Intelligent interaction method and device and electronic equipment

Abstract

The invention discloses an intelligent interaction method and device and electronic equipment. The method comprises the steps of obtaining interaction information input by a target object; intention classification is carried out on the interaction information through an intention classification model to obtain a classification result, and the classification result comprises one of a social communication class, a query class and a transaction class; processing the interaction information according to the classification result to obtain a processing result; and generating reply information of the interaction information according to the processing result, and returning the reply information to the target object. The technical problem of low accuracy of intelligent interaction in related technologies is solved.

Description

Technical Field

[0001] The present application relates to the field of natural language processing, and more specifically, to an intelligent interaction method, device and electronic device. Background Art

[0002] With the application of intelligent technology in the operation and maintenance interaction system, cloud-network intelligent interaction has achieved rapid development. For example, the intelligent operation workbench has achieved one-stop visualization of user perception and operation status, one-stop execution of command and dispatch and production operations, and one-stop control of operation quality and safe operation by promoting standardized and normalized operation of the whole network large screen. However, there is still a problem of low accuracy in intelligent interaction.

[0003] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0004] The embodiments of the present application provide an intelligent interaction method, device and electronic device to at least solve the technical problem of low accuracy of intelligent interaction in related technologies.

[0005] According to one aspect of an embodiment of the present application, an intelligent interaction method is provided, including: obtaining interaction information input by a target object; performing intent classification on the interaction information through an intent classification model to obtain a classification result, wherein the classification result includes one of the following: social communication class, query class, and transaction class; processing the interaction information based on the classification result to obtain a processing result; generating reply information for the interaction information based on the processing result, and returning the reply information to the target object.

[0006] Optionally, the intent classification model is trained in the following manner: obtaining training data for training the intent classification model, and classification annotation data of the training data, wherein the training data includes historical interaction data and artificially synthesized data, and the classification annotation data is used to represent the classification result to which each data in the training data belongs; training the original intent classification model with the training data and the classification annotation data to obtain the intent classification model.

[0007] Optionally, the interaction information is processed according to the classification result to obtain a processing result, including: when the classification result indicates that the interaction information belongs to the social communication category, using the cloud network knowledge big model to process the interaction information to obtain a first processing result; when the classification result indicates that the interaction information belongs to the query category, decomposing the interaction information to obtain a second processing result; when the classification result indicates that the interaction information belongs to the transaction category, decomposing the interaction information into interaction information of multiple query categories, and aggregating the interaction information of multiple query categories to obtain a third processing result.

[0008] Optionally, when the classification result indicates that the interactive information belongs to the query class, the interactive information is decomposed to obtain a second processing result, including: obtaining the original text for entity extraction in the i-th round, wherein i is a positive integer, and when i is 1, the original text is interactive information; determining multiple candidate probabilities for each candidate sub-entity in the original text by multiple methods, wherein each candidate sub-entity corresponds to a candidate probability in each method; determining a weighted candidate probability for each candidate sub-entity based on the multiple candidate probabilities; determining the candidate sub-entity whose weighted candidate probability meets the first condition as the first entity to be extracted in the i-th round, deleting the first entity from the original text to obtain an updated original text, and entering the i+1th round to perform entity extraction again, repeating the entity extraction steps until the second condition is met and stopping the entity extraction, to obtain the second processing result.

[0009] Optionally, multiple candidate probabilities for each candidate sub-entity in the original text are determined in a variety of ways, including: obtaining the minimum split entity in the original text through text regularization; filling the minimum split entity to obtain a first candidate sub-entity; calculating the similarity between each first candidate sub-entity and an entity in an entity library to determine a first candidate probability corresponding to each first candidate sub-entity; determining the second candidate sub-entities of the original text through an entity extraction model, and calculating the similarity between the second candidate sub-entities and the entities in the entity library in the order of the second candidate sub-entities to determine a second candidate probability corresponding to each second candidate sub-entity; vectorizing the original text and the entities in the entity library to obtain a first vector and a second vector; determining the similarity between the first vector and the second vector to obtain a third candidate probability; and determining multiple candidate probabilities for each candidate sub-entity in the original text based on the first candidate probability, the second candidate probability and the third candidate probability.

[0010] Optionally, based on multiple candidate probabilities, a weighted candidate probability of each candidate sub-entity is determined, including: determining a first weight corresponding to the first candidate probability, a second weight corresponding to the second candidate probability, and a third weight corresponding to the third candidate probability; based on the first weight, the first candidate probability, the second weight, the second candidate probability, the third weight, and the third candidate probability, the weighted candidate probability of each candidate sub-entity is determined.

[0011] Optionally, the method also includes: obtaining the historical entity inheritance state of the target object in the previous round of dialogue, the turn switching action of the previous round of dialogue, the probability state distribution result of the previous round of dialogue, and obtaining the output result of the historical dialogue and the current entity inheritance state to which the interaction information belongs; based on the historical entity inheritance state of the previous round of dialogue, the turn switching action of the previous round of dialogue, the probability state distribution result, the output result of the historical dialogue, the interaction information and the current entity inheritance state, determining the probability of the entity in the interaction information switching turns.

[0012] Optionally, based on the processing result, reply information of the interactive information is generated, including: obtaining the target entity in the second processing result, and obtaining the intention of the interactive information; determining the target prompt template with the highest similarity to the target entity and intention from the prompt template set through a text similarity algorithm; and generating reply information of the interactive information based on the target prompt template.

[0013] According to another aspect of an embodiment of the present application, an intelligent interaction device is also provided, including: an acquisition module, used to acquire interaction information input by a target object; a classification module, used to perform intent classification on the interaction information through an intent classification model to obtain a classification result, wherein the classification result includes one of the following: social communication class, query class, transaction class; a processing module, used to process the interaction information according to the classification result to obtain a processing result; a generation module, used to generate reply information of the interaction information according to the processing result, and return the reply information to the target object.

[0014] According to another aspect of the embodiments of the present application, an electronic device is also provided, including: a memory for storing program instructions; a processor, connected to the memory, for executing program instructions to implement the following functions: obtaining interaction information input by a target object; performing intent classification on the interaction information through an intent classification model to obtain a classification result, wherein the classification result includes one of the following: social communication, query, and transaction; processing the interaction information based on the classification result to obtain a processing result; generating reply information for the interaction information based on the processing result, and returning the reply information to the target object.

[0015] According to another aspect of the embodiments of the present application, a non-volatile storage medium is provided, which includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the above-mentioned intelligent interaction method by running the computer program.

[0016] According to another aspect of the embodiments of the present application, a computer program product is provided, including computer instructions, which implement the above-mentioned intelligent interaction method when executed by a processor.

[0017] In an embodiment of the present application, the interaction information input by the target object is obtained; the interaction information is classified by intent through an intent classification model to obtain a classification result, wherein the classification result includes one of the following: social communication class, query class, and transaction class; the interaction information is processed according to the classification result to obtain a processing result; based on the processing result, reply information of the interaction information is generated, and the reply information is returned to the target object, thereby achieving the purpose of accurately understanding and responding to user intentions, thereby achieving the technical effect of improving the accuracy and efficiency of the cloud-network intelligent interaction system, and further solving the technical problem of low accuracy of intelligent interaction in related technologies. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0019] Figure 1 is a hardware structure block diagram of a computer terminal for implementing an intelligent interaction method according to an embodiment of the present application;

[0020] Figure 2 is a flow chart of an intelligent interaction method according to an embodiment of the present application;

[0021] Figure 3 It is a structural diagram of an intelligent interaction device according to an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0023] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] The information collected in the embodiments of the present application is information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with the relevant laws, regulations and standards of the relevant regions, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or reject automated decision-making results; if the user chooses to reject, the expert decision-making process will be entered.

[0025] First, some nouns or terms that appear in the process of explaining the embodiments of the present application are subject to the following explanations:

[0026] BERT (Bidirectional Encoder Representation from Transformers) is a pre-training model for language representation. It emphasizes that it is no longer the traditional unidirectional model pre-training method, but instead adopts a new training strategy to generate a deep bidirectional language representation model.

[0027] LLM (Large Language Model): An artificial intelligence model designed to understand and generate human language. Its typical characteristics are large scale and billions of parameters.

[0028] BI (Business Intelligence) is an information system established for the purpose of providing operational data for decision-making analysis.

[0029] The relevant technologies still have the following problems in the intelligent interaction process: (1) At present, the data island phenomenon of each cloud network system is serious. Although each cloud network business system has the ability to store and query data, there are data barriers between systems, and it is difficult to correlate and analyze data between systems. (2) The existing cloud network data interaction systems are generally single-round or single-indicator modes, lacking multiple rounds of questioning, and it is difficult to ensure the accuracy of data retrieval; at the same time, the existing systems generally only include the query function of data values, lack of analysis and interpretation of data, which is not conducive to the operation and maintenance personnel to quickly analyze data.

[0030] The related technologies usually use template SQL, which can only query indicators or use large language models for knowledge question and answer. The template SQL method needs to determine the pre-made template according to the query frequency, which is time-consuming and labor-intensive, and has poor robustness and is not universal. Based on the large language model, it can usually only answer historical knowledge and cannot perform real-time indicator perception analysis. In addition, the existing intent recognition or multi-round question and answer are mostly based on general text, without considering the particularity of operator indicators. Operator entities are often nested in multiple layers, so the accuracy of such entity queries is poor.

[0031] In order to solve the problems existing in the related art, the embodiment of the present application provides an intelligent interaction method, which can be run on Figure 1 In the computer terminal shown, the computer terminal is explained below.

[0032] The intelligent interaction method embodiment provided in the embodiment of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 FIG. 1 shows a hardware structure block diagram of a computer terminal for implementing an intelligent interaction method. Figure 1 As shown, the computer terminal 10 may include one or more (102a, 102b, ..., 102n are used to illustrate) processors (the processor may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission module 106 for communication functions connected via a wired and / or wireless network. In addition, it may also include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, and a BUS bus. Those skilled in the art can understand that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations are shown.

[0033] It should be noted that the one or more processors and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 10. As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0034] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the intelligent interaction method in the embodiment of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, realizing the above-mentioned intelligent interaction method. The memory 104 may include a high-speed random access memory, and may also include a 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 104 may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0035] The transmission module 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission module 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission module 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0036] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 .

[0037] It should be noted that, in some optional embodiments, the above Figure 1 The computer terminal shown may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of hardware elements and software elements. It should be noted that Figure 1 This is merely one example of a particular embodiment and is intended to illustrate the types of components that may be present in the computer terminal described above.

[0038] In the above-mentioned operating environment, an embodiment of the present application provides an embodiment of an intelligent interaction method. 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, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0039] Figure 2 is a flow chart of an intelligent interaction method according to an embodiment of the present application. Figure 2 As shown, the method comprises the following steps:

[0040] Step S202: Acquire interaction information input by the target object.

[0041] In the above step S202, the user (ie, the target object) can input one or more interactive information to be analyzed in the form of a dialogue through the input box. In some embodiments of the present application, the interactive information can also be input by voice. The interactive information can, for example, be: asking about a certain performance indicator of the cloud network, such as "How many 5G users are there in Beijing?", or issuing an operation instruction, such as "Restart the 4G base station in Changping District", or chatting (ie, social communication), such as "How is the weather?", etc.

[0042] The input module of the system will capture the interactive information input by the target object, for example, through a software interface or hardware device (such as a microphone, touch screen, etc.). During the capture process, the system can also perform some pre-processing, such as voice-to-text conversion and text standardization (removing special characters, correcting spelling errors, etc.) to ensure the accuracy and uniform format of the input information.

[0043] Step S204, performing intent classification on the interaction information through an intent classification model to obtain a classification result, wherein the classification result includes one of the following: social communication class, query class, and transaction class.

[0044] In the above step S204, the intent classification model is a natural language processing technology used to identify and classify the specific purpose or behavior tendency behind the text or voice input by the user. The classification results usually include the following three types:

[0045] 1. Social communication intent: This type of intent is usually related to small talk, greetings, feedback, or non-business-related communication. For example, users can enter "How is the weather?" or "Hi, how is the system running?". Social communication intents are usually easy to process, and the system can use pre-trained general large language models to generate friendly and natural responses to improve the user experience.

[0046] 2. Query: This is the most common type of user intent in cloud network operation and maintenance scenarios, involving specific queries on cloud network status, indicator data, or historical records. For example, "How many online broadband users are there in Beijing?" or "How was the 5G network performance last week?" The processing of query intents requires calling a dedicated data query module, which can include real-time indicator retrieval, historical data query, data fusion analysis, etc., to provide accurate data feedback.

[0047] 3. Transaction type: Transaction type intents involve executing specific cloud network operation and maintenance operations or tasks, such as "restarting the 4G base station in Changping District" or "adjusting the bandwidth allocation of a certain server." The processing of this type of intent usually requires close integration with the cloud network operation execution system to ensure that the system can safely and effectively perform the corresponding operation and maintenance operations according to user instructions.

[0048] The working principle of the intent classification model includes: extracting key features such as keywords, phrases or contextual information from the text or voice input by the user. Based on the extracted features, the model maps the user intent to one of the three types of intent mentioned above. For example, it can be achieved through a pre-trained machine learning model, such as a support vector machine (SVM), a decision tree, a neural network (especially deep learning models such as BERT, LLM, etc.). The model outputs the classification result, and the system calls the corresponding processing module based on the classification result. For example, if the model predicts that the user intent belongs to the transaction class, the system will call the cloud network operation execution module; if the prediction belongs to the query class, the data query and analysis module will be called. Through effective intent classification, the cloud network intelligent interaction system can understand the user's request type more accurately, thereby improving response speed and accuracy, reducing misunderstandings and repeated inquiries, and improving operation and maintenance efficiency and user experience.

[0049] Step S206: Process the interactive information according to the classification result to obtain a processing result.

[0050] In the above step S206, processing the interactive information according to the classification result is the next action taken by the intelligent dialogue system after understanding the user's intention. This process is based on the type of user input (social communication, query, transaction), and the system will call different processing modules or strategies to generate responses to ensure that the reply meets the actual needs of the user. The following is a specific processing step:

[0051] 1. Social communication processing: If the user's intention is classified as social communication, the system will call the chat or social module to generate a response. This module usually uses a general language model to understand and generate natural language dialogues, providing a friendly, non-business communication experience. For example, if the user says "What's the weather like?", the system will query the weather information or respond with a preset weather-related response.

[0052] 2. Query processing: When the user's intention involves cloud network indicators or data query, the system will enter the cloud network data retrieval process. For example, it includes:

[0053] 1) Entity and indicator recognition: extract keywords and entities from user input, such as location, time, specific indicators, etc.;

[0054] 2) Data query: Based on the identified entities and indicators, the system sends query requests to the database or real-time data source to obtain relevant data;

[0055] 3) Data analysis and interpretation: The system can also analyze the acquired data, such as trend analysis, data comparison, etc., and then convert the analysis results into natural language form;

[0056] 4) Generate responses: Organize the queried data or analysis results into easy-to-understand responses and return them to the user.

[0057] 3. Transaction processing: If the user inputs an intention involving the execution of cloud network operation and maintenance operations, the system will call the cloud network operation execution module. For example, it includes:

[0058] 1) Operation instruction analysis: understand the detailed requirements of user instructions, such as the object of operation, operation type, operation parameters, etc.;

[0059] 2) Operation execution and verification: After security verification, the system performs corresponding operation and maintenance operations, such as restarting services, adjusting configurations, etc.

[0060] 3) Status feedback: After the operation is completed, the system will detect the operation results and feedback the status to the user to confirm whether the operation is successful.

[0061] Step S208: Generate reply information of the interaction information according to the processing result, and return the reply information to the target object.

[0062] In the above step S208, after obtaining the processing results, the system converts the structured data or operation status involved into a natural language description. In addition, the system will consider historical conversation information to more accurately understand the user's intention and context, and ensure that the generated reply or the performed operation is based on complete user needs. For example, in a multi-round conversation, if the user asks "What was the number of 5G users yesterday?" and then asks "What about today?", the system will directly query the relevant data of "today" based on the previous round of conversation status (known time "yesterday" and indicator "number of 5G users"), avoiding the need to repeatedly ask specific indicators and time. For another example, in a multi-round conversation, if the user asks for multiple time point data of a certain indicator in succession, the system will naturally link this information in the reply, providing trend analysis or comparison results, rather than just isolated data points. Finally, the system presents the generated reply information to the target object in a user-friendly manner, such as text, voice, chart or any suitable format.

[0063] Through the above steps S202 to S208, the purpose of accurately understanding and responding to the user's intention is achieved, thereby achieving the technical effect of improving the accuracy and efficiency of the cloud network intelligent interaction system, and further solving the technical problem of low accuracy of intelligent interaction in related technologies. The following is an explanation.

[0064] In step S204 of the above-mentioned intelligent interaction method, the intent classification model is trained in the following manner: obtaining training data for training the intent classification model, and classification annotation data of the training data, wherein the training data includes historical interaction data and artificially synthesized data, and the classification annotation data is used to indicate the classification result to which each data in the training data belongs; and training the original intent classification model by using the training data and the classification annotation data to obtain the intent classification model.

[0065] In the embodiment of the present application, the specific training steps are as follows:

[0066] 1. Training data collection: First, collect a large amount of training data for training the intent classification model. This data can come from historical interaction records, including various questions and instructions raised by users through the cloud network intelligent interaction system. In addition, in order to enhance the generalization ability of the model, you can also create artificial synthetic data, including simulating user input in different scenarios, to ensure that the model can cover a variety of possible intent expressions;

[0067] 2. Data preprocessing: Clean and preprocess the collected training data, such as removing irrelevant information, correcting text errors, standardizing expressions, etc., to ensure data quality so that the model can learn from clear and consistent input;

[0068] 3. Classification labeling: Classify and label each piece of training data, that is, assign one or more intent classification labels to each piece of data. For example, for the data "How many 5G users are there in Beijing?", the labeling can be set to "query category", "specific indicator query", etc.;

[0069] 4. Build a model: Select a suitable machine learning model as the original intent classification model. It can be a deep learning model, such as BERT, Transformer, etc., or a traditional machine learning model, such as support vector machine (SVM), decision tree, etc.;

[0070] 5. Training model: Use the collected training data and the corresponding classification annotated data to train the original intent classification model. During the training process, the model attempts to learn the association between the data and the intent label, and adjusts the model parameters to minimize the prediction error, that is, the difference between the classification annotated data and the model prediction results. Through multiple iterations, until the model reaches a satisfactory performance indicator, a trained intent classification model is obtained.

[0071] In step S206 of the above-mentioned intelligent interaction method, the interaction information is processed according to the classification result to obtain a processing result, including: when the classification result indicates that the interaction information belongs to the social communication category, the cloud network knowledge big model is used to process the interaction information to obtain a first processing result; when the classification result indicates that the interaction information belongs to the query category, the interaction information is decomposed to obtain a second processing result; when the classification result indicates that the interaction information belongs to the transaction category, the interaction information is decomposed into interaction information of multiple query categories, and the interaction information of multiple query categories is aggregated to obtain a third processing result.

[0072] In an embodiment of the present application, when the interactive information input by the user is classified as a social communication type, the system uses the cloud network knowledge model for processing. The cloud network knowledge model is a pre-trained language model that can understand and generate natural language dialogues, and is particularly suitable for non-business communications such as chatting, greetings, or feedback. This type of model usually contains a large amount of general knowledge and language patterns, which can communicate in a friendly manner and enhance the user experience. The processing steps are as follows: The system calls the cloud network knowledge model as a processing module, and based on the user's input, the cloud network knowledge model generates a natural language reply, which can be a direct answer to the user's question, or an invitation or feedback to continue the conversation.

[0073] If the interactive information belongs to the query category, the system will enter the cloud network data retrieval process. This type of processing usually involves specific queries on cloud network status, indicator data or historical records, and requires detailed analysis and data processing capabilities. The specific steps are as follows: Through natural language processing technologies, such as entity extraction and intent recognition models, key query intentions and entities such as location, time, specific indicators, etc. are identified from the text entered by the user. Based on the identified intentions and entities, the system sends a query request to the database or real-time data source to obtain relevant data or information. The system analyzes the acquired data, such as trend analysis, data comparison, etc., and then organizes the analysis results into natural language to provide the specific information required by the user.

[0074] For transaction-type input, the system decomposes it into multiple executable queries or operations and summarizes the results. This type of processing involves executing cloud network operation and maintenance operations or tasks, such as restarting services, configuration adjustments, etc. The specific processing method is as follows: decompose the transaction instructions entered by the user into a series of more specific queries or operations. For example, if the user requests to "adjust the 4G base station configuration in Changping District", it needs to be decomposed into steps such as querying the current configuration, calculating the new configuration, and performing configuration adjustments. For each decomposed query or operation, the system calls the corresponding module for processing, such as the data query module, the operation execution module, etc. The results of all queries or operations are summarized to generate a comprehensive transaction processing result, which may include, for example, status feedback on whether the operation is successful, changes in the cloud network status after the operation, etc.

[0075] For each processing result under each category, the system will generate corresponding reply information based on the processing logic and the acquired data, ensuring that the reply content is not only accurate but also able to meet the specific needs of users and provide an efficient service experience.

[0076] In the above steps, when the classification result indicates that the interactive information belongs to the query class, the interactive information is decomposed to obtain a second processing result, including: obtaining the original text for entity extraction in the i-th round, wherein i is a positive integer, and when i is 1, the original text is interactive information; determining multiple candidate probabilities for each candidate sub-entity in the original text by multiple methods, wherein each candidate sub-entity corresponds to a candidate probability in each method; determining a weighted candidate probability for each candidate sub-entity based on the multiple candidate probabilities; determining the candidate sub-entity whose weighted candidate probability meets the first condition as the first entity to be extracted in the i-th round, deleting the first entity from the original text to obtain an updated original text, and entering the i+1th round to perform entity extraction again, repeating the entity extraction steps until the second condition is met and stopping the entity extraction, and obtaining the second processing result.

[0077] In the embodiment of the present application, when the classification result indicates that the interactive information belongs to the query class, the system performs detailed entity extraction and intent understanding on the interactive information to generate accurate query instructions and obtain the required data. This process involves multiple rounds of entity extraction, each of which determines the entities to be extracted based on a specific strategy until the stop condition is met. The following explains this process in detail:

[0078] 1. Get the original text: For the i-th round of entity extraction, the system first gets the original text for entity extraction in this round. When i is 1, the original text is the initial user interaction information. In subsequent rounds, the original text is the text updated after the previous round of entity extraction.

[0079] 2. Probability determination of candidate sub-entities: The system determines multiple candidate probabilities for each candidate sub-entity in the original text in a variety of ways.

[0080] 3. The system performs weighted calculation on the candidate probabilities obtained by each of the above methods to obtain the weighted candidate probability of each candidate sub-entity. The weighting coefficient can reflect the accuracy and reliability of different methods to ensure that the final entity is more accurate.

[0081] 4. First entity determination and text update: The system selects the candidate sub-entity that meets the first condition (such as the candidate probability exceeds a certain threshold) based on the weighted candidate probability and determines it as the first entity to be extracted in the i-th round. Then, the determined first entity is deleted from the original text to obtain the updated original text. The updated text will be used for the next round of entity extraction (i+1th round).

[0082] 5. Multiple rounds of entity extraction and stop conditions: The system repeats the above steps of entity extraction until the second condition is met and entity extraction is stopped. The second condition may include one of the following: entity extraction is completed or there are no more entities to extract or the maximum round is reached.

[0083] 6. Finally, through multiple rounds of entity extraction, the system can accurately understand the user's query intent and specific indicator requirements and obtain the second processing result.

[0084] This multi-round entity extraction method can effectively handle complex queries, especially those involving multiple entities and indicators. It improves the query accuracy and efficiency of the cloud network intelligent interaction system by gradually refining and understanding user input. At the same time, by dynamically adjusting and updating the original text, the system can better adapt to multi-round dialogue scenarios and achieve continuous and coherent dialogue processing.

[0085] For example, due to the particularity of cloud network indicator entities, such as "number of 5G users", "number of broadband online users", etc., they usually contain multiple sub-entities such as "5G", "number of users", "online", etc., and the traditional entity extraction model cannot effectively extract them. At the same time, the same entity has multiple expressions, such as "number of 5G users" contains "number of 5G logged-in users", "number of 5G online users", "number of 5G active users" and other similar words. In response to this situation, the embodiment of the present application proposes a three-level model of "decomposition-combination-recognition" for entities, while integrating the traditional extraction model and ensuring the accuracy of recognition through the weighted model.

[0086] The following text input is "How many 5G outages and large-scale failures are there in Changping?" At the same time, the list of indivisible entities of cloud network indicators is simplified for easy explanation. The simplified indivisible entity list includes "5G", "number of users", "number of outages", "large area", "failure", "4G", "Changping" entities; the cloud network indicator entity library includes "number of 5G users", "number of 5G outages", "large-scale failures", "Changping", "5G base station failures" and other entities.

[0087] (a) First, the input sentence is matched with the indivisible entities to obtain five indivisible entities, namely “5G”, “broken station”, “large area”, “fault” and “Changping”. Then, the slot templates of the indivisible entities are filled to obtain multiple first candidate sub-entities such as “5G broken station”, “large area fault” and “5G fault”. Then, the text is vectorized and the similarity is calculated with the cloud network indicator entity library in turn to obtain entities p with a probability greater than 0.7. 1 (Changping) = 1, p 1 (5G base station failure) = 0.75, p 1 (5G disconnected sites) = 0.95, p 1 (large area failure) = 1;

[0088] (b) The second candidate sub-entity of the original text is determined through the entity extraction model. The first sentence start, end = 0, 0 is cyclically truncated and the correlation is calculated with the cloud network index entity library. The entity with higher correlation is selected as the candidate. The index extracted from this entity is p 2 (Changping) = 1, p 2 (5G base station failure) = 0.63, p 2 (5G disconnected sites) = 0.93, p 2 (large area failure) = 1;

[0089] (c) Directly vectorize the original text to obtain the first vector, and then perform similarity calculation with the entity list (i.e., the second vector corresponding to the entity in the entity library). In this example, p 3 (Changping) = 0.51, p 3 (5G base station failure) = 0.59, p 3 (5G disconnected sites) = 0.61, p 3 (large area failure) = 0.8;

[0090] (d) According to P(s i )=0.6*(s i )+0.2*p 2 (s i )+0.2*p 3 (s i ), the probability of large-scale failure is the highest in the first round, and the “large-scale failure” entity is extracted;

[0091] (e) Delete the “large-scale failure” entity in the original sentence and get “What is the total number of 5G disconnected stations in Changping?” Perform the second round and follow the above process to get the “Number of 5G disconnected stations” entity. Then perform the third round to get the “Changping” entity and complete this round.

[0092] In the above steps, multiple candidate probabilities of each candidate sub-entity in the original text are determined in a variety of ways, including: obtaining the minimum split entity in the original text through text regularization; filling the minimum split entity to obtain the first candidate sub-entity; calculating the similarity between each first candidate sub-entity and the entity in the entity library to determine the first candidate probability corresponding to each first candidate sub-entity; determining the second candidate sub-entity of the original text through the entity extraction model, and calculating the similarity between the second candidate sub-entities and the entities in the entity library in the order of the second candidate sub-entities to determine the second candidate probability corresponding to each second candidate sub-entity; vectorizing the original text and the entities in the entity library to obtain the first vector and the second vector; determining the similarity between the first vector and the second vector to obtain the third candidate probability; and determining multiple candidate probabilities for each candidate sub-entity in the original text based on the first candidate probability, the second candidate probability and the third candidate probability.

[0093] In the embodiment of the present application, in order to more accurately identify and understand user intent, text regularization, entity extraction model and vectorized similarity calculation are combined to comprehensively evaluate the relevance and possibility of candidate sub-entities and entities in the cloud network system. The specific steps are as follows:

[0094] 1. The system first uses predefined regular expressions to extract the smallest segmented entities in the original text. These entities are usually the most basic elements in cloud network indicator queries, such as "4G", "number of users", "online", etc.

[0095] 2. Fill the slots of the minimum segmentation entities obtained through text regularization, that is, combine these basic elements into possible query entities. For example, combine "4G" and "number of users" into "number of 4G users". Such a first candidate sub-entity is a preliminary guess based on the rules, which provides a candidate range for subsequent fine recognition;

[0096] 3. The system calculates the similarity between each first candidate sub-entity and the entities in the entity library (a database containing all known entities in the cloud network system). This can be achieved by calculating text similarity scores, such as edit distance, cosine similarity, etc., to determine the degree of match between each candidate sub-entity and the entity in the entity library, that is, the first candidate probability p 1 (s i ), s i is each candidate sub-entity, p is the probability, and in the embodiment of the present application, only entities with a probability greater than 0.7 may be selected;

[0097] 4. In addition to rule-based methods, the system also uses entity extraction models (such as Transformer-based UIE or BART models) to identify second candidate sub-entities in the original text. The model predicts the probability of an entity by understanding the text context, which can identify entities that are not covered by rule-based methods;

[0098] 5. For each second candidate sub-entity, the system calculates the similarity with the entities in the entity library in the order of their appearance in the original text, and obtains the candidate probability of each second candidate sub-entity, that is, the second candidate probability p 2 (s i ). This sequential calculation helps capture the contextual information of the entity and improves the accuracy of recognition. The specific formula is: 2 (s i )=[X start ;X end ;δ(s i )], where X start The feature vector representing the starting position of the second candidate sub-entity in the original text, X end The feature vector representing the end position of the second candidate sub-entity in the original text, which is extracted from the hidden layer of the Transformer model to represent the information of the entity boundary, δ(s i ) is the semantic vector of the second candidate sub-entity, which is also generated by the Transformer model and reflects the semantic features of the entity. After calculating the second candidate probability of each second candidate sub-entity, a threshold condition is set, for example, candidate entities with a probability greater than 0.6 are considered high confidence entities. This condition is used to filter out items that the model considers to be most likely to be key entities in the query intent, thereby reducing noise and improving recognition accuracy.

[0099] 6. The system vectorizes the original text and the entities in the entity library to obtain the first vector of the original text and the second vector of the entities in the entity library. Then the similarity between the first vector and the second vector is calculated as the third candidate probability p 3 (s i ), the vectorized similarity calculation can capture the semantic information of entities, for example, only taking entities with a probability greater than 0.4.

[0100] According to the candidate probabilities (first candidate probability, second candidate probability, third candidate probability) obtained in the above three ways, multiple candidate probabilities of each candidate sub-entity in the original text can be determined.

[0101] In the above steps, the weighted candidate probability of each candidate sub-entity is determined based on multiple candidate probabilities, including: determining a first weight corresponding to the first candidate probability, a second weight corresponding to the second candidate probability, and a third weight corresponding to the third candidate probability; based on the first weight, the first candidate probability, the second weight, the second candidate probability, the third weight, and the third candidate probability, the weighted candidate probability of each candidate sub-entity is determined.

[0102] In the embodiment of the present application, the weighted candidate probability P(s i ) is calculated as follows:

[0103] P(s i )=θ 1 *p 1 (s i )+θ 2 *p 2 (s i )+θ 3 *p 3 (s i )

[0104] Among them, θ 1 represents the first weight, θ 2 represents the second weight, θ 3 represents the third weight, p 1 (s i ) represents the first candidate probability, p 2 (s i ) represents the second candidate probability, p 3 (s i ) represents the third candidate probability. In an optional embodiment, θ 1 Take 0.6, θ 2 Take 0.2, θ 3 Take 0.2. After obtaining the weighted candidate probability, take the entity with the largest candidate probability, and the weighted candidate probability is greater than 0.9 as the entity to be extracted this time.

[0105] In the above-mentioned intelligent interaction method, the method also includes: obtaining the historical entity inheritance state of the target object in the previous round of dialogue, the turn switching action of the previous round of dialogue, the probability state distribution result of the previous round of dialogue, and obtaining the output result of the historical dialogue and the current entity inheritance state to which the interaction information belongs; based on the historical entity inheritance state of the previous round of dialogue, the turn switching action of the previous round of dialogue, the probability state distribution result, the output result of the historical dialogue, the interaction information and the current entity inheritance state, determine the probability of the entity in the interaction information switching turns.

[0106] In the embodiment of the present application, entities and intentions will be recorded for entities extracted in the previous round or several previous rounds, and the record template takes into account the particularity of cloud-network interaction. An example is as follows:

[0107] type (entity / intent type) indicator (specific recorded entity) gap (difference from the current round) Query-Intent Business development indicator query 1 index Number of active 5G users 1 Time-Entity 2024-05-11 20:00:00 2 Location-Entity Beijing 1

[0108] In the embodiment of the present application, the probability of an entity in the interaction information switching round can be determined by the following formula:

[0109]

[0110] Among them, η is a constant used to adjust the sensitivity of switching; e(s) represents the probability state distribution result of the previous round of dialogue, e′(s′) is the probability distribution of whether to switch rounds in this round, that is, the probability of entities in the interaction information switching rounds, w′ is the input of the current cloud-network interaction round, that is, the above-mentioned interaction information, is the output result of the historical round (or the output result of the historical conversation), s′ is the current state (whether an entity or intention from the previous round is inherited to this round), that is, the above-mentioned current entity inheritance state, s represents the historical state of the previous round, that is, the historical entity inheritance state of the previous conversation, and c represents the action of the previous round (whether to switch), that is, the turn switching action of the previous conversation.

[0111] According to the above formula, it is calculated whether each entity is saved to this round (currently it is set to be retained if the probability is greater than 0.85, otherwise it is not retained), and then a comprehensive evaluation is made on whether to switch in this round.

[0112] Example: Taking the above table as an example, the new round of questions is "What is the number of 4G active users?", and the queries on this entity such as "Beijing", "2024-05-11 20:00:00", "5G active users", and "Business development indicator query" in the historical rounds are calculated respectively. The entity extracted in this round is "4G active users". According to the matching, "Beijing" and "2024-05-11 20:00:00" are retained, and the current round is not switched.

[0113] In step S208 of the above-mentioned intelligent interaction method, reply information of the interactive information is generated based on the processing result, including: obtaining the target entity in the second processing result, and obtaining the intention of the interactive information; determining the target prompt template with the highest similarity to the target entity and intention from the prompt template set through a text similarity algorithm; and generating reply information of the interactive information based on the target prompt template.

[0114] In the embodiment of the present application, different prompt templates are designed for different business types by prefabricating prompt templates, so as to maximize the analysis potential of the large language model and realize the analysis of indicators.

[0115] Example: Taking the number of 5G users, the cloud network situation, and the RRC connection success rate as a brief case for reference, the example template is as follows:

[0116]

[0117] The system obtains the target entity from the second processing result, that is, the cloud network metrics or related entities finally determined through multiple rounds of entity extraction. At the same time, the system also obtains the intention of the interaction information. The system uses a text similarity algorithm to determine the target prompt template with the highest similarity to the target entity and intention from a preset set of prompt templates. The set of prompt templates contains prompt templates for different entities and intentions, which are designed to stimulate the analysis and response capabilities of the large language model (LLM).

[0118] According to the entities and intentions of the interaction information extracted previously, fill them into the corresponding SQL template slots, complete the corresponding SQL metric query, obtain real-time metrics, and call the vector similarity model to implement the retrieval of the Faiss vector database and extract the most similar business data knowledge.

[0119] The text similarity algorithm can include, for example, cosine similarity, Jaccard similarity, etc., which are used to calculate the semantic similarity between entities, intentions, and templates. The similarity calculation after vectorization is as follows, where A and B are the embedding vectors of the prompt template key and the interaction information respectively. The similarity formula is as follows:

[0120]

[0121] Calculate the similarity according to the above steps, extract the prompt template with the highest similarity by matching the key value, that is, the target prompt template. The system combines the target entity and the cloud network metrics queried from the database in real time with the target prompt template. For example, insert the entity and metric values into specific positions of the template to generate a complete and specific information-containing prompt. Send the prompt that integrates the target entity, real-time metrics, and target prompt template into the large language model (LLM). The large model uses its powerful language understanding and generation capabilities to generate natural language response information according to the entity, metric, and intention information in the prompt.

[0122] The generated response information can also go through post-processing steps, such as removing redundant information, formatting data, adding additional explanations, etc., to ensure the clarity and readability of the information. The post-processing steps may also include harmlessness detection to ensure the security and compliance of the response information.

[0123] The intelligent interaction method provided in the embodiment of the present application designs a hierarchical intent system and adopts a hierarchical intent classification structure to perform fine-grained intent decomposition on each sub-business section; constructs intent data by means of template construction and language model expansion, and recalls and sorts entities in combination with an indicator vector retrieval model to increase the accuracy of overall indicator queries. By designing a dialogue state template, the dialogue state is filled and maintained according to key information such as intent and entity extracted in each round of dialogue, and the missing information required for the current query is supplemented according to the historical dialogue state to achieve the association and correction between each round of dialogue; at the same time, a multi-round intent correction and update model is constructed to determine whether to correct the current intent and indicator and update the dialogue state according to the historical dialogue state and the current input statement. Based on the vector-based similarity semantic retrieval capability, static cloud network knowledge and dynamic business performance indicators are integrated, and through the dynamic prompt matching method, data intelligent analysis based on the large model is completed to realize intelligent perception analysis of cloud network business.

[0124] Figure 3 is a structural diagram of an intelligent interactive device according to an embodiment of the present application, such as Figure 3 As shown, the device comprises:

[0125] An acquisition module 30, used to acquire the interaction information input by the target object;

[0126] A classification module 32, configured to classify the interaction information by using an intention classification model to obtain a classification result, wherein the classification result includes one of the following: social communication class, query class, and transaction class;

[0127] A processing module 34 is used to process the interactive information according to the classification result to obtain a processing result;

[0128] The generating module 36 is used to generate reply information of the interactive information according to the processing result, and return the reply information to the target object.

[0129] The above-mentioned intelligent interaction device also includes a training module 38, which is used to train the intent classification model. Specifically, the intent classification model is trained in the following manner: obtaining training data for training the intent classification model, and classification annotation data of the training data, wherein the training data includes historical interaction data and artificially synthesized data, and the classification annotation data is used to indicate the classification result to which each data in the training data belongs; the original intent classification model is trained by the training data and the classification annotation data to obtain the intent classification model.

[0130] In the processing module in the above-mentioned intelligent interactive device, the processing module is also used to, when the classification result indicates that the interactive information belongs to the social communication category, use the cloud network knowledge big model to process the interactive information to obtain a first processing result; when the classification result indicates that the interactive information belongs to the query category, decompose the interactive information to obtain a second processing result; when the classification result indicates that the interactive information belongs to the transaction category, decompose the interactive information into interactive information of multiple query categories, and summarize the interactive information of multiple query categories to obtain a third processing result.

[0131] In the processing module in the above-mentioned intelligent interactive device, the processing module is also used to obtain the original text for entity extraction in the i-th round, wherein i is a positive integer, and when i is 1, the original text is interactive information; determine multiple candidate probabilities for each candidate sub-entity in the original text by multiple methods, wherein each candidate sub-entity corresponds to a candidate probability in each method; determine the weighted candidate probability of each candidate sub-entity based on the multiple candidate probabilities; determine the candidate sub-entity whose weighted candidate probability meets the first condition as the first entity to be extracted in the i-th round, delete the first entity from the original text to obtain an updated original text, and enter the i+1th round to extract the entity again, repeat the entity extraction steps until the second condition is met and stop extracting the entity to obtain a second processing result.

[0132] In the processing module in the above-mentioned intelligent interactive device, the processing module is also used to obtain the minimum split entity in the original text through text regularization; fill the minimum split entity to obtain the first candidate sub-entity; calculate the similarity between each first candidate sub-entity and the entity in the entity library, and determine the first candidate probability corresponding to each first candidate sub-entity; determine the second candidate sub-entities of the original text through the entity extraction model, and calculate the similarity between the second candidate sub-entities and the entities in the entity library in the order of the second candidate sub-entities, and determine the second candidate probability corresponding to each second candidate sub-entity; vectorize the original text and the entities in the entity library to obtain the first vector and the second vector; determine the similarity between the first vector and the second vector to obtain the third candidate probability; determine multiple candidate probabilities for each candidate sub-entity in the original text based on the first candidate probability, the second candidate probability and the third candidate probability.

[0133] In the processing module in the above-mentioned intelligent interaction device, the processing module is also used to determine a first weight corresponding to the first candidate probability, a second weight corresponding to the second candidate probability, and a third weight corresponding to the third candidate probability; based on the first weight, the first candidate probability, the second weight, the second candidate probability, the third weight, and the third candidate probability, the weighted candidate probability of each candidate sub-entity is determined.

[0134] The above-mentioned intelligent interaction device also includes a determination module 40, which is used to obtain the historical entity inheritance state of the target object in the previous round of dialogue, the turn switching action of the previous round of dialogue, the probability state distribution result of the previous round of dialogue, and the output result of the historical dialogue and the current entity inheritance state to which the interaction information belongs; based on the historical entity inheritance state of the previous round of dialogue, the turn switching action of the previous round of dialogue, the probability state distribution result, the output result of the historical dialogue, the interaction information and the current entity inheritance state, determine the probability of the entity in the interaction information switching the turn.

[0135] In the generation module in the above-mentioned intelligent interactive device, the generation module is also used to obtain the target entity in the second processing result, and to obtain the intention of the interactive information; determine the target prompt template with the highest similarity to the target entity and intention from the prompt template set through the text similarity algorithm; and generate reply information of the interactive information based on the target prompt template.

[0136] It should be noted that Figure 3 The intelligent interactive device shown is used to perform Figure 2 The intelligent interaction method shown, therefore the relevant explanations in the above intelligent interaction method are also applicable to the intelligent interaction device, and will not be repeated here.

[0137] An embodiment of the present application also provides an electronic device, which includes a memory and a processor, wherein the memory is used to store program instructions; the processor is connected to the memory, and is used to execute program instructions to implement the following functions: obtaining interaction information input by a target object; performing intent classification on the interaction information through an intent classification model to obtain a classification result, wherein the classification result includes one of the following: social communication, query, and transaction; processing the interaction information based on the classification result to obtain a processing result; generating reply information for the interaction information based on the processing result, and returning the reply information to the target object.

[0138] It should be noted that the above electronic equipment is used to execute Figure 2 The intelligent interaction method shown, therefore the relevant explanations and instructions in the above intelligent interaction method are also applicable to the electronic device and will not be repeated here.

[0139] An embodiment of the present application also provides a non-volatile storage medium, which includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the following intelligent interaction method by running the computer program: obtaining interaction information input by a target object; performing intent classification on the interaction information through an intent classification model to obtain a classification result, wherein the classification result includes one of the following: social communication, query, and transaction; processing the interaction information based on the classification result to obtain a processing result; generating reply information for the interaction information based on the processing result, and returning the reply information to the target object.

[0140] It should be noted that the above non-volatile storage medium is used to execute Figure 2 The intelligent interaction method shown, therefore the relevant explanations and instructions in the above intelligent interaction method are also applicable to the non-volatile storage medium and will not be repeated here.

[0141] An embodiment of the present application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the intelligent interaction method in each embodiment of the present application.

[0142] The embodiments of the present application also provide a computer program, which, when executed by a processor, implements the steps of the intelligent interaction method in each embodiment of the present application.

[0143] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0144] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0145] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0146] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0147] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0148] If the integrated unit is implemented in the form of 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 the present 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, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk and other media that can store program codes.

[0149] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. An intelligent interaction method, characterized in that: include: Obtain the interactive information input by the target object; Performing intent classification on the interaction information through an intent classification model to obtain a classification result, wherein the classification result includes one of the following: social communication class, query class, and transaction class; Processing the interactive information according to the classification result to obtain a processing result; Generate reply information of the interaction information according to the processing result, and return the reply information to the target object.

2. The method according to claim 1, characterized in that: The intent classification model is trained in the following way: Acquire training data for training the intent classification model, and classification annotation data of the training data, wherein the training data includes historical interaction data and artificially synthesized data, and the classification annotation data is used to indicate the classification result to which each data in the training data belongs; The original intent classification model is trained using the training data and the classification annotation data to obtain the intent classification model.

3. The method according to claim 1, characterized in that The interactive information is processed according to the classification result to obtain a processing result, including: When the classification result indicates that the interaction information belongs to the social communication category, the interaction information is processed using the cloud network knowledge big model to obtain a first processing result; When the classification result indicates that the interaction information belongs to the query class, decomposing the interaction information to obtain a second processing result; When the classification result indicates that the interaction information belongs to the transaction class, the interaction information is decomposed into interaction information of multiple query classes, and the interaction information of the multiple query classes is aggregated to obtain a third processing result.

4. The method according to claim 3, characterized in that When the classification result indicates that the interaction information belongs to the query class, the interaction information is decomposed to obtain a second processing result, including: Obtaining the original text for entity extraction in the i-th round, where i is a positive integer, and when i is 1, the original text is the interactive information; Determine multiple candidate probabilities for each candidate sub-entity in the original text in multiple ways, wherein each candidate sub-entity corresponds to a candidate probability in each way; Determining a weighted candidate probability of each candidate sub-entity based on the multiple candidate probabilities; The candidate sub-entity whose weighted candidate probability satisfies the first condition is determined as the first entity to be extracted in the i-th round, the first entity is deleted from the original text to obtain an updated original text, and the i+1-th round is entered to perform entity extraction again, and the entity extraction steps are repeated until the second condition is met, and the entity is stopped to obtain the second processing result.

5. The method according to claim 4, characterized in that Multiple candidate probabilities of each candidate sub-entity in the original text are determined in multiple ways, including: Obtaining the minimum segmentation entity in the original text through text regularization; Filling the minimum segmentation entity to obtain a first candidate sub-entity; Calculate the similarity between each first candidate sub-entity and the entities in the entity library to determine the first candidate probability corresponding to each first candidate sub-entity; Determine the second candidate sub-entities of the original text through the entity extraction model, calculate the similarity of the second candidate sub-entities and the entities in the entity library in the order of the second candidate sub-entities, and determine the second candidate probability corresponding to each second candidate sub-entity; Vectorizing the original text and the entities in the entity library to obtain a first vector and a second vector; Determine the similarity between the first vector and the second vector to obtain a third candidate probability; A plurality of candidate probabilities of each candidate sub-entity in the original text are determined according to the first candidate probability, the second candidate probability and the third candidate probability.

6. The method according to claim 5, characterized in that Determining a weighted candidate probability of each candidate sub-entity according to the multiple candidate probabilities includes: Determine a first weight corresponding to the first candidate probability, a second weight corresponding to the second candidate probability, and a third weight corresponding to the third candidate probability; A weighted candidate probability of each candidate sub-entity is determined according to the first weight, the first candidate probability, the second weight, the second candidate probability, the third weight, and the third candidate probability.

7. The method according to claim 1, characterized in that The method further comprises: Obtaining the historical entity inheritance state of the target object in the previous round of dialogue, the turn switching action of the previous round of dialogue, the probability state distribution result of the previous round of dialogue, and the output result of the historical dialogue and the current entity inheritance state to which the interaction information belongs; The probability of the entity in the interaction information switching the turn is determined according to the historical entity inheritance state of the previous round of dialogue, the turn switching action of the previous round of dialogue, the probability state distribution result, the output result of the historical dialogue, the interaction information and the current entity inheritance state.

8. The method according to claim 4, characterized in that Generating reply information of the interaction information according to the processing result, including: Acquire the target entity in the second processing result, and acquire the intention of the interaction information; Determine, from the prompt template set, a target prompt template having the highest similarity to the target entity and the intent by using a text similarity algorithm; Generate reply information of the interactive information according to the target prompt template.

9. An intelligent interactive device, characterized in that: include: An acquisition module, used to acquire the interactive information input by the target object; A classification module, configured to classify the interaction information by intent through an intent classification model to obtain a classification result, wherein the classification result includes one of the following: social communication, query, and transaction; A processing module, used for processing the interaction information according to the classification result to obtain a processing result; A generating module is used to generate reply information of the interaction information according to the processing result, and return the reply information to the target object.

10. An electronic device, characterized in that: include: A memory for storing program instructions; A processor, connected to the memory, is used to execute program instructions to implement the following functions: obtaining interactive information input by a target object; Performing intent classification on the interaction information through an intent classification model to obtain a classification result, wherein the classification result includes one of the following: social communication, query, and transaction; The interactive information is processed according to the classification result to obtain a processing result; based on the processing result, reply information of the interactive information is generated, and the reply information is returned to the target object.

11. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the intelligent interaction method described in any one of claims 1 to 8 by running the computer program.

12. A computer program product comprising computer instructions, characterized in that: When the computer instructions are executed by a processor, the intelligent interaction method described in any one of claims 1 to 8 is implemented.

Images