A multi-scene intelligent consultation system based on a pre-trained large model

Abstract

The application discloses a multi-scene intelligent consultation system based on a pre-training large model, and aims to solve the problems of the existing consultation system in multi-modal understanding, semantic consistency, adaptive modeling and expert ability calling. The system comprises a multi-modal input recognition module, a semantic state cognitive map construction module, an instruction translation module, a self-consistent cognitive feedback module, an expert model scheduling module, an instruction response generation module and a system scheduling and resource management module, and can realize multi-modal information fusion, semantic intention analysis, expert knowledge matching and intelligent response generation, and is widely applicable to intelligent customer service, medical diagnosis, financial consultation and other multi-scene intelligent interaction applications.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a multi-scenario intelligent consultation system based on a pre-trained large model. Background Technology

[0002] Against the backdrop of the convergence of artificial intelligence and natural language processing, constructing intelligent consultation systems with multimodal understanding and semantically consistent expression has become a key direction in research on intelligent human-computer interaction driven by complex tasks. In recent years, with the remarkable capabilities of pre-trained large models (such as GPT, BERT, and T5) in semantic understanding, language generation, and task generalization, academia and industry have been exploring their value in consultation assistance applications across various scenarios, including government services, medical consultations, and financial advice. Most existing intelligent consultation systems rely on rule bases, limited-domain corpora for training, or lightweight dialogue models, lacking the ability to model complex semantic structures. This makes it difficult to achieve unified understanding and context preservation under multimodal inputs, resulting in deficiencies in system response accuracy, generalization ability, and interactive coherence.

[0003] Current technologies have attempted to improve question-answering performance by introducing pre-trained models or by using scene templates and rule engines for intent recognition and task response. However, these methods are mostly limited to unimodal input parsing and static semantic graph management, making it difficult to support unified representation of multimodal information such as speech, text, and images, as well as dynamic scheduling of expert knowledge across multiple scenarios. Furthermore, existing systems generally lack cognitive consistency verification mechanisms, often failing to guarantee the consistency of semantic transmission and the reliability of response logic when dealing with user input ambiguity, contextual ambiguity, or command conflicts. Some studies have introduced dialogue state management strategies to achieve intent tracking, but their graph structures are coarse, and command translation lacks intermediate structure support, leading to significant performance degradation in complex interactive tasks.

[0004] In terms of model scheduling and system resource management, traditional methods mostly adopt a static service architecture, which cannot tailor and load expert models and optimize parameters based on task complexity, scene labels, and device status, thus limiting the deployment efficiency and edge adaptability of models. At the same time, existing intelligent systems lack a structured closed-loop mechanism in cognitive feedback, and cannot dynamically correct and reconstruct structures after identifying structural anomalies and semantic deviations, resulting in response deviations and increased interaction failure rates.

[0005] Therefore, how to provide a multi-scenario intelligent consultation system based on a pre-trained large model is a key technical problem that urgently needs to be solved. Summary of the Invention

[0006] One objective of this invention is to propose a multi-scenario intelligent consultation system based on a pre-trained large model. This invention establishes a closed-loop structure from user intent recognition to expert knowledge invocation and response generation by constructing modules such as multimodal input recognition, semantic state cognitive graph generation, instruction translation and consistency verification, expert model scheduling and response, and system resource management. It adopts semantic state vector to drive the construction of intermediate semantic instruction structure and introduces a self-consistent cognitive feedback mechanism to ensure semantic consistency. It has the advantages of strong multimodal understanding, accurate semantic parsing, wide scenario adaptability, and flexible model scheduling.

[0007] A multi-scenario intelligent consultation system based on a pre-trained large model according to an embodiment of the present invention includes the following modules:

[0008] A multimodal input recognition module is used to receive text data, voice data, and image data input by the user, and convert the data into structured input data in a unified format;

[0009] The semantic state cognitive graph construction module is used to construct semantic state vectors based on the structured input data and generate a semantic state cognitive graph containing user intent, context information and scene labels;

[0010] The instruction translation module is used to parse the semantic state vectors in the semantic state cognitive graph into intermediate instruction representations, and perform structured mapping on the intermediate instruction representations to generate intermediate semantic instruction structures.

[0011] The self-consistent cognitive feedback module is used to perform semantic feedback on the intermediate semantic instruction structure. Figure 1 Consistency verification, and dynamic updating of the intermediate semantic instruction structure based on a cognitive bias correction mechanism;

[0012] The expert model scheduling module is used to schedule the matched expert models according to the scene labels in the semantic state cognitive graph, and load the corresponding expert model parameters and response strategies.

[0013] The instruction response generation module is used to input the updated intermediate semantic instruction structure into the scheduled expert model, generate structured response content, and output it in a user-understandable format.

[0014] The system scheduling and resource management module is used to execute expert model pruning and parameter distillation strategies based on task complexity, scenario type, and device resource status.

[0015] Optionally, modules can be integrated using the following methods:

[0016] S1. The multimodal input recognition module receives text data, voice data, and image data input by the user, respectively, and performs format parsing and modal standardization processing on the input data to generate structured input data in a unified format.

[0017] S2. Input the structured input data into the semantic state cognitive graph construction module, extract semantic elements based on the structured input data, generate semantic state vectors, and construct a semantic state cognitive graph containing user intent, context information and scene labels;

[0018] S3. The instruction translation module receives the semantic state vector in the semantic state cognitive graph, parses the semantic state vector into an intermediate instruction representation, and performs a structured mapping on the intermediate instruction representation to generate an intermediate semantic instruction structure.

[0019] S4. Input the intermediate semantic instruction structure into the self-consistent cognitive feedback module. The self-consistent cognitive feedback module performs consistency verification and dynamic updates on the intermediate semantic instruction structure according to the preset cognitive bias correction mechanism, and outputs the updated semantic state vector.

[0020] S5. The expert model scheduling module receives scene labels from the semantic state cognitive graph, selects a matching expert model based on the scene labels, and loads the corresponding expert model parameters and response strategies.

[0021] S6. Input the intermediate semantic instruction structure into the scheduled expert model, and have the instruction response generation module generate structured response content and output the response result in a user-understandable format.

[0022] S7. The system scheduling and resource management module collects the computational load information and device resource status of the current task, and executes the pruning configuration and parameter distillation strategy related to the expert model.

[0023] Optionally, S2 specifically includes:

[0024] S21. Perform syntactic analysis and semantic segmentation on the structured input data to identify semantic elements corresponding to user task intent, context-related entities, and scene indicator words;

[0025] S22. The semantic elements are encoded into user intent vectors, context vectors and scene tag vectors according to their categories, and then standardized in a unified dimension to form sub-vectors of the three categories.

[0026] S23. The sub-vectors of the three categories are concatenated and combined according to the preset semantic slot order to generate a semantic state vector. The preset semantic slots include user intent slot, context slot and scene label slot in sequence.

[0027] S24. Based on the association and semantic coupling strength between slots in the semantic state vector, establish a structured node set and connection relationship to construct a semantic state cognitive graph.

[0028] S25. The semantic state cognitive graph includes: user intent nodes for describing task requirements, context information nodes for carrying the input semantic context, and scene label nodes for identifying the application environment.

[0029] S26. The above nodes are connected by directed edges. The edges are established based on the following criteria: semantic edges based on semantic dependencies, temporal edges based on contextual time order, and statistical edges based on semantic co-occurrence frequency. Each edge is accompanied by an edge weight attribute value.

[0030] Optionally, S23 specifically includes:

[0031] S231. During the system initialization phase, a semantic slot order is preset, which includes user intent slot, context slot and scene tag slot in sequence. The slot order remains fixed during system operation and is prohibited from being dynamically modified.

[0032] S232. Set a unified vector dimension standard for the three types of sub-vectors. All sub-vectors must be dimension aligned before splicing. If any semantic element of any category is missing, the corresponding slot is filled with zero vectors to maintain the structural stability of the semantic state vector.

[0033] S233. In the structure of the semantic state vector, each slot is an independent semantic domain. The system has a semantic consistency rule. If there is a conflict between the user intent slot and the context slot, the vector splicing process will be blocked and a correction request will be returned.

[0034] S234. In the splicing operation, a linear connection method is used to concatenate the sub-vectors of the three categories according to the preset semantic slot order to form a complete semantic state vector.

[0035] S235. The semantic state vector is constructed as a one-dimensional fixed-length multi-slot structure, wherein each slot is used as an independent node input in the downstream graph construction process and is not subjected to vector compression and fusion processing.

[0036] S236. The semantic state vector is appended with slot labels to represent the type, filling status and confidence of each slot, which is used to guide the generation of edge connection relationships and the allocation of node weights in graph construction.

[0037] Optionally, S3 specifically includes:

[0038] S31. The instruction translation module receives the semantic state vector from the semantic state cognitive graph, performs matching analysis on the semantic state vector, and extracts the semantic backbone for task recognition.

[0039] S32. Map the extracted semantic backbone content into an intermediate instruction representation, wherein the intermediate instruction representation includes an operation type field, an object identifier field, and a constraint condition field;

[0040] S33. Perform structured parsing on each field in the intermediate instruction representation to construct a slot-type intermediate semantic instruction structure;

[0041] S34. In the structured mapping process, a multi-dimensional mapping strategy is adopted. Based on the slot combination features of the semantic state vector, the predefined instruction template is matched first. If the match fails, the system's internally configured rule set is called to complete the structure parsing.

[0042] S35. The intermediate semantic instruction structure supports a cross-scenario instruction alignment mechanism, and standardizes the consistency of instruction expression in different scenarios through a semantic tag mapping relationship table.

[0043] Optionally, S33 specifically includes:

[0044] S331. The slot-type intermediate semantic instruction structure includes an operation type slot, an object identifier slot, and a constraint condition slot, which respectively correspond to the operation type field, object identifier field, and constraint condition field in the intermediate instruction representation.

[0045] S332. The above slots are arranged in a fixed order and encapsulated using a unified data structure. Each slot structure includes three fields: slot label, slot value, and slot position confidence.

[0046] S333. The slot value is filled by the instruction translation module based on the semantic element matching result in the semantic state vector. The slot position confidence is directly generated by the matching score and is used to represent the semantic reliability of the slot.

[0047] S334. When a slot value is missing, the system calls the corresponding default parameter to fill it according to the slot type. When a slot value exists but its confidence level is lower than the set threshold, a preset placeholder is filled into the slot to indicate that the slot is in an untrusted state.

[0048] S335. The slot structure supports a field expansion mechanism, allowing the addition of extended slots while maintaining structural consistency.

[0049] S336. The slot-type intermediate semantic instruction structure must pass the inter-slot semantic consistency check before output. If a semantic conflict is detected between any two slots, the system stops the downward transmission process of the intermediate semantic instruction structure and sends a structure reconstruction request to the upstream module.

[0050] Optionally, S4 specifically includes:

[0051] S41. Input the intermediate semantic instruction structure into the self-consistent cognitive feedback module, and perform semantic consistency analysis on the operation type slot, object identifier slot and constraint condition slot respectively.

[0052] S42. Based on the semantic similarity calculation results between the slot content and the corresponding semantic elements in the semantic state vector, construct a deviation scoring matrix to characterize the degree of semantic offset between the slot content and the current cognitive state of the system.

[0053] S43. When the deviation score of any slot exceeds the set threshold, the system triggers the cognitive deviation correction mechanism and reconstructs the slot content according to the preset semantic correction rule set.

[0054] S44. Slot content reconstruction includes slot value replacement, semantic weakening processing, and semantic conflict marking.

[0055] S45. The reconstructed intermediate semantic instruction structure will undergo structural verification again, including:

[0056] Slot structure integrity check is used to determine whether any required slots are missing;

[0057] Slot type matching validation is used to confirm whether the slot content conforms to the semantic scope defined by the slot.

[0058] Inter-slot logic consistency check is used to detect whether there are conflicting expressions, logical inversions, or mutual exclusion relationships between slots;

[0059] If any check fails, the system refuses to pass the structure to downstream modules and marks it as an abnormal structure state.

[0060] S46. The qualified intermediate semantic instruction structure is used to generate the updated semantic state vector. The update operation rewrites the corresponding slot representation in the semantic state vector based on the semantic elements in the current instruction structure.

[0061] Optionally, S42 specifically includes:

[0062] S421. Each score value in the deviation scoring matrix is ​​calculated using a multi-factor weighted scoring function, the formula of which is:

[0063] ;

[0064] in, Indicates the first The semantic deviation score between each slot and the corresponding semantic element in the semantic state vector. For slot embedding vectors, This is the semantic representation vector corresponding to the slot in the semantic state vector. Represents the vector dot product. and These are the L2 norms of the vectors. This represents the confidence score for the current slot. This indicates the historical number of semantic conflicts that have occurred between this slot and other slots. , , The pre-defined weighting factors for the system are used to control the contribution weights of semantic similarity scoring, confidence enhancement, and historical conflict penalty, respectively.

[0065] S422. Assign hierarchical weights to all slot deviation scores according to slot type to reflect the differences in the system's consistency judgment standards for different semantic slots. Among them, the operation type slot score has the highest weight, and the constraint condition slot score has the lowest weight.

[0066] S423. All scoring results are normalized to form the final deviation scoring matrix. The normalization adopts the maximum and minimum scaling method so that the scoring values ​​are distributed in the interval [0,1].

[0067] S424. The system sets a deviation scoring threshold for each type of slot, with the operation type slot having the smallest tolerance and the restriction condition slot having the largest tolerance. If the score of a slot exceeds the corresponding threshold, the slot is judged to have semantic offset.

[0068] S425. When the score of any slot in the deviation scoring matrix exceeds its set threshold, the system marks the slot as a high deviation state.

[0069] Optionally, S45 specifically includes:

[0070] S451. The structure verification is performed based on the system's preset set of required slots. The set of required slots is dynamically loaded according to the task type and scene label. The verification content includes whether the slot exists and whether the slot value is not empty.

[0071] S452. The slot type matching verification is performed based on the built-in slot semantic definition table of the system. The verification criteria include whether the data type and semantic category of the slot value are consistent with its defined type. If there is a mismatch, it is marked as a type anomaly.

[0072] S453, The inter-slot logic consistency check includes the judgment of the following three types of logical conflicts: semantic conflict, logical reversal, and mutual exclusion;

[0073] S454. The system performs each check in the order of integrity, type matching, and logical consistency. If any check fails, the subsequent process is terminated immediately, the system refuses to pass the intermediate semantic instruction structure to the downstream module, and marks the structure as an abnormal state.

[0074] Optionally, S7 specifically includes:

[0075] S71. The system scheduling and resource management module collects the expert model call request, task complexity identifier and device operating status of the current task, and constructs a task resource profile, which includes: task type, expert model calculation requirements, current system load and remaining computing power.

[0076] S72. Based on the task resource profile, perform layer-level pruning configuration on the expert model using a structural pruning strategy. The structural pruning strategy deletes low-importance nodes and redundant channels based on the node importance distribution in the expert model calculation graph, and generates a pruned expert model structure graph.

[0077] S73. Based on the task resource profile and combined with the semantic generality weight of the target scene, select the distillation source expert model and the distillation target expert model, and execute the parameter distillation process. The distillation includes two stages: label alignment distillation and intermediate representation matching distillation. The parameters of the distillation target expert model are updated using the feature matching loss function.

[0078] S74. The system schedules and sorts tasks based on the task priority of the expert model and the free window of the device, prioritizes loading the trimmed expert model structure, dynamically allocates video memory and computing threads, and performs matching and binding of software and hardware resources.

[0079] S75. When the system detects that the expert model fails to load or there is a runtime resource conflict, it automatically invokes the rollback mechanism to restore the default version of the expert model structure.

[0080] The beneficial effects of this invention are:

[0081] (1) This invention, by constructing a multimodal input recognition module and a semantic state cognitive graph construction module, has for the first time achieved unified parsing and structured expression of heterogeneous user inputs such as text, speech and images. It adopts a fixed-slot semantic state vector design, combined with syntax analysis, semantic segmentation and entity recognition operations, to encode user intent, context information and scene labels respectively and splice them into a one-dimensional multi-slot structure. Furthermore, it constructs a cognitive graph structure with associated edge weights, which greatly improves the semantic extraction accuracy and context modeling ability of the system for multimodal inputs, and solves the problems of weak context perception and limited input understanding ability of existing systems.

[0082] (2) The present invention designs an instruction translation module and a self-consistent cognitive feedback mechanism based on an intermediate semantic instruction structure. The system generates a standardized slot-type intermediate instruction structure based on the semantic state vector, and performs consistency verification on the operation type, object identifier and constraint condition slot after translation. The degree of instruction deviation is judged by constructing a semantic deviation scoring matrix and a slot position reliability mechanism. When the deviation score exceeds the set threshold, the cognitive deviation correction process is automatically triggered to reconstruct the slot value and verify its structural integrity, type matching and logical consistency. This ensures the reliability of the intermediate instruction transmission process and the stability of the downstream task execution, and solves the problems of instruction ambiguity and semantic conflict in multi-round interaction.

[0083] (3) Based on the scene labels in the semantic state graph, the present invention designs an expert model scheduling and system resource management mechanism, matches corresponding expert models for different task scenarios, and performs model structure pruning and parameter distillation in combination with the current device load and task complexity. The model inference structure is optimized by node importance pruning and feature alignment distillation strategies. While ensuring the accuracy of model response, the system's computing resource consumption and loading time are significantly reduced. The on-demand calling and lightweight operation of expert models are realized, which is suitable for edge device deployment environment and has good scalability and engineering practical value. Attached Figure Description

[0084] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0085] Figure 1 This is a schematic diagram of the overall architecture of a multi-scenario intelligent consultation system based on a pre-trained large model proposed in this invention;

[0086] Figure 2 This is a flowchart of the semantic state vector generation and graph structure construction of a multi-scenario intelligent consultation system based on a pre-trained large model proposed in this invention.

[0087] Figure 3 This is a flowchart of the intermediate semantic instruction generation and consistency verification process for a multi-scenario intelligent consultation system based on a pre-trained large model proposed in this invention. Detailed Implementation

[0088] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0089] refer to Figures 1-3 A multi-scenario intelligent consultation system based on a pre-trained large model includes the following modules:

[0090] A multimodal input recognition module is used to receive text data, voice data, and image data input by the user, and convert the data into structured input data in a unified format;

[0091] The semantic state cognitive graph construction module is used to construct semantic state vectors based on the structured input data and generate a semantic state cognitive graph containing user intent, context information and scene labels;

[0092] The instruction translation module is used to parse the semantic state vectors in the semantic state cognitive graph into intermediate instruction representations, and perform structured mapping on the intermediate instruction representations to generate intermediate semantic instruction structures.

[0093] The self-consistent cognitive feedback module is used to perform semantic feedback on the intermediate semantic instruction structure. Figure 1 Consistency verification, and dynamic updating of the intermediate semantic instruction structure based on a cognitive bias correction mechanism;

[0094] The expert model scheduling module is used to schedule the matched expert models according to the scene labels in the semantic state cognitive graph, and load the corresponding expert model parameters and response strategies.

[0095] The instruction response generation module is used to input the updated intermediate semantic instruction structure into the scheduled expert model, generate structured response content, and output it in a user-understandable format.

[0096] The system scheduling and resource management module is used to execute expert model pruning and parameter distillation strategies based on task complexity, scenario type, and device resource status.

[0097] This invention provides a multi-scenario intelligent consultation system based on a pre-trained large model. It can receive and uniformly process multimodal data such as text, voice, and images input by users, improving the system's adaptability to complex inputs. By constructing a semantic state cognitive graph, it achieves structured expression of user intent, context, and scene information, enhancing semantic understanding accuracy. The instruction translation module supports efficient conversion of semantics to instructions, while the self-consistent cognition module performs consistency verification and dynamic updates based on feedback, achieving cognitive closed-loop optimization. Combined with the expert model scheduling and instruction response generation modules, the system can dynamically match expert models according to semantic states and output reasonable responses, possessing advantages such as intelligence, scalability, and high responsiveness.

[0098] A multi-scenario intelligent consultation system based on a pre-trained large model, wherein the modules are implemented through the following methods:

[0099] S1. The multimodal input recognition module receives text data, voice data, and image data input by the user, respectively, and performs format parsing and modal standardization processing on the input data to generate structured input data in a unified format.

[0100] In this embodiment, step S1 specifically includes: the text recognition unit, speech recognition unit, and image recognition unit in the multimodal input recognition module respectively receive text data, speech data, and image data input by the user; the text recognition unit directly acquires structured text information; the speech recognition unit uses a speech transcription model to extract acoustic features and perform semantic recognition processing on the speech signal, outputting corresponding text information; the image recognition unit extracts target semantic tags and scene information from the image based on an image content recognition model, converting them into a parsable semantic representation; subsequently, the system uniformly deconstructs and labels the input data of the above different modalities, including identifying the input source, extracting keyword slots, constructing semantic vector representations, and generating structurally consistent multimodal fusion input data under a unified temporal framework, providing standardized input support for subsequent semantic graph construction and instruction generation. This step ensures the consistency and structural standardization of the system when facing multiple types of input, improving the accuracy and stability of the overall intelligent response process.

[0101] S2. Input the structured input data into the semantic state cognitive graph construction module, extract semantic elements based on the structured input data, generate semantic state vectors, and construct a semantic state cognitive graph containing user intent, context information and scene labels;

[0102] In this embodiment, S2 specifically includes:

[0103] S21. Perform syntactic analysis and semantic segmentation on the structured input data to identify semantic elements corresponding to user task intent, context-related entities, and scene indicator words;

[0104] S22. The semantic elements are encoded into user intent vectors, context vectors and scene tag vectors according to their categories, and then standardized in a unified dimension to form sub-vectors of the three categories.

[0105] S23. The sub-vectors of the three categories are concatenated and combined according to the preset semantic slot order to generate a semantic state vector. The preset semantic slots include user intent slot, context slot and scene label slot in sequence.

[0106] S24. Based on the association and semantic coupling strength between slots in the semantic state vector, establish a structured node set and connection relationship to construct a semantic state cognitive graph.

[0107] S25. The semantic state cognitive graph includes: user intent nodes for describing task requirements, context information nodes for carrying the input semantic context, and scene label nodes for identifying the application environment.

[0108] S26. The above nodes are connected by directed edges. The edges are established based on the following criteria: semantic edges based on semantic dependencies, temporal edges based on contextual time order, and statistical edges based on semantic co-occurrence frequency. Each edge is accompanied by an edge weight attribute value.

[0109] In this embodiment, S23 specifically includes:

[0110] S231. During the system initialization phase, a semantic slot order is preset, which includes user intent slot, context slot and scene tag slot in sequence. The slot order remains fixed during system operation and is prohibited from being dynamically modified.

[0111] S232. Set a unified vector dimension standard for the three types of sub-vectors. All sub-vectors must be dimension aligned before splicing. If any semantic element of any category is missing, the corresponding slot is filled with zero vectors to maintain the structural stability of the semantic state vector.

[0112] S233. In the structure of the semantic state vector, each slot is an independent semantic domain. The system has a semantic consistency rule. If there is a conflict between the user intent slot and the context slot, the vector splicing process will be blocked and a correction request will be returned.

[0113] S234. In the splicing operation, a linear connection method is used to concatenate the sub-vectors of the three categories according to the preset semantic slot order to form a complete semantic state vector.

[0114] S235. The semantic state vector is constructed as a one-dimensional fixed-length multi-slot structure, wherein each slot is used as an independent node input in the downstream graph construction process and is not subjected to vector compression and fusion processing.

[0115] S236. The semantic state vector is appended with slot labels to represent the type, filling status and confidence of each slot, which is used to guide the generation of edge connection relationships and the allocation of node weights in graph construction.

[0116] This invention utilizes a semantic state cognitive graph construction module to categorize semantic elements in structured input data into three types: user intent vectors, context vectors, and scene label vectors. Semantic slot information is extracted from each category, and after standardized processing according to a unified dimension, these are concatenated in the order of the semantic slots to form semantic state vectors, thus constructing a structurally consistent semantic state graph. This approach improves the logical coherence and expressive consistency of semantic vector combinations, effectively avoiding semantic conflicts and structural chaos. It ensures the stability and accuracy of connections between nodes in the graph, providing a high-precision and highly consistent semantic expression foundation for subsequent semantic reasoning and expert instruction generation.

[0117] S3. The instruction translation module receives the semantic state vector in the semantic state cognitive graph, parses the semantic state vector into an intermediate instruction representation, and performs a structured mapping on the intermediate instruction representation to generate an intermediate semantic instruction structure.

[0118] In this embodiment, S3 specifically includes:

[0119] S31. The instruction translation module receives the semantic state vector from the semantic state cognitive graph, performs matching analysis on the semantic state vector, and extracts the semantic backbone for task recognition.

[0120] S32. Map the extracted semantic backbone content into an intermediate instruction representation, wherein the intermediate instruction representation includes an operation type field, an object identifier field, and a constraint condition field;

[0121] S33. Perform structured parsing on each field in the intermediate instruction representation to construct a slot-type intermediate semantic instruction structure;

[0122] S34. In the structured mapping process, a multi-dimensional mapping strategy is adopted. Based on the slot combination features of the semantic state vector, the predefined instruction template is matched first. If the match fails, the system's internally configured rule set is called to complete the structure parsing.

[0123] S35. The intermediate semantic instruction structure supports a cross-scenario instruction alignment mechanism, and standardizes the consistency of instruction expression in different scenarios through a semantic tag mapping relationship table.

[0124] In this embodiment, S33 specifically includes:

[0125] S331. The slot-type intermediate semantic instruction structure includes an operation type slot, an object identifier slot, and a constraint condition slot, which respectively correspond to the operation type field, object identifier field, and constraint condition field in the intermediate instruction representation.

[0126] S332. The above slots are arranged in a fixed order and encapsulated using a unified data structure. Each slot structure includes three fields: slot label, slot value, and slot position confidence.

[0127] S333. The slot value is filled by the instruction translation module based on the semantic element matching result in the semantic state vector. The slot position confidence is directly generated by the matching score and is used to represent the semantic reliability of the slot.

[0128] S334. When a slot value is missing, the system calls the corresponding default parameter to fill it according to the slot type. When a slot value exists but its confidence level is lower than the set threshold, a preset placeholder is filled into the slot to indicate that the slot is in an untrusted state.

[0129] S335. The slot structure supports a field expansion mechanism, allowing the addition of extended slots while maintaining structural consistency.

[0130] S336. The slot-type intermediate semantic instruction structure must pass the inter-slot semantic consistency check before output. If a semantic conflict is detected between any two slots, the system stops the downward transmission process of the intermediate semantic instruction structure and sends a structure reconstruction request to the upstream module.

[0131] This invention achieves a structured mapping from semantic state vectors to intermediate instruction representations through an instruction translation module. It supports parsing the semantic backbone of user tasks into intermediate instruction representations containing operation types, object identifiers, and constraints, and performs slot-based structural parsing on each field to generate a unified and standardized intermediate semantic instruction structure. The system employs a multi-dimensional mapping mechanism and template matching strategy, supporting fault-tolerant filling and confidence expansion for different semantic slot types, enhancing the system's robustness and semantic coverage in multi-scenario task parsing. A semantic consistency verification mechanism is introduced into the instruction structure to ensure its stability under contextual evolution, providing semantic closed-loop support for subsequent scenario instruction generation and expert model invocation.

[0132] S4. Input the intermediate semantic instruction structure into the self-consistent cognitive feedback module. The self-consistent cognitive feedback module performs consistency verification and dynamic updates on the intermediate semantic instruction structure according to the preset cognitive bias correction mechanism, and outputs the updated semantic state vector.

[0133] In this embodiment, S4 specifically includes:

[0134] S41. Input the intermediate semantic instruction structure into the self-consistent cognitive feedback module, and perform semantic consistency analysis on the operation type slot, object identifier slot and constraint condition slot respectively.

[0135] S42. Based on the semantic similarity calculation results between the slot content and the corresponding semantic elements in the semantic state vector, construct a deviation scoring matrix to characterize the degree of semantic offset between the slot content and the current cognitive state of the system.

[0136] S43. When the deviation score of any slot exceeds the set threshold, the system triggers the cognitive deviation correction mechanism and reconstructs the slot content according to the preset semantic correction rule set.

[0137] S44. Slot content reconstruction includes slot value replacement, semantic weakening processing, and semantic conflict marking.

[0138] S45. The reconstructed intermediate semantic instruction structure will undergo structural verification again, including:

[0139] Slot structure integrity check is used to determine whether any required slots are missing;

[0140] Slot type matching validation is used to confirm whether the slot content conforms to the semantic scope defined by the slot.

[0141] Inter-slot logic consistency check is used to detect whether there are conflicting expressions, logical inversions, or mutual exclusion relationships between slots;

[0142] If any check fails, the system refuses to pass the structure to downstream modules and marks it as an abnormal structure state.

[0143] S46. The qualified intermediate semantic instruction structure is used to generate the updated semantic state vector. The update operation rewrites the corresponding slot representation in the semantic state vector based on the semantic elements in the current instruction structure.

[0144] In this embodiment, S42 specifically includes:

[0145] S421. Each score value in the deviation scoring matrix is ​​calculated using a multi-factor weighted scoring function, the formula of which is:

[0146] ;

[0147] in, Indicates the first The semantic deviation score between each slot and the corresponding semantic element in the semantic state vector. For slot embedding vectors, This is the semantic representation vector corresponding to the slot in the semantic state vector. Represents the vector dot product. and These are the L2 norms of the vectors. This represents the confidence score for the current slot. This indicates the historical number of semantic conflicts that have occurred between this slot and other slots. , , The pre-defined weighting factors for the system are used to control the contribution weights of semantic similarity scoring, confidence enhancement, and historical conflict penalty, respectively.

[0148] In this invention, a deviation scoring function is introduced, which is a fundamental computational mechanism necessary for achieving semantic slot-level semantic consistency assessment. This function comprehensively considers the cosine similarity between corresponding semantic elements in the slot's semantic representation vector and semantic state vector, the confidence weight of the current slot, and the historical conflict distance between the slot and other slots. It generates a numerical deviation score for each slot through a multi-factor weighted summation method. This score quantifies the reliability and consistency of the slot's semantics, providing a clear mathematical basis for the deviation correction mechanism.

[0149] Compared to traditional methods that rely on rule matching or static template judgment, this scoring function is computable, adjustable, and adaptive. It can dynamically assess deviation risks based on different contexts and semantic evolution, thereby achieving more accurate slot-level consistency determination. By introducing this function, the system can automatically detect semantic structure anomalies and trigger a feedback mechanism without interrupting the task flow, effectively supporting the closed-loop control process of "deviation detection—structure verification—semantic rewrite" in the self-consistent cognitive feedback module described in this invention.

[0150] S422. Assign hierarchical weights to all slot deviation scores according to slot type to reflect the differences in the system's consistency judgment standards for different semantic slots. Among them, the operation type slot score has the highest weight, and the constraint condition slot score has the lowest weight.

[0151] S423. All scoring results are normalized to form the final deviation scoring matrix. The normalization adopts the maximum and minimum scaling method so that the scoring values ​​are distributed in the interval [0,1].

[0152] S424. The system sets a deviation scoring threshold for each type of slot, with the operation type slot having the smallest tolerance and the restriction condition slot having the largest tolerance. If the score of a slot exceeds the corresponding threshold, the slot is judged to have semantic offset.

[0153] S425. When the score of any slot in the deviation scoring matrix exceeds its set threshold, the system marks the slot as a high deviation state.

[0154] In this embodiment, S45 specifically includes:

[0155] S451. The structure verification is performed based on the system's preset set of required slots. The set of required slots is dynamically loaded according to the task type and scene label. The verification content includes whether the slot exists and whether the slot value is not empty.

[0156] S452. The slot type matching verification is performed based on the built-in slot semantic definition table of the system. The verification criteria include whether the data type and semantic category of the slot value are consistent with its defined type. If there is a mismatch, it is marked as a type anomaly.

[0157] S453, The inter-slot logic consistency check includes the judgment of the following three types of logical conflicts: semantic conflict, logical reversal, and mutual exclusion;

[0158] S454. The system performs each check in the order of integrity, type matching, and logical consistency. If any check fails, the subsequent process is terminated immediately, the system refuses to pass the intermediate semantic instruction structure to the downstream module, and marks the structure as an abnormal state.

[0159] This invention establishes a cognitive bias correction mechanism based on semantic slots as the basic unit by performing consistency verification and dynamic updates on the intermediate semantic instruction structure through a self-consistent cognitive feedback module, thereby achieving precise iterative updates of the instruction-level semantic state. The system first constructs a multi-factor weighted bias scoring matrix based on semantic similarity, semantic confidence, and historical dependencies. This matrix evaluates the semantic consistency between the slot fields in the semantic state vector and the current knowledge state. When the score of any slot exceeds a preset threshold, the cognitive bias correction mechanism is triggered, performing semantic fuzzification, fault-tolerant reconstruction, and slot content replacement on the corresponding field. This enhances the semantic robustness and reasoning robustness of the system in inconsistent contexts. Furthermore, the system introduces three strategies in the structure verification process: structural constraint verification, slot type matching verification, and logical consistency verification. These strategies perform multiple screenings on the intermediate instruction structure from the dimensions of syntactic completeness, semantic standardization, and logical coherence, ensuring that the updated semantic state vector structure is legal, semantically reliable, and logically closed-loop. This module enables a leap from one-way instruction generation to two-way semantic verification, providing support for the system to build an adaptive, closed-loop intelligent semantic control mechanism. It significantly improves the accuracy of intelligent response, instruction generation, and cognitive consistency in complex contexts, which is the key innovation of this invention that distinguishes it from existing technologies.

[0160] S5. The expert model scheduling module receives scene labels from the semantic state cognitive graph, selects a matching expert model based on the scene labels, and loads the corresponding expert model parameters and response strategies.

[0161] In this embodiment, step S5 specifically includes: the expert model scheduling module receiving scene tags extracted from the semantic state cognitive graph. These scene tags reflect the operating environment, interaction target, and service context of the current semantic task. The system constructs a mapping relationship between scene tags and expert models, performs association screening on multiple candidate expert models, and selects the expert model with the highest matching degree based on a preset matching degree evaluation mechanism, comprehensively considering factors such as the model's service domain adaptability, resource consumption, response timeliness, and historical performance. Subsequently, the system loads the expert model's structural configuration, inference parameters, and response strategy, and completes interface adaptation with the current semantic state vector, ensuring stable operation of the expert model during task execution. Simultaneously, the system supports dynamically adjusting the inference path of the loaded model when the semantic state changes, or triggering an expert model replacement mechanism based on feedback signals, achieving adaptive updates to task processing strategies in complex semantic scenarios. This step improves the accuracy and flexibility of expert model invocation, ensuring the system's intelligent response efficiency and stability in multi-task, multi-scenario environments.

[0162] S6. Input the intermediate semantic instruction structure into the scheduled expert model, and have the instruction response generation module generate structured response content and output the response result in a user-understandable format.

[0163] In this embodiment, step S6 specifically includes: inputting the intermediate semantic instruction structure, which has undergone structural normalization, into the target expert model selected and loaded by the expert model scheduling module. Based on the operation type field, object identifier field, and constraint condition field in the intermediate instruction structure, the system constructs a task parsing graph within the expert model and triggers the corresponding inference execution path. The expert model performs structural mapping, target matching, and parameter solving on the intermediate semantic instructions according to its internally defined knowledge rules, inference engine, and strategy template, generating structured response content that matches the semantics of the current task. The response results include, but are not limited to, recommended options, operation schemes, control instructions, or knowledge explanation text, possessing clear semantic direction and task operability. Subsequently, the instruction response generation module performs format conversion and language reorganization on the structured response content, combining user context information and display preferences to output natural language, visual content, or multimodal response results, ensuring that the final result is understandable and practical for the user. This step achieves a complete mapping from structured semantic instructions to user-perceived content, effectively improving the interactive expression capability and result presentation quality of the intelligent system.

[0164] S7. The system scheduling and resource management module collects the computational load information and device resource status of the current task, and executes the pruning configuration and parameter distillation strategy related to the expert model.

[0165] In this embodiment, S7 specifically includes:

[0166] S71. The system scheduling and resource management module collects the expert model call request, task complexity identifier and device operating status of the current task, and constructs a task resource profile, which includes: task type, expert model calculation requirements, current system load and remaining computing power.

[0167] S72. Based on the task resource profile, perform layer-level pruning configuration on the expert model using a structural pruning strategy. The structural pruning strategy deletes low-importance nodes and redundant channels based on the node importance distribution in the expert model calculation graph, and generates a pruned expert model structure graph.

[0168] S73. Based on the task resource profile and combined with the semantic generality weight of the target scene, select the distillation source expert model and the distillation target expert model, and execute the parameter distillation process. The distillation includes two stages: label alignment distillation and intermediate representation matching distillation. The parameters of the distillation target expert model are updated using the feature matching loss function.

[0169] S74. The system schedules and sorts tasks based on the task priority of the expert model and the free window of the device, prioritizes loading the trimmed expert model structure, dynamically allocates video memory and computing threads, and performs matching and binding of software and hardware resources.

[0170] S75. When the system detects that the expert model fails to load or there is a runtime resource conflict, it automatically invokes the rollback mechanism to restore the default version of the expert model structure.

[0171] This invention utilizes a system scheduling and resource management module to dynamically perceive and evaluate the computational load information and device resource status of the current task in real time, constructing a task resource profile. Based on this profile, it executes expert model pruning and parameter distillation strategies, achieving lightweight deployment and resource adaptation optimization of the expert model. The system can automatically eliminate redundant parameters and adjust the model structure according to the importance of model execution nodes, significantly reducing the operating overhead of the expert model in computationally constrained environments. Simultaneously, the system supports adjusting the expert model's running priority, computational accuracy, and caching strategy based on dynamic resource changes during task execution, ensuring stable intelligent inference capabilities even in high-concurrency or resource-fluctuating scenarios. This module improves the overall resource scheduling flexibility and expert model operation stability of the system, enhancing its practicality and scalability in multi-task, multi-device environments.

[0172] Example:

[0173] To verify the practical application effect of the "Multi-Scenario Intelligent Consultation System Based on Pre-trained Large Model" proposed in this invention, the B-Government Cloud, a well-known domestic Class B intelligent government affairs platform, was selected as the verification environment. This platform covers multiple subsystems including government consultation, policy interpretation, and public affairs, handling approximately 215,000 user consultation requests daily. Its business scenarios are complex, data structures are heterogeneous, and it places extremely high demands on response timeliness, semantic accuracy, and the ability to coordinate multi-department model calls. For a long time, this platform has faced three major challenges: First, user consultation questions are diverse and their language is not standardized, making it difficult for traditional rule-based systems to accurately identify intent; second, redundant deployment of expert knowledge models and the lack of a dynamic calling mechanism lead to large response latency and high computational pressure; third, the system often experiences semantic drift and instruction misunderstanding in multi-turn dialogues, seriously affecting user satisfaction.

[0174] To address the aforementioned issues, the platform decided to deploy the intelligent consultation system described in this invention, integrating it into the platform's existing intelligent interfaces. Deployment began in December 2024 and was completed by the end of February 2025, with three consecutive months of field testing conducted in a district government data center environment in Beijing. The system as a whole adopts the modular structure proposed in this invention, including a multimodal input recognition module, a semantic state graph construction module, a self-consistent cognitive feedback module, an expert model scheduling module, and a system scheduling and resource management module.

[0175] During testing, the platform integrated 12 types of consultation scenarios, including social security inquiries, household registration transfers, business registration, and traffic violation processing, involving 47 expert knowledge models. The system first performs unified structured encoding on user-input voice, text, and image materials, and constructs a semantic state graph. This graph is further translated into an intermediate semantic instruction structure and fed into the expert model execution module, where the system intelligently selects the matching model and outputs the response result. To ensure response accuracy, the system automatically calls the self-consistent cognitive feedback module before response generation to evaluate the consistency between the slot semantics of the instruction structure and the knowledge graph state, and corrects and rewrites slots with detected deviations, achieving automatic semantic optimization and instruction self-correction. Simultaneously, the system dynamically adjusts the expert model structure and computing resource allocation based on the current CPU load and model response frequency, significantly improving response efficiency and stability.

[0176] Before system deployment, the average response time on platform B was 12.4 seconds, exceeding 25 seconds in some complex task scenarios, resulting in excessive user wait times and high churn rates. After system deployment, the average response time decreased to 4.1 seconds, a reduction of 66.9%. The average task execution latency decreased from 9.2 seconds to 2.6 seconds, significantly improving task processing throughput. In terms of semantic understanding, the system's expert model matching accuracy increased from 78.6% to 96.4%, with a near-zero false call rate. Particularly in parsing tasks involving non-standard language expressions, based on the structured semantic instruction generation and self-consistent cognitive feedback mechanism of this invention, the system can effectively identify and automatically correct semantic conflicts, self-referential contradictions, and historical dependencies in user input, increasing the success rate of multi-round semantic consistency verification from 81.5% to 97.2%. Furthermore, by introducing a system scheduling and resource management module, the inference parameter structure of the expert model was tailored as needed, avoiding redundant calculations, reducing system resource utilization from 89% before deployment to 72%, and maintaining stable response under high-concurrency scenarios. The relevant performance test data are shown in Table 1:

[0177] Table 1: Performance Improvement Data of Multi-Scenario Intelligent Consultation System

[0178]

[0179] For example, during the peak period of enterprise annual inspection business on January 15, 2025, the system processed 43,872 consultation requests continuously within 12 hours, including 8,134 cross-departmental model joint inference tasks, with a zero expert model scheduling failure rate. When the system detected the confusion between "annual review" and "annual inspection" in the user's semantics, it used a semantic slot deviation matrix evaluation mechanism to identify the incorrect substitution of "annual review" for "annual inspection" in this context, and automatically triggered slot correction logic to correctly route the request to the expert model in the "business annual inspection" domain. The system then used the response generation module to output accurate service process guidance, improving question-and-answer efficiency and user experience. The entire task took 1.9 seconds, with a user satisfaction score of 98.5%.

[0180] Another typical case occurred in early February 2025. A community office system in Beijing submitted a policy consultation request regarding "medical insurance enrollment for non-Beijing residents of advanced age" via a voice interface. The voice contained a mixed dialect expression: "I'm an old man, from out of town, over seventy years old in Beijing, asking if I can get medical insurance?" The traditional model failed to recognize the intention of "can get medical insurance," and "old man" and "in Beijing" were not correctly matched to the age and household registration slots. After deploying the system of this invention, the multimodal recognition module successfully completed the speech semantic transcription and fuzzy correction. The semantic state graph accurately extracted the three elements of "non-Beijing resident," "advanced age," and "medical insurance," and generated standardized intermediate semantic instructions. The expert model for "interpreting the urban residents' medical insurance enrollment policy" was then scheduled for processing. The system response time was 2.1 seconds, outputting the standard policy answer: "According to Beijing regulations, non-local residents aged 60 and above who have resided in Beijing for at least 6 months can apply for enrollment," effectively solving the semantic deviation problem.

[0181] After deployment, the system ran stably for three months, serving over 6.8 million users. The average daily call frequency of the expert model increased from 7,200 times before deployment to 19,300 times, and the system's automatic model switching success rate reached 99.6%. Particularly in complex task scenarios such as multi-semantic jumps and multi-instruction parallel processing, the closed-loop semantic verification and instruction self-updating mechanism of this invention demonstrated significant advantages. User feedback indicates that platform consultation satisfaction increased from 86.7% to 97.9%.

[0182] In summary, this invention not only solves the problems of weak semantic understanding, crude model scheduling mechanism and low response efficiency in traditional intelligent consultation systems, but also achieves a comprehensive improvement in the system's semantic control, expert model adaptability and task execution intelligence through structured semantic expression and cognitive feedback mechanism, and has broad engineering applicability and promotion value.

[0183] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-scene intelligent consultation system based on a pre-trained large model, characterized in that, The system comprises the following modules: A multi-modal input recognition module for receiving user inputted text data, voice data and image data, and converting the data into structured input data in a unified format; A semantic state cognitive graph construction module for constructing a semantic state vector based on the structured input data, and generating a semantic state cognitive graph containing user intent, context information and scene labels; An instruction translation module for parsing the semantic state vector in the semantic state cognitive graph into an intermediate instruction representation, and structurally mapping the intermediate instruction representation to generate an intermediate semantic instruction structure; A self-consistent cognitive feedback module for verifying the intent consistency of the intermediate semantic instruction structure, and dynamically updating the intermediate semantic instruction structure based on a cognitive bias correction mechanism; An expert model scheduling module for scheduling a matching expert model according to the scene label in the semantic state cognitive graph, and loading corresponding expert model parameters and response strategies; An instruction response generation module for inputting the updated intermediate semantic instruction structure into the scheduled expert model, generating structured response content, and outputting the content in a user understandable format; A system scheduling and resource management module for executing expert model pruning and parameter distillation strategies according to task complexity, scene type and device resource state; The semantic state cognitive graph construction module specifically comprises: Performing syntax analysis and semantic segmentation on the structured input data to identify semantic elements corresponding to user task intent, context-related entities and scene indicators; Encoding the semantic elements into user intent vectors, context vectors and scene label vectors according to categories, and performing unified dimension standardization processing to form sub-vectors of the three categories; Concatenating and combining the sub-vectors of the three categories in a preset semantic slot order to generate a semantic state vector, the preset semantic slots including user intent slots, context slots and scene label slots in sequence; Establishing a structured node set and connection relationship according to the association and semantic coupling strength between the slots in the semantic state vector, and constructing a semantic state cognitive graph; The semantic state cognitive graph comprises user intent nodes for describing task requirements, context information nodes for carrying input semantic context, and scene label nodes for identifying application environments; The nodes are connected by directed edges, and the establishment of the edges includes semantic edges based on semantic dependency relationships, time sequence edges based on context time sequences, and statistical edges based on semantic co-occurrence frequencies, with edge weight attribute values attached; Further specifically comprising: In the system initialization stage, the semantic slot order is preset to include user intent slots, context slots and scene label slots in sequence, and the slot order remains fixed during system operation and is prohibited from being dynamically modified; A unified vector dimension standard is set for the sub-vectors of the three categories, and all sub-vectors must be dimensionally aligned before concatenation. If any category of semantic elements is missing, a zero vector is filled in the corresponding slot to maintain the structural stability of the semantic state vector. The slots in the structure of the semantic state vector are independent semantic domains, the system is provided with a semantic consistency rule, and if there is a conflict between the user intent slot and the context slot, the vector splicing process will be blocked and a correction request will be returned; In the splicing operation, a linear connection method is used to connect the three types of sub-vectors in a predetermined semantic slot order to form a complete semantic state vector; The semantic state vector is a one-dimensional fixed-length multi-slot structure, where each slot is input as an independent node in the downstream graph construction process without vector compression and fusion processing; The semantic state vector is attached with a slot label to represent the type, filling state and confidence of each slot, which is used to guide the generation of edge connection relationship and the allocation of node weight in graph construction.

2. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 1, characterized in that, The modules are realized by the following methods: S1. The multi-modal input recognition module receives the text data, voice data and image data input by the user, performs format analysis and modality standardization processing on the input data, and generates structured input data in a unified format; S2. The structured input data is input into the semantic state cognitive graph construction module, the semantic elements are extracted based on the structured input data, the semantic state vector is generated, and the semantic state cognitive graph containing user intent, context information and scene label is constructed; S3. The instruction translation module receives the semantic state vector in the semantic state cognitive graph, and parses the semantic state vector into intermediate instruction representation, and performs structured mapping on the intermediate instruction representation to generate intermediate semantic instruction structure; S4. The intermediate semantic instruction structure is input into the self-consistent cognitive feedback module, which performs consistency checking and dynamic updating on the intermediate semantic instruction structure according to the preset cognitive bias correction mechanism, and outputs the updated semantic state vector; S5. The expert model scheduling module receives the scene label in the semantic state cognitive graph, selects the matching expert model based on the scene label, and loads the corresponding expert model parameters and response strategy; S6. The intermediate semantic instruction structure is input into the scheduled expert model, and the instruction response generation module generates structured response content and outputs the response result in a format understandable by the user; S7. The system scheduling and resource management module collects the computing load information and device resource state of the current task, and performs pruning configuration and parameter distillation strategy related to the expert model.

3. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 2, characterized in that, The S3 specifically includes: S31. The instruction translation module receives the semantic state vector in the semantic state cognitive graph, and performs matching analysis on the semantic state vector to extract the semantic backbone for task recognition; S32. The extracted semantic backbone content is mapped into intermediate instruction representation, which includes operation type field, object identification field and restriction condition field; S33. Structured parsing is performed on each field in the intermediate instruction representation to construct a slot-type intermediate semantic instruction structure; S34. In the structured mapping process, a multi-element mapping strategy is adopted, which preferentially matches the pre-defined instruction template based on the slot combination features of the semantic state vector, and calls the rule set configured internally by the system to complete the structure analysis when the matching is unsuccessful; S35, the intermediate semantic instruction structure supports a cross-scene instruction alignment mechanism, and consistency standardization of instruction expression in different scenes is performed through a semantic tag mapping relationship table.

4. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 3, characterized in that, The S33 specifically comprises: S331, the slot type intermediate semantic instruction structure comprises an operation type slot, an object identification slot and a restriction condition slot, which correspond to an operation type field, an object identification field and a restriction condition field in the intermediate instruction representation respectively; S332, the slots are arranged in a fixed order and are packaged by using a unified data structure, each slot structure comprises three fields of a slot tag, a slot value and a slot position confidence; S333, the slot value is filled by an instruction translation module according to a semantic element matching result in a semantic state vector, and the slot position confidence is directly generated from a matching score and is used to represent semantic reliability of the slot; S334, when a slot value is missing, the system fills the slot value by using a corresponding default parameter according to a slot type, and when the slot value exists but its confidence is lower than a set threshold, a preset placeholder is filled into the slot to identify that the slot is in an untrusted state; S335, the slot structure supports a field extension mechanism, which allows adding an extended slot under the premise of maintaining structure consistency; S336, the slot type intermediate semantic instruction structure must pass an inter-slot semantic consistency check before being output, and if semantic conflict relations exist between any two slots, the system stops the intermediate semantic instruction structure from being passed down and sends a structure reconstruction request to an upstream module.

5. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 4, characterized in that, The S4 specifically comprises: S41, inputting the intermediate semantic instruction structure into a self-consistent cognitive feedback module to perform semantic consistency analysis on the operation type slot, the object identification slot and the restriction condition slot respectively; S42, constructing a deviation score matrix according to a semantic similarity calculation result between the slot content and a corresponding semantic element in the semantic state vector, to represent a semantic deviation degree between the slot content and a current cognitive state of the system; S43, when the deviation score of any slot is higher than a set threshold, the system triggers a cognitive deviation correction mechanism to reconstruct the slot content according to a preset semantic correction rule set; S44, the slot content reconstruction comprises slot value replacement, semantic weakening processing and semantic conflict marking; S45, the reconstructed intermediate semantic instruction structure will be rechecked, and the checking comprises: slot structure integrity checking, used to judge whether any mandatory slot is missing; slot type matching checking, used to confirm whether the slot content conforms to a semantic category defined by the slot; inter-slot logic consistency checking, used to detect whether there are conflict expressions, logic inversion or mutual exclusion relations between the slots; when any checking fails, the system rejects the structure from being passed to a downstream module and marks the structure as an abnormal state; S46, the intermediate semantic instruction structure that passes the checking is used to generate an updated semantic state vector, and the updating operation rewrites a corresponding slot representation in the semantic state vector based on a semantic element in the current instruction structure.

6. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 5, characterized in that, The S42 specifically comprises: S421, each score value in the deviation score matrix is calculated by using a multi-factor weighted score function, and the formula is: ； wherein, denotes the semantic deviation score between the th slot and the corresponding semantic element in the semantic state vector, is the slot embedding vector, is the semantic representation vector in the semantic state vector corresponding to the slot, denotes the vector dot product, and are the L2 norm of the vectors, respectively, denotes the confidence score of the current slot, denotes the number of times the semantic conflict between the slot and other slots has occurred in history, , , are the system preset weighting factors for controlling the contribution weights of the semantic similarity score, the confidence enhancement term and the historical conflict penalty term, respectively. S422, all slot bias score values are assigned hierarchical weights according to slot types, which are used to reflect the differences in the system's consistency judgment standards for different semantic slots, wherein the operation type slot score weight is the highest, and the restriction condition slot score weight is the lowest; S423, all score results are normalized to form a final deviation score matrix, and the normalization adopts the maximum and minimum scaling method to make the score value distributed in the [0, 1] interval; S424, the system sets a deviation score threshold for each type of slot, wherein the operation type slot has the smallest tolerance, and the restriction condition slot has the largest tolerance, and if the score value of a slot exceeds the corresponding threshold, the slot is determined to have semantic drift; S425, when the score of any slot in the deviation score matrix exceeds the set threshold, the system marks the slot as a high deviation state.

7. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 6, characterized in that, The S45 specifically includes: S451, the structure verification is performed based on a system preset set of mandatory slots, the set of mandatory slots is dynamically loaded according to task types and scene labels, and the verification content includes whether the slot exists and whether the slot value is non-empty; S452, the slot type matching verification is performed according to a built-in slot semantic definition table of the system, and the verification standard includes whether the data type and semantic category of the slot value are consistent with the defined type, and if there is a mismatch, it is marked as a type exception; S453, the slot logic consistency verification includes the judgment of the following three types of logic conflicts: semantic conflict type, logic inversion type, and mutual exclusion relationship type; S454, the system sequentially performs each verification in the order of integrity, type matching, and logic consistency, and if any verification fails, the subsequent process is immediately terminated, the system refuses to pass the intermediate semantic instruction structure to the downstream module, and marks the structure as an abnormal structure state.

8. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 7, characterized in that, The S7 specifically includes: S71, the system scheduling and resource management module collects the expert model calling request, task complexity identification and device running state of the current task to build a task resource portrait, which includes: task type, expert model calculation demand, system current load and remaining computing power; S72, according to the task resource portrait, the structure pruning strategy is used to perform hierarchical pruning configuration on the expert model, the structure pruning strategy deletes low importance nodes and redundant channels according to the node importance distribution in the expert model calculation graph, and generates a pruned expert model structure graph; S73, according to the task resource portrait and combined with the semantic generality weight of the target scene, a distillation source expert model and a distillation target expert model are selected to perform a parameter distillation process, the distillation includes two stages of label alignment distillation and intermediate representation matching distillation, and a feature matching loss function is used to update the parameters of the distillation target expert model; S74, the system schedules and sorts according to the task priority of the expert model and the device idle window, preferentially loads the pruned expert model structure, dynamically allocates video memory and computing threads, and performs matching binding of software and hardware resources; S75, when the system detects that the expert model loading fails or the runtime resource conflicts, a fallback mechanism is automatically called to restore to the default version of the expert model structure.

Images