Smart park man-machine conversation intention recognition system based on natural language processing drive

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a smart park man-machine conversation intention recognition system based on natural language processing drive, comprising: step 1, a natural language processing module for receiving and preprocessing natural language input of a park user; 2, the intention recognition module communicates with the natural language processing module, modular design is adopted, preprocessing, intention recognition, RAG retrieval generation, chain reasoning and self-checking, interaction control and other parts are included, and parallel development and maintenance are facilitated; a multi-stage verification and rollback mechanism enables the result to be more reliable and the update to be more controllable; by means of weak / semi-supervised learning, intelligent word segmentation and multi-space embedding, rich semantics can be extracted even if labels are few; the RAG module gives consideration to high correlation and duplicate removal, and can dynamically access the newest laws and regulations and data; event triggering type updating is supported, regular increment updating is also achieved, and continuous self-adaption is achieved through closed-loop feedback; and the timely and stable delivery of the message is ensured by the channel reliability control and the push priority strategy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a wisdom park human-computer dialogue intention recognition system based on natural language processing driving. BACKGROUND

[0002] The current wisdom park human-computer dialogue intention recognition system is very complex, and requires tuning a large number of models and parameters, which is time-consuming and laborious. It relies on multi-step reasoning and graph convolution, and consumes a lot of computing power and memory. In real-time or high-concurrency scenarios, it is prone to delay. The underlying is like a black box and is difficult to explain. It is also highly dependent on data quality and parameter settings, and problems such as overfitting or result fluctuations may occur when there is insufficient data or improper parameter tuning.

[0003] Therefore, the present application provides a wisdom park human-computer dialogue intention recognition system based on natural language processing driving to solve the above-mentioned problems. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a wisdom park human-computer dialogue intention recognition system based on natural language processing driving to solve the problems raised in the background.

[0005] To achieve the above purpose, the present application realizes the following technical scheme: a wisdom park human-computer dialogue intention recognition system based on natural language processing driving, comprising the following steps:

[0006] Step 1: a natural language processing module for receiving and preprocessing natural language input of a park user;

[0007] Step 2: an intention recognition module in communication with the natural language processing module for extracting user intention from the preprocessed text based on a large model and a semantic analysis algorithm;

[0008] Step 3: a retrieval augmentation generation (RAG) module in communication with the intention recognition module for initiating retrieval to a knowledge base according to the user intention and generating a decision support result in combination with a large model;

[0009] Step 4: a chain reasoning and self-verification module in communication with the intention recognition module and the RAG module for splitting complex tasks into multi-step logical reasoning, verifying and correcting the result of each step of reasoning;

[0010] Step 5: an interaction control module in communication with the above modules for driving a robot or a display device to perform a corresponding task and feedback to the user according to the generated decision support result.

[0011] Preferably, the natural language processing module comprises a word segmentation unit, a word vector encoding unit and a context semantic representation unit to improve the context understanding ability of multi-turn dialogue;

[0012] Step 1.1. The word segmentation unit automatically determines the possibility of new words based on the segmentation probability of the Bayesian non-parametric model, considers the sentence as a word sequence, and models the word table using the Pitman-Yor process or Dirichlet process. The formula is as follows:

[0013] Wherein

[0014] n w : the number of times the word w has been observed; t w : the number of clusters; N: the total number of words processed so far; t: the sum of the total number of tables for all words; d: discount parameter, controls the long-tail distribution characteristics, usually 0≤d<1; θ: concentration parameter, controls the probability of new word generation; P base (w): baseline distribution, gives initial probability to unseen or low-frequency words;

[0015] Step 1.1.1. Energy minimization based on graph segmentation, considering the character sequence as a node chain, using a binary segmentation variable to determine whether to segment at a certain position, defining an energy function, and finding the optimal segmentation. The formula is as follows:

[0016]

[0017] λ: smoothing term weight, controls the segmentation consistency preference, avoids excessive segmentation or too little segmentation; T: switching penalty, adds additional cost when the segmentation decision of the two consecutive positions is different, used to encourage reasonable word length distribution; φ(w): intra-word score;

[0018] Step 1.1.2. Dynamic window segmentation based on information entropy, the formula is as follows:

[0019] ΔH(i)=H(c i-k+1:i |c 1:i-k )+H(c i+1:i+k |c i+1+k:N )-H(c 1:N ),

[0020] H(i): conditional entropy or empirical entropy calculated based on statistical corpus, evaluates the degree of freedom of left and right character combinations, k: window size, controls the estimation range of local context entropy; γ: penalty factor, used to avoid excessive segmentation at high entropy but short words.

[0021] Step 1.2. The word vector encoding unit is based on dynamic PPMI+ spectral decomposition smoothing embedding, the formula is as follows:

[0022] First, compute co-occurrence matrix M,

[0023] M ij = PMI(w i ,w j )-log(k),

[0024] Second, truncate the smoothed matrix

[0025]

[0026] Final word vector takes:

[0027]

[0028] Where probability is estimated by corpus statistics; k: smoothing constant, used in ShiftedPPMI to suppress extremely high PMI values; d: truncated dimension, control vector dimension; a: spectral value smoothing index, less direct adjustment, affect vector length distribution;

[0029] Step 1.2.1 double space hyper-sphere embedding, let the vector on the hyper-sphere surface, but capture ambiguity or word class through bidirectional projection: each word has two vectors: main vector and offset vector, combined and normalized to unit sphere Its formula is:

[0030] v w = Normalize(u w + g(p w )), where main vector, offset vector

[0031] g(·) can be a small feedforward network or a gating function,

[0032] g(p w ) = σ(W g p w + b g ) ⊙ tanh(V g p w ),

[0033] u w : main representation, learned from the context prediction task; p w : offset representation, capture specific attributes; W g , V g , b g : gating network parameters; σ: activation function;

[0034] Step 1.2.2 To capture hierarchical or tree-like semantic structure, embed the word into hyperbolic space, which can better express hierarchical or thematic hierarchy, its formula is:

[0035]

[0036] ||u||<1: guarantee inside Poincaré sphere, arcosh: hyperbolic geometry property;

[0037] Step 1.3 Contextual semantic representation unit based on memory-enhanced multi-step self-loop reasoning, maintains explicit dialogue memory vector in multi-round dialogue, updates through multi-step reasoning, cooperates with self-verification mechanism, and its formula is:

[0038]

[0039] L=1…L wherein is the last round final memory, h t is the input representation of this round, f att represents the attention interaction with h t and memory,

[0040] Self-verification: For the generated or inferred result r t , introduce the verification function g check (r t ), if not consistent, trigger re-reasoning: if g check (r t )<τ, then re-adjust the input or memory and re-reasoning the context output: the final context representation can be taken:

[0041]

[0042] h t : Current round user or system dialogue embedding; L-step internal memory state; L: internal reasoning step, which can be dynamic or fixed; f att (·): attention function, dot product attention plus gating; g check (·): verification score network, which can measure the consistency of inference results with historical memory or knowledge base facts; τ: threshold value, used to determine whether to return step correction;

[0043] Step 1.3.1 Construct a graph structure for each round of speech nodes and their semantic entities in multi-round dialogue, propagate contextual information through graph convolution, and its formula is:

[0044]

[0045] neighbor set of node v; d v : node degree, used for normalization; W (l) ,b (l) : the l-th layer trainable parameters;

[0046] Step 1.3.2 In multi-turn dialogue, different turns may contribute differently. Design cross-layer gating to dynamically select context information injection, whose formula is:

[0047] a. Turn embedding: For the t-th turn encoding vector sequence H t = [h t,1 ,...,h t,n ], obtained by the base Encoder;

[0048] b. Context aggregation gating: Each turn has a gating vector where is the pooled representation, c t-1 is the context representation of the previous turn;

[0049] c. Context update: c t = g t ⊙Transf(H t ,c t-1 )+(1-g t )⊙c t-1 where Transf(·) represents the fusion of the current turn input and the previous turn context through cross-attention, followed by a layer of Transformer;

[0050] W g ,b g : gating network parameters for determining the importance of the current turn input to the context; ⊙: element-wise multiplication; Transf(H t ,c t-1 ): can be implemented as multi-head attention, with H t as the query and c t-1 as the key-value; or vice versa, plus a feedforward; c t : updated context summary vector;

[0051] Step 1.3.3 For multi-turn intent recognition, treat intent classification and slot filling as multi-step reasoning, and use the self-validation principle to check consistency at the representation level, whose formula is:

[0052]

[0053] where is the prior based on the possibility of inferring slots from intent, P′ model (s i |c) is the model prediction distribution. If , add a correction gradient to the representation c or do another fusion reasoning:

[0054]

[0055] consistency loss; ∈: threshold value, controls when to trigger self-verification correction; η: correction step size; c: current context representation.

[0056] Preferably, the intention recognition module utilizes a strategy combining weak supervision and semi-supervised learning to train an intention classifier relying on unannotated data and a small amount of annotated data in a scenario where annotated data is scarce.

[0057] Step 2.1 introduces a pseudo-label "sharpening" index α and an unannotated loss weight λ to balance supervised and unsupervised signals in the case of annotation scarcity, and the formula is:

[0058]

[0059] where

[0060] annotated sample set; unannotated sample set; C: total number of intention categories; θ: classifier parameters.

[0061] Preferably, the RAG module includes:

[0062] A document retrieval unit is configured to establish a vector index for the park knowledge base and quickly retrieve relevant documents.

[0063] A generation unit is configured to fuse the retrieved document fragments and user intent using a large model to generate a natural language response.

[0064] Based on step 3.1, a document entropy penalty and retrieval intensity index are introduced to simultaneously consider "highly relevant documents" and "low redundancy documents", and the formula is:

[0065]

[0066] q: user query or dialogue context, d i (i = 1, …, K): the i-th retrieved document fragment.

[0067] Preferably, the chain reasoning and self-verification module includes:

[0068] A logical decomposition unit is configured to decompose the complex task corresponding to the user intent into N-step reasoning subtasks.

[0069] A self-verification unit is configured to perform consistency checking on the reasoning output of each subtask and correct back when inconsistency is detected.

[0070] Based on step 4.1, the logic decomposition generated each step sub-task representation zi (the vector of the ii-step reasoning sub-task) is jointly encoded with its self-verification score vi through an "attention-verification gating" hybrid mechanism, dynamically assigning weights whose formula is:

[0071]

[0072] x: input global context or user intent vector; z i : sub-task representation output by the i-step logic decomposition unit; v i ∈[0,1]: self-verification unit consistency score for the i-step reasoning result; h out : the final representation after fusion, used for subsequent generation or decision-making.

[0073] Preferably, the interaction control module further comprises a wireless connection unit for pushing the generated result to a display screen or a large TV screen in the park through wireless means;

[0074] Based on step 5.1, channel reliability entropy penalty and push priority index are introduced to coordinate the trade-off between latency, bandwidth and energy consumption, whose formula is:

[0075]

[0076] M: the total number of target screens for the current result to be pushed; φ: wireless scheduling strategy parameter set; P tx (i|φ): the probability of initiating a push to the i-th display terminal under strategy φ; u i ∈{0,1}: whether the i-th terminal is online and has subscribed to the push (1 represents online and has subscribed); E tx (i|φ): energy consumption estimate at the i-th terminal according to strategy φ; CQI i : the i-th channel quality indicator distribution, used to measure channel reliability;

[0077] H(CQI i )=-∑ k p(k|CQI i )lnp(k|CQI i ),

[0078] The entropy of this distribution is used to penalize terminals with overly uncertain channel quality; γ, η: energy consumption weight and entropy penalty coefficient, used to balance the push benefits and costs; α: push priority index, used to sharpen the push probability distribution of each terminal, so that "performance is better" or "user importance is higher" screens get higher priority.

[0079] Preferably, further comprising an online adaptive unit for continuously monitoring running data and updating model parameters of each module through weak supervision after system deployment, and a feedback loop unit for collecting user evaluation after task completion and using the evaluation results for subsequent model optimization.

[0080] Step 7.1 introduces adaptive balance coefficient ρt, feedback trust index τ and time decay factor δ to smoothly integrate weak supervision signals of running data and strong supervision signals of user evaluation after system deployment, with the formula being:

[0081]

[0082] θ t : current model parameters; D t : online running data collected in the latest time window; F t : user feedback pairs {(x j , e j )} collected in the latest time window, where e j is the evaluation vector of the user on sample x j ; “strong supervision” pseudo-labels mapped from user evaluation; p(k|x; θ): model prediction probability for class k; s(e): evaluation credibility scoring function.

[0083] Preferably, further comprising a periodic incremental update unit for performing small-batch incremental update on model parameters of each module according to a preset time interval to ensure the timeliness of the model and the park operation environment, and the periodic incremental update unit is further used to calculate the performance index of the model in the latest period before each update, and dynamically adjust the update frequency or batch size according to the performance decline amplitude;

[0084] Step 8.1 introduces the PID controller idea and batch / interval coupling mapping function to control the update cost while ensuring the timeliness of the model, with the formula being:

[0085] e k = P k-1 - P k ,$4pt](2) PID,

[0086]

[0087] Calculate prediction interval: u k = K P e,

[0088] P k : key performance indicator at the end of the kth period; ΔP k: Performance degradation: the performance degradation of this period compared to the last period, used to drive interval and batch changes; K P ,K I ,K D : PID controller's proportional, integral, and derivative coefficients; M: integral window length, used to accumulate the last M errors; u k : control output calculated by PID, used to predict the adjustment of the next update interval; κ: mapping bias term, determines which interval triggers significant batch growth.

[0089] Preferably, it also includes an event-triggered update unit for detecting major events occurring in the park and triggering incremental updates in real time to quickly adapt to sudden changes. The online adaptive unit works with the event-triggered update unit to learn the performance of the model under different scenarios by comparing historical events and performance data, and gradually improve the update strategy

[0090] Step 9.1 introduces the concepts of event severity, historical similarity suppression, trigger weight, and collaborative update loss, which can trigger updates in real time when major events occur in the park, while dynamically optimizing the update strategy by combining historical events and online performance data. Its formula is:

[0091]

[0092] S t =f sev (E t ),

[0093] E t : the detected event feature at time t; the i-th labeled major event in the historical event library; f sev (·): event severity evaluation function, which can combine the impact range, duration, and associated indicator mutation amplitude; κ(E, H) ∈ [0, 1]: event similarity measure, the larger the value, the more similar; S t : quantified event severity score, reflecting the "scale" and "influence" of the current event; N t : event novelty, the higher the value, the greater the difference between the event and historical events; β, γ: event severity and novelty weight coefficients, used to adjust their respective contributions to triggering; τ: trigger threshold bias, controls what level of "combined score" triggers updates significantly; w t ∈ (0, 1): event trigger weight, when w t ≈ 1, emphasizes event-driven updates; The weakly supervised loss of the online adaptive unit is based on the latest running data D t ; The strong supervision or special fine-tuning loss of the event triggered updating unit can be based on artificial labeling or expert rules on the event sample E t Computing.

[0094] Preferably, the updating rollback unit is further included for automatically judging and rolling back to the previous version model if necessary after the incremental updating by comparing the key indicators before and after the updating, so as to ensure the system stability.

[0095] Step 10.1 introduces the concepts of event severity, historical similarity suppression, trigger weight and collaborative updating loss, can trigger updating when a major event occurs in the park, and dynamically optimize the updating strategy by combining historical events and online performance data, and the formula is:

[0096] Delta P = P new -P old D KL = D KL p,

[0097] P old P new The key performance indicators of the pre-rollback / rollback model on the validation set or online monitoring data; Delta P performance difference: when Delta P < 0, the performance decreases after updating, and rollback needs to be considered; D KL KL divergence: measure the degree of drift of the model output distribution before and after updating, used to detect potential instability caused by "distribution shift"; w P Performance penalty weight: control the influence degree of performance decline on rollback decision; w D Drift penalty weight: control the influence degree of distribution drift on rollback decision; epsilon rollback threshold: when the rollback loss L rb exceeds the threshold, forced rollback; sigma (·) Sigmoid function: used to map the rollback loss to probability, supporting the "soft rollback" strategy.

[0098] The application provides a wisdom park man-machine dialogue intention recognition system based on natural language processing driving. Has the following beneficial effects:

[0099] 1、The application adopts modular design and is divided into preprocessing, intention recognition, RAG retrieval generation, chain reasoning and self-checking, interactive control and other parts, which is convenient for parallel development and maintenance; Multi-stage verification and rollback mechanism makes the result more reliable and the update more controllable; With weak / semi-supervised learning, intelligent word segmentation and multi-space embedding, even if the annotation is less, rich semantics can be extracted; The RAG module considers high correlation and deduplication, and can dynamically access the latest regulations and data; It supports event triggered updating and regular incremental updating, and realizes continuous self-adaptation through closed loop feedback; Channel reliability control and push priority strategy ensure that messages are delivered in time and stably.

[0100] 2、The application can save resources while ensuring the timeliness of the model by using a periodic incremental updating unit combined with performance monitoring and PID control; the event-triggered updating unit can quickly respond to sudden major events, and fuse historical similarity and online data optimization strategies; the rollback unit automatically or softly reverts through performance and distribution drift metrics, ensuring stable and reliable updates; the three work together to form a closed-loop feedback, achieving continuous self-adaptation and high robustness of the model, while improving controllability and explainability through quantitative indicators. BRIEF DESCRIPTION OF DRAWINGS

[0101] Figure 1 The flowchart of the application. DETAILED DESCRIPTION

[0102] To enable those skilled in the art to understand the application scheme, the technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings accompanying the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, other embodiments obtained by those skilled in the art without creative labor should be within the scope of the application.

[0103] The application will be described in detail below with reference to the drawings:

[0104] Embodiment:

[0105] Please refer to the accompanying drawings Figure 1 The embodiment of the application provides a wisdom park man-machine dialogue intention recognition system based on natural language processing driving, which comprises:

[0106] Step 1: a natural language processing module, used for receiving and preprocessing natural language input of a park user;

[0107] Step 2: an intention recognition module, in communication with the natural language processing module, used for extracting user intention from the preprocessed text based on a large model and a semantic analysis algorithm;

[0108] Step 3: a retrieval augmented generation (RAG) module, in communication with the intention recognition module, used for initiating retrieval to a knowledge base according to the user intention, and generating a decision support result in combination with a large model;

[0109] Step 4: a chain reasoning and self-verification module, in communication with the intention recognition module and the RAG module, used for splitting a complex task into multiple-step logical reasoning, verifying and correcting each step of reasoning result;

[0110] Step 5: an interaction control module, in communication with the above modules, used for driving a robot or a display device to perform a corresponding task and feed back to the user according to the generated decision support result.

[0111] Specifically, from text preprocessing, intent recognition, to knowledge retrieval generation, chain reasoning self-checking, and then interactive execution, it is easy to locate and maintain, and can be upgraded in parallel; Large models + RAG combine context and the latest knowledge in real time, making decisions more intelligent; Chain multi-step verification effectively prevents error accumulation, making output more reliable; Standardized interfaces make performance scalable and algorithms replaceable; Interactive control drives feedback end-to-end, ensuring smooth real-time dialogue and operation with users.

[0112] The natural language processing module includes a word segmentation unit, a word vector encoding unit, and a context semantic representation unit to improve the context understanding ability for multi-round dialogue;

[0113] Step 1.1. The word segmentation unit is based on the segmentation probability of the Bayesian non-parametric model, considering sentences as word sequences, modeling the word table using Pitman-Yor process or Dirichlet process, and automatically determining the possibility of new words, with the formula:

[0114] where

[0115] n w : the number of times the word w has been observed; t w : the number of clusters; N: the total number of words processed so far; t: the sum of the total number of tables for all words; d: the discount parameter, which controls the long-tail distribution characteristics, usually 0 ≤ d < 1; θ: the concentration parameter, which controls the probability of new word generation; P base (w): the base distribution, which gives the initial probability for unseen or low-frequency words;

[0116] Step 1.1.1. Energy minimization based on graph segmentation, considering character sequences as node chains, using binary segmentation variables to determine whether to segment at a certain position, defining an energy function, and finding the optimal segmentation, with the formula:

[0117]

[0118] λ: smoothing term weight, controlling segmentation consistency preference, avoiding excessive segmentation or insufficient segmentation; T: switching penalty, adding additional cost when the segmentation decision of consecutive two positions is different, used to encourage reasonable word length distribution; φ(w): intra-word score;

[0119] Step 1.1.2. Dynamic window segmentation based on information entropy, with the formula:

[0120] ΔH(i) = H(c i-k+1:i |c 1:i-k ) + H(c i+1:i+k |c i+1+k:N ) - H(c 1:N ),

[0121] H(i): conditional entropy or empirical entropy computed based on statistical corpus, evaluating the freedom of left and right character combination, k: window size, controlling the range of local context entropy estimation; γ: penalty factor, used to avoid excessive segmentation at high entropy but short words.

[0122] Step 1.2 Word vector encoding unit based on dynamic PPMI + spectral decomposition smoothing embedding, its formula is:

[0123] First, calculate the co-occurrence matrix M,

[0124] M ij = PMI(w i , w j ) - log(k),

[0125] Second, the truncated

[0126]

[0127] Finally, the word vector takes:

[0128]

[0129] Where the probability is estimated by corpus statistics; k: smoothing constant, used in ShiftedPPMI to suppress extremely high PMI values; d: truncated dimension, control vector dimension; α: spectral value smoothing index, less direct adjustment, affect vector length distribution;

[0130] Step 1.2.1 Double space hyper-spherical embedding, let the vector be on the hyper-spherical surface, but capture ambiguity or part of speech through bidirectional projection: each word has two vectors: main vector and offset vector, combined and normalized to unit sphere Its formula is:

[0131] v w = Normalize(u w + g(p w )), where Main vector, Offset vector,

[0132] g(·) can be a small feedforward network or a gating function,

[0133] g(p w ) = σ(W g p w + b g ) ⊙ tanh(V g p w ),

[0134] u w : main representation, learned from the context prediction task; pw : offset represents capturing specific attributes; W g ,V g ,b g : gating network parameters; σ: activation function;

[0135] Step 1.2.2 is a hierarchical or tree semantic structure, which puts word embedding points in hyperbolic space, which can better express hierarchical or thematic levels, and its formula is:

[0136]

[0137] ||u||<1: guarantee inside the Poincaré ball, arcosh: hyperbolic geometry characteristics;

[0138] Step 1.3 context semantic representation unit based on memory-enhanced multi-step self-loop reasoning, in multi-round dialogue, maintain explicit dialogue memory vector, update through multi-step reasoning, cooperate with self-verification mechanism, its formula is:

[0139]

[0140] L = 1…L where is the last round of final memory, h t is the input representation of this round, f att represents attention interaction with memory h t ,

[0141] Self-verification: for the generated or reasoning result r t , introduce the verification function g check (r t ), if not consistent, trigger re-reasoning: if g check (r t )<τ, then adjust the input or memory and re-reasoning the context output: the final context representation can be taken:

[0142]

[0143] h t : current round user or system dialogue embedding; Step L internal memory state; L: internal reasoning steps, which can be dynamic or fixed; f att (·): attention function, dot product attention plus gating; g check (·): verification score network, which can measure the consistency of reasoning results and historical memory or knowledge base facts; τ: threshold, used to judge whether it needs to return step correction;

[0144] Step 1.3.1 constructs a graph structure for each round of speech nodes and their semantic entities in multi-round dialogue, and propagates context information through graph convolution, and its formula is:

[0145]

[0146] Neighbour set of node v; d v : Node degree, for normalization; W (l) ,b (l) : The l-th layer trainable parameter;

[0147] Step 1.3.2 In multi-turn dialogue, different turns may contribute differently. We design a cross-layer gating mechanism to dynamically select the context information injection, whose formula is:

[0148] a. Turn embedding: For the sequence of turn encoding vectors H t = [h t,1 ,...,h t,n ] obtained by the base Encoder;

[0149] b. Context aggregation gating: Each turn has a gating vector where is the pooled representation, and c t-1 is the context representation of the previous turn;

[0150] c. Context update: c t = g t ⊙Transf(H t ,c t-1 ) + (1-g t )⊙c t-1 where Transf(·) represents the fusion of the current turn input and the previous turn context through cross-attention, followed by a layer of Transformer;

[0151] W g ,b g : Gating network parameters for determining the importance of the current turn input to the context; ⊙: Element-wise multiplication; Transf(H t ,c t-1 ): Can be implemented as multi-head attention, with H t as the query and c t-1 as the key-value; or vice versa, plus a feedforward; c t : Updated context summary vector;

[0152] Step 1.3.3 For multi-turn intent recognition, we treat intent classification and slot filling as multi-step reasoning, and use the self-validation principle to check consistency at the representation level, whose formula is:

[0153]

[0154] where is the prior based on the possibility of inferring slots from intent, P′model (s i |c) is a model predicted distribution, if then add a correction gradient on c or do a fusion inference again:

[0155]

[0156] Consistency loss; epsilon: threshold value, control when to trigger self-verification correction; eta: correction step; c: current context representation.

[0157] Specifically, new words are automatically discovered, polysemy and hierarchical relationship are considered, robust and rich word vectors are generated, dialogue context is maintained, inconsistent results are automatically corrected, multi-round interaction is ensured to be continuous and reliable, convolution and cross-layer gating accurately capture the dialogue context, the self-verification mechanism timely corrects the deviation, reduces the accumulation of errors, and overall, the system can quickly adapt, stably run in complex and variable scenarios, and provide high-quality semantic understanding and decision support.

[0158] The intent recognition module utilizes a strategy combining weak supervision and semi-supervised learning, and in the case of insufficient labeled data, relies on unlabeled data and a small amount of labeled data to train an intent classifier.

[0159] Step 2.1 introduces a pseudo-label "sharpening" index alpha and an unlabeled loss weight lambda to balance supervised and unsupervised signals in the case of insufficient labeling, and the formula is:

[0160]

[0161] Wherein

[0162] A set of labeled samples; A set of unlabeled samples; C: total number of intent categories; theta: classifier parameters.

[0163] Specifically, when the labeled data is scarce, combining a small amount of manual labeling with a large amount of unlabeled data can significantly reduce the dependence on high-cost labeling, utilize the potential information of unlabeled data to enhance the model's ability to recognize new intents or long-tail intents, and weak supervision signals and semi-supervised self-training complement each other, which can effectively resist the influence of noise labels or small data bias. Without large-scale new labeling, it can speed up model updating and optimization speed through continuous introduction of unlabeled samples for incremental training. Based on a small amount of labeled data, with the help of a large amount of unlabeled data, the intent classifier is trained through weak supervision + semi-supervised learning, realizing low-cost, high generalization, and strong robustness of intent recognition.

[0164] The RAG module includes:

[0165] The document retrieval unit is configured to establish a vector index for the park knowledge base and quickly retrieve relevant documents;

[0166] The generation unit is configured to fuse the retrieved document segments with the user intent using a large model to generate a natural language response.

[0167] Based on step 3.1, a document entropy penalty and a retrieval intensity index are introduced to simultaneously consider "highly relevant documents" and "low redundancy documents", and the formula is:

[0168]

[0169] q: user query or dialogue context, d i (i = 1, …, K): the i-th retrieved document segment.

[0170] Specifically, the document retrieval unit can quickly locate highly relevant knowledge based on the vector index; the generation unit deeply fuses these segments with the user intent to output coherent and natural responses; the introduced document entropy penalty can effectively filter information redundancy, and the retrieval intensity index dynamically balances the breadth and accuracy of retrieval, ensuring that key content is captured while avoiding repetition, thereby achieving precise, efficient, and concise knowledge-enhanced generation.

[0171] The chain reasoning and self-validation module includes:

[0172] The logical decomposition unit is configured to decompose the complex task corresponding to the user intent into N-step reasoning subtasks.

[0173] The self-validation unit is configured to check the consistency of the reasoning output of each subtask and correct it when inconsistency is detected.

[0174] Based on step 4.1, the "attention-validation gate" hybrid mechanism is used to jointly encode the subtask representation zi (vector of the ii-th step reasoning subtask) generated by logical decomposition and its self-validation score vi, and dynamically allocate weights, and the formula is:

[0175]

[0176] x: input global context or user intent vector; z i : subtask representation output by the i-th logical decomposition unit; v i ∈ [0, 1]: self-validation unit consistency score for the i-th reasoning result; h out : the final representation after fusion, used for subsequent generation or decision-making.

[0177] Specifically, the module decouples reasoning and fast error correction at a controllable granularity by decomposing complex intent into multiple-step sub-tasks and embedding consistency check and backtracking correction at each step, ensuring high reliability and explainability of the output.

[0178] The interaction control module further comprises a wireless connection unit for pushing the generated result to a display screen or a large TV screen in the park area through wireless means.

[0179] Based on step 5.1, a channel reliability entropy penalty and a push priority index are introduced to coordinate the trade-off between latency, bandwidth and energy consumption, and the formula is:

[0180]

[0181] M: the total number of target screens of the current result to be pushed; φ: a set of wireless scheduling strategy parameters; P tx (i|φ): the probability of initiating a push to the i-th display terminal under strategy φ; u i ∈{0,1}: whether the i-th terminal is online and subscribes to the push (1 represents online and subscribed); E tx (i|φ): the energy consumption estimate of the i-th terminal according to strategy φ; CQI i : the i-th channel quality indicator distribution, used to measure channel reliability;

[0182] H(CQI i )=-∑ k p(k|CQI i )lnp(k|CQI i ),

[0183] The entropy of this distribution is used to penalize terminals with overly uncertain channel quality; γ,η: energy consumption weight and entropy penalty coefficient, used to balance the push benefits and costs; α: push priority index, used to sharpen the push probability distribution of each terminal, so that screens with "better performance" or "higher user importance" have higher priority.

[0184] Specifically, after introducing the wireless connection unit, the interaction control module can flexibly and in real time push the generated result to various display terminals (such as large screens or TVs) in the park area, reducing wiring and maintenance costs, supporting rapid expansion and multi-device compatibility, and ensuring user experience and data security through low-latency transmission and security mechanisms.

[0185] Further comprising an online adaptive unit for continuously monitoring running data and updating module model parameters through weak supervision after system deployment, and a feedback loop unit for collecting user evaluations after task completion and using the evaluation results for subsequent model optimization.

[0186] Step 7.1 introduces adaptive balance coefficient ρt, feedback trust index τ and time decay factor δ to smoothly fuse weakly supervised signals from running data and strongly supervised signals from user ratings after system deployment, with the formula:

[0187]

[0188] θ t : current model parameters; D t : online running data collected in the recent time window; F t : user feedback pairs {(x j , e j )} collected in the recent time window, where e j is the evaluation vector of user on sample x j ; "strong supervision" pseudo-labels mapped from user ratings; p(k|x; θ): model's predicted probability of class k; s(e): rating credibility scoring function.

[0189] Specifically, after introducing the model online adaptive unit and feedback closed-loop unit, the system can continuously monitor running data and dynamically update module parameters through weakly supervised signals after deployment, while collecting user rating feedback for subsequent optimization after task completion, thereby quickly responding to environmental and demand changes without the need for large amounts of manual annotation, reducing maintenance costs, and continuously iterating to improve system performance and user experience.

[0190] It also includes a periodic incremental update unit for performing small batch incremental updates on module model parameters at preset time intervals to ensure the timeliness of the model and the park operating environment. The periodic incremental update unit is further configured to calculate the performance indicators of the model in the last period before each update, and dynamically adjust the update frequency or batch size according to the performance decline amplitude;

[0191] Step 8.1 introduces the PID controller idea and batch / interval coupling mapping function to control the update cost while ensuring the timeliness of the model, with the formula:

[0192] e k = P k-1 - P k , $4pt](2) PID,

[0193]

[0194] Calculate the prediction interval: u k = K P e,

[0195] P k : key performance indicator at the end of the kth period; ΔPk : Performance degradation: The performance degradation of this cycle compared to the previous cycle, used to drive interval and batch changes; K P ,K I ,K D : Proportional, integral, and differential coefficients of the PID controller; M: Integral window length, used to accumulate the most recent M errors; u k : The control output calculated by PID is used to predict the adjustment of the next update interval; κ: The mapping bias term that determines which interval triggers significant batch growth.

[0196] Specifically, the periodic incremental update unit automatically performs small-batch parameter updates through predetermined time intervals and performance monitoring, and dynamically adjusts the update frequency and batch size, ensuring that the model is synchronized with the park environment and achieving the best balance between efficiency and resources.

[0197] It also includes an event-triggered update unit for detecting major events occurring in the park and instantly triggering incremental updates to quickly adapt to sudden changes. The model online adaptive unit works in conjunction with the event-triggered update unit to conduct online learning of the model's performance in different scenarios by comparing historical events and performance data, and gradually improve the update strategy.

[0198] Step 9.1 introduces concepts such as event severity, historical similarity suppression, trigger weight, and collaborative update loss. This allows for instant triggering of updates when a major incident occurs in the park. It also dynamically optimizes the update strategy by combining historical events and online performance data. The formula is:

[0199]

[0200] S t =f sev (E t ),

[0201] E t : The characteristics of the emergency event detected at time t; The i-th marked major event in the historical event database; f sev (·): event severity evaluation function, which can be combined with the impact range, duration, and mutation amplitude of related indicators; κ(E,H)∈[0,1]: event similarity measurement, the larger the value, the more similar it is; S t : The quantified severity score of the event, reflecting the "scale" and "influence" of the current event; N t : event novelty, the higher the value, the greater the difference between the event and historical events; β, γ: event severity and novelty weight coefficients, used to adjust their respective contributions to triggering; τ: trigger threshold bias, which controls the degree of "combined score" to significantly trigger the update; w t∈(0,1): event trigger weight, when w t ≈1, emphasize event-driven update; Weakly supervised loss of online adaptive unit, based on recent running data D t ; Strongly supervised or specialized fine-tuning loss of event trigger update unit, calculated based on artificially labeled or expert rule event samples E t .

[0202] Specifically, the event trigger update unit cooperates with the online adaptive unit to immediately start incremental update and online learning by real-time detection of major events and comparison with historical performance data, constantly improving the update strategy to achieve rapid response to sudden changes and long-term robust optimization.

[0203] It also includes an update rollback unit to automatically determine and roll back to the previous version model if necessary to ensure system stability after incremental update by comparing key indicators before and after the update;

[0204] Step 10.1 introduces the concepts of event severity, historical similarity suppression, trigger weight, and collaborative update loss, which can trigger updates immediately when a major event occurs in the park, while dynamically optimizing the update strategy by combining historical events and online performance data. Its formula is:

[0205] ΔP=P new -P old ,$4pt]D KL =D KL [p,

[0206] P old ,P new Key performance indicators of pre-rollback / post-rollback models on validation set or online monitoring data; ΔP performance difference: when ΔP < 0, the performance decreases after the update and needs to be considered for rollback; D KL KL divergence: measures the degree of drift in the output distribution of the model before and after the update, used to detect potential instability caused by "distribution shift"; w P Performance penalty weight: controls the impact of performance decline on rollback decision; w D Drift penalty weight: controls the impact of distribution drift on rollback decision; ∈ rollback threshold: when the rollback loss L rb exceeds this threshold, it is forced to roll back; σ(·) Sigmoid function: used to map the rollback loss to a probability, supporting the "soft rollback" strategy.

[0207] Specifically, the update rollback unit automatically compares the key indicators before and after the incremental update, and if it finds performance degradation or abnormalities, it immediately rolls back to the previous stable version, achieving seamless switching, ensuring system stability and reliability, and reducing operational risks.

[0208] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A smart park human-computer dialogue intention recognition system based on natural language processing, characterized by: include: Step 1: Natural language processing module, used to receive and pre-process natural language input from park users; Step 2: an intent recognition module, communicating with the natural language processing module, for extracting user intent from the preprocessed text based on a large model and a semantic analysis algorithm; Step 3: The retrieval enhancement generation RAG module communicates with the intention recognition module to initiate a search in the knowledge base according to the user intention and generate a decision support result in combination with the large model; Step 4: Chain reasoning and self-verification module, which communicates with the intention recognition module and the RAG module, is used to split complex tasks into multiple steps of logical reasoning and verify and correct the results of each step of reasoning; Step 5: The interactive control module communicates with the above modules and is used to drive the robot or display device to perform corresponding tasks and provide feedback to the user based on the generated decision support results.

2. The natural language processing-driven smart park human-computer dialogue intention recognition system according to claim 1 is characterized in that: The natural language processing module includes a word segmentation unit, a word vector encoding unit, and a contextual semantic representation unit to improve the context understanding ability of multiple rounds of dialogue; Step 1.1: The word segmentation unit considers the sentence as a word sequence based on the segmentation probability of the Bayesian non-parametric model. The Pitman–Yor process or Dirichlet process is used to model the vocabulary and automatically determine the probability of new words. The formula is: in n w : The number of occurrences of word w that has been observed so far; t w : number of clusters; N: total number of words currently processed; t: sum of the total number of tables for all words; d: discount parameter, controlling the long-tail distribution characteristics, usually 0≤d<1; θ: concentration parameter, controlling the probability of new word generation; P base (w): baseline distribution, giving initial probabilities for unseen or low-frequency words; Step 1.1.1 is based on minimizing the energy of graph segmentation. Consider the character sequence as a node chain and use a binary segmentation variable to determine whether to segment at a certain position. Define the energy function and find the optimal segmentation. The formula is: λ: Smoothing term weight, controlling the segment consistency preference to avoid over-segmentation or under-segmentation; T: Switching penalty, when the segmentation decisions of two consecutive positions are different, an additional penalty is added to encourage a reasonable word length distribution; φ(w): Intra-word score; Step 1.1.2 is based on information entropy dynamic window segmentation, and its formula is: ΔH(i)=H(c i-k+1:i ∣c 1:i-k )+H(c i+1:i+k ∣c i+1+k:N )-H(c 1:N ), H(i): conditional entropy or empirical entropy calculated based on statistical corpus, which evaluates the degrees of freedom of left and right character combinations; k: window size, which controls the range of local context entropy estimation; γ: penalty factor, which is used to avoid excessive segmentation in high-entropy but short words. Step 1.2 The word vector encoding unit is based on the smooth embedding of dynamic PPMI+spectral decomposition, and its formula is: First, calculate the co-occurrence matrix M, M ij =PMI(w i ,w j )-log(k), Secondly, truncate the smoothed matrix The final word vector is: The probability is estimated by corpus statistics; k is a smoothing constant, which is used to suppress extremely high PMI values ​​in Shifted PPMI; d is the truncation dimension, which controls the vector dimension; α is the spectral value smoothing exponent, which is less directly adjusted and affects the vector length distribution; Step 1.2.1 Dual-space hypersphere embedding, let the vector be on the hypersphere, but capture polysemy or part of speech through bidirectional projection: each word has two vectors: the main vector and the offset vector, which are combined and normalized to the unit sphere. The formula is: v w =Normalize(u w +g(p w )),in principal vector, Offset quantity, g(·) can be a small feedforward network or a gating function, g(p w )6σ(W g p w +b g )⊙tanh(V g p w ), u w : Main representation, learning from context prediction task; p w : offset representation, capturing specific attributes; W g ,V g ,b g : gating network parameters; σ: activation function; Step 1.2.2 is to capture the hierarchical or tree-like semantic structure by placing the word embedding points in hyperbolic space, which can better express the level or topic hierarchy. The formula is: ||u||<1: guaranteed to be inside the Poincaré sphere, arcosh: hyperbolic geometric properties; Step 1.3 The contextual semantic representation unit is based on memory-enhanced multi-step self-loop reasoning. In multiple rounds of dialogue, it maintains an explicit dialogue memory vector and updates it through multi-step reasoning, combined with a self-verification mechanism. The formula is: L=1…Lwhere is the final memory of the previous round, h t is the input representation of this round, f att Indicates h t Interacting with memory for attention, Self-verification: for the generated or inference results r t , introduce the test function g check (r t ), if inconsistent, trigger further reasoning: if g check (r t )<τ, then re-adjust the input or memory and re-infer the context output: the final context representation is desirable: h t : The current round of user or system dialogue embedding; The internal memory state of the Lth step; L: the number of internal reasoning steps, which can be dynamic or fixed; f att (·): attention function, which is the dot product attention plus gating; g check (·): verification score network, which can measure the consistency of reasoning results with historical memory or knowledge base facts; τ: threshold, used to determine whether backtracking correction is needed; Step 1.3.1 constructs a graph structure of the speech nodes and their semantic entities in each round of multi-round dialogue, and propagates context information through graph convolution. The formula is: The neighbor set of node v; d v : Node degree, used for normalization; W (l) ,b (l) : trainable parameters of the lth layer; Step 1.3.2: In a multi-round dialogue, different rounds may contribute differently. We design cross-layer gating and dynamically select context information injection. The formula is: a. Round embedding: For the t-th round encoding vector sequence H t =[h t,1 ,...,h t,n ], obtained by the basic Encoder; b. Context aggregation gating: Each round has a gating vector in is the pooling representation, c t-1 Represents the context of the previous round; c.Context update: c t =g t ⊙Transf(H t ,c t-1 )+(1-g t )⊙c t-1 Transf(·) represents the cross-attention fusion of the current round input and the previous round context and then passes it through a Transformer layer; W g ,b g : Gating network parameters, used to determine the importance of the current round input to the context; ⊙: element-by-element multiplication; Transf(H t ,c t-1 ): can be implemented as multi-head attention: the query is H t , the key value is c t-1 ; or vice versa, add feedforward; c t : updated context summary vector; Step 1.3.3: For multi-round intent recognition, treat the subtasks of intent classification and slot filling as multi-step reasoning. Use the self-verification principle to perform consistency checks at the representation level. The formula is: in It is the priori of slot possibility based on intention inference, P′ model (s i |c) is the model prediction distribution, if Then add the correction gradient to the representation c or do another fusion inference: -consistency loss; ∈: threshold, controlling when to trigger self-verification correction; η: correction step size; c: current context representation.

3. The natural language processing-driven smart park human-computer dialogue intention recognition system according to claim 1 is characterized in that: The intent recognition module uses a strategy that combines weak supervision and semi-supervised learning to train the intent classifier based on unlabeled data and a small amount of labeled data in scenarios where labeled data is scarce. Step 2.1 introduces the pseudo-label "sharpening" exponent α and the unlabeled loss weight λ to balance the supervised and unsupervised signals when labels are scarce. The formula is: in There is a set of labeled samples; Unlabeled sample set; C: total number of intent categories; θ: classifier parameters.

4. The natural language processing-driven smart park human-computer dialogue intention recognition system according to claim 1 is characterized in that: The RAG module includes: Document retrieval unit, used to create vector indexes for the park knowledge base and quickly retrieve relevant documents; The generation unit is used to use the large model to fuse the retrieved document fragments with the user's intention to generate a natural language response. Based on step 3.1, the document entropy penalty and retrieval strength index are introduced to take into account both "high relevance documents" and "low redundancy documents". The formula is: q: user query or conversation context, d i (i=1,…,K): The i-th document fragment retrieved.

5. The natural language processing-driven smart park human-computer dialogue intention recognition system according to claim 1 is characterized in that: The chain reasoning and self-verification module includes: A logic decomposition unit, configured to decompose the complex task corresponding to the user intention into N-step reasoning subtasks; A self-verification unit is used to check the consistency of the reasoning output of each subtask and to correct it if an inconsistency is detected; Based on step 4.1, the "attention-verification gating" hybrid mechanism is used to jointly encode the sub-task representations zi (the vector of the reasoning sub-task in step ii) generated by the logical decomposition and their self-verification scores vi, and dynamically assign weights. The formula is: x: input global context or user intent vector; z i : The subtask representation output by the logical decomposition unit in step i; v i ∈[0,1]: the self-verification unit scores the consistency of the i-th step reasoning result; h out : The final representation after fusion is used for subsequent generation or decision making.

6. The natural language processing-driven smart park human-computer dialogue intention recognition system according to claim 1 is characterized in that: The interactive control module further includes a wireless connection unit for wirelessly pushing the generated results to a display screen or a large TV screen in the park; Based on step 5.1, the channel reliability entropy penalty and push priority index are introduced to coordinate the trade-off between latency, bandwidth and energy consumption. The formula is: M: the total number of target screens to which the results are currently being pushed; φ: wireless scheduling strategy parameter set; P tx (i|φ): the probability of initiating a push to the i-th display terminal under strategy φ; u i ∈{0,1}: whether the i-th terminal is online and subscribed to the push (1 means online and subscribed); E tx (i|φ): Energy consumption estimation at the i-th terminal according to strategy φ; CQI i : The i-th channel quality indicator distribution is used to measure channel reliability; H(CQI i )=-∑ k p(k∣CQI i )lnp(k∣CQI i ), The entropy of this distribution is used to penalize terminals with overly uncertain channel quality; γ, η: energy consumption weight and entropy penalty coefficient, used to balance push benefits and costs; α: push priority index, used to sharpen the push probability distribution of each terminal, giving higher priority to screens with "better performance" or "higher user importance".

7. The natural language processing-driven smart park human-computer dialogue intention recognition system according to claim 1 is characterized in that: It further includes a model online adaptive unit for continuously monitoring the operation data after the system is deployed and updating the model parameters of each module through weak supervision, and a feedback closed-loop unit for collecting user evaluations after the task is completed and using the evaluation results for subsequent model optimization; Step 7.1 introduces the adaptive balancing coefficient ρt, the feedback trust index τ, and the time decay factor δ to smoothly fuse the weak supervision signal of the running data and the strong supervision signal of the user evaluation after the system is deployed. The formula is: θ t is the current model parameter; D t : Online operation data collected in the most recent time window; F t : User feedback collected in the most recent time window {(x j ,e j )}, where e j For user sample x j The evaluation vector of "Strongly supervised" pseudo-labels obtained by mapping user evaluations; p(k|x;θ): the model's predicted probability for category k; s(e): the evaluation credibility scoring function.

8. The natural language processing-driven smart park human-computer dialogue intention recognition system according to claim 1 is characterized in that: It also includes a periodic incremental update unit for performing small batch incremental updates on the model parameters of each module at preset time intervals to ensure the timeliness of the model and the park operating environment. The periodic incremental update unit is further used to calculate the performance indicators of the model in the most recent cycle before each update, and dynamically adjust the update frequency or batch size according to the degree of performance degradation; Step 8.1 introduces the PID controller concept and batch / interval coupling mapping function to control the update cost while ensuring the timeliness of the model. The formula is: e k =P k-1 -P k ,$4pt](2)PID, Calculate the prediction interval: u k =K P e, P k : Key performance indicator at the end of the kth cycle; ΔP k : Performance degradation: The performance degradation of this cycle compared to the previous cycle, used to drive interval and batch changes; K P ,K I ,K D : Proportional, integral, and differential coefficients of the PID controller; M: Integral window length, used to accumulate the most recent M errors; u k : The control output calculated by PID is used to predict the adjustment of the next update interval; κ: The mapping bias term that determines which interval triggers significant batch growth.

9. The natural language processing-driven smart park human-computer dialogue intention recognition system according to claim 1 is characterized in that: It also includes an event-triggered update unit, which is used to detect major events occurring in the park and instantly trigger incremental updates to quickly adapt to sudden changes. The model online adaptive unit works in conjunction with the event-triggered update unit to conduct online learning of the model's performance in different scenarios by comparing historical events and performance data, and gradually improve the update strategy. Step 9.1 introduces concepts such as event severity, historical similarity suppression, trigger weight, and collaborative update loss. This allows for instant triggering of updates when a major incident occurs in the park. It also dynamically optimizes the update strategy by combining historical events and online performance data. The formula is: S t =f sev (E t ), E t : The characteristics of the emergency event detected at time t; The i-th annotated major event in the historical event database; f sev (·): event severity evaluation function, which can be combined with the impact range, duration, and mutation amplitude of related indicators; κ(E,H)∈[0,1]: event similarity measurement, the larger the value, the more similar it is; S t : The quantified severity score of the event reflects the "scale" and "influence" of the current event; N t : event novelty, the higher the value, the greater the difference between the event and historical events; β, γ: event severity and novelty weight coefficients, used to adjust their respective contributions to triggering; τ: trigger threshold bias, which controls the degree of "combined score" that significantly triggers an update; w t ∈(0,1): event trigger weight, when w t ≈1 emphasizes event-driven updates; The weakly supervised loss of the online adaptive unit is based on the most recent running data D t ; The event-triggered update unit has strong supervision or specialized fine-tuning loss, which can be used to update event samples E based on manual annotation or expert rules. t calculate.

10. The natural language processing-driven smart park human-computer dialogue intention recognition system according to claim 1 is characterized in that: It also includes an update rollback unit, which is used to automatically determine and roll back to the previous version of the model when necessary by comparing key indicators before and after the incremental update to ensure system stability; Step 10.1 introduces concepts such as event severity, historical similarity suppression, trigger weight, and collaborative update loss. This allows for instant triggering of updates when a major incident occurs in the park. It also dynamically optimizes the update strategy by combining historical events and online performance data. The formula is: ΔP=P new -P old ,$4pt]D KL =D KL [p, P old ,P new Key performance indicators of the model before / after rollback on the validation set or online monitoring data; ΔP performance difference: when ΔP < 0, it means that the performance has degraded after the update and a rollback should be considered; D KL KL divergence: measures the degree of drift of the model output distribution before and after the update, and is used to detect potential instability caused by "distribution shift"; P Performance penalty weight: controls the impact of performance degradation on rollback decisions; w D Drift penalty weight: controls the impact of distribution drift on rollback decision; ∈ rollback threshold: when the rollback loss L rb When the threshold is exceeded, rollback is forced; σ(·) Sigmoid function: used to map the rollback loss to probability, supporting the "soft rollback" strategy.

Images