Semantic understanding method and engine based on multi-model collaborative reasoning and self-supervised learning

Abstract

This invention discloses a semantic understanding method and engine based on multi-model collaborative reasoning and self-supervised learning, belonging to the fields of natural language processing and artificial intelligence. The method includes preprocessing input data to extract semantic features; intelligently scheduling at least two models from a large language model, task-oriented model, and rule engine for collaborative reasoning through a dynamic routing module based on multi-level policies and meta-learning networks; using a multi-model fusion module to perform hierarchical fusion and conflict resolution of the reasoning results, outputting the final semantic understanding result; and continuously optimizing the system through feedback learning. A self-supervised pre-training task is used to learn a general semantic representation from unlabeled data to enhance the model's generalization ability. The engine includes core modules such as input processing, dynamic routing, multi-model fusion, feedback learning, model repository, and monitoring and security, which work together to implement the above method. This invention achieves high-precision, high-efficiency, and adaptive semantic understanding.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and artificial intelligence, and in particular to semantic understanding methods and engines based on multi-model collaborative reasoning and self-supervised learning. Background Technology

[0002] With the widespread application of artificial intelligence technology, semantic understanding engines have become the core of intelligent interaction systems. However, current mainstream solutions still face significant challenges when dealing with complex real-world scenarios.

[0003] First, existing methods have limitations in understanding complex intents and adapting across domains. For example, while CN112036177A uses multi-model fusion to calculate text similarity, its static feature fusion method cannot achieve real-time dynamic routing. CN120597213A introduces the concept of dynamic routing, but its mechanism based on a pre-defined mapping table lacks adaptive learning capabilities and struggles to handle complex open-domain queries. Second, existing technologies heavily rely on labeled data, hindering rapid model adaptation in data-scarce scenarios. CN113011427A uses self-supervised learning for image segmentation, but its task design is unsuitable for text semantic understanding. CN120012907A reduces label dependency through atomic behavior orchestration, but still requires extensive prior knowledge to define the behavior library, failing to fully utilize unlabeled text data to learn general semantic representations. Finally, there are shortcomings in achieving efficient inference and system robustness while maintaining accuracy. Publication number CN119358687A improves performance through multi-model token-level collaboration, but parallel inference leads to high computational overhead; Publication number CN120012907A has a complex multi-branch tree orchestration mechanism and does not integrate model compression and efficient deployment technologies. In addition, existing solutions generally lack complete monitoring, interpretability, and security mechanisms, affecting system stability and reliability.

[0004] Therefore, there is an urgent need for a technical solution that can effectively utilize large-scale unlabeled text data, learn general semantic representations through specially designed self-supervised learning tasks, and combine them with multi-model collaborative reasoning to solve the problem of dependence on labeled data. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a semantic understanding method based on multi-model collaborative reasoning and self-supervised learning, comprising the following steps:

[0006] S101. Input data acquisition and standardization preprocessing: Receive and convert the original input data into standardized text, perform word segmentation, lexicalization, basic language feature extraction and shallow semantic parsing, and generate structured intermediate representation objects.

[0007] S102, Dynamic routing decision and intelligent model scheduling: Based on the intermediate representation object and context information, a model scheduling list is generated through multi-level routing decision, and at least one semantic understanding model is intelligently scheduled.

[0008] S103. Multi-model parallel inference and standardized result collection: Schedule the corresponding models for inference according to the model scheduling list, and collect the standardized inference results returned by each model.

[0009] S104. Multi-model fusion and conflict collaborative resolution: The collected inference results from multiple models are fused in a hierarchical manner, and the conflicts detected during the fusion process are resolved using multiple strategies to generate the final semantic understanding result;

[0010] S105. Output Results and Continuous Feedback Learning: Output the final semantic understanding results and perform feedback learning based on the log data of the processing link to optimize the routing strategy and update the model performance profile; at the same time, use the de-identified massive user query text to perform periodic incremental training on the system's shared basic semantic encoding model through a self-supervised pre-training task.

[0011] As a preferred embodiment of this application, the dynamic routing decision and intelligent model scheduling includes: a first-level coarse screening of intent types, which determines the initial intent type of the input through an intent classifier; a second-level fine-grained domain routing, which determines the subdivided domain through a domain classifier if the initial intent type belongs to a preset professional category; and a third-level adaptive routing decision based on meta-learning, which receives a comprehensive feature vector through a meta-learning routing network and outputs a weight vector pointing to each candidate model, wherein the comprehensive feature vector includes at least preliminary semantic features, intent probability distribution, domain probability distribution, and real-time system load indicators.

[0012] As a preferred embodiment of this application, the third-level meta-learning-based adaptive routing decision further includes: evaluating the prediction uncertainty of each candidate model for the current input; if the uncertainty of the optimal candidate model exceeds a preset threshold, the decision is to initiate multi-model parallel inference and schedule the top K models with the highest weights to process together.

[0013] As a preferred embodiment of this application, the layered fusion includes: semantic layer fusion, which performs weighted voting on the intent predictions of each model's output based on the weights and historical accuracy of each model to determine the initial fusion intent; structural layer fusion, which clusters the slot values ​​output by each model based on semantic similarity and determines the final filling value of each slot according to the weights and confidence of the source models; and learning-based fusion fine-tuning, which constructs a graph with the outputs of each model as nodes and learns the node adjustment coefficients through a graph neural network to fine-tune the initial fusion results.

[0014] As a preferred embodiment of this application, the multi-strategy resolution is executed in the following priority order: the judgment result of the rule engine is adopted first; if it cannot be resolved, the conflicting content is linked to an external knowledge base or knowledge graph for verification; if it still cannot be resolved, the conflict is submitted to a manual review platform for adjudication, and the adjudication result is used to optimize the system.

[0015] As a preferred embodiment of this application, the continuous feedback learning includes: incrementally updating the parameters of the meta-learning routing network using reinforcement learning or supervised learning methods based on user feedback or task execution results; and dynamically updating the performance indicators of each model according to its performance in a large number of requests, so as to influence the calculation of future routing decision weights.

[0016] This application also provides a semantic understanding engine based on multi-model collaborative reasoning and self-supervised learning, including:

[0017] The input processing module is used to receive and standardize multimodal raw input, and transform it into a structured intermediate representation object rich in linguistic features;

[0018] A dynamic routing module, connected to the input processing module, is used to intelligently schedule at least one semantic understanding model based on the intermediate representation object, context, and system state through multi-level routing decisions.

[0019] The multi-model fusion module, connected to the dynamic routing module, is used to perform hierarchical fusion and conflict resolution on the inference results of multiple models to generate the final semantic understanding result.

[0020] The model repository module is used for versioned storage, containerized deployment, and service management of semantic understanding models;

[0021] The feedback learning module, connected to the dynamic routing module and the model repository module, is used to collect and process link logs, optimize routing strategies, and update model performance profiles.

[0022] The monitoring and security module provides system observability, security, and compliance assurance.

[0023] As a preferred embodiment of this application, the dynamic routing module includes:

[0024] The routing policy execution engine is used to execute multi-level routing decision logic in a preset order;

[0025] The intent classifier component and the domain classifier component are built into the routing strategy execution engine and are used to perform coarse screening of the input intent type and fine-grained domain discrimination when the intent type belongs to a preset professional category.

[0026] The meta-learning routing network component is a lightweight neural network trained by a model-independent meta-learning algorithm. It is built into the routing policy execution engine and is used to receive a comprehensive feature vector that integrates semantic features, classification probabilities and system load information, and output a weight vector pointing to each candidate model.

[0027] The model registry is used to dynamically maintain the metadata, real-time health status, and historical performance metrics of all available model inference services.

[0028] The model scheduling executor component is used to schedule the target model through a remote service call interface based on the decision result of the routing strategy execution engine, and to implement timeout and circuit breaker control.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0030] 1) Improve the accuracy and reliability of understanding complex intents. Through dynamic routing and multi-model collaboration, the system can integrate the general knowledge of large language models, the domain accuracy of task-oriented models, and the business determinism of rule engines to achieve a more accurate understanding of fuzzy and complex user inputs and effectively reduce the error rate of understanding.

[0031] 2) Enhance the model's generalization ability and reduce data dependence. By adopting self-supervised pre-training and online learning techniques, the model can be continuously optimized using large-scale unlabeled data and online interactive data, which significantly improves its adaptability in scenarios with few samples and new domains and reduces its dependence on labeled data.

[0032] 3) Achieve efficient inference and elastic resource utilization. Through compression technologies such as model quantization and knowledge distillation, combined with an elastic deployment architecture that integrates edge cloud collaboration, the system can achieve dynamic optimization of resource allocation. While ensuring high accuracy, it can significantly improve response speed and throughput, and reduce overall deployment and operation costs.

[0033] 4) Construct an evolvable, explainable, secure, and controllable system. The system integrates a feedback learning mechanism for self-optimization and provides end-to-end monitoring and interpretability analysis. Simultaneously, built-in privacy protection, security defense, and compliance detection mechanisms ensure data security and the reliability of algorithmic decisions, meeting the application requirements of highly sensitive fields. Attached Figure Description

[0034] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0035] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is a flowchart of the semantic understanding method provided in the embodiments of the present invention.

[0037] Figure 2 This is a flowchart of input data acquisition and standardized preprocessing provided in an embodiment of the present invention.

[0038] Figure 3 This is a flowchart of dynamic routing decision-making and intelligent model scheduling provided in an embodiment of the present invention.

[0039] Figure 4 This is a flowchart of multi-model parallel inference and standardized result collection provided in the embodiments of the present invention.

[0040] Figure 5 This is a flowchart of multi-model fusion and conflict collaborative resolution provided in the embodiments of the present invention.

[0041] Figure 6 This is a flowchart of the result output and continuous feedback learning provided in the embodiments of the present invention.

[0042] Figure 7 This is an architecture diagram of the semantic understanding engine provided in an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0044] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship and movement of each component in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indication will also change accordingly.

[0045] Furthermore, the use of terms such as "first" and "second" in this invention is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" and "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but only on the basis of being achievable by those skilled in the art. When the combination of technical solutions is contradictory or impossible to implement, such a combination of technical solutions should be considered non-existent and not within the scope of protection claimed by this invention.

[0046] Example 1

[0047] This invention provides a semantic understanding method based on multi-model collaborative reasoning and self-supervised learning. The method achieves high-precision and high-efficiency semantic understanding of user input through the collaborative work of an input processing module, a dynamic routing module, a multi-model fusion module, and a feedback learning module. The entire process begins with data input, proceeds through preprocessing, intelligent routing, multi-model reasoning, result fusion, and conflict resolution, ultimately outputting a structured semantic understanding result and forming a feedback loop to continuously optimize the system. Figure 1 As shown, the specific steps include:

[0048] S101 Input Data Acquisition and Standardization Preprocessing

[0049] This step provides a standardized input basis for subsequent routing decisions. For example... Figure 2 As shown, the input processing module receives raw input data from the user client. The raw input data is mainly divided into two categories: the first category is directly input text data, such as text from keyboard input, instant messaging software, or OCR recognition output; the second category is raw signal data, such as voice audio signals or image signals containing text.

[0050] When the input is a raw signal, the input processing module first calls the corresponding conversion module to convert it into standard text. For example, for speech signals, the system processes them through an integrated Automatic Speech Recognition (ASR) component. The ASR component receives audio streams or audio files, decodes them using acoustic and language models, converts the audio feature sequences into the most probable word sequences, and finally outputs transcribed text with timestamps and confidence scores. For image signals, the system processes them through an Optical Character Recognition (OCR) component. The OCR component first performs image preprocessing (such as binarization and denoising), and then uses a deep learning-based text detection and recognition model (such as CRNN or a Transformer-based model) to locate and recognize text regions in the image, finally outputting text arranged in reading order.

[0051] Subsequently, the input processing module performs standardization preprocessing on the obtained text. The first step of preprocessing is word segmentation and tokenization. The input processing module loads the appropriate word segmenter based on the target language (such as Chinese or English). For Chinese, a sub-word segmenter based on BPE or WordPiece is used to cut the sentence into a sequence of tokens, while adding special start and delimiter characters. For English, sub-word segmentation is performed similarly, and capitalization is handled.

[0052] The second step of preprocessing is basic language feature extraction. The input processing module takes the same word sequence obtained in the first step and feeds it into a lightweight, pre-trained context encoding model (such as Tiny BERT or a distilled BERT model) to obtain the context vector representation of each word and the aggregate semantic vector of the entire sentence. The context encoding model is pre-trained on large-scale unlabeled text through self-supervised pre-training tasks such as masked language modeling, which can effectively capture the semantic relationships and contextual information between words; the aggregate semantic vector is a fixed-length dense vector, which serves as the initial semantic feature for subsequent routing decisions.

[0053] The third step in preprocessing is shallow semantic parsing. The input processing module performs named entity recognition and dependency parsing in parallel. The named entity recognition model uses a model based on conditional random fields or sequence labeling BERT, which benefits from general semantic representations learned through self-supervised pre-training. This model identifies entities in the text such as names of people, places, organizations, times, dates, and currencies, and labels their types and boundaries. The dependency parser parses the grammatical dependencies between words in the sentence, constructing a dependency tree to understand the sentence's core structure. Finally, the intermediate representation encapsulation component encapsulates the lexical sequence, entity list, syntax tree, and preliminary semantic vectors into a structured intermediate representation object, serving as the unified input format for all subsequent processing steps.

[0054] S102, Dynamic Routing Decision and Model-Based Intelligent Scheduling

[0055] After preprocessing, the generated intermediate representation object, along with the context information of the current dialogue (such as historical dialogue records, user session state, and user profile tags), is sent to the dynamic routing module. Figure 3 As shown, the core responsibility of the dynamic routing module is to intelligently determine which semantic understanding model(s) should handle the current request based on the characteristics of the current input and the system state, so as to achieve the optimal balance between efficiency and accuracy.

[0056] Routing decision-making is a multi-level, hierarchical process. The first level is coarse screening of intent types. The dynamic routing module has a built-in intent classifier component, which is a text classification model fine-tuned on a general dialogue dataset. The intent classifier component takes a preliminary semantic vector and a sequence of words as input and outputs a probability distribution on predefined intent types. These intent types typically include general question answering, business processing, information querying, transaction operations, and casual conversation. The system selects the intent type with the highest probability as the initial routing basis. For example, if the input is "Book me a flight to Beijing tomorrow," it might be classified as a "business processing" intent.

[0057] The second level is fine-grained domain routing. If the first level determines that the intent belongs to a category requiring specialized knowledge (such as business processing or information retrieval), the domain classifier component is activated. This component is trained on a large amount of domain-labeled data and can distinguish between sub-domains such as finance, healthcare, travel, government affairs, and e-commerce. The domain classifier component concatenates the complete input text with historical context and then classifies it to obtain a domain probability distribution. Combined with the intent type from the first level, the system can more accurately locate the task scenario, such as "business processing - travel".

[0058] The third level is adaptive routing decision-making based on meta-learning. The dynamic routing module maintains a meta-learning routing network component, which is a lightweight neural network with a 3-layer fully connected network structure, 128 nodes per layer, and ReLU activation function. The meta-learning routing network component is trained offline on massive meta-tasks using a model-independent meta-learning algorithm. Each meta-task simulates a few-shot scheduling scenario under a specific intent-domain combination, with the goal of enabling the routing network to quickly adapt to new tasks or newly added models. During online inference, the meta-learning routing network submodule receives a comprehensive feature vector, which is composed of a preliminary semantic vector, intent probability distribution, domain probability distribution, context state encoding (such as dialogue turn number, query length), and real-time system load metrics (such as the length of each model's service queue and GPU memory utilization). The meta-learning routing network outputs a weight vector, where each element corresponds to a candidate model in the model library, and the weight value represents the suitability of scheduling that model in the current scenario.

[0059] Routing decisions consider not only model capability matching but also cost and uncertainty awareness. Cost awareness means that during training, the loss function of the routing network includes estimated latency and computational resource consumption for model inference, guiding the network to balance accuracy and efficiency. Uncertainty awareness refers to the system evaluating the prediction uncertainty of each model for the current input. For deep learning models, Monte Carlo Dropout is enabled, and multiple forward propagation samplings are performed to calculate the variance of the prediction results as an uncertainty estimate. If the uncertainty of the optimal candidate model is too high (exceeding a threshold), the dynamic routing module will decide to initiate multi-model parallel inference, that is, simultaneously scheduling the top K models with the highest weights to process the data together, in order to aggregate multi-source information and reduce risk.

[0060] Finally, the dynamic routing module generates a model scheduling list based on the decision results. The model scheduling list specifies the model identifier to be scheduled, the scheduling type (serial or parallel), and the input formatting instructions customized for that model. For example, for large language models, the instructions require the input to be constructed as a prompt containing system role settings and dialogue history; for task-specific models, the instructions require the input to be constructed as a sequence with a specific format; and for rule engines, the instructions require extracting key entities and trigger words for pattern matching.

[0061] S103, Multi-model Parallel Inference and Standardized Results Collection

[0062] After the scheduling command is issued, the model scheduler sends the formatted input concurrently or sequentially to the target model through a unified model service interface (such as gRPC or high-speed HTTP interface). Figure 4 As shown, each model in the model repository is deployed as a microservice and provides a standardized inference API.

[0063] Upon receiving a request, each model performs inference computation within a set timeout limit (e.g., 500 milliseconds) and returns a structured inference result. The result for each model must adhere to a unified output specification, containing at least the following fields: 1) Predicted intent: one or more intent labels with confidence levels; 2) Filled slots: a key-value dictionary where the key is the slot name and the value is the extracted text fragment and its confidence level; 3) Model metadata, including model ID, version, and inference time. For deep learning models, the internally calculated uncertainty score must also be provided (if supported).

[0064] The system uses a global clock for timeout control. If a model fails to return a result after the timeout, the model scheduler marks it as a failure, and this result will not be included in subsequent fusion, potentially triggering a health degradation for that model. All successfully returned results are collected into a result pool, awaiting further fusion processing. Simultaneously, details of all model calls (such as response time and success rate) are recorded to update model performance profiles and provide data support for future routing decisions.

[0065] S104, Multi-model fusion and collaborative conflict resolution

[0066] Once the results pool is ready, the multi-model fusion module integrates the outputs of multiple models that may be complementary or contradictory into a more accurate and reliable final semantic understanding result. For example... Figure 5 As shown, the fusion process adopts a layered and progressive strategy.

[0067] First, semantic layer fusion is performed. The multi-model fusion module collects the intent predictions and their confidence scores from all models. A weighted voting method based on model weights is used for fusion. The voting weight of each model is dynamically determined by its routing scheduling weight and its historical accuracy. The weighted sum of the scores for each unique intent label is calculated, and the intent with the highest score is selected as the initial fusion intent. If the scores of the first and second highest intents are very close, it is considered an "intent conflict," triggering a conflict handling flag.

[0068] Next, structural layer fusion is performed. For each slot type that needs to be filled, the multi-model fusion module gathers the slot values ​​and confidence scores given by all models. Since different models may have slightly different representations of the same entity (e.g., "Shanghai" and "Shanghai Municipality"), direct voting may not be effective. Therefore, the system adopts a semantic similarity-based clustering fusion method: First, a sentence embedding model is used to convert all candidate slot values ​​into vectors; then, the cosine similarity between vectors is calculated, and hierarchical clustering or DBSCAN clustering is performed to group semantically similar values ​​into one class; finally, for the values ​​within each class, the comprehensive score of the class is calculated based on the weights and confidence scores of its source models, and the "center value" of the class with the highest score (which can be determined by the average of the vectors within the class or the corresponding original text) is selected as the final filling value for that slot. If a slot has two or more classes with similar scores but large semantic differences, then the slot is marked as having a "slot conflict".

[0069] Then, the system introduces a learning-based fusion network for fine-tuning. This learning-based fusion network is a small graph neural network. It is constructed as follows: the output of each model is used as a node in the graph. Node features include the intent confidence of the model's output, the confidence of each slot position, and the uncertainty score. The edge weights between nodes are determined by the similarity between models (e.g., whether they are pre-trained on the same architecture, the overlap of training data) and the historical co-occurrence accuracy statistically analyzed and stored in the model registry during historical interactions. This graph is input into the GNN, and after several rounds of message passing, the GNN outputs an adjustment coefficient for each node. This coefficient reflects the amount of confidence correction for the model's result under the current input and graph structure. Finally, the adjusted weights are used to fine-tune the fusion results from the first two steps, resulting in a fine-tuned set of intents and slots.

[0070] For conflicts detected during the fusion process (intent conflicts or critical slot conflicts), the system initiates a multi-strategy conflict resolution process. Resolution is attempted in priority order:

[0071] 1) Prioritize business strategy and check if any conflicting parties are supported by rule engine outputs. Rule engines typically encode deterministic business rules and key constraints (such as compliance reviews and fixed processes). If so, prioritize the rule engine's judgments due to their high accuracy and business security.

[0072] 2) External knowledge verification: If the business strategy cannot resolve the issue (e.g., both conflicting parties are statistical models), the conflicting entities or statements are linked to an external knowledge base or knowledge graph for verification. For example, for the conflicting slot values ​​"Beijing" and "Nanjing," a geographical knowledge graph is queried to verify which city has the corresponding airport; for the conflicting intents "query balance" and "transfer," a business knowledge base is queried to verify which operation is feasible in the current context. The verification results returned by the knowledge graph have high authority.

[0073] 3) Human-Machine Collaborative Resolution: When the above-mentioned automated methods fail to resolve the conflict, and the conflict directly impacts critical business processes, the system automatically pushes the current dialogue segment, conflict result, and context to the human review platform. Simultaneously, the system sends a friendly prompt to the user (e.g., "Your issue is complex; we are contacting a specialist for you. Please wait."). The human review result, once returned, is not only used to resolve the current conflict but also serves as a high-quality feedback sample for optimizing routing networks and convergence strategies.

[0074] After all fusion and resolution steps are completed, the multi-model fusion module generates the final semantic understanding result. This semantic understanding result is a structured JSON object that clearly contains metadata such as the defined intent, the fully populated slot list, the overall confidence level, and whether conflict resolution has been performed.

[0075] S105, Results Output and Continuous Feedback Learning

[0076] The final semantic understanding results are output in real time to downstream business processing modules, such as dialogue management systems, task execution engines, database query interfaces, or content recommendation systems, driving subsequent intelligent interactions or business processes.

[0077] like Figure 6 As shown, the entire semantic understanding process was recorded in detail by the system, forming a traceable log. This log includes not only the final output, but also the original input text, preprocessed features, routing decision basis, original outputs of each model, intermediate fusion processes, conflict resolution paths, and final user interaction feedback (e.g., whether the user made corrections in the next round, whether the task was successfully executed).

[0078] This log data is fed into the system's feedback learning pool. The feedback learning module periodically (e.g., daily) samples data from the learning pool for online optimization. Optimization primarily targets two core components:

[0079] 1) The routing strategy optimization component fine-tunes the dynamic routing network, using sessions where the user is ultimately satisfied or the task is successfully completed as positive samples, and sessions where the user explicitly corrects the error or the task fails as negative samples. Taking whether the routing decisions in a session lead to good results as the learning objective, reinforcement learning or supervised learning methods are used to incrementally update the parameters of the meta-learning routing network, ensuring that its decision-making strategy continuously reflects real business benefits.

[0080] 2) The performance monitoring and profile update component dynamically updates the performance metrics (accuracy, response speed, stability) of each model in the model registry based on their performance in a large number of requests. These updated metrics will directly affect the model weight calculation in future routing decisions, enabling adaptive resource allocation.

[0081] Through the iterative execution and feedback optimization of the above five core steps, the semantic understanding method described in this invention can form a continuously self-improving intelligent system, maintaining a high level of understanding accuracy and robustness in complex and ever-changing real-world application scenarios.

[0082] Example 2

[0083] This invention also provides a semantic understanding engine based on multi-model collaborative reasoning and self-supervised learning. The engine consists of multiple modules with clearly defined functions, interfaces, and independent deployment capabilities, working collaboratively. These modules interact through service calls, message passing, or shared storage to jointly complete semantic understanding tasks. For example... Figure 7 As shown, it specifically includes:

[0084] Input processing module

[0085] As the unified data entry point of the system, the input processing module's core function is to receive and standardize multimodal raw inputs from the outside, transforming them into structured data objects that are unified within the system and rich in linguistic features, thus providing a reliable foundation for all subsequent analysis and processing.

[0086] The input processing module consists of a set of horizontally scalable, stateless service instances, internally employing a pluggable, modular design. For voice input, the automatic speech recognition component handles the processing, encapsulating streaming and non-streaming recognition engines to convert audio signals into text with timestamps and confidence assessments in real time or in batches. For image input, the optical character recognition component is responsible, integrating an image preprocessing unit and a deep learning-based text detection and recognition model to extract and logically organize text information from images.

[0087] After obtaining the plain text, the text standardization pipeline first calls the word segmentation and lexicalization components to divide the input text into a sequence of basic units that the model can process, based on the linguistic characteristics of the input text. Subsequently, the basic semantic feature encoding component loads a lightweight pre-trained context encoding model to map the lexical sequence into a high-dimensional semantic vector sequence and generate an aggregate vector representing the semantics of the entire sentence. The parallel-running shallow language parsing component then uses its internal named entity recognition model and dependency parser to extract entity information and syntactic structural relationships from the text, respectively.

[0088] Finally, the intermediate representation encapsulation component is responsible for integrating and serializing all generated information into a structured data object. To improve processing efficiency, the input processing module also has a caching component, which can directly return cached preprocessed results for repeated input text.

[0089] Dynamic routing module

[0090] The dynamic routing module is the system's intelligent scheduling center. Its core function is to dynamically decide and schedule one or more of the most suitable semantic understanding models to execute the current task based on the current input features, context, and real-time system status, thereby achieving an optimal balance between understanding accuracy and processing efficiency. The dynamic routing module is deployed in the form of a stateless service cluster, which contains a series of components for collaborative decision-making.

[0091] The core of the dynamic routing module is the routing policy execution engine. This engine executes decision logic in a preset order. The first level of decision-making is performed by the intent classifier component, a classification model fine-tuned on general dialogue data, used to quickly determine the macro-level intent category of the user input. If the macro-level intent points to a specific professional domain, the second level of decision-making is triggered, where the domain classifier component performs fine-grained domain discrimination.

[0092] The core third-level decision-making is handled by the meta-learning routing network component. This component is a lightweight neural network trained using a meta-learning algorithm. During online service, the meta-learning routing network component receives a comprehensive feature vector that integrates the input semantic features, intent probability distribution, domain probability distribution, and real-time load information obtained from the system resource monitor, and outputs weight vectors pointing to each candidate model.

[0093] The model registry is the core data storage within this module, dynamically maintaining metadata for all available model inference services, including service endpoints, health status, capability descriptions, historical performance metrics, and resource consumption profiles. During routing decisions, the execution engine verifies and corrects the original weights output by the routing network based on real-time data from the registry. After the decision is made, the model scheduler executor component is responsible for calling the target model service through a high-performance RPC framework, implementing strict timeout and circuit breaker controls. The performance data generated during its execution is fed back to the model registry in real time to update the service profile.

[0094] Multi-model fusion module

[0095] The function of the multi-model fusion module is to integrate and arbitrate the inference results of multiple heterogeneous models on the same input, which may be inconsistent. Through a hierarchical and learnable fusion and conflict resolution mechanism, it generates a consistent, reliable and high-quality final semantic understanding result.

[0096] The multi-model fusion module is designed as a configurable, multi-stage processing pipeline. The semantic layer fusion unit first collects the intent prediction results from each model, performs weighted voting based on model weights and historical performance to determine the initial fusion intent, and identifies potential intent conflicts. Subsequently, the structural layer fusion unit processes slot filling. It uses a built-in semantic encoder component to vectorize text slot values ​​and aggregates similar values ​​through a clustering analysis component. Based on the comprehensive score of the clusters, it determines the final value of each slot and simultaneously marks slot conflicts.

[0097] To further enhance the robustness of the results, the learning-based fusion network unit constructs a graph with the outputs of each model as nodes. Node features include their confidence levels, and edge weights reflect the similarity between models. Adjustment coefficients for the initial fusion results are generated through graph neural network learning. When the pipeline detects a conflict marker, the conflict resolver is activated. The conflict resolver will sequentially attempt three strategies to resolve the conflict: business rule priority, external knowledge verification, and human-machine collaboration. After all processes are completed, the result assembly and output component encapsulates the final determined intent, slots, and metadata into a standard format and delivers it to downstream systems.

[0098] Model repository module

[0099] The model repository module is the asset management hub for all semantic understanding models in the system. It is responsible for the versioned storage, containerized deployment, service encapsulation, and full lifecycle management of the models, ensuring that model assets are traceable, reusable, and can evolve smoothly.

[0100] The core of the model repository module is the model repository, which persistently stores different versions of model files and related assets. Through the model containerization build component, these assets can be automatically packaged into Docker images containing a complete runtime environment. The service deployment and orchestration component integrates with the underlying container orchestration platform, responsible for deploying model images as elastically scalable inference service instances and automatically registering them to the model registry center of the dynamic routing module.

[0101] To support risk-free model iteration, the model repository module provides complete A / B testing and canary release capabilities. New model versions can be deployed with small-scale deployments, and the model performance evaluation component can compare their differences from the baseline version in business and performance metrics in real time, providing data-driven support for release decisions.

[0102] Feedback Learning Module

[0103] The feedback learning module is the "growth engine" driving the system's continuous self-optimization and evolution. It automatically optimizes the system's core decision-making components and basic models by collecting and analyzing real-world feedback data from the production environment. Specifically, it includes the following sub-modules:

[0104] 1) Data collection and processing pipeline, continuously extracting feedback samples containing the complete processing chain from system logs;

[0105] 2) The routing strategy optimization component periodically fine-tunes the meta-learning routing network in the dynamic routing module using feedback samples;

[0106] 3) Model performance monitoring and profile update component, continuously analyzes the online performance of each model and dynamically updates performance indicators;

[0107] 4) Self-supervised training component: Utilizing a massive amount of anonymized user query text, the basic semantic encoding model is incrementally trained through a specially designed self-supervised task.

[0108] Furthermore, the self-supervised training component within the feedback learning module can automatically construct various pre-training tasks using massive amounts of anonymized user query text, including: a) a masked language modeling task, randomly masking 15% of the words and predicting the masked words; b) a sentence order prediction task, determining whether two sentences represent consecutive dialogue turns; and c) a contrastive learning task, learning semantic similarity through positive and negative samples. These tasks collectively perform periodic incremental training on the system's shared basic semantic encoding model, enabling it to continuously adapt to new linguistic phenomena.

[0109] Monitoring and Security Module

[0110] The monitoring and security module provides comprehensive observability, stability assurance, and security and compliance support for the entire system, serving as the cornerstone for the reliable and trustworthy operation of the system.

[0111] The system monitoring submodule aggregates metrics, logs, and tracking data from the entire process, providing real-time insights through a visual dashboard and timely notifications when anomalies occur via an intelligent alert engine. The interpretability and analysis submodule provides in-depth diagnostic tools, supporting backtracking of the decision-making process for a single prediction and analysis of key influencing factors, thus aiding in algorithm debugging and optimization.

[0112] The security and privacy protection submodule is integrated throughout the entire data processing workflow, implementing entry content filtering, differential privacy protection for training data, and encryption for data transmission and storage. The fairness audit component regularly evaluates the performance consistency of the core model across different user groups to ensure the fairness and compliance of algorithmic decisions.

[0113] In summary, through the meticulous design and collaborative work of the above six modules, a highly available, high-performance semantic understanding engine with continuous learning capabilities has been constructed, which can effectively meet the high requirements of semantic understanding technology in complex and ever-changing real-world application scenarios.

[0114] Example 3

[0115] The present invention also provides an electronic device, including: a processor, a transmitting device, an input device, an output device, and a memory. The processor may be implemented using a general-purpose CPU (Central Processing Unit), a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory may be implemented using a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM), and is used to store computer program code. The computer program code includes computer instructions. When the processor executes the computer instructions, the electronic device executes a method as described in any of the above possible implementation methods.

[0116] Example 4

[0117] The present invention also provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor of an electronic device, cause the processor to perform a method as described in any of the above possible implementations.

[0118] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0119] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A semantic understanding method based on multi-model collaborative reasoning and self-supervised learning, characterized in that, Includes the following steps: S101. Input data acquisition and standardization preprocessing: Receive and convert the original input data into standardized text, perform word segmentation, lexicalization, basic language feature extraction and shallow semantic parsing, and generate structured intermediate representation objects. S102, Dynamic routing decision and intelligent model scheduling: Based on the intermediate representation object and context information, a model scheduling list is generated through multi-level routing decision, and at least one semantic understanding model is intelligently scheduled. S103. Multi-model parallel inference and standardized result collection: Schedule the corresponding models for inference according to the model scheduling list, and collect the standardized inference results returned by each model. S104. Multi-model fusion and conflict collaborative resolution: The collected inference results from multiple models are fused in a hierarchical manner, and the conflicts detected during the fusion process are resolved using multiple strategies to generate the final semantic understanding result; S105. Outputting Results and Continuous Feedback Learning: Outputting the final semantic understanding result and performing feedback learning based on the log data of the processing link to optimize the routing strategy and update the model performance profile. In step S104, the multi-strategy resolution is performed in the following priority order: the judgment result of the rule engine is adopted first; if it cannot be resolved, the conflict content is linked to an external knowledge base or knowledge graph for verification; if it still cannot be resolved, the conflict is submitted to a manual review platform for adjudication, and the adjudication result is used to optimize the system.

2. The method according to claim 1, characterized in that, In step S102, the dynamic routing decision and model intelligent scheduling include: The first level of intent type coarse screening determines the initial intent type of the input through an intent classifier; The second level is fine-grained routing within the domain. If the initial intent type belongs to a preset professional category, the sub-domain is determined by a domain classifier. The third level is adaptive routing decision based on meta-learning. The meta-learning routing network receives a comprehensive feature vector and outputs a weight vector pointing to each candidate model. The comprehensive feature vector includes at least the initial semantic features, intent probability distribution, domain probability distribution, and real-time system load index.

3. The method according to claim 2, characterized in that, The third-level meta-learning-based adaptive routing decision also includes: evaluating the prediction uncertainty of each candidate model for the current input; if the uncertainty of the optimal candidate model exceeds a preset threshold, the decision is to initiate multi-model parallel inference and schedule the top K models with the highest weights to process together.

4. The method according to claim 1, characterized in that, In step S104, the hierarchical fusion includes: semantic layer fusion, which performs weighted voting on the intent predictions of each model's output based on the weights and historical accuracy of each model to determine the initial fusion intent; structural layer fusion, which clusters the slot values ​​output by each model based on semantic similarity and determines the final filling value of each slot according to the weights and confidence of the source models; and learning-based fusion fine-tuning, which constructs a graph with the outputs of each model as nodes and learns the node adjustment coefficients through a graph neural network to fine-tune the initial fusion results.

5. The method according to claim 2, characterized in that, In step S105, the continuous feedback learning includes: incrementally updating the parameters of the meta-learning routing network using reinforcement learning or supervised learning methods based on user feedback or task execution results; dynamically updating the performance indicators of each model according to its performance in a large number of requests to influence the calculation of future routing decision weights; and incrementally training the system's shared basic semantic encoding model using de-identified unlabeled user query text through a self-supervised pre-training task.

6. A semantic understanding system based on multi-model collaborative reasoning and self-supervised learning, wherein the system applies the method as described in any one of claims 1 to 5, characterized in that, include: The input processing module is used to receive and standardize multimodal raw input, and transform it into a structured intermediate representation object rich in linguistic features; A dynamic routing module, connected to the input processing module, is used to intelligently schedule at least one semantic understanding model based on the intermediate representation object, context, and system state through multi-level routing decisions. The multi-model fusion module, connected to the dynamic routing module, is used to perform hierarchical fusion and conflict resolution on the inference results of multiple models to generate the final semantic understanding result. The model repository module is used for versioned storage, containerized deployment, and service management of semantic understanding models; The feedback learning module, connected to the dynamic routing module and the model repository module, is used to collect and process link logs, optimize routing strategies, and update model performance profiles. The monitoring and security module provides system observability, security, and compliance assurance.

7. The system according to claim 6, characterized in that, The dynamic routing module includes: The routing policy execution engine is used to execute multi-level routing decision logic in a preset order; The intent classifier component and the domain classifier component are built into the routing strategy execution engine and are used to perform coarse screening of the input intent type and fine-grained domain discrimination when the intent type belongs to a preset professional category. The meta-learning routing network component is a lightweight neural network trained by a model-independent meta-learning algorithm. It is built into the routing policy execution engine and is used to receive a comprehensive feature vector that integrates semantic features, classification probabilities and system load information, and output a weight vector pointing to each candidate model. The model registry is used to dynamically maintain the metadata, real-time health status, and historical performance metrics of all available model inference services. The model scheduling executor component is used to schedule the target model through a remote service call interface based on the decision result of the routing strategy execution engine, and to implement timeout and circuit breaker control.

Images