Large model-based meaningless input processing optimization method

Abstract

The invention discloses a meaningless input processing optimization method based on a large model, and relates to the cross technical field of large model input processing, natural language processing and multi-modal fusion, and the method comprises the steps: obtaining text, voice and image multi-modal input, and carrying out basic cleaning, multi-modal feature extraction and unified representation fusion; core features of keyword confusion degree, grammar destruction mode and logic consistency are mined in a linkage mode through an intention understanding module, and accurate matching of the features and scenes is achieved by means of a two-way attention mechanism; hierarchical filtering is carried out according to a three-level framework of coarse screening, fine judgment and decision making, and the processing efficiency and the recognition precision are balanced; according to the method, scene five-dimensional data modeling is combined to generate personalized response, scene adaptation weight adjustment and safety verification mechanism processing are used for different scenes such as customer service, children and finance, accurate recognition, high efficiency and low consumption of meaningless input, cross-scene adaptation and self-evolution processing are finally achieved, and result compliance and user satisfaction are guaranteed.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of natural language processing, multimodal learning, and large-scale artificial intelligence model optimization, specifically a method for optimizing the processing of meaningless input based on a large model. Background Technology

[0002] Handling meaningless input from large models is a core aspect of ensuring effective interaction, efficient resource utilization, and a positive user experience. However, traditional processing techniques and existing optimization solutions struggle to balance recognition accuracy, processing efficiency, and scenario adaptability, thus failing to meet the demands of batch interaction across multiple scenarios. Specific limitations are as follows:

[0003] Traditional large-scale models often employ rule filtering based on keyword matching, regular expressions, or simple statistical features such as word frequency and sentence length. They lack an intent understanding module and common-sense verification mechanism, making them ineffective at recognizing semantically confusing inputs that conform to surface grammatical structures. Some solutions introduce language model perplexity as a criterion, but their ability to recognize complex situations such as logical contradictions and cross-modal inconsistencies is limited. Furthermore, traditional methods are typically designed only for text modality, lacking a unified and robust ability to handle meaningless information in multimodal inputs such as speech and images (e.g., garbled speech or meaningless text layout in images), leading to cross-modal misjudgments and missed judgments, posing a significant risk of cross-modal misjudgment.

[0004] To address the aforementioned issues, some studies have attempted to introduce deep learning models for end-to-end meaningless input recognition. While this improves accuracy, it typically requires full inference for each input, resulting in lengthy processing times per input, low efficiency, and high computational costs, making it difficult to meet the real-time requirements of high-concurrency scenarios. Other solutions employ lightweight models or caching mechanisms to improve efficiency, but sacrifice accuracy in recognizing novel and complex meaningless patterns, resulting in insufficient accuracy and an inability to simultaneously achieve rapid response and accurate recognition. Furthermore, these solutions suffer from rigid resource allocation strategies, failing to dynamically schedule computing resources based on input confidence levels. Excessive storage and power consumption make them unsuitable for batch input scenarios with a large daily volume of inquiries.

[0005] The lack of cross-scenario response adaptability and the traditional solution adopts a uniform response mode without differentiation, which does not take into account the characteristics of the scenario (user profile, business type, environmental information) and emotional and courtesy needs. This results in a lack of effective input guidance in customer service scenarios, weak interactivity in children's scenarios, insufficient risk warnings in financial scenarios, and insufficient user satisfaction, failing to meet the differentiated interaction needs of different scenarios.

[0006] Therefore, there is an urgent need for an optimization method that can accurately identify multimodal meaningless inputs, balance processing efficiency and system resources, and have the ability to adapt to different scenarios. Summary of the Invention

[0007] The purpose of this invention is to provide an optimization method for processing meaningless input based on large models, in order to solve the problems of high misjudgment rate, incomplete multimodal coverage, imbalance between efficiency, accuracy and resource consumption, and insufficient cross-scene response adaptability of existing large model meaningless input processing technologies.

[0008] To address this, the present invention provides an optimization method for handling meaningless input based on a large model, comprising the following steps:

[0009] S10. Acquire multimodal inputs of text, speech, and images, complete basic cleaning, multimodal feature extraction and unified representation fusion, and unify the input format;

[0010] S20. Call the intent understanding module to perform preliminary intent interpretation on the input, and mine three core features of meaningless input: keyword confusion, syntax violation pattern, and logical consistency. Achieve accurate contextual matching between features and feature library through a bidirectional attention mechanism.

[0011] S30 executes hierarchical filtering according to a three-level architecture of coarse screening, fine judgment and decision-making, quickly diverts high-confidence meaningless inputs, performs in-depth verification of suspected inputs, and allocates CPU / GPU computing resources as needed.

[0012] S40. Based on five-dimensional data of users, environment, and business, model scenario types, generate scenario-based response strategies, and output the optimal response template after dynamic weighted sorting.

[0013] S50. Through feature tracing and decision visualization, interpretable output is achieved, and the judgment basis is explained to the user in a synchronized manner;

[0014] S60. Collect user feedback and manually labeled data, update the feature library through incremental comparative learning, and optimize the response strategy using PPO reinforcement learning to form a closed-loop iteration.

[0015] If the judgment result does not meet the scenario compliance requirements or user feedback indicates room for optimization, return to the hierarchical filtering step to readjust the parameters.

[0016] In one embodiment, the input access and preprocessing in step S10 are specifically operated as follows:

[0017] (1) Compatible with three types of input: text, voice and image: text input directly captures character sequences; voice input is transcribed into text strings, preserving rhythm and pause acoustic features; image input extracts text content through OCR, and simultaneously collects visual features such as layout, character size and color contrast.

[0018] (2) Targeted basic cleaning: At the text level, meaningless symbols and long strings of characters are removed using regular expressions, and common spelling errors are corrected; the speech-to-text verification is performed to check the coherence and mark abnormally segmented paragraphs; the image-extracted text is filtered to remove incomplete characters and record the location distribution features.

[0019] (3) Multimodal feature extraction: Text semantic vectors are generated by the Sentence-BERT model, speech feature vectors are generated by the MFCC+BiLSTM model, and image visual feature vectors are generated by the ResNet-50 model.

[0020] (4) Multimodal fusion: The ViLT / FLAVA multimodal Transformer is used to map the three types of feature vectors to the same 768-dimensional semantic space to eliminate the differences in modal representation.

[0021] In one embodiment, the specific operation method for performing intent understanding and linkage recognition on random input in step S20 is as follows:

[0022] (1) The intent understanding module outputs preliminary interpretation results such as query requirements and no clear intent;

[0023] (2) Extracting core features: Keyword confusion is calculated by combining Word2Vec / BERT semantic similarity algorithm with TF-IDF weighting, grammatical violation patterns are identified by dependency parsing, information entropy calculation and n-gram language model, and logical consistency is detected by combining TransE / TransR knowledge graph reasoning with COMET-ATOMIC common sense verification.

[0024] (3) The bidirectional attention mechanism aligns the initial intent with the typical misjudged intent in the feature library. If the matching degree reaches the set standard, it is judged as meaningless input and misjudged case and included in the feature library update queue. If it does not reach the standard, it enters the hierarchical filtering mechanism.

[0025] In one embodiment, the specific operation method of the hierarchical filtering mechanism in step S30 is as follows:

[0026] (1) Coarse screening layer (efficiency priority): The TinyBERT / ALBERT distillation model, RE2 multi-mode regular matching, and BM25 fast retrieval are used to match meaningless input templates and high-frequency fragments; a high threshold and a low threshold are set. If the matching degree is not lower than the high threshold, it is marked as a high-confidence meaningless input. If the matching degree is between the low threshold and the high threshold, it is marked as a suspected meaningless input. If the matching degree is lower than the low threshold, it is marked as a normal input candidate.

[0027] (2) Fine-grained judgment layer (accuracy priority): Validation is carried out through Cross-Attention multimodal fusion, FGSM+PGD adversarial training, and context coherence algorithm; if the multimodal fusion is judged to be meaningless, the change in intent after adversarial perturbation reaches the set standard, or the context BLEU value is lower than the set standard, it is judged to be meaningless; otherwise, full intent inference is started.

[0028] (3) Decision layer (resource optimization): a greedy resource allocation algorithm and WMRR weighted sorting are adopted. High-confidence meaningless input calls the CPU lightweight response template, and normal input calls the GPU full-precision inference resources.

[0029] In one embodiment, the specific operation method for cross-scene response adaptation output in step S40 is as follows:

[0030] (1) Scene feature modeling: Using LightGBM / XGBoost gradient boosting tree and Doc2Vec / Paragraph-BERT, the five-dimensional data of user profile (age, device), environment (time, location) and business are processed to output scene type and adaptation weight;

[0031] (2) Contextualized response strategy generation: Based on Transformer Seq2Seq generation model, VADER sentiment analysis, and Politeness Score algorithm; generate polite prompts and effective input guidance content for customer service scenarios, generate fun interactive and positive guidance content for children's scenarios, and generate risk prompts and manual verification guidance content for financial scenarios;

[0032] (3) Response output: The optimal template is dynamically sorted and selected according to WMRR. The sorting weight factors include scenario adaptability, user preference, response time and security. If the security does not meet the standard, the weight of the template will be reduced.

[0033] In one embodiment, the specific operation of the interpretability output in step S50 is as follows:

[0034] (1) Feature tracing: By analyzing LIME and SHAP values, the input segments that play a key role in the determination of meaninglessness are located, and the contribution of each segment to the determination result is quantified;

[0035] (2) Decision visualization: The matching focus of the bidirectional attention mechanism is presented through the attention weight heatmap, and the full decision path of hierarchical filtering is presented through the Graphviz decision tree;

[0036] (3) Output format: The response text is accompanied by a simplified source explanation, which clearly defines the core basis for determining that it is meaningless input.

[0037] In one embodiment, the specific operation method of the feedback optimization closed loop in step S60 is as follows:

[0038] (1) Feedback collection: Obtain feedback data through subsequent user behavior (including resending valid input and dwell time) and manual annotation;

[0039] (2) Feature library update: Incremental contrastive learning (Triplet Loss) is adopted to construct triplets with new misjudged cases as positive samples and normal inputs as negative samples for training;

[0040] (3) Response strategy optimization: By combining PPO reinforcement learning with Thompson Sampling multi-armed slot machine, the strategy weights are adjusted with response efficiency, user satisfaction, security and interpretability as reward functions.

[0041] The present invention proposes an optimization method for processing meaningless input based on a large model, the advantages of which are as follows:

[0042] This large-scale model's optimized method for handling meaningless input leverages intent understanding and multimodal feature fusion to accurately extract core features such as keyword confusion, grammatical violations, and logical conflicts. A hierarchical filtering mechanism facilitates efficient coarse screening and in-depth verification, while lightweight quantization compression and dynamic resource allocation balance response speed and recognition accuracy. Cross-scenario feature modeling and personalized response strategies adapt to the functional, emotional, and compliance requirements of different scenarios. Feedback optimization loops and incremental comparative learning enable dynamic self-evolution of the feature library. Interpretability enhancement technology facilitates traceability and visualization of the judgment process. For special cases such as multimodal garbled text, novel meaningless inputs, and high-risk compliance scenarios, cross-modal unified representation, adaptive strategy iteration, and security weight constraints are used. Ultimately, this achieves high accuracy, high efficiency, high adaptability, and high transparency in meaningless input recognition, ensuring low-power operation and improved user experience in batch scenarios. Attached Figure Description

[0043] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0044] Figure 1 This is a closed-loop flowchart of the first embodiment of this application;

[0045] Figure 2 This is a flowchart of the three-level decision-making process for hierarchical filtering according to the first embodiment of this application;

[0046] Figure 3This is a flowchart of the cross-scenario response adaptation architecture of the first embodiment of this application. Detailed Implementation

[0047] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described in this specification are merely illustrative of the invention and are not intended to limit the invention.

[0048] Example:

[0049] Reference Figure 1 , Figure 1 This is a closed-loop flowchart of the first embodiment of a meaningless input processing optimization method based on a large model according to this application. The meaningless input processing optimization method based on a large model may include steps S10 to S50:

[0050] Step S10: Obtain multimodal inputs of text, speech and images, and perform basic cleaning, multimodal feature extraction and unified representation fusion.

[0051] In one feasible implementation, step S10 may include steps S101 to S104:

[0052] Step S101: Receive text and multimodal input.

[0053] Specifically, for text input, the character sequence entered by the user is captured directly; for voice input, the audio signal is first converted into a text string through speech-to-text technology, while preserving the original speech's rhythm, pauses, and other acoustic features; for image input (such as screenshots containing text or handwritten images), the text content is first extracted through optical character recognition (OCR) technology, and then visual features such as the text layout, character size, and color contrast in the image are simultaneously acquired.

[0054] Step S102: Perform basic cleaning on the received data.

[0055] Specifically, at the text level, regular expressions are used to remove special symbols without semantic value, consecutive whitespace characters, and meaningless long strings of characters, while common spelling errors (such as mixing pinyin and Chinese characters, and confusing similar-looking characters) to ensure the initial semantic integrity of the text. For speech-to-text transcription, the coherence of the transcription results needs to be verified. If a large number of abnormal sentence breaks or semantically fragmented segments appear, they are marked as "speech segments to be verified" to avoid misjudgment caused by transcription errors. For image-extracted text, incomplete characters caused by image blurring or occlusion need to be filtered out, and the positional distribution characteristics of the text in the image are recorded simultaneously to provide visual dimension basis for subsequent multimodal fusion.

[0056] Step S103, multimodal feature extraction.

[0057] Specifically, the Sentence-BERT model transforms text into 768-dimensional semantic vectors. It captures the overall semantic information of the text; MFCC technology extracts the spectral features of speech, and then combines it with a BiLSTM model to learn the temporal prosodic rules of speech, generating speech feature vectors. The ResNet-50 model extracts visual features from images (such as edges and textures of text regions) to form image feature vectors. .

[0058] Step S104, multimodal fusion.

[0059] Specifically, the text (semantic vectors extracted by Sentence-BERT) ), speech (prosodic features extracted by MFCC+BiLSTM) Image text (layout features extracted by ResNet-50) The ViLT / FLAVA multimodal Transformer model is directly input, and a cross-modal attention mechanism maps features from different modalities to the same 768-dimensional semantic space. This model maps three types of feature vectors to a unified semantic space, eliminating representational differences between different modalities and resulting in a unified fused feature vector. .

[0060] Step S20 involves using the intent understanding module to mine core features such as keyword confusion, grammatical violation patterns, and logical consistency, and then using a bidirectional attention mechanism to achieve accurate matching between features and scenarios.

[0061] Specifically, the intent understanding module is associated with the feature extraction process. It identifies typical features from meaningless inputs, matches them to the feature library, and updates the feature library dynamically. The initial interpretation by the intent understanding module provides direction for feature extraction. Combined with a bidirectional attention mechanism, it achieves accurate matching between features and scenarios, ensuring the targeting and effectiveness of feature extraction. At the same time, it adds new misjudgment cases to the feature library, improving the accuracy of subsequent recognition.

[0062] It should be noted that adversarial example generation involves constructing meaningless samples in batches through character-level, syntactic-level, and semantic-level perturbations, and actively learning and optimizing annotations. An annotation committee composed of SVM, LR, and LightGBM prioritizes the annotation of ambiguous samples, providing high-quality training data for the feature library, which is the foundation for feature library updates.

[0063] In one feasible implementation, step S20 may include steps S201 to S203:

[0064] Step S201: Call the intent understanding module to perform preliminary intent interpretation on the input (such as query needs or no clear intent).

[0065] Step S202: Extract the three core features of meaningless input: keyword confusion, syntax violation pattern, and logical consistency.

[0066] Specifically, keyword disorder is calculated using the Word2Vec / BERT semantic similarity algorithm plus TF-IDF weighting, and the pairwise cosine similarity of the top 5 core words is used to determine keyword disorder. The mathematical expression is:

[0067]

[0068] In the formula, The top 5 core words of the input text after TF-IDF sorting, among which The range of values ​​is and , The cosine similarity is calculated based on Word2Vec / BERT semantic vectors, and its value range is [value range missing]. , The calculation method is as follows ,in It is in the input The ratio of the number of occurrences of a word to the total number of words in the input. yes The total corpus consists of 500,000 normal scenario texts (covering customer service, children, and financial scenarios).

[0069] Syntax violation patterns include dependency parsing + information entropy calculation and n-gram language models. The mathematical expression for information entropy is:

[0070]

[0071] In the formula, Text syntactic entropy characterizes the degree of disorder in the syntactic structure of the input text. ,in This represents the number of syntactic structure types, based on the Stanford dependency parser definition, covering 12 core syntactic structures including subject-verb, verb-object, and modifier-head structures. To determine the syntactic structure probabilities, we analyzed the syntax of 500,000 normal scenario texts and statistically determined the frequency of occurrence of each structure as the baseline probability. Entropy is a measure of the degree to which the prosodic pattern of a speech input deviates from the normal standard. , The number of normal rhythmic patterns is dynamically adjusted according to different scenarios (such as children's scenarios where voice input is the primary method). The value will be adapted to common children's intonation types. The probability distribution of normal prosodic patterns is defined based on speech features extracted using MFCC+BiLSTM, covering the standard prosody of the corresponding scenarios. Image text layout entropy is used to evaluate whether the layout of extracted text in an image conforms to normal standards. , The probability distribution for normal typesetting patterns is defined based on image text layout features extracted by ResNet-50, including standard typesetting forms such as left alignment and no character overlap. The number of normal layout modes covers mainstream and compliant image and text layout types. The text, speech, and image modal information are assigned weight coefficients, the sum of which is 1. The smaller the entropy value, the greater the modality weight. The specific calculation expression is as follows: , ,when hour, ;when hour, ;when hour, , The maximum entropy value for all modalities in the historical full scenario is used. In the n-gram language model, n takes the value of 2 / 3. When n=2, it is a bigram, which captures basic grammatical collocations. By statistically analyzing the occurrence probability of bigram pairs in normal corpus, it quantifies the degree of violation of basic grammatical collocations in the input and quickly identifies obvious grammatical confusion. When n=3, it is a trigram, which supplements the contextual logical connection, makes up for the limitations of bigram, captures logical conflicts at the short sentence level, and improves the ability to identify complex grammatical violation patterns.

[0072] Logical consistency is achieved through the collaboration of two major technologies: knowledge graph reasoning (TransE / TransR) and common sense base verification (COMET-ATOMIC). It judges whether there is a logical conflict in the input from two dimensions: the rationality of entity relationships and common sense in physics and mathematics, and ultimately helps to identify meaningless input.

[0073] Step S203: The bidirectional attention mechanism aligns the initial intent with the historical misjudged intents in the feature library, making the alignment of intent and misjudged features more focused on the current scene and reducing cross-scene mismatches.

[0074] Specifically, the mathematical expression for the bidirectional attention algorithm is:

[0075]

[0076] The mathematical expression for matching degree is:

[0077]

[0078] In the formula, All are 768-dimensional vectors, adapted to the output dimension of the BERT-base model, where Generated by the intent understanding module through input encoding. This is the encoding vector of historical misjudgment intentions in the feature library. for The corresponding meaningless feature vector; Five-dimensional vector of the scene It is encoded into a 768-dimensional scene feature vector using Doc2Vec; , ,in Let j be the vector of the j-th historical scene. Data is taken from historical scene data in the feature library. The cosine similarity between the current scene and the historical scene. It is a diagonal matrix, and the elements are the percentages of scene similarity. As the normalization factor, Its function is to avoid An excessively large inner product value causes the gradient of the Softmax function to vanish; the Softmax function will... Mapped to Weight distribution, the higher the weight, the more likely it is to be a weight. and The stronger the connection.

[0079] It should be noted that, by using cosine similarity to filter out the top 100 misjudged cases in the same or highly similar scenes with the current scene vector S, the average historical matching degree of similar historical cases is calculated, and the value of this average plus 0.9 is taken as the matching degree threshold. If the current matching degree is ≥ If the match is found to be "satisfactory", the system will proceed to the hierarchical filtering mechanism for further evaluation; otherwise, the system will proceed to the hierarchical filtering mechanism for further evaluation.

[0080] Step S30, refer to Figure 2 , Figure 2 The first embodiment of the meaningless input processing optimization method based on a large model provided in this application presents a three-level hierarchical filtering process flowchart. Through a three-level architecture of coarse screening, fine judgment, and decision-making, the method efficiently distinguishes between meaningless and normal inputs, balancing processing efficiency and recognition accuracy.

[0081] In one feasible implementation, step S30 may include steps S301 to S303:

[0082] Step S301, coarse screening layer, rapid feature matching, prioritizing efficiency.

[0083] The coarse-screening layer aims to quickly filter out high-confidence meaningless inputs, and its core lies in using lightweight techniques to achieve efficient recognition. Utilizing the TinyBERT-6L distillation model, while retaining BERT's core semantic recognition capabilities, the model parameters and computational load are significantly reduced, enabling rapid initial semantic judgment of the input text. Through RE2 multi-pattern regularization matching and BM25 fast retrieval technology, it matches predefined meaningless input templates (such as continuous symbol strings, fixed meaningless character combinations, and semantically conflicting sentence structures), while simultaneously retrieving frequently occurring meaningless input fragments from the feature library.

[0084] Specifically, set a low threshold. With high threshold The mean matching degree with standard deviation The mathematical expression is:

[0085]

[0086]

[0087] In the formula, This represents the total number of historical matching samples in the feature library. Let be the matching degree of the i-th sample.

[0088] It should be noted that when Mark the input as a high-confidence meaningless input; When the input is in the specified state, mark it as potentially meaningless; when... Mark the input as a normal input candidate.

[0089] Step S302, fine-grained judgment layer, deep semantic verification, prioritizing accuracy.

[0090] The precision judgment layer conducts in-depth validation targeting ambiguous cases between suspected meaningless input and normal input candidates, with the core objective of improving recognition accuracy. Multimodal fusion validation further explores the correlation between text, speech, and image features through the Cross-Attention mechanism; adversarial training validation (FGSM+PGD) observes changes in the model's judgment of input intent by applying small perturbations to the input features; and contextual coherence validation combines historical dialogue scenarios (if they exist) to determine semantic coherence.

[0091] Specifically, calculation With meaningless input typical feature vectors in the feature library The similarity is then combined with a scene-dynamic weighted average to obtain the final result. :

[0092]

[0093] In the formula, Dynamic scene weights, determined by LightGBM based on scene vectors. Output, satisfying , These are the typical modal features corresponding to meaningless inputs in the feature library.

[0094] Specifically, use and The method applies strong perturbations to the features of each modality, generates perturbed inputs, encodes them as perturbed intent vectors, and then calculates the single-modal intent deviation: Combined with the scene dynamic weights output by the LightGBM model The weighted fusion yields the final intentional change after adversarial perturbation:

[0095]

[0096] In the formula, The similarity between the original intent and the perturbated intent. This represents the similarity between the original intent and the normal intent.

[0097] Specifically, the mathematical expression for the context BLEU value:

[0098]

[0099] In the formula, The reference text length (reasonable subsequent input based on historical dialogues generated by GPT-2). , This represents the weight of the bigram and trigram syntax, with a value of 0.5. This is the bigram precision (the percentage of bigram matches between the current input and the reference text). This is the ternary syntax precision (the percentage of ternary syntax matches between the current input and the reference text).

[0100] Extracting scene-related features from the feature library The top 1000 meaningless inputs with the highest cosine similarity are used as the positive sample set. , and scene The top 1000 normal inputs based on cosine similarity are used as the negative sample set. ,calculate All samples The mean, standard deviation and All samples The mean and standard deviation are denoted as . , , and Constrain the candidate threshold range (95% confidence interval): Design the Youden Index ,in for middle The percentage of samples (true positive rate). for middle The percentage of samples (false positive rate) that ultimately determines the candidate range. The largest ,Right now:

[0101]

[0102] In the formula, The number of elements in the set.

[0103] Build a Apply strength The meaningless adversarial sample set after FGSM / PGD perturbation and pairs Apply strength The normal adversarial sample set after FGSM / PGD perturbation is denoted as follows: and ,calculate middle Mean, standard deviation and middle The mean, denoted as , and The final threshold was set to normal input adversarial examples. Upper limit of the 95% confidence interval: .

[0104] Build a scene-specific normal dialogue set (Statistics on normal dialogue logs in scenario S and round N from the feature library) The mean and standard deviation of the contextual BLEU scores for all dialogues are denoted as . and The mathematical expression for user feedback weight is:

[0105]

[0106] In the formula, For the scene Rounds Average user dwell time For the scene Rounds Average number of resends per user.

[0107] The lower the BLEU score, the less coherent the input (the higher the probability of it being meaningless). The BLEU score threshold is set as the lower bound of the 95% confidence interval of the BLEU distribution in normal dialogue. The mathematical expression is:

[0108]

[0109] Specifically, if the multimodal fusion is deemed meaningless, the intention change after adversarial perturbation reaches a set range, or the context BLEU value is lower than a set standard, then the input is deemed meaningless; otherwise, it is considered normal input, and full intention reasoning is initiated.

[0110] Step S303: Dynamically decide on resource scheduling, focusing on optimizing resource allocation, and select the optimal computing resources and response method based on the judgment results and input type.

[0111] Specifically, the mathematical expression for the greedy allocation logic in step S303 is:

[0112]

[0113] In the formula, For resource consumption costs, the "calculation time" method is used. Video memory usage Power consumption "Weighted summation, , The confidence score for meaningless input is obtained by weighting the coarse-screen matching score and the fine-screen probability. Force the invocation of the GPU threshold objective expression:

[0114]

[0115] Among them, the true positive rate True negative rate , Confidence threshold The number of true positives These are false negative numbers. It is a true negative number. These are false positives. Positive: meaningless input; Negative: normal input. The threshold corresponding to the maximum value of the Youden exponent on the ROC curve, satisfying... Force GPU usage at times.

[0116] By using dynamic weight calculation and sorting rules, accurate matching and efficient output of response templates are achieved. The mathematical expression for WARR weighted sorting is:

[0117]

[0118] In the formula, The number of candidate templates. , For the first The dynamic weight of each template, For the first Security constraint factors for each template For the first The ranking of each template.

[0119] Specifically, the mathematical expression for dynamic weights is:

[0120] ,

[0121] In the formula, The correlation coefficient for scene adaptability is based on templates in the feature library. The Pearson correlation coefficient between reuse rate and user satisfaction in the current scenario. The template is calculated based on the user's historical click / stay data to determine the correlation coefficient of user preferences. The Spearman correlation coefficient between selection frequency and user retention rate The response time correlation coefficient, The correlation coefficient is the overall correlation coefficient, output based on the scenario risk level model.

[0122] The mathematical expression for the security constraint factor is:

[0123]

[0124] In the formula, The matching rate is calculated from the scenario-compliant vocabulary in the feature library. For the scene The safety threshold is determined by taking the minimum compliance score of the template with no user complaints in scenario S from the feature library. .

[0125] according to Preliminary sorting in descending order, denoted as , computational scenario Standard deviation of historical template weights , For the scene Total number of historical templates The historical weighted average is used for two templates. , satisfy Then, they are rearranged in ascending order of response time, with the shorter response time being the first to arrive. Better.

[0126] Step S40: Combine five-dimensional data modeling of the scene to generate personalized responses, and achieve interpretable output through LIME+SHAP feature tracing and decision visualization.

[0127] In one feasible implementation, step S40 may include steps S401 to S402:

[0128] Step S401: The core attributes of the scene are mined through scene feature modeling to provide a basis for the response strategy; the appropriate response content is generated in combination with the scene requirements, and the optimal template is selected through dynamic sorting.

[0129] In one specific implementation, step S401 may include steps S4011 to S4013:

[0130] Step S4011: Scene feature modeling. Core attributes of the scene are mined from five dimensions: age, device, time, location and industry. The scene type is classified and output through the LightGBM algorithm. The scene description is transformed into a 768-dimensional semantic vector through the Paragraph-BERT algorithm, which supports cross-scene policy transfer.

[0131] Step S4012, Personalized response generation, uses the Transformer Seq2Seq model, takes scene vectors and meaningless input features as input, and outputs adapted text, where the encoder is BERT-base and the decoder is a lightweight version of GPT-2.

[0132] It should be noted that children's scenarios must meet the VADER positive emotion-related standard, while customer service scenarios must meet the Politeness Score-related standard.

[0133] In step S4013, the optimal template is dynamically sorted by WMRR to obtain real-time feedback. If the security does not meet the standard, the weight of the template is reduced to 0.1, which greatly reduces the probability of the template being selected for output.

[0134] Step S402: Interpretable output is achieved through LIME+SHAP feature tracing and decision visualization. The final output response text is accompanied by a simplified tracing explanation.

[0135] In one specific implementation, step S402 may include steps S4021 to S4022:

[0136] Step S4021: Quantify the impact of the input segment on the judgment result through feature tracing to help the user understand the reason for the judgment.

[0137] Specifically, LIME locally interpretable model-agnostic interpretation constructs a local linear model for a single meaningless input to locate the input segment that plays a key role in the judgment result; SHAP value analysis is based on game theory principles to quantify the contribution weight of each key segment to the meaningless judgment, providing users and reviewers with specific evidence of why it is judged as meaningless, and also accurately locating the reasons for misjudgment, providing targeted optimization directions for feature library updates (incremental comparative learning).

[0138] Step S4022 presents the entire decision-making process through decision visualization, facilitating review and iteration, and ensuring that the decision-making process is transparent and traceable.

[0139] Specifically, decision visualization includes attention weight heatmaps and Graphviz decision trees: attention weight heatmaps visualize the matching focus of the bidirectional attention mechanism, and Graphviz decision trees visualize the entire path of coarse screening, fine judgment, and decision-making, intuitively presenting the thresholds and rules of each step of the judgment, making the judgment process transparent from a black box, which is convenient for manual review and provides observable optimization nodes for model iteration.

[0140] Step S50: Collect user feedback and manually labeled data, and dynamically update the feature library and response strategy through incremental comparative learning and PPO reinforcement learning; for different scenarios such as customer service, children and finance, use scenario adaptation weight adjustment and security verification mechanism to process, and finally achieve accurate recognition of meaningless input, high efficiency and low consumption, cross-scenario adaptation and self-evolution processing.

[0141] In one feasible implementation, step S50 may include steps S501 to S503:

[0142] Step S501: Collect feedback by obtaining feedback data through subsequent user behavior (re-sending valid input / stay duration) and manual annotation.

[0143] Step S502: Feature library update. Through incremental contrastive learning (Triplet Loss), new misjudged cases are treated as positive samples and normal inputs are treated as negative samples.

[0144] Step S503: Response strategy optimization. Through PPO reinforcement learning and Thompson Sampling multi-armed slot machine, the strategy weights are adjusted with response efficiency, user satisfaction, security and interpretability as reward functions.

[0145] Specifically, the mathematical expression for the PPO reinforcement learning reward function is:

[0146]

[0147] In the formula, The scene dynamic weights are calculated by the scene feature modeling module based on the scene's five-dimensional vector. Using the LightGBM gradient boosting tree model, scene-specific historical data is analyzed to output adaptive weights for efficiency, satisfaction, and security, ensuring that the weights match the scene's priority. Weighting based on user feedback , The longer the dwell time and the fewer the number of resends, the closer the weight is to 1.2. For efficiency, ,in For the current input Actual response time (unit: ms). / For the scene Download the minimum / maximum historical response times. The shorter the time taken, the higher the score. For user satisfaction, In response to security, ,in For the scene Total number of compliant keywords For the current response The number of compliant keywords hit. The more hits you make, the higher your score.

[0148] Specifically, define a scenario-specific sample set: positive sample set. Scene The sample size is as follows: Historical interaction samples from which users explicitly expressed "satisfaction" (positive reviews, no complaints, and no resubmissions). negative sample set Scene The sample size is the historical interaction data from which users explicitly expressed dissatisfaction. and the full sample set .right Label ,right Label Define a set of behavioral indicators. From the feedback collection module, indicators strongly correlated with satisfaction are selected and defined differently according to different scenarios:

[0149] Financial scenarios:

[0150] Children's scene:

[0151] Customer service scenario:

[0152] Pearson correlation coefficient was used to quantify the indicators. With satisfaction tags The correlation strength is expressed mathematically as:

[0153]

[0154] For the sample In terms of indicators The value on, For the scene Lowering the target The mean, For the scene The historical average satisfaction level. The closer the value is to 1, the stronger the positive correlation between the indicator and satisfaction.

[0155] Only positively correlated indicators are retained, and after normalization, the weights are ensured to sum to 1. The mathematical expression is:

[0156] Standardized score To eliminate dimensional differences, Z-score standardization is adopted, and the mathematical expression is: , For the scene Lowering the target The standard deviation.

[0157] Ultimately, the satisfaction level was determined. The mathematical expression is:

[0158]

[0159] In the formula, Map the standardized scores to The interval, the final .

[0160] With reward function value To optimize the objective, the response policy parameters are updated using the PPO clip constraint mechanism: Let the policy before the update be... After the update, it is Calculate the strategy ratio , In response to the action, To define the scene state, the clip function is used to limit the amplitude: ,in, , , The step size is set to 0.001, based on the historical average reward for the same scenario. This is applied when the reward fluctuates for three consecutive rounds. or improve the accuracy of the scenario test set. Stop when the time comes.

[0161] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing the processing of meaningless input based on a large model, characterized in that: Includes the following steps: S10. Acquire multimodal inputs of text, speech, and images, complete basic cleaning, multimodal feature extraction and unified representation fusion, and unify the input format; S20. Call the intent understanding module to perform preliminary intent interpretation on the input, and mine three core features of meaningless input: keyword confusion, syntax violation pattern, and logical consistency. Achieve accurate contextual matching between features and feature library through a bidirectional attention mechanism. S30 executes hierarchical filtering according to a three-level architecture of coarse screening, fine judgment and decision-making, quickly diverts high-confidence meaningless inputs, performs in-depth verification of suspected inputs, and allocates CPU / GPU computing resources as needed. S40. Based on five-dimensional data of users, environment, and business, model scenario types, generate scenario-based response strategies, and output the optimal response template after dynamic weighted sorting. S50. Through feature tracing and decision visualization, interpretable output is achieved, and the judgment basis is explained to the user in a synchronized manner; S60. Collect user feedback and manually labeled data, update the feature library through incremental comparative learning, and optimize the response strategy using PPO reinforcement learning to form a closed-loop iteration. If the judgment result does not meet the scenario compliance requirements or user feedback indicates room for optimization, return to the hierarchical filtering step to readjust the parameters.

2. The method for optimizing meaningless input processing based on a large model according to claim 1, characterized in that: In step S10, the specific operation method for input access and preprocessing is as follows: (1) Compatible with three types of input: text, voice and image: text input directly captures character sequences; voice input is transcribed into text strings, preserving rhythm and pause acoustic features; image input extracts text content through OCR, and simultaneously collects visual features such as layout, character size and color contrast. (2) Targeted basic cleaning: At the text level, meaningless symbols and long strings of characters are removed using regular expressions, and common spelling errors are corrected; the speech-to-text verification is performed to check the coherence and mark abnormally segmented paragraphs; the image-extracted text is filtered to remove incomplete characters and record the location distribution features. (3) Multimodal feature extraction: Text semantic vectors are generated by the Sentence-BERT model, speech feature vectors are generated by the MFCC+BiLSTM model, and image visual feature vectors are generated by the ResNet-50 model. (4) Multimodal fusion: The ViLT / FLAVA multimodal Transformer is used to map the three types of feature vectors to the same 768-dimensional semantic space to eliminate the differences in modal representation.

3. The method for optimizing meaningless input processing based on a large model according to claim 1, characterized in that: In step S20, the specific operation method for performing intent understanding and linkage recognition on random input is as follows: (1) The intent understanding module outputs preliminary interpretation results for query requests and queries without clear intent; (2) Extracting core features: Keyword confusion is calculated by combining the Word2Vec / BERT semantic similarity algorithm with TF-IDF weighting. Syntax violation patterns are identified by dependency parsing, information entropy calculation and n-gram language model. Logical consistency is detected by combining TransE / TransR knowledge graph reasoning with COMET-ATOMIC common sense verification. (3) The bidirectional attention mechanism aligns the initial intent with the typical misjudged intent in the feature library. If the matching degree reaches the set standard, it is judged as meaningless input and misjudged case and included in the feature library update queue. If it does not reach the standard, it enters the hierarchical filtering mechanism.

4. The method for optimizing meaningless input processing based on a large model according to claim 1, characterized in that: In step S30, the specific operation method of the hierarchical filtering mechanism is as follows: (1) Coarse screening layer: The TinyBERT / ALBERT distillation model, RE2 multi-mode regular matching, and BM25 fast retrieval are used to match meaningless input templates and high-frequency fragments; a high threshold and a low threshold are set. If the matching degree is not lower than the high threshold, it is marked as a high-confidence meaningless input. If the matching degree is between the low threshold and the high threshold, it is marked as a suspected meaningless input. If the matching degree is lower than the low threshold, it is marked as a normal input candidate. (2) Fine-grained judgment layer (accuracy priority): Validation is carried out through Cross-Attention multimodal fusion, FGSM+PGD adversarial training, and context coherence algorithm; if the multimodal fusion is judged to be meaningless, the change in intent after adversarial perturbation reaches the set standard, or the context BLEU value is lower than the set standard, it is judged to be meaningless; otherwise, full intent inference is started. (3) Decision layer: The greedy resource allocation algorithm and WMRR weighted sorting are adopted. High-confidence meaningless input calls the CPU lightweight response template, and normal input calls the GPU full-precision inference resources.

5. The method for optimizing meaningless input processing based on a large model according to claim 1, characterized in that: In step S40, the specific operation method for cross-scene response adaptation output is as follows: (1) Scene feature modeling: The five-dimensional data of user profile, environment, and business are processed by LightGBM / XGBoost gradient boosting tree and Doc2Vec / Paragraph-BERT to output scene type and adaptation weight; (2) Contextualized response strategy generation: Based on the Transformer Seq2Seq generation model, VADER sentiment analysis, and Politeness Score algorithm, guide content is generated; (3) Response output: The optimal template is dynamically sorted and selected according to WMRR. The sorting weight factors include scenario adaptability, user preference, response time and security. If the security does not meet the standard, the weight of the template will be reduced.

6. The method for optimizing meaningless input processing based on a large model according to claim 1, characterized in that: In step S50, the specific operation method for interpretable output is as follows: (1) Feature tracing: By analyzing LIME and SHAP values, the input segments that play a key role in the determination of meaninglessness are located, and the contribution of each segment to the determination result is quantified; (2) Decision visualization: The matching focus of the bidirectional attention mechanism is presented through the attention weight heatmap, and the full decision path of hierarchical filtering is presented through the Graphviz decision tree; (3) Output format: The response text is accompanied by a simplified source explanation, which clearly defines the core basis for determining that it is meaningless input.

7. The method for optimizing meaningless input processing based on a large model according to claim 1, characterized in that: In step S60, the specific operation method of the feedback optimization closed loop is as follows: (1) Feedback collection: Obtain feedback data through subsequent user behavior and manual annotation; (2) Feature library update: Incremental contrastive learning is adopted, and triples are constructed with new misjudged cases as positive samples and normal inputs as negative samples for training; (3) Response strategy optimization: By combining PPO reinforcement learning with Thompson Sampling multi-armed slot machine, the strategy weights are adjusted with response efficiency, user satisfaction, security and interpretability as reward functions.

Images