AI Agent intention recognition method and system and electronic equipment

Abstract

The invention provides an AI Agent intention recognition method and system and electronic equipment, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining input text data of a user, carrying out the multi-dimensional feature extraction and fusion of the input text data, and generating a comprehensive feature representation containing semantic, grammar and context information; performing parallel processing on the multiple intention understanding paths based on the comprehensive feature representation to obtain candidate intention information corresponding to each intention understanding path; performing multi-factor confidence fusion and dynamic calibration on all the candidate intention information, and determining the comprehensive confidence of each candidate intention information; matching is carried out according to the confidence degree of the candidate intention information, and a decision path matched with the candidate intention information is determined; and generating an intention recognition result for the user through the decision path matched with each piece of candidate intention information. According to the method, misjudgment and missed judgment are effectively reduced, a user can obtain more accurate and timely response, and the intention recognition accuracy is effectively enhanced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to an AI Agent intent recognition method, system, and electronic device. Background Technology

[0002] With the development of artificial intelligence technology, AI agents are widely used in fields such as intelligent customer service, smart homes, and intelligent assistants. Intent recognition, as the core link in AI agents' understanding of user needs, directly affects the user experience. Typically, most systems rely on a single pre-trained model such as BERT or GPT for intent classification and determine user intent through static thresholds.

[0003] In related technologies, single models have a low recognition rate for unfamiliar expressions and often ignore or fail to effectively utilize historical dialogue information when processing multi-turn dialogues, leading to errors in the recognition of intent switching or nested intents. At the same time, using a fixed confidence threshold to determine whether an intent is valid cannot adapt to the dynamic changes of different dialogue scenarios and user habits, resulting in misjudgments or omissions, which in turn affect the user intent recognition effect. Summary of the Invention

[0004] The problem addressed by this invention is how to improve the accuracy of recognizing user intent.

[0005] To address the aforementioned issues, this invention provides an AI Agent intent recognition method, system, and electronic device.

[0006] In a first aspect, the AI ​​Agent intent recognition method of the present invention includes: The system acquires user input text data and performs multi-dimensional feature extraction and fusion on the input text data to generate a comprehensive feature representation that includes semantic, syntactic and contextual information. By performing parallel processing on multiple intent understanding paths based on the comprehensive feature representation, candidate intent information corresponding to each intent understanding path is obtained; Multi-factor confidence fusion and dynamic calibration are performed on all the candidate intent information to determine the comprehensive confidence of each candidate intent information; Matching is performed based on the confidence level of the candidate intent information to determine a decision path that matches the candidate intent information; An intent recognition result for the user is generated by matching the decision path with each candidate intent information.

[0007] Optionally, the step of extracting and fusing multi-dimensional features from the input text data to generate a comprehensive feature representation containing semantic, syntactic, and contextual information includes: The input text data is subjected to standardization preprocessing to obtain normalized text data. The standardization preprocessing includes encoding unification, word segmentation, part-of-speech tagging, and named entity recognition. Feature extraction is performed on the standardized text data to obtain its semantic features, syntactic structure features, and sentiment features. The semantic features, the grammatical structure features, and the sentiment features are fused to generate the comprehensive feature representation.

[0008] Optionally, the intent understanding path includes a fuzzy matching path; the step of performing parallel processing on multiple intent understanding paths based on the comprehensive feature representation to obtain candidate intent information corresponding to each intent understanding path includes: Based on the semantic features in the comprehensive feature representation, the semantic similarity between the input text data and the historical intent template is obtained through the fuzzy matching path. Based on the semantic features and syntactic structure features in the comprehensive feature representation, the fuzzy membership degree of the input text data corresponding to multiple preset intent categories is obtained; Based on the semantic similarity and the fuzzy membership degree, a first intent probability distribution is obtained, and the first intent probability distribution is used as the candidate intent information of the fuzzy matching path.

[0009] Optionally, the intent understanding path includes an exact matching path; the step of processing multiple intent understanding paths in parallel based on the comprehensive feature representation to obtain candidate intent information corresponding to each intent understanding path includes: Based on the semantic features in the comprehensive feature representation, the deep semantic vector representation of the input text data is obtained through the precise matching path. Based on the deep semantic vector representation, the initial classification probability of the input text data corresponding to each preset intent category is obtained; Based on the syntactic structure features in the comprehensive feature representation, the initial classification probability is modified in a context-aware manner to obtain a second intent probability distribution, and the second intent probability distribution is used as the candidate intent information of the precise matching path.

[0010] Optionally, the intent understanding path includes a context-aware path; the step of processing multiple intent understanding paths in parallel based on the comprehensive feature representation to obtain candidate intent information corresponding to each intent understanding path includes: The context-aware path is used to obtain the context encoding vector of the user's historical dialogues. Based on the semantic features in the comprehensive feature representation and the context encoding vector, the relevance weight between the input text data and the historical dialogue context is obtained; The comprehensive feature representation is enhanced based on the relevance weights, and reasoning is performed in conjunction with a preset knowledge graph to obtain a third intent probability distribution, which is then used as the candidate intent information for the context-aware path.

[0011] Optionally, the step of performing multi-factor confidence fusion and dynamic calibration on all the candidate intent information to determine the comprehensive confidence of each candidate intent information includes: Based on the intent probability distribution corresponding to each candidate intent information, the base model prediction probability factor is obtained; Based on the user's historical dialogues, we obtain the context consistency factor and the historical pattern similarity factor; The initial confidence level is obtained by weighting and fusing the base model prediction probability factor, the context consistency factor, the historical pattern similarity factor, and the preset knowledge matching factor. The initial comprehensive confidence level is dynamically calibrated using a Bayesian update mechanism to obtain the comprehensive confidence level for each candidate intent information.

[0012] Optionally, the step of matching based on the confidence level of the candidate intent information to determine the decision path that matches the candidate intent information includes: The overall confidence level of each candidate intent information is matched and compared with preset high confidence thresholds and medium confidence thresholds; If the overall confidence level is greater than or equal to the high confidence threshold, then the high confidence decision path is taken as the decision path for the candidate intent information. If the overall confidence level is less than the high confidence threshold but greater than or equal to the medium confidence threshold, then the medium confidence decision path will be used as the decision path for the candidate intent information. If the overall confidence level is less than the medium confidence threshold, then the low confidence decision path will be used as the decision path for the candidate intent information.

[0013] Optionally, generating an intent recognition result for the user through the decision path matched with each candidate intent information includes: If the decision path is the high-confidence decision path, then an intent recognition result for the user is generated based on the intent probability distribution in the candidate intent information. If the decision path is the medium confidence decision path, then the intent probability distribution in the candidate intent information is optimized and reordered through a multi-model integrated voting process to obtain an optimized intent probability distribution; based on the optimized intent probability distribution, an intent recognition result for the user is generated. If the decision path is the low-confidence decision path, then based on the intent probability distribution in the candidate intent information, a clarification statement is generated to interact with the user and obtain user feedback text, and based on the user feedback text, an intent recognition result for the user is generated.

[0014] Secondly, the AI ​​Agent intent recognition system of the present invention includes: The data acquisition unit is used to acquire the user's input text data; The fusion unit is used to extract and fuse multi-dimensional features from the input text data to generate a comprehensive feature representation that includes semantic, syntactic and contextual information. The parallel processing unit is used to perform parallel processing based on the comprehensive feature representation through multiple intent understanding paths to obtain candidate intent information corresponding to each intent understanding path. The confidence determination unit is used to perform multi-factor confidence fusion and dynamic calibration on all the candidate intent information to determine the comprehensive confidence of each candidate intent information; A matching unit is used to match the candidate intent information based on the confidence level of the candidate intent information and determine a decision path that matches the candidate intent information. The generation unit is configured to generate an intent recognition result for the user by matching the decision path with each of the candidate intent information.

[0015] Thirdly, the electronic device of the present invention includes: a processor and a memory, the memory being used to store a computer program; When the computer program is loaded by the processor, it causes the processor to execute the AIAgent intent recognition method as described above. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the AI ​​Agent intent recognition method according to an embodiment of the present invention; Figure 2 This is a timing diagram of the AI ​​Agent intent recognition processing flow according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the AI ​​Agent intent recognition system according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0018] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0019] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0020] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0021] The names of the messages or information exchanged between the multiple devices in the embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of these messages or information.

[0022] Combination Figure 1 As shown in the figure, an AI Agent intent recognition method provided by an embodiment of the present invention includes: The system acquires user input text data and performs multi-dimensional feature extraction and fusion on the input text data to generate a comprehensive feature representation that includes semantic, syntactic and contextual information.

[0023] Specifically, the system first receives user input requests through a unified front-end interaction layer. These requests may originate from intelligent customer service, intelligent assistants, or various dialogue interfaces. The received text data undergoes standardized preprocessing, including character encoding to UTF-8 format, case conversion, and special character filtering. Subsequently, a multi-dimensional preprocessing module performs deep feature extraction. At the semantic level, pre-trained models such as BERT are used to map the text to a 768-dimensional or 1024-dimensional high-dimensional semantic space, generating feature vectors rich in semantic information. Simultaneously, a semantic density metric is calculated to assess the clarity of expression. At the grammatical level, word segmentation tools are used for part-of-speech tagging and named entity recognition, and a syntactic dependency tree is constructed to analyze sentence structure. At the contextual level, the most recent N rounds of historical dialogue records (where N≤10) are extracted, and dialogue state vectors are generated through temporal model encoding, with a decay mechanism assigning different weights to different time steps. Finally, a feature fusion algorithm integrates semantic vectors, grammatical structure, and sentiment polarity obtained from a sentiment detector into a unified comprehensive feature representation, providing a comprehensive input foundation for subsequent processing.

[0024] By performing parallel processing on multiple intent understanding paths based on the comprehensive feature representation, candidate intent information corresponding to each intent understanding path is obtained.

[0025] Specifically, a Diamond parallel processing architecture is adopted, simultaneously launching five independent intent understanding paths for parallel processing. Specifically, the fuzzy matching path achieves imprecise matching based on fuzzy logic and semantic similarity algorithms, handling low-confidence cases by constructing an intent fuzzy relation matrix R, where elements R[i,j] represent the semantic relevance between intent i and intent j. The precise matching path uses a deep learning model combining LSTM and attention mechanisms to capture long-distance dependencies in text sequences and performs hierarchical matching through intent trees. The context-aware path achieves contextual reasoning through attention mechanisms and knowledge graphs, establishing a dynamically updated context vector pool. The LSTM analysis path uses a bidirectional LSTM-CRF model for sequence labeling, extracting structured intent information. The rule engine path performs rule reasoning based on a domain knowledge base. Each path runs independently, analyzing user intent from different perspectives and outputting its own candidate intents and initial confidence levels, achieving multi-path collaborative verification and intent coverage. This parallel architecture improves processing efficiency and enhances system robustness.

[0026] Multi-factor confidence fusion and dynamic calibration are performed on all the candidate intent information to determine the comprehensive confidence of each candidate intent information.

[0027] Specifically, a dynamic confidence calibrator comprehensively evaluates the candidate intents generated by all paths. The confidence calculator employs a weighted fusion strategy, comprehensively considering four core factors: base model prediction probability (P_model), context consistency score (P_context), knowledge matching degree (P_knowledge), and historical pattern similarity (P_history). The weights (w1-w4) of each factor are dynamically adjusted according to the real-time scenario; for example, in scenarios with ambiguous expressions, the model weight is reduced while the context weight is increased. Based on this, a temperature scaling technique is used to calibrate the original confidence score, improving the reliability of the probability output. Simultaneously, a Bayesian update mechanism is introduced, dynamically adjusting the confidence score distribution by incorporating prior knowledge such as historical success rates. The final calibrated comprehensive confidence score not only reflects the reliability of the current recognition but also incorporates historical experience and scenario features, providing a precise basis for hierarchical decision-making.

[0028] Matching is performed based on the confidence level of the candidate intent information to determine the decision path that matches the candidate intent information.

[0029] Specifically, by setting three levels of confidence thresholds—high ≥ 0.85, medium 0.6-0.85, and low < 0.6—a tiered processing approach is adopted. High-confidence intents directly output results and update the intent cache, employing a fast response path. Medium-confidence intents initiate an enhanced processing flow, optimizing through multi-model ensemble inference and semantic re-matching, and using weighted voting or bundle search algorithms to select the optimal intent. Low-confidence intents trigger a three-level strategy for intelligently resetting the processing layer. First, the hint chain is reconstructed and re-analyzed; second, the processing priority is adjusted by reordering modules; and finally, an interactive confirmation mechanism is initiated to proactively clarify the issue to the user. Ultimately, the results of all paths are integrated through a conflict resolution strategy to generate a structured output containing the identified intent, confidence level, relevant parameters, and contextual labels. Simultaneously, the feedback learning module continuously collects user feedback, dynamically adjusting thresholds and weight configurations to form a complete closed-loop optimization system from user interaction to model updates.

[0030] An intent recognition result for the user is generated by matching the decision path with each candidate intent information.

[0031] Specifically, for high-confidence paths (confidence ≥ 0.85), the system directly outputs the intent recognition result and updates the intent cache, recording the currently successfully recognized pattern for subsequent optimization. The output result contains complete structured information, including the identified intent type, post-calibration confidence level, intent hierarchy path, extracted slot information such as product type, purchase time, and problem type, contextual tags such as conversation stage, urgency, user sentiment, and recommended action suggestions. For example, in an intelligent customer service scenario, after the system identifies a product quality complaint intent, it generates a structured response containing an apology template and solution options.

[0032] For paths with medium confidence (between 0.6 and 0.85), the system initiates an enhancement processing flow, optimizing the results through multi-model ensemble inference and semantic re-matching. Specifically, a weighted voting strategy is used to fuse the outputs of multiple weak classifiers, and a re-ranking based on a greedy algorithm or bundle search is performed to select the optimal intent combination. The system expands the query scope, such as re-matching vague expressions as "product feature failure complaints," thereby improving matching accuracy through semantic expansion. The optimized results undergo a second confidence assessment; those reaching high confidence are transferred to the fast output channel, otherwise, the enhanced processing marker is retained for subsequent analysis.

[0033] For low-confidence paths (confidence < 0.6), the system activates a three-tiered intelligent reset processing layer strategy: First, it reconstructs the cue chain, reorganizes the preprocessing flow, and adjusts the focus of feature extraction; second, it reorders execution modules, changes the priority allocation of parallel processing paths, and strengthens the weight of contextual inference paths; finally, it initiates an interactive confirmation mechanism, generating targeted clarification questions to proactively obtain supplementary information from the user. For example, when identifying a vague complaint, the system will generate a clarification question: "Are you dissatisfied with the product quality, delivery, or service?" and re-match intent based on user feedback.

[0034] Finally, the outputs of the three paths are integrated through the result fusion module, and a conflict resolution strategy is used to resolve conflicts between multiple candidate intents. In specific implementation, arbitration is performed based on the inheritance relationship of the intent hierarchy tree, context consistency score, and historical success rate. The generated structured output not only includes the finally determined intent information, but also processing metadata: total latency, time consumed by each module, triggered reset strategy, processing path type, and cache hit status. Simultaneously, the feedback learning module collects recognition results and user feedback from actual applications in real time, samples samples with high uncertainty based on the active learning mechanism for manual annotation, dynamically updates model parameters through incremental training, and continuously optimizes the knowledge base content, forming a complete closed loop from intent recognition to system evolution.

[0035] In a preferred embodiment of the present invention, the above-mentioned AI Agent intent recognition method can be implemented through a multilingual text processing system based on modular design. The technical architecture of the multilingual text processing system includes the following core components: a front-end interaction layer is used to receive multilingual text input through a unified user access interface and perform initialization verification; a semantic space mapping module is responsible for mapping the input text to a high-dimensional semantic feature space; a parallel processing engine adopts a three-way parallel processing architecture: a fuzzy matching branch: based on fuzzy logic and semantic similarity algorithms, it achieves imprecise matching; a precise matching branch: through deep learning models such as LSTM and Transformer, it achieves semantic-level precise matching; a context association branch: using attention mechanisms and knowledge graphs, it achieves context-related reasoning; and a result fusion and optimization module is used for inverse semantic restoration, mapping the processing results from the feature space back to the natural language space, for multi-path fusion, using weighted voting or ensemble learning methods to fuse the three-way processing results, and iterative optimization using a feedback mechanism based on confidence evaluation to achieve iterative optimization of the results.

[0036] In another preferred embodiment of the present invention, the original text input by the user is first uniformly encoded and converted into the standard UTF-8 encoding format. At the same time, the case of English characters is standardized, and special characters and meaningless symbols are filtered or converted. Simultaneously, a word segmentation tool based on statistics and deep learning is used to segment the text and perform part-of-speech tagging and named entity recognition to extract key linguistic features. Furthermore, a multi-granularity text representation is constructed, including character-level, word-level, and sentence-level vector representations, to provide a multi-level feature foundation for subsequent semantic modeling. For example, in a scenario where a user inputs, "This stuff is really too much, I don't even know what to say," the system uses a word segmentation tool (such as HanLP) based on the BiLSTM-CRF deep learning model to intelligently segment the ambiguous complaint text. It accurately identifies "enough" as a complete sentiment adjective unit rather than incorrectly segmenting it into discrete characters. At the same time, it segments "things" into independent noun units. This accurate word segmentation result lays a structured foundation for the subsequent part-of-speech tagging module to label it as an adjective (VA) and a noun (NN), the named entity recognition module to classify "things" as a product entity (PRODUCT), and the semantic space construction module to calculate the sentiment vector of "enough." Ultimately, this supports the ambiguous intent disambiguation module to correctly parse the focus of the user's complaint, thereby achieving accurate identification of ambiguous complaint intent.

[0037] For example, an e-commerce platform's intelligent customer service system receives a vague complaint from user U987654321 in dialogue session S20240115_001: "This stuff is really enough, I don't even know what to say anymore." At this point, the system needs to accurately identify the user's intent and provide an appropriate response. Specifically, regarding the vague complaint: "This stuff is really enough, I don't even know what to say anymore," the system performs deep analysis of the text using multi-dimensional feature extraction technology. At the semantic level, it detects low semantic density (0.42), indicating ambiguity; at the emotional level, it identifies strong negative emotions (frustration 0.82). Simultaneously, it combines this with contextual information from a phone purchase three days prior in historical dialogues to construct a comprehensive feature representation encompassing semantic, emotional, and contextual features. Based on this comprehensive feature representation, the system initiates multiple intent understanding paths in parallel: the fuzzy matching path calculates the probability of a product quality complaint as 0.68 using membership degrees; the LSTM precise path uses an attention mechanism to highlight the keyword "enough"; and the contextual path combines purchase records to strengthen the reasoning about product quality issues. After generating candidate intents with varying confidence levels for each of the three paths, the system employs a multi-factor fusion strategy for dynamic calibration. It considers four dimensions: model prediction probability (0.67), contextual consistency (0.85), knowledge matching degree (0.72), and historical patterns (0.78), and obtains a calibrated confidence level of 0.82 through Bayesian updates. Since the confidence level is in the intermediate range, the system initiates an enhancement processing mechanism, fusing the outputs of multiple models and performing semantic re-matching, ultimately raising the confidence level to 0.87 and confirming the intent graph as a product quality complaint. The system generates a structured response suggesting customer service inquire about specific details, while simultaneously recording the successfully identified pattern. When the user subsequently confirms the intent, the feedback learning module automatically adjusts the weight configuration, increasing the contextual importance from 0.35 to 0.38, and adding features combining fuzzy expressions and contextual clues to the successful pattern library, achieving online optimization and continuous evolution.

[0038] In another example, in a multi-turn dialogue scenario between a user and a smart home assistant, the user first requests: "Turn on the living room lights," then cancels, saying: "No, never mind," and finally issues a compound command: "Wait, turn it to warm light, and check the weather tomorrow." When processing the third round of complex input, the system identifies two independent intent segments through multi-dimensional preprocessing: the device control intent to turn on warm light and the information query intent to check the weather tomorrow. Semantic analysis shows the two intent densities are 0.85 and 0.82 respectively, syntactic analysis confirms a parallel relationship, and sentiment detection indicates the user is in a decisive state. The system then initiates a diamond-shaped parallel architecture to simultaneously process five paths: a fuzzy matching path identifies light adjustment (similarity 0.94) and weather query (0.91) through template matching; an LSTM attention path assigns different attention weights to the two intents; a context-aware path tracks the complete logical flow of the dialogue state from turning on the lights to canceling and resuming modification; a sequence labeling path accurately labels warm light as a color entity and tomorrow as a time entity; and a rule engine path triggers the sequential execution of rules for multiple intents. All five results confirm the existence of two independent intents: adjusting the light color and querying the weather. In the confidence assessment phase, multi-dimensional scores were calculated for both intents. The intent to adjust lighting received a model score of 0.91-0.95, reaching 0.98 after context enhancement, 0.89 for knowledge matching, and 0.93 for historical patterns. After Bayesian calibration and temperature scaling, the overall confidence score was 0.96. The intent to query the weather scored 0.88-0.93, with a weaker contextual association at only 0.90, resulting in a final confidence score of 0.91. Since the confidence scores for both intents exceeded the high threshold of 0.85, the system directly executed the fast track. First, it inherited the device information from the first round of dialogue—the living room light—and inferred that the device needed to be turned on before adjusting the color to 3000K warm light. Simultaneously, it queried the weather forecast for Haidian District, Beijing, for tomorrow. The system generated a composite response: "Turning on the living room light and adjusting it to warm light. Tomorrow, Haidian District, Beijing will be cloudy turning sunny, with temperatures ranging from 2-10℃." After processing, the system identified the behavioral pattern of adjusting warm light + checking the weather in the evening, proactively suggested creating a routine, and added this multi-intent sequence pattern to the knowledge base, completing a full closed loop from accurate identification to personalized service.

[0039] The AI ​​Agent intent recognition method in this embodiment acquires user input text data and then extracts and fuses multi-dimensional features to generate a comprehensive feature representation that includes semantic, syntactic, and contextual information. Semantic information helps understand the core meaning of the text, syntactic information helps analyze the text structure, and contextual information considers historical dialogue content, thus more comprehensively grasping the user's intent. For example, in a multi-turn dialogue, the user may have mentioned some key information in the earlier rounds, which is crucial for understanding the intent of subsequent dialogues. By comprehensively considering these multi-dimensional features, the user's true intent can be captured more accurately, improving the accuracy of intent recognition compared to single-dimensional feature extraction. This embodiment employs multiple intent understanding paths for parallel processing based on the comprehensive feature representation to obtain candidate intent information corresponding to each intent understanding path. This parallel processing approach is equivalent to understanding and analyzing the user's input text from multiple perspectives. Different paths will focus on different features or processing methods, thereby uncovering more possible intent information. For example, one path may focus more on semantic analysis, while another path may focus more on contextual relationships. In this way, various possible intents can be more comprehensively covered, providing more basis for subsequent accurate judgment and effectively improving the accuracy of intent recognition.

[0040] Furthermore, this embodiment performs multi-factor confidence fusion and dynamic calibration on all candidate intent information to determine the comprehensive confidence level of each candidate intent information. Multi-factor fusion considers the impact of various factors on intent confidence, such as the matching degree between intent and context, and the semantic clarity of intent, rather than simply relying on a fixed threshold. Dynamic calibration can adjust the confidence judgment criteria in real time according to specific dialogue scenarios and user habits. For example, in a more formal customer service dialogue scenario, the requirement for intent confidence will be higher; while in a more casual intelligent assistant dialogue scenario, the confidence judgment criteria will be relatively more flexible. In this way, the reliability of each candidate intent information can be judged more accurately, thereby more accurately determining the user's intent and further improving the accuracy of intent recognition. After determining the comprehensive confidence level of each candidate intent information, matching is performed based on the confidence level of the candidate intent information to determine the matching decision path, and the user-specific intent recognition result is generated through the matching decision path of each candidate intent information. This process is equivalent to filtering and optimizing candidate intent information, selecting the decision path that best matches the current dialogue scenario and user intent, thereby generating the most accurate intent recognition result. Compared to traditional methods that directly judge based on a single confidence threshold, this method can better consider various factors, generate intent recognition results that better meet the user's actual needs, and effectively improve the accuracy of intent recognition.

[0041] Meanwhile, in multi-turn dialogues, users may express complex intentions, such as intent switching or nested intents. This embodiment, through multi-dimensional feature extraction and fusion, considers contextual information, enabling better understanding and processing of complex intents in multi-turn dialogues. Furthermore, parallel processing of multiple intent understanding paths, multi-factor confidence fusion, and dynamic calibration techniques can more accurately identify these complex intents. For example, when interacting with a smart home system, a user might first inquire about the weather and then decide whether to turn on the air conditioner based on the weather. This intent switching is more accurately identified in this embodiment, providing a more coherent and logical dialogue experience and improving user satisfaction with the AI ​​Agent. Different users may have different expression habits and dialogue styles when communicating with the AI ​​Agent, and different dialogue scenarios have different requirements for intent recognition. The multi-factor confidence fusion and dynamic calibration technology in this embodiment can adjust the intent recognition strategy and judgment criteria in real time according to different user habits and dialogue scenarios. For example, for users who prefer colloquial and vague expressions, the system can more flexibly identify their intentions through dynamic calibration; for more specialized dialogue scenarios, such as technical inquiries in intelligent customer service, the system can more accurately identify professional intentions by increasing the confidence level requirements. This adaptability allows the AI ​​Agent to better meet the needs of different users and improve the user experience in various scenarios.

[0042] In summary, this embodiment effectively reduces false positives and false negatives, enabling users to receive more accurate and timely responses, enhancing user trust in the AI ​​Agent and their willingness to use it, thereby optimizing the overall user experience. Simultaneously, the system can integrate multi-dimensional features to deeply understand user intent from different perspectives, making it more adaptable and flexible in complex dialogue scenarios and diverse user needs. Dynamic calibration technology further strengthens this adaptability, enabling the system to adjust its strategies in real time according to different scenarios and user habits, improving its level of intelligence.

[0043] Optionally, the step of extracting and fusing multi-dimensional features from the input text data to generate a comprehensive feature representation containing semantic, syntactic, and contextual information includes: The input text data is subjected to standardization preprocessing to obtain normalized text data. The standardization preprocessing includes encoding unification, word segmentation, part-of-speech tagging, and named entity recognition. Feature extraction is performed on the standardized text data to obtain its semantic features, syntactic structure features, and sentiment features. The semantic features, the grammatical structure features, and the sentiment features are fused to generate the comprehensive feature representation.

[0044] Specifically, the standardization preprocessing module first receives raw user input text from different channels. The system performs a unified encoding operation, converting various character encodings to the standard UTF-8 format to ensure encoding consistency in subsequent processing. Next, word segmentation is performed, using a deep learning-based word segmentation tool to accurately segment Chinese text, while standardizing word segmentation and unifying capitalization for English text. In the part-of-speech tagging stage, a pre-trained sequence tagging model is used to label the grammatical role of each word, such as noun, verb, adjective, etc., to help understand sentence structure. The named entity recognition module runs concurrently, identifying specific entities in the text such as names of people, places, organizations, and products, and linking these entities to corresponding nodes in the knowledge graph. The entire preprocessing process also includes special character filtering, removal of meaningless symbols, and text normalization, ultimately generating structured and standardized text data to provide standardized input for subsequent feature extraction.

[0045] Furthermore, based on standardized text data, the system performs feature extraction in parallel across three dimensions. Semantic feature extraction employs a pre-trained language model such as BERT to map the segmented text into a high-dimensional vector space, generating 768-dimensional or 1024-dimensional semantic vectors. This vector not only captures the contextual meaning of words but also understands the semantic relationships within sentences through a multi-layered attention mechanism. Simultaneously, the system calculates a semantic density index to quantify the clarity of the text's expression. Grammatical structure feature extraction uses a dependency parser to construct a sentence dependency tree, identifying core components such as subject, verb, and object, and their modification relationships, extracting syntactic pattern features and grammatical structure labels. Sentiment feature extraction uses a sentiment analysis model to identify the text's emotional polarity (positive, negative, neutral), emotional intensity (0-1 points), and specific emotional types (anger, joy, disappointment, etc.). Combined with a tone detection module, it analyzes the urgency and interactive intent of the expression. These three feature dimensions comprehensively depict the deeper meaning of user input from different perspectives.

[0046] In the feature fusion stage, the system employs a multimodal fusion strategy to integrate feature information from three dimensions. First, the semantic vector, syntactic feature vector, and sentiment vector are dimensionally aligned and standardized to ensure they are on the same scale. Then, an attention-weighted fusion mechanism is used to dynamically adjust the weight ratio of each feature according to the current processing scenario: syntactic structure is given higher weight when handling technical issues; sentiment features are given increased weight in customer service dialogue scenarios; and semantic features are emphasized in information query scenarios. In practice, a cross-attention mechanism is used to calculate the correlation matrix between features, achieving deep interactive fusion of features. The final comprehensive feature representation is a unified high-dimensional vector that simultaneously contains semantic intent, syntactic structural constraints, and sentiment-level user state information. This comprehensive feature representation serves as the unified input for the subsequent parallel processing engine, providing a comprehensive and accurate feature foundation for multi-path intent understanding.

[0047] In this optional embodiment, a unified feature representation system is constructed through multi-level standardized preprocessing and multi-dimensional feature collaborative extraction. This system can reflect the deep semantics of the text, capture the language structure rules and the user's emotional state. It effectively solves the problems of semantic ambiguity, structural misjudgment and emotional neglect caused by the reliance on a single feature in traditional intent recognition. Specifically, standardized preprocessing establishes a unified text analysis foundation and eliminates coding differences and noise interference. The three-way parallel feature extraction obtains complementary feature information from the three dimensions of semantic understanding, syntactic parsing and sentiment analysis. The intelligent fusion mechanism realizes the organic integration of different feature dimensions through dynamic weight allocation and attention interaction. This enables the generated comprehensive feature representation to simultaneously carry complete information about what the user wants to express, how to express it, and with what emotion. This provides a high-information-density and high-discrimination input foundation for subsequent multi-path intent recognition, fundamentally improving the system's ability to understand complex expressions, ambiguous intents and multi-emotional texts.

[0048] Optionally, the intent understanding path includes a fuzzy matching path; the step of performing parallel processing on multiple intent understanding paths based on the comprehensive feature representation to obtain candidate intent information corresponding to each intent understanding path includes: Based on the semantic features in the comprehensive feature representation, the semantic similarity between the input text data and the historical intent template is obtained through the fuzzy matching path. Based on the semantic features and syntactic structure features in the comprehensive feature representation, the fuzzy membership degree of the input text data corresponding to multiple preset intent categories is obtained; Based on the semantic similarity and the fuzzy membership degree, a first intent probability distribution is obtained, and the first intent probability distribution is used as the candidate intent information of the fuzzy matching path.

[0049] Specifically, the fuzzy matching path is based on semantic feature vectors extracted from the comprehensive feature representation, typically high-dimensional vectors of 768 or 1024 dimensions, which are then compared with the similarity of these vectors to the historical intent template library stored in the system. The system maintains a dynamically updated intent template database containing validated standard intent expressions and their corresponding semantic vector representations for various scenarios. By calculating the cosine similarity or Euclidean distance between the input text semantic vector and each template vector, the system obtains multiple sets of similarity scores.

[0050] For example, in intelligent customer service scenarios, vague expressions like "enough is enough" might generate similarity matches between 0.65 and 0.85 with multiple templates such as "quality problem complaint," "service dissatisfaction," and "product defect." Nearest neighbor algorithms or clustering analysis are used to filter out candidate templates with similarity scores higher than a threshold (e.g., 0.6), serving as the basis for initial intent matching. Simultaneously, semantic and syntactic structural features are further combined, and fuzzy logic theory is used to softly classify the text into intent categories. In specific implementation, a membership function is defined for each preset intent category, such as "product quality complaint," "return application," "exchange request," and "general dissatisfaction." These functions comprehensively consider semantic relevance, syntactic pattern matching, and contextual fit. During the calculation process, the system analyzes the syntactic structural features of the input text, such as sentence types (declarative, interrogative, exclamatory), core verbs (e.g., complaint, request, inquiry), and modifiers (e.g., negative adjectives, degree adverbs). Combined with the keyword weight distribution in semantic features, the membership value of the text to each intent category is calculated. For example, in the input "This stuff is really too much", the system calculates that the membership degree for "product quality complaint" is 0.68 and the membership degree for "general dissatisfaction" is 0.82, forming a continuous fuzzy judgment rather than a binary classification.

[0051] Finally, the semantic similarity results and fuzzy membership results are fused to generate the first intent probability distribution. In the specific implementation, a weighted fusion algorithm is used, where semantic similarity reflects the closeness of the text to the standard expression, and fuzzy membership reflects the text's tendency to belong to a category boundary. The system uses an adaptive weight adjustment mechanism to assign higher weight to semantic similarity when processing clear expressions and increase the weight of fuzzy membership when processing fuzzy expressions. The fused probability distribution presents the confidence level of each candidate intent. These are sorted according to their probability values, and the Top-K, typically K=3-5 intents, are selected as the output of the fuzzy matching path. In this embodiment, the fuzzy matching path ultimately outputs the intent distributions for "product quality complaint" (probability 0.68), "general dissatisfaction" (0.62), and "return application" (0.45). These results will be collaboratively verified and integrated with the output of the subsequent precise matching path for decision-making.

[0052] In this optional embodiment, a dual verification mechanism combining semantic similarity calculation and fuzzy logic reasoning is employed to construct an intent recognition capability that effectively handles non-standard expressions, ambiguous descriptions, and boundary cases, significantly improving the system's adaptability to imprecise user input. Specifically, semantic similarity calculation compares the current input with a historical intent template library, ensuring the correlation between the recognition result and verified patterns; fuzzy membership analysis characterizes intent boundaries through continuous probability distributions, achieving soft discrimination of ambiguous expressions; the weighted fusion of these two methods preserves the stability of pattern matching while also considering the flexibility of contextual understanding, enabling the system to generate a reasonable set of candidate intents even when user expressions are incomplete, unclear, or ambiguous. This design in this embodiment not only improves the recognition accuracy in fuzzy scenarios but also provides a diverse input foundation for subsequent multi-path collaborative decision-making, enhancing the system's robustness in handling complex linguistic phenomena.

[0053] Optionally, the intent understanding path includes an exact matching path; the step of processing multiple intent understanding paths in parallel based on the comprehensive feature representation to obtain candidate intent information corresponding to each intent understanding path includes: Based on the semantic features in the comprehensive feature representation, the deep semantic vector representation of the input text data is obtained through the precise matching path. Based on the deep semantic vector representation, the initial classification probability of the input text data corresponding to each preset intent category is obtained; Based on the syntactic structure features in the comprehensive feature representation, the initial classification probability is modified in a context-aware manner to obtain a second intent probability distribution, and the second intent probability distribution is used as the candidate intent information of the precise matching path.

[0054] Specifically, the precise matching path first extracts the semantic feature portion from the comprehensive feature representation. This feature, a high-dimensional vector containing rich contextual information, is further refined using a deep learning encoder, such as a Transformer-based deep neural network. Through multi-layered self-attention mechanisms and residual connections, the network can capture long-distance dependencies between words, distinguishing between core semantic information and decorative content. In the intelligent assistant scenario, when processing compound sentences like "Let's turn on the lights to warmer" and "Let's check the weather tomorrow," the encoder generates highly abstract representations of two semantic fragments: "Turn on the lights to warmer" and "Tomorrow's weather," while preserving the logical relationship between them. This deep semantic vector representation in this embodiment not only encodes lexical meaning but also includes sentence-level semantic structure and discourse-level semantic associations. Further, the deep semantic vector is input into a multi-layered classification network for intent classification. This classification network typically includes fully connected layers and a softmax output layer, outputting probability distributions for a preset set of intent categories, such as device control, information query, and service request. In specific implementations, the system employs a multi-label classification strategy, allowing one input to correspond to multiple intent categories. For example, in compound instruction scenarios, adjusting to warm light corresponds to the device control category, and checking tomorrow's weather corresponds to the information query category; both receive high probability values. The initial classification probability directly reflects the relative distance between the deep semantic vector and the prototype of each intent category in the semantic space, providing a basic score for subsequent accurate judgment. The initial classification probability is further corrected and optimized using syntactic structural features in the comprehensive feature representation. In specific implementation, the system analyzes syntactic dependency relations, phrase structure, and sentence type information to identify syntactic constraints.

[0055] For example, in a smart assistant scenario, the system detects that "adjusting to warm light" is a verb-object structure, strengthening the confidence in the device control category. It also identifies the conjunction "by the way," indicating an additional rather than a substitution relationship between the two intentions. The system further performs cross-turn consistency verification by incorporating grammatical patterns from historical dialogues, ensuring that the current intention remains grammatically coherent with the dialogue history. Through these grammatical analyses, the system weights and adjusts the initial classification probabilities to form a final second intention probability distribution. This distribution reflects both semantic matching and grammatical constraints, achieving more accurate intention recognition.

[0056] In a preferred embodiment of the present invention, the precise intent recognition module is activated when it detects that the user input has a high initial confidence level (e.g., ≥0.85). It performs deep semantic analysis of the text using a fusion model of LSTM and an attention mechanism. First, it uses a Long Short-Term Memory (LSTM) network to process the input sequence step-by-step, generating a hidden state h_t at each time step t to capture long-distance dependencies in the text. Then, an attention mechanism is introduced, calculating the attention weight α_t for each hidden state using learnable parameter matrices W_a, W_h, and a bias b_h, and normalizing it using a softmax function, enabling the model to automatically focus on key semantic segments such as "enough" and "warm light". Finally, the hidden states at all time steps are weighted and summed according to the attention weights to obtain a context vector c that gathers global key information, serving as a high-quality semantic representation for subsequent intent classification and slot extraction, thereby achieving the transformation from precise sequence understanding to explicit intent determination.

[0057] In this optional embodiment, a high-level intent understanding capability is constructed through a two-layer optimization mechanism of deep semantic encoding and grammatical structure correction. This capability can accurately parse complex expressions, identify nested intents, and handle multi-turn dialogues, significantly improving the system's recognition accuracy for structurally complete and semantically explicit expressions. Specifically, deep semantic vector representation extracts deep semantic features from the text through a multi-layer neural network, achieving accurate characterization of complex semantic relationships and logical structures. Initial classification probability calculation provides an accurate benchmark assessment for intent recognition. The grammatical structure feature correction mechanism further integrates syntactic constraints and contextual consistency into the decision-making process, ensuring that the recognition results not only conform to semantic logic but also meet grammatical rules and dialogue coherence requirements. This two-layer processing architecture enables the system to maintain high-precision recognition performance when processing compound instructions, nested intents, and multi-turn dialogues. It performs particularly well in scenarios requiring accurate understanding of complete user expressions, such as intelligent assistants and complex customer service, effectively compensating for comprehension biases that may arise from single semantic matching.

[0058] Optionally, the intent understanding path includes a context-aware path; the step of processing multiple intent understanding paths in parallel based on the comprehensive feature representation to obtain candidate intent information corresponding to each intent understanding path includes: The context-aware path is used to obtain the context encoding vector of the user's historical dialogues. Based on the semantic features in the comprehensive feature representation and the context encoding vector, the relevance weight between the input text data and the historical dialogue context is obtained; The comprehensive feature representation is enhanced based on the relevance weights, and reasoning is performed in conjunction with a preset knowledge graph to obtain a third intent probability distribution, which is then used as the candidate intent information for the context-aware path.

[0059] Specifically, the context-aware path first retrieves and integrates the user's historical dialogue records in the current session, typically including the dialogue content from the most recent N rounds (e.g., 5-10 rounds), each round containing user input and system response. The system uses a sequence encoder, such as an LSTM or Transformer encoder, to encode these historical dialogues, compressing each round into a fixed-dimensional dialogue state vector. During the encoding process, the system introduces a temporal decay mechanism, assigning different weights to dialogues at different points in time, with recent dialogues having higher weights than earlier ones, to reflect the temporal evolution of the dialogue focus. In the intelligent assistant scenario, the user's first round, "Please turn on the living room lights," and the third round, "Please adjust the lights to warmer," are encoded as a temporally related dialogue state sequence, forming a context encoding vector that reflects the dialogue's evolution. By calculating the dot product similarity between the current semantic vector and each historical dialogue state vector, a set of attention weight values ​​is obtained. These weights reflect the degree of correlation between the current input and various parts of the historical dialogue. For example, in the intelligent customer service scenario, the user's current vague complaint "I'm fed up" will have a high relevance weight to information in the historical dialogue such as "I bought a phone 3 days ago" or "I received it yesterday." The system uses a softmax function to convert these similarities into a normalized attention distribution. Historical dialogue fragments with higher weights have a greater impact on the understanding of the current intent. This mechanism allows the system to dynamically focus on the most relevant historical information, avoiding interference from irrelevant contextual information. Historical contextual information is fused into the current feature representation through weighted summation, forming a context-enhanced feature vector. Simultaneously, the system accesses a domain knowledge graph to query concept nodes and relational edges related to the current semantics. In the intelligent assistant scenario, the system queries knowledge associations such as warm light and light color adjustment, and living room lights and smart home devices. Based on graph neural networks or rule-based reasoning engines, the system extracts relevant reasoning clues from the knowledge graph, such as warm light typically being used at night and living room lights supporting color adjustment. Combining the context-enhanced features with the knowledge reasoning results, a third intent probability distribution is calculated through a multilayer perceptron network. This distribution reflects a comprehensive intent judgment based on historical dialogue context and domain knowledge, providing the system with important candidate intent information.

[0060] In this optional embodiment, a context-aware intent recognition capability with cross-turn memory and domain knowledge support is constructed through the deep integration of historical dialogue encoding and knowledge graph reasoning. This significantly improves the system's understanding coherence and accuracy in multi-turn dialogues, complex tasks, and domain-specific scenarios. Specifically, the context encoding vector establishes a continuous representation of the dialogue state by temporally modeling historical dialogues, enabling the system to remember and understand the dialogue evolution process. The dynamic calculation mechanism of relevance weights enables intelligent filtering of historical information, ensuring that the system focuses on the context fragments most relevant to the current input. The introduction of knowledge graphs provides domain common sense and logical constraints for intent reasoning, placing user expressions within a richer semantic network for understanding. This design allows the system to maintain high-precision intent recognition even when user expressions rely on previous dialogues, involve professional knowledge, or require logical reasoning. It effectively solves the problems of intent fragmentation, information forgetting, and lack of domain knowledge that often occur in cross-turn dialogues using traditional methods, significantly improving the continuous interaction capability and task completion rate of the intelligent dialogue system.

[0061] Optionally, the step of performing multi-factor confidence fusion and dynamic calibration on all the candidate intent information to determine the comprehensive confidence of each candidate intent information includes: Based on the intent probability distribution corresponding to each candidate intent information, the base model prediction probability factor is obtained; Based on the user's historical dialogues, we obtain the context consistency factor and the historical pattern similarity factor; The initial confidence level is obtained by weighting and fusing the base model prediction probability factor, the context consistency factor, the historical pattern similarity factor, and the preset knowledge matching factor. The initial comprehensive confidence level is dynamically calibrated using a Bayesian update mechanism to obtain the comprehensive confidence level for each candidate intent information.

[0062] Specifically, the base model prediction probability factor (P_model) is directly derived from the original intent probability distribution output by each intent understanding path. In practice, the system extracts the predicted probability value of the path for the most likely intent category from the results of each processing path, including fuzzy matching, exact matching, and context awareness. For example, in an intelligent customer service scenario, the probability of a fuzzy matching path outputting a product quality complaint is 0.68, the probability of an LSTM exact matching path outputting the same intent is 0.67, and the probability of a context-aware path outputting the same intent is 0.85. The system uses the probability values ​​output by each path as independent sources of evidence for subsequent confidence fusion. This factor reflects the direct judgment of user intent by different analysis models based on their algorithmic characteristics and is the most fundamental component of confidence assessment. By analyzing the logical coherence between the current candidate intent and historical dialogue content, the context consistency factor (P_context) is obtained. The system uses a dialogue state tracking model to embed candidate intents into the entire dialogue flow and evaluate whether they conform to the development logic of the dialogue. For example, in a dialogue flow where the user first inquires about a product and then expresses dissatisfaction, the context consistency of the complaint intent will be higher than that of the consultation intent. The system calculates the historical pattern similarity factor (P_history) by comparing the current input with the user's or similar user groups' historical behavior patterns. It extracts the user's typical intent patterns, expression habits, and problem-solving paths from the user profile database and calculates the similarity between the current candidate intent and these historical patterns. If a user has a high history of product complaints, the historical pattern similarity factor for their current complaint intent will be correspondingly higher.

[0063] The system obtains a knowledge matching factor (P_knowledge) by querying the domain knowledge graph. This factor measures the degree of alignment between the candidate intent and domain common sense. In practice, the system first automatically adjusts the weights of each factor based on the current scenario characteristics: reducing model weights and increasing context weights in fuzzy expression scenarios, and increasing knowledge matching weights in technical query scenarios. Then, the initial confidence score is calculated using the formula: Confidence = w1·P_model + w2·P_context + w3·P_knowledge + w4·P_history. Where: P_model represents the probability value predicted by the base model; P_context is the context consistency score, reflecting the consistency between the intent and the current dialogue environment; P_knowledge is the knowledge matching degree, reflecting the degree of alignment between the intent and entity relationships in the knowledge graph; P_history is the historical pattern similarity, measuring the similarity between the current intent and previously successfully recognized intents; w1, w2, w3, and w4 are dynamic weight parameters that are adaptively adjusted according to the real-time scenario. In the intelligent customer service case, the system assigns a weight of 0.35 to P_context, 0.25 to P_model, 0.20 to P_knowledge, and 0.20 to P_history, and calculates an initial confidence level of 0.7645.

[0064] Finally, the system employs a Bayesian update mechanism to dynamically calibrate the initial confidence level, thereby improving the reliability and adaptability of probability estimation. Specifically, the system uses the historical success rate of each intent category as the prior probability, the current initial confidence level as the likelihood function, and combines real-time feedback data, such as user behavior characteristics and environmental parameters, as evidence to calculate the posterior probability using the Bayesian formula. In the intelligent customer service case, the historical prior probability of the product quality complaint intent is 0.15. Combining evidence such as the current user complaint rate (0.12) and product defect rate (0.03), the confidence level is calibrated from 0.7645 to 0.82 using a Bayesian update. Simultaneously, the system uses temperature scaling technology to smooth the probability distribution, ultimately outputting a calibrated comprehensive confidence level. This confidence level reflects both the intensity of the current analysis and incorporates historical experience and real-time scenario characteristics, resulting in higher discriminative reliability.

[0065] In this optional embodiment, a two-layer optimization mechanism combining multi-factor weighted fusion and Bayesian dynamic calibration is employed to construct an adaptive confidence assessment system that integrates model prediction, contextual logic, historical patterns, and domain knowledge, significantly improving the credibility and scenario adaptability of intent recognition results. Specifically, the multi-factor fusion mechanism systematically integrates evidence from four dimensions: base model prediction, contextual consistency, historical behavioral patterns, and knowledge matching, overcoming the limitations of relying solely on model output. Dynamic weight allocation intelligently adjusts the contribution of each factor according to different scenario characteristics, ensuring the assessment system maintains optimal discriminative balance. Furthermore, a Bayesian update mechanism integrates historical experience and real-time data into the confidence calculation. Through continuous iteration of prior probabilities and current evidence, the confidence assessment not only reflects instantaneous judgment strength but also possesses long-term reliability based on statistical learning. This design enables the system to output robust and interpretable confidence indicators even when facing fuzzy expressions, complex scenarios, or sparse data, providing accurate and reliable basis for subsequent hierarchical decision-making, significantly reducing the risk of misjudgment, and enhancing the system's adaptability to different user groups and scenario changes.

[0066] Optionally, the step of matching based on the confidence level of the candidate intent information to determine the decision path that matches the candidate intent information includes: The overall confidence level of each candidate intent information is matched and compared with preset high confidence thresholds and medium confidence thresholds; If the overall confidence level is greater than or equal to the high confidence threshold, then the high confidence decision path is taken as the decision path for the candidate intent information. If the overall confidence level is less than the high confidence threshold but greater than or equal to the medium confidence threshold, then the medium confidence decision path will be used as the decision path for the candidate intent information. If the overall confidence level is less than the medium confidence threshold, then the low confidence decision path will be used as the decision path for the candidate intent information.

[0067] Specifically, the high confidence threshold is usually set to 0.85, and the medium confidence threshold is usually set to 0.60. In this embodiment, these thresholds can be dynamically adjusted based on historical performance data and scenario characteristics. The comprehensive confidence of each candidate intention obtained through multi-factor fusion and Bayesian calibration is compared step by step with these two thresholds. In the intelligent customer service case, the final confidence of the product quality complaint intention after enhancement processing is 0.87, and the system compares it with the high confidence threshold of 0.85. In the multi-round dialogue scenario, the confidence of the intention to adjust the light color is 0.96, and the same threshold comparison is also performed. This comparison process provides a clear quantitative basis for the subsequent decision path allocation. When the comprehensive confidence of the candidate intention reaches or exceeds the high confidence threshold of 0.85, it is assigned to the high confidence decision path. Under this path, the system directly adopts the recognition result and performs the corresponding operation, and at the same time updates the intention cache to record the successful pattern. In specific implementation, the extra verification steps are skipped, and a structured response is directly generated and returned to the user.

[0068] Exemplarily, in the intelligent assistant case, the confidence of the intention to adjust the light color is 0.96, and the system directly controls the living room light to be adjusted to warm light; the confidence of the intention to query the weather is 0.91, and the system directly calls the weather API to obtain the tomorrow's forecast. The high confidence path aims to ensure high efficiency in processing while maintaining high accuracy, and is applicable to scenarios with clear intentions and clear expressions.

[0069] When the comprehensive confidence of the candidate intention is between 0.60 and 0.85, it is assigned to the medium confidence decision path. Under this path, an enhancement processing flow is initiated to further optimize the recognition result. The specific implementation includes technical means such as multi-model integrated reasoning and semantic re-matching.

[0070] In the intelligent customer service case, the confidence of the product quality complaint intention obtained in the initial evaluation is 0.82, and the system immediately initiates enhancement processing. It is re-evaluated through a multi-model voting strategy and deeply matched with the quality problem complaint by combining semantic extension technology. Finally, after the confidence is increased to 0.87, it is transferred to the high confidence path. The medium confidence path improves the reliability of the result through additional processing steps, balancing the requirements of processing efficiency and accuracy.

[0071] When the overall confidence level of a candidate intent falls below the medium confidence threshold of 0.60, it is assigned to a low-confidence decision path. Under this path, a three-tiered intelligent reset processing mechanism is activated. First, the prompt chain is reconstructed, reorganizing the analysis process. Second, the execution modules are reordered, adjusting the priorities of each processing module. Finally, interactive confirmation is initiated, proactively requesting supplementary information from the user. In practice, the system generates targeted clarification questions, such as "Which aspect of the product are you dissatisfied with?", and re-matches intent based on user feedback. The low-confidence path resolves extremely ambiguous or insufficiently informed expressions through interactive learning. Simultaneously, the collected interaction data is used for subsequent model iterations and knowledge base expansion, forming a continuous evolution mechanism for the system.

[0072] In this optional embodiment, a dynamic triage mechanism guided by confidence thresholds constructs an adaptive decision-making system capable of intelligently balancing processing efficiency, resource allocation, and recognition accuracy, achieving optimized processing of high, medium, and low-quality recognition results. Specifically, the three-level threshold division precisely guides intent recognition results to differentiated processing flows. The high-confidence path enables rapid response to clear intents, significantly reducing processing latency and ensuring execution accuracy. The medium-confidence path achieves a balance between efficiency and accuracy through an enhanced processing mechanism, effectively improving the recognition accuracy of boundary cases. The low-confidence path employs an interactive reset strategy, transforming recognition difficulties into learning opportunities, avoiding misjudgment risks and accumulating valuable training data. In this embodiment, this hierarchical design not only optimizes computing resource allocation, enabling simple requests to receive rapid responses while complex problems are processed in depth, but more importantly, it forms a complete capability spectrum from rapid execution to enhanced optimization to active learning. This allows the system to intelligently select the most suitable processing strategy based on the actual quality of intent recognition, significantly improving the overall system's applicability, economy, and evolutionary capabilities.

[0073] Optionally, generating an intent recognition result for the user through the decision path matched with each candidate intent information includes: If the decision path is the high-confidence decision path, then an intent recognition result for the user is generated based on the intent probability distribution in the candidate intent information. If the decision path is the medium confidence decision path, then the intent probability distribution in the candidate intent information is optimized and reordered through a multi-model integrated voting process to obtain an optimized intent probability distribution; based on the optimized intent probability distribution, an intent recognition result for the user is generated. If the decision path is the low-confidence decision path, then based on the intent probability distribution in the candidate intent information, a clarification statement is generated to interact with the user and obtain user feedback text, and based on the user feedback text, an intent recognition result for the user is generated.

[0074] Specifically, under the high-confidence decision path, the intent with the highest confidence in the intent probability distribution is directly adopted as the final recognition result. In specific implementation, the intent category with a probability value greater than or equal to the high-confidence threshold (usually 0.85) and ranked first is selected from the probability distribution to directly generate the structured output result. This result not only includes the recognized intent type but also related slot information, context labels, and recommended response templates.

[0075] Exemplarily, in the smart home assistant scenario, the confidence of the intent to adjust the light color is 0.96, and the confidence of the intent to query the weather is 0.91. Since both exceed the threshold, these two intents are simultaneously taken as the final results, and a composite instruction execution plan is generated. The system also updates the intent cache, records this successful recognition pattern, and provides a reference basis for subsequent similar scenarios.

[0076] Under the medium-confidence decision path, the multi-model integration voting process is initiated to further optimize the candidate intents. In specific implementation, the system aggregates the recognition results of different paths such as fuzzy matching, exact matching, and context awareness, and adopts a weighted voting or averaging strategy for integrated decision-making.

[0077] For example, in the intelligent customer service case, the system synthesizes the outputs of the BERT classifier (confidence 0.79), the LSTM model (confidence 0.85), and the rule engine (confidence 0.72). After weighted average calculation, the confidence of the enhanced intent of product quality complaint is 0.84. At the same time, the system performs semantic re-matching, expands the original expression into a more standardized query form, such as expanding "enough is enough" into "complaint about product quality problems", to further improve the matching accuracy. Finally, the candidate intents are re-ranked according to the optimized probability distribution, and the intent with the highest confidence is selected to generate the final recognition result.

[0078] Under the low-confidence decision path, the interactive clarification mechanism is activated to obtain more user information. In specific implementation, the system analyzes the candidate intents in the intent probability distribution and generates targeted clarification questions for the intent categories with relatively high probabilities but still not reaching the confidence threshold.

[0079] For example, in the intelligent customer service scenario, the system may generate a clarification statement such as "Are you dissatisfied with the product quality, delivery, or service?" to guide the user to supplement key information. The user feedback text is re-entered into the intent recognition process, and the system re-performs multi-dimensional feature extraction and intent recognition by combining the original input and the feedback information. After obtaining sufficient information, the system recalculates the intent probability distribution. If the confidence reaches the threshold, the final recognition result is generated; if it still does not reach the threshold, it may continue to clarify or be recorded as a pending case. This process not only solves the current recognition difficulty but also accumulates valuable learning data for the system for subsequent model optimization and knowledge base expansion.

[0080] In this optional embodiment, a multi-layered result generation system is constructed through a differentiated result generation strategy guided by decision paths. This system can intelligently adjust the output accuracy and interaction depth based on the confidence level, achieving full-scenario coverage from direct execution to integrated optimization and interactive clarification. Specifically, the direct generation mechanism of the high-confidence path ensures rapid response and accurate execution of clear intentions, significantly improving the processing efficiency of clear expressions. The integrated optimization process of the medium-confidence path enhances the reliability of boundary case recognition through multi-model voting and semantic re-matching, significantly improving the recognition accuracy of complex expressions while maintaining reasonable processing latency. The interactive clarification mechanism of the low-confidence path transforms recognition difficulties into active learning opportunities, guiding users to supplement key information through intelligent questioning, solving the current recognition problem while accumulating training data. This hierarchical generation strategy enables the system to intelligently select the most appropriate output method based on recognition quality. It not only optimizes the user experience, such as quickly responding to clear requests, deeply processing complex problems, and actively clarifying ambiguous expressions, but also forms a closed loop of capabilities from execution to learning, significantly improving the overall performance and continuous evolution capability of the system at different confidence levels.

[0081] In a preferred embodiment of the present invention, combined with Figure 2As shown, the system first receives text input from the user. This text is converted into a unified semantic vector representation by a multi-dimensional semantic encoding module. This semantic vector is then fed in parallel into three core processing modules. Specifically, the fuzzy module performs fuzzy logic reasoning to generate fuzzy confidence; the precise module uses a deep learning model for precise matching to generate precise confidence; and the context module combines historical dialogues to perform contextual association analysis to generate contextual confidence. The outputs of these three modules are simultaneously fed into a dynamic calibrator. This calibrator, based on a multi-factor fusion strategy, comprehensively considers model prediction, contextual consistency, historical patterns, knowledge matching, and other dimensions for weighted calculation, and dynamically adjusts the results using a Bayesian update mechanism to generate a comprehensive confidence score for each candidate intent. The comprehensive confidence score is then processed hierarchically by the decision engine. High-confidence intents are returned directly; medium-confidence intents trigger an enhancement processing flow, optimizing the results through multi-model integration and semantic re-matching before returning them; and low-confidence intents initiate an interactive clarification mechanism, requesting supplementary information from the user. All processing results enter the feedback learning module, where the system continuously optimizes model parameters, adjusts confidence thresholds, and expands the knowledge base based on actual user feedback, forming a complete closed loop from recognition execution to autonomous evolution. In this embodiment of the invention, through modular design and parallel processing architecture, the system can simultaneously utilize the flexibility of fuzzy inference, the accuracy of precise models, and the coherence of context awareness. After dynamic calibration and intelligent decision-making, it can ultimately achieve efficient, accurate, and adaptive intent recognition. The entire system realizes intelligent routing of processing paths and optimized resource allocation, ensuring that the most appropriate processing strategy can be adopted under different confidence levels.

[0082] Combination Figure 3 As shown, the AI ​​Agent intent recognition system of this invention includes: The data acquisition unit is used to acquire the user's input text data; The fusion unit is used to extract and fuse multi-dimensional features from the input text data to generate a comprehensive feature representation that includes semantic, syntactic and contextual information. The parallel processing unit is used to perform parallel processing based on the comprehensive feature representation through multiple intent understanding paths to obtain candidate intent information corresponding to each intent understanding path. The confidence determination unit is used to perform multi-factor confidence fusion and dynamic calibration on all the candidate intent information to determine the comprehensive confidence of each candidate intent information; A matching unit is used to match the candidate intent information based on the confidence level of the candidate intent information and determine a decision path that matches the candidate intent information. The generation unit is configured to generate an intent recognition result for the user by matching the decision path with each of the candidate intent information.

[0083] The AI ​​Agent intent recognition system of the present invention has the same advantages over the prior art as the AI ​​Agent intent recognition method described above, and will not be repeated here.

[0084] Combination Figure 4 As shown, the electronic device of this embodiment includes: a processor and a memory, wherein the memory is used to store computer programs; When the computer program is loaded by the processor, it causes the processor to execute the AIAgent intent recognition method as described above.

[0085] The electronic device of the present invention has the same advantages over the prior art as the AI ​​Agent intent recognition method described above, and will not be repeated here.

[0086] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for recognizing the intent of an AI agent, characterized in that, include: The system acquires user input text data and performs multi-dimensional feature extraction and fusion on the input text data to generate a comprehensive feature representation that includes semantic, syntactic and contextual information. By performing parallel processing on multiple intent understanding paths based on the comprehensive feature representation, candidate intent information corresponding to each intent understanding path is obtained; Multi-factor confidence fusion and dynamic calibration are performed on all the candidate intent information to determine the comprehensive confidence of each candidate intent information; Matching is performed based on the confidence level of the candidate intent information to determine a decision path that matches the candidate intent information; An intent recognition result for the user is generated by matching the decision path with each candidate intent information.

2. The AI ​​Agent intent recognition method according to claim 1, characterized in that, The step of extracting and fusing multi-dimensional features from the input text data to generate a comprehensive feature representation that includes semantic, syntactic, and contextual information includes: The input text data is subjected to standardization preprocessing to obtain normalized text data. The standardization preprocessing includes encoding unification, word segmentation, part-of-speech tagging, and named entity recognition. Feature extraction is performed on the standardized text data to obtain its semantic features, syntactic structure features, and sentiment features. The semantic features, the grammatical structure features, and the sentiment features are fused to generate the comprehensive feature representation.

3. The AI ​​Agent intent recognition method according to claim 1, characterized in that, The intent understanding path includes a fuzzy matching path; the parallel processing of multiple intent understanding paths based on the comprehensive feature representation to obtain candidate intent information corresponding to each intent understanding path includes: Based on the semantic features in the comprehensive feature representation, the semantic similarity between the input text data and the historical intent template is obtained through the fuzzy matching path. Based on the semantic features and syntactic structure features in the comprehensive feature representation, the fuzzy membership degree of the input text data corresponding to multiple preset intent categories is obtained; Based on the semantic similarity and the fuzzy membership degree, a first intent probability distribution is obtained, and the first intent probability distribution is used as the candidate intent information of the fuzzy matching path.

4. The AI ​​Agent intent recognition method according to claim 1, characterized in that, The intent understanding path includes an exact matching path; the parallel processing of multiple intent understanding paths based on the comprehensive feature representation to obtain candidate intent information corresponding to each intent understanding path includes: Based on the semantic features in the comprehensive feature representation, the deep semantic vector representation of the input text data is obtained through the precise matching path. Based on the deep semantic vector representation, the initial classification probability of the input text data corresponding to each preset intent category is obtained; Based on the syntactic structure features in the comprehensive feature representation, the initial classification probability is modified in a context-aware manner to obtain a second intent probability distribution, and the second intent probability distribution is used as the candidate intent information of the precise matching path.

5. The AI ​​Agent intent recognition method according to claim 1, characterized in that, The intent understanding path includes a context-aware path; the parallel processing of multiple intent understanding paths based on the comprehensive feature representation to obtain candidate intent information corresponding to each intent understanding path includes: The context-aware path is used to obtain the context encoding vector of the user's historical dialogues. Based on the semantic features in the comprehensive feature representation and the context encoding vector, the relevance weight between the input text data and the historical dialogue context is obtained; The comprehensive feature representation is enhanced based on the relevance weights, and reasoning is performed in conjunction with a preset knowledge graph to obtain a third intent probability distribution, which is then used as the candidate intent information for the context-aware path.

6. The AI ​​Agent intent recognition method according to claim 1, characterized in that, The step of performing multi-factor confidence fusion and dynamic calibration on all the candidate intent information to determine the comprehensive confidence of each candidate intent information includes: Based on the intent probability distribution corresponding to each candidate intent information, the base model prediction probability factor is obtained; Based on the user's historical dialogues, we obtain the context consistency factor and the historical pattern similarity factor; The initial confidence level is obtained by weighting and fusing the base model prediction probability factor, the context consistency factor, the historical pattern similarity factor, and the preset knowledge matching factor. The initial comprehensive confidence level is dynamically calibrated using a Bayesian update mechanism to obtain the comprehensive confidence level for each candidate intent information.

7. The AI ​​Agent intent recognition method according to claim 1, characterized in that, The step of matching based on the confidence level of the candidate intent information to determine the matching decision path includes: The overall confidence level of each candidate intent information is matched and compared with preset high confidence thresholds and medium confidence thresholds; If the overall confidence level is greater than or equal to the high confidence threshold, then the high confidence decision path is taken as the decision path for the candidate intent information. If the overall confidence level is less than the high confidence threshold but greater than or equal to the medium confidence threshold, then the medium confidence decision path will be used as the decision path for the candidate intent information. If the overall confidence level is less than the medium confidence threshold, then the low confidence decision path will be used as the decision path for the candidate intent information.

8. The AI ​​Agent intent recognition method according to claim 7, characterized in that, The step of generating an intent recognition result for the user by matching the decision path with each candidate intent information includes: If the decision path is the high-confidence decision path, then an intent recognition result for the user is generated based on the intent probability distribution in the candidate intent information. If the decision path is the medium confidence decision path, then the intent probability distribution in the candidate intent information is optimized and reordered through a multi-model integrated voting process to obtain an optimized intent probability distribution; based on the optimized intent probability distribution, an intent recognition result for the user is generated. If the decision path is the low-confidence decision path, then based on the intent probability distribution in the candidate intent information, a clarification statement is generated to interact with the user and obtain user feedback text, and based on the user feedback text, an intent recognition result for the user is generated.

9. An AI Agent intent recognition system, characterized in that, include: The data processing unit is used to acquire the user's input text data, and to perform multi-dimensional feature extraction and fusion on the input text data to generate a comprehensive feature representation containing semantic, syntactic and contextual information; The parallel processing unit is used to perform parallel processing based on the comprehensive feature representation through multiple intent understanding paths to obtain candidate intent information corresponding to each intent understanding path. The confidence determination unit is used to perform multi-factor confidence fusion and dynamic calibration on all the candidate intent information to determine the comprehensive confidence of each candidate intent information; A matching unit is used to match the candidate intent information based on the confidence level of the candidate intent information and determine a decision path that matches the candidate intent information. The generation unit is configured to generate an intent recognition result for the user by matching the decision path with each of the candidate intent information.

10. An electronic device, characterized in that, include: Processor and memory, the memory being used to store computer programs; When the computer program is loaded by the processor, it causes the processor to execute the AI ​​Agent intent recognition method as described in any one of claims 1-8.

Images