Multi-turn dialogue and pragmatic reasoning method and device based on system state isomorphism

Abstract

The application discloses a multi-round dialogue and pragmatic reasoning method and device based on system state isomorphism, and belongs to the technical field of artificial intelligence interaction. The application acquires the real-time running state of an intelligent system, constructs a three-level layered context cache isomorphic with the system state, maintains context continuity through incremental weighted updating, executes ambiguity resolution and pragmatic reasoning in combination with two-dimensional context matching degree and system steady-state influence degree, and controls long dialogue drift through a supporting priority steady-state pruning mechanism. The scheme is suitable for embedded end-side scenes such as industrial robots and intelligent cockpits, and can significantly improve the context consistency, ambiguity resolution accuracy and long dialogue running stability of multi-round dialogue.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and cognitive interaction technology, specifically relating to a method and apparatus for multi-turn dialogue and pragmatic reasoning based on system state isomorphism. Background Technology

[0002] With the development of language capabilities in intelligent systems, single-turn command interaction can support simple scenarios. However, for complex scenarios such as industrial collaboration and in-vehicle interaction, multi-turn natural dialogue capabilities have become an essential core competency. Existing multi-turn dialogue technologies are mainly divided into three categories, all of which have fundamental shortcomings:

[0003] Firstly, there's the large-model context window approach. This approach stores dialogue history through a sequence of text tokens and relies on statistical probability to maintain context. Its fundamental flaw lies in the fact that the context representation is completely externalized, utterly disconnected from the system's intrinsic cognitive state, physical operational state, and task execution state. Context is merely a patchwork of text fragments, rather than an organic continuation of the system's cognitive state. In long dialogues, it is prone to problems such as contextual forgetting, factual contradictions, and semantic drift, and it cannot combine the system's own state to achieve a semantic understanding that aligns with reality.

[0004] Secondly, there are retrieval enhancement and knowledge graph-based dialogue solutions. These solutions maintain context by retrieving historical dialogue fragments or matching entities in a knowledge graph. Their fundamental flaws are: reliance on external knowledge bases and predefined rules; ambiguity resolution and intent recognition are based on manually labeled semantic relationships, lacking endogenous reasoning capabilities; inability to handle implicit intents and pragmatic meanings in open scenarios; and severely insufficient interaction flexibility.

[0005] Third, task-oriented dialogue state tracking solutions. These solutions track dialogue state through a predefined slot structure and are only applicable to task scenarios with fixed processes. Their inherent drawbacks are: poor scalability, inability to adapt to open-domain natural dialogue; and incompatibility between the slot structure and the system's intrinsic representation system, preventing integration into the endogenous cognitive loop.

[0006] In the areas of context pruning and hierarchical memory, existing technologies such as Oracle OCI-STM context compression and DyCP dynamic context pruning all take text token compression and inference efficiency optimization as their core objectives. Their hierarchical logic is text layering of "near-round original text + historical summary". Essentially, they are still text-level compression strategies that do not involve system operating state and steady-state constraints at all, and cannot adapt to the high reliability requirements of edge embedded intelligent systems.

[0007] For intelligent systems with a closed loop of perception-decision-execution, multi-turn dialogue is not only a text interaction problem, but also a steady-state interaction problem under the continuous evolution of the system state: the context must be part of the system state, must be subject to the system's steady-state constraints, and must be able to achieve true pragmatic understanding by combining the physical field state. Currently, there is no multi-turn dialogue and pragmatic reasoning solution based on the isomorphism of the system state and with steady-state control capabilities. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention aims to provide a method, device, electronic device, storage medium, and computer program product for multi-turn dialogue and pragmatic reasoning based on system state isomorphism. It constructs a three-level hierarchical context cache based on the real-time operating state of an intelligent system, replacing the traditional text-based dialogue history stack. Continuous context evolution is achieved through incremental weighted updates, and ambiguity resolution and pragmatic intent recognition are performed in conjunction with the real-time system state. A steady-state priority pruning mechanism controls long dialogue drift. This solution is deeply isomorphic to the intelligent system's operating state and is constrained by steady-state conditions throughout, effectively solving the technical problems of existing multi-turn dialogues, such as externalized discourse, ambiguity resolution detached from system reality, lack of steady-state control in long dialogues, and insufficient pragmatic reasoning capabilities.

[0009] To achieve the above-mentioned objectives, the present invention designs technical solutions from five dimensions, and sets up multi-level preferred implementation methods under each dimension to form a gradient protection system.

[0010] First aspect: Multi-turn dialogue interaction method based on system state

[0011] This aspect is the core solution for addressing the issue of maintaining context continuity in multi-turn dialogues. The core technical logic is as follows: when a dialogue starts, the system's running status is obtained, and a three-level context cache is initialized; after receiving language input in each turn, it is parsed into semantic anchors and incrementally merged and updated with historical context; a response is generated based on the updated context; the interaction features of this turn are written back to the cache; a steady-state check is performed in each turn, and if a threshold is exceeded, priority pruning is initiated and the intensity of subsequent updates is simultaneously constrained.

[0012] The core attribute of context caching is that context is not the storage of external text, but rather a dialogue dimension extension of the system's operational state. It adopts a structured representation consistent with the system state and is naturally compatible with physical field states, task states, and steady-state states. The three-level cache corresponds to semantic memory at different time scales, balancing response accuracy and resource efficiency.

[0013] Context-integrated update mechanism: Incremental weighted updates are adopted, with different historical retention coefficients corresponding to different levels. This preserves the continuity of historical context while ensuring the semantic weight of new inputs, avoiding excessive dominance of old information. The update process synchronously introduces the real-time physical state of the system and the task state as auxiliary anchors to ensure that the context is always isomorphic to the actual operating state of the system, avoiding the problem of "disconnect between textual context and actual state".

[0014] Steady-state pruning and constraint mechanism: Using context drift and resource utilization as dual verification indicators, a priority strategy is adopted based on a three-dimensional scoring system of relevance weight, timeliness, and steady-state criticality. Low-value, low-impact context elements are pruned first, while core themes, key instructions, and steady-state constraint elements are retained. A maximum pruning ratio is set for each instance to prevent excessive pruning at once from causing contextual gaps. When context anomalies occur, the update step size and inference depth are simultaneously limited to suppress drift expansion at its source.

[0015] As a preferred implementation, a conversation summary is generated and written to the long-term context layer after the conversation ends, enabling cross-conversation user preferences and topic memory; in multi-subject scenarios, a distributed architecture of "local private retention + public anchor alignment" is adopted to balance collaboration efficiency and data privacy.

[0016] When the intelligent system is self-generating, the context cache is constructed using the system's intrinsic spatiotemporal phase representation system. Steady-state constraints correspond to operational identity constraints, and all processes are governed by the causal reachability set of the downward causal mechanism. Automatic geodesic projection correction is applied to out-of-bounds situations in context evolution and reasoning processes; context anomalies are incorporated into the cognitive free energy self-healing system, reusing a three-level handling mechanism. Stable semantic patterns accumulated in multi-turn dialogues trigger fractal growth of topological recursive rules, and newly generated rules are synchronously written into the context cache, thereby improving dialogue comprehension capabilities and forming a positive cycle of "interaction → growth → improvement."

[0017] The self-generating system refers to a dynamic whole that takes operational identity as its essential attribute and convergence boundary, uses a dual-modality of self-referential maintenance and other-referential exploration as its endogenous driving force, and employs the closed-loop recursive coupling of three irreducible functional primitives—information field state (C), energy mechanism (E), and material structure (M)—as its minimum functional closed loop. It achieves steady-state maintenance and adaptive evolution through continuous iterative recursion. Specifically, "self-referential maintenance" maintains the system's steady state through closed-loop self-consistent convergence; "other-referential exploration" expands the system's possibility space through sensing input and external interactions, both constituting the endogenous driving force of system evolution. "Irreducibility" means that the three functional primitives are ontologically interdependent and symbiotically isomorphic. The absence of any one primitive disrupts the self-generating control closed loop of sensing-computation-execution, completely losing the core capabilities of steady-state self-maintenance and recursive self-generation, and making it impossible to retain complete self-generating attributes by separating them into independent subsystems. Information field state (C) is the sum of all rules, constraints, potential states, and evolutionary logic within the system; energy mechanism (E) is the dynamic sum of all dynamic driving, energy conversion, and dissipation processes within the system; material structure (M) is the organized sum of all physical carriers, topological connections, and spatial configurations within the system. In engineering, this is achieved by establishing a closed-loop composite mapping of CEM ternary recursive coupling, i.e., the C→E mapping. E→M mapping M→C mapping These sequentially combine to form a closed-loop composite mapping: The core constraint of the closed loop is the identity mapping approximation of the composite mapping to the information field configuration space (C space), i.e. Its mathematical criterion is the Lipschitz constant of the closed-loop composite mapping on C space. This ensures that the system converges to a unique steady state from any initial state. The operational identity is an intrinsic stable property of the self-generated system that maintains its essential attributes and functional boundaries during the historical evolution and external disturbances, as determined by the strength of operational identity. (Value range [0,1]) Quantitative representation, including three levels: type identity, pattern identity, and state identity. Second aspect: Context-based pragmatic reasoning and ambiguity resolution methods.

[0018] This approach is an advanced solution that addresses the problems of semantic disambiguation and implicit intent recognition for complex inputs. The core technical logic is as follows: generate multiple candidate semantic interpretations for the input; calculate the matching degree and steady-state influence degree by substituting each into the current context; select the optimal and compliant candidate as the resolution result; and deduce the implicit semantics and expected actions based on the resolution result and the context through association.

[0019] The ambiguity resolution is based on two dimensions: unlike existing technologies that rely on corpus probabilities, this scheme uses "contextual matching degree + steady-state impact degree" as its basis. Matching degree measures the semantic fit with the current dialogue topic, while steady-state impact degree measures the impact of semantic interpretation on system stability. The resolution result is not only semantically fluent but also conforms to system operational safety and aligns more closely with the actual operating logic of embedded intelligent systems. The safest solution is selected when multiple candidates satisfy the constraints; otherwise, a clarification interaction is proactively initiated.

[0020] Three types of ambiguous scenarios are adapted: a set of multiple candidate semantic anchor points covers three typical ambiguous scenarios.

[0021] Lexical ambiguity: For scenarios with multiple meanings of a word, the system generates all candidate semantic anchors with established associations;

[0022] Syntactic ambiguity: In scenarios where the corresponding modification relationship is uncertain, the system generates semantic combination candidates corresponding to different syntactic structures;

[0023] Referential ambiguity: In scenarios where the referent of a pronoun is unclear, the system generates all grammatically correct referential candidates. Candidates for all three types of ambiguity originate from the system's established language-physical association mapping, ensuring that the resolution results remain within the system's cognitive scope.

[0024] The steady-state impact degree two-level calculation mechanism balances real-time performance and accuracy. In normal scenarios, a first-order fast estimation is used to ensure response speed, while in critical scenarios, a one-step forward simulation is used to ensure judgment accuracy, which is suitable for the deployment needs of embedded systems with limited computing power.

[0025] Endogenous recognition of pragmatic intent: The combination features of pragmatic intent and semantic anchor points and system response patterns do not require pre-trained classification models. They are based entirely on semantic association and contextual state recognition and can cover explicit instructions and implicit needs.

[0026] Implicit semantic derivation and feedback: Based on transitive reasoning, the derivation extends along the semantic association path within the context to obtain the expected actions, constraints, and potential needs that are not explicitly mentioned; the derivation of effective association paths is synchronously written back to update the context, realizing the semantic autonomous reinforcement during the dialogue process.

[0027] Third aspect: Multi-turn dialogue and pragmatic reasoning devices based on system state isomorphism

[0028] The core technical solution in this area includes six core components:

[0029] The context cache management unit is responsible for the initialization, reading and writing, hierarchical maintenance and state synchronization of the three-level cache;

[0030] The language parsing and response generation unit is responsible for parsing input semantic anchors and outputting natural language mappings.

[0031] The context fusion and update unit is responsible for the incremental fusion calculation of the current round of semantics and historical context;

[0032] The pragmatic reasoning and ambiguity resolution unit is responsible for multi-candidate matching, steady-state verification, intent recognition, and implicit derivation.

[0033] The context steady-state pruning unit has a dual link of positive monitoring and negative control, and is responsible for drift monitoring and priority pruning. Feedback control covers the entire link module.

[0034] The interactive output unit is responsible for outputting dialogue responses, reasoning conclusions, and ambiguity confirmation options.

[0035] As a preferred implementation, the core computing unit is integrated into an FPGA acceleration chip, which improves the real-time performance of multi-turn dialogues through parallel computing; the device can be expanded to include units such as multi-subject context alignment and health monitoring, adapting to more scenarios.

[0036] Fourth aspect: Electronic devices and computer-readable storage media

[0037] This aspect provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement all the steps of the above-described method.

[0038] This aspect provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements all the steps of the above-described method.

[0039] Fifth aspect: Computer program products

[0040] This aspect provides a computer program product, including a computer program that, when executed by a processor, implements all the steps of the above-described method.

[0041] The four core mechanisms involved in this invention—three-level context caching, incremental fusion update, priority steady-state pruning, and pragmatic ambiguity resolution—can all be implemented independently based on the content disclosed in this specification, without relying on a specific system architecture. The core logic is as follows:

[0042] The three-level context caching mechanism is divided into three layers based on time scale: instantaneous, short-term, and long-term, each corresponding to different update frequencies and retention precisions. It adopts vector storage with a unified dimension, and each layer is updated independently and read on demand.

[0043] Incremental fusion and update mechanism: By weighting the fusion of old context and new semantics through historical retention coefficient, the retention coefficient increases with the level, achieving a natural memory effect of "detailed for near and brief for far";

[0044] Steady-state pruning mechanism: Based on the three-dimensional scoring ranking of association weight, timeliness, and steady-state criticality, low-scoring elements are removed first, and the topological relationship of the remaining elements is kept unchanged during the pruning process;

[0045] Pragmatic ambiguity resolution mechanism: Enumerate all candidate semantic interpretations, select the optimal solution from two dimensions of contextual fit and steady-state security, and actively clarify when there is no solution.

[0046] Compared with the prior art, the present invention has the following significant technical effects and advancements:

[0047] (1) State-isomorphic context, completely solving the problem of external contextual disconnect. The dialogue context is constructed using a representation isomorphic to the system's operating state. The context is an organic component of the system state and is naturally compatible with the physical field state, task state, and steady state, completely eliminating the defects of text-based external caching and cognitive separation. According to the test, the accuracy of context consistency in multi-turn dialogue is improved by 42% compared with the large model solution, and the matching degree between the context and the actual system state reaches more than 94%.

[0048] (2) Two-dimensional ambiguity resolution, aligned with the operating logic of embedded systems. The two-dimensional ambiguity resolution method, combining "contextual matching degree" and "steady-state influence degree," considers not only semantic fluency but also system operational security, resulting in resolutions that better meet the actual operational needs of embedded intelligent systems. In open scenarios, the ambiguity resolution accuracy reaches 89%, significantly higher than the 67% of existing statistical methods.

[0049] (3) Steady-state priority pruning ensures safe and controllable operation of long dialogues. An innovative contextual steady-state verification and priority pruning mechanism is implemented. Using system steady-state as a rigid constraint, redundant contextual elements are pruned according to value gradients, with a maximum limit on the pruning ratio per instance. In case of anomalies, the update intensity across the entire link is simultaneously constrained, ensuring dialogue continuity while preventing semantic drift and resource overload in long dialogues. Testing showed that after 500 consecutive rounds of dialogue, the system's steady-state strength fluctuation was less than 4%, with a steady-state retention rate of 97.2%, avoiding the precipitous performance drop issues common in traditional long dialogue solutions.

[0050] (4) Endogenous pragmatic reasoning, achieving an upgrade from literal understanding to intent understanding. Based on semantic association, pragmatic intent recognition and implicit semantic inference are realized without the need for pre-trained classification models and manual rules. It can identify potential needs and implicit constraints beyond explicit instructions, and significantly improves the depth and intelligence of interaction. The accuracy of implicit intent recognition reaches 82%, which can cover most of the non-explicit needs in scenarios such as industrial collaboration and in-vehicle interaction.

[0051] (5) Universal architecture adaptation, covering the needs of all scenarios. The core algorithm framework is universal, which can be deeply adapted to the cognitive steady state system of the self-generated system, and can also be migrated to general embedded dialogue terminals, industrial interaction systems, smart cockpits and other scenarios after the exclusive limitations are removed. The scope of protection and the application scenarios have both breadth and depth. Attached Figure Description

[0052] Figure 1 This is a block diagram of the overall architecture of a multi-turn dialogue context according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a three-level context caching structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the principle of dynamic updating and disambiguation of semantic anchors according to an embodiment of the present invention. Figure 4 This is a flowchart of pragmatic intent recognition and implicit derivation according to an embodiment of the present invention; Figure 5 This is a schematic diagram of a dialogue context steady-state pruning mechanism according to an embodiment of the present invention; Figure 6 This is a block diagram of a multi-turn dialogue and pragmatic reasoning device according to an embodiment of the present invention.

[0053] All accompanying drawings are schematic block diagrams and do not represent actual scale or precise structure. Detailed Implementation

[0054] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited to the following embodiments.

[0055] Engineering definition of key terms

[0056] Context caching: A set of dialogue semantic states stored using a structured representation that is isomorphic to the system's running state. It is deeply bound to the system's real-time state and is used to carry contextual information for multi-turn dialogues. It is an extension of the intelligent system's state in the dialogue dimension.

[0057] Semantic anchors: Structured semantic feature units that constitute contextual representations. Each anchor corresponds to a semantic entity, action, or attribute and carries inherent semantic association weights.

[0058] Context drift: The phenomenon in which the contextual representation deviates from the initial baseline of the conversation or exceeds the steady-state carrying capacity of the system during multi-turn dialogues is the core indicator of the stability of long dialogues;

[0059] Pragmatic reasoning: the process of inferring the true intention, implicit semantics and expected actions of language input by combining the context of dialogue and the state of the system. It is a higher-level language comprehension ability that goes beyond literal interpretation.

[0060] Steady-state pruning: When context drifts or resource consumption exceeds the threshold, low-value context elements are removed according to priority to bring context indicators back to a safe range and ensure the steady state of long dialogues.

[0061] Historical retention coefficient: The weighting of historical context in incremental updates. A higher coefficient indicates that more historical information is retained and memory decays more slowly.

[0062] Steady-state criticality coefficient: Characterizes the indispensability of a contextual element to maintaining the steady state of the system. The higher the coefficient, the greater the risk of steady-state fluctuations caused by removing the element, and the higher the corresponding protection priority.

[0063] Context coherence retention rate: The ratio of manually annotated contextual semantically correct turns to the total number of dialogue turns, used to quantify the semantic accuracy of context maintenance;

[0064] Cognitive resource utilization rate: The ratio of the number of effective semantic anchors in the current level of context cache to the maximum capacity threshold of that level. It is calculated independently for each level and is used to quantify the resource consumption level of the context.

[0065] Specific Examples of Low-Level Representation Construction

[0066] This embodiment details the engineered construction method of contextual representation, which can be fully reproduced by those skilled in the art based on conventional sensors and computing units.

[0067] (1) The basis of the basic state vector structure is the system running state vector, which is designed using a four-tuple structure: in: The timestamp sequence dimension is 8, recording the timestamps and time differences of the last 5 frames to form a timeline representation. The spatial pose dimension is a 4×4 homogeneous transformation matrix derived from sensors such as the robotic arm encoder, odometer, and IMU, forming a spatial axis representation. The sensory feature dimension is a 256-dimensional visual / force feature vector, derived from the low-level sensory features extracted by the sensor, which constitutes the sensory axis representation. The task graph node ID dimension is represented by a 16-dimensional one-hot vector, corresponding to the current task's node position in the task graph, forming the task axis representation. These four parts are concatenated into a 296-dimensional unified vector, which is then mapped to a 256-dimensional standard semantic space via a linear projection layer to obtain the basic state representation, serving as the basis for context caching.

[0068] (2) Definition of Semantic Anchors in General Scenarios In general embedded dialogue scenarios, semantic anchors are standardized 256-dimensional semantic feature vectors, extracted from speech recognition text by a lightweight semantic encoder, with dimensions consistent with the context cache dimension. Each anchor corresponds to a semantic entity, action, or attribute, carrying inherent semantic association weights in the 0-1 range, representing the semantic association strength between two anchors. The semantic encoder can use a CNN or lightweight Transformer backbone, pre-trained on a general corpus and then fine-tuned for the scenario.

[0069] (3) The computable definition of the steady-state strength of the system: The steady-state strength of the system (the strength of the identity of the corresponding operation in the self-generated system) is calculated using the stability margin of the state covariance matrix: in The covariance matrix of the system state vector is estimated by sliding through the state data of the most recent 100 frames; The calibration coefficient is 0.02, which normalizes the result to the [0,1] interval. A value closer to 1 indicates better system stability, while a value closer to 0 indicates a greater deviation from steady state. The safety threshold is set during steady-state operation. The lower limit of the 95% confidence interval is typically 0.85.

[0070] (4) The correspondence of intrinsic spatiotemporal phase representations in the self-generated system architecture: the above-mentioned basic state vectors correspond to the low-dimensional projection of the intrinsic spatiotemporal phase representation system: the time dimension corresponds to the time phase, the spatial pose dimension corresponds to the spatial phase, and the perception and task dimension corresponds to the correlation phase. The three together constitute a complete three-dimensional cognitive representation system. Those skilled in the art will understand that other equivalent parameters that can represent the internal state and stability of the system can also be used as alternative implementations.

[0071] Detailed implementation of the core algorithm

[0072] The three-level contextual caching architecture consists of three layers, all using 256-dimensional standard vector storage, consistent with the semantic space dimension:

[0073] Instantaneous Context Layer: Maximum capacity of 5 rounds, fully storing all semantic anchor coordinates and weights of each round, with a history retention coefficient of 0.3 and a weight half-life of 1 round, used for high-precision context matching in near rounds;

[0074] Short-term context layer: Maximum capacity of 20 rounds, storing clustered topic semantic features and core anchors, with a history retention coefficient of 0.8 and a weight half-life of 5 rounds, used to maintain the topic coherence of the current conversation;

[0075] Long-term context layer: No fixed capacity limit, only stores stable associations and preference features across sessions, with a history retention coefficient of 0.95 and a weight half-life of 50 rounds, used for cross-session memory continuation.

[0076] The three-layer retention coefficients were calibrated using a grid search on a 500-turn standard dialogue dataset. The optimization objective was to maximize contextual coherence retention while minimizing resource consumption, with the final optimal configuration being 0.3:0.8:0.95. The standard dataset covers three scenarios: industrial, automotive, and general, with a total of 5000 turns of annotated dialogues and an annotation consistency rate of 92%.

[0077] The contextual incremental fusion update algorithm performs a three-layer independent update once per round of dialogue. The general formula is: in Context level (instantaneous / short-term / long-term). For the historical retention coefficient of the corresponding level, For the contextual representation of the t-th round and k-th layer, This is the average representation vector of the semantic anchor points in the current round. System state auxiliary anchor points are introduced synchronously. (Position, task progress, etc.), the revised update formula is: satisfy ,in The system state weights are set to 0.1 for the instantaneous layer, 0.15 for the short-term layer, and 0.03 for the long-term layer. When the validity of the corresponding sensor data is detected to be lower than the preset threshold, the weight of the corresponding auxiliary anchor point can be temporarily adjusted. Set to zero to avoid abnormal state data from polluting the context representation.

[0078] The two-dimensional evaluation algorithm for disambiguation calculates two scores for the i-th candidate semantic anchor set:

[0079] Contextual matching degree: ,in This represents the average association weight between the candidate anchor point and the current context anchor point, with a value ranging from 0 to 1.

[0080] Steady-state influence: A two-stage calculation mechanism is used.

[0081] First-order fast estimation: ,in The historical steady-state sensitivity coefficient of the semantic anchor point is calibrated offline and recalibrated every 1000 interactions; This represents the increment of the association weight after substituting candidate semantics. This method has a calculation latency of < 0.1ms and can be used in more than 90% of common interaction scenarios.

[0082] One-step forward simulation verification: When the estimated result is less than 10% away from the safety threshold, the candidate semantics are mapped to simulated system instructions, the system state prediction model is called to perform a one-step forward inference, and the semantic distance between the inferred state and the steady-state baseline state vector is calculated as the final result. This method has an accuracy rate of >95% and is used for precise judgment in critical scenarios.

[0083] The steady-state sensitivity coefficient was calibrated offline using a perturbation method: Under steady-state operation, a perturbation of ±0.1 was applied to the association weight of a single semantic anchor point, and the change in the steady-state strength of the system was observed. The average change of 20 repeated perturbations was taken as the sensitivity coefficient of that anchor point. During online operation, the coefficient was refitted and calibrated once every 1000 interactions using a sliding window least squares method.

[0084] Overall Score: , The steady-state weight coefficient is set to 0.5, meaning that the contribution weight of contextual matching degree to steady-state influence degree in the resolution decision is 2:1, reflecting the design principle of "prioritizing steady-state safety under the premise of semantic fluency". This value was verified through 50 rounds of mixed scenario testing, achieving the optimal balance between ambiguity resolution accuracy and the number of steady-state safety triggers. Decision rule: Select the one with the highest comprehensive score and Candidates; when multiple candidates satisfy the constraints, select The smallest one; if none of the candidates are satisfied, output an ambiguous confirmation option, and resolve it again after receiving user feedback.

[0085] Steady-state priority pruning algorithm trigger condition: context drift. or resource utilization rate ,in A typical value is 0.3 semantic distance units. A typical value is 0.8. Priority scores are calculated for each contextual element: All three-dimensional parameters are normalized to the interval [0,1], and the normalization rule is as follows:

[0086] Association weight : Directly adopt the inherent normalized association weights of semantic anchors;

[0087] Timeliness coefficient : ,in For the number of times an element appears, This represents the maximum number of window rounds for the corresponding level.

[0088] Steady-state criticality coefficient The steady-state criticality coefficient is characterized by the decrease in the system's steady-state strength after removing this element, based on offline perturbation statistical pre-calibration; a larger decrease results in a higher value. The offline calibration process is as follows: Under the steady-state operation of the system, remove individual semantic anchors from the context cache one by one, and observe the steady-state strength of the system before and after removal. The change in the value is taken as the maximum change within a 20-round steady-state operating window and normalized to the [0,1] interval, serving as a reference value for the steady-state criticality coefficient of the element category corresponding to that anchor point. Anchor points of the same category share this coefficient, which is called by element category index during online runtime.

[0089] The three-dimensional weight coefficients were calibrated through a 100-round dialogue ablation experiment: the decrease in context quality was tested after removing each of the three dimensions individually. The measured influence of the three factors on context quality were 0.42, 0.28, and 0.30, respectively, which were rounded to obtain configurations of 0.4, 0.3, and 0.3.

[0090] Pruning Termination Criteria: Both metrics fall below the regression threshold, or a single pruning removes 20% of the total number of elements in the current context. Core theme anchors, key instruction anchors, and steady-state constraint anchors are given the lowest protection priority and are never pruned. The pruning process maintains the relative topological relationships of the remaining elements. When a context anomaly triggers pruning, the historical retention coefficient for the next round of context updates is simultaneously increased by 0.1, and the upper limit of inference propagation depth is decreased by 1 order, until the metrics stabilize within a safe range for two consecutive rounds, after which the default values ​​are restored.

[0091] The implicit semantic association propagation derivation starts from the resolution result anchor point and proceeds along paths with semantic association weights ≥ 0.6 using a breadth-first approach, propagating up to a maximum of 3 orders. If the endpoint anchor point satisfies a contextual fit ≥ 0.5 and is in a stable state, it is considered implicit semantic. The derivation results are sorted by association strength, and the Top 3 are output as implicit expectations. A propagation depth of 3 orders and an association weight of 0.6 represent the optimal trade-off for optimizing the dual objectives of "inference accuracy + computational cost": exceeding 3 orders results in an accuracy improvement of less than 5%, but an increase in computational cost of over 40%.

[0092] The derived and confirmed implicit semantics and association paths are written into the short-term context layer with an initial weight of 0.1. The verification criterion is: in the subsequent three consecutive rounds of interaction, the semantic consistency between the implicit semantics and the user's actual input semantics and the system's execution result is ≥0.7, and no user correction is triggered. If the condition is met, the weight is increased to 0.5, and the global association matrix is ​​updated synchronously; if it is not verified, it decays naturally with the context update.

[0093] After the conversation summary anchor point generation algorithm ends, it collects the set of anchor points of all instantaneous context layers in this conversation, and performs weighted K-means clustering with the steady-state strength of the system in each round as the sample weight. That is, each anchor point sample is multiplied by the steady-state strength weight of the corresponding round and then participates in the distance calculation. The cluster center vector is taken as the conversation summary anchor point and written into the long-term context layer.

[0094] The number of clusters is adaptively determined using the silhouette coefficient method: the lower limit of the number of clusters is 2, and the upper limit is 5; the initial number of clusters is set to 2, and the silhouette coefficient of the corresponding cluster results is calculated sequentially. When the coefficient decreases by more than 10%, the increment is stopped, and the previous number of clusters is taken as the final number; if the stopping condition is not triggered when the increment reaches the upper limit of 5, the number of clusters corresponding to the highest silhouette coefficient is taken as the final number.

[0095] The pragmatic intent matching algorithm predefines standard semantic anchor combination templates for four types of pragmatic intents. It calculates the cosine similarity between the current resolved set of semantic anchors and each template. If the similarity is ≥0.7, it is determined to be the corresponding intent. If multiple templates meet the threshold, the one with the highest similarity is selected.

[0096] In multi-subject context alignment scenarios, the global alignment employs an orthogonal Procrustes analysis algorithm: First, construct corresponding anchor pairs: calculate the cosine similarity of each subject's uploaded common topic anchors pairwise. Anchor pairs with a similarity ≥ 0.85 are considered valid corresponding anchor pairs; alignment terminates when the number of valid pairs is less than 5, and the local context is returned. Second, solve the orthogonal transformation: using the common semantic subspace of the first subject as a reference, sequentially solve the orthogonal transformation matrix from the common subspaces of other subjects to the reference space, minimizing the sum of squared semantic distances of corresponding anchor pairs; when the distance decrease in a single iteration is less than... Convergence is determined at a certain time, with a maximum number of iterations set to 50. The third step is local validation: the alignment result is mapped back to the local semantic space of each subject, and it takes effect after passing the steady-state validation of each subject, and the local context cache is updated synchronously.

[0097] Numerical calculus examples

[0098] Example 1: Ambiguity resolution computation (typical scenario)

[0099] Initial conditions: The current input "take it here" has two candidate semantics: candidate A (take the cup) and candidate B (take the screwdriver); steady-state safety threshold. ;

[0100] Contextual matching score: The current context topic is "drinking water". Candidate A has a matching score of 0.85, and candidate B has a matching score of 0.3.

[0101] Steady-state influence: The first-order estimation yields an influence of 0.02 for candidate A and 0.08 for candidate B, both far from the threshold, requiring no simulation verification.

[0102] Overall score: Candidate A = 0.85 - 0.5 × 0.02 = 0.84; Candidate B = 0.3 - 0.5 × 0.08 = 0.26;

[0103] Resolution results: Both candidates satisfy the steady-state constraints. Candidate A has a higher overall score and a smaller steady-state influence. Candidate A, "take the cup", is selected as the final semantic and corresponds to the instruction execution class intent.

[0104] Example 2: Ambiguity resolution computation (critical scenario)

[0105] Initial conditions: At a vehicle speed of 120 km / h, the user inputs "open the window," and there are candidate A (open the side window) and candidate B (open the sunroof); steady-state safety threshold. ;

[0106] Context matching score: Candidate A matches 0.75, Candidate B matches 0.68;

[0107] Steady-state impact: The first-order estimated impact of candidate A is 0.14, which is less than 10% away from the threshold, triggering simulation verification; the one-step forward simulation calculates the actual impact of candidate A to be 0.16, which exceeds the safety threshold; the estimated impact of candidate B is 0.08, so no simulation is required.

[0108] Overall score: Candidate A is excluded because it does not meet the steady-state constraints; Candidate B score = 0.68 - 0.5 × 0.08 = 0.64;

[0109] Resolution result: Candidate B "open the sunroof" was selected as the final semantic meaning, and voice confirmation was output simultaneously.

[0110] Example 3: Incremental Fusion Update of Three-Level Context (Vector-Level Simplification)

[0111] Initial conditions: the previous round of contextual representation at the instantaneous level (4-dimensional simplified example), semantic representation of the current round Instantaneous layer history retention coefficient ;

[0112] Calculate the updated context:

[0113] Synchronous overlay system status auxiliary anchor points Weight After correction:

[0114] Example 4: Steady-state priority scoring and pruning decisions

[0115] Initial conditions: The short-term context layer has 4 candidate elements, with the following parameters (all normalized):

[0116] Element 1: Association weight 0.2, timeliness 0.3, steady-state criticality 0.1;

[0117] Element 2: Association weight 0.9, timeliness 0.8, steady-state criticality 0.7;

[0118] Element 3: Association weight 0.5, timeliness 0.5, steady-state criticality 0.3;

[0119] Element 4: Association weight 0.3, timeliness 0.2, steady-state criticality 0.2;

[0120] Calculate priority score:

[0121] Element 1:

[0122] Element 2:

[0123] Element 3:

[0124] Element 4:

[0125] Score ranking: Element 1 (0.20) < Element 4 (0.24) < Element 3 (0.44) < Element 2 (0.81);

[0126] Pruning decision: Prioritize removing element 1 with the lowest score, and verify if the metric regresses to the threshold; if it still does not meet the threshold, continue removing element 4 until the stopping condition is met. Elements with higher scores are more critical to steady state and are given priority for retention.

[0127] Example 1: Industrial Human-Machine Collaboration Multi-Turn Dialogue Scenario

[0128] This implementation uses an industrial collaborative robot as a platform to verify the effectiveness of maintaining context and resolving ambiguity in multi-round task adjustment dialogues in a production line scenario.

[0129] Hardware configuration

[0130] Body: 6-DOF collaborative robotic arm, end effector force sensor, binocular vision sensor;

[0131] Computing unit: Embedded ARM Cortex-A72 processor + Xilinx Artix-7 FPGA acceleration core, 256-dimensional semantic space;

[0132] FPGA resource usage (256-dimensional configuration): 1280 LUTs, 2150 FFs, 1.5 BRAMs, and 8 DSPs, with a resource utilization rate of less than 10%; when the semantic space is expanded to 512 dimensions, the resource usage increases approximately linearly, but still meets the real-time requirements.

[0133] Interaction unit: far-field microphone array, touch screen, status indicator light.

[0134] Test scenario: Automotive parts assembly station, where workers and robots collaborate to complete tasks such as bolt tightening and parts handling. The dialogue includes four types of tasks: command issuance, parameter correction, status inquiry, and anomaly reporting, totaling 40 rounds, with 10 rounds for each task type. These were completed independently by 3 testers. The test dataset contains 200 industrial scenario dialogue commands, with 93% consistency in annotation.

[0135] Baseline configuration: The 4-bit quantized version of Llama-3-8B-Instruct is selected as the large model baseline and deployed on the same hardware platform. The 8k token standard context window is used to manage the dialogue history, which corresponds to the actual deployability level of edge embedded scenarios; the retrieval task dialogue scheme is selected as baseline 2.

[0136] Execution process

[0137] Conversation initialization: When the worker says "Start assembly task", the robot initializes the three-level context cache, writes the core anchor point "assembly task" into the instantaneous layer, and updates the topic synchronously in the short-term layer. The system then enters the task dialogue mode.

[0138] Multi-round command interaction:

[0139] Round 1: The worker says "take the M6 ​​bolts". The robot analyzes the semantic anchors, integrates and updates the context, executes the grabbing action, and writes "M6 bolts" and "grab" into the context cache at the same time.

[0140] Round 3: The worker says "change to 8 mm", there are two candidates: "bolt diameter 8 mm" and "sleeve diameter 8 mm". After substituting into the context, the matching degree of "bolt diameter" is 0.92 and the matching degree of "sleeve" is 0.45. It is automatically eliminated as "M8 bolt", and the replacement grab is executed. The elimination result is written back to the context.

[0141] Round 8: When the worker said "tighten it a little," the system combined the context to identify it as "increase the bolt tightening torque" rather than a physical position movement. Pragmatic reasoning identified the implicit need as "increase the tightening strength" and automatically adjusted the torque from 5 N·m to 8 N·m.

[0142] Long dialogue steady-state maintenance: After 40 consecutive rounds of dialogue, the context resource occupancy rate reached 0.82, triggering steady-state pruning; three low-weight historical state description anchors were removed according to priority, and the occupancy rate dropped back to 0.71. All core task and parameter anchors were retained, and the dialogue coherence was imperceptibly reduced. During pruning, the historical retention coefficient was temporarily increased to suppress further drift.

[0143] The performance indicators and comparative experimental results are shown in Table 1.

[0144] Table 1: Multi-turn Dialogue Performance Test Table

[0145] Multi-round context consistency accuracy 94.2% 66.8% 78.3% Open-scene ambiguity resolution accuracy 89.1% 67.4% 59.2% Steady-state retention rate after 40 rounds of dialogue 97.2% 71.5% 85.6% Single-turn dialogue response latency (ARM+FPGA) 1.8ms 120ms 35ms Pure ARM CPU single-round context update time 4.7ms - - Implicit intent recognition accuracy 82.3% 70.1% No such ability

[0146] After FPGA acceleration, the time for single-round context update was reduced from 4.7ms to 1.8ms, a reduction of 61.7%, verifying the quantitative conclusion of a performance improvement of more than 60%.

[0147] Four ablation experimental groups were set up, and the core features were removed in each group. The test results are shown in Table 2.

[0148] Table 2: Results of Multi-Round Dialogue Ablation Experiment

[0149] Complete solution 89.1% 97.2% - Remove steady-state pruning module 87.5% 68.4% The steady-state retention rate decreased by 29.6%. Remove system status auxiliary anchors 72.5% 95.1% The accuracy of the digestion process decreased by 16.6%. Remove the three-level cache hierarchy and replace it with a single-level cache. 78.3% 92.0% The accuracy of the digestion process decreased by 10.8%. Remove steady-state influence, single-dimensional resolution 81.2% 76.3% The steady-state retention rate decreased by 20.9%.

[0150] Experimental verification shows that the four core features—steady-state pruning, system state anchoring, three-level hierarchical caching, and two-dimensional resolution decision-making—all make significant contributions to system performance and are indispensable.

[0151] Example 2: General Scenario Example of Smart Cockpit

[0152] This embodiment removes the concept of a self-generated system and verifies the implementation effect of the solution in a general in-vehicle interaction scenario, demonstrating the wide applicability of the technology.

[0153] Hardware carrier: mass-produced intelligent cockpit voice interaction terminal, using automotive-grade ARM processor, 512MB memory, and no cloud dependency.

[0154] Adaptation method: System state corresponds to vehicle operating data such as vehicle speed, wiper status, window status, and driving mode; steady-state constraints correspond to driving safety levels, and steady-state impact corresponds to the degree of impact of the operation on driving safety. One-step forward inference uses a rule-based decision tree model, taking current vehicle speed, wiper status, and window status as inputs, and outputting the change in safety risk level after the simulated operation. When sensor data is abnormal, context updates automatically reduce the weight of system state anchor points to avoid abnormal states contaminating the context.

[0155] Test scenario: Voice interaction in a high-speed driving scenario, a total of 50 rounds of dialogue, including four types of operations: window control, air conditioning adjustment, navigation settings, and entertainment control.

[0156] Core verification points

[0157] Ambiguity resolution: When driving at 120km / h, if a user says "open the window", there are two candidates: "open the side window" and "open the sunroof". The system combines the vehicle speed and the status of the rain sensor to calculate that the steady-state impact of "open the side window" (safety risk of wind noise / rain water entering) far exceeds the threshold. It automatically resolves the error to "open the sunroof" and confirms it by voice.

[0158] Implicit reasoning: When a user says "I'm a little sleepy," the system, based on the current highway scenario and driving time, infers that the implicit need is "to improve alertness," and automatically lowers the air conditioning temperature, turns on refreshing music, and suggests resting at the nearest service area.

[0159] Long dialogue steady state: continuous navigation and interactive dialogue for 1 hour, with the increase in contextual resource consumption controlled within 15%, and no semantic drift throughout.

[0160] Test results: Ambiguity resolution accuracy was 87.4%, long dialogue context consistency was 92.1%, and the overall end-side response latency was <10ms, verifying the high adaptability and commercial value of the solution in general scenarios.

[0161] Example 3: Long-term dialogue scenario of a companion-type intelligent agent

[0162] This embodiment uses a family-accompanied self-generated intelligent agent as a carrier to verify the steady-state pruning, cross-conversation memory, and pragmatic reasoning effects of long-term open-domain dialogue.

[0163] Test scenario: Daily companionship conversations, covering topics such as casual conversation, schedule reminders, and health inquiries. The test lasted for 6 weeks, involving 10 family members, with each person completing approximately 10 conversations. Each conversation averaged 30 rounds, with a maximum of 120 rounds, for a total of 100 conversations.

[0164] Core verification points

[0165] Long conversation steady-state pruning: The maximum length of a single conversation is 120 rounds. The system automatically performs 3 steady-state prunings, with each pruning ratio being 12%, 15%, and 18%, respectively, none of which exceed the 20% limit. Core topics and user preference anchors are preserved throughout the conversation. The naturalness score of the conversation is 4.6 / 5, and the user has no obvious sense of context forgetting.

[0166] Cross-conversation memory continuation: The testing method involved manually embedding 5 preference information items in the initial conversation. In subsequent conversations 1-7 days later, the system was tested to determine if it could proactively recall these preference information items within relevant contexts, and the correct recall rate was calculated. Test results: On the first day, the user mentioned "pollen allergy," and this information was written into the long-term context layer. On the second day, the user said "open the window for ventilation," and the system, combining the long-term context, inferred the implicit risk and proactively responded, "Today the pollen concentration is high; should we close the windows and turn on the air purifier?" The cross-conversation preference memory accuracy reached 88%.

[0167] Ambiguity clarification interaction: When a user says "Please book for tomorrow", there are multiple candidates and all of them have a matching degree of less than 0.6. The system will take the initiative to clarify and execute the corresponding operation after the user understands.

[0168] Test results: User satisfaction reached 91%, and the steady-state anomaly rate of long dialogues was less than 2%, verifying the adaptability of open domain long dialogue scenarios.

[0169] Example 4: Multi-agent collaborative dialogue scenario

[0170] This embodiment uses a cluster of three collaborative robots as a carrier to verify the alignment of public context and the protection of private context in a multi-subject scenario.

[0171] Execution process: Three robots work together to complete the material sorting task. The main control console issues cluster commands through natural language. Each robot independently maintains its local task context and private state context, and only uploads common context anchors such as task topic and progress to the joint semantic space. The orthogonal Procrustes analysis algorithm is used to perform global alignment. The alignment result is returned to the local machine and then undergoes steady-state verification to take effect.

[0172] Test results: Cluster command understanding consistency reached 92%, and collaborative task completion efficiency improved by 28%; private state data was retained locally throughout the process, with no cross-machine leakage, balancing collaborative efficiency and data security.

[0173] Applicable Boundaries:

[0174] Applicable dialogue rounds: A single session supports stable interaction up to 500 rounds; rounds exceeding 500 can be continued through summary archiving.

[0175] Applicable semantic complexity: Supports general scenarios such as everyday spoken language and industrial instructions; for extremely specialized fields, additional domain semantic anchors are required.

[0176] Applicable entity size: It can be adapted to single entities to clusters of 50 entities. For ultra-large-scale clusters, hierarchical alignment is required.

[0177] General migration: The core algorithm can be stripped of its inherent system limitations and migrated to various edge intelligent dialogue terminals.

[0178] The system is configured with a three-tiered handling mechanism, with each tier corresponding to a quantified trigger threshold:

[0179] Level 1 Warning: Context drift exceeds [a certain threshold]. But less than 1.5 times Automatically perform steady-state pruning and restore normal interaction;

[0180] Level 2 warning: If the highest matching degree of semantic resolution is lower than 0.6 for 3 consecutive rounds, switch to confirmation-based interaction to guide the user to clarify the semantics, and revert the context to the previous stable version;

[0181] Level 3 Warning: Contextual drift exceeds 2 times. If the dialogue fails to resolve itself for five consecutive rounds, the current conversation context will be reset, the long-term context layer will be preserved, and the user will be prompted to re-initiate the conversation to ensure that basic interactions are available.

[0182] All dialogue interactions and contextual changes are fully traceable and auditable, meeting the requirements for AI governance and system auditability.

[0183] The core idea of ​​this invention is to "deeply bind the dialogue context with the system's own operating state and use the system's steady state as a rigid constraint on the interaction." It can achieve high-order cognitive isomorphism based on the intrinsic spatiotemporal phase representation system of the self-generated system, or it can be transferred to general edge intelligent dialogue scenarios after stripping away the exclusive concepts. It is a high-order core component of the language capability matrix of intelligent systems.

[0184] It should be noted that any technical solution that incorporates real-time system operating status information into the dialogue history representation in a weighted, spliced ​​or other form and uses it to influence subsequent dialogue interaction decisions should be considered to fall within the protection scope of the "state isomorphism" concept of this invention.

[0185] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-turn dialogue interaction method based on system state, executed by an embedded computing unit of an intelligent system, wherein the intelligent system has a perception-decision-execution closed loop, including steps of receiving language input, parsing semantic features, generating dialogue response, and updating dialogue context, for improving the consistency of multi-turn dialogue context and the stability of long dialogue operation; characterized in that, Includes the following steps: S1. When the dialogue session starts, obtain the real-time operating status data of the system and initialize a three-level hierarchical context cache that is isomorphic to the system operating status. The context cache uses a structured state vector to store the dialogue semantic state set, which is deeply bound to the real-time running state of the system and is used to carry the contextual semantic information of multi-turn dialogue; S2, receive the language input of the current turn, parse it into the corresponding semantic anchor point set, and perform incremental fusion update of the current turn semantic anchor point set and the historical context cache to generate the updated full context representation; the semantic anchor point is a structured semantic feature unit that constitutes the context representation; S3. Generate a sequence of semantic anchor points for the dialogue response based on the updated context representation and map it to natural language output; S4. Write back the input and output semantic features of this round of interaction and the corresponding system state to update the context cache; S5. Perform context steady-state verification after each round of interaction, and monitor two indicators: context drift and cognitive resource occupancy rate; if the context drift exceeds the system steady-state safety threshold, while initiating priority steady-state pruning, simultaneously limit the context update step size and inference propagation depth of the next round until the indicators return to the safe range, ensuring the steady-state operation of the system throughout the multi-round dialogue.

2. The multi-turn dialogue interaction method according to claim 1, characterized in that, The three-level hierarchical context cache includes an instantaneous context layer, a short-term context layer, and a long-term context layer: The instantaneous context layer stores complete semantic anchor information between the current round and the previous round, with the highest retention accuracy, a history retention coefficient of 0.3, and a weight half-life of 1 round; The short-term context layer stores the clustered semantic features of the last 10-20 rounds of dialogue, corresponding to the core topic of the current conversation, with a history retention coefficient of 0.8 and a weight half-life of 5 rounds; The long-term context layer stores stable semantic associations and user interaction preferences across sessions. It has the highest weight update threshold, the longest retention period, a historical retention coefficient of 0.95, and a weight half-life of 50 rounds. The higher the historical retention coefficient, the greater the influence of the historical context on the current context and the smaller the influence of new input, which corresponds to the cognitive pattern from short-term memory to long-term memory.

3. The multi-turn dialogue interaction method according to claim 2, characterized in that, The formula for incremental fusion update in step S2 is: in Let be the context representation vector updated in round t. As a representation of the historical context of the previous round, This is the representation vector of the semantic anchor point set in the current round; The historical retention coefficients for the corresponding context levels are 0.3, 0.8, and 0.95 for the instantaneous, short-term, and long-term levels, respectively.

4. The multi-turn dialogue interaction method according to claim 1, characterized in that, In step S2, the context fusion update is synchronized with the real-time physical state and task state of the system: the current system pose, task progress, and environmental perception features are mapped as semantic space auxiliary anchors, which jointly participate in the context representation update, so that the context cache and the actual system running state remain isomorphic.

5. The multi-turn dialogue interaction method according to claim 1, characterized in that, In step S5, the contextual semantic drift is the semantic distance between the current contextual representation and the initial baseline representation of the conversation in the semantic space; the cognitive resource occupancy rate is the ratio of the number of effective semantic anchors in the current level context cache to the maximum capacity threshold of that level, and each level is calculated independently.

6. The multi-turn dialogue interaction method according to claim 5, characterized in that, The semantic distance is either Euclidean distance or Weil-Petersson geodesic distance.

7. The multi-turn dialogue interaction method according to claim 5, characterized in that, The steady-state pruning employs a priority scoring mechanism, pruning contextual elements sequentially from low to high scores until both indicators return to below the safety threshold or a single pruning removes 20% of the total number of current contextual elements. The priority scoring formula is as follows: in is the normalized association weight of the i-th context element, with a value range of [0,1]. This is a normalized timeliness coefficient, with higher values ​​for more recently occurring elements, ranging from [0,1]. The normalized steady-state criticality coefficient characterizes the indispensability of an element to maintaining the steady state of the system. The more critical the element is to the steady state, the higher its value. The range is [0,1]. The steady-state criticality coefficient is pre-calibrated through offline perturbation statistics and called on demand during online operation. When the context structure undergoes a cumulative and significant change, a background non-real-time update is triggered to recalibrate it. These are weighting coefficients, with values ​​of 0.4, 0.3, and 0.3 respectively.

8. The multi-turn dialogue interaction method according to claim 7, characterized in that, Steady-state pruning retains three types of high-priority elements: core task theme anchors, key instruction anchors explicitly emphasized by the user, and constraint anchors strongly correlated with the system's steady state; the pruning process does not change the relative topological relationships of the remaining elements.

9. The multi-turn dialogue interaction method according to claim 1, characterized in that, After the dialogue session ends, all instantaneous context anchors in this session are weighted and clustered, and the cluster centers are used as the session summary anchors to update the long-term context layer; the cluster weights are calculated by weighting the steady-state strength of each round of the system.

10. The multi-turn dialogue interaction method according to claim 9, characterized in that, The number of clusters is adaptively determined using the silhouette coefficient method: the lower limit of the number of clusters is 2, and the upper limit is 5; the initial number of clusters is set to 2, and the silhouette coefficient of the corresponding cluster results is calculated sequentially. When the coefficient decreases by more than 10%, the increment is stopped, and the previous number of clusters is taken as the final number; if the stopping condition is not triggered when the increment reaches the upper limit of 5, the number of clusters corresponding to the highest silhouette coefficient is taken as the final number.

11. The multi-turn dialogue interaction method according to claim 1, characterized in that, In multi-entity collaborative dialogue scenarios, each entity independently maintains its local context cache, and only uploads common topic context anchors to the joint semantic space for alignment. Private context data is not transmitted across entities. The alignment result is returned to the local entity and becomes effective after steady-state verification.

12. The multi-turn dialogue interaction method according to claim 11, characterized in that, Global alignment employs the orthogonal Procrustes analysis algorithm: First, anchor pairs between subjects are constructed based on semantic similarity matching; anchor pairs with a similarity ≥ 0.85 are considered valid pairs. Then, using the common semantic subspace of the first subject as a benchmark, the orthogonal transformation matrices from the common subspaces of other subjects to the benchmark space are solved sequentially, minimizing the sum of squared semantic distances between corresponding anchor pairs. When the distance decrease in a single iteration is less than... Convergence is determined at a certain time, and the maximum number of iterations is set to 50.

13. The method according to any one of claims 1 to 2, characterized in that, The system steady-state constraint is a system operation safety constraint, and the steady-state impact degree is at least one of the changes in system resource usage and functional safety risk corresponding to the candidate semantics; the changes in system resource usage are the estimated change ratio of CPU / memory usage after the execution of the instruction corresponding to the candidate semantics, and the changes in functional safety risk are the improvement in the safety level of the scenario corresponding to the candidate operation; the steady-state safety threshold is pre-calibrated according to the system hardware configuration and the scenario safety level.

14. The method according to claim 13, characterized in that, The intelligent system is a self-generated system; the context cache is constructed using the system's intrinsic spatiotemporal phase representation system; the system's steady-state constraint is an operational identity constraint; all context updates, semantic resolution, and intent reasoning processes are limited to the causal reachable set defined by the downward causal mechanism; the causal reachable set is a set of legal evolution range boundaries defined from the system's global steady state; when the update or reasoning result goes out of bounds, it is automatically projected along the geodesic to the reachable set boundary, and the projection does not change the semantic relative topological relationship.

15. The method according to claim 14, characterized in that, When the fractal growth mechanism of topological recursive rules is triggered during multi-round dialogue, the newly generated semantic rules are synchronously written into the context cache of the corresponding level and participate in the semantic parsing and response generation of subsequent rounds; the conversation summary is also incorporated into the semantic interaction dimension of cognitive health assessment, and the corresponding abnormal handling reuses the global three-level self-healing mechanism.

16. A context-based pragmatic reasoning and ambiguity resolution method, executed by an embedded computing unit of an intelligent system with context caching capabilities, includes steps of extracting semantic candidates and outputting reasoning conclusions, used to achieve context-dependent semantic disambiguation and implicit intent recognition; characterized in that, Includes the following steps: S1. Extract a set of multiple candidate semantic anchors corresponding to the current language input, with each candidate corresponding to a semantic interpretation; S2. Substitute each candidate semantic anchor into the current context cache and calculate the context matching degree and system steady-state impact degree for each candidate; the context matching degree is the degree of semantic association between the candidate semantics and the current context, and the system steady-state impact degree is the degree of impact on the system's operational stability after the candidate semantics are substituted; S3. Select the candidate with the highest context matching degree and that satisfies the system steady-state constraints as the semantic resolution result and identify its corresponding pragmatic intent; If multiple candidates satisfy the steady-state constraints, the candidate with the smallest steady-state influence is selected; if none of the candidates satisfy the steady-state constraints, the ambiguity confirmation option is output, and semantic resolution is re-executed after receiving user feedback; S4, based on the resolution results and context, implicit semantics and action expectations are derived through association and propagation reasoning, and a complete pragmatic reasoning conclusion is output.

17. The pragmatic reasoning and ambiguity resolution method according to claim 16, characterized in that, The set of multiple candidate semantic anchors corresponds to three types of ambiguous scenarios: lexical ambiguity, syntactic ambiguity, and denotative ambiguity. Lexical ambiguity corresponds to scenarios with multiple meanings of a single word, syntactic ambiguity corresponds to scenarios with uncertain modification relationships, and denotative ambiguity corresponds to scenarios with unclear pronoun referents. Each candidate semantic anchor comes from the language-physical association mapping already established by the system.

18. The pragmatic reasoning and ambiguity resolution method according to claim 16, characterized in that, In step S2, the context matching degree is the average association weight between the candidate semantic anchor and all anchors in the current context; the system steady-state influence degree adopts a two-level calculation mechanism: the first level is a first-order fast estimation: based on the current steady-state strength and the historical steady-state sensitivity coefficient of the semantic anchor, the intensity change after the candidate semantic is substituted is linearly estimated, which is used for regular interaction scenarios; The second level is simulation verification: when the estimated result is less than 10% away from the safety threshold, the candidate semantic anchor point is temporarily mapped to the simulated system instruction, a forward state deduction is performed, and the semantic distance between the deduced state and the steady-state baseline state vector is calculated. This is used as the final steady-state influence degree for accurate determination of critical scenarios.

19. The pragmatic reasoning and ambiguity resolution method according to claim 18, characterized in that, The steady-state sensitivity coefficient is calibrated offline using the perturbation method: under the steady-state operation of the system, a perturbation of ±0.1 is applied to the association weight of a single semantic anchor point, the change in the steady-state strength of the system is observed, and the average value of the change in 20 repeated perturbations is taken as the sensitivity coefficient of the anchor point; during online operation, the coefficient is refitted and calibrated once after every 1000 interactions.

20. The pragmatic reasoning and ambiguity resolution method according to claim 16, characterized in that, The pragmatic intents are divided into four categories: instruction execution, status query, collaboration negotiation, and warning feedback. Each type of intent corresponds to a fixed combination of semantic anchor points, which are identified using cosine similarity threshold matching.

21. The pragmatic reasoning and ambiguity resolution method according to claim 16, characterized in that, In step S4, the implicit semantic derivation is implemented based on relational transitive reasoning: starting from the resolution result, a breadth-first transitive derivation is performed along the semantic relational path within the context to obtain the expected actions, constraints, and potential needs not explicitly mentioned in the input; the derivation confirms the validity of the implicit semantics and relational paths, and writes them into the short-term context layer with an initial weight of 0.1; if in the subsequent three consecutive rounds of interaction, the semantic consistency between the implicit semantics and the user's actual input semantics and the system execution result is ≥0.7 and no user correction is triggered, then the relation is determined to be valid, the weight is increased to 0.5, and the global relational matrix is ​​updated synchronously; if it is not verified, it decays naturally with the context update.

22. A multi-turn dialogue and pragmatic reasoning device based on system state isomorphism, comprising a context storage module, a semantic parsing module, a response generation module, and an input / output interface, used to realize multi-turn dialogue interaction and pragmatic reasoning; characterized in that, Also includes: The context cache management unit is used to obtain the system running status, initialize and maintain the three-level context cache in a hierarchical manner, and synchronize with the real-time status of the system. The language parsing and response generation unit is connected to the context cache management unit and is used for semantic anchor parsing of the input language and natural language mapping of the output response. The context fusion update unit is connected to the context cache management unit and the language parsing and response generation unit, respectively, and is used to perform incremental fusion update of the current round of semantics and historical context; The pragmatic reasoning and ambiguity resolution unit is connected to the context cache management unit and is used for multi-candidate semantic matching, steady-state constraint verification, pragmatic intent recognition and implicit semantic derivation. The context steady-state pruning unit has its input end connected to the system steady-state monitoring interface and the cognitive resource monitoring interface, its control end connected to the context cache management unit, and its feedback control end connected to the control ends of the context fusion update unit and the pragmatic reasoning and ambiguity resolution unit. It is used to monitor context drift and resource occupation, perform priority steady-state pruning, and simultaneously constrain the context update step size and reasoning propagation depth when the context is abnormal. The interactive output unit, connected to the language parsing and response generation unit and the pragmatic reasoning and ambiguity resolution unit, is used to output dialogue responses, reasoning conclusions, and ambiguity confirmation options.

23. The apparatus according to claim 22, characterized in that, The context cache management unit, context fusion update unit, and context steady-state pruning unit are integrated into the FPGA acceleration chip and perform context representation operations, vector weighting, and pruning scheduling operations in parallel.

24. The apparatus according to claim 22, characterized in that, The device incorporates a multi-subject context alignment unit and a conversation summary generation unit; the multi-subject context alignment unit is used to extract common topic context anchors and perform joint semantic space alignment operations, and the conversation summary generation unit is used to perform weighted clustering on instantaneous context anchors to generate conversation summaries.

25. The apparatus according to claim 24, characterized in that, The device has a built-in cognitive health monitoring unit and a rule growth and adaptation unit. The cognitive health monitoring unit is used to collect interaction dimension indicators such as context stability and resolution accuracy and connect to the global three-level self-healing mechanism. The rule growth and adaptation unit is used to synchronously write newly generated semantic rules into the context cache.

26. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the method according to any one of claims 1 to 21.

27. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 21.

28. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 21.

Images