343 results about "Confusion" patented technology

The invention relates to a test case generation method and system based on multi-agent efficient collaboration, belongs to the technical field of software testing, and solves the problems of incomplete scene coverage, logic disorder and the like when a single large model processes a complex task. The method comprises the following steps: constructing a multi-modal professional field knowledge base; the task planning agent module generates a test case generation task based on a project development document and a software source code file of a to-be-tested project and distributes the test case generation task to the test demand analysis agent module; based on the test case generation task, extracting a test demand point related to the tested software configuration item, describing the test demand point, and labeling a corresponding tracking relationship between a test demand point ID and a code snippet or a function in the software source code file to construct a test demand-code snippet / function set; and generating a test case and a test description document based on the test demand-code snippet / function set and the multi-modal professional domain knowledge base. And the ability of generating the test case of the large model is enhanced through organic and efficient cooperation of multiple agents.

The invention discloses a business English scene knowledge graph generation and intelligent retrieval method, and relates to the technical field of natural language processing, and the method comprises the steps: constructing a document time sequence index and anaphora alignment mapping table, and registering an evidence anchor point index; extracting event units under the constraint and merging the event units into event clusters in a cross-document manner; arranging an event chain according to unified time and generating a process graph, and binding evidence anchor points to relation edges; analyzing the query into a query intention expression and a path constraint, and generating an evidence presentation sub-plan and an image-text combined execution plan; and positioning answers in the atlas, integrally returning the answers, paths and evidences, and meanwhile, forming replayable audit tracks. The method solves the problems of cross-channel time dislocation and confusion of references, and supports interpretable and auditable retrieval. The method is suitable for scenes such as mails, conference summary, instant communication and the like, guarantees locatable sources, path replay and evidence verification, and is convenient for internal control review.

The invention provides a text segmentation method and related equipment. The method comprises the steps of obtaining a to-be-processed text; segmenting the to-be-processed text into ordered statement sequences to obtain an initial statement set of the to-be-processed text; wherein the ordered statement sequence comprises a plurality of statements; the semantic variation, the confusion degree variation and the information entropy variation of a first target text block are calculated when a to-be-decided statement in an initial statement set of the to-be-processed text is added into the first target text block, and the first target text block is a set of multiple statements meeting a merging condition; determining the collaboration degree of the statement to be decided and the first target text block based on the semantic variable quantity, the confusion variable quantity and the information entropy variable quantity; and partitioning the to-be-processed text based on the collaboration degree of the to-be-decided statement and the first target text block to obtain a target partitioned text of the to-be-processed text. The text segmentation quality can be improved.

The invention provides a multi-modal data desensitization method, system and device and a medium, and belongs to the technical field of information security. The method comprises the following steps: firstly, acquiring multi-modal original data based on preset access information and acquisition frequency, and after standardized preprocessing, positioning sensitive information by using a classification recognition algorithm and generating an analysis report. Then, according to a report, matching a strategy from a desensitization strategy library bound with the sensitivity level, adjusting a dynamic confusion algorithm parameter to desensitize, integrating and generating desensitized multi-modal data, and recording and storing a mapping relationship between the original data and the desensitized data; and if a restoration request is received, after authorization verification is passed, reverse desensitization restoration data is carried out by using the mapping relation. Besides, full-process feedback information is collected, the desensitization effect is quantitatively evaluated according to a preset index, the desensitization strategy and algorithm are iteratively optimized accordingly, and effective desensitization and safety management of multi-modal data are achieved.

The invention discloses an intelligent perception analysis system for the real-time participation degree and cognitive state of a mathematics classroom. The intelligent perception analysis system comprises a multi-source data acquisition module, a fusion analysis engine module, a real-time feedback module and an offline optimization module. The method has the beneficial effects that a dynamic participation index is innovatively designed, and an attention attenuation resetting mechanism is introduced; the real-time feedback module generates a cognitive thermodynamic diagram of HSL color mapping to position group obstacle points, and triggers self-adaptive question pushing; and the off-line optimization module dynamically updates the edge weight of the knowledge graph through error co-occurrence analysis, and early warns a cognitive confusion relationship without textbook association. According to the method, the limitation of single-mode perception is broken through, the problem solving step quality visual diagnosis and the cross-cycle cognitive impairment prediction are realized, a teaching closed loop of'multi-source perception-hierarchical quantification-real-time intervention-knowledge evolution 'is formed, and the mathematical classroom cognitive state analysis precision and the teaching intervention timeliness are remarkably improved.

The invention relates to the technical field of online education, and discloses an online course learning management method based on a knowledge graph. The method comprises the following steps: acquiring multi-modal learning behavior data of a learner, and extracting a deep learning state vector reflecting knowledge understanding depth, learning input degree and cognitive confusion through semantic fusion; and dynamically calculating and updating the logical relationship strength among the knowledge points in the course knowledge graph by using the vector, so that the knowledge structure can adaptively evolve along with the actual cognitive state of the learning group. And generating a real-time personalized learning path based on the updated knowledge graph and the current state vector of the learner. Meanwhile, according to cognitive confusion features in the state vector, intervention measures such as pushing of remedial resources, adjusting of content sequence or starting of self-adaptive testing are triggered in real time. According to the method, the dynamic optimization of the knowledge graph and the accurate and immediate response of learning intervention are realized, and the adaptability and management efficiency of online learning are improved.

The invention provides an experience knowledge graph construction and question answering method based on heuristic self-question answering, and the method comprises the steps: randomly dividing a text corpus into small batches, inputting the small batches into a large language model, and obtaining a heuristic rule set through induction processing and quality evaluation; constructing question and answer pairs by using the heuristic rule set and the text content cues, converting the question and answer pairs, and then carrying out confidence calculation to obtain a structured experience triple with confidence; constructing a knowledge graph by using the structured experience triad with confidence; extracting the core fragment of the original narrative query by using the heuristic rule set and generating a refined query text; obtaining an empirical path by using the embedded vector of the refined query text and the knowledge graph; and inputting the original narrative query and the experience path into the large language model to generate a final answer. According to the method, logic confusion and content contradiction possibly caused by a traditional RAG method are effectively avoided, the generated answer is more reliable, and the reasoning process is more interpretable.

The invention discloses a semantic matching-based field name standard abbreviation mapping method and system, and particularly relates to the technical field of data governance, a source field set to be subjected to abbreviation and a standard field set in a standard field library are acquired, field names and field descriptions are normalized, and field text representation is generated; using the pre-training language model to generate a semantic vector and recalling a candidate mapping pair set; calculating character string similarity, synonym expansion matching score, data type consistency score and sample value pattern consistency score for the candidate mapping pairs; generating a neighbor confusion risk index based on semantic neighbor density, entity category distribution deviation, synonym expansion and context anti-verification, performing risk gating correction, and judging a score in combination with a difficult case; and applying table-level consistency constraints in the same service table or entity domain to carry out joint consistency solution to obtain a table-level optimal mapping scheme, generating confirmation confidence, and outputting an automatic confirmation or manual recheck mapping result.

The invention provides a course dynamic optimization method and system, an electronic device and a storage medium, and relates to the technical field of online education, and the method achieves the quantitative association of a course unit-knowledge fragment through a soft mapping matrix, enables the perplexity to be bidirectionally transmitted between the course unit and the knowledge fragment, and achieves the quantitative traceability. The comprehensive confusion degree is obtained by fusing the confusion degrees of the course unit layer and the knowledge fragment layer, so that the high-confusion course unit can be positioned more accurately and explainably, and fine optimization of the course can be realized; through a course optimization decision with cost constraint, automatic generation of a course optimization scheme with controllable cost is realized; besides, personalized learning path optimization of a course structure layer is realized, and a course optimization strategy continuously acts on subsequent learners, so that dynamic evolution of course contents and personalized learning paths is realized.

The invention relates to the technical field of mobile application continuous integration automation, and discloses a multi-channel APP automatic packaging system which comprises the steps that byte code files in a released mobile terminal APP installation package are acquired, and an occupied symbol space set is generated; generating an available confusion namespace by using set complementary operation; extracting method-level code change by using an abstract syntax tree difference algorithm, and generating a change method set; generating a patch special confusion rule for the change method set by using a namespace allocation algorithm; compiling and obfuscating the change method set by applying a patch special obfuscating rule to generate a safe obfuscating patch byte code; performing packaging and signature operation on the safe obfuscation patch byte code, and outputting a minimized obfuscation safe hot repair patch pack; according to the method, the technical problems that class definition conflicts or method calling errors occur after the hot repair patch is loaded, and the downloading success rate of a user is affected due to the overlarge size of the hot repair patch are solved.

The invention relates to the technical field of elevator traction system parameter optimization design, discloses an elevator traction system parameter driving optimization method, and aims to solve the problems that existing design depends on experiences of engineers, constraints are once, soft constraint violation is difficult to quantify, discrete parameters are poor in adaptability, and multi-objective decision making is disordered. Comprising the following steps: presetting engineering constraint conditions and a discrete design parameter library containing wheel bodies, rope bodies and matching parameters, and dividing constraints into hard constraints and soft constraints according to a safety compliance correlation degree; constructing a penalty cost model, and quantifying a soft constraint violation quantity and a weight coefficient to output a penalty cost; generating a full-amount parameter combination and verifying hard constraints, and reserving an effective combination; and a comprehensive cost objective function is constructed, and an optimal combination with the minimum comprehensive cost is screened in combination with the physical cost and the penalty cost. Safety compliance is guaranteed, cost optimization space is expanded, quantitative decision basis is provided, optimization efficiency is improved, the scheme is ensured to adapt to engineering practice, and the method is suitable for design and transformation of the elevator traction system.

The invention relates to an anatomy intelligent tutoring system and method based on term alignment and multi-modal verification. The method comprises the following steps: acquiring multi-modal data of a learner and implementing time axis alignment and term alignment to generate a standardized time sequence data set for feature extraction; inputting the feature vector into a long short-term memory network model, and analyzing the time sequence relevance to generate a cognitive state vector; when the conceptual error probability in the cognitive state vector exceeds a threshold value, extracting a term set to be verified based on a three-dimensional anatomical model interaction log; the term pair with the highest confusion degree is accurately positioned as an error attribution result by performing knowledge graph topological correlation analysis on the term set and calculating the confusion degree between the terms; and finally, in combination with the historical ability portrait of the learner, performing dynamic matching from the multi-modal strategy library to generate personalized tutoring content, and embedding the personalized tutoring content into a three-dimensional model interface to execute intervention, so that the effects of automatically identifying anatomical concepts from the multi-modal behavior data to understand error roots and generating targeted tutoring strategies are realized.

The invention discloses an encrypted traffic adaptive update classification method and system oriented to an open network environment. The method comprises the following steps: firstly, extracting endogenous semantic features and exogenous environment features based on a causal decoupling mechanism, and stripping an environment confusion factor through anti-fact disturbance and invariance constraint to obtain invariant semantic representation; constructing a macroscopic drift state vector representing a network situation, and inputting a trained meta-learning super-network intelligent decision adaptive control hyper-parameter; performing cross-modal element calibration by utilizing large language model thinking chain reasoning, and calculating the subspace direction consistency of an instantaneous gradient vector of a candidate sample and a category optimization trajectory prototype so as to screen credible samples; and in combination with the capacity-limited playback queue, gradient orthogonal projection constraints are introduced to update low-rank adaptation layer parameters. According to the method, the concept drift problem is solved through causal decoupling and meta-learning decision, forgetting prevention is achieved through orthogonal projection updating, and the online adaptability and robustness of the model in the open environment can be improved without additional manual annotation.

The invention relates to a model training method and device, computer equipment, a storage medium and a program product. The method comprises the steps of obtaining at least one symbol set from a medical special symbol library, and generating a corresponding training sample set according to each symbol set; for each symbol set, testing the response of the medical decision model to the symbol set through the training sample set corresponding to the symbol set to obtain response data corresponding to the symbol set; determining error distribution of error response of the medical decision model to the symbol set according to the response data corresponding to the symbol set, and calculating a confusion entropy value corresponding to the symbol set according to the error distribution; grading the symbol sensitivity of the medical decision model to at least one symbol set according to the confusion entropy value corresponding to each symbol set, and determining a symbol set scheduling strategy according to a sensitivity grading result; and performing reinforcement learning training on the medical decision model according to the symbol set scheduling strategy. By adopting the method, the performance of the medical decision model can be improved.

The invention discloses a low-voltage uninterruptible power operation personnel risk early warning method fused with space-time behavior analysis, and relates to the technical field of intelligent early warning of electric power operation risks, and the method comprises the following steps: collecting the position coordinate data of a plurality of operation personnel at each time point in an operation area in real time, and constructing a space trajectory matrix; in a set time window, continuous attitude data acquisition is carried out on each operator, and the data is converted into a standardized action feature vector. According to the method, a self-adaptive multi-modal recognition mechanism is formed by constructing a space trajectory matrix and an action feature matrix, fusing position and attitude information, quantifying the behavior confusion degree between personnel, introducing a behavior confusion index to dynamically adjust a modal weight, and enhancing individual difference recognition in combination with voice positioning. Compared with a traditional method, the method can effectively solve the problem of affiliation confusion in high-overlap and high-similarity scenes, achieves the precise affiliation of high-risk behaviors, and remarkably improves the accuracy and real-time intervention capability of an early warning system.

The invention discloses a multi-mode voice interaction system and method for an intelligent cabin. Firstly, various vehicle working condition parameters in a CAN bus are read in real time, a wavelet domain Wiener filtering and judgment statistic enhancement strategy is dynamically adjusted, and self-adaptive voice purification based on vehicle state prior is achieved. Secondly, a bidirectional long-short-term memory network is adopted as a semantic understanding core, semantic feature vectors are extracted from enhanced signals through segmented maximum pooling operation, and the problem of semantic confusion caused by manifold collapse in a traditional method is effectively solved. And the feature vector and a safety state signal are synchronously input into an improved LLM-CoT engine and a dynamic arbiter, so that context accurate understanding and safety-up priority response are realized. According to the method, vehicle working condition perception and biLSTM depth time sequence modeling are deeply adapted to a vehicle-mounted complex acoustic scene, the error rate of semantic understanding is remarkably reduced in actual measurement under multi-person dialect dialogue, and the accuracy, robustness and safety of intelligent cabin interaction are systematically improved.

The invention relates to the technical field of artificial intelligence volunteer filling, and discloses an artificial intelligence-based volunteer filling recommendation method. The method comprises the steps of obtaining academic score data and interest preference data of a target student, and determining an initial feature vector of the target student according to the academic score data and interest preference data; meanwhile, acquiring a professional information database of the target school, and constructing a professional feature model based on the database; calculating a matching degree matrix between the target student and the specialty according to the initial feature vector and a specialty feature model; and generating a volunteer filling recommendation sequence based on the matching degree matrix, and transmitting the recommendation sequence to the user terminal. According to the method, personal characteristics of students and professional information are deeply combined with the help of an artificial intelligence technology, personalized recommendation is provided for volunteer filling, blindness in the filling process is reduced, the students are helped to more accurately select adaptive majors and schools, the volunteer filling rationality and efficiency are improved, and the volunteer filling experience is improved. And the information combing pressure and decision confusion of students and parents in voluntary selection are relieved.

The invention relates to the field of man-machine interaction, and particularly provides a task processing method, device, equipment and system, a medium and a product. The method comprises the following steps: in response to a first session instruction of a user for a target type task, executing the target type task; the target type task belongs to a task type capable of receiving supplementary information input by a user during task execution; in the task execution period of the target type task, responding to a received second session instruction input by the user, and executing a corresponding task operation according to the intention of the second session instruction; wherein the intention is used for indicating to input supplementary information of the target type task or initiate a new task. Therefore, on the premise of not interrupting the original target type task, the new task instruction of the user can be responded in parallel or the supplementary input of the user to the target type task can be processed, and the problems of interaction blockage and context chaos caused by the exclusive session state under the existing architecture are solved.

The invention relates to the technical field of advertisement prediction methods, in particular to an advertisement overflow value prediction method and device and a storage medium, and the method specifically comprises the following steps: 1, constructing user-behavior-scene-environment four-dimensional data collection, and synchronously carrying out cleaning and standardization processing; step 2, depth feature engineering construction including time sequence behavior features, behavior interval features, behavior intensity features, cross-platform association features and user heterogeneity features; step 3, an overflow behavior prediction model based on Transform is constructed; step 4, user heterogeneity adaptation based on meta learning; 5, introducing a causal inference method to quantify the influence of external factors on advertisement overflow; according to the method, for the causal confusion problem, a causal attention mechanism and tendency score matching method is adopted, the association between advertisement exposure and natural conversion is accurately stripped, the real influence of advertisement overflow is clearly defined, and value misjudgment of a traditional attribution method is avoided.

The invention provides a confused expression recognition method and device fusing lightweight double-branch attention, equipment and a storage medium. Relates to the technical field of image classification processing. The method comprises the steps of adopting a lightweight neural network MobileNetV3 as a backbone network, and performing preliminary feature extraction on an input face image to obtain a feature map; respectively inputting the feature map into the first branch and the second branch; the first branch is based on a KRSA-ViT module of sparse attention of a target area, modeling a global dependency relationship between face target areas, and obtaining a first feature; the second branch fuses channel attention and space attention to enhance local feature expression based on a grouping attention rearrangement module to obtain a second feature; and fusing the first feature and the second feature to obtain a fused feature, inputting the fused feature into a full connection layer, and outputting a confusion expression classification result. According to the invention, the confusion expressions of students in the classroom teaching scene can be accurately identified.

The invention provides a root fault positioning scheme applied to an industrial system. In the scheme of the invention, confusion factors corresponding to initialized unobservable equipment in a preset historical time period and causal relationship diagrams corresponding to all the equipment are configured as a model of learnable parameters, and the model is subjected to multi-round iterative training until convergence so as to obtain the causal relationship diagrams of all the equipment; and then obtaining a root fault positioning result of the fault equipment in the industrial system when the fault occurs based on the causal relationship graph corresponding to all the equipment in the industrial system. According to the technical scheme of the invention, the more accurate causal relationship among the devices in the industrial system can be obtained, so that the reliability of the root fault positioning result of the industrial system is improved, and the problem of false alarm or missing report in root fault positioning is avoided.

The invention discloses a TEE-GPU collaborative model credible training method and device based on parameter confusion, and the method comprises the steps: enabling a client to encrypt training data and a model architecture in a preprocessing stage, and uploading the encrypted training data and model architecture to a server; and the trusted execution environment of the server side decrypts the model architecture, carries out confusion processing on a linear layer and then deploys the linear layer on an external GPU (Graphics Processing Unit). In the training stage, initial forward propagation calculation of training data is firstly completed in the TEE, and then intermediate results are confused and then transmitted to the GPU so as to execute subsequent forward propagation calculation. In the back propagation process, extra confusion is applied to the gradient by the TEE, and the gradient is issued to the GPU to calculate a new gradient; and after receiving the confusion gradient, the GPU updates the parameters in a confusion form. And randomly sampling calculation data in the TEE, and carrying out integrity verification to ensure the integrity of a calculation result. Compared with an existing credible training scheme, the method has the advantages that the time overhead can be remarkably reduced while the calculation accuracy, integrity and privacy of the model are ensured.

The invention discloses a tumor medical record time axis generation method based on multi-modal recognition and dynamic cue words, which thoroughly solves the problem of disordered stage division of a traditional method by introducing a dynamic cue word mechanism into an operation discrimination node. Through a multi-mode OCR + NLP fusion strategy, the accuracy of medical record information extraction is greatly improved; the medical compliance of the time axis is ensured through time sequence verification and field completion driven by the knowledge graph; and through front-end mind map or medical history outline visualization, the complex medical record can be clearly displayed. Practice shows that compared with manual arrangement, the time consumption of sequential output of the complete medical record is reduced by more than 90%, the item missing rate is reduced by more than 80%, and the method can be widely applied to scenes such as clinical decision support, follow-up visit management, scientific research data integration and medical insurance quality control.

The invention provides a data security sharing method under an AI platform, and belongs to the technical field of AI platforms, and the method comprises the steps: building a distributed architecture of a confusion center node and a plurality of confusion edge nodes in an AI platform management server, carrying out the knowledge vector extraction of to-be-shared data, and converting the to-be-shared data into a high-dimensional vector for representation; generating a plurality of ellipsoidal confusion bubbles based on vector distribution characteristics and constructing a multi-dimensional authorization strategy, adopting a game boundary model comprising an upper-layer game model and a lower-layer game model to perform separation processing on a crossed and overlapped region, establishing a data training reasoning mapping relation according to a game solving result, and calculating confusion parameters through a confusion intensity optimization function; and finally, real-time confusion protection in the data transmission process is executed, and data flow is continuously monitored. The technical problem of data leakage risk caused by mismatching of a confusion protection strategy and an actual security demand in the data sharing process under the AI platform is solved.

The invention discloses a model generalization ability optimization method and system based on hybrid experts, and aims to solve the problems of weak generalization ability, difficulty in knowledge migration, low semantic alignment efficiency and the like when a multi-modal large model performs fine adjustment on downstream tasks in a new field. The method comprises the following steps: embedding a hybrid expert module in a low-rank adaptive fine tuning framework of a self-attention module in the visual encoder, introducing a heterogeneous hybrid convolution structure, innovatively realizing transverse and longitudinal independent scaling through bilinear interpolation in the convolution structure, supporting a multi-scale convolution kernel, and realizing dynamic combination through an MoE framework; constructing a cross-modal expert adapter of channel perception, and adopting dynamic channel recombination, cross-modal conditional convolution and weighted spatial feature aggregation methods; according to the method, the problem of local semantic confusion can be well solved, meanwhile, semantic fusion of image features and text features is enhanced, more effective spatial information and priori knowledge are injected into the model, and therefore the generalization of the model is improved.

The invention relates to the technical field of intelligent models, and discloses an image generation method and device, computer equipment, a storage medium and a program product. Similar preset prompt words, a potential classification list of a to-be-generated image and confusion category pairs are determined on the basis of original prompt words, so that the confusion category pairs are utilized to carry out classification correction; misjudgment of the subject category of the to-be-generated image is avoided, and the accuracy of the subject category is improved. And if the theme category is a character theme, optimizing the original cue word through the scene type of the to-be-generated image to obtain a target cue word, so that the image generation model performs image generation by using the target cue word with rich details and the scene type. When iteration of the image generation model is not finished, face restoration is carried out on the intermediate image generated by the image generation model, so that generation time prolonging caused by image restoration after iteration is finished is avoided, image generation efficiency is improved, a face collapse phenomenon is avoided, and image generation quality is improved.

The invention belongs to the field of equipment fault diagnosis, and discloses an unsupervised cross-equipment fault diagnosis method fusing dual alignment and pseudo-label learning. Through collaborative optimization of three mechanisms of global domain adaptation, conditional domain confrontation and pseudo-label learning, depth feature alignment and refining of three levels of global-local-instance are realized, and the accuracy, robustness and generalization ability of unsupervised cross-equipment fault diagnosis are significantly improved. The core technical problems of confusion of different fault category features on a target domain and low diagnosis precision due to the fact that an existing domain adaptation method only pays attention to global distribution alignment and ignores a category structure are solved, and the technical bottleneck that when a traditional intelligent diagnosis model is applied to new equipment or new working conditions, the performance can be guaranteed only by depending on label data is solved.

The invention belongs to the technical field of computer information security, particularly discloses a content security protection method and device based on semantic consistency, and aims to solve the problem that high-level cue word injection attacks are difficult to effectively recognize in the prior art. Carrying out real-time interception on dominant illegal texts through a content filtering module; quantizing generation rationality differences of prompt words between attack and normal language models by using a word vector confusion degree calculation module; evaluating the logic coherence of the sentence structure of the prompt word through a statement semantic consistency judgment module; integrating the multi-dimensional feature data and carrying out risk classification by adopting a machine learning algorithm; and routing the suspected attack request to a value fine tuning model for processing according to a judgment result. According to the technical scheme, high-precision recognition and response to complex cue word injection attacks are achieved, and the content safety protection capacity and compliance guarantee level of a large model in an interaction scene are remarkably improved.

The invention discloses a causal modeling-based remote sensing instance segmentation method, which comprises the following steps of: explicitly describing a causal relationship among a target, a context and a segmentation mask by establishing a structural causal model (SCM); causal intervention is simulated through a local object transformation module, and the confusion influence of non-causal variables is eliminated; a large segmentation model (SAM) is guided to pay attention to causal correlation features in a model training stage through attention invariance constraint, so that a more accurate image segmentation result is realized. In the invention, causal interpretation is strong: a structural causal model is introduced into a remote sensing segmentation task for the first time; eliminating non-causal interference: simulating causal interference through a local object transformation module, and weakening the influence of false correlation factors such as background and appearance; improving feature consistency: ensuring that the attention of the model is consistent with the same object in different environments through attention invariance constraint; generalization ability is enhanced, the method is obviously superior to a mainstream method on various remote sensing data sets, and average segmentation precision is improved by 3%-5%.

The invention discloses a dynamic random language input generation method based on large model architecture and training data features, and relates to the technical field of large model testing and natural language processing crossing. Comprising the steps of obtaining a target large model type, training data and a test scene demand, and performing large model-classical feature collaborative preparation; the noise tolerance is corrected by training data field concentration degree and grammar specification degree quantification in combination with a model type coefficient, and the confusion degree is finely adjusted according to a trigger rate; effective conflicts are screened by means of cross-modal consistency loss, and input effectiveness is verified in a three-dimensional mode through the trigger rate, the strength and the influence degree; calculating power and resource efficiency quantitative feedback iteration is carried out, stage adaptive input is generated based on time sequence features, and cross-task feature multiplexing is realized through task similarity; performing low-sensitive word disturbance and confrontation-random balance enhancement as required, verifying a result, and returning to a parameter adaptation link for readjustment if the result does not reach the standard; finally, high input adaptation and high-efficiency generation are realized, and the test is ensured to be accurate and efficient.