Metacognitive Reasoning and Dynamic Adaptive Intent Understanding System Based on Human-Machine Collaboration

Abstract

This application discloses a metacognitive reasoning and dynamic adaptive intent understanding system based on human-computer collaboration, belonging to the field of artificial intelligence technology. It includes: an input processing module that receives user input; a two-stage knowledge retrieval module that uses the MN-BERT model to classify user intents by difficulty to narrow down the retrieval scope, and then uses a Rerank model to refine the candidate intent set and filter out relevant intents; an LLM reasoning module that, based on the retrieval results, uses a large language model to perform metacognitive reasoning, locates the final target intent, and intelligently infers and completes the required parameters; an intent integrity assessment module that evaluates the completeness of the reasoning results; if incomplete, a multi-turn dialogue guidance module is triggered, which dynamically generates guidance questions through LLM and interacts with the user to complete the information; and a continuous learning and knowledge base update module that continuously collects interaction data, optimizes the model, and updates the knowledge base. This system solves the problems of balancing intent retrieval efficiency and accuracy, insufficient depth of reasoning for complex intents, and rigid interactive guidance.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a metacognitive reasoning and dynamic adaptive intent understanding system based on human-computer collaboration. Background Technology

[0002] Currently, accurately and efficiently understanding user intent is a core challenge in fields such as intelligent dialogue systems, intelligent customer service, and fault diagnosis. Traditional intent recognition methods often fall short when user expressions are complex or ambiguous, or when intent evolves and shifts across multiple rounds of interaction. The main problems faced by these systems include: how to quickly and accurately locate the user's true intent from a vast array of possible intents; how to infer the required parameters that the user has not explicitly expressed when contextual information is limited or incomplete; and how to design a dynamic and adaptive interaction mechanism that, with the fewest possible rounds and the most natural guidance, completes missing information to smoothly and accurately fulfill the user's instructions or requests.

[0003] To address the above challenges, the industry has mainly developed the following intent recognition methods: (1) Rule-based methods match based on predefined keywords and scales. This method is simple and direct, but lacks flexibility and struggles to handle the diversity and complexity of language. (2) Machine learning methods use algorithms such as Support Vector Machines (SVM) and Random Forests to classify intents. Compared to rule-based methods, it can handle a certain degree of language variation, but its performance heavily relies on manually designed features. (3) Deep learning methods utilize models such as Recurrent Neural Networks (RNN) and Convolutional Neural Networks (CNN) to automatically learn semantic representations. This method improves representation capabilities, but its generalization ability remains limited due to the inherent limitations of the model structure. (4) Large Language Model (LLM)-based methods leverage the powerful contextual understanding and generation capabilities of large language models based on the Transformer architecture to directly perform intent recognition and reasoning. For example, as described in patent CN117786064A, LLM is used to identify user intents and call relevant functions to generate answers.

[0004] Despite the progress made in existing technologies, significant limitations and shortcomings remain, including: The contradiction between retrieval efficiency and performance: For the limited context length of LLMs, the conventional combination of embedding (dual encoder) and rerank (cross encoder) models has bottlenecks. Embedding models have limited performance ceilings because queries and passages don't interact during "offline" encoding; while using a rerank model alone can fully utilize interactive information to improve accuracy, its computational cost is extremely high. Combining both methods increases cost and complexity because the embedding model has poor recall capabilities, requiring dynamic adjustment of thresholds or recall numbers. Insufficient depth and adaptability in multi-turn dialogues: Methods described in patent CN112256825B only use natural language understanding models to extract entity and intent information from the current and previous turns for completion. This method suffers from insufficient information utilization (ignoring earlier dialogue history), insufficient depth of intent understanding (difficulty in handling complex implicit intents), and a lack of adaptability (inability to dynamically adjust strategies based on different users and scenarios).

[0005] In summary, existing technologies have significant shortcomings in balancing retrieval efficiency and accuracy, deep semantic reasoning, and adaptive interactive guidance when dealing with complex and dynamic real-world intent understanding scenarios. Therefore, there is an urgent need for a more intelligent, efficient, and human-like intent understanding system. Summary of the Invention

[0006] The present invention aims to solve the problems in the prior art that it is impossible to effectively balance the efficiency and accuracy of intent retrieval, and that it is difficult to perform deep reasoning and intelligent, dynamic guidance and completion for complex and incomplete user intents, thereby improving the accuracy of intent recognition and interaction efficiency, and improving user experience.

[0007] To achieve the above objectives, this invention provides a metacognitive reasoning and dynamic adaptive intent understanding system and method based on human-computer collaboration, which can accurately locate intent, actively reason and complete information, and achieve complete intent capture through natural multi-turn dialogue.

[0008] In a first aspect, the present invention provides a metacognitive reasoning and dynamic adaptive intent understanding system based on human-computer collaboration, the system comprising: The input processing module is used to receive and process user input information; A two-stage knowledge retrieval module, connected to the input processing module, is used to retrieve candidate intents from the intent knowledge base based on the processed user input. The two-stage knowledge retrieval module includes: The first-level retrieval submodule, built on the MN-BERT model, is used to classify the difficulty of user intent, thereby narrowing the search scope. The second-level retrieval submodule, built on the Rerank model, is used to perform fine-grained retrieval of the intent knowledge base subset after being filtered by the first-level retrieval submodule, and to select one or more of the most relevant candidate intents. The LLM reasoning module is connected to the two-stage knowledge retrieval module. It is used to receive user input and candidate intents, perform metacognitive reasoning through the large language model LLM, determine the final target intent, and perform inference to complete its required parameters. An intent integrity assessment module, connected to the LLM inference module, is used to assess the integrity of the intent information output by the LLM inference module based on rules and machine learning methods. A multi-turn dialogue guidance module, connected to the intent integrity assessment module, is used to initiate an interaction process to generate guidance questions and conduct multi-turn dialogues with the user to complete the missing information when the intent information is determined to be incomplete. The continuous learning and knowledge base update module is used to collect interaction data and optimize model parameters and update the intent knowledge base.

[0009] Specifically, in the two-stage knowledge retrieval module, the MN-BERT model used in the first-level retrieval submodule is an improved model based on the BERT model. Its downstream tasks in the pre-training stage include the Masked Language Modeling (MLM) task and the Named Entity Recognition (NER) task with custom intent parameter labels as entities. Its loss function... The loss function is the sum of the MLM task loss and the NER task loss. The expression is,

[0010] In the formula, For MLM task loss, For NER mission losses, For the set of words that are obscured, The total number of words that are masked. For the obscured word, This is the word representation obtained through the last layer of the BERT model; The number of words in the sentence. For the first One-hot encoding of the true label of each word. The model predicts the first The probability distribution of the labels for each word.

[0011] Specifically, in the two-stage knowledge retrieval module, the reranking model used in the second-level retrieval submodule is a model based on the Cross Encoder architecture, and is trained using a two-stage training method including a general corpus training stage and a professional intent understanding corpus training stage; wherein, the positive examples of the professional intent understanding corpus are pairs of user questions and corresponding intents, and the negative examples are pairs of user questions and the highest-scoring incorrect intent samples calculated by the reranking model.

[0012] Furthermore, the re-ranking model is trained using a local contrastive estimation method, and its loss function is... The expression is,

[0013] In the formula, It is a query set. Is it related to query? Related document collection, The target search engine is for the query Found The top-ranked collection of documents, It is a positive sample. It is a query and documents Match scores between them.

[0014] Furthermore, the re-ranking model calculates the semantic similarity between the prompt and each intent description in the knowledge base, and the system identifies and filters out the three most relevant intents and their corresponding parameters as a candidate set.

[0015] Specifically, the LLM inference module achieves precise positioning of intent through parameter inference via two key steps: Intent selection reflection: The LLM comprehensively evaluates and compares the candidate intents of the input, and simulates metacognitive thinking to determine the optimal target intent; Parameter completion reasoning: For the selected target intent, the LLM compares the required parameters with the parameters already provided by the user, and uses contextual reasoning and knowledge transfer to deduce the missing parameters that the user has not explicitly provided.

[0016] Specifically, in the multi-turn dialogue guidance module, when the intent information is determined to be incomplete, the multi-turn dialogue mechanism is activated, and the interactive intent information completion execution process includes the following steps: Step A: Identify missing parameters in the current intent information; Step B: Based on the missing parameters, the LLM dynamically generates one or more targeted guiding questions; Step C: Receive user feedback on the guided question and integrate the feedback information into the intent information; Step D: Return the updated intent information to the intent integrity assessment module for reassessment; Step E: Repeat steps A through D until the intent information reaches the preset integrity threshold.

[0017] Secondly, the present invention provides a metacognitive reasoning and dynamic adaptive intent understanding method based on human-computer collaboration, applied to a metacognitive reasoning and dynamic adaptive intent understanding system based on human-computer collaboration provided in the first aspect. The method includes the following steps: Step S1: Initial input processing and intent candidate generation; Step S2: Intent Precision Localization and Parameter Inference Driven by LLM; Step S3: Intent information integrity assessment; Step S4: If the evaluation result is incomplete, perform multiple rounds of interactive intent-guided completion. Step S5: Continuous learning and knowledge base updates.

[0018] Furthermore, in step S1, the initial input processing and intent candidate generation include: Step S101: Receive the user's input question; Step S102: Use the MN-BERT model to classify the difficulty of user intent and determine the subset of the intent knowledge base to be retrieved; Step S103: Use the re-ranking model to perform fine-grained retrieval on the subset of the intent knowledge base, and select the top N most relevant candidate intents and their corresponding parameters, where N is an integer greater than or equal to 1.

[0019] Furthermore, in step S4, if the evaluation result is incomplete, then performing multi-round interactive intent-guided completion includes: Step S401: Identify missing information in the intent information; Step S402: The LLM generates one or more boot questions based on the missing information; Step S403: Receive the user's answer to the guidance question and integrate the new information into the intent information; Step S404: Repeat the assessment and guidance until the intent information is complete.

[0020] Thirdly, the present invention provides a computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a metacognitive reasoning and dynamic adaptive intent understanding method based on human-computer collaboration as provided in the second aspect of the present invention.

[0021] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a metacognitive reasoning and dynamic adaptive intent understanding method based on human-computer collaboration as provided in the second aspect of the present invention.

[0022] Compared with the prior art, the present invention has the following beneficial effects: 1. Optimization of retrieval efficiency and accuracy: By combining the intent difficulty classification (coarse ranking) of the MN-BERT model with the fine ranking of the Rerank model, a highly efficient two-stage knowledge retrieval mechanism is formed. This not only greatly narrows the scope of precise retrieval and reduces computational costs, but also ensures the accuracy and relevance of retrieval results, effectively resolving the contradiction between efficiency and performance in traditional retrieval schemes.

[0023] 2. Deep Intelligent Reasoning Capability: Innovatively applying LLM to the "metacognitive" level of intent reflection and parameter derivation, enabling it to understand complex and implicit intents, perform knowledge transfer and contextual reasoning, and achieve intelligent filling of unknown parameters from known information, greatly improving the system's depth of intent understanding for non-standard and incomplete information queries.

[0024] 3. Dynamically Adaptive Interactive Experience: The proposed multi-turn dialogue guidance mechanism based on intent integrity assessment enables the system to proactively identify information gaps and generate the most effective guiding questions, much like a human expert. This dynamic strategy significantly improves the efficiency of information completion and the fluency of dialogue, providing a more intelligent and personalized interactive experience.

[0025] 4. Enhanced Human-Machine Collaboration Cycle: This invention constructs a "human-machine collaborative intelligent enhancement cycle." The system guides human users to supplement information through multi-round dialogues, while continuously learning and optimizing itself using this high-quality interactive data, achieving the co-evolution of system intelligence and human wisdom in the interaction process.

[0026] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0027] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a schematic diagram of the MN-BERT model architecture in one embodiment of the present invention.

[0029] Figure 2 This is a schematic diagram of a method for metacognitive reasoning and dynamic adaptive intent understanding based on human-computer collaboration in one embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0031] To make the embodiments of the present invention clearer, the terms involved in the present invention are now defined and explained, as shown in Table 1 below.

[0032] Table 1: Definitions and Explanations of Terms.

[0033]

[0034] Example 1: In one embodiment, a metacognitive reasoning and dynamic adaptive intent understanding system based on human-computer collaboration is provided. The system includes an input processing module, a two-stage knowledge retrieval module, an LLM reasoning module, an intent integrity assessment module, a multi-turn dialogue guidance module, and a continuous learning and knowledge base update module, wherein: The input processing module is used to receive and process user input information.

[0035] The two-stage knowledge retrieval module, connected to the input processing module, is used to retrieve candidate intents from the intent knowledge base based on the processed user input. The two-stage knowledge retrieval module includes a first-level retrieval submodule and a second-level retrieval submodule, specifically: The first-level retrieval submodule, which is built based on the MN-BERT model, is used to classify the difficulty of user intent, thereby narrowing the retrieval scope. The second-level retrieval submodule, which is built based on the Rerank model, is used to perform fine-grained retrieval on the subset of intent knowledge base filtered by the first-level retrieval submodule, and select one or more of the most relevant candidate intents.

[0036] Furthermore, in the two-stage knowledge retrieval module, the MN-BERT model used in the first-level retrieval submodule is an improved model based on the BERT model, with the architecture as follows: Figure 1 As shown. To facilitate understanding of the meaning of the symbols in the model diagram, they are defined as follows: This indicates the result after processing by the BERT model, corresponding to the input label. The final hidden layer output vector. A special marker indicating the start of the input sequence, and its corresponding output vector. or It is often used as an aggregate semantic representation of the entire input sentence. This indicates the result after processing by the BERT model, corresponding to the input label. The final embedding vector representation. Tok1, Tok2, ..., TokN This represents the first to Nth word units obtained after segmenting the original input text. ) respectively correspond to the input word unit Tok1, Tok2, ..., TokN The final embedding vector representation obtained after processing by the BERT model. Corresponding to the input word units respectively Tok1, Tok2, ..., TokN This is the final hidden layer output vector used in the BERT model processing for downstream tasks (such as named entity recognition). The downstream tasks during the model's pre-training phase include the Masked Language Modeling (MLM) task and the Named Entity Recognition (NER) task with custom intent parameter labels as entities, and its loss function... The loss function is the sum of the MLM task loss and the NER task loss. The expression is,

[0037] In the formula, For MLM task loss, For NER mission losses, For the set of words that are obscured, The total number of words that are masked. For the obscured word, This is the word representation obtained through the last layer of the BERT model; The number of words in the sentence. For the first One-hot encoding of the true label of each word. The model predicts the first The probability distribution of the labels for each word.

[0038] Furthermore, in the two-stage knowledge retrieval module, the reranking model used in the second-level retrieval submodule is a model based on the Cross Encoder architecture, and is trained using a two-stage training method including a general corpus training stage and a professional intent understanding corpus training stage; wherein, the positive examples of the professional intent understanding corpus are pairs of user questions and corresponding intents, and the negative examples are pairs of user questions and the top three erroneous intent samples with the highest scores calculated by the reranking model.

[0039] Furthermore, the re-ranking model is trained using a local contrastive estimation method, and its loss function is... The expression is,

[0040] In the formula, It is a query set. Is it related to query? Related document collection, The target search engine is for the query Found The top-ranked collection of documents, It is a positive sample. It is a query and documents Match scores between them.

[0041] Furthermore, the re-ranking model calculates the semantic similarity between the prompt and each intent description in the knowledge base, and the system identifies and filters out the three most relevant intents and their corresponding parameters as a candidate set.

[0042] The LLM reasoning module is connected to the two-stage knowledge retrieval module. It is used to receive the user input and candidate intent, perform metacognitive reasoning through the large language model LLM, determine the final target intent, and perform inference to complete the required parameters. Furthermore, in the LLM inference module, precise positioning of intent is achieved through parameter inference via two key steps: Intent selection reflection: The LLM comprehensively evaluates and compares the candidate intents of the input, and simulates metacognitive thinking to determine the optimal target intent; Parameter completion reasoning: For the selected target intent, the LLM compares the required parameters with the parameters already provided by the user, and uses contextual reasoning and knowledge transfer to deduce the missing parameters that the user has not explicitly provided.

[0043] The intent integrity assessment module is connected to the LLM inference module and is used to assess the integrity of the intent information output by the LLM inference module based on rules and machine learning methods. The multi-turn dialogue guidance module is connected to the intent integrity assessment module and is used to initiate an interaction process to generate guidance questions and conduct multi-turn dialogues with the user to complete the missing information when the intent information is determined to be incomplete. Furthermore, when the intent information is determined to be incomplete, the multi-turn dialogue guidance module initiates a multi-turn dialogue mechanism to implement an interactive intent information completion process, which includes the following steps: Step A: Identify missing parameters in the current intent information; Step B: Based on the missing parameters, the LLM dynamically generates one or more targeted guiding questions; Step C: Receive user feedback on the guided question and integrate the feedback information into the intent information; Step D: Return the updated intent information to the intent integrity assessment module for reassessment; Step E: Repeat steps A through D until the intent information reaches the preset integrity threshold.

[0044] The continuous learning and knowledge base update module is used to collect interaction data and optimize model parameters and update the intent knowledge base.

[0045] Example 2: In this embodiment, a metacognitive reasoning and dynamic adaptive intent understanding method based on human-computer collaboration is provided, such as... Figure 2 As shown, the method includes the following steps: Step S1: Initial input processing and intent candidate generation; Furthermore, the initial input processing and intent candidate generation include: Step S101: Receive the user's input question; Step S102: Use the MN-BERT model to classify the difficulty of user intent and determine the subset of the intent knowledge base to be retrieved; Step S103: Use the re-ranking model to perform fine-grained retrieval on the subset of the intent knowledge base, and select the top N most relevant candidate intents and their corresponding parameters, where N is an integer greater than or equal to 1.

[0046] Step S2: Intent Precision Localization and Parameter Inference Driven by LLM; Step S3: Intent information integrity assessment; Furthermore, the intent information integrity assessment involves the system evaluating the integrity of the intent information generated by the LLM (including the target intent, required parameters, user input parameters, and derived non-input parameters). The assessment employs a combination of rule-based and machine learning methods, comprehensively considering multi-dimensional indicators such as parameter coverage and confidence. If the intent information is deemed complete, the process terminates, and the system persistently stores the relevant information, providing a foundation for subsequent processing.

[0047] Step S4: If the evaluation result is incomplete, perform multiple rounds of interactive intent-guided completion. Furthermore, if the evaluation result is incomplete, multi-round interactive intent-guided completion is performed, including: Step S401: Identify missing information in the intent information; Step S402: The LLM generates one or more boot questions based on the missing information; Step S403: Receive the user's answer to the guidance question and integrate the new information into the intent information; Step S404: Repeat the assessment and guidance until the intent information is complete.

[0048] Step S5: Continuous learning and knowledge base updates.

[0049] Furthermore, the continuous learning and knowledge base update described in step S5 refers to the system continuously collecting user interaction data and feedback during operation. Through unsupervised learning and transfer learning techniques, the system can continuously optimize the re-ranking model, enrich the LLM's knowledge reserves, and dynamically update the intent knowledge base. This self-improvement mechanism ensures the long-term stability and adaptability of the system's performance.

[0050] Example 3: This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a metacognitive reasoning and dynamic adaptive intent understanding method based on human-computer collaboration as provided in the second aspect of the present invention.

[0051] Example 4: This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a metacognitive reasoning and dynamic adaptive intent understanding method based on human-computer collaboration as provided in the second aspect of the present invention.

[0052] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A metacognitive reasoning and dynamic adaptive intent understanding system based on human-computer collaboration, characterized in that, The system includes: The input processing module is used to receive and process user input information; A two-stage knowledge retrieval module, connected to the input processing module, is used to retrieve candidate intents from the intent knowledge base based on the processed user input. The two-stage knowledge retrieval module includes: The first-level retrieval submodule, built on the MN-BERT model, is used to classify the difficulty of user intent, thereby narrowing the search scope. The second-level retrieval submodule, which is built based on the Rerank model, is used to perform fine-grained retrieval of the intent knowledge base subset after being filtered by the first-level retrieval submodule, and to select one or more most relevant candidate intents. The LLM reasoning module is connected to the two-stage knowledge retrieval module. It is used to receive user input and candidate intents, perform metacognitive reasoning through the large language model LLM, determine the final target intent, and perform inference to complete its required parameters. An intent integrity assessment module, connected to the LLM inference module, is used to assess the integrity of the intent information output by the LLM inference module based on rules and machine learning methods. A multi-turn dialogue guidance module, connected to the intent integrity assessment module, is used to initiate an interaction process to generate guidance questions and conduct multi-turn dialogues with the user to complete the missing information when the intent information is determined to be incomplete. The continuous learning and knowledge base update module is used to collect interaction data and optimize model parameters and update the intent knowledge base.

2. The metacognitive reasoning and dynamic adaptive intent understanding system based on human-computer collaboration as described in claim 1, characterized in that: The MN-BERT model used in the first-level retrieval submodule is an improved version of the BERT model. Its downstream tasks during the pre-training phase include Masked Language Modeling (MLM) and Named Entity Recognition (NER) with custom intent parameter labels as entities. Its loss function... The loss function is the sum of the MLM task loss and the NER task loss. The expression is, In the formula, For MLM task loss, For NER mission losses, For the set of words that are obscured, The total number of words that are masked. For the obscured word, This is the word representation obtained through the last layer of the BERT model; The number of words in the sentence. For the first One-hot encoding of the true label of each word. The model predicts the first The probability distribution of the labels for each word.

3. The metacognitive reasoning and dynamic adaptive intent understanding system based on human-computer collaboration as described in claim 1, characterized in that: The re-ranking model used in the second-level retrieval submodule is a model based on the Cross Encoder architecture, which is trained using a two-stage training method including a general corpus training stage and a professional intent understanding corpus training stage. The positive examples of the professional intent understanding corpus are pairs of user questions and corresponding intents, while the negative examples are pairs of user questions and the highest-scoring incorrect intent samples calculated by the re-ranking model.

4. The metacognitive reasoning and dynamic adaptive intent understanding system based on human-computer collaboration as described in claim 3, characterized in that: The reordering model is trained using a local contrastive estimation method, and its loss function is... The expression is, In the formula, It is a query set. Is it related to query? Related document collection, The target search engine is for the query Found The top-ranked collection of documents, It is a positive sample. It is a query and documents Match scores between them.

5. The metacognitive reasoning and dynamic adaptive intent understanding system based on human-computer collaboration as described in claim 1, characterized in that, The LLM inference module achieves precise positioning of intent through parameter inference via two key steps: Intent selection reflection: The LLM comprehensively evaluates and compares the candidate intents of the input, and simulates metacognitive thinking to determine the optimal target intent; Parameter completion reasoning: For the selected target intent, the LLM compares the required parameters with the parameters already provided by the user, and uses contextual reasoning and knowledge transfer to deduce the missing parameters that the user has not explicitly provided.

6. The metacognitive reasoning and dynamic adaptive intent understanding system based on human-computer collaboration as described in claim 1, characterized in that, In the multi-turn dialogue guidance module, when the intent information is determined to be incomplete, the module initiates a multi-turn dialogue mechanism to implement an interactive intent information completion process, which includes: Step A: Identify missing parameters in the current intent information; Step B: Based on the missing parameters, the LLM dynamically generates one or more targeted guiding questions; Step C: Receive user feedback on the guided question and integrate the feedback information into the intent information; Step D: Return the updated intent information to the intent integrity assessment module for reassessment; Step E: Repeat steps A through D until the intent information reaches the preset integrity threshold.

7. A metacognitive reasoning and dynamic adaptive intent understanding method based on human-computer collaboration, applied to the system described in any one of claims 1 to 6, characterized in that, The method includes: Step S1: Initial input processing and intent candidate generation; Step S2: Intent Precision Localization and Parameter Inference Driven by LLM; Step S3: Intent information integrity assessment; Step S4: If the evaluation result is incomplete, perform multiple rounds of interactive intent-guided completion. Step S5: Continuous learning and knowledge base updates.

8. The metacognitive reasoning and dynamic adaptive intent understanding method based on human-computer collaboration as described in claim 7, characterized in that, The initial input processing and intent candidate generation include: Receive user-input questions; The MN-BERT model is used to classify the difficulty of user intent and determine the subset of the knowledge base for the retrieval intent. The intent knowledge base subset is finely ranked and retrieved using a re-ranking model to select the top N most relevant candidate intents and their corresponding parameters, where N is an integer greater than or equal to 1.

9. The metacognitive reasoning and dynamic adaptive intent understanding method based on human-computer collaboration as described in claim 7, characterized in that, If the evaluation result is incomplete, then multiple rounds of interactive intent-guided completion will be performed, including: Identify missing information in intent information; The LLM generates one or more boot questions based on the missing information; Receive user responses to guidance questions and integrate the new information into intent information; Repeat the assessment and guidance until the intent information is complete.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a metacognitive reasoning and dynamic adaptive intent understanding method based on human-computer collaboration as described in any one of claims 7 to 9.

11. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a human-computer collaborative metacognitive reasoning and dynamic adaptive intent understanding method as described in any one of claims 7 to 9.

Images