Intelligent reply method for user service, medium and electronic equipment

Abstract

The invention provides an intelligent reply method for a user service, a medium and electronic equipment, and can input current input information and historical context information of a user into a deep learning model to obtain an identified user intention, and then input the user intention and a question decomposition instruction into the deep learning model to obtain decomposed sub-intentions, thereby improving the user service reply efficiency. Acquiring skill indication information according to the sub-intention, inputting the sub-intention, the skill indication information and the user associated information into a deep learning model to acquire reply content, inputting the reply content and a content retouching instruction into the deep learning model, acquiring the reply content after combined retouching, and feeding back the reply content to the user, so that the user intention can be accurately understood; and accurate reply contents and services related to the business are provided according to the user association information in combination with the user demand, so that the problem that the intelligent customer service cannot accurately reply the intention of the user in combination with the business information in the prior art is solved, man-machine conversation can be more natural, smoother and more concise, and the conversation efficiency and service experience of the user are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an intelligent reply method, medium and electronic device for user services. Background Art

[0002] At present, artificial intelligence technology (AI) has penetrated into all aspects of social life. Typical artificial intelligence solutions include intelligent voice assistants, intelligent recommendations, intelligent customer service, autonomous driving, etc. Among them, intelligent customer service is a customer service solution based on artificial intelligence technology. Through functions such as natural language processing, knowledge base management, business process management, and dialogue management, it realizes the recognition of user voice, text and other information, understands the user's intention, and gives corresponding intelligent replies to provide users with automated and intelligent services.

[0003] Due to technical limitations and data dependence, when intelligent customer service solutions handle industry-specific problems in different industries, misjudgments or inaccurate answers may occur. For example, in the hotel service industry, in the face of customer inquiries, delivery, meal delivery, customer experience and other needs, intelligent customer service solutions often cannot provide targeted replies and effective services, resulting in poor user experience.

[0004] Therefore, there is a need to provide an intelligent customer service solution that can accurately understand the user's needs in a specific industry and can provide reply content that combines the user's specific needs. Summary of the Invention

[0005] An object of the present application is to provide an intelligent reply method for user services to solve the problem that it is difficult for intelligent customer service in the prior art to accurately understand the user's intention and unable to provide a reply that combines the specific needs of the user.

[0006] To achieve the above object, some embodiments of the present application provide an intelligent reply method for user services, which is used for an electronic device. The method includes:

[0007] Obtain the current input information of the user;

[0008] Input the current input information and the user's historical context information into a deep learning model to obtain the recognized user intention;

[0009] Input the user intention and a preset question decomposition instruction into a deep learning model to obtain the decomposed sub-intentions;

[0010] Obtain the corresponding skill indication information according to the sub-intentions, and input the sub-intentions, skill indication information and user association information into a deep learning model to obtain the reply content;

[0011] Input the reply content and the content polishing instruction into the deep learning model, obtain the combined and polished reply content, and feedback it to the user.

[0012] Further, after obtaining the user's current input information, it includes:

[0013] Normalize and clean the current input information;

[0014] Intercept keywords from the cleaned input information, and use the processed input information as the new current input information.

[0015] Further, the user's current input information includes one or more of the following combinations: text, voice, picture, video.

[0016] Further, input the current input information and the user's historical context information into the deep learning model to obtain the recognized user intent, including:

[0017] Input the preset intent recognition indication information, the current input information, and the user's historical context information into the deep learning model, so that the deep learning model performs user intent recognition according to the intent recognition indication information. Among them, the intent recognition indication information is used to control the deep learning model to recognize the user intent according to the current input information when the current input information is irrelevant to the historical context information, and to recognize the user intent by combining the current input information and the historical context information when they are relevant.

[0018] Further, the sub-intent includes an intent category and intent content. The intent categories include accommodation and catering - consultation, accommodation and catering - delivering items, accommodation and catering - ordering meals, accommodation and catering - customer experience, fitness - consultation, fitness - reservation, fitness - complaints and feedback, fitness - others, film and drama - consultation, film and drama - purchase, film and drama - feedback, film and drama - complaints, scenic spots - information consultation, scenic spots - ticket related, scenic spots - play experience feedback, scenic spots - special needs.

[0019] Further, obtain the corresponding skill indication information according to the sub-intent, including:

[0020] If there is used skill indication information, obtain the used skill indication information;

[0021] If there is no used skill indication information, determine the corresponding skill indication information through a recommendation algorithm according to the intent category and intent content of the sub-intent;

[0022] If the skill indication information cannot be determined, use the fallback skill indication information.

[0023] Further, the user association information includes user check-in information and item inventory information. The user check-in information includes the user's surname, gender, check-in date, and guest room number information. The item inventory information includes the item inventory quantity, the quantity of free items provided in the guest room, and the unit price for charging.

[0024] Further, inputting the sub-intention, skill indication information, and user association information into a deep learning model to obtain a reply content further includes:

[0025] Providing intervention strategy information to the deep learning model so that the reply content generated by the deep learning model is restricted within the range of the skill indication information. The intervention strategy information includes prompt strategy information, error correction strategy information, recommendation optimization strategy information, and auto-completion strategy information.

[0026] Some embodiments of the present application also provide a computer-readable medium, on which computer-readable instructions are stored. The computer-readable instructions can be executed by a processor to implement the intelligent reply method for the foregoing user service.

[0027] Some embodiments of the present application also provide an electronic device, which includes a memory for storing computer program instructions and a processor for executing the computer program instructions. Wherein, when the computer program instructions are executed by the processor, the electronic device executes the intelligent reply method for the foregoing user service.

[0028] Compared with the prior art, the solution provided by the present application can obtain the user's current input information, input the current input information and the user's historical context information into a deep learning model to obtain the recognized user intention, then input the user intention and the preset question decomposition instruction into the deep learning model to obtain the decomposed sub-intention, obtain the corresponding skill indication information according to the sub-intention, input the sub-intention, skill indication information, and user association information into the deep learning model to obtain the reply content, input the reply content and the content polishing instruction into the deep learning model to obtain the merged and polished reply content and feedback it to the user, so as to accurately understand the user intention, and provide accurate business-related reply content and services according to the user association information combined with the user's needs, solve the problem that the intelligent customer service in the prior art cannot accurately reply to the user intention in combination with business information, and can also make the human-computer dialogue more natural, fluent, concise and clear, effectively improving the user dialogue efficiency and service experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] By reading the detailed description of the non-limiting embodiments with reference to the following drawings, other features, purposes, and advantages of the present application will become more obvious:

[0030] Figure 1 It is a flowchart of an intelligent reply method for a user service provided by some embodiments of the present application. Detailed implementation manners

[0031] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented on the premise of the technical solution of the present invention, and gives detailed implementation manners and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.

[0032] Here, the intelligent reply method for user services in the embodiments of the present application is suitable for scenarios where timely and accurate responses are made to the demands put forward by users in industries involving customer service, such as the hotel industry.

[0033] In this scenario, the user can put forward questions or requests for items in various forms, such as text, voice, image, etc., to the intelligent customer service system. After receiving the input information of the user, the intelligent customer service system needs to accurately understand the user's needs and intentions, and make a targeted and specific reply to the user's intentions according to the service information provided by the current service place, so as to provide the user with automated and intelligent customer service, meeting the various needs of the user 24 hours a day.

[0034] The intelligent reply method for user services provided by the present application can obtain the current input information of the user, input the current input information and the user's historical context information into a deep learning model to obtain the recognized user intention, then input the user intention and a preset question decomposition instruction into the deep learning model to obtain the decomposed sub-intentions, obtain the corresponding skill indication information according to the sub-intentions, input the sub-intentions, skill indication information and user association information into the deep learning model to obtain the reply content, input the reply content and a content polishing instruction into the deep learning model to obtain the merged and polished reply content and feedback it to the user, so as to accurately understand the user's intention, and provide accurate business-related reply content and services according to the user association information combined with the user's needs, solving the problem that the intelligent customer service in the prior art cannot accurately reply to the user's intention in combination with business information, and also making the human-machine dialogue more natural, fluent, concise and clear, effectively improving the user's dialogue efficiency and service experience.

[0035] Figure 1 Some embodiments of the present application show the flow of executing the intelligent reply method for user services by an electronic device. The electronic device is the execution subject of the method, such as Figure 1 As shown, the method may include the following steps:

[0036] Step S101, obtain the current input information of the user.

[0037] It will be appreciated that electronic devices may include, but are not limited to, laptop computers, desktop computers, tablet computers, mobile phones, wearable devices, head-mounted displays, servers, mobile email devices, portable game consoles, portable music players, reader devices, televisions having one or more processors embedded or coupled thereto, or other electronic devices capable of accessing a network, etc.

[0038] In some embodiments of the present application, the user's current input information may be in various forms, including but not limited to text, voice, picture, video, or a combination of various forms, etc. The text may be information expressed in different languages, and the languages may include but are not limited to Chinese, English, and other languages, as long as they are languages that can be supported by the deep learning model used in the embodiments of the present application.

[0039] In some embodiments of the present application, after obtaining the user's current input information, the electronic device may also perform standardized cleaning on the current input information, perform keyword interception on the cleaned input information, and use the processed input information as new current input information.

[0040] Here, the user's current input information may contain non-standard data, special characters, illegal harmful content, etc. Before the current input information is subsequently processed, it is first normalized and cleaned. Through normalization and cleaning, the current input information is normalized and special characters and other data are removed. After the current input information is normalized and cleaned, the keyword library is searched through the keywords in the input information to shield the illegal harmful content in the input information and exclude inappropriate interference information. Finally, the obtained information is used as the new current input information and provided to the deep learning model for subsequent processing.

[0041] In addition, the electronic device can set a keyword library for interception and use a keyword matching algorithm to intercept keywords in the keyword library to filter out false, misleading, rights-infringing, illegal and irregular content. It can be understood that the keyword matching algorithm can be any algorithm that can be used for keyword matching, such as a Deterministic Finite Automaton (DFA) algorithm.

[0042] Step S102: input the current input information and the user's historical context information into a deep learning model to obtain the recognized user intent.

[0043] Here, the deep learning model can be any machine learning model capable of natural language understanding and reasoning capabilities, such as large language models (LLMs), generative pre-trained transformer models (GPTs), etc., and can also be specific products of the corresponding machine learning models, such as Tongyi Qianwen, DeepSeek, etc.

[0044] The current input information and the user's historical context information are input into the deep learning model. The deep learning model calls its language understanding ability to analyze the input content, identify the user's intention, and output it. It can be understood that the current input information is the latest piece of information input by the user, and the historical context information is the historical information input by the user in the same session.

[0045] In some embodiments of the present application, the electronic device inputs preset intention recognition indication information, the current input information, and the user's historical context information into the deep learning model, so that the deep learning model performs user intention recognition according to the intention recognition indication information. Here, the intention recognition indication information is used to control the deep learning model to identify the user's intention based on the current input information when the current input information is irrelevant to the historical context information, and to identify the user's intention by combining the current input information and the historical context information when they are relevant.

[0046] The following takes the intelligent customer service in the hotel industry as an example for specific illustration.

[0047] For example, the user's current input information is: Can I check out late tomorrow? I'm not feeling well. The result obtained by the deep learning model through intention recognition of the current input information is: The guest is not feeling well and wants to delay the check-out.

[0048] Another example is the conversation process between the user and the intelligent customer service as follows:

[0049] User: Deliver takeout to my room.

[0050] Intelligent customer service: What is the last 4 digits of the mobile phone number for your takeout?

[0051] User: 3345.

[0052] The user's current input information is "3345". The result of the deep learning model's user intention recognition by combining the current input information and the historical context information "the last 4 digits of the mobile phone number for takeout" is "The last 4 digits of my mobile phone number are 3345".

[0053] In addition, the intent recognition indication information can also control the deep learning model to use different processing logic for specific words during the intent recognition process. For example, words such as "OK" and "OK" that simply indicate confirmation are judged based on context information; if there is no content that the intelligent customer service needs to reply to, they are ignored to streamline the conversation; if there are only numbers or letters, the actual meaning of the numbers or letters is judged based on the context information; if the content is not within the scope of the hotel customer service, it is directly ignored, etc.

[0054] For example, the conversation process between a user and an intelligent customer service representative is as follows:

[0055] 22:00 User: The air conditioner is too cold, I need someone to take a look.

[0056] 22:01 Intelligent customer service: I will immediately notify the engineering staff to come to your room.

[0057] 08:00 User: I didn't sleep well all night. The temperature can't be adjusted. Give me an explanation.

[0058] The deep learning model identifies the user intent based on the current input information and historical context information, and the result is: the guest complained that the air conditioning problem affected his sleep and asked for compensation. In addition, the deep learning model can also extract the key information in the user intent, combine it with the intervention strategy information, and filter out irrelevant data. For example, the key information in the user intent is: air conditioning problem, asking for compensation, and the irrelevant information is: find someone to take a look, I didn't sleep well last night.

[0059] Step S103: input the user intention and the preset question decomposition instructions into the deep learning model to obtain the decomposed sub-intentions.

[0060] Here, the question decomposition instructions are a set of pre-established prompt instructions for specifying relevant rules used by the deep learning model to decompose the identified user intent.

[0061] The rules included in the problem decomposition instructions can be described as follows:

[0062] - Analyze whether there is any correlation between different intentions. If there is any correlation, treat them as the same intention;

[0063] - Split unrelated intents, and the split content must contain complete information;

[0064] -Correct text errors in user input;

[0065] -Directly filter input content that cannot be recognized;

[0066] - If the user expresses multiple and conflicting intents, identify the user's final intent;

[0067] - If it is a delivery requirement, different item contents must be split into different intents.

[0068] After inputting the problem decomposition instruction into the deep learning model, the deep learning model decomposes the problems involved in the user's intent by invoking the language understanding ability, decomposes the problem into multiple independent sub-problems corresponding to different intent types, and then obtains the sub-intents corresponding to the sub-problems.

[0069] In some embodiments of the present application, the sub-intent includes an intent category and intent content. The intent category includes different intent types in multiple industries, and the intent content includes specific service requirements. Industries may include, but are not limited to: accommodation and catering, fitness, film and drama, scenic spots, etc.

[0070] The intent category may include, but is not limited to, accommodation and catering - consultation, accommodation and catering - delivery, accommodation and catering - reservation, accommodation and catering - customer experience, fitness - consultation, fitness - reservation, fitness - complaints and feedback, fitness - others, film and drama - consultation, film and drama - purchase, film and drama - feedback, film and drama - complaints, scenic spots - information consultation, scenic spots - ticketing related, scenic spots - play experience feedback, scenic spots - special requirements, etc.

[0071] The accommodation and catering - consultation intent category is used to describe information consultation services in the accommodation and catering industry, such as hotel basic information, facilities and service consultation, price and promotion information, special requirements, etc.; the accommodation and catering - delivery intent category is used to describe item supply services in the accommodation and catering industry, such as water delivery, towel delivery, slipper delivery, etc.; the accommodation and catering - reservation intent category is used to describe catering reservation services in the accommodation and catering industry, such as table reservation, menu introduction, etc.; the accommodation and catering - customer experience intent category is used to describe services related to customer experience in the accommodation and catering industry, such as service complaints, guest room facility maintenance, etc.

[0072] The fitness - consultation intent category is used to describe information consultation services in the fitness industry, such as course consultation, price consultation, venue facility consultation, etc.; the fitness - reservation intent category is used to describe reservation services in the fitness industry, such as course reservation, personal trainer reservation, etc.; the fitness - complaints and feedback intent category is used to describe complaints and feedback services in the fitness industry, such as service complaints, facility complaints, course feedback, etc.; the fitness - others intent category is used to describe other services in the fitness industry, such as membership benefit consultation, renewal and upgrade consultation, etc.

[0073] The movie and drama - consultation intention category is used to describe information consultation services in the movie and drama industry, such as movie / TV drama information, playing platforms and channels, movie news and events, etc.; the movie and drama - purchase intention category is used to describe purchase - related services in the movie and drama industry, such as membership purchase consultation, paid movie / TV drama purchase consultation, additional product purchase consultation, etc.; the movie and drama - feedback intention category is used to describe information feedback services in the movie and drama industry, such as content feedback, playing problem feedback, platform experience feedback, etc.; the movie and drama - complaint intention category is used to describe relevant complaint services in the movie and drama industry, such as infringement complaint, service quality complaint, etc.

[0074] The scenic spot - information consultation intention category is used to describe information consultation services in the scenic spot industry, such as basic scenic spot information, details of scenic spots, supporting facilities, etc.; the scenic spot - ticket - related intention category is used to describe ticket - related services in the scenic spot industry, such as ticket price, ticket - purchasing method, ticket policy, etc.; the scenic spot - play experience feedback intention category is used to describe relevant experience feedback services in the scenic spot industry, such as service quality feedback, facility experience feedback, environment feedback, etc.; the scenic spot - special needs intention category is used to describe special needs services in the scenic spot industry, such as group reservation, customized service, barrier - free facility requirements, etc.

[0075] The sub - intentions can be classified through a classification algorithm into the above - mentioned multiple intention categories. The classification algorithm can be, for example, the Support Vector Machine (SVM) algorithm, the RandomForest (RF) algorithm, etc.

[0076] Here, since the existing intelligent customer service solutions use small models and keyword recognition technologies, they cannot recognize the user input content that exceeds the keywords, let alone accurately recognize the user's intention. In addition, the existing intelligent customer service solutions can only handle one user question in the latest piece of information input by the user. When there are multiple user questions in the latest piece of information, it is difficult to handle them correctly. Even if the historical context information of the user is saved, it will lead to a mismatch of the historical context information when encountering multiple user questions because it does not correspond to the actual intention of the user.

[0077] Some embodiments of the present application overcome the shortcoming that only one user question can be solved in one piece of user input information in the existing intelligent customer service solutions by decomposing the user question into sub - questions and recording the historical context information of different sub - questions in the corresponding sub - questions and sub - intentions, thus avoiding the mismatch of different user question data.

[0078] For example, the user's current input information is "May I ask where to have breakfast, please send two bottles of mineral water, the air conditioner is too cold and I haven't slept well all night, please find someone to take a look". The deep learning model decomposes it into three sub-questions: 1) Where to have breakfast; 2) Send two bottles of mineral water to the room; 3) The air conditioner is too cold and I haven't slept well all night, please find someone to check. After decomposing into sub-questions, the intention recognition is performed on each sub-question respectively to obtain sub-intentions.

[0079] Step S104, obtain the corresponding skill instruction information according to the sub-intention, and input the sub-intention, skill instruction information and user association information into the deep learning model to obtain the reply content.

[0080] Here, the skill instruction information is a set of text instruction information, also known as a knowledge base tool, which is an information model established in advance according to actual business experience and is used to control the deep learning model to output corresponding results according to the indicated requirements.

[0081] The knowledge base tools can be divided into four types according to the intention types of user intentions: consultation, delivery, food delivery, and customer experience. For each type, corresponding skill text instructions are set for specific user intentions to handle different user problems.

[0082] In some embodiments of the present application, obtaining the corresponding skill instruction information according to the sub-intention may include the following situations:

[0083] 1) If there is used skill instruction information, obtain the used skill instruction information;

[0084] 2) If there is no used skill instruction information, determine the corresponding skill instruction information according to the intention category and intention content of the sub-intention through a recommendation algorithm. The recommendation algorithm can be, for example, a collaborative filtering algorithm (Collaborative Filtering, CF);

[0085] 3) If the skill instruction information cannot be determined, use the fallback skill instruction information, and the fallback skill instruction information is used to indicate to transfer to manual processing.

[0086] For example, the sub-intention is "Where to have breakfast". The electronic device obtains the preset skill instruction information "Consultation - Catering Consultation Tool" through the recommendation algorithm according to the sub-intention, and inputs the relevant information of the "Consultation - Catering Consultation Tool" into the deep learning model to control the output result of the deep learning model.

[0087] An example of the Consultation - Catering Consultation Tool is as follows:

[0088] #The following content is for the hotel dining consultation tool. You can only reply to guests' questions according to the existing content below (except for polite remarks). If the questions or requests from guests are not included in the following content, you can only answer: "Sorry, please wait a moment. I will inform the dining staff to contact you." Remember not to casually promise to meet guests' requirements or answer unknown questions.

[0089] #1. Prerequisites

[0090] - Please note that brunch includes lunch.

[0091] #2. According to the following judgment conditions, for conversations with guests.

[0092] ##2.1 Hotel restaurant information

[0093] - The hotel has two restaurants, an all-day dining restaurant and a Chinese restaurant.

[0094] ##2.2 All-day dining restaurant

[0095] - Location of the all-day dining restaurant (where breakfast is served): On the basement floor of the hotel.

[0096] - Business hours of the all-day dining restaurant are from 6:00 to 22:00

[0097] - Provide children's seats.

[0098] - The restaurant is non-smoking.

[0099] - The restaurant has private rooms. One can accommodate 10 people and the other can accommodate 6 people. Advance reservation is required.

[0100] - When guests ask if they can reserve a private room, you cannot directly agree.

[0101] - Except for breakfast, after guests ask about the restaurant information, you need to ask if you can help the guests make a reservation. Do not inform the guests of the reservation process first.

[0102] - If guests want to make a reservation, then jump to the restaurant reservation process.

[0103] - If guests ask what there is to eat for the buffet breakfast, buffet brunch, buffet dinner, or if there are items such as seafood, you need to tell the guests: "The buffet menu will be adjusted. If you need to know more information, can I let the restaurant staff contact you?"

[0104] 2.21 All-day dining restaurant - Buffet breakfast

[0105] - The time is from 6:30 to 10:30

[0106] - Is reservation required: No reservation is required.

[0107] - Breakfast Charging Standard

[0108] - Self-service breakfast charging standard for adults: 181 yuan per adult.

[0109] - Self-service breakfast charging standard for children: When two breakfasts are included in the room or two adults have paid, two children under 12 years old (including 12 years old) can have breakfast. If only one breakfast is included in the room, only one child under 12 years old (including 12 years old) can have breakfast. Children over 12 years old are charged 99 yuan per person. If the child is already 18 years old or older, they will be charged according to the adult price, 181 yuan per person.

[0110] #2.211 Whether to Check the Room Number for Breakfast

[0111] - Whether the room includes breakfast.

[0112] - If the guest's room includes self-service breakfast, no reservation is required. Just go to the restaurant and report the room number to have breakfast.

[0113] - If the guest's room does not include self-service breakfast, tell the guest to go directly to the restaurant to pay for breakfast. At the same time, inform the guest of the self-service breakfast charging standard for adults and children.

[0114] #2.212 Guests Need to Add Extra Breakfast

[0115] - You can just go directly to the restaurant to pay.

[0116] - If you have already had breakfast at the restaurant, can you pack the breakfast? No. (Do not need to tell the guest actively)

[0117] After inputting the sub-intent, skill indication information, and user association information into the deep learning model, the example of the reply result output by the deep learning model is as follows:

[0118] Breakfast is served at the all-day dining restaurant on the first basement floor of the hotel. The self-service breakfast time is from 6:30 to 10:30, and no reservation is required. If the room includes breakfast, just report the room number to have breakfast. If not, you can go to the restaurant to pay. Adults are charged 181 yuan, and the children's charging standard depends on the actual situation.

[0119] By inputting the sub-intent, skill indication information, and user association information into the deep learning model, the reply content is obtained, thus solving the problem that the existing intelligent customer service solution cannot combine the user's business information and cannot accurately reply to the user's intention.

[0120] Here, the user-related information includes the user's check-in information and item inventory information. The user's check-in information may include, but is not limited to, the user's surname, gender, check-in date, and guest room number information, etc. The item inventory information may include, but is not limited to, the item inventory quantity, the quantity freely distributed in the guest room, the charging unit price, etc. The user's check-in information and item inventory information can be used when the deep learning model generates response content using the knowledge base tool.

[0121] It can be understood that in different industries such as accommodation and catering, fitness, scenic spots, and movies and TV dramas, the user-related information can be presented in different forms. The corresponding user-related information can be determined according to the specific related information of the user in different industries. For example, the user-related information in the fitness industry can be the user's arrival information at the store, the user's corresponding personal trainer information, etc.

[0122] When the electronic device obtains the user's check-in information, it can call the hotel order data interface API according to the hotel order number bound by the user and read the user's check-in information.

[0123] When there is a user's demand for delivering items in the sub-intent, the electronic device needs to judge information such as the inventory of the item, the free quantity, and the charging standard, obtain the item inventory information, and then compare it with the item demand quantity. If the inventory exceeds the demand quantity and is within the free distribution limit, a delivery instruction is directly generated; if it exceeds the free distribution limit, the excess quantity and price are calculated, and a charging standard statement is generated; if the inventory is less than the demand quantity, a statement of insufficient inventory is generated and the purchaser is reminded to replenish the stock.

[0124] In some embodiments of the present application, when inputting the sub-intent, skill indication information, and user-related information into the deep learning model to obtain the response content, intervention strategy information can also be provided to the deep learning model to limit the response content generated by the deep learning model within the range of the skill indication information. Here, the intervention strategy information may include, but is not limited to, prompt strategy information, error correction strategy information, recommendation optimization strategy information, and automatic completion strategy information, etc.

[0125] The prompt strategy information is used to generate a response content that prompts the user to supplement information when the information input by the user is incomplete. For example, when the user books a restaurant, the user is prompted to tell the dining time and the number of people. The error correction strategy information is used to automatically detect and correct when there are semantic, spelling, or logical errors in the information input by the user. The recommendation optimization strategy information is used to dynamically adjust the response content according to the user's historical context information. The automatic completion strategy information is used to intelligently complete the incomplete input information of the user and generate the response content.

[0126] For example, an example of the intervention strategy information is as follows:

[0127] - You are a hotel employee, and the content you answer can only be based on the information of the tool.

[0128] - You can only select the corresponding tools from the tools and do not generate any text.

[0129] - If the guest needs to deliver guest room items and the required items are not in the delivery tool, then call the [Delivery - Others] tool. Otherwise, call the [Consultation] fallback tool.

[0130] By using this example intervention strategy information, it can be ensured that the response content generated by the deep learning model is within the scope of the content of the knowledge base tools used.

[0131] By using the intervention strategy information to assist the deep learning model in analyzing the user's intention, it can guide the user to improve the requirement information, generate accurate response content, and improve the user's conversation efficiency and satisfaction.

[0132] In addition, the knowledge base tools can be continuously increased according to the user requirements of the industry and can reach thousands or even tens of thousands of items. Some examples of knowledge base tools such as the Delivery - Water Delivery tool and the Customer Experience - Greeting tool are as follows:

[0133] Delivery - Water Delivery tool

[0134] # The following content is for the hotel water delivery tool and can only be replied according to the following rules. If the requirements raised are not in the following content, only reply with words like "Sorry, I'll ask my colleague to reply to you right away." Remember not to casually promise to meet the guest's requirements or answer unknown questions.

[0135] # 1. Prerequisites

[0136] - It is necessary to know whether the status in the current guest information is not arrived at the store, already checked in or already checked out

[0137] - No matter whether the guest says mineral water or something else, only "bottled water" can be used in the reply.

[0138] # 2. Related content

[0139] - Location of bottled water: On the writing desk

[0140] # 3. Delivery rules

[0141] - Reply according to the reference script.

[0142] ## 3.1 The guest needs water

[0143] - There are 2 bottles of water in the room. If the guest needs additional water, it needs to be charged according to the bottled water price. Ask the guest if they want to buy it.

[0144] ## 3.1.1 The guest asks how many bottles of water are in the room

[0145] - There are 2 bottles of water in the room.

[0146] ##3.2 The guest did not receive the bottled water in the guest room

[0147] - We are very sorry. We will urge the staff to send it as soon as possible

[0148] ##3.3 The guest complains or complains that the bottled water is charged

[0149] - If the guest complains due to the charge of the hotel bottled water, we will express our understanding and ask the guest to wait a moment. We will immediately report to the supervisor.

[0150] #4. Post - condition, the generated content must meet the following conditions.

[0151] - The room number can only appear in the reply when the guest status is checked - out.

[0152] - The guest cannot be addressed in the reply content, such as Mr. / Ms. X.

[0153] Customer experience - greeting tool

[0154] - You are a new employee in a hotel, humorous and law - abiding. You must not answer questions that violate laws, ethics, and are anti - social.

[0155] - The content you answer can only be based on the rules of the tool. You cannot fabricate non - existent information.

[0156] - You must maintain a polite, professional, and friendly tone. Do not repeat and list in the reply content.

[0157] - Try to streamline the reply to about 50 words.

[0158] In some embodiments of this application, after the electronic device obtains the reply content output by the deep - learning model, it can also intercept keywords in the reply content through a keyword matching algorithm to ensure that the output reply content does not include illegal and harmful content. Similarly, the keyword matching algorithm can be, for example, the DFA algorithm, etc.

[0159] Step S105, input the reply content and the content polishing instruction into the deep - learning model, obtain the merged and polished reply content and feedback it to the user.

[0160] Similarly, the content polishing instruction is a set of pre - established prompt instructions used to specify the relevant rules for the deep - learning model to analyze, merge, and polish the reply content corresponding to multiple sub - intents.

[0161] The rules in the content polishing instruction can be described as follows:

[0162] # Basic rules

[0163] - The polished reply cannot change the meaning of the original reply and cannot infer or derive content that does not exist.

[0164] - Only polish the "reply content" that is prone to ambiguity, otherwise directly quote it.

[0165] # Polishing Rules

[0166] - Determine the type of conversation based on the "reply content".

[0167] ## Unable to meet user needs

[0168] - If the user's needs cannot be met, directly quote the original reply for explanation. For example, "We will contact you after the employee confirms."

[0169] ## User orders food

[0170] - If the reply involves an amount, be sure to calculate the total payment price based on the quantity and price in the reply.

[0171] ## User consults or gives feedback on problems

[0172] - Directly send the "content to reply to the user" to the user without changing the content.

[0173] ## User needs items to be delivered

[0174] - If the user's delivery needs can be met, judge the time interval between the user's requests. If multiple delivery requests are made within 30 seconds, the reply content needs to be combined into one reply.

[0175] ## User makes a complaint

[0176] - It is necessary to combine the employee's replies related to the complaint in the "historical conversation content" and merge them into one sentence while keeping the meaning of the "current conversation content" complete.

[0177] - If the "reply content" is to confirm or give the user's needs to the staff or notify the employee to handle, directly use the "reply content" without changing or adding any content.

[0178] ## Ask for the user's room number and name

[0179] - Regardless of whether there is already user information, if the "reply content" needs to ask for the user's room number, name, and contact information, it still needs to be asked after polishing and cannot directly use the existing [user information] to reply to the question.

[0180] # Style

[0181] -Text requirements: Use a natural conversation tone, be short, friendly, enthusiastic, and easy-going. Unlike a machine, use clear, accurate, and polite language to ensure the message is conveyed correctly. Try to keep the response within 60 words.

[0182] -Make semantic judgments based on the "historical conversation content"

[0183] -In informal situations or when intimacy needs to be enhanced, a relaxed and friendly tone will be used, which may include daily slang or catchphrases to make the conversation more natural and fluent.

[0184] After receiving multiple response contents and content polishing instructions, the deep learning model merges and polishes the multiple response contents according to the content polishing instructions, and outputs the merged and polished response content, making the conversation between the user and the intelligent customer service more natural, fluent, concise, and clear, and improving user satisfaction.

[0185] The electronic device obtains the merged and polished response content and feeds it back to the user for the user to read and solve the user's service problem.

[0186] This application also provides a computer-readable medium, on which computer-readable instructions are stored, and the computer-readable instructions can be executed by a processor to implement the intelligent response method for the foregoing user service.

[0187] This application also provides an electronic device, which includes a memory for storing computer program instructions and a processor for executing the computer program instructions. When the computer program instructions are executed by the processor, the electronic device executes the intelligent response method for the foregoing user service.

[0188] In addition, the intelligent response method for the user service of this application can be implemented as a multi-skill collaboration general framework with a hierarchical architecture. The framework includes a main control module and multiple sub-modules. The main control module is used for monitoring and receiving tasks, responsible for task planning and allocation, coordinating the running order of each sub-module, processing the return results of each sub-module, outputting the final processing results, etc. The sub-modules include multiple intelligent modules with specific skills, which are used for processing tasks with specific skills, and the data can cooperate horizontally with each other.

[0189] The hierarchical architecture design of the framework decomposes the task into subtasks. Each subtask is responsible for processing the requirements in a specific field. The subtasks are relatively independent and do not interfere with each other. The advantages are: each sub-module only needs to remember the context related to it, avoiding context overload; the function of each sub-module is convenient for independent development and debugging; it supports the distributed development process.

[0190] For example, the main control module can perform functions such as user input processing and intent recognition, user question decomposition, and merging reply results. The sub-modules can include a consultation module, a delivery module, and a customer experience module. The user input processing and intent recognition function performs data specification processing on the received user input text information, clears data such as special characters, and calls the language understanding ability of the large language model to analyze the content of the user input, recognize the user intent, and extract key information. The user question decomposition function decomposes the question into independent sub-tasks of different categories such as consultation, delivery, food delivery, and customer experience according to the extracted key information, and assigns the sub-tasks of different categories to specific skill sub-modules for processing. The merging reply results function merges the reply results generated by each sub-module, outputs the same question once, preserves the reply content generated by each sub-module without tampering, checks the generated results and automatically corrects them to make the conversation more natural, fluent, concise and clear. Finally, the merged reply result is output to the user. The sub-module is responsible for processing and generating the reply content. For the assigned sub-tasks, it retrieves the preset relevant knowledge base tools according to the category and requirements, combines the user information for processing, and generates the reply content. The consultation module is used to call the consultation skill knowledge base to generate the reply content. The delivery module is used to call the delivery skill knowledge base to generate the reply content and generate a delivery work order. The customer experience module is used to call the customer experience knowledge base, process customer complaints to generate a complaint handling work order, and generate the reply content.

[0191] In summary, the solution provided by this application can obtain the user's current input information, input the current input information and the user's historical context information into the deep learning model to obtain the recognized user intent, then input the user intent and the preset question decomposition instruction into the deep learning model to obtain the decomposed sub-intents, obtain the corresponding skill indication information according to the sub-intents, input the sub-intents, skill indication information and user association information into the deep learning model to obtain the reply content, input the reply content and the content polishing instruction into the deep learning model to obtain the merged and polished reply content and feedback it to the user. Thus, it can accurately understand the user intent, and provide accurate business-related reply content and services according to the user association information combined with the user needs, solve the problem that the intelligent customer service in the prior art cannot accurately reply to the user intent in combination with the business information, and can also make the human-machine conversation more natural, fluent, concise and clear, effectively improving the user conversation efficiency and service experience.

[0192] It should be noted that the present application can be implemented in software and / or a combination of software and hardware. For example, it can be implemented using an application specific integrated circuit (ASIC), a general purpose computer, or any other similar hardware device. In one embodiment, the software program of the present application can be executed by a processor to implement the steps or functions described above. Similarly, the software program (including related data structures) of the present application can be stored in a computer-readable recording medium, such as a RAM memory, a magnetic or optical drive, or a floppy disk and similar devices. Additionally, some steps or functions of the present application can be implemented using hardware, for example, as a circuit that cooperates with the processor to execute each step or function.

[0193] In a typical configuration of the present application, both the terminal and the network device include one or more processors (CPUs), an input / output interface, a network interface, and a memory.

[0194] The memory may include non-permanent memory in the form of computer-readable media, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of computer-readable media.

[0195] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape disk storage or other magnetic storage devices, or any other non-transitory media that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transitory media such as modulated data signals and carrier waves.

[0196] In addition, a part of this application can be applied as a computer program product, such as computer program instructions. When executed by a computer, through the operations of the computer, it can call or provide the methods and / or technical solutions according to this application. The program instructions for calling the methods of this application may be stored in a fixed or removable recording medium, and / or transmitted through a data stream in a broadcast or other signal-bearing medium, and / or stored in the working memory of a computer device that runs according to the program instructions. Here, an embodiment according to this application includes a device, which includes a memory for storing computer program instructions and a processor for executing the program instructions. When the computer program instructions are executed by the processor, the device is triggered to run based on the methods and / or technical solutions according to the foregoing multiple embodiments of this application.

[0197] For those skilled in the art, it is obvious that this application is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of this application, this application can be implemented in other specific forms. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of this application is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in this application. Any reference signs in the claims should not be construed as limiting the claimed rights. In addition, obviously, the word "including" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units or devices stated in the apparatus claims can also be implemented by one unit or device through software or hardware. First, second, etc. are used to denote names and do not denote any particular order.

Claims

1. An intelligent reply method for user services, used in an electronic device, characterized in that, The method includes: Obtaining the user's current input information; Inputting the current input information and the user's historical context information into a deep learning model to obtain the recognized user intention; Inputting the user intention and a preset question decomposition instruction into a deep learning model to obtain the decomposed sub-intentions; Obtaining the corresponding skill instruction information according to the sub-intentions, and inputting the sub-intentions, the skill instruction information, and the user association information into a deep learning model to obtain the response content; Inputting the response content and a content polishing instruction into a deep learning model to obtain the merged and polished response content and feedback it to the user.

2. The method according to claim 1, wherein After obtaining the user's current input information, it includes: Normalizing and cleaning the current input information; Performing keyword interception on the cleaned input information, and using the processed input information as the new current input information.

3. The method according to claim 1, characterized in that The user's current input information includes one or more of the following combinations: text, voice, picture, video.

4. The method according to claim 1, characterized in that Inputting the current input information and the user's historical context information into a deep learning model to obtain the recognized user intention, including: Inputting the preset intention recognition instruction information, the current input information, and the user's historical context information into a deep learning model, so that the deep learning model performs user intention recognition according to the intention recognition instruction information, where the intention recognition instruction information is used to control the deep learning model to recognize the user intention according to the current input information when the current input information and the historical context information are irrelevant, and to recognize the user intention by combining the current input information and the historical context information when they are relevant.

5. The method according to claim 1, wherein The sub-intentions include intention categories and intention contents. The intention categories include Accommodation and Catering - Consultation, Accommodation and Catering - Delivery, Accommodation and Catering - Ordering Food, Accommodation and Catering - Customer Experience, Fitness - Consultation, Fitness - Appointment, Fitness - Complaint and Feedback, Fitness - Others, Movies and TV - Consultation, Movies and TV - Purchase, Movies and TV - Feedback, Movies and TV - Complaint, Scenic Spots - Information Consultation, Scenic Spots - Ticket Related, Scenic Spots - Play Experience Feedback, Scenic Spots - Special Requirements.

6. The method according to claim 5, wherein Obtaining the corresponding skill instruction information according to the sub-intentions, including: If there is used skill instruction information, obtaining the used skill instruction information; If there is no used skill instruction information, determining the corresponding skill instruction information according to the intention category and intention content of the sub-intentions through a recommendation algorithm; If the skill instruction information cannot be determined, using the fallback skill instruction information.

7. The method according to claim 1, characterized in that, The user association information includes user check-in information and item inventory information. The user check-in information includes the user's surname, gender, check-in date, and guest room number information. The item inventory information includes the item inventory quantity, the free distribution quantity in the guest room, and the charging unit price.

8. The method according to claim 1, wherein Inputting the sub-intentions, the skill instruction information, and the user association information into a deep learning model to obtain the response content, further including: Providing intervention strategy information to the deep learning model so that the response content generated by the deep learning model is restricted within the range of the skill instruction information, where the intervention strategy information includes prompt strategy information, error correction strategy information, recommendation optimization strategy information, and auto-completion strategy information.

9. A computer-readable medium having computer-readable instructions stored thereon, the computer-readable instructions being executable by a processor to implement the method according to any one of claims 1 to 8.

10. An electronic device, the electronic device includes a memory for storing computer program instructions and a processor for executing the computer program instructions, wherein, When the computer program instructions are executed by the processor, cause the electronic device to execute the method according to any one of claims 1 to 8.

Images