Intelligent assistant intention analysis method and device, equipment, storage medium and product

Abstract

The invention discloses an intention analysis method and device of an intelligent assistant, equipment, a storage medium and a product, and relates to the technical field of intelligent question answering, and the method comprises the steps: obtaining a user question input by a user; extracting slot position information associated with each agent from the user question based on a preset slot position extraction strategy; performing slot position standardization on the user question according to the slot position information; based on the standardized slot position of the user question, recalling a candidate agent matched with the user question; and performing intention understanding on the candidate agents through a large language model to determine an optimal agent. According to the technical scheme, the intent understanding of multi-intent sorting of slot extraction, semantic correlation and a large language model is combined, the difficulty of intent understanding is reduced, the intent understanding accuracy and answer effectiveness of the intelligent question and answer assistant are improved, and high robustness and high availability of intent understanding of the intelligent assistant are achieved.

Description

Technical Field

[0001] The present application relates to the field of intelligent question-answering technology, and in particular to an intent parsing method, device, equipment, storage medium and product for an intelligent assistant. Background Art

[0002] With the rapid development of large language models, it is possible to quickly build intelligent agents in many different fields, covering multiple aspects such as domain question-answering RAG assistants and task-based intelligent assistants. An intelligent question-answering assistant that integrates multiple intelligent agents aims to provide users with direct and accurate answers. For such an intelligent question-answering assistant, intent understanding, that is, how to accurately understand the user's intention and select the correct and appropriate intelligent agent to respond to the user's question, is the biggest challenge.

[0003] There are still the following difficulties in understanding the intention of intelligent question-answering assistants:

[0004] Difficulty understanding user questions: When users use intelligent question-answering assistants to ask questions, they often use keywords, professional terms, aliases, etc., which increases the difficulty of understanding the semantics of user questions. At the same time, users' colloquial expressions and dialect voice input also increase the difficulty of understanding their intentions.

[0005] Complexity of domain knowledge: In certain specific domain problems, relevant domain knowledge is required to accurately understand the user's intention. For example, in professional fields such as finance and law, user questions may involve professional terms or specific concepts, and lack of relevant domain knowledge may lead to misunderstanding of intent. In addition, some domain knowledge is related but independent of each other, and the model needs to correctly distinguish the differences in different domain knowledge for selection.

[0006] Agent update and iteration: Currently, large language models and their applications are still in a stage of rapid development, and the functions of each agent are often changed. The capabilities of domain question-answering agents are highly related to the underlying knowledge base, and the real-time update of the knowledge base leads to real-time changes in the scope of agent content. Therefore, it is impossible to make the model clearly understand the scope of capabilities, knowledge scope, and some limitations of each agent through fine-tuning.

[0007] Therefore, it is necessary to propose an intent understanding method to optimize the above difficulties, reduce the difficulty of intent understanding, and improve the accuracy of intent understanding of the intelligent question-answering assistant. Summary of the invention

[0008] The main purpose of this application is to provide an intention parsing method, device, equipment, storage medium and product for an intelligent assistant, aiming to reduce the difficulty of intention understanding and improve the accuracy of intention understanding and the effectiveness of answers of the intelligent question-and-answer assistant.

[0009] To achieve the above objectives, the present application proposes a method for analyzing the intention of an intelligent assistant, the method comprising:

[0010] Get the user input question;

[0011] Extracting slot information associated with each agent from the user question based on a preset slot extraction strategy;

[0012] Performing slot standardization on the user question according to the slot information;

[0013] Based on the standardized slots of the user question, recall candidate agents matching the user question;

[0014] The intention of the candidate agents is understood through a large language model to determine the best agent.

[0015] In one embodiment, the step of extracting slot information associated with each agent from the user question based on a preset slot extraction strategy includes:

[0016] Preprocessing the user question;

[0017] According to the preset type of slot, the preset slot matching rules, deep learning-based NER entity recognition rules or pre-created entity extraction models are used to extract the slot information associated with each agent from the pre-processed user questions.

[0018] In one embodiment, the step of normalizing the slots of the user questions according to the slot information includes:

[0019] According to the slot information, the user is asked a corresponding identifier to distinguish similar slots of different agents.

[0020] In one embodiment, the step of recalling candidate agents matching the user question based on the standardized slots of the user question includes:

[0021] Convert the slot-normalized user questions into vectors;

[0022] The vectorized user questions are compared with the corpus data in the preset corpus database for vector similarity calculation. Based on the vector similarity calculation results, several intelligent agents corresponding to the corpus data with the most relevant similarity are obtained as candidate intelligent agents.

[0023] In one embodiment, the step of understanding the intention of the candidate agents through a large language model to determine the best agent also includes:

[0024] The candidate agents are initially screened, specifically including:

[0025] Based on the relevance score of each candidate agent and a preset threshold, the candidate agents are divided into a precise group, a fuzzy group and an irrelevant group;

[0026] The precise group is retained, the irrelevant group is removed, and the fuzzy group is used as the input of the large language model for intent understanding.

[0027] In one embodiment, the step of understanding the intention of the candidate agents through a large language model to determine the best agent includes:

[0028] The name of the candidate agent and the description of the relevant field are parsed through a large language model, prompt words are constructed, and related content in the learning field is combined with similar historical queries to understand the functions of the candidate agent and obtain several intent categories to determine the best agent. The functions at least include the responsibilities and scope of the candidate agent.

[0029] In one embodiment, the method further comprises:

[0030] The several intent categories are sorted individually to obtain customized final sorting results of the agents and form an agent list.

[0031] In one embodiment, the step of performing personalized sorting on the plurality of intent categories to obtain customized agent final sorting results includes:

[0032] According to the user background information and preset customized factors, the ranking of several intent categories of intent understanding is fine-tuned to obtain a customized final ranking result of the intelligent agent. The customized factors include: the relevance of the original user question to the intelligent agent, user authority and one or more of user preferences.

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

[0034] When an agent is added, updated or deleted, a preset modification method is selected to modify the description and / or offline corpus of the agent.

[0035] In one embodiment, before the step of obtaining the user question input by the user, the step further includes:

[0036] Slot extraction and standardization are performed on each pre-collected user question corpus, wherein the user question corpus is related to the function of the corresponding intelligent agent;

[0037] Clean and vectorize the user question corpus after slot extraction and standardization;

[0038] The cleaned and vectorized user question corpus is labeled with corresponding agent tags to create an offline corpus database.

[0039] The present application also proposes an intention parsing device for an intelligent assistant, the device comprising:

[0040] The acquisition module is used to obtain the user question input by the user;

[0041] An extraction module, used to extract slot information associated with each agent from the user question based on a preset slot extraction strategy;

[0042] A standardization module, used for performing slot standardization on the user question according to the slot information;

[0043] A screening module, for recalling candidate agents matching the user question based on the standardized slots of the user question;

[0044] The intention understanding module is used to understand the intention of the candidate agents through a large language model to determine the best agent.

[0045] The present application also proposes an intent parsing device for an intelligent assistant, the device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the intent parsing method for an intelligent assistant as described above.

[0046] The present application also proposes a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the steps of the intelligent assistant intention parsing method as described above are implemented.

[0047] The present application also proposes a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the intelligent assistant intention parsing method as described above.

[0048] One or more technical solutions proposed in this application have at least the following technical effects: obtaining user questions input by users by providing an intent parsing method, device, equipment, storage medium and product for an intelligent assistant; extracting slot information associated with each agent from the user question based on a preset slot extraction strategy; standardizing the user question according to the slot information; recalling candidate agents that match the user question based on the standardized slots of the user question; and understanding the intent of the candidate agents through a large language model to determine the best agent. The technical solution of this application combines slot extraction, semantic relevance and a multi-intent understanding method of a large language model to reduce the difficulty of intent understanding, improve the accuracy of intent understanding and answer validity of the intelligent question-answering assistant, and achieve high robustness and high availability of intent understanding of the intelligent assistant. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0050] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0051] Figure 1 A flowchart of the first embodiment of the method for analyzing the intention of the intelligent assistant of the present application;

[0052] Figure 2 A flowchart of the second embodiment of the method for parsing intent of an intelligent assistant of the present application;

[0053] Figure 3 A flowchart diagram of the third embodiment of the method for parsing intent of an intelligent assistant of the present application;

[0054] Figure 4 A flowchart of the fourth embodiment of the method for parsing intent of an intelligent assistant of the present application;

[0055] Figure 5 A complete flowchart of the method for analyzing the intention of an intelligent assistant provided in an embodiment of the present application;

[0056] Figure 6 This is a schematic diagram of the module structure of the intention parsing device of the intelligent assistant in the embodiment of the present application;

[0057] Figure 7 This is a schematic diagram of the device structure of the hardware operating environment involved in the intent parsing method of the intelligent assistant in the embodiment of the present application.

[0058] The purpose, features and advantages of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0059] It should be understood that the specific embodiments described herein are only used to explain the technical solutions of the present application and are not used to limit the present application.

[0060] In order to better understand the technical solution of the present application, a detailed description will be given below in conjunction with the accompanying drawings and specific implementation methods.

[0061] The main solution of the embodiment of the present application is: by providing an intention parsing method, device, equipment, storage medium and product for an intelligent assistant, the user question input by the user is obtained; based on a preset slot extraction strategy, the slot information associated with each agent is extracted from the user question; the slot information of the user question is standardized according to the slot information; based on the standardized slots of the user question, the candidate agents that match the user question are recalled; the intention of the candidate agents is understood through a large language model to determine the best agent. The technical solution of the present application combines slot extraction, semantic relevance and a multi-intention understanding method of a large language model, which reduces the difficulty of intention understanding, improves the intention understanding accuracy and answer validity of the intelligent question-answering assistant, and achieves high robustness and high availability of the intelligent assistant in understanding intentions.

[0062] Due to existing technologies, there are still the following difficulties in understanding the intentions of intelligent question-answering assistants:

[0063] Difficulty understanding user questions: When users use intelligent question-answering assistants to ask questions, they often use keywords, professional terms, aliases, etc., which increases the difficulty of understanding the semantics of user questions. At the same time, users' colloquial expressions and dialect voice input also increase the difficulty of understanding their intentions.

[0064] Complexity of domain knowledge: In certain specific domain problems, relevant domain knowledge is required to accurately understand the user's intention. For example, in professional fields such as finance and law, user questions may involve professional terms or specific concepts, and lack of relevant domain knowledge may lead to misunderstanding of intent. In addition, some domain knowledge is related but independent of each other, and the model needs to correctly distinguish the differences in different domain knowledge for selection.

[0065] Agent update and iteration: Currently, large language models and their applications are still in a stage of rapid development, and the functions of each agent are often changed. The capabilities of domain question-answering agents are highly related to the underlying knowledge base, and the real-time update of the knowledge base leads to real-time changes in the scope of agent content. Therefore, it is impossible to make the model clearly understand the scope of capabilities, knowledge scope, and some limitations of each agent through fine-tuning.

[0066] Therefore, this application provides a solution and proposes an intent understanding method that combines slot extraction, semantic relevance, and multi-intent sorting of a large language model, optimizes the above-mentioned difficulties, reduces the difficulty of intent understanding, and improves the intent understanding accuracy and answer effectiveness of the intelligent question and answer assistant.

[0067] Specifically, this application can focus on proposing solutions in the following aspects:

[0068] 1. Fuzzy language understanding: ① Deal with the incompleteness or colloquial characteristics of user expressions. ② Solve the errors that may be caused by automatic speech recognition (ASR).

[0069] 2. Complexity of domain knowledge: ① How to reduce the impact of professional domain knowledge on the complexity of intent analysis. ② Distinguish between domains with similar but different functions.

[0070] 3. Continuous evolution of the intelligent agent: ① When domain knowledge is updated, ensure that intent understanding can quickly adapt to the new domain scope.

[0071] Based on this, the embodiment of the present application provides a method for analyzing the intention of an intelligent assistant, referring to Figure 1 , Figure 1 This is a flow chart of the first embodiment of the intention parsing method of the intelligent assistant of the present application.

[0072] In this embodiment, the intelligent assistant intention parsing method includes steps S10 to S50, each of which is explained in detail below.

[0073] like Figure 1 As shown, the first embodiment of the present application proposes a method for analyzing the intention of an intelligent assistant, the method comprising:

[0074] Step S10, obtaining a user question input by a user;

[0075] The technical solution of this embodiment can reduce the difficulty of intent understanding, improve the accuracy of intent understanding and the effectiveness of answers of the intelligent question-and-answer assistant, and achieve high robustness, high availability and high creativity in the intelligent assistant's understanding of intent.

[0076] The execution subject of this embodiment may be a computing service device with data processing, network communication and program running functions, such as a tablet computer, a personal computer, a mobile phone, etc., or an electronic device capable of realizing the above functions, an intention parsing device of an intelligent assistant, etc. The following takes the intention parsing device of an intelligent assistant as an example to illustrate this embodiment.

[0077] The solution of this embodiment can be applied to a variety of intelligent question-answering scenarios, such as intelligent assistants in the financial field.

[0078] Specifically, first, the user's input question is obtained. The user can enter the question to be asked in the input box of the operation interface of the intelligent assistant. The intelligent assistant receives the user's input question to provide a corresponding answer to the user question. The user question can be text, and can also be information such as pictures and voice.

[0079] Step S20, extracting slot information associated with each agent from the user question based on a preset slot extraction strategy;

[0080] In a dialogue system, an agent refers to a software program or system that can interact with a user. This agent usually has the following characteristics and capabilities:

[0081] Natural Language Understanding (NLU): The agent can understand the user's natural language input, including spoken or written language. This usually involves technologies such as language model, grammatical analysis, and word sense disambiguation.

[0082] Dialogue Management: Dialogue management is the core part of the agent, which is responsible for deciding the next action of the agent. This includes intent recognition (determining the user's purpose) and dialogue policy (deciding how to respond).

[0083] Slot filling: In a conversation, the agent needs to extract key information (slots) from the user's utterance, such as time, place, quantity, etc., in order to better understand and respond to the user's needs.

[0084] Natural Language Generation (NLG): The agent converts the results of internal processing into natural language output to communicate with the user. This process requires generating responses that are grammatically correct, semantically clear, and contextually appropriate.

[0085] Context Management: The agent is able to keep track of the context of the conversation so that it remains coherent and consistent across multiple turns.

[0086] Multi-turn conversation capability: The agent is able to handle conversations involving multiple back-and-forth interactions, rather than just a one-off question and answer session.

[0087] Learning ability: Some advanced agents can learn from conversations through machine learning algorithms to continuously improve their conversational capabilities.

[0088] Integration and connectivity: Agents often need to integrate with other systems or services, such as databases, APIs, or external services, to obtain information or perform tasks.

[0089] Here are some common types of dialogue system agents:

[0090] Task-oriented agents: focus on completing specific tasks, such as booking flights, ordering meals, etc.

[0091] Chatty Agents: Used to engage in informal social conversations, provide entertainment or companionship.

[0092] Question-answering agent: mainly provides fact-based question-answering services.

[0093] Assistive agents: help users complete daily tasks, such as schedule management, reminders, etc.

[0094] Among them, the application of dialogue system agents is very extensive, including customer service robots, virtual assistants, smart home control interfaces, etc. With the development of artificial intelligence technology, these agents are becoming more and more intelligent and can provide a more natural and smooth user experience.

[0095] In this embodiment, the intelligent agent is mainly used as an intelligent assistant in the financial field.

[0096] In natural language processing (NLP) and dialogue systems, slots refer to specific information extracted from the user's speech, which is usually used to fill in the parameters after intent recognition. For example, in a dialogue for booking air tickets, the departure city, destination, date, etc. may all be different slots.

[0097] In a dialogue system, slots are a mechanism for extracting specific information from user input. This information is usually related to completing a specific task or intent. The following are the applications of slots in a dialogue system:

[0098] Information extraction: The main function of slots is to extract key information from the user's natural language expression. For example, in a restaurant reservation conversation, slots can be used to identify information such as the date, time, number of people, and restaurant name mentioned by the user.

[0099] Intent understanding: Slots help the dialogue system better understand the user's intent. By filling in the corresponding slots, the system can confirm the specific action the user wants to perform and obtain all the necessary information to perform the action.

[0100] Dialogue Management: In dialogue management, the state (filled or not) and value of a slot affect the flow of the dialogue. The system can decide the next action based on the state of the slot, such as requesting missing information or confirming the collected information.

[0101] Context maintenance: Slots help the dialogue system maintain the context of the conversation. By tracking the value of the slot, the system can reference and utilize this information at different stages of the conversation, thereby providing a coherent conversation experience.

[0102] Task Completion: Once all necessary slots are filled, the dialogue system can use this information to complete a specific task, such as booking, enquiry, purchase, etc.

[0103] The specific application steps are as follows:

[0104] Recognizing intent: First, the dialogue system needs to recognize the user’s intent (e.g., booking a flight).

[0105] Extracting slots: After determining the intent, the system attempts to extract slots related to the intent (e.g., departure city, destination, date) from the user's input.

[0106] Slot filling: The system uses natural language understanding (NLU) technology to determine the value of each slot (for example, the departure city is "Beijing").

[0107] Slot confirmation: The system may ask the user to confirm whether the drawn slot value is correct to ensure the accuracy of the information.

[0108] Task Execution: Once all necessary slots are filled and confirmed, the system can use this information to execute the corresponding task.

[0109] Among them, effective slot management is a key component in building an efficient and user-friendly dialogue system.

[0110] In this embodiment, after receiving the user question input by the user, the intelligent assistant extracts the slot information associated with each agent from the user question based on a preset slot extraction strategy.

[0111] As an implementation mode, the slot extraction strategy may be pre-configured, for example, may be configured according to a preset type of slot.

[0112] As an implementation method, the slot extraction strategy may include slot matching rules, NER entity recognition rules based on deep learning, or a slot information extraction strategy of a pre-created entity extraction model, through which slot information associated with each agent can be extracted from the user question.

[0113] As an implementation manner, the step of extracting slot information associated with each agent from the user question based on a preset slot extraction strategy may include:

[0114] Preprocessing the user question;

[0115] According to the preset type of slot, the preset slot matching rules, deep learning-based NER entity recognition rules or pre-created entity extraction models are used to extract the slot information associated with each agent from the pre-processed user questions.

[0116] The solution of this embodiment can accurately extract the slot information associated with each agent from the user's question. First, the user question input by the user is preprocessed, wherein the preprocessing includes: eliminating the influence of capitalization difference and punctuation marks on the user question.

[0117] Then, according to the specific type of slot, rule matching, NER entity recognition or advanced entity extraction model is flexibly used to extract slot-related information associated with each agent from the user question.

[0118] Step S30, performing slot standardization on the user question according to the slot information;

[0119] Slot standardization is the process of formatting user input information in the dialogue system to ensure that the information can be consistently understood and processed by the system. The following are the key steps of slot standardization:

[0120] Identify slots: First, the system needs to identify the slot information in the user input through the natural language understanding (NLU) component.

[0121] Extract slot values: Extract specific slot values ​​from the user's words. For example, if the user says "I want to book a flight from Beijing to Shanghai tomorrow", the system needs to extract slot values ​​such as "Departure: Beijing" and "Destination: Shanghai".

[0122] Clean data: remove irrelevant characters in slot values, such as punctuation marks, extra spaces, etc.

[0123] Unified format: Convert slot values ​​to a unified format. For example:

[0124] Date: "Tomorrow" is converted to a date in the format "YYYY-MM-DD".

[0125] Time: "3:00 PM" is converted to "HH:MM" format.

[0126] Place: "Beijing" is unified as "Beijing City" or the corresponding standard place name code.

[0127] Map to standard value: Map the slot value entered by the user to the standard value within the system. For example, the user may enter "Beijing", and the standard value within the system may be "Beijing" or a specific city code.

[0128] Handle synonyms and aliases: Map synonyms or aliases to standard terms. For example, "Capital Airport" and "Beijing Capital International Airport" are both mapped to "Beijing Capital International Airport".

[0129] Error handling: If a slot value is unrecognizable or malformed, the system should be able to ask the user for clarification or to provide correct information.

[0130] Here is a simplified example of slot normalization:

[0131] User input: I want to book a flight from Beijing Capital Airport to Shanghai Pudong International Airport at 3:00 PM tomorrow.

[0132] Slot standardization process:

[0133] Date: "Tomorrow" → "2023-04-15"

[0134] Time: "3:00 pm" → "15:00"

[0135] Departure place: "Beijing Capital Airport" → "Beijing Capital International Airport" (or corresponding code)

[0136] Destination: "Shanghai Pudong International Airport" → "Shanghai Pudong International Airport" (or corresponding code)

[0137] After slot standardization, the system can process user requests more accurately and perform corresponding operations, such as querying flight information, booking air tickets, etc.

[0138] In this embodiment, the user questions are subjected to slot standardization processing, and similar slots of different agents can be distinguished by corresponding identifiers for the user questions.

[0139] Compared with the prior art, this embodiment implements a universal slot extraction strategy to ensure slot independence based on the characteristics of different agents, enhance the generalization ability of intent understanding, and strengthen the recognition of slot differences between similar agents.

[0140] Step S40, based on the standardized slots of the user question, recalling candidate agents matching the user question;

[0141] This stage mainly focuses on recalling candidate agents that match the user's question.

[0142] As an implementation manner, the step of recalling candidate agents matching the user question based on the standardized slots of the user question may include:

[0143] Convert the slot-normalized user questions into vectors;

[0144] The vectorized user questions are compared with the corpus data in the preset corpus database for vector similarity calculation. Based on the vector similarity calculation results, several intelligent agents corresponding to the corpus data with the most relevant similarity are obtained as candidate intelligent agents.

[0145] In this embodiment, an offline corpus database is pre-built, and different user question corpora and corresponding intelligent agents are stored in the corpus database.

[0146] Among them, the user question corpus is related to the function of the corresponding intelligent agent. The user question corpus is stored in a vectorized manner. Slot extraction and standardization can be performed on each pre-collected user question corpus, and an offline corpus database can be created based on the user question corpus and the corresponding intelligent agent.

[0147] As an implementation method, vector similarity calculation is performed between the vectorized user question and the corpus data in the preset corpus database, and based on the vector similarity calculation result, several agents corresponding to the corpus data with the most similarity are obtained as candidate agents.

[0148] For example: convert user questions into vectors, compare the vectorized user questions with the corpus data in the preset corpus database, obtain the most relevant topK entries and corresponding agents by calculating the cosine similarity, and thus obtain candidate agents.

[0149] Step S50, understanding the intention of the candidate agents through a large language model to determine the best agent.

[0150] As an implementation method, the step of understanding the intention of the candidate agents through a large language model to determine the best agent may include:

[0151] The name of the candidate agent and the description of the relevant field are parsed through a large language model, prompt words are constructed, and similar historical queries are combined to learn field-related content to understand the function of the candidate agent and obtain several intent categories to determine the best agent.

[0152] The functions of the candidate agent at least include the responsibilities and scope of the candidate agent.

[0153] Specifically, this embodiment can use a large language model (LLM) to deeply understand the candidate agents to determine the best agent. Since these candidate agents are related to user questions, the large language model needs to parse the user's intentions and agent functions. The key is to ensure that the large language model can accurately understand the responsibilities and scope of the agent. This embodiment uses the agent name, description and similar history query to enable the large language model to deepen its understanding.

[0154] This embodiment adopts the above scheme, specifically by obtaining the user question input by the user; extracting the slot information associated with each agent from the user question based on the preset slot extraction strategy; standardizing the slot of the user question according to the slot information; recalling the candidate agents that match the user question based on the standardized slot of the user question; and understanding the intention of the candidate agents through a large language model to determine the best agent. The technical solution of this application combines slot extraction, semantic relevance and a multi-intention understanding method of a large language model, which reduces the difficulty of intention understanding, improves the intention understanding accuracy and answer validity of the intelligent question-answering assistant, and achieves high robustness and high availability of the intelligent assistant in understanding intentions.

[0155] Based on the first embodiment of the present application, in the second embodiment of the present application, the same or similar contents as those in the first embodiment can refer to the above introduction, and will not be repeated in the following. Figure 2 In step S50, the intention of the candidate agents is understood by a large language model to determine the best agent, and the method also includes step S501:

[0156] Step S501, performing preliminary screening of the candidate agents.

[0157] Compared with the first embodiment described above, this embodiment can also perform a preliminary screening of the candidate agents after recalling the candidate agents that match the user's question. Subsequently, the intention of the initially screened candidate agents is understood through a large language model to determine the best agent.

[0158] This embodiment takes into account that among the recalled candidate agents that match the user's question, some candidate agents are exactly matched with the user's question and no longer need to be understood for intent; or some candidate agents are irrelevant agents and need to be removed, while some candidate agents may be fuzzy matches and need to be understood and identified for intent.

[0159] As an implementation manner, step S501, preliminarily screening the candidate agents may specifically include:

[0160] Based on the relevance score of each candidate agent and a preset threshold, the candidate agents are divided into a precise group, a fuzzy group and an irrelevant group;

[0161] The precise group is retained, the irrelevant group is removed, and the fuzzy group is used as the input of the large language model for intent understanding.

[0162] In the specific implementation, first, the intention parsing device of the intelligent assistant is used to recall candidate agents that match the user's question. The user's question is converted into a vector, and the vectorized user question is compared with the corpus data in the preset corpus database. The most relevant topK entries and corresponding agents are obtained by calculating the cosine similarity. Then, based on the relevance score of each agent and the preset threshold, the candidate agents are divided into precise group, fuzzy group and irrelevant group.

[0163] Among them, the precise group is the final selection after multiple verifications and returns the results directly; the fuzzy group is related but uncertain and will be used for intent analysis of the subsequent large language model; the irrelevant group has no direct connection with the user's question and will not enter the subsequent process.

[0164] Then, a large language model (LLM) can be used to deeply understand the candidate set of fuzzy group agents to determine the best agent. Since these candidate agents are related to user questions, the large language model needs to parse user intentions and agent functions. The key is to ensure that the large language model can accurately understand the responsibilities and scope of the agent. This embodiment uses agent name, description and similar history queries to enable the large language model to have a deeper understanding.

[0165] When performing online queries through the solution of this embodiment, the function call of the large language model can be used, combined with the similarity model results. The large language model can not only describe the relevant fields, but also learn field-related content based on a few shots, thereby reducing the uncertainty of understanding field knowledge.

[0166] This embodiment uses the above scheme, specifically by obtaining the user question input by the user; extracting the slot information associated with each agent from the user question based on the preset slot extraction strategy; standardizing the slot of the user question according to the slot information; recalling the candidate agents matching the user question based on the standardized slots of the user question; preliminarily screening the candidate agents, and using the large language model to understand the intent of the fuzzy groups in the candidate agents to determine the best agent. The technical solution of this application combines slot extraction, semantic relevance and the multi-intention understanding method of the large language model, which reduces the difficulty of intent understanding, improves the accuracy of intent understanding and the effectiveness of answers of the intelligent question-and-answer assistant, and achieves high robustness and high availability of intent understanding of the intelligent assistant. In addition, after recalling the candidate agents matching the user question, this embodiment can also preliminarily screen the candidate agents, and subsequently, use the large language model to understand the intent of the preliminarily screened fuzzy candidate agents to determine the best agent, thereby reducing the processing complexity and computing resources of the large language model, and further improving the accuracy and high reliability of intent understanding.

[0167] Based on the second embodiment of the present application, in the third embodiment of the present application, the same or similar contents as those of the above-mentioned second embodiment can be referred to the above introduction and will not be repeated later.

[0168] Among them, the name of the candidate intelligent agent and the description of the relevant field are parsed through a large language model, prompt words are constructed, and related content in the learning field is combined with similar historical queries to understand the function of the candidate intelligent agent and obtain several intent categories to determine the best intelligent agent.

[0169] On this basis, please refer to Figure 3 In step S50, after the intention of the candidate agents is understood by a large language model to determine the best agent, the method further includes:

[0170] Step S60, perform personalized sorting on several intent categories to obtain customized final sorting results of the intelligent agents and form an intelligent agent list.

[0171] In this embodiment, after parsing the names of candidate agents and related field descriptions through a large language model, constructing prompt words, and combining similar historical queries to learn field-related content, the functions of the candidate agents are understood, and several intent categories are obtained. The several intent categories are personalized and sorted to obtain customized final ranking results of the agents. Then, a list of agents can be formed based on the final ranking results of the agents, so that the best agent can be determined.

[0172] As an implementation method, the step of performing personalized sorting on the plurality of intent categories to obtain customized agent final sorting results may include:

[0173] According to the user background information and preset customization factors, the ranking of several intent categories of intent understanding is fine-tuned to obtain the customized final ranking result of the intelligent agent.

[0174] The customized factors may include: one or more of the relevance of the original user question to the agent, user authority, and user preference.

[0175] Specifically, as an application example, based on the previous stage, the user's question and candidate agents are understood through a large language model to obtain several intent categories, and then the ranking of the intent understanding is fine-tuned based on the user's background information. In addition to the relevance of the original user question and the agent, this embodiment also considers the user's authority area, the desired agent type, and the preference of the position or department, etc., to provide a more customized final ranking result.

[0176] Compared with the prior art, this embodiment introduces multi-intent sorting output to replace the traditional single intent output. The multi-intent sorting basis includes not only the relevance of the intent to the user's question, but also user permissions and department preferences, so as to improve the accuracy of intent recognition and the applicability of the response.

[0177] Furthermore, when an agent is added, updated or deleted, a preset modification method is selected to modify the description and / or offline corpus of the agent, thereby improving the high availability of the intent understanding of the intelligent question and answer assistant.

[0178] Based on the third embodiment of the present application, in the fourth embodiment of the present application, the same or similar contents as those in the third embodiment can refer to the above description, and will not be repeated in the following. Figure 4 The step S10, before obtaining the user question input by the user, also includes:

[0179] Step S101, extracting and standardizing slots for each pre-collected user question corpus, wherein the user question corpus is related to the function of the corresponding intelligent agent;

[0180] Step S102, cleaning and vectorizing the user question corpus after slot extraction and standardization;

[0181] Step S103, labeling the cleaned and vectorized user question corpus with corresponding agent tags to create an offline corpus database.

[0182] Among them, this embodiment constructs an offline corpus database, and stores different user question corpora and corresponding intelligent agents in the corpus database, so that vector similarity calculation can be performed on the user questions and the corpus data in the corpus database based on the corpus database, and based on the vector similarity calculation results, several intelligent agents corresponding to the corpus data with the most relevant similarity are obtained as candidate intelligent agents.

[0183] Among them, the user question corpus is related to the function of the corresponding intelligent agent. The user question corpus is stored in a vectorized manner. Slot extraction and standardization can be performed on each pre-collected user question corpus, and the user question corpus after slot extraction and standardization can be cleaned and vectorized. Then, the cleaned and vectorized user question corpus is marked with the corresponding intelligent agent label, thereby creating an offline corpus database.

[0184] This embodiment establishes and optimizes an offline corpus, extracts slots and standardizes each corpus, and ensures that the function-related corpus of all agents is fully covered.

[0185] This embodiment uses the above scheme, specifically by creating an offline corpus database, and when obtaining the user question input by the user, extracts the slot information associated with each agent from the user question based on the preset slot extraction strategy; the user question is slot-standardized according to the slot information; based on the standardized slots of the user question, the candidate agents matching the user question are recalled; the candidate agents are preliminarily screened, and the fuzzy groups in the candidate agents are understood through the large language model to determine the best agent. The technical solution of this application combines slot extraction, semantic relevance and the multi-intention understanding method of the large language model, which reduces the difficulty of intention understanding, improves the accuracy of intention understanding and the effectiveness of answers of the intelligent question-answering assistant, and realizes the high robustness and high availability of the intelligent assistant for intention understanding. In addition, after recalling the candidate agents matching the user question, this embodiment can also preliminarily screen the candidate agents, and subsequently, the fuzzy candidate agents after the preliminarily screening are understood through the large language model to determine the best agent, thereby reducing the processing complexity and computing resources of the large language model, and further improving the accuracy and high reliability of intention understanding.

[0186] The intention understanding process of the intelligent assistant in this embodiment can be referred to Figure 5 shown.

[0187] It should be noted that the above embodiments can be implemented in combination according to actual conditions, and the embodiments of the present application do not make specific limitations on this.

[0188] Compared with the prior art, the embodiment of the present application evaluates the slots in different situations, compares similar slots, and customizes a unique extraction strategy for each slot based on the actual application, to ensure more accurate slot identification and differentiation, thereby achieving effective optimization of slot extraction. For the agent function, the embodiment of the present application introduces similar historical queries related to the user query, and enhances the large language model's understanding of the association between user intent and agent functions, thereby significantly improving the accuracy of selecting agents. For the selection of agents, unlike the conventional single agent selection, the embodiment of the present application introduces multi-intention sorting output to replace the traditional single intent output. The sorting basis includes not only the relevance of the intent to the user's question, but also covers user permissions and department preferences, etc., so as to improve the accuracy of intent recognition and the applicability of the response. Moreover, the output of this embodiment is a multi-intention result sorted based on the relevance of the user question to the agent, user permissions and preferences, which effectively improves the accuracy of intent understanding and the effectiveness of the answer.

[0189] In addition, the embodiment of the present application uses a large language model Function Call for intent understanding, adds a few shots of similar user questions, and allows the large language model to select multiple intent categories, effectively improving the accuracy of model distribution and the bottom-up processing of online abnormal situations, and improving the high robustness of the intelligent question-answering assistant's intent understanding. Moreover, this application includes slots, ES similar user questions, agent descriptions, agent filtering and other methods to optimize the large language model for intent understanding. When an agent is added, updated or deleted, the appropriate method is selected to fine-tune the agent's description and / or offline corpus, thereby improving the high availability of the intelligent question-answering assistant's intent understanding.

[0190] The embodiment of the present application designs a new framework that integrates slot extraction, vector similarity calculation, intelligent distribution of large language models, and an innovative prompt strategy, which effectively reduces the error in the model's understanding of the agent's capabilities and significantly improves the accuracy of the overall selection.

[0191] In addition, the present application also proposes an intention parsing device for an intelligent assistant, such as Figure 6 As shown, the device comprises:

[0192] An acquisition module 10 is used to acquire a user question input by a user;

[0193] An extraction module 20, configured to extract slot information associated with each agent from the user question based on a preset slot extraction strategy;

[0194] A standardization module 30, configured to perform slot standardization on the user question according to the slot information;

[0195] A screening module 40, for recalling candidate agents matching the user question based on the standardized slots of the user question;

[0196] The intention understanding module 50 is used to understand the intention of the candidate agents through a large language model to determine the best agent.

[0197] The intention parsing device for the intelligent assistant provided in this application adopts the intention parsing method for the intelligent assistant in the above-mentioned embodiment, which can reduce the difficulty of understanding the intention, improve the accuracy of the intention understanding and the effectiveness of the answer of the intelligent question-and-answer assistant, and achieve high robustness and high availability of the intelligent assistant's understanding of the intention. Compared with the prior art, the beneficial effects of the intention parsing device for the intelligent assistant provided in this application are the same as the beneficial effects of the intention parsing method for the intelligent assistant provided in the above-mentioned embodiment, and the other technical features in the intention parsing device for the intelligent assistant are the same as the features disclosed in the above-mentioned embodiment method, which will not be repeated here.

[0198] In addition, to achieve the above-mentioned purpose, the present application also proposes an intention parsing device for an intelligent assistant, the device comprising: a memory, a processor, and a computer program stored on the memory and executable on the processor, the computer program being configured to implement the steps of the intention parsing method for an intelligent assistant as described above.

[0199] Reference below Figure 7 , which shows a schematic diagram of the structure of the intention parsing device of the intelligent assistant suitable for implementing the embodiment of the present application. The intention parsing device of the intelligent assistant in the embodiment of the present application may include but is not limited to mobile terminals such as mobile phones, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle-mounted terminals (such as vehicle-mounted navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 7 The intent parsing device of the smart assistant shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0200] like Figure 7As shown, the intention parsing device of the intelligent assistant may include a processing device 1001 (such as a central processing unit, a graphics processor, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM: Read Only Memory) 1002 or a program loaded from a storage device 1003 to a random access memory (RAM: Random Access Memory) 1004. In RAM1004, various programs and data required for the operation of the intention parsing device of the intelligent assistant are also stored. The processing device 1001, ROM1002 and RAM1004 are connected to each other via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touch screen, a touchpad, a keyboard, a mouse, an image sensor, a microphone, an accelerometer, a gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; storage devices 1003 including, for example, a magnetic tape, a hard disk, etc.; and communication devices 1009. The communication device 1009 can allow the intent parsing device of the intelligent assistant to communicate wirelessly or wired with other devices to exchange data. Although the figure shows the intent parsing device of the intelligent assistant with various systems, it should be understood that it is not required to implement or have all the systems shown. More or fewer systems may be implemented or provided alternatively.

[0201] In particular, according to the embodiments disclosed in the present application, the process described above with reference to the flowchart can be implemented as a computer software program. For example, the embodiments disclosed in the present application include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication device, or installed from a storage device 1003, or installed from a ROM 1002. When the computer program is executed by the processing device 1001, the above-mentioned functions defined in the method of the embodiment disclosed in the present application are executed.

[0202] The intention parsing device of the intelligent assistant provided in this application adopts the intention parsing method of the intelligent assistant in the above-mentioned embodiment, which can reduce the difficulty of understanding the intention, improve the accuracy of the intention understanding and the effectiveness of the answer of the intelligent question-and-answer assistant, and achieve high robustness and high availability of the intelligent assistant's understanding of the intention. Compared with the prior art, the beneficial effects of the intention parsing device of the intelligent assistant provided in this application are the same as the beneficial effects of the intention parsing method of the intelligent assistant provided in the above-mentioned embodiment, and the other technical features in the intention parsing device of the intelligent assistant are the same as the features disclosed in the method of the previous embodiment, which will not be repeated here.

[0203] It should be understood that the various parts disclosed in this application can be implemented by hardware, software, firmware or a combination thereof. In the description of the above embodiments, specific features, structures, materials or characteristics can be combined in any one or more embodiments or examples in a suitable manner.

[0204] In addition, to achieve the above-mentioned purpose, the present application also proposes a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the intelligent assistant intention parsing method as described above are implemented.

[0205] The computer-readable storage medium provided in the present application may be, for example, a USB flash drive, but is not limited to electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, systems or devices, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, system or device. The program code contained on the computer-readable storage medium may be transmitted using any appropriate medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination of the above.

[0206] The above-mentioned computer-readable storage medium may be included in the intent parsing device of the intelligent assistant; or it may exist independently without being assembled into the intent parsing device of the intelligent assistant.

[0207] The above-mentioned computer-readable storage medium carries one or more programs. When the above-mentioned one or more programs are executed by the intention parsing device of the intelligent assistant, the intention parsing device of the intelligent assistant: obtains the user question input by the user; extracts the slot information associated with each intelligent agent from the user question based on a preset slot extraction strategy; standardizes the slots of the user question according to the slot information; recalls the candidate intelligent agents that match the user question based on the standardized slots of the user question; and understands the intention of the candidate intelligent agents through a large language model to determine the best intelligent agent.

[0208] Computer program code for performing the operations of the present application may be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0209] The flow chart and block diagram in the accompanying drawings illustrate the possible architecture, function and operation of the system, method and computer program product according to various embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a module, a program segment or a part of a code, and the module, the program segment or a part of the code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two square boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0210] The modules involved in the embodiments described in this application may be implemented by software or hardware, wherein the name of the module does not constitute a limitation on the unit itself in some cases.

[0211] The readable storage medium provided in this application is a computer-readable storage medium, which stores computer-readable program instructions (i.e., computer programs) for executing the above-mentioned intelligent assistant's intention parsing method, which can reduce the difficulty of intention understanding, improve the accuracy of the intelligent question-and-answer assistant's intention understanding and the effectiveness of the answer, and achieve high robustness and high availability of the intelligent assistant's intention understanding. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as the beneficial effects of the intelligent assistant's intention parsing method provided in the above-mentioned embodiment, which will not be repeated here.

[0212] In addition, to achieve the above-mentioned purpose, the present application also provides a computer program product, which includes a computer program, and when the computer program is executed by a processor, it implements the steps of the intelligent assistant intention parsing method as described above.

[0213] The computer program product provided by this application can reduce the difficulty of understanding intent, improve the accuracy of the intent understanding and the effectiveness of the answers of the intelligent question-answering assistant, and achieve high robustness and high availability of the intelligent assistant in understanding intent. Compared with the prior art, the beneficial effects of the computer program product provided by this application are the same as the beneficial effects of the intent parsing method of the intelligent assistant provided in the above embodiment, which will not be repeated here.

[0214] One or more technical solutions proposed in this application have at least the following technical effects: obtaining user questions input by users by providing an intent parsing method, device, equipment, storage medium and product for an intelligent assistant; extracting slot information associated with each agent from the user question based on a preset slot extraction strategy; standardizing the user question according to the slot information; recalling candidate agents that match the user question based on the standardized slots of the user question; and understanding the intent of the candidate agents through a large language model to determine the best agent. The technical solution of this application combines slot extraction, semantic relevance and a multi-intent understanding method of a large language model to reduce the difficulty of intent understanding, improve the accuracy of intent understanding and answer validity of the intelligent question-answering assistant, and achieve high robustness and high availability of intent understanding of the intelligent assistant.

[0215] The above descriptions are only some embodiments of the present application, and are not intended to limit the patent scope of the present application. All equivalent structural changes made using the contents of the present application specification and drawings under the technical concept of the present application, or direct / indirect applications in other related technical fields are included in the patent protection scope of the present application.

Claims

1. A method for analyzing the intention of an intelligent assistant, characterized in that: The method comprises: Get the user input question; Extracting slot information associated with each agent from the user question based on a preset slot extraction strategy; Performing slot standardization on the user question according to the slot information; Based on the standardized slots of the user question, recall candidate agents matching the user question; The intention of the candidate agents is understood through a large language model to determine the best agent.

2. The method according to claim 1, characterized in that The step of extracting slot information associated with each agent from the user question based on a preset slot extraction strategy includes: Preprocessing the user question; According to the preset type of slot, the preset slot matching rules, deep learning-based NER entity recognition rules or pre-created entity extraction models are used to extract the slot information associated with each agent from the pre-processed user questions.

3. The method according to claim 1, characterized in that The step of normalizing the slots of the user questions according to the slot information comprises: According to the slot information, the user is asked a corresponding identifier to distinguish similar slots of different agents.

4. The method according to claim 1, characterized in that The step of recalling candidate agents matching the user question based on the standardized slots of the user question comprises: Convert the slot-normalized user questions into vectors; The vectorized user questions are compared with the corpus data in the preset corpus database for vector similarity calculation. Based on the vector similarity calculation results, several intelligent agents corresponding to the corpus data with the most relevant similarity are obtained as candidate intelligent agents.

5. The method according to claim 4, characterized in that The step of understanding the intention of the candidate agents through a large language model to determine the best agent also includes: The candidate agents are initially screened, specifically including: Based on the relevance score of each candidate agent and a preset threshold, the candidate agents are divided into a precise group, a fuzzy group and an irrelevant group; The precise group is retained, the irrelevant group is removed, and the fuzzy group is used as the input of the large language model for intent understanding.

6. The method according to any one of claims 1 to 5, characterized in that The step of understanding the intention of the candidate agents through a large language model to determine the best agent includes: The name of the candidate agent and the description of the relevant field are parsed through a large language model, prompt words are constructed, and related content in the learning field is combined with similar historical queries to understand the functions of the candidate agent and obtain several intent categories to determine the best agent. The functions at least include the responsibilities and scope of the candidate agent.

7. The method according to claim 6, characterized in that The method further comprises: The several intent categories are sorted individually to obtain customized final sorting results of the agents and form an agent list.

8. The method according to claim 7, characterized in that The step of performing personalized sorting on the plurality of intent categories to obtain customized agent final sorting results comprises: According to the user background information and preset customized factors, the ranking of several intent categories of intent understanding is fine-tuned to obtain a customized final ranking result of the intelligent agent. The customized factors include: the relevance of the original user question to the intelligent agent, user authority and one or more of user preferences.

9. The method according to claim 7, characterized in that The method further comprises: When an agent is added, updated or deleted, a preset modification method is selected to modify the description of the agent and / or the offline corpus.

10. The method according to claim 4, characterized in that The step of obtaining the user question input by the user also includes: Slot extraction and standardization are performed on each pre-collected user question corpus, wherein the user question corpus is related to the function of the corresponding intelligent agent; Clean and vectorize the user question corpus after slot extraction and standardization; The cleaned and vectorized user question corpus is labeled with corresponding agent tags to create an offline corpus database.

11. An intention analysis device for an intelligent assistant, characterized in that: The device comprises: The acquisition module is used to obtain the user question input by the user; An extraction module, used to extract slot information associated with each agent from the user question based on a preset slot extraction strategy; A standardization module, used for performing slot standardization on the user question according to the slot information; A screening module, for recalling candidate agents matching the user question based on the standardized slots of the user question; The intention understanding module is used to understand the intention of the candidate agents through a large language model to determine the best agent.

12. An intention parsing device for an intelligent assistant, characterized in that: The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the intention parsing method of the intelligent assistant as described in any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the steps of the intention parsing method of an intelligent assistant as described in any one of claims 1 to 10.

14. A computer program product, characterized in that The computer program product includes a computer program, which, when executed by a processor, implements the steps of the intention parsing method of the intelligent assistant as described in any one of claims 1 to 10.

Images