Multi-scene intelligent consultation system based on pre-training large model

Abstract

The invention discloses a multi-scene intelligent consultation system based on a pre-trained large model, and aims to solve the defects of an existing consultation system in the aspects of 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. The method is widely applicable to multi-scene intelligent interaction application of intelligent customer service, medical inquiry, financial consultation and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to a multi-scene intelligent consultation system based on a pre-trained large model. BACKGROUND

[0002] Under the background of the fusion development of artificial intelligence and natural language processing, how to build an intelligent consultation system with multi-modal understanding ability and semantic consistency expression has become a key direction in the research of intelligent human-computer interaction under complex task driving. In recent years, with the pre-trained large models (such as GPT, BERT, T5, etc.) showing excellent ability in semantic understanding, language generation and task generalization, the academic and industrial circles have explored their application value in consultation assistance in multi-scene such as government service, medical diagnosis, financial consultation, etc. The existing intelligent consultation systems mostly rely on rule base, limited domain corpus training or lightweight dialogue model, lack the modeling ability of complex semantic structure, and are difficult to realize unified understanding and context maintenance under multi-modal input, resulting in deficiencies in system response accuracy, generalization ability and interaction continuity.

[0003] Some existing technologies attempt to introduce pre-trained models to improve the performance of question and answer, or to identify intent and respond to tasks through scene templates and rule engines, but these methods are mostly limited to single-modal input analysis and static semantic graph management, and are difficult to support unified representation of voice, text and image and dynamic scheduling of multi-scene expert knowledge. In addition, the existing systems generally lack cognitive consistency verification mechanism, and when dealing with user input ambiguity, context ambiguity or instruction conflict, it is often difficult to guarantee the consistency of semantic transmission and the reliability of response logic. Some research introduces dialogue state management strategy to realize intent tracking, but its graph structure is extensive and lacks intermediate structure support for instruction translation, resulting in significant performance decline in complex interactive tasks.

[0004] In terms of model scheduling and system resource management, traditional methods mostly use static service architecture, which cannot load and optimize parameters of expert models according to task complexity, scene label and device state, restricting the deployment efficiency and edge adaptation ability of the model. At the same time, the existing intelligent systems lack a structured closed-loop mechanism in cognitive feedback, and cannot dynamically correct and restructure the structure after identifying structural abnormalities and semantic deviations, causing response deviation and increasing the failure rate of interaction.

[0005] Therefore, how to provide a multi-scene intelligent consultation system based on a pre-trained large model is a key technical problem that needs to be solved at present. SUMMARY

[0006] One purpose of the present application is to propose a multi-scene intelligent consulting system based on a pre-trained large model, which establishes a closed-loop structure from user intent recognition to expert knowledge calling to response generation by constructing modules such as multi-modal input recognition, semantic state cognitive graph generation, instruction translation and consistency verification, expert model scheduling response, and system resource management, adopts semantic state vector to drive intermediate semantic instruction structure, and introduces a self-consistent cognitive feedback mechanism to ensure semantic consistency, and has the advantages of strong multi-modal understanding, accurate semantic analysis, wide scene adaptation, and flexible model scheduling.

[0007] A multi-scene intelligent consulting system based on a pre-trained large model according to an embodiment of the present application comprises the following modules: A multi-modal input recognition module is configured to receive text data, voice data and image data input by a user, and convert the data into structured input data in a unified format; A semantic state cognitive graph construction module is configured to construct a semantic state vector based on the structured input data, and generate a semantic state cognitive graph containing user intent, context information and scene labels; An instruction translation module is configured to parse the semantic state vector in the semantic state cognitive graph into an intermediate instruction representation, and perform structured mapping on the intermediate instruction representation to generate an intermediate semantic instruction structure; A self-consistent cognitive feedback module is configured to perform intent verification on the intermediate semantic instruction structure, and perform dynamic updating on the intermediate semantic instruction structure based on a cognitive bias correction mechanism; Figure One An expert model scheduling module is configured to schedule a matching expert model according to the scene label in the semantic state cognitive graph, and load corresponding expert model parameters and response strategies; An instruction response generation module is configured to input the updated intermediate semantic instruction structure into the scheduled expert model, generate structured response content, and output the content in a format understandable by the user; A system scheduling and resource management module is configured to perform expert model pruning and parameter distillation strategies according to task complexity, scene type and device resource state.

[0008] Optionally, the modules are connected through the following methods: S1, the multi-modal input recognition module receives text data, voice data and image data input by a user, performs format analysis and modality standardization processing on the input data, and generates structured input data in a unified format; S2, input the structured input data into the semantic state cognitive graph construction module, extract semantic elements based on the structured input data, generate a semantic state vector, and construct a semantic state cognitive graph containing user intent, context information and scene labels; ​S3, receiving a semantic state vector in the semantic state cognitive graph by the instruction translation module, and parsing the semantic state vector into an intermediate instruction representation, and performing a structured mapping on the intermediate instruction representation to generate an intermediate semantic instruction structure; S4, inputting the intermediate semantic instruction structure into a self-consistent cognitive feedback module, and performing consistency checking and dynamic updating on the intermediate semantic instruction structure according to a preset cognitive bias correction mechanism, and outputting an updated semantic state vector; S5, receiving a scene label in the semantic state cognitive graph by the expert model scheduling module, selecting a matched expert model based on the scene label, and loading corresponding expert model parameters and response strategies; S6, inputting the intermediate semantic instruction structure into the scheduled expert model, generating a structured response content by the instruction response generation module, and outputting a response result in a user understandable format; S7, collecting the computing load information and the device resource state of the current task by the system scheduling and resource management module, and performing pruning configuration and parameter distillation strategies related to the expert model.

[0009] Optionally, the S2 specifically includes: S21, performing syntax analysis and semantic segmentation on the structured input data, and identifying semantic elements corresponding to user task intent, context related entities and scene indicators; S22, encoding the semantic elements into user intent vectors, context vectors and scene label vectors according to categories respectively, and performing unified dimension standardization processing to form sub-vectors of three categories; S23, splicing and combining the three categories of sub-vectors according to a preset semantic slot order to generate a semantic state vector, the preset semantic slot including a user intent slot, a context slot and a scene label slot in turn.

[0010] S24, establishing a structured node set and a connection relationship according to the association and semantic coupling strength between each slot in the semantic state vector, and constructing a semantic state cognitive graph; S25, the semantic state cognitive graph includes a user intent node for describing task requirements, a context information node for carrying input semantic context, and a scene label node for identifying an application environment; S26, the nodes are connected through directed edges, and the establishment of the edges includes semantic edges established based on semantic dependency relationship, time sequence edges established based on context time sequence, and statistical edges established based on semantic co-occurrence frequency, the edges being attached with edge weight attribute values.

[0011] Optionally, the S23 specifically includes: S231, preset semantic slot order in the system initialization stage, including user intent slot, context slot and scene label slot in turn, the slot order remains fixed during system operation, and dynamic modification is prohibited; S232, set uniform vector dimension standard for three categories of sub-vectors, all sub-vectors must be dimensionally aligned before splicing, if any category semantic element is missing, the corresponding slot is filled with zero vector to keep the structure of semantic state vector stable; S233, each slot in the structure of the semantic state vector 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; S234, in the splicing operation, a linear connection method is used to connect the three categories of sub-vectors in the preset semantic slot order to form a complete semantic state vector; S235, the result of the semantic state vector is a one-dimensional fixed-length multi-slot structure, wherein each slot is input as an independent node in the downstream graph construction process without vector compression and fusion processing; S236, the semantic state vector is attached with 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.

[0012] Optionally, the S3 specifically comprises: S31, receiving the semantic state vector in the semantic state cognitive graph by the instruction translation module, and performing matching analysis on the semantic state vector to extract the semantic backbone for task recognition; S32, mapping the extracted semantic backbone content to intermediate instruction representation, the intermediate instruction representation including operation type field, object identification field and restriction condition field; S33, structurally analyzing each field in the intermediate instruction representation to construct a slot-type intermediate semantic instruction structure; S34, in the structural mapping process, a multi-element mapping strategy is adopted, based on the slot combination characteristics of the semantic state vector, the pre-defined instruction template is preferentially matched, and when the matching is unsuccessful, the rule set configured in the system is called to complete the structure analysis; S35, the intermediate semantic instruction structure supports a cross-scene instruction alignment mechanism, and the consistency standardization of instruction expression in different scenes is realized through the semantic label mapping relationship table.

[0013] Optionally, the S33 specifically comprises: S331, the slot-type intermediate semantic instruction structure includes operation type slot, object identification slot and restriction condition slot, respectively corresponding to operation type field, object identification field and restriction condition field in the intermediate instruction representation; S332, each slot is arranged in a fixed order and encapsulated by a unified data structure, and each slot structure includes a slot label, a slot value, and a slot position reliability; S333, the slot value is filled by the instruction translation module according to the matching result of the semantic elements in the semantic state vector, and the slot position reliability is directly generated from the matching score, indicating the semantic reliability of the slot; S334, when a slot value is missing, the system calls the corresponding default parameter according to the slot type to fill it, and when the slot value exists but its confidence is lower than the 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 expansion mechanism, allowing the addition of extended slots while maintaining consistent structure; S336, the slot type intermediate semantic instruction structure must pass through inter-slot semantic consistency verification before output, and if semantic conflict relations are detected between any two slots, the system stops the intermediate semantic instruction structure from passing down the flow and sends a structure reconstruction request to the upstream module.

[0014] Optionally, the S4 specifically includes: 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 identification slot, and restriction condition slot respectively; S42, construct a deviation score matrix according to the semantic similarity calculation results between the slot content and the corresponding semantic elements in the semantic state vector, to represent the semantic deviation degree between the slot content and the current cognitive state of the system; S43, when the deviation score of any slot is higher than the set threshold, the system triggers the cognitive deviation correction mechanism and reconstructs the slot content according to the preset semantic correction rule set; S44, slot content reconstruction includes slot value replacement, semantic weakening processing, and semantic conflict marking; S45, the reconstructed intermediate semantic instruction structure will be re-verified, and the verification includes: Slot structure integrity verification is used to determine whether each required slot is missing; Slot type matching verification is used to confirm whether the slot content conforms to the semantic category defined by the slot; Inter-slot logic consistency verification is used to detect whether there are conflict expressions, logic inversion, or mutual exclusion between slots; When any verification fails, the system refuses to pass the structure to the downstream module and marks it as an abnormal structure state; S46, the verified intermediate semantic instruction structure 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.

[0015] Optionally, the S42 specifically includes: S421, each score value in the bias score matrix is calculated using a multi-factor weighted score function, and the formula is: ; Wherein, represents the semantic bias score between the i-th slot and the corresponding semantic element in the semantic state vector, is a slot embedding vector, is a semantic representation vector corresponding to the slot in the semantic state vector, represents the dot product of vectors, and are the L2 norm of the vector respectively, represents the confidence score of the current slot, represents the number of times of semantic conflict between the slot and other slots in the history, , , , is a system preset weighting factor for controlling the contribution weight of semantic similarity score, confidence enhancement item and historical conflict penalty item respectively; S422, all slot bias score values are given hierarchical weights according to slot types, which are used to reflect the differences in consistency judgment standards of different semantic slots by the system, 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 bias score matrix, and the normalization adopts the maximum and minimum scaling method, so that the score value is distributed in the [0, 1] interval; S424, the system sets a bias score threshold for each type of slot, wherein the operation type slot tolerance is the smallest, and the restriction condition slot tolerance is the largest. If the score value of a slot exceeds the corresponding threshold, the slot is determined to have semantic deviation; S425, when the score of any slot in the bias score matrix exceeds the set threshold, the system marks the slot as a high bias state.

[0016] Optionally, 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 the task type and the scene label, 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 slot semantic definition table built in 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, the type is marked as an exception; S453, the inter-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, if any verification fails, the subsequent process is terminated immediately, the system refuses to pass the intermediate semantic instruction structure to the downstream module, and the structure is marked as an abnormal state.

[0017] Optionally, the S7 specifically comprises: 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, constructs a task resource portrait, and the task resource portrait 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 layer-level 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 the semantic generality weight of the target scene, a distillation source expert model and a distillation target expert model are selected, a parameter distillation process is performed, 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.

[0018] The beneficial effects of the application are: (1) The application realizes unified analysis and structured expression of heterogeneous user inputs such as text, voice and image by constructing a multi-modal input recognition module and a semantic state cognitive graph construction module, adopts a fixed slot semantic state vector design, combines syntax analysis, semantic segmentation and entity recognition operations, encodes and splices user intent, context information and scene labels into a one-dimensional multi-slot structure, further constructs a cognitive graph structure with associated edge weights, greatly improves the semantic extraction accuracy and context modeling capability of the system for multi-modal input, and solves the problems of weak context awareness and limited input understanding of existing systems.

[0019] (2) The application designs an instruction translation module and a self-consistent cognitive feedback mechanism based on an intermediate semantic instruction structure, generates a standardized slot type intermediate instruction structure based on the semantic state vector, and performs consistency checking on the operation type, object identification and restriction condition slot after translation, judges the instruction deviation degree through the construction of a semantic deviation score matrix and a slot position reliability mechanism, automatically triggers the cognitive deviation correction process when the deviation score exceeds the set threshold, reconstructs the slot value and checks its structural integrity, type matching and logical consistency, 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.

[0020] (3) The application designs an expert model scheduling and system resource management mechanism based on the scene label in the semantic state graph, matches the corresponding expert model for different task scenes, and performs model structure pruning and parameter distillation combined with the current device load and task complexity, optimizes the model inference structure through node importance pruning and feature alignment distillation strategies, significantly reduces the calculation resource consumption and loading time of the system while ensuring the accuracy of the model response, realizes on-demand calling and lightweight running of the expert model, adapts to the edge device deployment environment, has good scalability and engineering practical value. BRIEF DESCRIPTION OF DRAWINGS

[0021] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0022] Figure 1 A schematic diagram of the overall architecture of a multi-scene intelligent consulting system based on a pre-trained large model is proposed for the application; Figure 2 A semantic state vector generation and graph structure building flowchart of a multi-scene intelligent consulting system based on a pre-trained large model is proposed for the application; Figure 3 An intermediate semantic instruction generation and consistency checking process flowchart of a multi-scene intelligent consulting system based on a pre-trained large model is proposed for the application. Detailed Implementation

[0023] 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.

[0024] refer to Figures 1-3 A multi-scenario intelligent consultation system based on a pre-trained large model includes the following modules: 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; 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. 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. The self-consistent cognitive feedback module is used to perform semantic feedback on the intermediate semantic instruction structure. Figure One Consistency verification, and dynamic updating of the intermediate semantic instruction structure based on a cognitive bias correction mechanism; 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. 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. 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.

[0025] 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.

[0026] A multi-scenario intelligent consultation system based on a pre-trained large model, wherein the modules are implemented through the following methods: S1, a multi-modal input recognition module receives text data, voice data and image data input by a user respectively, performs format analysis and modality standardization processing on the input data, and generates structured input data in a unified format; In the embodiment, step S1 specifically includes: a text recognition unit, a voice recognition unit and an image recognition unit in the multi-modal input recognition module receive text data, voice data and image data input by a user respectively; the text recognition unit directly acquires structured text information; the voice recognition unit performs acoustic feature extraction and semantic recognition processing on voice signals by using a voice transcription model, and outputs corresponding text information; the image recognition unit extracts target semantic labels and scene information in an image based on an image content recognition model, and converts them into a parseable semantic representation; then, the system uniformly performs format deconstruction and state labeling on the input data in different modalities, including identifying input sources, extracting keyword slots, constructing semantic vector representations, and generating structured multi-modal fusion input data in a unified time sequence framework, thereby 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, and improves the accuracy and stability of the overall intelligent response process.

[0027] S2, input the structured input data into a semantic state cognitive graph construction module, extract semantic elements based on the structured input data, generate a semantic state vector, and construct a semantic state cognitive graph containing user intent, context information and scene labels; In the embodiment, S2 specifically includes: S21, perform syntax analysis and semantic segmentation on the structured input data, and identify semantic elements corresponding to user task intent, context-related entities and scene indicators; S22, encode the semantic elements into user intent vectors, context vectors and scene label vectors according to categories respectively, and perform unified dimension standardization processing to form sub-vectors of the three categories; S23, splice and combine the sub-vectors of the three categories in a preset semantic slot order to generate a semantic state vector, wherein the preset semantic slots include user intent slots, context slots and scene label slots in sequence.

[0028] S24, establish a structured node set and a connection relationship according to the association and semantic coupling strength between each slot in the semantic state vector, and construct a semantic state cognitive graph; S25, the semantic state cognitive graph includes a user intent node for describing task requirements, a context information node for carrying input semantic context, and a scene label node for identifying an application environment; S26, the nodes are connected by directed edges, and the establishment of the edges is based on semantic edges established based on semantic dependency relationships, timing edges established based on context time sequences, and statistical edges established based on semantic co-occurrence frequencies, and the edges are attached with edge weight attribute values.

[0029] In the embodiment, the S23 specifically comprises: S231, presetting a semantic slot order in the system initialization stage, sequentially including a user intent slot, a context slot and a scene label slot, the slot order remains fixed during system operation and is prohibited from being dynamically modified; S232, setting a unified vector dimension standard for the three categories of sub-vectors, and all sub-vectors must be subjected to dimension alignment processing before splicing, and if any category of semantic element is missing, the corresponding slot is kept stable in structure of the semantic state vector by filling a zero vector; S233, each slot in the structure of the semantic state vector is an independent semantic domain, the system is provided with a semantic consistency rule, and if there is a conflict expression between the user intent slot and the context slot, the vector splicing process will be blocked and a correction request will be returned; S234, in the splicing operation, a linear connection mode is adopted to sequentially connect the three categories of sub-vectors in the preset semantic slot order to form a complete semantic state vector; S235, the semantic state vector constitutes a one-dimensional fixed-length multi-slot structure, wherein each slot is input as an independent node in the downstream graph construction process without vector compression and fusion processing; S236, the semantic state vector is attached with a slot label to represent the type, filling state and confidence of each slot, for guiding the generation of edge connection relationship and the allocation of node weight in graph construction.

[0030] The semantic state cognitive graph construction module divides the semantic elements in the structured input data into three categories of user intent vector, context vector and scene label vector, extracts semantic slot information respectively, splices the semantic state vector after unified dimension standardization processing according to the semantic slot order, and constructs a structured semantic state graph. This mode improves the logical coherence and expression consistency of the semantic vector combination, effectively avoids semantic conflicts and structural chaos, ensures the connection stability and accuracy between the slot nodes in the graph, and provides a high-precision and high-consistency semantic expression basis for subsequent semantic reasoning and expert instruction generation.

[0031] S3, receiving, by the instruction translation module, the semantic state vector in the semantic state cognitive graph, and parsing the semantic state vector into an intermediate instruction representation, and performing structured mapping on the intermediate instruction representation to generate an intermediate semantic instruction structure; In the embodiment, the S3 specifically comprises: S31, receiving a semantic state vector in the semantic state cognition graph by the instruction translation module, and performing matching analysis on the semantic state vector to extract a semantic stem for task recognition; S32, mapping the extracted semantic stem content into an intermediate instruction representation, the intermediate instruction representation including an operation type field, an object identification field, and a restriction condition field; S33, structurally analyzing each field in the intermediate instruction representation to construct a slot-type intermediate semantic instruction structure; S34, in the structural mapping process, a multi-element mapping strategy is adopted, based on the slot combination characteristics of the semantic state vector, a pre-defined instruction template is preferentially matched, and when the matching is unsuccessful, a rule set configured in the system is called to complete the structural analysis; S35, the intermediate semantic instruction structure supports a cross-scene instruction alignment mechanism, and different scene instruction expressions are standardized by a semantic label mapping relationship table.

[0032] In the embodiment, the S33 specifically includes: S331, the slot-type intermediate semantic instruction structure includes an operation type slot, an object identification slot, and a restriction condition slot, corresponding to the operation type field, the object identification field, and the restriction condition field in the intermediate instruction representation, respectively; S332, the slots are arranged in a fixed order and encapsulated by a unified data structure, and each slot structure includes a slot label, a slot value, and a slot position confidence three fields; S333, the slot value is filled by the instruction translation module according to the semantic element matching result in the semantic state vector, and the slot position confidence is directly generated from the matching score, used to represent the semantic reliability of the slot; S334, when a slot value is missing, the system fills it with the corresponding default parameter according to the 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 expansion mechanism, allowing the addition of an extended slot while maintaining structural consistency; S336, the slot-type intermediate semantic instruction structure must pass an inter-slot semantic consistency check before output, and if a semantic conflict relationship is detected between any two slots, the system stops the intermediate semantic instruction structure from passing down the process and sends a structure reconstruction request to the upstream module.

[0033] The application realizes the structured mapping of the semantic state vector to the intermediate instruction expression by the instruction translation module, supports the parsing of the user task semantic main content into the intermediate instruction representation containing the operation type, object identification and limitation condition, and the slot structure analysis of each field, and generates the uniform standard intermediate semantic instruction structure. The system adopts the multi-element mapping mechanism and the template matching strategy, supports the fault tolerance filling and confidence expansion of different semantic slot types, and enhances the task analysis robustness and semantic coverage ability of the system in multiple scenes. The semantic consistency verification mechanism is introduced in the instruction structure, ensures the expression stability of the instruction structure under the context evolution, and provides the semantic closed loop support for the subsequent scene instruction generation and expert model calling.

[0034] S4, input the intermediate semantic instruction structure to a self-consistent cognitive feedback module, the self-consistent cognitive feedback module performs consistency verification and dynamic update on the intermediate semantic instruction structure according to a preset cognitive bias correction mechanism, and outputs an updated semantic state vector; In the embodiment, the S4 specifically comprises: S41, input the intermediate semantic instruction structure to the self-consistent cognitive feedback module, and perform semantic consistency analysis on the operation type slot, object identification slot and limitation condition slot respectively; S42, construct a bias score matrix according to the semantic similarity calculation results between the slot content and the corresponding semantic elements in the semantic state vector, for representing the semantic deviation degree between the slot content and the current cognitive state of the system; S43, when the bias score of any slot is higher than a set threshold, the system triggers the cognitive bias correction mechanism, and reconstructs the slot content according to a preset semantic correction rule set; S44, the slot content reconstruction includes slot value replacement, semantic weakening processing and semantic conflict marking; S45, the reconstructed intermediate semantic instruction structure will be re-verified, and the verification includes: slot structure integrity verification, for judging whether each mandatory slot is missing or not; slot type matching verification, for confirming whether the slot content conforms to the semantic category defined by the slot or not; inter-slot logic consistency verification, for detecting whether there is a conflict expression, logic inversion or mutual exclusion relationship between the slots or not; When any verification fails, the system refuses to pass the structure to the downstream module, and marks it as an abnormal structure state; S46, the intermediate semantic instruction structure that passes the verification is used to generate an updated semantic state vector, and the updating operation rewrites the corresponding slot representation in the semantic state vector based on the semantic elements in the current instruction structure.

[0035] In the embodiment, the S42 specifically comprises: S421、the score value in the deviation score matrix is calculated by using a multi-factor weighted score function, and the formula is: ; wherein, represents the semantic deviation score between the i-th slot and the corresponding semantic element in the semantic state vector, is a slot embedding vector, is a semantic representation vector corresponding to the slot in the semantic state vector, represents a vector dot product, and are L2 norms of vectors respectively, represents the confidence score of the current slot, represents the historical number of times of semantic conflicts between the slot and other slots, , , , is a system preset weighting factor for controlling the contribution weight of the semantic similarity score, the confidence enhancement term and the historical conflict penalty term respectively; In the present application, the deviation score function is introduced, which is a necessary basic computer mechanism for realizing semantic slot level semantic consistency evaluation. The function comprehensively considers the cosine similarity between the slot semantic representation vector and the corresponding semantic element in the semantic state vector, the confidence weight of the current slot and the historical conflict distance between the slot and other slots, and generates a numerical deviation score for each slot through multi-factor weighted summation. The score value can quantify the reliability and consistency of the slot semantics, and provide a clear mathematical basis for the deviation correction mechanism.

[0036] Compared with the traditional rule matching or static template judgment method, the scoring function has the characteristics of computability, controllability and adaptability, and can dynamically evaluate the deviation risk according to different contexts and semantic evolution, so as to realize more accurate slot level consistency judgment. Through the introduction of the function, the system can automatically detect semantic structure abnormalities and trigger the feedback mechanism without interrupting the task process, effectively supporting the closed-loop control process of "deviation detection-structure checking-semantic rewriting" in the self-consistent cognitive feedback module described in the present application.

[0037] S422、all slot deviation score values are given hierarchical weights according to slot types, which are used to reflect the differences in consistency judgment standards of different semantic slots by the system, 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, so that the score value is distributed in the [0, 1] interval; S424、the system sets a deviation score threshold for each type of slot, wherein the operation type slot tolerance is the smallest, and the limit condition slot tolerance is the largest, and if the score of a slot exceeds the corresponding threshold, the slot is determined to have semantic deviation; 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.

[0038] In this embodiment, the S45 specifically includes: S451、the structure verification is performed based on a set of mandatory slots preset by the system, the set of mandatory slots is dynamically loaded according to the task type and the scene label, 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 the 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 logical conflicts: semantic conflict type, logical 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 a structure abnormal state.

[0039] The application establishes a cognitive bias correction mechanism with semantic slots as the basic unit by self-consistent cognitive feedback module for consistency verification and dynamic update of intermediate semantic instruction structure, and realizes precise iterative update of instruction-level semantic state. The system first constructs a multi-factor weighted deviation score matrix based on semantic similarity, semantic confidence and historical dependency relationship, evaluates the semantic consistency of the slot field in the semantic state vector and the current knowledge state, and when the score of any slot exceeds the preset threshold, triggers the cognitive bias correction mechanism to perform semantic fuzzification, fault tolerance reconstruction and slot content replacement on the corresponding field, thereby enhancing the semantic robustness and reasoning robustness of the system in inconsistent context. Further, the system introduces three types of strategies, structure constraint verification, slot type matching verification and logic consistency verification, in the structure verification, and performs multiple screening on the intermediate instruction structure from the aspects of grammatical integrity, semantic normativity and logical coherence, thereby ensuring that the updated semantic state vector is structurally legal, semantically reliable and logically closed. The module realizes the transition from one-way instruction generation to two-way semantic verification, provides support for the system to build an adaptive and closed-loop intelligent semantic control mechanism, significantly improves the intelligent response accuracy, instruction generation accuracy and cognitive consistency in complex context, and is a key innovation point of the application that distinguishes it from the prior art.

[0040] S5, receiving, by the expert model scheduling module, a scene label in the semantic state cognitive graph, selecting a matched expert model based on the scene label, and loading corresponding expert model parameters and response strategies; In the embodiment, step S5 specifically includes: receiving, by the expert model scheduling module, a scene label extracted in the semantic state cognitive graph, the scene label being used to reflect a running environment, an interaction target and a service context in which a current semantic task is located. The system performs associated screening on a plurality of candidate expert models by constructing a mapping relationship between the scene label and the expert model, and optimizes an expert model with the highest matching degree according to a preset matching degree evaluation mechanism and by comprehensively considering factors such as service field adaptability, resource occupation, response timeliness and historical performance of the model. Then, the system loads structure configuration, reasoning parameters and response strategies of the expert model, and completes interface adaptation with a current semantic state vector, so as to ensure that the expert model can stably operate in a task execution process. Meanwhile, the system supports dynamically adjusting a reasoning path of a loaded model when a semantic state changes, or triggering an expert model replacement mechanism according to a feedback signal, so as to realize adaptive updating of a task processing strategy in a complex semantic scene. This step improves the precision and flexibility of expert model calling, and guarantees intelligent response efficiency and stability of the system in a multi-task and multi-scene environment.

[0041] S6, inputting the intermediate semantic instruction structure into the scheduled expert model, generating, by an instruction response generation module, a structured response content, and outputting a response result in a user understandable format; In the embodiment, step S6 specifically includes: inputting the intermediate semantic instruction structure that has completed structure standardization processing into a target expert model selected and loaded by the expert model scheduling module, constructing a task analysis graph in the expert model according to an operation type field, an object identification field and a restriction condition field in the intermediate instruction structure, and triggering a corresponding reasoning execution path. The expert model performs structure mapping, target matching and parameter solving on the intermediate semantic instruction according to knowledge rules, reasoning engines and strategy templates defined in the expert model, to generate a structured response content matched with a current task semantic. The response result includes but is not limited to a recommended option, an operation scheme, a control instruction or a knowledge explanation text, and has clear semantic direction and task operability. Then, the instruction response generation module performs format conversion and language recombination on the structured response content, outputs a natural language, visual content or multi-modal response result in combination with user context information and display preferences, so as to ensure that the final result is understandable and practical for the user. This step realizes complete mapping from the structured semantic instruction to the user perception content, and effectively improves the interactive expression ability and result presentation quality of the intelligent system.

[0042] S7, the system scheduling and resource management module collects the current task computing load information and device resource state, and executes the pruning configuration and parameter distillation strategy related to the expert model.

[0043] In this embodiment, S7 specifically includes: S71, the system scheduling and resource management module collects the expert model invocation request, task complexity identification and device running state of the current task, constructs a task resource portrait, and the task resource portrait includes: task type, expert model computing demand, system current load and remaining computing power; S72, according to the task resource portrait, the structure pruning strategy is used to perform layer-level 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 computation graph, and generates a pruned expert model structure graph; S73, according to the task resource portrait and the semantic generality weight of the target scene, select the distillation source expert model and the distillation target expert model, execute the parameter distillation process, the distillation includes two stages of label alignment distillation and intermediate representation matching distillation, and the 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, the fallback mechanism is automatically called, and the expert model structure of the default version is restored.

[0044] The present application dynamically perceives and real-time evaluates the computing load information and device resource state of the current task through the system scheduling and resource management module, constructs a task resource portrait, and executes the pruning configuration and parameter distillation strategy of the expert model based on the portrait, realizes the lightweight deployment and resource adaptation optimization of the expert model. The system can automatically remove redundant parameters according to the importance of the model execution node, adjust the model structure, and significantly reduce the running overhead of the expert model in the computing resource limited environment. At the same time, the system supports adjusting the expert model running priority, computing accuracy and cache strategy according to the dynamic change of resources during task execution, and guarantees the stable intelligent inference ability in high concurrency or resource fluctuation scene. This module improves the overall resource scheduling flexibility of the system and the stability of the expert model running, enhances the practicality and scalability in multi-task and multi-device environment.

[0045] Embodiment: To verify the practical application effect of the "multi-scene intelligent consultation system based on pre-trained large model" proposed in the present application, a domestic well-known B-level intelligent government platform B-Government Cloud is selected as the verification environment. The platform covers multiple subsystems such as government consultation, policy interpretation, livelihood affairs, etc. It receives about 215,000 user consultation requests per day, with complex business scenarios, heterogeneous data structures, and high requirements for response timeliness, semantic accuracy, and multi-department model calling coordination ability. For a long time, the platform has faced three major problems: first, user consultation questions are diverse and language expression is not standardized, making it difficult for traditional rule systems to accurately identify intent; second, expert knowledge models are redundantly deployed, lacking a dynamic calling mechanism, resulting in large response delay and high computing pressure; third, the system often has problems such as semantic drift and instruction understanding deviation in multi-turn dialogue, seriously affecting user satisfaction.

[0046] To solve the above problems, the platform decides to deploy the intelligent consultation system described in the present application and integrate it into the original intelligent interface of the platform. The deployment and implementation started in December 2024 and was completed by the end of February 2025, and a continuous 3-month field test was conducted in the government data center environment of a certain district in Beijing. The system as a whole adopts the modular structure proposed in the present application, including a multi-modal input recognition module, a semantic state graph construction module, a self-consistent cognitive feedback module, an expert model scheduling module, a system scheduling and resource management module, etc.

[0047] During the test, the platform accessed 12 types of consultation scenarios, including social security inquiries, household registration migration, enterprise registration, traffic violation handling, etc., a total of 47 expert knowledge models were called. The system first uniformly structures the user input voice, text and picture materials, and constructs a semantic state graph. The graph is further translated into an intermediate semantic instruction structure and accessed to 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 assess the consistency of the slot semantics and knowledge graph state of the instruction structure, and correct and rewrite the slots detected with deviations, achieving automatic optimization of semantics and self-correction of instructions. The system dynamically adjusts the expert model structure and computing resource ratio according to the current CPU load and model response frequency, significantly improving the response efficiency and stability.

[0048] Before the deployment of the system, the average response time of the B platform was 12.4 seconds, and in some complex task scenarios, it even exceeded 25 seconds, causing long user waiting time and high attrition rate. After the deployment of the system, the average response time decreased to 4.1 seconds, with a decrease of 66.9%. The average task execution time delay decreased from 9.2 seconds to 2.6 seconds, significantly improving the task processing throughput. In terms of semantic understanding, the expert model matching accuracy of the system increased from 78.6% to 96.4%, with almost zero false call rate. In particular, in the analysis of non-standard language expressions, based on the structured semantic instruction generation and self-consistent cognitive feedback mechanism of the application, the system can effectively identify and automatically correct semantic conflicts, self-referential contradictions and historical dependencies in user input, making the multi-round semantic consistency verification success rate jump from 81.5% to 97.2%. In addition, by introducing the system scheduling and resource management module, the reasoning parameter structure of the expert model is trimmed on demand to avoid redundant calculation, reducing the system resource occupancy rate from 89% before deployment to 72%, and maintaining stable response in high concurrency scenarios. The relevant performance test data is shown in Table 1:

[0049] Table 1: Performance improvement data of multi-scenario intelligent consulting system

[0050] For example, on January 15, 2025, during the "business annual inspection business peak period", the system continuously processed 43,872 consulting requests in 12 hours, including 8,134 cross-department model joint reasoning tasks, with a failure rate of 0 for expert model scheduling. When the system identified the mixed use of "annual inspection" and "annual check" in user semantics, it recognized the misrepresentation of "annual inspection" as "annual check" in this context through the semantic slot deviation matrix evaluation mechanism, and automatically triggered the slot correction logic to correctly route to the "business annual check" domain expert model. The system then used the response generation module to output accurate business process guidance, improving the efficiency of question and answer and user experience. The entire task took 1.9 seconds, and the user satisfaction score was 98.5%.

[0051] Another typical case occurred in early February 2025. A community office system in a certain street in Beijing submitted a policy consultation request about "non-Beijing elderly medical insurance participation" through a voice interface. The voice contains mixed dialect expressions: "My old man, an outsider, has been in Beijing for more than 70 years, and asks if he can participate in medical insurance?" The traditional model failed to identify the "medical insurance" intent, and "old man" and "in Beijing" did not correctly match the age and household registration slots. After deploying the system of the present application, the multi-modal recognition module successfully completed the voice semantic transcription and fuzzy correction, the semantic state atlas accurately extracted the three elements of "non-Beijing", "elderly", and "medical insurance", and generated standardized intermediate semantic instructions, and dispatched the "urban resident medical insurance participation policy explanation" expert model for processing. The system response time is 2.1 seconds, and the output is "According to the regulations of Beijing, non-local residents who are 60 years old or older and have lived in Beijing for more than 6 months can apply for medical insurance." The semantic deviation problem is effectively solved.

[0052] After deployment, the system has been running stably for three months, with a cumulative service user of more than 6.8 million times. The average daily call frequency of the expert model increased from 7,200 times before deployment to 19,300 times. The system automatic model switching success rate reached 99.6%. Especially in complex task scenarios such as multi-semantic jumping and multi-instruction parallel processing, the closed-loop semantic verification and instruction self-updating mechanism of the present application show significant advantages. User feedback shows that the platform consultation satisfaction has increased from 86.7% to 97.9%.

[0053] In summary, the present application not only solves the problems of weak semantic understanding ability, rough model scheduling mechanism, and low response efficiency of traditional intelligent consultation systems, but also realizes the overall improvement of system semantic control, expert model adaptability, and task execution intelligence through structured semantic expression and cognitive feedback mechanism, with wide engineering applicability and promotion value.

[0054] The above describes only the preferred embodiments of the present application, but the protection scope of the present application is not limited thereto. Any skilled person in the art can make equivalent substitutions or changes to the technical solutions and inventive concepts of the present application within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application.

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 input 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 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 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 response 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.

2. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 1, characterized in that, The modules are connected through the following methods: S1. The multi-modal input recognition module receives user input text data, voice data and image data, and performs format analysis and modality standardization processing on the input data to generate structured input data in a unified format; S2. The structured input data is input into the semantic state cognitive graph construction module, semantic elements are extracted based on the structured input data, a semantic state vector is generated, and a semantic state cognitive graph containing user intent, context information and scene labels is constructed; S3. The instruction translation module receives the semantic state vector in the semantic state cognitive graph, parses the semantic state vector into intermediate instruction representation, and performs structured mapping on the intermediate instruction representation to generate an intermediate semantic instruction structure; S4. The intermediate semantic instruction structure is input into the self-consistent cognitive feedback module, which performs consistency verification and dynamic updating of the intermediate semantic instruction structure according to a pre-set cognitive bias correction mechanism, and outputs an updated semantic state vector; S5. The expert model scheduling module receives the scene label in the semantic state cognitive graph, selects a matching expert model based on the scene label, and loads corresponding expert model parameters and response strategies; 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 user understandable format; S7. The system scheduling and resource management module collects the current task computing load information and device resource state, and executes pruning configuration and parameter distillation strategies 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 S2 specifically comprises: S21. Perform 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; S22, encode the semantic elements according to categories into user intention vector, context vector and scene label vector respectively, and perform unified dimension standardization processing to form three categories of sub-vectors; S23, the three categories of sub-vectors are spliced and combined according to the preset semantic slot order to generate a semantic state vector, and the preset semantic slot sequentially includes a user intention slot, a context slot and a scene label slot. 4.S24, according to the association and semantic coupling strength between each slot in the semantic state vector, a structured node set and a connection relationship are established to construct a semantic state cognitive graph; S25, the semantic state cognitive graph includes: a user intention node for describing task requirements, a context information node for carrying input semantic context, and a scene label node for identifying an application environment; S26, the nodes are connected through directed edges, and the establishment of the edges includes: semantic edges established based on semantic dependency relationship, time sequence edges established based on context time sequence, and statistical edges established based on semantic co-occurrence frequency, and the edges are attached with edge weight attribute values.

5. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 2, characterized in that, The S23 specifically includes: S231, preset the semantic slot order in the system initialization stage, which includes user intention slot, context slot and scene label slot in sequence, the slot order remains fixed during system operation and is prohibited from being dynamically modified; S232, set a unified vector dimension standard for the three categories of sub-vectors, all sub-vectors must be dimensionally aligned before splicing, and if any category of semantic elements is missing, the corresponding slot is filled with zero vector to maintain the structural stability of the semantic state vector; S233, each slot in the structure of the semantic state vector is an independent semantic domain, the system has a semantic consistency rule, and if there is a conflict expression between the user intention slot and the context slot, the vector splicing process will be blocked and a correction request will be returned; S234, in the splicing operation, a linear connection method is used to connect the three categories of sub-vectors in the preset semantic slot order to form a complete semantic state vector; S235, the semantic state vector constitutes a one-dimensional fixed-length multi-slot structure, wherein each slot is input as an independent node in the downstream graph construction process without vector compression and fusion processing; S236, the semantic state vector is attached with 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.

6. 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 a semantic backbone for task recognition; S32, map the extracted semantic backbone content to an intermediate instruction representation, which includes an operation type field, an object identification field and a restriction condition field; S33, structure analysis 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, a slot combination feature based on a semantic state vector is used, a predefined instruction template is preferentially matched, and a rule set configured in the system is called to complete 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 label mapping relationship table.

7. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 5, characterized in that, The S33 specifically includes: S331, the slot type intermediate semantic instruction structure includes 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 includes a slot label, a slot value and a slot position confidence three fields; S333, the slot value is filled by an instruction translation module according to a semantic element matching result in the semantic state vector, and the slot position confidence is directly generated from the matching score and is used to represent the semantic reliability of the slot; S334, when a slot value is missing, the system fills the slot value according to a corresponding default parameter according to the slot type, and when the slot value exists but the confidence is lower than a set threshold, a preset placeholder is filled in 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 the consistency of the structure; S336, the slot type intermediate semantic instruction structure must pass through slot semantic consistency verification before being output, and if semantic conflict relations exist between any two slots, the system stops the intermediate semantic instruction structure from passing down the process and sends a structure reconstruction request to an upstream module.

8. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 2, characterized in that, The S4 specifically includes: 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; 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 the semantic deviation degree between the slot content and the 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 includes slot value replacement, semantic weakening processing and semantic conflict marking; S45, the reconstructed intermediate semantic instruction structure will be re-verified, and the verification includes: slot structure integrity verification, used to determine whether each mandatory slot is missing; slot type matching verification, used to confirm whether the slot content conforms to the semantic category defined by the slot; slot logic consistency verification, used to detect whether there is a conflict expression, a logic inversion or a mutual exclusion relationship between the slots; when any verification fails, the system refuses to pass the structure to a downstream module and marks it as an abnormal structure state; S46, the intermediate semantic instruction structure that is passed through the check 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.

9. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 7, characterized in that, The S42 specifically includes: S421, each score value in the bias score matrix is calculated 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 given 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 has the highest score weight, and the constraint condition slot has the lowest score weight; S423, all score results are normalized to form a final bias score matrix, and the normalization adopts a maximum and minimum scaling method, so that the score values are distributed in the [0, 1] interval; S424, the system sets a bias score threshold for each type of slot, wherein the operation type slot has the smallest tolerance, and the constraint condition slot has the largest tolerance, and if a slot score value exceeds the corresponding threshold, the slot is determined to have semantic deviation; S425, when any slot score in the bias score matrix exceeds the set threshold, the system marks the slot as a high bias state.

10. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 7, characterized in that, The S45 specifically includes: S451, the structure check 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 check content includes whether the slot exists and whether the slot value is non-empty; S452, the slot type matching check is performed according to a built-in slot semantic definition table of the system, and the check 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 inter-slot logic consistency check 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 check in the order of integrity, type matching, and logic consistency, and if any check 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 a structure exception state.

11. The multi-scene intelligent consulting system based on a pre-trained large model according to claim 2, 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, and the task resource portrait 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 a pruned expert model structure graph is generated; 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 sequences according to the task priority of the expert model and the device idle window, preferentially loads the pruned expert model structure, dynamically allocates the video memory and the computing thread, and performs the matching binding of the software and hardware resources; S75, when the system detects that the expert model loading fails or the runtime resource conflicts, the fallback mechanism is automatically called, and the expert model structure of the default version is restored.

Images