Question answering method and device based on multi-agent collaboration and reinforcement learning optimization

Abstract

The present application relates to a method and device for answering questions based on multi-agent collaboration and reinforcement learning optimization. The method includes: obtaining the user's question information and identifying the user's intention information corresponding to the question information; extracting the user's question demand information and core question content through a demand extraction strategy, and querying the question answer content corresponding to each of the question demand information; when the completeness of the question answer with core question content is lower than a preset completeness threshold, obtaining the supplementary answer information corresponding to each of the question demand information through an information improvement calling program; generating the user's target answer information based on the question answer content corresponding to each of the question demand information and the supplementary answer information corresponding to each of the question demand information. The use of this method can improve the user satisfaction after answering the user's question when answering complex questions.

Description

Technical Field

[0001] The present application relates to the field of natural language processing and deep learning technology, and in particular to a question answering method and device based on multi-agent collaboration and reinforcement learning optimization. Background Art

[0002] With the rapid development of natural language processing and deep learning technologies, intelligent question-answering systems and conversational artificial intelligence are widely used in various business scenarios. However, traditional question-answering systems mostly rely on a single data source or knowledge base, and it is difficult to flexibly call multiple external APIs to meet complex and multi-dimensional problem requirements. At the same time, when the problem involves multi-step logical links (depth) or needs to obtain information from multiple fields at the same time (breadth), traditional systems are prone to lengthy calling processes and excessive resource consumption. In addition, in the presence of authority restrictions, resource quotas, and ambiguous problem scenarios, traditional systems lack effective negotiation and decision-making mechanisms. Therefore, how to improve the quality of answers to complex questions and the efficiency of answers is the current research focus.

[0003] The traditional way of answering user questions is to directly answer the user's questions after identifying the content of the user's question. However, when it comes to complex questions, the answering efficiency is too low, and there are cases where the question cannot be answered completely. Moreover, when answering questions, it is impossible to give priority to the content of the question that meets the user's most core concerns, which leads to poor user satisfaction after answering user questions when answering complex questions. Summary of the invention

[0004] Based on this, it is necessary to provide a question answering method, device, computer equipment, computer-readable storage medium and computer program product based on multi-agent collaboration and reinforcement learning optimization to address the above technical problems.

[0005] In a first aspect, the present application provides a question answering method based on multi-agent collaboration and reinforcement learning optimization, comprising:

[0006] Obtaining user question information and identifying user intention information corresponding to the question information;

[0007] Based on the question information and the user intention information, extract the user's question demand information and the core question content in each question demand information through a demand extraction strategy, and query the question answer content corresponding to each question demand information;

[0008] Based on the answer content of each of the questions, identify the completeness of the answer to each of the core questions, and when the completeness of the answer to the core question content is lower than the preset completeness threshold, obtain the supplementary answer information corresponding to the requirement information of each of the questions through the information completion calling program;

[0009] The user's target answer information is generated based on the question answer content corresponding to each of the question demand information and the supplementary answer information corresponding to each of the question demand information.

[0010] Optionally, the identifying user intention information corresponding to the question information includes:

[0011] Extracting the question text content of the question information, and identifying the text semantic information corresponding to the question text content;

[0012] Based on the text semantic information and the question text content, each user intention feature of the question information is analyzed through an intention recognition program, and each user intention feature is used as the user intention information corresponding to the question information.

[0013] Optionally, based on the question information and the user intention information, extracting each question demand information of the user and the core question content in each question demand information through a demand extraction strategy includes:

[0014] Extracting each key question text in the question text content, and identifying the sub-text semantics corresponding to each key question text based on the text semantic information;

[0015] Based on the sub-text semantics corresponding to each of the key question texts, the text feature information corresponding to each of the key question texts is identified through a question requirement identification model, and the question requirement information corresponding to each of the text feature information is searched in a question requirement database;

[0016] Based on each of the user intention features and each of the text feature information, the target text feature information corresponding to each user intention feature is screened through a key indicator database;

[0017] The key question text corresponding to each target question feature information is used as the core question content in each of the question requirement information.

[0018] Optionally, the querying of the answer content corresponding to each of the question requirement information includes:

[0019] Acquire the question type corresponding to each of the question requirement information and the question query method of each question query type, and identify the question query method corresponding to each of the question requirement information based on each of the question types;

[0020] Based on each of the question demand information, the question answer content corresponding to each of the question demand information is queried through the question query method corresponding to each of the question demand information.

[0021] Optionally, the step of identifying the completeness of the answers to the core questions based on the answers to the questions includes:

[0022] In the answer to each question, identify the core question answer content corresponding to each core question content, and for each question answer content, identify each core problem point of the question answer content;

[0023] Through the gap information calculation strategy, the unanswered core problem points in each of the core problem points are identified, and based on the number of each of the unanswered core problem points and the number of each of the core problem points, the completeness of the answer to the core problem content is calculated.

[0024] Optionally, the information is used to complete the calling program to obtain the supplementary reply information corresponding to each of the question requirement information, including:

[0025] Based on the unanswered core issues, generate a list of core missing information and identify the unanswered issues for each non-core issue content;

[0026] Generate a missing information list based on each of the unanswered question points, and construct an information collection strategy based on the core missing information list and the missing information list;

[0027] Based on the core missing information list and the missing information list, the supplementary response information corresponding to each of the problem requirement information is obtained through the information collection strategy.

[0028] In a second aspect, the present application also provides a question answering device based on multi-agent collaboration and reinforcement learning optimization, comprising:

[0029] An acquisition module, used to acquire user's question information and identify user intention information corresponding to the question information;

[0030] A query module, for extracting the user's question demand information and the core question content in each question demand information based on the question information and the user intention information through a demand extraction strategy, and querying the question answer content corresponding to each question demand information;

[0031] An identification module is used to identify the completeness of the answers to the core questions based on the answers to the questions, and when the completeness of the answers to the core questions is lower than a preset completeness threshold, obtain the supplementary answer information corresponding to the requirement information of the questions through an information completion calling program;

[0032] A generation module is used to generate the user's target answer information based on the question answer content corresponding to each question requirement information and the supplementary answer information corresponding to each question requirement information.

[0033] Optionally, the acquisition module is specifically used to:

[0034] Extracting the question text content of the question information, and identifying the text semantic information corresponding to the question text content;

[0035] Based on the text semantic information and the question text content, each user intention feature of the question information is analyzed through an intention recognition program, and each user intention feature is used as the user intention information corresponding to the question information.

[0036] Optionally, the query module is specifically used to:

[0037] Extracting each key question text in the question text content, and identifying the sub-text semantics corresponding to each key question text based on the text semantic information;

[0038] Based on the sub-text semantics corresponding to each of the key question texts, the text feature information corresponding to each of the key question texts is identified through a question requirement identification model, and the question requirement information corresponding to each of the text feature information is searched in a question requirement database;

[0039] Based on each of the user intention features and each of the text feature information, the target text feature information corresponding to each user intention feature is screened through a key indicator database;

[0040] The key question text corresponding to each target question feature information is used as the core question content in each of the question requirement information.

[0041] Optionally, the query module is specifically used to:

[0042] Acquire the question type corresponding to each of the question requirement information and the question query method of each question query type, and identify the question query method corresponding to each of the question requirement information based on each of the question types;

[0043] Based on each of the question demand information, the question answer content corresponding to each of the question demand information is queried through the question query method corresponding to each of the question demand information.

[0044] Optionally, the identification module is specifically used to:

[0045] In the answer to each question, identify the core question answer content corresponding to each core question content, and for each question answer content, identify each core problem point of the question answer content;

[0046] Through the gap information calculation strategy, the unanswered core problem points in each of the core problem points are identified, and based on the number of each of the unanswered core problem points and the number of each of the core problem points, the completeness of the answer to the core problem content is calculated.

[0047] Optionally, the identification module is specifically used to:

[0048] Based on the unanswered core issues, generate a list of core missing information and identify the unanswered issues for each non-core issue content;

[0049] Generate a missing information list based on each of the unanswered question points, and construct an information collection strategy based on the core missing information list and the missing information list;

[0050] Based on the core missing information list and the missing information list, the supplementary response information corresponding to each of the problem requirement information is obtained through the information collection strategy.

[0051] In a third aspect, the present application provides a computer device, wherein the computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of any one of the methods in the first aspect are implemented.

[0052] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of any one of the methods in the first aspect are implemented.

[0053] In a fifth aspect, the present application provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the steps of any one of the methods in the first aspect are implemented.

[0054] The above-mentioned method and device for answering questions based on multi-agent collaboration and reinforcement learning optimization obtains the user's question information and identifies the user's intention information corresponding to the question information; based on the question information and the user's intention information, extracts the user's question demand information and the core question content in each of the question demand information through the demand extraction strategy, and queries the question answer content corresponding to each of the question demand information; based on the question answer content, identifies the completeness of the question answer of each of the core question content, and when the completeness of the question answer of the core question content is lower than the preset completeness threshold, obtains the supplementary answer information corresponding to each of the question demand information through the information improvement calling program; based on the question answer content corresponding to each of the question demand information and the supplementary answer information corresponding to each of the question demand information, generates the user's target answer information. This scheme combines the user's intention information to filter the core question content in the user's question information, so that when analyzing the completeness of the question answer, the completeness of the core question content is used as the analysis target, so as to give priority to analyzing the core content of the user, improve the completeness of the answer to the core question content when answering the user's question information, and the priority of the answer, and avoid the problem of failing to give priority to the question content that satisfies the user's most core concerns. Then, when actually answering, this solution can self-check, optimize, and supplement the payment content, thereby ensuring the completeness of the answers to users' questions and avoiding the situation where the answer efficiency is too low and the question cannot be fully answered when complex questions are involved. Especially when the question content is designed to include more than one question, including multiple questions, it is possible to answer and supplement the questions according to the information required by each question, thereby comprehensively improving the completeness and accuracy of the answers to the questions. Finally, this solution does not require manual answers, but can answer and supplement intelligently, thereby improving the efficiency of answering complex questions, so that when answering the user's question information, it can answer the question content efficiently and comprehensively, and can give priority to complete answers to the core question content, and ensure the comprehensiveness and accuracy of the answers to the needs of each question, thereby comprehensively improving the user satisfaction after answering the user's questions when answering complex questions. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related technologies, the drawings required for use in the embodiments or the related technical descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0056] Figure 1 A flowchart of a question answering method based on multi-agent collaboration and reinforcement learning optimization in one embodiment;

[0057] Figure 2 A flowchart of an example of question answering based on multi-agent collaboration and reinforcement learning optimization in one embodiment;

[0058] Figure 3 It is a structural block diagram of a question answering device based on multi-agent collaboration and reinforcement learning optimization in one embodiment;

[0059] Figure 4 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0061] The question answering method based on multi-agent collaboration and reinforcement learning optimization provided in the embodiment of the present application can be applied to the application environment of question answering based on multi-agent collaboration and reinforcement learning optimization. Among them, the method can be applied to a terminal, a server, or a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. Among them, the terminal can be, but is not limited to, various personal computers, laptops, medium-sized computers, etc. Among them, the terminal combines the user's intention information to screen the core question content in the user's question information, so that when analyzing the completeness of the question answer, the completeness of the core question content is used as the analysis target, so as to give priority to the analysis of the user's core content, improve the completeness of the answer to the core question content when answering the user's question information, and the priority of the answer, and avoid the problem of not being able to give priority to the question content that satisfies the user's most core concerns. Then, this solution can self-check, optimize, and supplement the answer to the payment content when actually answering, thereby ensuring the integrity of the answer to the user's question, and avoiding the situation where the answer efficiency is too low when it comes to complex questions, and the question cannot be answered completely. In particular, when the question content is designed to include more than one question, it is possible to answer and supplement the questions according to the information required by each question, thereby comprehensively improving the completeness and accuracy of the question answers. Finally, this solution does not require manual answers, but can provide intelligent answers and intelligent supplements, thereby improving the efficiency of answering complex questions, so that when answering user question information, it can answer the question content efficiently and comprehensively, and can give priority to complete answers to core question content, and ensure the comprehensiveness and accuracy of the answers to each question requirement, thereby comprehensively improving the user satisfaction after answering the user's questions when answering complex questions.

[0062] In an exemplary embodiment, Figure 1As shown, a question answering method based on multi-agent collaboration and reinforcement learning optimization is provided, and the method is applied to a terminal as an example for explanation, including the following steps S101 to S104. Among them:

[0063] Step S101, obtaining user question information and identifying user intention information corresponding to the question information.

[0064] In this embodiment, the terminal responds to the user's question input operation and obtains the user's question content, wherein the question content can be text content or audio content. After obtaining the content of each question, the terminal normalizes the format of each question content to obtain the question information of each text content. Then, the terminal identifies the user's intention information in the question information through the intention recognition module preset in the terminal, wherein the intention recognition module is an (IntentAgent) intention recognition agent program. Among them, the user's intention information recognized by the intention recognition module is the user's intention information on the question type and the business module to which each question belongs. Among them, question types are such as (width, depth, mixed, fuzzy, etc.), and business modules are such as (inventory management, financial analysis, sales recognition, credit inquiry, etc.). The specific recognition process will be described in detail later.

[0065] Step S102, based on the question information and the user intention information, through the demand extraction strategy, extract the user's question demand information and the core question content in each question demand information, and query the question answer content corresponding to each question demand information.

[0066] In this embodiment, the terminal extracts the user's question demand information and the core question content in each question demand information through a demand extraction strategy based on the question information and the user's intention information, and queries the question answer content corresponding to each question demand information. Among them, each question demand information is the information that needs to be queried for each question content in the user's question information, and the query information can be queried in the business database, knowledge graph, historical cache / memory file (memory.txt), and internal data API database stored in the terminal. Each question demand information includes the core question content, which is the key question content that the user is mainly concerned about in each question demand information, that is, there is question demand information in each question demand information that does not contain the core question content, and the core question content exists between one or more question demand information. The specific identification process will be described in detail later.

[0067] Step S103, based on the answer content of each question, identify the completeness of the answer to each core question content, and when the completeness of the answer to the core question content is lower than the preset completeness threshold, obtain the supplementary answer information corresponding to the requirement information of each question through the information completion call program.

[0068] In this embodiment, the terminal identifies the completeness of the answers to the questions of each core question based on the answer content of each question, and when the completeness of the answers to the core question content is lower than the preset completeness threshold, the terminal obtains the supplementary answer information corresponding to each question requirement information through the information completion calling program. The question answer completeness is the completeness of the answer to the core question content. Among them, the completeness of the answer to the core question content is the ratio between the number of core information points of the corresponding answer information and the number of all core information points in each core information point of the core question content. The specific process of obtaining the supplementary answer information corresponding to each question requirement information will be described in detail later.

[0069] Step S104, based on the question answer content corresponding to each question requirement information and the supplementary answer information corresponding to each question requirement information, the user's target answer information is generated.

[0070] In this embodiment, the terminal generates the user's target answer information based on the question answer content corresponding to each question demand information and the supplementary answer information corresponding to each question demand information. Specifically, the terminal arranges the question answer content and the supplementary answer information corresponding to the core question content in the front, and arranges the question answer content and the supplementary answer information corresponding to the non-core question content in the back, and each question answer content and the supplementary answer information are provided with the corresponding question content, thereby constructing a target answer summary table of the user's question information, and the target answer summary table is used as the user's target answer information.

[0071] Based on the above scheme, by combining the user's intention information, the core question content in the user's question information is screened, so that when analyzing the completeness of the question answer, the completeness of the core question content is used as the analysis target, so as to give priority to analyzing the user's core content, improve the completeness of the answer to the core question content when answering the user's question information, and the priority of the answer, avoiding the problem of not being able to give priority to the question content that meets the user's most core concerns. Then, this scheme can self-check, optimize, and supplement the payment content when actually answering, thereby ensuring the completeness of the answer to the user's question, and avoiding the situation where the answer efficiency is too low and the question cannot be fully answered when complex questions are involved. In particular, when the question content is designed to include more than one question, including multiple questions, the question can be answered and supplemented according to the information required by each question, thereby comprehensively improving the completeness and accuracy of the question answer. Finally, this solution does not require manual replies, but can provide intelligent replies and intelligent supplements, thereby improving the efficiency of answering complex questions. When answering users' question information, it is possible to answer the question content efficiently and comprehensively, and give priority to complete replies to core questions, and ensure the comprehensiveness and accuracy of the responses to each question, thereby comprehensively improving user satisfaction after answering user questions when answering complex questions.

[0072] Optionally, identifying user intention information corresponding to the question information includes: extracting question text content of the question information, and identifying text semantic information corresponding to the question text content; based on the text semantic information and the question text content, parsing each user intention feature of the question information through an intention recognition program, and using each user intention feature as the user intention information corresponding to the question information.

[0073] In this embodiment, the terminal extracts the question text content of the question information and identifies the text semantic information corresponding to the question text content. The terminal identifies the text semantic information corresponding to the text content by using a large language model based on natural language processing technology to identify the text semantic information corresponding to the text content.

[0074] Then, based on the text semantic information and the question text content, the terminal analyzes the user intention features of the question information through the intention recognition program, and uses the user intention features as the user intention information corresponding to the question information. Among them, the text semantic information includes the sub-semantic information corresponding to each sub-question text content, that is, when the text semantic information includes the question semantic information of each question type, the text content corresponding to each question type corresponds to a sub-semantic information, and the sub-semantic information corresponding to each sub-question text content is obtained. Then, based on the sub-semantic information corresponding to each sub-question text content, the terminal queries the question type (i.e., user intention feature) to which each sub-question text content belongs in the question database, and uses the question type to which each sub-question text content belongs and the sub-semantic information corresponding to each sub-question text content as the user intention information corresponding to the question information.

[0075] Based on the above solution, the question type corresponding to each question information is identified based on the text semantic content of the question information, and the content of each question is split and analyzed, thereby improving the comprehensiveness and accuracy of the recognition of user intentions.

[0076] Optionally, based on the question information and the user intention information, a demand extraction strategy is used to extract the user's question demand information and the core question content in each question demand information, including: extracting each key question text in the question text content, and identifying the sub-text semantics corresponding to each key question text based on the text semantic information; based on the sub-text semantics corresponding to each key question text, identifying the text feature information corresponding to each key question text through a question demand identification model, and querying the question demand information corresponding to each text feature information in a question demand database; based on each user intention feature and each text feature information, filtering the target text feature information corresponding to each user intention feature through a key indicator database; using the key question text corresponding to each target question feature information as the core question content in each question demand information.

[0077] In this embodiment, the terminal extracts each key question text in the question text content, and identifies the sub-text semantics corresponding to each key question text based on the text semantic information. Among them, the sub-text semantics is the text semantic information of each question type. Then the terminal identifies the text feature information corresponding to each key question text through the question requirement recognition model based on the sub-text semantics corresponding to each key question text, and queries the question requirement information corresponding to each text feature information in the question requirement database. Among them, the text feature information is the question information point corresponding to each key question text. Among them, the question information point is the lowest granularity data unit required to answer the question. For example, "current inventory", "last month's sales", and "user authority list" are all information points. The terminal uses natural language understanding (NLU) technology and predefined business domain knowledge to perform semantic analysis on user questions. For example, the key question text of the user question is: "What is the sales volume last month and the inventory situation as of this month?" The information points that may be extracted are: sales (last month); inventory (this month as of now); overall y set = {last month's sales information point, this month's inventory information point); the terminal can use the pre-trained language model and rule combination method:

[0078] The terminal uses the NLU model to identify contextual features such as business entities (such as "sales", "inventory") and time conditions ("last month", "this month so far") involved in the question. Then, the domain knowledge graph is used to map the identified entities, determine their corresponding actual data fields or API dependencies, and obtain a clearer set of y. Then, among the acquired y information points, the terminal needs to mark the parts that are critical to the answer as core information points z.

[0079] Judgment basis: Some information points can be defined as "key points" based on business strategy, user problem orientation and common sense. If the user's primary demand is to understand the key KPI (such as sales) before drawing any conclusions, then these information points should belong to z.

[0080] Policy example:

[0081] If the question is "To determine the effectiveness of this quarter's marketing strategy, I need to know this quarter's sales and user conversion rate", then "this quarter's sales" and "user conversion rate" are z, because without these two data points, the rest of the information analysis is out of the question.

[0082] Technically, z can be automatically extracted by matching a pre-defined list of key indicators, user intent features, and priority rules annotated in the training set.

[0083] Finally, we can conclude that the set z is a subset of y (zcy). Usually, the number of z is no more than y. Sometimes z may even be a true subset of y. In some problems, z=y (all information points are equally important).

[0084] Based on the above solution, by using natural language understanding (NLU) technology and predefined business domain knowledge to perform semantic analysis on user questions, the core problem content in the problem demand information is obtained, which improves the accuracy and comprehensiveness of identifying the core problem content.

[0085] Optionally, query the question answer content corresponding to each question demand information, including: obtaining the question type corresponding to each question demand information, and the question query method of each question query type, and based on each question type, identifying the question query method corresponding to each question demand information; based on each question demand information, query the question answer content corresponding to each question demand information through the question query method corresponding to each question demand information.

[0086] In this embodiment, the terminal obtains the question type corresponding to each question demand information and the question query method of each question query type, and identifies the question query method corresponding to each question demand information based on each question type. Then, based on each question demand information, the terminal queries the question answer content corresponding to each question demand information through the question query method corresponding to each question demand information. Among them, the terminal determines how much information the current system already has that can directly answer or partially answer the user's needs. This step uses the following data sources:

[0087] Knowledge graph: A business knowledge base built in advance by the system, which contains entity relationships, known facts, and cached data.

[0088] Historical cache / memory file (memory.txt): Previously collected and stored information data points in the form of key-value pairs or structured records.

[0089] Quick query of internal data AP (optional): quickly access the internal database to obtain immediately available information (such as some existing permanent data of the company, such as the exchange rate of the day and yesterday's inventory). This is a lightweight query and is not equivalent to the business API that is planned to be called later.

[0090] The following steps can be used for technical implementation:

[0091] Match existing data for each information point in y: According to the identifier of the information point (for example, "last month's sales" is mapped to the database field monthly_sales(last_month)), search the knowledge graph or cache to see if there is a ready-made value.

[0092] The terminal adds the successfully found and real-time valid information points to the x set, which is the currently known and available information point set.

[0093] Based on the above solution, question queries are performed according to different question query methods based on the question demand information of different question types, thereby ensuring the accuracy and comprehensiveness of the query for each question.

[0094] Optionally, based on the answer content of each question, the completeness of the answer to each core question content is identified, including: in the answer content of each question, the core question answer content corresponding to each core question content is identified, and for each question answer content, each core problem point of the question answer content is identified; through a gap information calculation strategy, the unanswered core problem points in each core problem point are identified, and based on the number of each unanswered core problem points and the number of each core problem point, the completeness of the answer to the core question content is calculated.

[0095] In this embodiment, the terminal identifies the core question answer content corresponding to each core question content in each question answer content, and identifies each core question point of the question answer content for each question answer content, wherein the core question point is the core information point in the core question content.

[0096] Then, the terminal identifies the unanswered core problem points among the core problem points through the gap information calculation strategy, and calculates the completeness of the answer to the core problem content based on the number of unanswered core problem points and the number of each core problem point.

[0097] Specifically, after the terminal completes the construction of the x set, it compares y with x to determine what information is still missing:

[0098] z needed =z-(z∩x) if z needed is an empty set, it means that all core information points z have been found in x and no additional collection is needed. needed Contains those core information points that are not met and need to be collected as a priority.

[0099] y needed =y-(y∩x)

[0100] Among them, y needed Represents the missing information in non-core information points, as well as the missing parts in core information (in fact, z needed Yes needed A subset or partial intersection of the responses) is used as a reference in subsequent information collection so as to decide whether to further improve the answers after the core needs are met.

[0101] Then, the terminal checks z needed Whether it is empty is used to determine whether the core requirements have been met.

[0102] If needed Is empty:

[0103] core_satisfied=True

[0104] This means that the information required to answer the user's core question is available, and there is no need to call external APIs to further collect other non-essential information.

[0105] If needed Not empty:

[0106] core_satisfied=False

[0107] This means that you need to enter the subsequent information collection plan and API call process to obtain z needed Information points in .

[0108] The above solution is implemented by adopting the "model + rule" fusion method: first use the NLU model to automatically extract possible information points (core problem points), and then use business rules to mark the core information according to priority and context.

[0109] Explainability and Scalability:

[0110] y, z, x, z output by FactAgent needed ,y needed It is a transparent and explainable collection with clear and structured requirements for problems. It is very convenient to add new information points or modify business priority rules.

[0111] Performance optimization:

[0112] Frequently used information points can be cached in advance to speed up the construction of x sets. For knowledge graph queries, high-performance graph databases (such as Neo4j) or in-memory databases (Redis) can be used to speed up retrieval.

[0113] (IV) Information Planning and Decision-making Module (PlannerAgent)

[0114] PlannerAgent plans based on FactAgent output and core_satisfied status:

[0115] 1. If core_satisfied=Tue, it means that the core needs of users can be met without obtaining additional information. PlannerAgent can directly end the information collection phase and move to the final answer generation phase; or decide whether to supplement non-core information in y based on business strategies.

[0116] 2. If core_satisfied=False, PlannerAgent formulates information collection strategy and calling route:

[0117] Consider multi-dimensional information acquisition in parallel or sequential calling.

[0118] According to z needed The priority determines the calling order, and the information required by z is obtained first.

[0119] Check user permissions and resource quotas. If the key AP permissions are insufficient, trigger a rejection or request the user to provide more permission information.

[0120] Output the call plan path diagram, including the API call sequence, parallel relationship, and data dependency. The following is a comprehensive description of the technical implementation details of the information planning and decision module (PiannerAgent). This module is the hub of the entire strategy execution link. Its core function is to convert information requirements into executable resource acquisition paths based on the output of FactAgent, thereby achieving efficient, controllable and fast answers. This link is also the key from abstract problem requirements to specific API call sequences (question map answer path).

[0121] Core logic and process details:

[0122] 1. Input and Preconditions

[0123] When PlannerAgent receives the results from FactAgent, it can obtain:

[0124] The structured information point set (y, z) of the user's question, and z needed and needed Information gaps

[0125] core_satisfied status (True / False), indicating whether the core information point is ready.

[0126] Existing permissions and resource information (may come from user context or permission tags during the intent identification phase).

[0127] In addition, PlannerAgent has the ability to access the enterprise's internal AP1 directory, data dependency mapping, and call cost model.

[0128] 2. Core decision-making ideas:

[0129] Quick response strategy: prioritize user core information points needed Only when z needed When it is not empty, Planner needs to formulate information collection strategies to fill the gaps in these key information points.

[0130] If core_satisfied=Tiue: means z neededis empty, the core demand point has been met. At this time, PlannerAgent can directly stop information collection according to the business strategy and enter the answer generation stage to achieve rapid response. If the business strategy requires the answer to be as comprehensive as possible, you can also consider continuing to collect y needed The remaining information in the answer can enhance the completeness of the answer.

[0131] If core_satisfied=False: It means that there is still a gap in the core information necessary to answer the user's question, and it needs to be obtained from an external data source (AP). At this time, PlannerAgent needs to develop an access roadmap to call the appropriate API to obtain z needed Information points in .

[0132] 3. Information planning and problem map construction:

[0133] y needed With z needed Transformed into a "Problem Graph" or "Information Needs Graph". In this graph:

[0134] Nodes represent points of information (such as "last month's sales", "current inventory").

[0135] Edges represent data dependencies (such as “getting information point A before calling AP1_B to get information point B”).

[0136] Some information points can be obtained in parallel (such as inventory information and weather data are not related and can be obtained at the same time); others have a sequence requirement (such as the need to obtain the store ID before calling the inventory API).

[0137] 4. Priority and scheduling strategy:

[0138] Based on z needed For the core:

[0139] Right needed The information points in the z set are prioritized for path planning to ensure that the API call sequence that satisfies the z set is obtained as quickly as possible. needed There are multiple information points in , and it is necessary to determine whether they can be obtained in parallel or in sequence.

[0140] If some information points have no dependencies, multiple APIs can be called in parallel to speed up the acquisition process.

[0141] For scenarios that require deep chaining, such as obtaining information C requires obtaining information A first and then information B before finally obtaining C, PlannerAgent will build an ordered API call chain (e.g., first call API_A → obtain A → call API_B with parameter A → obtain B → call APIC with parameter B → obtain C).

[0142] 5. Permission and resource check:

[0143] When generating a plan, PlannerAgent needs to check user permissions and resource limits:

[0144] If a key API requires advanced permissions and the current user does not have them, z cannot be satisfied. needed This will trigger a policy that refuses to answer or prompts the user to escalate permissions.

[0145] If the resource quota is insufficient (for example, the number of API calls has reached the upper limit), Planner can try to use alternative information sources, or prompt the user to confirm additional consumption before making the call.

[0146] Through the pre-configured permission mapping table and call cost model, PlannerAgent prunes infeasible paths during planning to avoid wasting subsequent calculations and attempts.

[0147] 6. Call the planned path output:

[0148] The output of PlannerAgent is a clear and structured call plan graph (CallPlanGraph), describing:

[0149] Which APIs need to be called?

[0150] . The order and parallel relationship of calls;

[0151] . Parameter passing and data dependency.

[0152] Example:

[0153] If z needed It contains two information points: "last month's sales" and "this month's inventory":

[0154] To obtain "last month's sales", you need to call the financial data API: AP1_Finance (no pre-dependency, can be called immediately);

[0155] "This Month's Inventory" needs to obtain the warehouse ID (known as x) before calling the inventory AP1: The API_Inventory path diagram can be expressed as:

[0156] Start-->CallAPI_Finance->GetSalesInfo-->CallAPI_Inventory-->GetInventoryInfo-->End.

[0157] If neither has a prerequisite, the two can be called in parallel. If there is a dependency, the order of precedence will be clearly indicated in the path with directed edges.

[0158] 7. Time, cost and alternatives:

[0159] PlannerAgent can also optimize the call path, such as selecting an API version with lower latency or lower cost, or preparing an alternative API path in case of failure.

[0160] An "API cost model (CostMadel)" and "API reliability statistics table" can be maintained internally to learn the best path from historical data. In the subsequent reinforcement learning phase, better strategies will be continuously applied.

[0161] 8. Iteration and adjustment:

[0162] If the call to AP1 fails in the subsequent Action,Agents, the ObservationAgent will report the situation back to the PlannerAgent.

[0163] PlannerAgent can choose an alternate path, retry, or ultimately reject the request.

[0164] This process is dynamic, and PlannerAgcnt also has the ability to re-plan if an exception is encountered after the initial planning.

[0165] 9. Quick response and decision points:

[0166] When core_satisfied=True, PlannerAgent can stop collecting information immediately, allowing the system to quickly generate answers when it has mastered the core information points, thereby improving response efficiency.

[0167] If the business policy or user explicitly requires as complete an answer as possible, you can choose to continue collecting y on this basis. needed Non-core information points to increase the richness of the answer.

[0168] Based on the above solution, by identifying the core problem points, it is determined whether the core problem content has been answered, which improves the completeness and priority of the answers to the core problem content, ensures that the user's core problem content can receive priority and complete answers, and improves user satisfaction.

[0169] Optionally, through the information improvement calling program, supplementary response information corresponding to each problem requirement information is obtained, including: based on each unanswered core problem point, a core missing information list is generated, and each unanswered problem point of each non-core problem content is identified; based on each unanswered problem point, a missing information list is generated, and based on the core missing information list and the missing information list, an information collection strategy is constructed; based on the core missing information list and the missing information list, supplementary response information corresponding to each problem requirement information is obtained through the information collection strategy.

[0170] In this embodiment, the terminal generates a core missing information list based on each unanswered core problem point, and identifies each unanswered problem point of each non-core problem content. Then. The terminal generates a missing information list based on each unanswered problem point, and constructs an information collection strategy based on the core missing information list and the missing information list. After that, the terminal obtains the supplementary reply information corresponding to each problem requirement information through the information collection strategy based on the core missing information list and the missing information list. This information collection strategy is centered on each unanswered problem point and the unanswered core problem point, relies on graph search and optimization, and selects the strategy of preferentially calling the API agent in the order of first unanswered core problem points and then unanswered problem points through a heuristic or greedy strategy, so as to query the reply content corresponding to each problem point through each reply question query method.

[0171] Based on the above solution, after screening each problem point, targeted information collection of the problem point is carried out through information collection strategy, which improves the collection efficiency and accuracy of supplementary reply information.

[0172] This application also provides an example of question answering based on multi-agent collaboration and reinforcement learning optimization, such as Figure 2 As shown, the specific processing process includes the following steps:

[0173] Step S201, obtaining user's question information.

[0174] Step S202: extract the question text content of the question information, and identify the text semantic information corresponding to the question text content.

[0175] Step S203, based on the text semantic information and the question text content, the user intention features of the question information are analyzed through the intention recognition program, and each user intention feature is used as the user intention information corresponding to the question information.

[0176] Step S204: extract each key question text in the question text content, and identify the sub-text semantics corresponding to each key question text based on the text semantic information.

[0177] Step S205, based on the sub-text semantics corresponding to each key question text, the text feature information corresponding to each key question text is identified through the question requirement identification model, and the question requirement information corresponding to each text feature information is searched in the question requirement database.

[0178] Step S206, based on each user intention feature and each text feature information, the target text feature information corresponding to each user intention feature is screened through the key indicator database.

[0179] Step S207: Use the key question text corresponding to each target question feature information as the core question content in each question requirement information.

[0180] Step S208, obtaining the question type corresponding to each question requirement information and the question query method of each question query type, and identifying the question query method corresponding to each question requirement information based on each question type.

[0181] Step S209, based on each question demand information, query the question answer content corresponding to each question demand information through the question query method corresponding to each question demand information.

[0182] Step S210, in each question answer content, identifying the core question answer content corresponding to each core question content, and for each question answer content, identifying each core problem point of the question answer content.

[0183] Step S211, through the gap information calculation strategy, identify the unanswered core problem points in each core problem point, and calculate the question answer completeness of the core problem content based on the number of each unanswered core problem point and the number of each core problem point.

[0184] Step S212: Generate a core missing information list based on each unanswered core problem point, and identify each unanswered problem point of each non-core problem content.

[0185] Step S213, generating a missing information list based on each unanswered question point, and constructing an information collection strategy based on the core missing information list and the missing information list.

[0186] Step S214, based on the core missing information list and the missing information list, the supplementary response information corresponding to each question requirement information is obtained through the information collection strategy.

[0187] Step S215, based on the question answer content corresponding to each question requirement information and the supplementary answer information corresponding to each question requirement information, the user's target answer information is generated.

[0188] It should be understood that, although the various steps in the flowcharts involved in the above-mentioned embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-mentioned embodiments can include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0189] Based on the same inventive concept, the embodiment of the present application also provides a question answering device based on multi-agent collaboration and reinforcement learning optimization for implementing the question answering method based on multi-agent collaboration and reinforcement learning optimization involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more embodiments of the question answering device based on multi-agent collaboration and reinforcement learning optimization provided below can refer to the limitations of the question answering method based on multi-agent collaboration and reinforcement learning optimization above, and will not be repeated here.

[0190] In an exemplary embodiment, Figure 3 As shown, a question answering device based on multi-agent collaboration and reinforcement learning optimization is provided, including: an acquisition module 310, a query module 320, an identification module 330 and a generation module 340, wherein:

[0191] The acquisition module 310 is used to acquire the user's question information and identify the user's intention information corresponding to the question information;

[0192] The query module 320 is used to extract the user's question demand information and the core question content in each question demand information based on the question information and the user intention information through the demand extraction strategy, and query the question answer content corresponding to each question demand information;

[0193] The identification module 330 is used to identify the completeness of the answer to each of the core questions based on the answer content of each of the questions, and when the completeness of the answer to the core question content is lower than a preset completeness threshold, obtain the supplementary answer information corresponding to each of the question requirement information through the information completion calling program;

[0194] The generation module 340 is used to generate the user's target answer information based on the question answer content corresponding to each of the question requirement information and the supplementary answer information corresponding to each of the question requirement information.

[0195] Optionally, the acquisition module 310 is specifically configured to:

[0196] Extracting the question text content of the question information, and identifying the text semantic information corresponding to the question text content;

[0197] Based on the text semantic information and the question text content, each user intention feature of the question information is analyzed through an intention recognition program, and each user intention feature is used as the user intention information corresponding to the question information.

[0198] Optionally, the query module 320 is specifically used to:

[0199] Extracting each key question text in the question text content, and identifying the sub-text semantics corresponding to each key question text based on the text semantic information;

[0200] Based on the sub-text semantics corresponding to each of the key question texts, the text feature information corresponding to each of the key question texts is identified through a question requirement identification model, and the question requirement information corresponding to each of the text feature information is searched in a question requirement database;

[0201] Based on each of the user intention features and each of the text feature information, the target text feature information corresponding to each user intention feature is screened through a key indicator database;

[0202] The key question text corresponding to each target question feature information is used as the core question content in each of the question requirement information.

[0203] Optionally, the query module 320 is specifically used to:

[0204] Acquire the question type corresponding to each of the question requirement information and the question query method of each question query type, and identify the question query method corresponding to each of the question requirement information based on each of the question types;

[0205] Based on each of the question demand information, the question answer content corresponding to each of the question demand information is queried through the question query method corresponding to each of the question demand information.

[0206] Optionally, the identification module 330 is specifically used to:

[0207] In the answer to each question, identify the core question answer content corresponding to each core question content, and for each question answer content, identify each core problem point of the question answer content;

[0208] Through the gap information calculation strategy, the unanswered core problem points in each of the core problem points are identified, and based on the number of each of the unanswered core problem points and the number of each of the core problem points, the completeness of the answer to the core problem content is calculated.

[0209] Optionally, the identification module 330 is specifically used to:

[0210] Based on the unanswered core issues, generate a list of core missing information and identify the unanswered issues for each non-core issue content;

[0211] Generate a missing information list based on each of the unanswered question points, and construct an information collection strategy based on the core missing information list and the missing information list;

[0212] Based on the core missing information list and the missing information list, the supplementary response information corresponding to each of the problem requirement information is obtained through the information collection strategy.

[0213] Each module in the above-mentioned question answering device based on multi-agent collaboration and reinforcement learning optimization can be implemented in whole or in part by software, hardware and their combination. Each of the above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to each of the above modules.

[0214] In an exemplary embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as shown in FIG. Figure 4As shown. The computer device includes a processor, a memory, an input / output interface, a communication interface, a display unit and an input device. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface, the display unit and the input device are connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and the external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be realized through WIFI, a mobile cellular network, NFC (near field communication) or other technologies. When the computer program is executed by the processor, a question answering method based on multi-agent collaboration and reinforcement learning optimization is realized. The display unit of the computer device is used to form a visually visible picture, which can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device shell, or an external keyboard, touchpad or mouse.

[0215] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0216] In an exemplary embodiment, a computer device is provided, including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, steps corresponding to a question answering method based on multi-agent collaboration and reinforcement learning optimization are implemented.

[0217] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, steps corresponding to the question answering method based on multi-agent collaboration and reinforcement learning optimization are implemented.

[0218] In one embodiment, a computer program product is provided, comprising a computer program, which, when executed by a processor, implements steps corresponding to a question answering method based on multi-agent collaboration and reinforcement learning optimization.

[0219] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0220] A person of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. Non-relational databases may include distributed databases based on blockchains, etc., but are not limited to this. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but are not limited to this.

[0221] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0222] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A question answering method based on multi-agent collaboration and reinforcement learning optimization, characterized in that: The method comprises: Obtaining user question information and identifying user intention information corresponding to the question information; Based on the question information and the user intention information, extract the user's question demand information and the core question content in each question demand information through a demand extraction strategy, and query the question answer content corresponding to each question demand information; In the answer to each question, identify the core question answer content corresponding to each core question content, and for each question answer content, identify each core problem point of the question answer content; By using the gap information calculation strategy, the unanswered core problem points in each of the core problem points are identified, and based on the number of each of the unanswered core problem points and the number of each of the core problem points, the completeness of the answer to the core problem content is calculated; When the completeness of the answers to the questions with core question content is lower than the preset completeness threshold, the information completion calling program is used to obtain the supplementary answer information corresponding to the required information of each question; The user's target answer information is generated based on the question answer content corresponding to each of the question demand information and the supplementary answer information corresponding to each of the question demand information.

2. The method according to claim 1, characterized in that The identifying user intention information corresponding to the question information includes: Extracting the question text content of the question information, and identifying the text semantic information corresponding to the question text content; Based on the text semantic information and the question text content, each user intention feature of the question information is analyzed through an intention recognition program, and each user intention feature is used as the user intention information corresponding to the question information.

3. The method according to claim 2, characterized in that The extracting of each question demand information of the user and the core question content in each question demand information based on the question information and the user intention information through a demand extraction strategy includes: Extracting each key question text in the question text content, and identifying the sub-text semantics corresponding to each key question text based on the text semantic information; Based on the sub-text semantics corresponding to each of the key question texts, the text feature information corresponding to each of the key question texts is identified through a question requirement identification model, and the question requirement information corresponding to each of the text feature information is searched in a question requirement database; Based on each of the user intention features and each of the text feature information, the target text feature information corresponding to each user intention feature is screened through a key indicator database; The key question text corresponding to each target question feature information is used as the core question content in each of the question requirement information.

4. The method according to claim 1, characterized in that: The query of the answer content corresponding to each of the question requirement information includes: Acquire the question type corresponding to each of the question requirement information and the question query method of each question query type, and identify the question query method corresponding to each of the question requirement information based on each of the question types; Based on each of the question demand information, the question answer content corresponding to each of the question demand information is queried through the question query method corresponding to each of the question demand information.

5. The method according to claim 1, characterized in that The information is used to complete the calling program, and the supplementary reply information corresponding to each of the question requirement information is obtained, including: Based on the unanswered core issues, generate a list of core missing information and identify the unanswered issues for each non-core issue content; Generate a missing information list based on each of the unanswered question points, and construct an information collection strategy based on the core missing information list and the missing information list; Based on the core missing information list and the missing information list, the supplementary response information corresponding to each of the problem requirement information is obtained through the information collection strategy.

6. A question answering device based on multi-agent collaboration and reinforcement learning optimization, characterized in that: The device comprises: An acquisition module, used to acquire user question information and identify user intention information corresponding to the question information; A query module, for extracting the user's question demand information and the core question content in each question demand information based on the question information and the user intention information through a demand extraction strategy, and querying the question answer content corresponding to each question demand information; An identification module is used to identify the core question answer content corresponding to each core question content in the answer content of each question, and identify each core problem point of the answer content for each question answer content; identify the unanswered core problem points in each core problem point through a gap information calculation strategy, and calculate the completeness of the answer to the core question content based on the number of each unanswered core problem point and the number of each core problem point; when the completeness of the answer to the core question content is lower than a preset completeness threshold, obtain the supplementary answer information corresponding to each question requirement information through an information completion calling program; A generation module is used to generate the user's target answer information based on the question answer content corresponding to each question requirement information and the supplementary answer information corresponding to each question requirement information.

7. The device according to claim 6, characterized in that The acquisition module is specifically used for: Extracting the question text content of the question information, and identifying the text semantic information corresponding to the question text content; Based on the text semantic information and the question text content, each user intention feature of the question information is analyzed through an intention recognition program, and each user intention feature is used as the user intention information corresponding to the question information.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images