CN-122019570-A - Large model dialogue information collection method and system integrating service inquiry

CN122019570ACN 122019570 ACN122019570 ACN 122019570ACN-122019570-A

Abstract

The invention discloses a large model dialogue information collection method and system for fusing service inquiry, comprising the steps of receiving initial input of a user, initializing a dialogue variable, loading a predefined prompt word template, generating a call prompt word by combining a dialogue state filling placeholder, calling a large model to extract candidate structured variables, executing multi-level fuzzy inquiry according to accurate, fuzzy and auxiliary matching priorities when judging that the service inquiry needs to be executed, adjusting dialogue strategies according to the results, confirming the variables, generating a clear list or inquiring, converting the user prompt word and calculating the confidence coefficient of the active list to determine target items when updating the dialogue variable, and generating response iteration until all target variables are extracted. Based on the construction, the system solves the problems of insufficient fusion of information extraction and service verification, fuzzy references and multi-list confusion in the prior art, improves the variable extraction accuracy and dialogue efficiency, and adapts to complex multi-round service dialogue scenes.

Inventors

SUN HAIDONG
WANG YING
ZHANG XIAO
ZHU MINGYU
SHEN YI
LU YIMING
SONG JIAYI

Assignees

苏州瑞泰信息技术有限公司

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. A large model dialogue information collection method integrating service inquiry is characterized by comprising the following steps: receiving initial input of a user, and initializing a session variable; Loading a predefined prompt word template and a session variable, filling placeholders according to the state of the session variable, and generating a large language model call prompt word; invoking a large language model to analyze user input, and extracting candidate structured variables according to the prompt words; judging whether service inquiry is required to be executed or not based on the candidate structured variable attribute; when service inquiry is needed, executing multi-stage fuzzy inquiry according to the association relation and matching field of the candidate structured variables; dynamically adjusting dialogue strategies according to the execution result of the multilevel fuzzy query, including confirming variable values, generating a list to be clarified or pursuing a dialogue; Updating a session variable according to the adjustment result of the dialogue strategy, and when a user inputs a list reference word, calculating the confidence level of the active list through converting the reference word to determine a target list and items; Generating a system response and iteratively executing the steps until all target variables are extracted.
2. The method for collecting large model dialogue information with integrated business query as claimed in claim 1, wherein the multi-level fuzzy query comprises: Constructing query conditions according to the priorities of the accurate matching field, the fuzzy matching field and the auxiliary matching field in sequence, and gradually relaxing the matching conditions; wherein the exact match field performs a full-equal comparison; If the accurate matching does not have a result, the second-stage fuzzy matching is carried out, and the comparison is carried out through a character string similarity algorithm; triggering a third-level auxiliary matching when the fuzzy matching still has no result or the result quantity is insufficient, and reducing the query range by combining the verified information of the related variable; And when no result exists in the continuous two-stage query, the threshold is reduced, the matching range of the auxiliary field is enlarged, the query result is reordered and then stored in the current query result cache.
3. The method for collecting large model dialogue information with integrated business query as claimed in claim 1, wherein calculating confidence of active list comprises: Calculating a match confidence score based on a weighted combination of one or more of a most recent priority policy, a semantic association policy, a dialog focus policy, and a dialog turn policy; The latest priority strategy calculates a time correlation score according to the difference value between the generated time stamp of the list to be clarified and the current dialogue time stamp; the semantic association strategy carries out semantic matching on the entity type clues identified in the user input and preset content categories of the list to be clarified; the dialogue focus strategy is evaluated by combining the relevance of the current dialogue theme and the content category of the list to be clarified; The dialog turn policy calculates a relevance score based on the dialog turns for which the list to be clarified is generated or referenced, and the distance from the current dialog turn.
4. The method for collecting large model dialogue information with integrated business query as claimed in claim 1, wherein determining the target list and the target item through the conversion reference word comprises: extracting ordinal or location description referents from user input; Converting the extracted reference words into uniform internal ordinal numbers or position representations; calculating the matching confidence of each active list to be clarified by combining a multi-clarification list context manager; and determining a target list based on the matching confidence score, matching the internal ordinal representation with the ordinal identification of the list item in the target list, and locking the specific list item.
5. The method for collecting large model dialogue information for business query fusion according to claim 1, wherein generating large language model call prompt words comprises: Converting the target variable list into a JSON format, and filling the JSON list placeholder of the variable to be extracted; converting the content in the collected variable mapping into a JSON format, and filling the JSON list placeholders of the current extracted variables; Filling the latest user input text into the latest user input text placeholder; And converting the result summary of the latest service query in the current query result cache into a JSON format and filling the JSON format into a corresponding placeholder.
6. The method for collecting large model dialogue information with integrated business query as claimed in claim 5, wherein when generating large language model call prompt words, the method is characterized in that the method comprises the steps of dynamically adding targeted processing instructions in combination with session states: When the number of the overtaking of the variable approaches to the predefined limit of the number of the active overtaking, a cautious overtaking guide is added into the prompt word; when there are multiple active clarification lists in the multi-clarification list context manager, adding an indication that prioritizes recently generated clarification lists; And supplementing the priority description of variable extraction according to the dependency relationship diagram among the variables maintained in the parent-child variable dependency relationship.
7. The method for collecting large model dialogue information with integrated business query as claimed in claim 1, wherein updating the dialogue variables comprises: when the query result is single exact match, updating the variable verification state to verified and storing the unique identifier of the service system; When generating a list to be clarified, adding the list to a set of lists to be clarified and updating a multi-clarification list context manager; When no matching record exists, updating the statistics of the number of the overtaking questions and marking the variable state; and dynamically maintaining a dialogue history record and a current query result cache.
8. A business query-converged large model dialogue information collection system, based on the business query-converged large model dialogue information collection method according to any one of claims 1 to 7, comprising: the user interface module receives and preprocesses natural language input of a user and transmits a system response to the user; The session management module creates a session instance and manages session variables, including a target variable list, a collected variable map, a list set to be clarified and a multi-clarification list context manager; The prompt word generation module loads a predefined template, dynamically fills placeholders in combination with session variables and adds targeted processing instructions to generate large language model call prompt words; the large language model interaction module integrates the prompt words and user input, invokes the large language model to extract candidate structured variables and performs format verification and safety filtration; the business inquiry decision module judges whether to execute business inquiry based on the candidate variable attribute, constructs a variable queue to be inquired and sorts the variable queue according to the business logic priority; The business inquiry executing module executes multi-level fuzzy inquiry according to the variable queue to be inquired, dynamically adjusts a matching threshold value and processes father-son variable dependency relationship; The dialogue strategy adjusting module generates a clarification list, a questionnaire or a confirmation instruction according to the query result, analyzes the reference words in the user input and determines target items; The response generation module generates natural language response based on the session state, wherein the natural language response comprises a confirmation type text, a clarification type text, a challenge type text and a flow ending type text; and the session storage module is used for storing the session instance and the state information and supporting state recovery after session interruption.
9. An electronic device, comprising: at least one processor, and A memory communicatively coupled to the at least one processor, wherein, The memory stores a computer program executable by the at least one processor to enable the at least one processor to perform the large model dialog information collection method of fusion service queries of any of claims 1-7.
10. A computer readable storage medium storing computer instructions for causing a processor to execute the large model dialogue information collection method of fusion business query according to any one of claims 1-7.

Description

Large model dialogue information collection method and system integrating service inquiry Technical Field The invention relates to the technical field of electric digital data processing, in particular to a large model dialogue information collection method and system integrating service inquiry. Background Currently, with the continuous evolution of natural language processing technology, a Large Language Model (LLM) shows excellent understanding and generating capability in a dialogue information extraction task, and becomes an important technical support for scenes such as intelligent customer service, business consultation and automatic form filling. Traditional methods based on rules or statistical models have limited generalization capability in the face of diversified and unstructured user expressions, while large models can more flexibly identify and extract key entities, relations and instructions from conversations by virtue of strong semantic understanding and context reasoning capability, so that the coverage range and interaction naturalness of information collection are remarkably improved. In the prior art, some schemes have been attempted to combine a large model with a business system, such as adjusting and optimizing information collection flow by dynamic prompt words, or using large model agents to call APIs to achieve automatic form filling. Although the existing large model dialogue information collection method for fusing business queries advances in semantic understanding and flow automation, the method still has significant defects, and particularly has limitation in complex multi-round dialogue scenes. The large model is usually guided by the prompt to extract variables from the dialog and to introduce a business query interface for data validation or form population. For example, part of the schemes employ search enhancement generation (RAG) techniques to access external knowledge bases or convert natural language into database queries via NL2SQL to enhance the business relevance of information. However, the core defect is that the real-time fusion of the information extraction process of the large model and the logic and data verification of the service system is insufficient, so that the extraction result lacks service accuracy. Specifically, the large model only relies on semantic matching to extract variables, but cannot verify the existence, uniqueness or validity of the name in the business database in real time, and is easy to generate "illusion" or error data. Accordingly, there is a need for improvements in the art for large model dialogue information collection methods and systems that integrate business queries to solve the above-described problems. Disclosure of Invention The invention overcomes the defects of the prior art, provides a large model dialogue information collection method and a large model dialogue information collection system for fusing business query, and aims to solve the problems of fuzzy references and multi-list context confusion caused by multi-round dialogue when large model information is extracted in the prior art. In order to achieve the purpose, the technical scheme adopted by the invention is that the large model dialogue information collecting method for integrating service inquiry comprises the following steps: receiving initial input of a user, and initializing a session variable; Loading a predefined prompt word template and a session variable, filling placeholders according to the state of the session variable, and generating a large language model call prompt word; invoking a large language model to analyze user input, and extracting candidate structured variables according to the prompt words; judging whether service inquiry is required to be executed or not based on the candidate structured variable attribute; when service inquiry is needed, executing multi-stage fuzzy inquiry according to the association relation and matching field of the candidate structured variables; dynamically adjusting dialogue strategies according to the execution result of the multilevel fuzzy query, including confirming variable values, generating a list to be clarified or pursuing a dialogue; Updating a session variable according to the adjustment result of the dialogue strategy, and when a user inputs a list reference word, calculating the confidence level of the active list through converting the reference word to determine a target list and items; Generating a system response and iteratively executing the steps until all target variables are extracted. In a preferred embodiment of the present invention, the multi-level fuzzy query comprises: Constructing query conditions according to the priorities of the accurate matching field, the fuzzy matching field and the auxiliary matching field in sequence, and gradually relaxing the matching conditions; wherein the exact match field performs a full-equal comparison; If the accurate matching does not have a resu