Intelligent complex intention recognition method and system based on large model

Abstract

The invention discloses an intelligent complex intention recognition method and system based on a large model, and the method comprises the following steps: (1) carrying out the preprocessing of user input data, and recognizing and marking a plurality of intentions; (2) when the intention quantity is greater than a preset threshold value, triggering a reasoning process; (3) acquiring a context record and historical state information of the current session; (4) dynamically generating and optimizing a prompt text based on the intention, the context and the historical state, and inputting the prompt text into a large language model for reasoning; (5) storing a reasoning result to a state management system and a context storage system; (6) distributing the reasoning result to a single-round semantic module and terminal equipment, and obtaining an intermediate state result; (7) integrating intermediate state results to form structured data, and feeding back the structured data to a prompt management module after natural language conversion; and (8) iteratively executing the steps (4)-(7) until a final reasoning result is output and issued to the terminal equipment.

Description

Technical Field

[0001] This invention relates to the field of large model optimization and intelligent interaction technology in complex scenarios, and more specifically to a method and system for intelligent complex intent recognition based on a large model. Background Technology

[0002] Currently, the field of complex intent recognition commonly employs FunctionCall technology based on large models. The core process includes: voice or text input parsing: converting user input into text using ASR (Automatic Speech Recognition) or NLP models. Intent classification and slot filling: leveraging the semantic understanding capabilities of large models to identify user intents (such as "check the weather" or "set reminders") and extract key information (such as location and time).

[0003] Tool call decision: Based on the intent, match the preset FunctionCall list and select the corresponding tool interface (such as map API, alarm reminder).

[0004] Multi-turn dialogue management: Optimize subsequent interactions through context caching (such as dialogue history).

[0005] In the aforementioned complex scenarios, the FunctionCall model exhibits poor response performance, specifically manifested in inaccurate intent recognition, high latency, and high resource costs. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent complex intent recognition method and system based on a large model, in order to solve the technical problems existing in the background art. It adopts a multi-module interaction mechanism and a modular collaborative architecture: through the distributed collaboration of cloud modules, central control modules, interactive large model modules, and inference modules, it achieves layered processing of complex intents.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for intelligent complex intent recognition based on a large model includes the following steps:

[0009] (1) Preprocess user input data using a lightweight intent recognition model to identify and label multiple intents;

[0010] (2) When the number of intents exceeds a preset threshold, the inference module is triggered;

[0011] (3) Obtain the context record and historical state information of the current session;

[0012] (4) Dynamically generate and optimize prompt text based on intent, context and historical state, and input it into a large language model for inference;

[0013] (5) Store the reasoning results in the state management system and the context storage system;

[0014] (6) The reasoning results are distributed to the single-round semantic module and the terminal device through the central control module to obtain the intermediate state results;

[0015] (7) Integrate intermediate state results to form structured data, and feed it back to the prompt management module after natural language conversion;

[0016] (8) Iterate through steps (4)-(7) until the final inference result is output and sent to the terminal device.

[0017] In some embodiments, the lightweight intent recognition model in step (1) is the 7B function retrieval algorithm, and its processing includes:

[0018] Clean the input text of special characters, redundant spaces, and punctuation marks, and perform simplified / traditional Chinese conversion and synonym normalization;

[0019] Multiple intent keywords are identified through keyword matching and semantic similarity calculation.

[0020] In some embodiments, the preset threshold in step (2) is a single intent threshold.

[0021] In some embodiments, in step (3), context records and historical states are retrieved from the distributed cache using the user session identifier.

[0022] In some embodiments, dynamically generating the prompt text in step (4) includes:

[0023] Filter contextual and historical state information related to the intent;

[0024] Information is encapsulated and spliced ​​according to a preset format, which includes intent priority, context association logic, and historical state reference identifier.

[0025] In some embodiments, obtaining the intermediate state result in step (6) includes:

[0026] Extract intermediate semantic analysis data from the single-round semantic module;

[0027] Extract and analyze results and status feedback data from terminal devices.

[0028] In some embodiments, the natural language conversion in step (7) employs a natural language generation algorithm to map structured data fields into semantically coherent text.

[0029] This embodiment also provides an intelligent complex intent recognition system based on a large model, including:

[0030] A lightweight intent recognition module is used to preprocess user input data and recognize multiple intents;

[0031] The central control module counts the number of intentions and triggers the reasoning process, coordinating data interaction among multiple modules;

[0032] Context storage system stores session history and operation status;

[0033] The prompt management module dynamically generates and optimizes the prompt text for input large language models;

[0034] The large language model module performs deep inference and outputs the results.

[0035] The state management system stores structured data of inference results;

[0036] The natural language conversion module converts structured data into natural language text.

[0037] A single-round semantic module and terminal device interface are used to obtain intermediate state results.

[0038] In some embodiments, the lightweight intent recognition module employs the 7B function retrieval algorithm, which includes a text cleaning unit and an intent keyword extraction unit.

[0039] In some embodiments, the prompt management module includes:

[0040] The information filtering unit filters redundant information based on intent relevance.

[0041] Encapsulate and splice units to generate structured prompt text according to a preset logical format.

[0042] In some embodiments, the central control module is further configured to:

[0043] Integrate the intermediate state results returned by the single-round semantic module and the terminal device;

[0044] Update the session context to the context storage system in real time.

[0045] In some embodiments, a distributed architecture is adopted, including a cloud module, an interactive large model module, and an inference module, wherein the cloud module provides underlying support for context storage and state management.

[0046] The beneficial effects of this invention compared to the prior art are:

[0047] In intent processing, a 7B function retrieval system accurately identifies multiple intents, replacing the cumbersome FunctionCall tool list and avoiding the limitations of single-intent processing. During the inference process, dynamic prompt management is implemented by combining context and historical states, making large-scale model inference more aligned with user needs. In the data interaction phase, intermediate state results are obtained through multi-module interaction, ensuring information integrity and accuracy. The conversion of structured data to natural language makes the results easier to understand. Overall, this achieves efficient, intelligent, and user-friendly interactive processing, improving user experience and system processing efficiency. Attached Figure Description

[0048] Figure 1 This is a flowchart of this embodiment. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar components or components having the same or similar functions throughout. The described embodiments are some, but not all, of the embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0050] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0051] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, an indirect connection through an intermediate medium, or the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0052] In the description of this application, it should be understood that the terms "upper", "lower", "front", "back", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this application.

[0053] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or display that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or display.

[0054] The following will combine Figure 1 This application provides a detailed description of an intelligent complex intent recognition method based on a large model, as illustrated in the embodiments of this application. It is worth noting that the following embodiments are merely illustrative of this application and do not constitute a limitation thereof.

[0055] Example 1:

[0056] like Figure 1 As shown, an intelligent complex intent recognition method based on a large model includes the following steps:

[0057] Step 1.1: 7B function retrieval identifies multiple intents and preprocesses input data.

[0058] The system receives text input from the user, removes special characters, redundant spaces and punctuation marks through the text cleaning module, and performs simplified / traditional Chinese conversion and synonym normalization.

[0059] Step 1.2: The 7B function searches and identifies intent, recognizing multiple intent keywords.

[0060] Step 2.1: Count the number of central control intentions. The central control determines the number of intentions and enters the reasoning module.

[0061] The central control module receives the intent list output by the 7B retrieval model and counts the number of intents using a counter. For example, the intent "If it rains tomorrow in Chengdu, remind me to bring an umbrella at 7 am" would be output as ["WEATHER","REMINDER"]. The count would be 2.

[0062] Step 3.1: Obtain context records and historical states

[0063] The central control module retrieves the context record of the current session from a distributed cache (such as Redis) based on the user session ID (cot_id). Simultaneously, it queries the cot_status based on the cot_id and performs different logical actions according to the status.

[0064] Step 4.1: Prompt management encapsulation and splicing

[0065] The prompt management module searches for the corresponding tool description in the database based on the current intent list (such as "WEATHER" or "REMINDER") and context information, in order to construct the most suitable prompt word.

[0066] Step 5.1: Large Model Reception and Processing. The concatenated prompt text is sent to the Doubao large model. After receiving it, the large model uses its internal algorithms and training parameters to perform deep analysis and inference calculations on the input content.

[0067] Step 5.2: Result Storage. After the large model completes inference, the output results are stored in the state management system and the context storage system respectively, for use and traceability in subsequent steps.

[0068] Step 6.1: Extracting the Reasoning Results

[0069] Step 6.2: Transmit the results to the central control unit

[0070] The results are sent to the central control module, which then parses them and generates a task execution plan, such as "first check today's weather, and then infer whether a reminder is needed based on the results returned by the weather interface".

[0071] Step 7.1: Semantic Interconnection between Central Control and Single-Wheel Control System

[0072] The central control module sends the current reasoning result to the single-round semantic module, such as "Query the weather in Chengdu today".

[0073] Step 8.1: Connecting the central control unit and the terminal

[0074] The central control module converts the follow-up questions into a terminal-compatible format (such as APP pop-ups or voice broadcasts) and sends them to the user terminal.

[0075] Step 8.2: Obtain the intermediate state result of the terminal

[0076] The system receives input from the user terminal (such as "set the alarm to 7 o'clock"), parses it into structured data (such as "date":"2025-06-24","alarm":"alarm","time":"19:00:00"), and transmits it to the central control module.

[0077] Step 9.1: Organizing Intermediate Results

[0078] Integrating single-round semantics with intermediate results returned by the terminal, structured data is formed: "src_txt":"Query Chengdu's weather today","feedback":"Heavy rain in Chengdu today, temperature 19 to 22 degrees Celsius, air quality excellent.","semantic":{"datetime":"2025-06-24","location":"Chengdu City"},

[0079] Step 9.2: Store the results in context management.

[0080] The structured data is stored in a distributed database, and the session context is updated for use in subsequent steps.

[0081] Step 10.1: Structured Data Parsing. After receiving the structured data, the Natural Language Transformation module performs in-depth analysis of the data's structure and content, understanding the meaning and function of each data field.

[0082] Step 10.2: Convert to Natural Language. Based on the parsing results, a natural language generation algorithm is used to convert the structured data into natural language text that is easy for humans to understand, ensuring that the converted text is semantically accurate and fluently expressed.

[0083] Step 11.1: Pass text to prompt management

[0084] The text is sent to the prompt management module as input reference for the next round of inference in the large model.

[0085] Step 12.1: Secondary filtering and supplementation of the prompt. After receiving the natural language text, the prompt management module performs secondary filtering and supplementation on the text based on the current context information and task requirements to further optimize the prompt content.

[0086] Step 12.2: Re-encapsulation and concatenation. The optimized natural language text, along with other relevant information (such as context records, historical states, etc.), is re-encapsulated and concatenated according to the input requirements of the large model to generate a new prompt text.

[0087] Step 13.1: New round of processing for the large model. The newly generated prompt text is sent to the Doubao large model again. The large model performs a new round of inference calculation based on the new input (e.g., "feedback": "Heavy rain in Chengdu today, temperature 19 to 22 degrees Celsius, air quality excellent."), and obtains the next inference result (because the result is rainy, I need to be reminded to bring an umbrella at 7 o'clock).

[0088] Step 13.2: Result Storage and Update. Store the next inference result output by the large model into the state management system and context storage system, overwriting or updating previous result information to maintain the real-time nature and accuracy of the data.

[0089] Step 14.1: Extraction and organization of inference results. The state management system extracts the next inference result data from the stored new results and performs formatting and necessary preprocessing on the data to make it meet the receiving requirements of the central control module.

[0090] Step 14.2: Transmit the results to the central control module. The compiled results of the next step of reasoning are transmitted to the central control module, enabling it to obtain the latest reasoning progress for subsequent decision-making and processing.

[0091] Step 15.1: Central Control Result Preparation. The central control module performs a final check and adjustment on the received inference results. Based on the characteristics and requirements of the terminal devices, it performs adaptation processing on the results, such as format conversion and content simplification.

[0092] Step 15.2: Sending inference results to the terminal. Through the established communication link and interface, the adapted inference results are sent to the terminal device, enabling the terminal device to display or further process the final inference results.

[0093] This embodiment also provides an intelligent complex intent recognition system based on a large model, including:

[0094] A lightweight intent recognition module is used to preprocess user input data and recognize multiple intents;

[0095] The central control module counts the number of intentions and triggers the reasoning process, coordinating data interaction among multiple modules;

[0096] Context storage system stores session history and operation status;

[0097] The prompt management module dynamically generates and optimizes the prompt text for input large language models;

[0098] The large language model module performs deep inference and outputs the results.

[0099] The state management system stores structured data of inference results;

[0100] The natural language conversion module converts structured data into natural language text.

[0101] A single-round semantic module and terminal device interface are used to obtain intermediate state results.

[0102] In some embodiments, the lightweight intent recognition module employs the 7B function retrieval algorithm, which includes a text cleaning unit and an intent keyword extraction unit.

[0103] In some embodiments, the prompt management module includes:

[0104] The information filtering unit filters redundant information based on intent relevance.

[0105] Encapsulate and splice units to generate structured prompt text according to a preset logical format.

[0106] In some embodiments, the central control module is further configured to:

[0107] Integrate the intermediate state results returned by the single-round semantic module and the terminal device;

[0108] Update the session context to the context storage system in real time.

[0109] In some embodiments, a distributed architecture is adopted, including a cloud module, an interactive large model module, and an inference module, wherein the cloud module provides underlying support for context storage and state management.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent complex intent recognition based on a large model, characterized in that, Includes the following steps: (1) Preprocess user input data to identify and label multiple intents; (2) When the number of intents exceeds a preset threshold, the inference process is triggered; (3) Obtain the context record and historical state information of the current session; (4) Dynamically generate and optimize prompt text based on intent, context and historical state, and input it into a large language model for inference; (5) Store the reasoning results in the state management system and the context storage system; (6) Distribute the reasoning results to the single-round semantic module and the terminal device to obtain intermediate state results; (7) Integrate intermediate state results to form structured data, and feed it back to the prompt management module after natural language conversion; (8) Iterate through steps (4)-(7) until the final inference result is output and sent to the terminal device.

2. The method according to claim 1, characterized in that, The lightweight intent recognition model mentioned in step (1) is the 7B function retrieval algorithm, and its processing includes: Clean the input text of special characters, redundant spaces, and punctuation marks, and perform simplified / traditional Chinese conversion and synonym normalization; Multiple intent keywords are identified through keyword matching and semantic similarity calculation.

3. The method according to claim 1, characterized in that, The preset threshold mentioned in step (2) is a single intent threshold.

4. The method according to claim 1, characterized in that, In step (3), context records and historical states are retrieved from the distributed cache using the user session identifier.

5. The method according to claim 1, characterized in that, Step (4) involves dynamically generating the prompt text, including: Filter contextual and historical state information related to the intent; Information is encapsulated and spliced ​​according to a preset format, which includes intent priority, context association logic, and historical state reference identifier.

6. The method according to claim 1, characterized in that, The intermediate state results obtained in step (6) include: Extract intermediate semantic analysis data from the single-round semantic module; Extract and analyze results and status feedback data from terminal devices.

7. The method according to claim 1, characterized in that, In step (7), natural language conversion uses a natural language generation algorithm to map structured data fields into semantically coherent text.

8. A complex intent recognition system based on a large model, characterized in that, include: A lightweight intent recognition module is used to preprocess user input data and recognize multiple intents; The central control module counts the number of intentions and triggers the reasoning process, coordinating data interaction among multiple modules; Context storage system stores session history and operation status; The prompt management module dynamically generates and optimizes the prompt text for input large language models; The large language model module performs deep inference and outputs the results. The state management system stores structured data of inference results; The natural language conversion module converts structured data into natural language text. A single-round semantic module and terminal device interface are used to obtain intermediate state results.

9. The system according to claim 8, characterized in that, The lightweight intent recognition module adopts the 7B function retrieval algorithm, which includes a text cleaning unit and an intent keyword extraction unit.

10. The system according to claim 8, characterized in that, The prompt management module includes: The information filtering unit filters redundant information based on intent relevance. Encapsulate and splice units to generate structured prompt text according to a preset logical format.

Images