Large model dialogue information collection method and system fusing business query

Abstract

The invention discloses a large model dialogue information collection method and system fused with business query. The method comprises the steps that initial input of a user is received, a session variable is initialized, a predefined cue word template is loaded and combined with a session state filling placeholder to generate a calling cue word, and a large model is called to extract a candidate structured variable; when it is judged that service query needs to be executed, multi-level fuzzy query is executed according to accurate, fuzzy and auxiliary matching priorities, a dialogue strategy is adjusted according to a result, variables are confirmed, and a clear list or a question is generated; and when the session variables are updated, converting the user anaphora and calculating the confidence of the active list to determine a target item, and generating response iteration until all target variables are extracted. Based on the construction of the system, the problems of insufficient fusion of information extraction and business verification, fuzzy reference and multi-list confusion in the prior art are solved, the variable extraction accuracy and the dialogue efficiency are improved, and the system is adaptive to complex multi-round business dialogue scenes.

Description

Technical Field

[0001] This invention relates to the technical field of electronic digital data processing, and more particularly to a method and system for collecting large-scale dialogue information that integrates business queries. Background Technology

[0002] Currently, with the continuous evolution of natural language processing technology, large language models (LLMs) have demonstrated superior understanding and generation capabilities in dialogue information extraction tasks, becoming a crucial technical support for scenarios such as intelligent customer service, business consultation, and automated form filling. Traditional rule-based or statistical model-based methods have limited generalization ability when faced with diverse and unstructured user expressions, while large models, with their powerful semantic understanding and contextual reasoning capabilities, can more flexibly identify and extract key entities, relationships, and instructions from dialogues, significantly improving the coverage of information collection and the naturalness of interaction. Existing technologies have already explored integrating large models with business systems, such as optimizing information collection processes through dynamic prompts or using large model agents to call APIs for automatic form filling.

[0003] While existing large-scale model-based dialogue information collection methods for integrating business queries have made progress in semantic understanding and process automation, they still have significant shortcomings, especially in complex multi-turn dialogue scenarios. Typically, large models are guided by prompts to extract variables from the dialogue and then introduced into business query interfaces for data validation or form filling. For example, some solutions use Retrieval Augmentation (RAG) technology to access external knowledge bases or use NL2SQL to convert natural language into database queries to enhance the business relevance of information. However, their core flaw lies in the insufficient real-time integration of the large model's information extraction process with the logic and data validation of the business system, resulting in a lack of business accuracy in the extraction results. Specifically, large models rely solely on semantic matching to extract variables but cannot verify the existence, uniqueness, or validity of the name in the business database in real time, easily leading to "illusory" or erroneous data.

[0004] Therefore, it is necessary to improve the existing methods and systems for collecting large-scale dialogue information for integrated business queries in order to solve the above problems. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a method and system for collecting large-scale model dialogue information that integrates business queries. It aims to solve the problems of ambiguous referencing and confusion of multiple list contexts caused by multiple rounds of dialogue when extracting large-scale model information in the prior art.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: a method for collecting large-scale model dialogue information that integrates business queries, comprising:

[0007] Receive initial input from the user and initialize session variables;

[0008] Load predefined prompt word templates and session variables, fill in placeholders based on the state of session variables, and generate prompt words for calling the large language model.

[0009] The large language model is invoked to parse user input and candidate structured variables are extracted based on prompt words;

[0010] Based on the candidate structured variable attributes, determine whether a business query needs to be executed;

[0011] When business queries are required, multi-level fuzzy queries are performed based on the relationships between candidate structured variables and matching fields;

[0012] Based on the execution results of multi-level fuzzy queries, the dialogue strategy is dynamically adjusted, including confirming variable values, generating a list of questions to be clarified, or using follow-up questions.

[0013] Based on the results of the dialogue strategy adjustment, the conversation variables are updated. When the user inputs a list referent, the active list confidence is calculated by transforming the referent, and the target list and items are determined.

[0014] Generate a system response and iteratively execute the above steps until all target variables are extracted.

[0015] In a preferred embodiment of the present invention, multi-level fuzzy query includes:

[0016] Query conditions are constructed in order of priority: exact match fields, fuzzy match fields, and auxiliary match fields, and the matching conditions are gradually relaxed.

[0017] Among them, exact match fields are subjected to full equality comparison;

[0018] If no exact match is found, the process proceeds to the second level of fuzzy matching, which compares strings using a similarity algorithm.

[0019] If fuzzy matching still yields no results or the number of results is insufficient, a third-level auxiliary matching is triggered, which combines the verified information of the associated variables to narrow down the query range.

[0020] Among them, the matching threshold is dynamically adjusted. When two consecutive queries have no results, the threshold is lowered and the matching range of auxiliary fields is expanded. The query results are deduplicated, sorted, and then stored in the current query result cache.

[0021] In a preferred embodiment of the present invention, calculating the confidence level of the active list includes:

[0022] The matching confidence score is calculated based on a weighted combination of one or more of the following strategies: nearest-to-nearest strategy, semantic association strategy, dialogue focus strategy, and dialogue turn strategy.

[0023] Among them, the nearest priority strategy calculates a time relevance score based on the difference between the generation timestamp of the list to be clarified and the current conversation timestamp;

[0024] The semantic association strategy performs semantic matching based on entity type clues identified in user input and preset content categories in the list to be clarified;

[0025] The dialogue focus strategy is evaluated by combining the relevance of the current dialogue topic with the categories of content in the clarification list;

[0026] The dialogue turn strategy calculates a relevance score based on the distance between the current dialogue turn and the dialogue turn in which the clarification list was generated or referenced.

[0027] In a preferred embodiment of the present invention, determining the target list and items by converting pronouns includes:

[0028] Extract ordinal or positional descriptive pronouns from user input;

[0029] Convert the extracted pronouns into uniform internal ordinal numbers or positional representations;

[0030] By combining a multi-clarification list context manager, the matching confidence of each active clarification list is calculated;

[0031] The target list is determined based on the matching confidence score, and the internal ordinal representation is matched with the ordinal identifier of the list item in the target list to lock the specific list item.

[0032] In a preferred embodiment of the present invention, generating large language model invocation prompts includes:

[0033] After converting the target variable list to JSON format, fill in the JSON list placeholders for the variables to be extracted;

[0034] After converting the contents of the collected variable mappings into JSON format, fill in the JSON list placeholders of the currently extracted variables;

[0035] Fill the latest user input text into the latest user input text placeholder;

[0036] Convert the summary of the most recent business query result in the current query result cache into JSON format and then fill in the corresponding placeholders.

[0037] In a preferred embodiment of the present invention, when generating prompts for calling a large language model, targeted processing instructions are dynamically added based on the session state:

[0038] When the number of follow-up questions about a variable approaches the predefined limit for the number of proactive follow-up questions, add a prompt to the prompt word indicating caution in asking follow-up questions;

[0039] When multiple active clarification lists exist in a multi-clarification list context manager, add an instruction to prioritize processing the most recently generated clarification list;

[0040] Based on the dependency graph maintained in the parent-child variable dependency relationship, supplement the explanation of the priority of variable extraction.

[0041] In a preferred embodiment of the present invention, updating session variables includes:

[0042] When the query result is a single exact match, the variable validation status is updated to validated and a unique identifier for the business system is stored.

[0043] When a list of clarifications is generated, the list is added to the collection of lists of clarifications and the context manager of the multiple clarification lists is updated.

[0044] When no matching record is found, update the follow-up question count and mark the variable status.

[0045] Dynamically maintain the dialogue history and current query result cache.

[0046] This invention provides a large-scale model dialogue information collection system that integrates business queries, comprising:

[0047] The user interface module receives and preprocesses the user's natural language input, and then sends the system response to the user.

[0048] The session management module creates session instances and manages session variables, including a list of target variables, a map of collected variables, a set of lists to be clarified, and a context manager for multiple clarification lists.

[0049] The prompt word generation module loads predefined templates, dynamically fills in placeholders with session variables, and adds targeted processing instructions to generate prompt words for calling the large language model.

[0050] The large language model interaction module integrates prompt words and user input, calls the large language model to extract candidate structured variables, and performs format validation and security filtering;

[0051] The business query decision module determines whether to execute a business query based on the attributes of candidate variables, constructs a queue of variables to be queried, and sorts them according to the priority of business logic.

[0052] The business query execution module performs multi-level fuzzy queries based on the queue of variables to be queried, dynamically adjusts the matching threshold, and handles the parent-child variable dependency relationship.

[0053] The dialogue strategy adjustment module generates a clarification list, follow-up questions, or confirmation instructions based on the query results, and parses pronouns in the user input to determine the target item.

[0054] The response generation module generates natural language responses based on the session state, including confirmation, clarification, follow-up questions, and process termination texts.

[0055] The session storage module stores session instances and state information, and supports state recovery after a session is interrupted.

[0056] This invention provides an electronic device, comprising:

[0057] At least one processor; and

[0058] A memory communicatively connected to the at least one processor; wherein,

[0059] The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the large-scale dialogue information collection method for fusion business queries as described in any one of claims 1-7.

[0060] This invention provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the large-scale dialogue information collection method for fusion business queries.

[0061] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0062] (1) This invention combines a multi-clarification list context manager with a referential resolution process. The multi-clarification list context manager dynamically records the lifecycle, content category, and generation sequence of each clarification list and assigns an incrementing ordinal identifier to each list item. The referential resolution process, on the other hand, transforms the ordinal or positional descriptions of the user input and calculates the confidence of the active list based on time decay, semantic association, dialogue focus, and dialogue turn strategies, thereby achieving accurate determination of the target list and items. The combination of the two enables the system to intelligently track and manage the context state of multiple lists in large-scale model dialogues without relying on end-to-end processing of large models, significantly improving the accuracy and efficiency of referential resolution and reducing repeated questioning and dialogue interruptions caused by unclear referentials.

[0063] (2) This invention achieves multi-level fuzzy query and business query decision module collaboration. Multi-level fuzzy query adopts a hierarchical query strategy, gradually relaxing the conditions according to the priority of exact matching, fuzzy matching and auxiliary matching, and dynamically adjusting the matching threshold. The business query decision module intelligently judges the necessity of the query based on the attributes of the candidate variables and constructs a priority queue. This significantly improves the success rate of variable verification and reduces query failures caused by non-standard data or fuzzy user input. Compared with the existing technology, which is prone to mismatch or no results due to fixed thresholds or lack of correlation checks, this invention optimizes the allocation of query resources through dynamic adjustment and priority management.

[0064] (3) This invention helps solve the problem of insufficient integration between information extraction from large models and business system logic by interacting with dynamic prompt words and large language models. The dynamic prompt word generation module dynamically fills in placeholders and adds targeted instructions such as cautious follow-up questions and priority processing instructions based on the conversation state, such as collected variables, number of follow-up questions, and clarification list status. The large language model interaction module uses this rich context to guide the large model to perform deep semantic analysis and variable extraction, reducing illusion or erroneous data. This invention enhances the guidance capability of prompt words through dynamism and state awareness.

[0065] (4) This invention ensures the consistency of dialogue state and the relevance of system responses through the cooperation of the session variable management and response generation modules. The session management module initializes and maintains various session variables, while the response generation module generates different types of responses based on the real-time state of these variables. This enables the system to track the progress of the dialogue and user intent throughout, maintaining the continuity and accuracy of multi-turn dialogues. Compared with the isolated state management of existing technologies, which leads to context loss or response mismatch, this invention achieves complete traceability of dialogue context through integrated variable management and state persistence. Attached Figure Description

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a flowchart of the stages of a dialogue collection system according to a preferred embodiment of the present invention;

[0068] Figure 2 This is a flowchart of the dialogue collection process (initial dialogue) of a preferred embodiment of the present invention.

[0069] Figure 3 This is a flowchart of a preferred embodiment of the present invention. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0072] Application Overview:

[0073] This application addresses the issues of ambiguous referencing and contextual confusion among multiple lists in existing large-scale model dialogue information collection methods for integrated business queries in complex multi-turn dialogue scenarios. While existing technologies integrate business query interfaces, their limitations are profoundly reflected in the following aspects: 1. When multiple clarification lists are generated sequentially or in parallel during a dialogue, the system cannot track the lifecycle, type, and time sequence of each list. This leads to confusion when users use ordinal references, making it difficult for the system to determine which list to refer to, resulting in referencing confusion. 2. Existing methods rely on the natural language understanding capabilities of the large-scale model to directly parse user input, but they lack a dedicated module to map ordinal or positional descriptions to specific items in the lists, and they lack semantic encoding of the list structure, resulting in low parsing accuracy. These limitations cause the system to frequently encounter errors in multi-turn dialogues, forcing repeated follow-up questions and severely impairing efficiency and user experience.

[0074] The fundamental reason why existing technologies fail to solve the above problems lies in the fragmentation at the architectural level and the lack of core functional modules: they do not deeply integrate the context state of multiple lists with the dialogue management mechanism, but instead treat business queries, list generation, and reference parsing as independent tasks, simply leaving them to large models to process through end-to-end prompts. They neither assign traceable semantic ordinal identifiers to each dynamic list item nor have a dedicated manager capable of globally managing the lifecycle, contextual relationships, and generation sequence of multiple lists. From a technical adaptability perspective, while large models excel at semantic understanding, their training data is mainly continuous text, lacking the ability to accurately process dynamic structured data and making it difficult to reliably parse the positional relationships of discrete list elements. Rule matching or simple state machines can manage lists, but require pre-setting all dialogue paths, failing to adapt to the flexibility and diversity of multi-turn dialogues. Attempting to compensate for these deficiencies solely through enhanced prompts leads to intermediate forgetting due to the surge in context length, diluting early list information and causing a sharp drop in reference parsing accuracy with each turn. Ultimately, this architectural flaw and technical incompatibility cause the system to be unable to dynamically focus and infer intent when faced with ambiguous references such as "the first one" or "the middle one" by users. Instead, it can only rely on repeated calls to large models or pre-set prompts for generalized parsing, which not only causes response delays but also inevitably leads to problems such as failed reference resolution, confusion between multiple lists, and delays in business validation.

[0075] This application constructs a multi-clarification list context manager, which records the lifecycle, content category, and generation sequence of each list, and assigns an incrementing ordinal identifier to each list item; it converts user pronouns into a unified internal representation; and it achieves dynamic focusing and accurate reference resolution of multiple lists without relying on end-to-end processing of a large model, thus solving the problems of reference ambiguity and context confusion.

[0076] Exemplary method:

[0077] like Figure 2 As shown, a method for collecting large-scale model dialogue information that integrates business queries includes the following steps:

[0078] S1. Receive the user's initial input and initialize session variables;

[0079] S2. Load the predefined prompt word template and session variables, fill in the placeholders according to the state of the session variables, and generate the large language model call prompt words;

[0080] S3. Call the large language model to parse user input and extract candidate structured variables based on prompt words;

[0081] S4. Based on the candidate structured variable attributes, determine whether a business query needs to be executed;

[0082] S5. When a business query is required, perform a multi-level fuzzy query based on the association relationship of candidate structured variables and the matching fields.

[0083] S6. Dynamically adjust the dialogue strategy based on the execution results of multi-level fuzzy queries:

[0084] If the query result is a single exact match, then directly confirm the variable value and update the session variable;

[0085] When the query result is multiple matching records, a list item ordinal identifier is assigned to each matching item, a list to be clarified is constructed and stored in the multi-clarification list context manager, and a clarification script containing the list is generated at the same time.

[0086] If the query result is no matching record, then a follow-up question will be generated;

[0087] S7. Update the conversation variables based on the results of the dialogue strategy adjustment;

[0088] When a user inputs a list referent, the referent is transformed, and combined with a multi-clarification list context manager, the active list confidence is calculated to determine the target list and items.

[0089] S8. Generate a system response and send it to the user. Iterate through steps S2-S7 until all target variables are extracted.

[0090] like Figure 1 As shown, in step S1, the system creates a new session instance and assigns a unique session identifier to the session; at the same time, it initializes a set of session variables used to manage the entire dialogue process.

[0091] Specifically, the initialization process includes:

[0092] When a user initiates a new interaction, the user interface module receives the user's initial natural language input, preprocesses and standardizes it;

[0093] A unique session identifier (session_id) is created for this dialogue session, and session initialization is performed; when the session is a brand new session, the system creates a new session instance and initializes all session variables to their default state.

[0094] Based on this, the system initializes the following session variables:

[0095] The target variable list is used to store a predefined set of information for all target variables that need to be collected, including variable name, data type and business relevance, and is initialized as an ordered list structure;

[0096] The variable mapping has been collected. Create an empty key-value pair mapping structure to record the variable key-value pairs that have been successfully extracted and verified and their verification status. Each variable contains a variable name, variable value, verification status and timestamp.

[0097] The set of clarification lists is initialized with an empty list container to store a list of multiple options that require clarification from the user. Each clarification list contains a list ID, generation time, associated variables, and option details.

[0098] The conversation history is created by establishing an empty double-ended queue structure to store all user and system interactions in chronological order, including the timestamp, speaker, and content text for each record.

[0099] Follow-up question count: Initialize an empty counter mapping to record the number of times each variable has been actively followed up, with an initial value of 0 for each variable;

[0100] The multi-clarification list context manager is an empty manager object that manages the context information of all active clarification lists, including the generation sequence, content category, and activity status of each list.

[0101] The current query result cache is set to null and is used to store the original result of the most recent business query.

[0102] Session status flags are used to identify the current session status, including: in progress, completed, and terminated.

[0103] Dynamic prompt word status: Initializes the status information required for prompt word generation, used to record the current prompt word generation status and version information;

[0104] Parent-child variable dependency relationships are used to maintain a dependency graph between variables.

[0105] Timestamp information is used to record the time of the operation;

[0106] User identity identifiers are used to store user authentication and identification information.

[0107] After all session variables are initialized, the system persists the session instance and its initial state to the session storage to ensure state consistency in subsequent dialogue flows. The initialization process lays a complete initial state foundation for subsequent information extraction, business queries, and multi-turn dialogue management.

[0108] In step S2, the system loads the prompt word templates and session variables configured in the predefined phase.

[0109] The prompt word template includes:

[0110] Character setting clarifies the identity and stance of the large language model in the dialogue;

[0111] Skill requirements: Describe the abilities needed to build a large language model;

[0112] The thought chain construction guides the large language model to carry out reasoning in logical steps, including understanding user intent, identifying variables, judging business query needs, processing query results and deciding on subsequent operations;

[0113] Constraints on the response of large models constrain the generated content and behavior of large language models;

[0114] Placeholder replacement rules define the placeholders that can be dynamically filled in the prompt and their corresponding filling logic.

[0115] The system dynamically fills in the placeholders in the prompt word template based on the current state of the session variables.

[0116] Dynamic filling follows placeholder replacement rules, including two:

[0117] After converting the target variable list to JSON format, fill in the JSON list placeholders for the variables to be extracted;

[0118] After converting the contents of the collected variable mappings into JSON format, fill in the JSON list placeholders for the currently extracted variables;

[0119] Fill the latest user input text into the latest user input text placeholder;

[0120] After converting the summary of the most recent business query result in the current query result cache into JSON format, fill in the placeholder of the summary of the most recent business query result to ensure that the large language model can obtain the complete dialogue context and business state.

[0121] At the same time, the system dynamically adds targeted processing instructions based on key data in the session state:

[0122] When the number of follow-up questions for a certain variable in the follow-up question count is close to the predefined limit for the number of active follow-up questions for that variable, a prompt word of caution for follow-up questions is added to guide the large language model to obtain information with more accurate expression;

[0123] When multiple active clarification lists exist in a multi-clarification list context manager, add an instruction to prioritize the most recently generated clarification list, thus clarifying the response priority of the large language model;

[0124] Based on the dependency graph maintained in the parent-child variable dependency relationship, supplement the priority description of variable extraction to guide the large language model to first extract core variables without parent variables, and then extract child variables whose parent variables have been verified.

[0125] The generated prompts integrate complete contextual information and clear processing guidelines, providing a structured understanding and generation framework for subsequent large language models. The large language model itself possesses powerful context processing capabilities and domain knowledge, enabling it to strictly adhere to the prompt requirements for information extraction and to reasonably follow up on missing information. Prompts are a key element in realizing dialogue collection within the large language model. After prompt generation, the system updates the dynamic prompt status in the session variables, recording the version hash value and generation timestamp of the prompt, ensuring that the prompt status is traceable and manageable.

[0126] In step S3, a pre-trained large language model is invoked to perform in-depth analysis of the user's input, and candidate structured variables are identified and extracted from the user's input based on the generated prompt words.

[0127] Specifically, the dynamic prompts generated in step S2 are integrated with the latest user input text to form complete model input data. During the integration process, it is ensured that the contextual information in the prompts is consistent with the conversation state, including historical data in the collected variable mapping, business feedback in the current query result cache, and interaction context in the dialogue history, so as to provide a comprehensive decision basis for the large language model.

[0128] The integrated input data is sent to a remotely deployed large language model service for processing. The large language model, as the core processing engine, performs in-depth semantic analysis and intent recognition on user input based on its powerful natural language understanding capabilities and the clear role settings, skill requirements, thought chain construction guidance and constraints in the prompt words.

[0129] The specific content includes:

[0130] Perform deep semantic analysis on user input and combine it with predefined information in the target variable list to identify potential variable candidates that are relevant to the business in the text;

[0131] By comparing with the collected variable mappings, duplicate or validated variables are eliminated, and the focus is finally on the new variables to be extracted.

[0132] At the same time, the model will pay special attention to targeted instructions dynamically added in the prompts, including: requests for careful follow-up on specific variables or instructions for prioritizing multiple clarification lists.

[0133] When extracting candidate structured variables, the large language model follows the constraints in the prompt words to ensure that only variable information that conforms to the predefined structure is output;

[0134] Candidate structured variables refer to variable information that conforms to a predefined structure and is initially identified by the model from user input. These variables have not yet been verified by the backend business system and may be fictitious, mismatched, or ambiguous. Their validity needs to be confirmed through subsequent business data verification steps.

[0135] The extraction results are presented in JavaScript object abbreviation format. Each candidate variable contains three core elements: variable name, variable value, and confidence score. The variable name corresponds to a predefined identifier in the target variable list, the variable value is the specific content parsed from the user input, and the confidence score reflects the model's certainty assessment of the extraction result. The value ranges from 0 to 1, with the closer to 1 indicating higher reliability.

[0136] After receiving the results returned by the model, the system performs a dual verification mechanism: the format verification stage ensures that the returned JavaScript object's simplified spectrum structure is complete, variable names match the predefined list, and data types conform to the target variable's data type requirements; the security filtering stage cleans the variable values, removing potentially malicious characters or incorrectly formatted content to prevent abnormal data from entering subsequent processes. For candidate variables with a confidence level below a preset threshold, the system marks them as pending verification and does not add them to the collected variable mapping. Instead, they are further confirmed through multi-level fuzzy queries in subsequent business query steps.

[0137] After verification, the system will pass the valid candidate structured variables to the next step, and at the same time update the current query result cache in the session variables, storing the original results extracted this time and the processing status.

[0138] In step S4, the system uses a multi-dimensional evaluation mechanism based on the attribute characteristics of the candidate structured variables to determine whether it is necessary to interact with the backend business system to verify the validity of the variables.

[0139] Candidate structured variable attributes refer to a series of characteristic descriptions assigned to each variable to be extracted during the pre-definition stage, including: whether it needs to be associated with business data validation, data type, and parent-child variable dependency relationship. These attributes determine how the variable is processed in the information collection process.

[0140] The business query necessity assessment process combines these attributes with the current session state to determine whether a business query needs to be initiated through layer-by-layer verification.

[0141] First, the system performs a correlation check on each candidate structured variable extracted in step S3; each variable is marked as related to business queries during the predefined stage. This mark is a Boolean value, which indicates whether the value of the variable must be verified by real-time data from the backend business system.

[0142] For example, core business variables such as customer name and product model are marked as requiring association with business queries, while auxiliary information such as contact information is configured as not requiring association depending on the scenario; the system traverses all candidate variables, filters out variables marked as requiring association with business queries, and initially determines potential business query objects.

[0143] Secondly, the system performs an integrity check. The collected variable mapping stores all variable information that has been successfully extracted and verified by the business. The system compares the current candidate variable with the collected variable mapping.

[0144] If the name and value of a candidate variable already exist in the collected variable map and the verification status is verified, then it is determined that the variable does not need to be queried again and the subsequent evaluation is skipped directly.

[0145] If a variable exists but its verification status is pending or invalid, it still needs to be included in the business query scope to update the verification status.

[0146] Next, the system performs a parent-child variable dependency check; some variables have predefined parent-child dependency relationships, meaning that the validity of the child variable depends on the prior verification of the parent variable.

[0147] The parent-child variable dependency relationship is described by a predefined parent variable attribute. For example, the product model is a child variable, and its corresponding parent variable may be the order number. The business query of the product model is only meaningful after the order number is successfully verified.

[0148] The system checks whether each candidate variable has a parent variable. If it does, it further verifies whether the parent variable is already in the collected variable mapping and whether its status is verified. If the parent variable is not verified or does not exist, the business query request for the child variable is temporarily suspended until the parent variable is verified and then re-evaluated.

[0149] After completing the above checks, the system constructs a variable queue for pending business queries and sorts the variables in the queue by priority.

[0150] The sorting rules follow the order of business logic:

[0151] First, handle independent variables that do not depend on parent variables to ensure that core business entities are verified first;

[0152] Next, process the child variables of the parent variable that have been validated to ensure the integrity of the dependency relationship;

[0153] Finally, based on the importance of variables in the business process, prioritize handling variables marked as necessary for the process.

[0154] For example, in an order query scenario, the order number is processed first as an independent and necessary variable. After it is verified, sub-variables such as the product model are then entered into the query queue.

[0155] After priority sorting is completed, the system passes the queue of variables to be queried to the next step, updates the current query status in the session variables, and records the name, quantity, and priority order of the variables to be queried, providing a clear processing sequence for the execution of subsequent multi-level fuzzy queries.

[0156] This evaluation process ensures that business queries are initiated only for necessary variables, avoiding invalid requests, while dependency management and priority sorting guarantee the consistency of business logic and query efficiency.

[0157] like Figure 3 As shown, in step S5, when the system determines that a business query needs to be executed based on the evaluation in step S4, it will start a multi-level fuzzy query mechanism based on the association relationship of candidate structured variables and the preset matching fields to achieve accurate data interaction with the backend business system.

[0158] Multi-level fuzzy search is a hierarchical query strategy that gradually relaxes matching conditions. By applying matching rules of different precision in sequence, it can improve the recall rate while ensuring the accuracy of the query and avoid query failures caused by fuzzy user input or non-standard data.

[0159] The association between candidate structured variables refers to the predefined association between variables based on business logic. For example, predefined associations include the mapping relationship between customer name and customer unique identifier, and the subordinate relationship between order number and product model. These relationships are pre-configured through the business association attributes in the target variable list to guide field association and data aggregation during queries.

[0160] Matching fields refer to the specific data items used for data retrieval in the backend business system, including: exact matching fields, fuzzy matching fields, and auxiliary matching fields. Their priority and matching rules are preset according to the business scenario during the system initialization phase.

[0161] The system first extracts the set of matching fields associated with the first variable in the queue of variables to be queried, that is, the variable sorted according to the priority determined in step S4.

[0162] Each candidate structured variable is associated with a set of matching fields during the pre-definition phase; the system constructs query conditions in descending order of field priority:

[0163] The first-level query prioritizes exact matching fields, comparing the candidate variable values ​​with the corresponding fields in the business system for full equality.

[0164] If no exact match is found, the process proceeds to the second level of fuzzy matching. The fuzzy matching fields are compared using a string similarity algorithm to filter out records with similarity scores higher than the threshold.

[0165] In a preferred embodiment, the user input value is processed by Chinese word segmentation. The word segmentation process adopts a word segmentation algorithm based on BPE or WordPiece, and combines a predefined industry dictionary and statistical model to decompose the user input into multiple semantic units.

[0166] The core keywords from these segmented results are used as query parameters to construct multiple or combined fuzzy query requests. Fuzzy segmentation queries aim to address situations where user input contains imprecise descriptions, partial matches, or word order adjustments. Query results are initially scored based on matching degree, relevance, and keyword coverage. If several highly relevant records are returned, they are considered candidate results and passed to the dialogue strategy adjustment unit.

[0167] If fuzzy matching still yields no results or the number of results is insufficient, a third-level auxiliary matching is triggered. This involves combining the verified information of the associated variables to narrow down the query range and further filter potential matches.

[0168] In a preferred embodiment, if the word segmentation fuzzy query still fails to effectively resolve ambiguity or obtain matching results, the user input value is converted to pinyin, including full pinyin and the first letter of pinyin.

[0169] Using the converted Pinyin, a Pinyin fuzzy query request is constructed, primarily to address issues such as typos, homophones, or abbreviations in user input. The Pinyin fuzzy query combines an edit distance algorithm or an N-gram model to evaluate the similarity of Pinyin matches and returns the top few business records with the highest matching scores.

[0170] When processing multivariate queries with related relationships, the system follows the logic of first master and then follow in relational queries.

[0171] When the queue of variables to be queried contains variables with parent-child dependencies, the system will first perform a multi-level fuzzy query on the parent variable. After the parent variable is verified by the business system and a unique identifier is obtained, the identifier will be passed as an association condition to the query request of the child variable.

[0172] If a parent variable query does not yield a unique result, the child variable query will be paused until the parent variable is clarified through the clarification process before it is restarted.

[0173] For scenarios where there are multiple candidate values ​​for the same variable or multiple variables are queried in parallel, the system adopts a processing method that combines batch aggregation with step-by-step verification.

[0174] When the candidate structured variables extracted in step S3 contain multiple variables to be verified, and these variables have no direct dependency relationship, the system will construct a batch query request, send the query conditions of multiple variables to the business system at one time, and aggregate the related query results through the business relevance attributes of the variables.

[0175] After the batch query results are returned, the system will verify the variables in order of priority, processing independent variables first and then related variables to ensure efficient use of query resources.

[0176] During multi-level fuzzy search, the system monitors the search status in real time and dynamically adjusts the matching threshold.

[0177] In the initial stage, the thresholds for exact matching and fuzzy matching are set relatively high to ensure the accuracy of the results;

[0178] If two consecutive queries return no results, the system will automatically lower the threshold and expand the matching range of auxiliary fields, while recording the threshold adjustment log in the current query result cache.

[0179] Once the query yields matching results, the system will deduplicate and prioritize the results. The ranking criteria include the precision of the matching fields, the similarity score, and the completeness of the associated variables.

[0180] Finally, the system stores the sorted query result set in the current query result cache of the session variable and marks the query completion status, providing a data basis for the dialogue strategy adjustment in step S6.

[0181] In step S6, the system, based on the characteristics of the results returned by the multi-level fuzzy query, combined with the preset business rules and the current session state, activates the dynamic dialogue strategy adjustment mechanism. Through differentiated processing logic, it achieves accurate response to the variable verification results, while ensuring the continuity of the dialogue process and the smoothness of the user experience.

[0182] When the query result is a single exact match, the system executes the variable confirmation and status update process.

[0183] Specifically, a single exact match refers to a result set returned by a multi-level fuzzy query that contains only one record, and the similarity between the core matching field of the record and the candidate structured variable value reaches the preset exact match threshold. At the same time, the record has a unique business system identifier and can be directly associated with the real entity data in the backend business system.

[0184] At this point, the system first binds the unique identifier of the business system corresponding to the matching record with the candidate variable value to confirm the validity and uniqueness of the variable value; then it updates the collected variable mapping, which is an empty key-value pair mapping structure that is created to record the variable key-value pairs that have been successfully extracted and verified and their verification status. Each variable contains a variable name, variable value, verification status and timestamp. The system stores the variable name, the variable value bound to the unique identifier of the business system, the verification status marked as verified and the current timestamp in the mapping structure.

[0185] Based on this, the system checks the parent-child variable dependency relationship. The parent-child variable dependency relationship is a session variable used to maintain the dependency graph between variables. The system views the child variables associated with this variable.

[0186] If there are unextracted or unverified sub-variables, and the parent variable dependency condition of the sub-variable has been met, the extraction priority of the target variable list is updated. The target variable list is a predefined set of information used to store all target variables that need to be collected, including variable name, data type and business relevance. It is initialized as an ordered list structure, and the system puts the relevant sub-variables first to provide guidance for the subsequent information collection process.

[0187] Simultaneously, the system updates the current query result cache, which is a session variable set to null and used to store the original result of the most recent business query. The system records the details and verification status of the exact match result and adds a variable confirmation log to the dialogue history. The dialogue history is an empty doubly queue structure that stores all user and system interactions in chronological order, including the timestamp, speaker role, and content text for each record. The added log includes the confirmed variable name, the unique identifier of the business system, and the operation time.

[0188] When the requirements of some process nodes are met due to the confirmation of this variable, the session status flag will be updated synchronously to the in-progress status of the corresponding stage. The session status flag is a variable used to identify the current session status, including in progress, completed, and terminated. The updated status provides an accurate status basis for the generation of prompts for the next round of dialogue.

[0189] When the query result contains multiple matching records, the system initiates a multi-option clarification process.

[0190] Specifically, multiple matching records refer to the result set returned by a multi-level fuzzy query containing two or more records that meet the matching conditions. Although these records are all related to the candidate structured variables, it is impossible to determine the unique target based on the existing information, and further clarification is required from the user.

[0191] The system standardizes multiple matching records, extracting a unique business system identifier, a display name for user identification, and auxiliary information to help distinguish between them from each record. The auxiliary information includes the entity's associated attributes; for example, the customer's city, product release time, and order creation date.

[0192] The system sorts these records according to preset list presentation rules, which can be configured according to business scenarios. Examples include sorting in descending order by similarity score of matching fields, sorting in reverse order by the most recent update time of the records, or sorting in alphabetical order by entity name, to ensure the logical presentation of the list and the ease of user identification.

[0193] After sorting, the system assigns a list item ordinal identifier to each matching item. This identifier is an incrementing integer starting from 1, and each identifier corresponds to a unique matching item, which is used for accurate parsing when the user refers to it later.

[0194] The system constructs a list to be clarified, which is a structured data collection containing multiple list item objects. Each list item object integrates a unique identifier of the business system, a display name, auxiliary information, and a list item ordinal identifier.

[0195] After the construction is completed, the system stores the list to be clarified into the multi-clarification list context manager. The multi-clarification list context manager is an empty initialized manager object used to manage the context information of all active clarification lists, including the generation sequence, content category and activity status of each list.

[0196] The stored information includes: a unique session-level list ID assigned to the list to be clarified, the generation timestamp, the corresponding associated variable name, the content category, list item details, and the current state initialized as active.

[0197] At the same time, the system updates the set of the list to be clarified in the session variables, adds the newly generated list to be clarified to the set, updates the current query result cache, and records the association between the original data of multiple matching records and the list to be clarified.

[0198] In the process of generating clarification scripts, the system combines the content of the list to be clarified with the dialogue history to construct natural and clear guiding text. The scripts must clearly inform the user that there are multiple matching results, and present the display name and auxiliary information of each list item in order of its ordinal number. Finally, the system guides the user to make a selection by ordinal number or position description with a clear question. The generated clarification scripts will be stored in the dialogue history and marked as the interaction content of the system role.

[0199] When the user enters the information again and the system recognizes that the input contains ordinal or positional descriptive pronouns for the clarification list, the system executes the pronoun parsing process.

[0200] S61. Extract ordinal or positional descriptive pronouns from user input;

[0201] Examples include "the first one", "the top one", "the second to last", "the third customer", etc.

[0202] Extraction methods include: rule-based pattern matching, deep learning-based sequence labeling models, or directly utilizing the contextual understanding capabilities of large language models;

[0203] S62. Convert the extracted ordinal or positional descriptive pronouns into a unified internal ordinal or positional representation; the conversion method is implemented through a predefined mapping table or language model parsing;

[0204] Furthermore, the conversion steps are as follows:

[0205] When the pronoun is a specific ordinal number, the ordinal number of the list item is directly converted to the corresponding integer;

[0206] For example, “first”, “second”, “third”, and “4” can be transformed into: 1, 2, 3, and 4.

[0207] When the pronoun is a positional descriptor, it is converted into an integer offset relative to the first or last position of the list, depending on the displayed length of the list to be clarified or the actual storage order.

[0208] For example, "top", "middle", "bottom", "top", "bottom"; "top" is converted to the number 1, and "bottom" is converted to the total length of the list. For "middle", it might be the result of rounding up the list length and dividing by 2;

[0209] When the pronoun is a relative ordinal number, it is converted to an integer ordinal number relative to the last element of the list, based on the display length of the list to be clarified.

[0210] For example, the second to last is converted to the list length minus 1, and the last is converted to the list length;

[0211] List item ordinal numbers can also handle compound pronouns, for example, "the third customer" or "that one from Shanghai," where "customer" or "Shanghai" is passed as a type clue to the multi-clarification list context manager to assist in subsequent list focusing.

[0212] S63. Simultaneously, retrieve all currently active clarification lists from the multi-clarification list context manager. An active list refers to a list that has not yet been explicitly selected by the user or determined by the system to be resolved. For each active clarification list, calculate its matching confidence score with the current user's intended meaning.

[0213] Match confidence is calculated based on a weighted combination of one or more of the following strategies:

[0214] The nearest-to-first strategy assigns a time relevance score to each list item based on the difference between its generation timestamp and the current conversation timestamp; lists generated closer to the current conversation time receive higher scores. Specifically, this is calculated using an exponential decay function.

[0215] The semantic association strategy identifies entity type clues related to pronouns in user input and then semantically matches these clues with preset content categories in each list to be clarified. Higher matching scores result in higher scores.

[0216] Semantic matching can be achieved by comparing predefined ontology dictionaries, thesauruses, or by using large language models to calculate vector similarity. Specifically, type clues and list categories are embedded into the same vector space, and cosine similarity is calculated.

[0217] The dialogue focus strategy combines the overall theme of the current dialogue with the most recently discussed topics or recently filled slot information in the historical dialogue rounds to infer the dialogue focus of the current user's intent.

[0218] The focus of the conversation is evaluated in relation to the content categories of each item on the list to be clarified; the stronger the relevance, the higher the score.

[0219] Relevance assessment can be achieved through dialogue topic models, including LDA, BERT topic models, or contextual reasoning capabilities based on large language models;

[0220] The dialogue turn strategy assigns a dialogue turn relevance score to the current dialogue turn based on the distance between the list of items to be clarified and the dialogue turn in which it was generated or most recently referenced; the closer the list is, the higher its score.

[0221] The results of the above strategies are weighted and summed to obtain the overall matching confidence score;

[0222] Furthermore, as a preferred embodiment of the present invention, the weight configuration of the matching confidence calculation strategy is dynamically adjustable; the dynamic adjustment is achieved through a reinforcement learning-based policy optimizer.

[0223] The strategy optimizer iteratively optimizes the weights based on the success rate of referential resolution in historical dialogues and user feedback, in order to maximize the accuracy of referential resolution and user satisfaction.

[0224] The reinforcement learning model treats the dialogue context and referential type as the state, the weight adjustment of different strategies as the action, and successful resolution as the reward signal.

[0225] Through continuous online or offline training, the model can learn the optimal weight combination for different dialogue scenarios.

[0226] S64. Based on the comprehensive matching confidence score, the multi-clarification list context manager determines the specific list to be clarified that the user's referential intent points to.

[0227] If the overall match confidence score of a list to be clarified is significantly higher than that of other lists and the score itself also exceeds a preset absolute threshold, then that list is identified as the target list.

[0228] When multiple lists of clarification have similar overall matching confidence scores that all exceed a preset threshold, indicating ambiguity, a disambiguation question is generated that includes multiple clarification lists, requesting the user to specify the object being referred to. The generated disambiguation question includes a brief description of the ambiguous list and its respective number.

[0229] When the overall matching confidence score of all the lists to be clarified is lower than the preset threshold, it indicates that the referent cannot be effectively parsed, and a general clarification follow-up question is generated to request the user to specify the referent.

[0230] S65. After determining the target list, the unified internal ordinal or positional representation is matched with the ordinal identifier of the list item in the target list to accurately locate the specific list item. The matching result includes: the unique identifier of the business system and the display name.

[0231] If a single list item cannot be identified, generate clarifying follow-up questions targeting the target list, requesting the user to make more specific choices. These follow-up questions include further descriptions of the ambiguous items in the target list.

[0232] The system stores the unique business system identifier of the selected list item in the current dialogue state, marks the corresponding list to be clarified as resolved or selected, updates its status in the multi-clarification list context manager, and generates a reply confirming the selection or guiding subsequent information.

[0233] When the query result is no matching record, the system executes the information supplementation and follow-up question process;

[0234] No matching record means that after a multi-level fuzzy query has tried to match at all levels, no record that meets the minimum matching threshold has been found. This indicates that the candidate structured variable values ​​may be incorrect, incomplete, or unrelated to the data in the backend business system.

[0235] The system first retrieves the follow-up question count statistics from the session variables. The follow-up question count statistics are an initialized empty counter mapping used to record the number of times each variable has been actively followed up. The initial value is 0. The system finds the counter corresponding to the current candidate variable and checks whether its follow-up question count has reached the predefined active follow-up question count limit. This limit is the maximum number of follow-up questions set according to the business scenario, usually 2 to 3 times, to avoid excessive follow-up questions that lead to a decline in user experience.

[0236] If the number of follow-up inquiries has not reached the limit, the system will generate a follow-up question script. The script must combine the historical information in the collected variable mapping and the dialogue history to clearly inform the user that no matching results were found and guide the user to provide more accurate or complete information. For example, if no customer named Li Siqiang was found, please confirm whether the name is correct or provide more identity-related information.

[0237] If the number of follow-up questions has approached the limit, the wording will be further optimized to a more cautious approach.

[0238] For example, "No relevant information found. This is the last confirmation. Please provide the accurate customer name."

[0239] While generating the dialogue, the system increments the follow-up question counter for the candidate variable by 1, updates the follow-up question count statistics, and stores the follow-up question dialogue in the dialogue history.

[0240] If the limit for the number of follow-up questions has been reached, the system will generate a message to terminate the follow-up questioning, informing the user that no matching results can be obtained with the existing information, and providing guidance on subsequent solutions.

[0241] For example, users are advised to check their information and resubmit the query or be transferred to a human agent. For instance, "If you still cannot find a match after multiple attempts, you can check your information and consult again, or contact customer service for assistance."

[0242] At this point, the system will mark the candidate variable as invalid, store it in the collected variable mapping, and update the session status flag. If the variable is a necessary variable for the process, the status will be marked as pending manual processing; otherwise, the status will remain in progress. Simultaneously, the system will update the current query result cache, recording the status information of no matching results and terminated follow-up questions, providing a basis for decision-making in subsequent dialogue processes.

[0243] In step S7, the system updates the conversation variables in a targeted manner based on the dialogue strategy results dynamically adjusted in S6, ensuring that the conversation state is synchronized with the dialogue flow. At the same time, when the user inputs a list of pronouns, the system completes the state synchronization and variable maintenance after the pronoun parsing.

[0244] When the query result in S6 is a single exact match, the system executes the session variable update process.

[0245] When a multi-level fuzzy query returns a single exact match result, the system first updates the collected variable mapping. Specifically, it binds the unique identifier of the matching item's business system to the variable value, marks the verification status as "verified", records the current timestamp, and stores it in the collected variable mapping.

[0246] At the same time, the system checks the parent-child variable dependency relationship. If there is a child variable that depends on the current variable and the child variable has not been extracted, the extraction priority of the target variable list is updated, and the child variable is moved to the front.

[0247] In addition, the system updates the session status flags: if all target variables have been validated, it is marked as "Completed"; otherwise, it remains "In Progress". The current query result cache synchronously records the "Single Match Successful" status and match details.

[0248] When the query results in multiple matching records and a list of records to be clarified is generated, the system updates the list of records to be clarified, adds the newly generated list of records to the set, and assigns a unique list ID to it.

[0249] Meanwhile, the multi-clarification list context manager records the list's generation timestamp, content category, associated variables, and activity status.

[0250] The current query results cache stores the original data of multiple matching records and their association with the list of records to be clarified. The dialogue history adds system-generated clarification messages and interaction content marked as system roles.

[0251] If the query results have no match and the limit for follow-up questions has not been reached, the system updates the follow-up question count statistics, increments the current variable's follow-up question count by 1, and updates the current query result cache to the "no match - follow-up question pending" state.

[0252] If the limit for the number of follow-up inquiries has been reached, the system will mark the variable as "invalid" and store it in the collected variable mapping, and update the session status flag: if the variable is a necessary variable for the process, it will be marked as "pending manual processing"; otherwise, it will remain "in progress".

[0253] Add a phrase to terminate follow-up questions in the conversation history and synchronize it to the current query result cache.

[0254] When a user inputs ordinal or positional descriptive pronouns for a clarification list, the system determines the target list and specific items through pronoun parsing, dynamic focusing, and confidence calculation, and updates session variables.

[0255] Regardless of the strategy adjustment result, the system will update the dynamic prompt word status. The dynamic prompt word status is a variable used to record the current prompt word generation status and version information. The system will record the prompt word generation version requirements and status update time after this session variable update, ensuring that prompt word generation in the next round of S2 can be based on the latest session status. All updated content of session variables will be persisted to session storage, ensuring the consistency and traceability of the status in subsequent dialogue rounds, and providing an accurate status foundation for information extraction, business query, and multi-turn dialogue management.

[0256] In step S8, the system generates and sends a natural language response that matches the dialogue scenario and user needs based on the updated session state in S7. At the same time, it judges the current information collection progress to decide whether to start the next round of dialogue process until all target variables are extracted.

[0257] The system first obtains the core session variable data after the S7 update, including the collected variable mapping, the set of lists to be clarified, the session state markers, the dialogue history, and the current query result cache. These data together form the basis for response generation.

[0258] The collected variable map is a key-value pair structure that records verified variables, the clarification list set is a container that stores a list of multiple options, and the session state flag indicates the current stage of the session. The combination of these three elements can clarify the core information that needs to be conveyed in the current dialogue.

[0259] The system constructs targeted response text based on different dialogue scenarios:

[0260] When S7 updates a variable due to a single exact match, the response is a confirmation text. It should clearly inform the user that the target variable has been verified, and guide subsequent information collection based on the extraction priority of the target variable list.

[0261] When S7 adds a list of items to be clarified due to multiple matching records, the response is a clarification text. It is necessary to integrate the ordinal identifier, display name and auxiliary information of the list items in the list to be clarified, present them in order and guide the user to select by ordinal number or position description.

[0262] When S7 executes the follow-up question process due to no matching record, the response is a follow-up question text. The tone of the expression needs to be adjusted according to the status of the follow-up question count. If the limit has not been reached, guide the user to supplement information. If the limit is approaching, prompt for final confirmation.

[0263] Once S7 completes the reference parsing and updates the status, the response should be a reference confirmation text. This text should clearly inform the user of the selected list item and guide the next step of the operation in conjunction with the variables to be extracted.

[0264] When the collected variable mapping contains all variables of the target variable list and all of them are verified, the response should be a process end text, which should inform the user that the information collection is complete and the subsequent business processing actions.

[0265] When the session status is marked as pending manual processing or terminated, the response should be an exception handling text that explains the current status and provides guidance on solutions.

[0266] Once constructed, the system will store the response text in the dialogue history, mark it as the interaction content of the system role, and attach the current timestamp to ensure that the dialogue context is traceable.

[0267] The system then sends the generated response to the user through the user interface module. The user interface module is responsible for realizing the interaction between the system and the user, supporting the adaptation and conversion of various input and output formats such as text and voice, and ensuring that the response can be accurately delivered to the user.

[0268] After the response is sent, the system executes an iterative judgment process:

[0269] First, check whether all variables in the target variable list already exist in the collected variable mapping and whether their verification status is verified.

[0270] If this condition is met, the system will update the session status to "completed", generate a process termination response, end the dialogue, and persist the session instance and its final state to session storage to provide complete data support for subsequent business processing.

[0271] If the condition is not met, it is necessary to further determine whether the session status flag is terminated or pending manual processing: if yes, generate the corresponding termination response and end the process; if no, return to S2, reload the prompt word template and the current session variables, start the next round of information collection loop, until all target variables are extracted.

[0272] Exemplary system:

[0273] A large-scale model dialogue information collection system integrating business queries includes: a user interface module, a session management module, a prompt word generation module, a large language model interaction module, a business query decision module, a business query execution module, a dialogue strategy adjustment module, a response generation module, and a session storage module.

[0274] The user interface module is used to realize the interaction between the system and the user. It receives the user's natural language input and performs preprocessing and standardization, including text cleaning, format conversion and multimodal input adaptation. At the same time, it transmits the system-generated response text or voice output to the user, supports the conversion and adaptation of various interaction forms, and ensures that the user input can be accurately entered into the system processing flow and the system response can be clearly conveyed to the user.

[0275] The session management module is responsible for the creation, maintenance, and state management of sessions. It generates a unique session identifier (session_id) for each new interaction, initializes and manages all session variables, including the target variable list, collected variable mappings, a set of clarification lists, dialogue history, follow-up question count statistics, a multi-clarification list context manager, current query result cache, session state markers, dynamic prompt word status, parent-child variable dependencies, timestamp information, and user identification. It also updates the state of each session variable in real time to ensure data consistency and traceability during the session and persists the session instance and state to the session storage module.

[0276] The prompt generation module loads a predefined prompt template, which includes role settings, skill requirements, thought chain construction guidelines, constraints for large model responses, and placeholder replacement rules. Based on the current session variable status provided by the session management module, it dynamically populates the placeholders in the template, including converting the target variable list, collected variable mappings, the latest user input text, and a summary of the most recent business query results into JSON format and filling them into the corresponding positions. It also dynamically adds targeted processing instructions based on key data in the session status, such as instructions for cautious follow-up questions, instructions for prioritizing clarification lists, and explanations of variable extraction priority. Finally, it generates structured large language model call prompts and updates the dynamic prompt status in the session variables.

[0277] The large language model interaction module integrates the dynamic prompts generated by the prompt generation module with the user's latest input text into complete model input data, and sends it to the pre-trained large language model service; it receives the processing results returned by the large language model, extracts candidate structured variables from them, and the candidate structured variables include variable names, variable values, and confidence scores, presented in JavaScript object solfège format; it performs double validation on the extracted results, including format validation to ensure structural integrity and data type matching, security filtering to remove malicious characters or erroneous content, passing the valid candidate structured variables to subsequent modules, and updating the current query result cache.

[0278] The business query decision module evaluates the candidate structured variable attributes output by the large language model interaction module from multiple dimensions to determine whether a business query needs to be executed. It performs a correlation check to filter out variables marked as needing to be associated with business queries; it performs a completeness check to exclude verified duplicate variables; it performs a parent-child variable dependency check to suspend unverified child variables of parent variables; it constructs a variable queue for business queries to be executed and sorts them according to business logic priority, passes the sorted queue to the business query execution module, and updates the current query status.

[0279] The business query execution module performs multi-level fuzzy queries based on the queue of variables to be queried provided by the business query decision module, combined with the relationships between candidate structured variables and preset matching fields. It constructs query conditions in order of priority—exact matching fields, fuzzy matching fields, and auxiliary matching fields—gradually relaxing the matching conditions. When handling variables with parent-child dependencies, it follows a master-to-child relationship query logic. For multiple variables without dependencies, it uses batch aggregation combined with step-by-step verification. It dynamically adjusts the matching threshold, and stores the deduplicated and sorted query results in the current query result cache, providing a data foundation for adjusting the dialogue strategy.

[0280] The dialogue strategy adjustment module dynamically adjusts the dialogue strategy based on the characteristics of the query results returned by the business query execution module. When the result is a single exact match, it confirms the variable value and triggers a session variable update. When the result is multiple matching records, it assigns a list item ordinal identifier to each matching item, constructs a list to be clarified and stores it in the multi-clarification list context manager, and generates clarification scripts. When the result is no matching record, it generates corresponding follow-up question scripts or terminates follow-up question scripts based on the number of follow-up questions. It processes list pronouns in user input, determines the target list and items by converting pronouns and calculating the confidence of active lists, and completes the reference parsing and status update.

[0281] The response generation module constructs system response text that meets the needs of the scenario based on the updated session state of the session management module, combined with information such as collected variable mappings, a list of items to be clarified, and session state markers. It generates confirmation, clarification, follow-up question, referential confirmation, process termination, or exception handling responses for different scenarios, ensuring that the expression is natural and clear. The response text is stored in the dialogue history, sent to the user through the user interface module, and triggers process iteration judgment.

[0282] The session storage module is used to persistently store all session instances and their state information, including session identifiers, the historical and current states of each session variable, and dialogue history. It supports state recovery after session interruption, provides a consistent data access basis for all modules of the system, and ensures the continuity and traceability of the dialogue process.

[0283] Exemplary device:

[0284] The information collection device provided in this application includes a memory, a controller, and a computer program stored in the memory and executable on the controller. When the controller executes the program, it implements the large-scale dialogue information collection method for fusion business queries described in any of the above embodiments.

[0285] The information collection device may also include a memory; the controller in the device may be one or more, with one controller being used as an example in the figure; the memory is used to store one or more programs; the one or more programs are executed by the one or more controllers, enabling the one or more controllers to implement the large-model dialogue information collection method for fusion business queries as described in any of the above embodiments.

[0286] Information collection equipment also includes input devices and output devices.

[0287] The controller, memory, input device, and output device in the information collection device can be connected by a bus or other means. The figure shows an example of connection via a bus.

[0288] The input device can be used to receive input natural language information and interactive commands, as well as generate signal inputs related to user settings and function control of the information collection device. The output device may include interactive devices such as a display screen and a voice output module.

[0289] The memory, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to the large-model dialogue information collection method for integrated business queries described in the embodiments of this application. The memory may include a program storage area and a data storage area, wherein the program storage area may store the operating system and at least one application program required for a function; the data storage area may store session data, variable mappings, query results, etc., created based on the use of the information collection device. Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some instances, the memory may further include memory remotely configured relative to the controller, and these remote memories can be connected to the information collection device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0290] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard parts (ASSPs), systems-on-chip (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable controller, which may be a dedicated or general-purpose programmable controller, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0291] Based on the above embodiments, this embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a controller, implements the large-model dialogue information collection method for fusion business query in any of the above embodiments of the present invention.

[0292] Computer programs used to implement the methods of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The computer programs can be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0293] In the context of this invention, a computer-readable storage medium stores computer instructions that, when executed by a controller, implement the large-scale dialogue information collection method for converged business queries provided by this invention. The computer-readable storage medium can be a tangible medium that may contain or store computer programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer-readable storage medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, the computer-readable storage medium can be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0294] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device for displaying information to the user (e.g., a cathode ray tube (CRT) or a liquid crystal display (LCD monitor); and a keyboard, microphone, and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0295] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or computing systems that include middleware components (e.g., application servers), or computing systems that include frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0296] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system. It addresses the shortcomings of traditional physical hosts and Virtual Private Server (VPS) services, such as high management difficulty and weak business scalability.

[0297] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0298] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for collecting large-scale model dialogue information that integrates business queries, characterized in that, include: Receive initial input from the user and initialize session variables; Load predefined prompt word templates and session variables, fill in placeholders based on the state of the session variables, and generate prompt words for calling the large language model. The large language model is called to parse user input and candidate structured variables are extracted based on prompt words; Based on the candidate structured variable attributes, determine whether a business query needs to be executed; When business queries are required, multi-level fuzzy queries are performed based on the relationships between candidate structured variables and matching fields; Based on the execution results of multi-level fuzzy queries, the dialogue strategy is dynamically adjusted, including confirming variable values, generating a list of questions to be clarified, or using follow-up questions. Based on the results of the dialogue strategy adjustment, the conversation variables are updated. When the user inputs a list referent, the active list confidence is calculated by transforming the referent, and the target list and items are determined. Generate a system response and iteratively execute the above steps until all target variables are extracted.

2. The method for collecting large-scale model dialogue information for integrated business queries according to claim 1, characterized in that: Multi-level fuzzy queries include: Query conditions are constructed in order of priority: exact match fields, fuzzy match fields, and auxiliary match fields, and the matching conditions are gradually relaxed. Among them, exact match fields are subjected to full equality comparison; If no exact match is found, the process proceeds to the second level of fuzzy matching, which compares strings using a similarity algorithm. If fuzzy matching still yields no results or the number of results is insufficient, a third-level auxiliary matching is triggered, which combines the verified information of the associated variables to narrow down the query range. Among them, the matching threshold is dynamically adjusted. When two consecutive queries have no results, the threshold is lowered and the matching range of auxiliary fields is expanded. The query results are deduplicated, sorted, and then stored in the current query result cache.

3. The method for collecting large-scale dialogue information for integrated business queries according to claim 1, characterized in that: Calculating the confidence score for the active list includes: The matching confidence score is calculated based on a weighted combination of one or more of the following strategies: nearest-to-nearest strategy, semantic association strategy, dialogue focus strategy, and dialogue turn strategy. Among them, the nearest priority strategy calculates a time relevance score based on the difference between the generation timestamp of the list to be clarified and the current conversation timestamp; The semantic association strategy performs semantic matching based on entity type clues identified in user input and preset content categories in the list to be clarified; The dialogue focus strategy is evaluated by combining the relevance of the current dialogue topic with the categories of content in the clarification list; The dialogue turn strategy calculates a relevance score based on the distance between the current dialogue turn and the dialogue turn in which the clarification list was generated or referenced.

4. The method for collecting large-scale model dialogue information for integrated business queries according to claim 1, characterized in that: The target list and items are determined by transforming pronouns, including: Extract ordinal or positional descriptive pronouns from user input; Convert the extracted pronouns into uniform internal ordinal numbers or positional representations; By combining a multi-clarification list context manager, the matching confidence of each active clarification list is calculated; The target list is determined based on the matching confidence score, and the internal ordinal representation is matched with the ordinal identifier of the list item in the target list to lock the specific list item.

5. The method for collecting large-scale dialogue information for integrated business queries according to claim 1, characterized in that: The generated large language model invocation prompts include: After converting the target variable list to JSON format, fill in the JSON list placeholders for the variables to be extracted; After converting the contents of the collected variable mappings into JSON format, fill in the JSON list placeholders of the currently extracted variables; Fill the latest user input text into the latest user input text placeholder; Convert the summary of the most recent business query result in the current query result cache into JSON format and then fill in the corresponding placeholders.

6. The method for collecting large-scale dialogue information for integrated business queries according to claim 5, characterized in that: When generating prompts for a large language model, dynamically add targeted processing instructions based on the session state: When the number of follow-up questions about a variable approaches the predefined limit for the number of proactive follow-up questions, add a prompt to the prompt word indicating caution in asking follow-up questions; When multiple active clarification lists exist in a multi-clarification list context manager, add an instruction to prioritize processing the most recently generated clarification list; Based on the dependency graph maintained in the parent-child variable dependency relationship, supplement the explanation of the priority of variable extraction.

7. The method for collecting large-scale model dialogue information for integrated business queries according to claim 1, characterized in that: Updating session variables includes: When the query result is a single exact match, the variable validation status is updated to validated and a unique identifier for the business system is stored. When a list of clarifications is generated, the list is added to the collection of lists of clarifications and the context manager of the multiple clarification lists is updated. When no matching record is found, update the follow-up question count and mark the variable status. Dynamically maintain the dialogue history and current query result cache.

8. A large-scale model dialogue information collection system for integrated business queries, based on the large-scale model dialogue information collection method for integrated business queries as described in any one of claims 1-7, characterized in that, include: The user interface module receives and preprocesses the user's natural language input, and then sends the system response to the user. The session management module creates session instances and manages session variables, including a list of target variables, a map of collected variables, a set of lists to be clarified, and a context manager for multiple clarification lists. The prompt word generation module loads predefined templates, dynamically fills in placeholders with session variables, and adds targeted processing instructions to generate prompt words for calling the large language model. The large language model interaction module integrates prompt words and user input, calls the large language model to extract candidate structured variables, and performs format validation and security filtering; The business query decision module determines whether to execute a business query based on the attributes of candidate variables, constructs a queue of variables to be queried, and sorts them according to the priority of business logic. The business query execution module performs multi-level fuzzy queries based on the queue of variables to be queried, dynamically adjusts the matching threshold, and handles the parent-child variable dependency relationship. The dialogue strategy adjustment module generates a clarification list, follow-up questions, or confirmation instructions based on the query results, and parses pronouns in the user input to determine the target item. The response generation module generates natural language responses based on the session state, including confirmation, clarification, follow-up questions, and process termination texts. The session storage module stores session instances and state information, and supports state recovery after a session is interrupted.

9. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the large-scale dialogue information collection method for fusion business queries as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the large-scale dialogue information collection method for fusion business queries as described in any one of claims 1-7.

Images