial expression text;

[0013] Correct or delete erroneous information and abnormal information in the initial expression text;

[0014] Remove information in the first expression text that is irrelevant to the intent recognition of the government affairs industry.

[0015] In one embodiment, the step of converting the cleaned initial expression text into an initial expression vector in numerical form includes: using a bidirectional encoder representation model of a transformation network to convert the cleaned initial expression text into the initial expression vector.

[0016] In one embodiment, before inputting the initial expression vector into a pre-trained intent recognition model, the method includes:

[0017] Collect multiple demand text materials from the government affairs industry;

[0018] Clean each demand text material;

[0019] Providing the cleaned demand text material to the annotator, so that the annotator annotates the user intention on each demand text material to obtain an annotated demand text material;

[0020] Convert each labeled requirement text material into an expression vector in numerical form, use the expression vector corresponding to each labeled requirement text material as a training sample, and form a training sample set with multiple training samples;

[0021] Choose the model architecture;

[0022] According to the training sample set, model training is performed on the basis of the selected model architecture to obtain a converged intent recognition model;

[0023] Performing a performance evaluation on the intent recognition model;

[0024] If the performance evaluation is qualified, the intention recognition model is connected to the government affairs industry knowledge base, so that the intention recognition model can call the government affairs industry knowledge from the government affairs industry knowledge base;

[0025] Deploy the intent recognition model to the application scenario.

[0026] In one embodiment, the method further comprises:

[0027] The question template is adjusted according to the feedback content of the user during multiple rounds of interaction with the user.

[0028] In one embodiment, the method further comprises:

[0029] Regularly obtain incremental data generated during the intent recognition process in the government affairs industry; wherein each piece of incremental data includes the first expression text, the initial intent, and the final intent;

[0030] Filter out first incremental data from the incremental data, wherein the similarity between the initial intent and the final intent is higher than a first preset value, and use the first incremental data as a positive sample;

[0031] Filter out second incremental data from the incremental data, in which the similarity between the initial intent and the final intent is lower than a second preset value, and use the second incremental data as a negative sample; wherein the first preset value is greater than the second preset value;

[0032] The positive samples and the negative samples are used to perform incremental training on the intent recognition model.

[0033] In one embodiment, the method further comprises:

[0034] After each incremental training, performing a performance evaluation on the intent recognition model;

[0035] If the performance evaluation is qualified, the intent recognition model is applied to the application scenario; otherwise, the intent recognition model continues to be trained until the performance evaluation of the intent recognition model is qualified.

[0036] According to the second aspect, the intention recognition device based on the government affairs big model provided by the embodiment of the present invention includes:

[0037] A text acquisition module is used to acquire the user's first expression text;

[0038] A text cleaning module, used for cleaning the first expression text;

[0039] A vector conversion module, used for converting the cleaned primary expression text into a primary expression vector in numerical form;

[0040] A model recognition module, used for inputting the initial expression vector into a pre-trained intention recognition model to obtain the initial intention of the user;

[0041] A template determination module, used to determine a corresponding question template according to the initial intention; wherein the question template is designed with multiple rounds of question information with the strategy of gradually narrowing the scope of the question;

[0042] The multi-round interaction module is used to perform multi-round interactions with the user according to the multi-round question information in the question template, and determine the user's final intention according to the interaction results.

[0043] According to a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, enables the computer to execute the method provided in the first aspect.

[0044] According to a fourth aspect, an embodiment of the present invention provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, the method provided in the first aspect is implemented.

[0045] The intent recognition method, device, medium, and equipment based on the government affairs big model provided by the embodiment of the present invention cleans the initial expression text and converts it into the initial expression vector, and then inputs the initial expression vector into the intent recognition model, so that the initial intent can be obtained, and then the multi-round question information in the question template is used to interact with the user in multiple rounds, so as to determine the final intent of the user according to the interaction result. It can be seen that through multiple rounds of interaction and step-by-step questioning, the user's needs can be understood more deeply, and the scope of intent can be gradually narrowed until the user's intent is accurately located. This gradually convergent strategy greatly improves the accuracy of intent recognition and reduces the possibility of misjudgment and missed judgment. Moreover, the automated execution based on the embodiment of the present invention enables the government service system to respond quickly and provide corresponding services, which greatly improves the efficiency of government services, reduces user waiting time, and improves the overall performance of the government service system. In addition, accurate user intent recognition can ensure that the government service system provides users with services or information that are more in line with their needs. Users no longer need to spend a lot of time on tedious queries or descriptions, but can quickly and accurately obtain the required services, thereby optimizing the user experience. By introducing intelligent intent recognition technology, a large number of user inquiries and service applications can be automatically processed, reducing dependence on manual customer service, which reduces the labor cost of government services and improves work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flowchart of an intention recognition method based on a government affairs big model in one embodiment of the present invention;

[0047] Figure 2 It is a structural block diagram of an intention recognition device based on a government affairs big model in one embodiment of the present invention. DETAILED DESCRIPTION

[0048] Government services involve many fields and subdivided scenarios, and users' questioning methods and expression habits vary greatly, which makes the identification of user intentions extremely complicated. Although traditional rule-based or template-based intention identification methods are simple and intuitive, they are often unable to cope with complex and changeable user questions, and their accuracy and flexibility are difficult to meet actual needs. Therefore, the government service industry urgently needs to introduce advanced technologies such as deep learning to accurately identify users' true intentions in order to provide more intimate and efficient services.

[0049] To this end, in the first aspect, the embodiment of the present invention provides an intention recognition method based on a government affairs big model, see Figure 1The method comprises the following steps S110 to S160:

[0050] S110, obtaining the user's first expression text;

[0051] In actual scenarios, when users encounter government affairs problems, they will express their needs or problems to the government affairs system, and the initial expression information will be used as the initial expression text. If the user expresses in the form of voice, the voice will be converted into text as the initial expression text. Since the user's initial expression may not be accurate, it is necessary to further clarify the user's intention.

[0052] S120, cleaning the initial expression text;

[0053] In one embodiment, the cleaning process of the first expression text in S120 may include:

[0054] 1. Removing redundant information from the first expression text;

[0055] Among them, duplicate and similar data are deduplicated to avoid introducing unnecessary noise in the subsequent process.

[0056] 2. Correct or delete any erroneous or abnormal information in the initial expression text;

[0057] Among them, errors and outliers in the initial expression text are checked and corrected or deleted to ensure the accuracy and consistency of the data.

[0058] 3. Remove information in the initial expression text that is irrelevant to the intention recognition of the government affairs industry.

[0059] Among them, features related to government industry intent recognition are extracted from the data, and redundant information irrelevant to the task is removed, thereby improving the accuracy of model recognition. In the training scenario, removing irrelevant information can also improve the training efficiency and performance of the model.

[0060] S130, converting the cleaned initial expression text into an initial expression vector in numerical form;

[0061] In actual scenarios, text vectorization technologies such as Word2Vec and BERT are used to convert the first expression text into a numerical representation suitable for large models. These technologies can convert words or sentences in the text into points in a high-dimensional vector space, which is convenient for the model to calculate and reason.

[0062] Specifically, you can choose a suitable text vectorization model type based on the characteristics and needs of the government industry. For example, the BERT model performs well in natural language processing tasks and is suitable for intent recognition tasks in the government industry.

[0063] After selecting the type of text vectorization model, adjust the parameters according to the clear indicators of vectorization, such as vector dimension, number of training rounds, etc., and train the text vectorization model to optimize the performance and efficiency of the text vectorization model. At the same time, you can use cross-validation and other techniques to evaluate and select the trained text vectorization model to ensure that the text vectorization model has good generalization ability and robustness.

[0064] To this end, in one embodiment, converting the cleaned initial expression text into an initial expression vector in numerical form may include: using a bidirectional encoder representation model of a transformation network to convert the cleaned initial expression text into the initial expression vector.

[0065] Among them, the bidirectional encoder representation model of the transformation network is the BERT model.

[0066] S140, inputting the initial expression vector into a pre-trained intention recognition model to obtain the initial intention of the user;

[0067] It can be seen that when the initial expression vector is input into the intent recognition model, the intent recognition model will output the user's initial intention.

[0068] S150, determining a corresponding question template according to the initial intention; wherein the question template is designed with multiple rounds of question information with the strategy of gradually narrowing the scope of the question;

[0069] In actual scenarios, you can set up a variety of question templates corresponding to the initial intentions in advance. The question templates contain multiple rounds of question information. These question information are set around the initial intention, and the question strategy of the multiple rounds of question information is to gradually narrow the scope of the question, which makes it easier to gradually determine the user's final intention.

[0070] When designing a question template, you need to have a deep understanding of the user consultation characteristics and common scenarios in the government affairs industry, and analyze the core content and characteristics of user needs. Based on the results of the demand analysis, design a series of targeted question templates to ensure that the question templates can cover most user needs. In addition, considering the user's feedback speed and experience, design concise, clear, and easy-to-understand question templates to improve user friendliness.

[0071] When designing question templates, analyze the user's possible intentions and demand changes based on the user's possible answers to provide a basis for the gradual convergence strategy. Based on each user feedback, design a gradual convergence question strategy to accurately locate the user's intention by gradually narrowing the scope of the question.

[0072] S160: Perform multiple rounds of interactions with the user according to the multiple rounds of question information in the question template, and determine the user's final intention according to the interaction results.

[0073] That is, based on the multiple rounds of question information in the question template, questions are asked to the user, and further questions are asked based on the user's feedback. After multiple rounds of questions and feedback, the user's final intention can be determined.

[0074] In one embodiment, the method provided by the embodiment of the present invention may further include:

[0075] The question template is adjusted according to the feedback content of the user during multiple rounds of interaction with the user.

[0076] In one embodiment, before inputting the initial expression vector into the pre-trained intent recognition model, the intent recognition model needs to be trained and deployed. The training and deployment of the intent recognition model may include the following 9 steps:

[0077] 1. Collect multiple demand text materials from the government affairs industry;

[0078] In actual scenarios, it is necessary to clarify the scope of data collection: collect diverse data related to the government affairs industry, including user consultation records, service application information, policy documents, and legal interpretations. These data should cover all aspects of government services to ensure that the intent recognition model can fully understand the characteristics and needs of the government affairs industry. The data must be multi-dimensional, including data at different times, in different scenarios, and from different user groups to simulate complex situations in the real world. In addition, data quality must be guaranteed: when collecting data, pay attention to the source and quality of the data, give priority to authoritative and reliable data sources, and avoid introducing noise and erroneous data.

[0079] 2. Clean each demand text material;

[0080] For example, redundant information in the demand text material is removed, erroneous information and abnormal information in the demand text material is corrected or deleted, and information in the demand text material that is irrelevant to the intention recognition of the government affairs industry is removed.

[0081] 3. Provide the cleaned demand text material to the annotation personnel, so that the annotation personnel can mark the user's intention on each demand text material to obtain a marked demand text material;

[0082] That is, the annotators annotate the user's intention on each requirement text material, thereby obtaining the corresponding annotated requirement text material.

[0083] Among them, the annotators are those with knowledge and experience in the government affairs industry. To ensure the annotation specifications, detailed annotation specifications can be formulated to clarify the definitions and classification standards of various user intentions, and ensure the consistency and accuracy of the annotation results.

[0084] 4. Convert each labeled requirement text material into an expression vector in numerical form, use the expression vector corresponding to each labeled requirement text material as a training sample, and form a training sample set with multiple training samples;

[0085] Since the annotation requirement text material is in text form, it is vectorized to obtain an expression vector in numerical form. An expression vector is used as a training sample, and each training sample forms a training sample set. The training sample set provides strong support for subsequent model training. At the same time, as the data increases and the model performance improves, the training sample set can be continuously updated and improved.

[0086] 5. Choose the model architecture;

[0087] In actual scenarios, first, evaluate industry adaptability: conduct in-depth analysis of the characteristics and needs of the government industry, and evaluate the applicability of different large model architectures in the government field. Large model architectures, such as Transformer structure, GPT, BERT, etc. Then, based on the evaluation results, select or customize the large model architecture that best suits the characteristics of the government industry, including considering the complexity of the model, training efficiency, inference speed, and processing capabilities for government data. If a customized model architecture is chosen, customize and optimize the model structure based on the unique business needs and data characteristics of the government industry.

[0088] 6. According to the training sample set, model training is performed on the basis of the selected model architecture to obtain a converged intent recognition model;

[0089] In the actual scenario, set the training objectives, such as language modeling, text classification, entity recognition, etc., to ensure that the model can learn common language knowledge and structure in the pre-training stage. Perform training and use the prepared training sample set to train the model to ensure that the model has basic language understanding and generation capabilities.

[0090] In an embodiment of the present invention, the training process of the model integrates the government industry data with the training process to ensure that the intent recognition model can learn specific knowledge and patterns in the government field.

[0091] 7. Evaluate the performance of the intent recognition model;

[0092] Among them, the performance of the intent recognition model on government industry data can be evaluated, including indicators such as accuracy, recall rate, F1 value, etc., to obtain the performance evaluation results.

[0093] If the performance evaluation fails, the model parameters need to be set according to the performance evaluation results to ensure the long-term stable operation and performance optimization of the model. After the model is deployed and applied, you can also continue to pay attention to changes in the industry and user needs, and continuously improve and optimize the model.

[0094] 8. If the performance evaluation is qualified, the intention recognition model is connected to the government affairs industry knowledge base, so that the intention recognition model can call government affairs industry knowledge from the government affairs industry knowledge base;

[0095] That is to say, if the performance of the intent recognition model meets the requirements, a connection will be established between the intent recognition model and the government affairs industry knowledge base, so that the intent recognition model can query the required government affairs industry knowledge from the government affairs industry knowledge base when it is executed.

[0096] 9. Deploy the intent recognition model to the application scenario.

[0097] It can be seen that after the above 9 steps, the intent recognition model can be deployed in the application scenario, thereby applying the intent recognition model.

[0098] It can be seen that in the embodiment of the present invention, the trained intent recognition model and question template are deployed to the government service system to ensure the stability and reliability of the system. During the deployment process, the scalability and maintainability of the system should be fully considered for subsequent upgrades and optimizations. Through the front-end interface or voice interaction method of the government service system, multiple rounds of interaction are carried out with users to obtain user needs and feedback in real time. At the same time, the system should support multiple interaction methods to meet the needs of different users. According to the user's feedback, the questions asked are adjusted in real time and the final user intention is identified to provide users with accurate government services or information push. In this process, issues such as user privacy protection and data security should be fully considered to ensure the compliance and security of the system.

[0099] In one embodiment, the method provided by the embodiment of the present invention may further include the following 4 steps:

[0100] 1. Regularly obtain incremental data generated during the intent recognition process in the government affairs industry; each piece of incremental data includes the first expression text, the initial intent, and the final intent;

[0101] 2. Filter out first incremental data from the incremental data, wherein the similarity between the initial intent and the final intent is higher than a first preset value, and use the first incremental data as a positive sample;

[0102] 3. Filter out second incremental data from the incremental data, the similarity between the initial intent and the final intent being lower than a second preset value, and use the second incremental data as a negative sample; wherein the first preset value is greater than the second preset value;

[0103] 4. Use the positive samples and the negative samples to perform incremental training on the intent recognition model.

[0104] It can be seen that in the intent recognition scenario, the incremental data generated is obtained, and then the positive samples and negative samples are determined based on the incremental data, and then the intent recognition model is incrementally trained based on the positive samples and negative samples. Positive samples are samples that the intent recognition model recognizes more accurately, and negative samples are samples that the intent recognition model recognizes with lower accuracy. Based on the positive and negative samples, the intent recognition model is fine-tuned.

[0105] In one embodiment, the method provided by the embodiment of the present invention may further include:

[0106] After each incremental training, performing a performance evaluation on the intent recognition model;

[0107] If the performance evaluation is qualified, the intent recognition model is applied to the application scenario; otherwise, the intent recognition model continues to be trained until the performance evaluation of the intent recognition model is qualified.

[0108] It can be seen that after each fine-tuning, the performance of the intent recognition model on the government industry data is evaluated, including indicators such as accuracy, recall rate, and F1 value, to ensure that the intent recognition model meets business needs.

[0109] The embodiment of the present invention adopts a deep learning model of Transformer or similar architecture, and uses word embedding and sentence embedding technology to convert text into numerical vectors. In addition, a reasonable question template is designed, such as providing multiple-choice questions and fill-in-the-blank questions, to guide users to clarify their question intentions. At the same time, a user feedback mechanism is established to collect user evaluation information on answers and services to optimize the performance of models and templates and improve user satisfaction.

[0110] The embodiment of the present invention proposes a method for intent recognition based on question templates and a large model of the government industry, aiming to solve the problems of inaccurate and inefficient user intent recognition in the government industry, and to provide intelligent and efficient support for government services.

[0111] It is understandable that the embodiments of the present invention involve natural language processing technology and deep learning technology.

[0112] Among them, natural language processing technology is one of the key technologies in the field of artificial intelligence. It studies how to make computers understand and generate human language. In recent years, with the advancement of algorithms, the surge in data volume and the improvement of computing power, natural language processing technology has made great progress. Today's natural language processing technology can already accurately identify semantics, emotions and other information in texts, and can even conduct complex conversational interactions. The maturity of this technology provides solid technical support for the realization of functions such as automated question-and-answer and intelligent recommendations in government services.

[0113] Among them, deep learning is a branch of machine learning, which simulates the learning process of the human brain by building a deep neural network. Deep learning technology has powerful feature extraction and pattern recognition capabilities, and can automatically learn useful information from massive data. In the field of government services, deep learning technology is widely used in user behavior analysis, intent recognition, intelligent recommendation and other aspects. Especially in user intent recognition, the deep learning model can automatically extract the key features for identifying user intent by learning a large number of samples of user questions, thereby achieving high-precision intent recognition.

[0114] In government services, efficiency and speed are key indicators for measuring service quality. By building an intent recognition model and integrating deep learning technology and natural language processing technology, the government service system can automatically process and analyze user questions. This automated processing not only reduces the need for manual intervention and the risk of human error, but also significantly shortens the time users have to wait for responses. The high-speed computing power of the large model ensures fast and accurate responses, greatly improving the overall efficiency of government services.

[0115] User experience is an important part of government services that cannot be ignored. Question templates help users express their needs more accurately through intelligent guidance in a gradually convergent manner, ensuring that questions are answered accurately. This interactive method not only improves the accuracy of question answers, but also increases the interactivity between users and the system, making users feel more cared for and respected. This service method not only improves user satisfaction, but also enhances users' trust and loyalty to government services.

[0116] In government services, the operating costs of manual customer service account for a considerable proportion. By introducing automated processing and intelligent question-and-answer systems, the number and workload of manual customer service can be greatly reduced, thereby reducing the operating costs of government services. The efficient operation of the big model enables the system to provide services to users around the clock and uninterruptedly, reducing the waste of human resources. At the same time, the big model has the ability to continuously learn and optimize, and can continuously improve its recognition accuracy and service quality, further reducing operating costs. This intelligent service method not only reduces operating costs, but also improves the sustainability and competitiveness of government services.

[0117] The embodiments of the present invention combine advanced deep learning technology and natural language processing technology to bring revolutionary changes to government services. This intelligent development model not only improves the efficiency and quality of government services, but also promotes the overall upgrading and transformation of the government industry. Through continuous technological innovation and application practice, government services will be more intelligent, personalized and precise, providing society with more efficient and convenient government services. This will help enhance the overall competitiveness and social influence of the government industry, and promote the government industry to develop in a more intelligent and efficient direction.

[0118] Based on the above analysis, the beneficial effects of the embodiments of the present invention are as follows:

[0119] 1. Improve the accuracy of intent recognition:

[0120] Through multiple rounds of interaction and step-by-step questioning, we can gain a deeper understanding of user needs, gradually narrow down the scope of intent, and finally accurately locate user intent. This gradual convergence strategy greatly improves the accuracy of intent recognition and reduces the possibility of misjudgment and missed judgment.

[0121] 2. Optimize user experience:

[0122] Accurate user intent recognition can ensure that the government service system provides users with services or information that better meet their needs. Users no longer need to spend a lot of time on tedious queries or descriptions, but can quickly and accurately obtain the required services, thereby optimizing the user experience.

[0123] 3. Improve the efficiency of government services:

[0124] The intention recognition method based on the gradual convergence of questions can quickly and accurately identify user intentions, allowing the government service system to respond quickly and provide corresponding services. This greatly improves the efficiency of government services, reduces user waiting time, and improves the overall performance of the government service system.

[0125] 4. Strong adaptability:

[0126] The large model architecture used has strong generalization capabilities and can adapt to changes in different government scenarios and user needs. Through continuous data collection and model training, the intent recognition model can be continuously optimized and improved to adapt to the increasingly complex government industry environment.

[0127] 5. Strong scalability:

[0128] The intent recognition model and question templates can be flexibly adjusted and expanded according to actual needs. New question templates can be easily added, question strategies can be optimized, or new data sources can be introduced to adapt to different government scenarios and changes in user needs. This scalability enables the system to continue to provide strong support for government services.

[0129] 6. Reduce labor costs:

[0130] By introducing intelligent intent recognition technology, a large number of user inquiries and service applications can be automatically processed, reducing the reliance on manual customer service. This reduces the labor cost of government services and improves work efficiency.

[0131] 7. Promote the intelligent development of government services:

[0132] It will promote the development of government services towards intelligence and efficiency. By continuously optimizing and improving functions and performance, the government service system will be able to provide services to users more conveniently and efficiently, and improve the overall level of government services.

[0133] In summary, the embodiments of the present invention improve the efficiency and quality of government services by accurately identifying user intentions. The embodiments of the present invention adopt natural language processing and deep learning technologies to achieve preliminary identification of user intentions by building an intention recognition model. The question template provides a way to refine questions or options, guiding users to gradually clarify their intentions. It can be seen that the present invention has a high recognition accuracy rate. The embodiments of the present invention are suitable for various intelligent question-and-answer scenarios in the government industry, such as policy consultation, administrative approval, online services, etc. Through automated processing and rapid response, manual intervention and waiting time can be significantly reduced, and the efficiency of government services can be improved. At the same time, accurate intention recognition based on deep learning and natural language processing technology can ensure accurate answers to questions and improve user satisfaction. In addition, it can also reduce the operating costs of government services and reduce the number and working hours of manual customer service.

[0134] In a second aspect, an embodiment of the present invention provides an intention recognition device based on a government affairs big model, see Figure 2 , the device 100 comprises:

[0135] A text acquisition module 110 is used to acquire the user's first expression text;

[0136] A text cleaning module 120, used for cleaning the first expression text;

[0137] A vector conversion module 130, used to convert the cleaned primary expression text into a primary expression vector in numerical form;

[0138] A model recognition module 140 is used to input the initial expression vector into a pre-trained intention recognition model to obtain the initial intention of the user;

[0139] The template determination module 150 is used to determine a corresponding question template according to the initial intention; wherein the question template is designed with multiple rounds of question information with the strategy of gradually narrowing the scope of the question;

[0140] The multi-round interaction module 160 is used to perform multi-round interactions with the user according to the multi-round question information in the question template, and determine the user's final intention according to the interaction results.

[0141] In one embodiment, the text cleaning module is specifically used to: remove redundant information in the initial expression text; correct or delete erroneous information and abnormal information in the initial expression text; and remove information in the initial expression text that is irrelevant to the intention recognition of the government affairs industry.

[0142] In one embodiment, the vector conversion module is specifically used to: use a bidirectional encoder representation model of a transformation network to convert the cleaned initial expression text into the initial expression vector.

[0143] In one embodiment, the apparatus further comprises:

[0144] A model training module is used to collect multiple demand text materials in the government affairs industry; clean each demand text material; provide the cleaned demand text materials to the annotation personnel so that the annotation personnel can mark the user intention on each demand text material to obtain a labeled demand text material; convert each labeled demand text material into an expression vector in numerical form, use the expression vector corresponding to each labeled demand text material as a training sample, and form a training sample set with multiple training samples; select a model architecture; according to the training sample set, perform model training on the basis of the selected model architecture to obtain a converged intent recognition model; perform performance evaluation on the intent recognition model; if the performance evaluation is qualified, connect the intent recognition model with the government affairs industry knowledge base so that the intent recognition model can call the government affairs industry knowledge from the government affairs industry knowledge base; deploy the intent recognition model to the application scenario.

[0145] In one embodiment, the apparatus further comprises:

[0146] The template adjustment module is used to adjust the question template according to the user's feedback content during multiple rounds of interaction with the user.

[0147] In one embodiment, the apparatus further comprises:

[0148] A model adjustment module is used to regularly obtain incremental data generated during the intent recognition process in the government affairs industry; wherein each incremental data includes an initial expression text, an initial intent, and the final intent; the first incremental data whose similarity between the initial intent and the final intent is higher than a first preset value is screened out from the incremental data, and the first incremental data is used as a positive sample; the second incremental data whose similarity between the initial intent and the final intent is lower than a second preset value is screened out from the incremental data, and the second incremental data is used as a negative sample; wherein the first preset value is greater than the second preset value; the intent recognition model is incrementally trained using the positive samples and the negative samples.

[0149] In one embodiment, the apparatus further comprises:

[0150] The model evaluation module is used to perform performance evaluation on the intent recognition model after each incremental training; if the performance evaluation is qualified, the intent recognition model is applied to the application scenario; otherwise, the intent recognition model continues to be trained until the performance evaluation of the intent recognition model is qualified.

[0151] It is understandable that the explanation, specific implementation, beneficial effects, examples, etc. of the relevant contents in the device provided in the embodiment of the present invention can be found in the corresponding parts of the method provided in the first aspect, and will not be repeated here.

[0152] In a third aspect, an embodiment of the present invention provides a computer-readable medium, on which computer instructions are stored. When the computer instructions are executed by a processor, the processor executes the method provided in the first aspect.

[0153] Specifically, a system or device equipped with a storage medium can be provided, on which software program code that implements the functions of any of the above-mentioned embodiments is stored, and a computer (or CPU or MPU) of the system or device can be enabled to read and execute the program code stored in the storage medium.

[0154] In this case, the program code itself read from the storage medium can realize the function of any one of the above-mentioned embodiments, and thus the program code and the storage medium storing the program code constitute a part of the present invention.

[0155] The storage medium embodiments for providing the program code include a floppy disk, a hard disk, a magneto-optical disk, an optical disk (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), a magnetic tape, a non-volatile memory card, and a ROM. Alternatively, the program code can be downloaded from a server computer by a communication network.

[0156] In addition, it should be clear that the functions of any of the above embodiments can be implemented not only by executing the program code read by the computer, but also by enabling an operating system operating on the computer to complete part or all of the actual operations based on instructions from the program code.

[0157] In addition, it can be understood that the program code read from the storage medium is written to a memory provided in an expansion board inserted into the computer or to a memory provided in an expansion module connected to the computer, and then based on the instructions of the program code, a CPU installed on the expansion board or expansion module is enabled to perform part or all of the actual operations, thereby realizing the functions of any of the above-mentioned embodiments.

[0158] It is understandable that the explanation, specific implementation methods, beneficial effects, examples, etc. of the relevant contents in the computer-readable medium provided in the embodiment of the present invention can be found in the corresponding parts of the method provided in the first aspect, and will not be repeated here.

[0159] In a fourth aspect, an embodiment of the present specification provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the method in any embodiment of the specification.

[0160] It is understandable that the explanation, specific implementation, beneficial effects, examples, etc. of the relevant contents in the computing device provided in the embodiment of the present invention can be found in the corresponding parts of the method provided in the first aspect, and will not be repeated here.

[0161] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0162] Those skilled in the art should be aware that in one or more of the above examples, the functions described in the present invention can be implemented by hardware, software, widgets, or any combination thereof. When implemented by software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium.

[0163] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made on the basis of the technical solution of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intention recognition method based on a big government model, It is characterized in that include: Get the user's first expression text; Performing cleaning processing on the initial expression text; Convert the cleaned initial expression text into an initial expression vector in numerical form; Inputting the initial expression vector into a pre-trained intention recognition model to obtain the initial intention of the user; Determine a corresponding question template according to the initial intention; wherein the question template is designed with multiple rounds of question information with the strategy of gradually narrowing the scope of the question; Perform multiple rounds of interactions with the user according to the multiple rounds of question information in the question template, and determine the user's final intention according to the interaction results.

2. The method according to claim 1, It is characterized in that The cleaning process of the first expression text includes: Removing redundant information from the initial expression text; Correct or delete erroneous information and abnormal information in the initial expression text; Remove information in the first expression text that is irrelevant to the intent recognition of the government affairs industry.

3. The method according to claim 1, It is characterized in that The method of converting the cleaned initial expression text into an initial expression vector in numerical form includes: using a bidirectional encoder representation model of a transformation network to convert the cleaned initial expression text into the initial expression vector.

4. The method according to claim 1, It is characterized in that Before inputting the initial expression vector into the pre-trained intention recognition model, the method includes: Collect multiple demand text materials from the government affairs industry; Clean each demand text material; Providing the cleaned demand text material to the annotator, so that the annotator annotates the user intention on each demand text material to obtain an annotated demand text material; Convert each labeled requirement text material into an expression vector in numerical form, use the expression vector corresponding to each labeled requirement text material as a training sample, and form a training sample set with multiple training samples; Choose the model architecture; According to the training sample set, model training is performed on the basis of the selected model architecture to obtain a converged intent recognition model; Performing a performance evaluation on the intent recognition model; If the performance evaluation is qualified, the intention recognition model is connected to the government affairs industry knowledge base, so that the intention recognition model can call the government affairs industry knowledge from the government affairs industry knowledge base; Deploy the intent recognition model to the application scenario.

5. The method according to claim 1, It is characterized in that The method further comprises: The question template is adjusted according to the feedback content of the user during multiple rounds of interaction with the user.

6. The method according to claim 1, It is characterized in that The method further comprises: Regularly obtain incremental data generated during the intent recognition process in the government affairs industry; wherein each piece of incremental data includes the first expression text, the initial intent, and the final intent; Filter out first incremental data from the incremental data, wherein the similarity between the initial intent and the final intent is higher than a first preset value, and use the first incremental data as a positive sample; Filter out second incremental data from the incremental data, in which the similarity between the initial intent and the final intent is lower than a second preset value, and use the second incremental data as a negative sample; wherein the first preset value is greater than the second preset value; The positive samples and the negative samples are used to perform incremental training on the intent recognition model.

7. The method according to claim 6, It is characterized in that The method further comprises: After each incremental training, performing a performance evaluation on the intent recognition model; If the performance evaluation is qualified, the intent recognition model is applied to the application scenario; otherwise, the intent recognition model continues to be trained until the performance evaluation of the intent recognition model is qualified.

8. An intention recognition device based on a large government model, It is characterized in that include: A text acquisition module is used to acquire the user's first expression text; A text cleaning module, used for cleaning the first expression text; A vector conversion module, used for converting the cleaned primary expression text into a primary expression vector in numerical form; A model recognition module, used for inputting the initial expression vector into a pre-trained intention recognition model to obtain the initial intention of the user; A template determination module, used to determine a corresponding question template according to the initial intention; wherein the question template is designed with multiple rounds of question information with the strategy of gradually narrowing the scope of the question; The multi-round interaction module is used to perform multi-round interactions with the user according to the multi-round question information in the question template, and determine the user's final intention according to the interaction results.

9. A computer-readable storage medium, It is characterized in that A computer program is stored thereon, and when the computer program is executed in a computer, the computer is caused to execute the method according to any one of claims 1 to 7.

10. A computing device, It is characterized in that The method comprises a memory and a processor, wherein the memory stores executable codes, and when the processor executes the executable codes, the method according to any one of claims 1 to 7 is implemented.

Citation Information

Patent Citations

  • Intention recognition method and device

    CN117151114A

  • Intelligent problem optimization dialogue method and system based on intention recognition of structural information

    CN117933271A

  • Intention identification method and device, equipment and medium

    CN118070790A

  • Electric power business hall intelligent customer service system based on domestic large model and use method

    CN118505237A