752 results about "Lexicon" patented technology

The invention relates to the technical field of large language models, and discloses a voice data interaction feedback control processing method based on a large language model. The method comprises the following steps: acquiring an industrial voice instruction through an AMR main controller, and obtaining a standardized vector through industrial lexicon matching and intention classification; inputting an industrial large language model for reasoning processing, and generating an AMR execution scheme; carrying out distributed coordination and task allocation on the AMR cluster to form a control instruction sequence; and the motion controller executes monitoring, processes exceptions and outputs a feedback strategy. Accurate classification and semantic understanding of complex industrial instructions are achieved, the processing capacity of professional knowledge in the industrial field is improved, and meanwhile high-precision real-time state monitoring is achieved.

The invention relates to a knowledge-enhanced large model data analysis agent visualization method, which comprises the following steps of: performing syntactic analysis by using a PCFG algorithm, identifying syntactic components and relationships of sentences, and splitting the sentences into sub-sentences with complete logic; extracting keywords and entity words of the sentences by combining a word segmentation technology with a business word segmentation library; recommending and summarizing candidate segmented words through a recall strategy, carrying out rough arrangement, fine arrangement and fine arrangement on the candidate segmented words, and correcting and complementing a segmented word result; splitting and analyzing the natural language problem based on the fine-tuned large model, and generating SQL statements meeting business requirements; analyzing the data set generated by the SQL, extracting field information and business meanings, and identifying an association relationship between fields; and automatically recommending chart styles according to data features and user preferences, assembling an analysis report and distributing the analysis report according to a specified frequency. According to the method, knowledge enhancement and large model technologies are combined, so that the efficiency and effect of data analysis and visualization are remarkably improved.

The invention provides an agricultural field large language model training method and device and a medium, and the method comprises the steps: collecting agricultural field data book information, and carrying out the preprocessing of the information, and obtaining a first Chinese data set and a first minority language data set, collecting the first Chinese data set and the first minority language data set, generating a first mixed language data set, processing the first mixed language data set by adopting a word segmentation library, performing secondary processing by adopting a preset special character processing method, constructing a word segmentation list, and constructing a WordPiece word piece model to process and divide the word segmentation list; converting each word in the word segmentation list into a word vector by adopting a word embedding algorithm; the method comprises the following steps of: constructing a coding component and a decoder component which are based on a Transform architecture and adopt a self-attention mechanism, generating an agricultural field large language model, and training the agricultural field large language model by taking a word vector as an input to obtain a trained agricultural field large language model. The problem of supporting a large language model of Chinese and minority languages is solved.

Systems and techniques are provided for automatically analyzing and processing domain-specific image artifacts and document images. A process can include training a domain-adapted ML network using a domain-specific training dataset including a plurality of training data inputs corresponding to a lexicon of domain-specific terminology. A first fine-tuning training of the domain-adapted ML network yields a domain-adapted general QA ML network, based on using a first question answering (QA) dataset comprising a first plurality of question-answer training pairs that do not correspond to the lexicon of domain-specific terminology. A second fine-tuning training of the domain-adapted general QA ML network yields a fine-tuned domain-adapted general QA ML network, and uses a second QA dataset comprising a second plurality of question-answer pairs generated based on a corpus of text narratives utilizing the lexicon of the domain-specific terminology.

A system for improving employee engagement and automating the integration of business processes. The system includes: a user interaction module configured to receive multimodal inputs, including natural language text and voice requests from employees via enterprise portals, mobile applications, voice-activated devices, and collaboration platforms, and to generate real-time responses via conversational output channels; a context processor that is operationally coupled with the user interaction module, wherein the context processor stores the interaction history, employee preferences, organizational role metadata, and task results in both short-term and long-term memory and dynamically derives context for controlling responses and initiating workflows; A natural language and intent processor that is communicatively linked to the context processor. The intent engine consists of large language models trained on company-specific lexicons to perform intent detection, entity extraction, ambiguity resolution, and sentiment or urgency classification from employee queries; a workflow orchestration layer in communication with the intent engine and the context processor, wherein the workflow orchestration layer includes a rule-based and AI-powered process execution engine configured to automatically initiate, route, and complete cross-functional business tasks, including approvals, escalations, and compliance checks, with workflows defined using modular templates that include conditional logic and time-based triggers; Webhooks and authentication protocols provide secure interoperability between the workflow orchestration layer and external enterprise platforms; a feedback and learning module that is operationally coupled with the workflow orchestration layer and the natural language understanding engine, wherein the feedback and learning module is configured to analyze task completion rates, response accuracy, latency metrics, and user feedback signals, and to retrain underlying language and workflow models for adaptive improvement in real time; and an administration console with role-based access controls, compliance dashboards, audit trails and interfaces for workflow configuration, whereby the administration console enables authorized personnel to monitor system operations, adjust interaction rules and enforce data protection restrictions across departments.

The invention provides a large model text generation method and system based on self-adaptive cue words, and belongs to the technical field of large model text generation. The technical problem to be solved is that the current large language model text generation has the defects of selection deviation and uncertainty. According to the technical scheme, hierarchical clustering processing is carried out on problems contained in a large-scale multi-task language understanding MMLU data set, a text embedding and similarity calculation matrix is adopted, and a plurality of problem clusters containing similar features are constructed and obtained; randomly extracting representative problem examples from each problem cluster, and constructing excellent cue words with remarkable guide effects for each problem based on domain knowledge and potential semantic association; a Qwen model is adopted to generate candidate cue words with diversity and pertinence for various question clusters divided in advance, a cue word bank is constructed, and each question and cue word pair in the cue word bank is manually scored; the method is applied to text generation.

The invention discloses a sensitive word auditing method, which belongs to the field of auditing, and comprises the following steps of: firstly, preprocessing an input text, such as removing spaces, symbols, stop words and the like, so as to generate a word segmentation result; thirdly, constructing a dynamic synonym and a deformation word library, combining user behaviors and NLP technology updating, and expanding word meanings by applying a word vector model Word2Vec; then, based on BERT, a text context is analyzed, and metaphors and biguan languages are detected. Performing sensitive word primary screening and secondary screening by using a static word bank and an SVM (Support Vector Machine), and analyzing complex semantics by LSTM (Long Short Term Memory); the model is optimized based on a feedback self-learning mechanism, difficult contents are manually checked, an analysis report is generated, and finally a word bank and an algorithm are periodically updated to adapt to hotspots. The method has the beneficial effects that sensitive word auditing is improved through a self-learning mechanism, synonyms and deformed word banks can be dynamically updated, and private expressions are identified. BERT and LSTM model multi-level auditing are adopted, secondary screening is carried out in combination with SVM, and recognition of complex contexts and biguan languages is improved.

The invention discloses an intelligent medical and invasive project evaluation method and system based on natural language processing. The method comprises the steps that professional corpora of multiple sources of an intelligent medical and invasive project are collected; the knowledge graph construction module comprises a term library construction module, a term clustering module and a relationship construction module; taking the professional corpus as a core to construct a professional word library; segmenting the professional lexicon into a plurality of clusters with professional terms as the center; constructing a knowledge graph by taking vocabularies in the clusters as nodes; and constructing an intelligent evaluation module which comprises an intelligent scoring module and an intelligent evaluation module, and outputting AI comments. According to the invention, the project plan is automatically processed through a natural language processing technology, so that the auditing efficiency is greatly improved; and in combination with a knowledge graph technology, a professional knowledge base is systematically constructed, background knowledge support is provided for a large language model, and the professionality and consistency of evaluation results are ensured.

Methods and systems for generating training data for training a contrastive language-audio machine-learning model. A plurality of audio segments are retrieved from a speech emotion recognition (SER) database along with metadata associated with the audio segments. The metadata of each audio segment includes an emotion class. Words or terms associated with emotions are retrieved from a lexicon. A large language model (LLM) is executed on (i) the classes of emotion associated with the audio segments and (ii) the words or terms from the lexicon. This generates a plurality of text captions associated with emotion, which are stored in a caption pool. For each audio segment retrieved from the SER database, that audio segment is paired with one or more of the text captions from the caption pool that were generated based on the emotion class associated with that audio segment. This yields audio-text pairs for training a contrastive learning model.

The present application relates to the technical field of electronics, and provides a method for generating a commodity label on the basis of a large language model, and an electronic device. The method specifically comprises: acquiring commodity description information related to a target commodity; performing structured processing on the commodity description information to obtain an objective property feature of the target commodity; and using a fine-tuned large language model to process the objective property feature on the basis of a language prompt lexicon to obtain one or more commodity labels of the target commodity. According to the method, a commodity selling point label can be generated by means of the commodity description information acquired from multiple sources and multiple paths, the source of the commodity description information is expanded, commodity information of different sources can be identified and acquired by means of image and text details, etc., which can assist the model in better understanding the commodity, and generating a more reliable selling point label having a higher confidence level and greater traceability.

The invention discloses a receipt identification method and device, equipment and a medium, and relates to the technical field of bill content identification. The method comprises the following steps: acquiring a receipt image through a preset acquisition mode, and identifying the receipt image by using a preset optical character identification model to determine text content and corresponding position information in the receipt image; performing information matching on the receipt image in a preset multi-mode large model prompt word bank to obtain a first prompt word, and inputting the first prompt word, the text content and the position information into a preset multi-mode large model to obtain a key information key value pair; and performing information matching on the receipt image in a preset language large model prompt word library to obtain a second prompt word, and inputting the second prompt word, the text content, the position information and the key information key value pair into a preset language large model to obtain a target key information key value pair. Through the technical scheme of the invention, the accuracy of identifying the key field content of the non-standard receipt can be improved.

The invention relates to the technical field of large models, in particular to a large model cue word automatic optimization method and device and electronic device.The large model cue word automatic optimization method comprises the steps that a question input by a user is obtained, and a plurality of candidate cue words are generated based on the question; performing effect evaluation on each candidate cue word by utilizing a large model, performing descending sorting on each candidate cue word according to an effect evaluation result to determine a plurality of optimal cue words, and storing the optimal cue words; carrying out iteration on the optimal cue word in the cue word bank, carrying out distillation evaluation on the optimal cue word after iteration, and if a distillation evaluation result reaches a set threshold value, updating the number of iterations; and performing descending sorting on the optimal cue words completing the set number of iterations according to the distillation evaluation result, and outputting the first optimal cue word as a final cue word. According to the method, the cue word can be automatically optimized, the cue word quality is improved, and the key problem generation accuracy of a large model is improved.

The invention relates to a retrieval method and device based on static word embedding, computer equipment, a computer readable storage medium and a computer program product. The method comprises the steps of performing word segmentation on an original text to obtain a first word segmentation result, training a static word embedding model by utilizing the first word segmentation result to obtain word vectors, and generating a synonym word library; the method comprises the following steps: establishing a full-text inverted index by utilizing an original text, and expanding query words by utilizing a synonym library in a retrieval stage; encoding the original text into a semantic vector by using a semantic generation model, and constructing a vector index based on the semantic vector; based on a to-be-queried text in a user query request, performing retrieval by using the full-text inverted index to obtain a first candidate document, and performing retrieval by using the vector index to obtain a second candidate document; performing fusion processing on the first candidate document and the second candidate document to obtain a target candidate document; and inputting the target candidate document into the text generation model to obtain a retrieval result. By adopting the method, the accuracy of text retrieval can be improved.

The invention discloses a big language model generation content security test system and method in a black box scene. The system comprises a jailbreak prompt word library module used for storing jailbreak prompt words for performing security test on a big language model; the violation question and answer pair module is used for storing violation question and answer pairs covering different types; the response acquisition module is used for obtaining a data request packet according to query content formed by the jailbreak prompt word and the query request; the security analysis module is used for calculating the similarity between response data corresponding to the query request and an expected violation answer, taking the similarity as a security score, and inputting the security score into the adaptive optimization module; and the adaptive optimization module is used for optimizing the jailbreak prompt words output by the jailbreak prompt word bank module by using a genetic algorithm according to the security score output by the security analysis module. According to the method and the device, the security of the large language model generation content can be effectively tested.

The invention relates to a steel bridge disease detection and identification method based on a large language model, and belongs to the technical field of artificial intelligence and civil engineering crossing. According to the method, a cross-modal feature alignment mechanism is constructed through a pre-trained multi-modal large language model by fusing a steel bridge image and a field customized text prompt, and a cascade detection process of'component identification-disease classification-region segmentation 'is realized. Comprising the following steps: designing a structured text prompt word bank to enhance semantic consistency, and dynamically fusing general knowledge and instance features in combination with a mixed prompt mechanism; a multi-level cross-modal alignment strategy is adopted to generate an anomaly graph, and a disease area is accurately positioned; a visual prompt enhancement module is introduced to improve the multi-scale feature discrimination ability, and the robustness in a complex environment is adjusted and optimized through data self-adaption. Under the condition of few samples or even zero samples, high-sensitivity detection and pixel-level segmentation of steel bridge cracks, corrosion and other diseases are achieved, and the problems that a traditional method is low in efficiency, poor in generalization, high in labor cost and the like are effectively solved.

The invention discloses a privacy protection text generation method and system based on a large language model. The method comprises the following steps: presetting a sensitive word library with sensitive word scores; performing sensitive word matching on each word in the input text and the sensitive word library, calculating sensitivity scores of the matched sensitive words, and summing the sensitivity scores to obtain a total sensitivity score of the input text; combining the total sensitivity score with a grading threshold to obtain a sensitivity level of the input text; establishing a mapping function between the sensitivity level and a differential privacy parameter, and automatically generating the differential privacy parameter in a differential privacy mechanism according to the sensitivity level; and adaptively controlling the privacy protection intensity of the text generated by the large language model by using a differential privacy parameter corresponding to the sensitivity level. According to the method and the system disclosed by the invention, the privacy protection intensity can be effectively adjusted when different sensitive information is coped with, the generation fluency is ensured to be kept in the low-sensitive content, stronger privacy protection is provided in the high-sensitive content, and accurate privacy control is realized.

The invention provides a sensitive information detection method and content display method in a question and answer scene, and the method comprises the steps: obtaining the current output information of a large language model, and storing the output information to a preset buffer area; wherein the output form of the output information is streaming output; taking the output information and original information stored in a preset buffer area as a current to-be-processed text; in response to the fact that the data length of the current to-be-processed text exceeds a preset length, taking the current to-be-processed text as a target detection text; wherein the preset length is determined according to a pre-constructed sensitive word library; and performing sensitive word detection on the target detection text to obtain the detection result, thereby improving the detection speed of the sensitive information, reducing the waiting time of the user and improving the user experience.

The invention relates to an A / B test question answering method and device based on a large model agent, and relates to the technical field of large models, agents and computers. The method comprises the following steps: displaying an intelligent interaction page associated with an intelligent agent; displaying a plurality of prompt word options in response to a trigger operation on a prompt word bank in the intelligent interaction page; in response to a trigger operation on a target cue word option in the plurality of cue word options, at least displaying a target cue word template corresponding to the target cue word option; in response to a content editing operation on the target cue word template, filling a target content corresponding to the content editing operation into the target cue word template to obtain a target cue word; and in response to a sending operation on the target cue word, displaying a target answer of the intelligent agent to the target cue word through the large model on the intelligent interaction page. The expression logic of the user for the intention can be optimized through the cue word, the content quality of questions is improved, then the accuracy of answers generated by the large model is improved, and the use experience of the user is improved.

The invention discloses a component identifier matching method and system based on NLP semantic segmentation and a multi-level word library, and relates to the field of constructional engineering informatization, and the method comprises the steps: building a standard main word library and a preset compensation word library, carrying out the semantic segmentation of an input component identifier, and obtaining a segmented lexical element sequence, calculating semantic similarity and morphological similarity between the segmented lexical element sequence and entries in a standard main word bank, and performing standard term mapping based on the semantic similarity and the morphological similarity to obtain a standard term mapping result; carrying out structured data conversion on the segmented lexical element sequence which does not complete the standard term mapping by utilizing a regular expression mode to obtain structured data, and carrying out AI extension library matching on the segmented lexical element sequence which does not complete the structured data conversion to obtain a synonym matching result; and combining the standard term mapping result, the structured data and the synonym matching result to obtain component identification description. By implementing the method, the matching accuracy of the component identifier names in the engineering project can be improved.

The invention relates to a text extraction data similarity matching method based on semantic similarity, and belongs to the field of natural language processing. According to the method, entity data is obtained and processed to serve as a standard word bank of the entity data, initial semantic similarity model training is conducted, candidate texts in documents are selected, document topics and distribution of the document topics are obtained, and the document categories are determined by calculating the topic similarity between the document topics and all event fields. Creating an event extraction template, extracting the document data to the corresponding event extraction template according to the document subject category, converting the unstructured data into structured data, identifying entity data from the structured data as a to-be-matched text, and determining the similarity between the to-be-matched text and the target text by using a semantic similarity model. According to the method and the device, automatic matching of the noun text and the standard entity text in the unstructured document is realized.

The invention belongs to the technical field of sea wave height prediction, and discloses a multi-scale sea wave height prediction method and system based on a large language model, and the method comprises the steps: S1, carrying out the decoupling interaction based on an amplitude phase, and obtaining enhanced sea wave height data; s2, extracting based on multi-scale semantic information: constructing a feature pyramid structure for the enhanced sea wave height data, outputting sea wave self-encoding features of different scales, performing self-encoding on a lexicon of a large language model, outputting word vector features, and obtaining the multi-scale semantic information by using a cross-modal attention mechanism and a dynamic fusion mechanism; and S3, a step of large-scale language model prediction based on sea wave height domain cue word driving: inputting the multi-scale semantic information and the domain specific semantic features into a large-scale language model, and outputting prediction features. According to the method, sea wave complex spatial-temporal change information is effectively mined, and accurate prediction of the sea wave height is realized.

The invention provides a multi-robot cooperative control method and system based on edge computing, and relates to the technical field of multi-robot control, and the method comprises the steps: dividing a search and rescue region into a plurality of sub-regions, generating a sub-region search and rescue task, and carrying out the search and rescue of the sub-regions according to the sub-region search and rescue task and the capability parameters and current states of ground robots. An optimal matching scheme of ground robots and sub-region search and rescue tasks is generated, different search and rescue strategies are adopted for different types of ground robots, and a heterogeneous robot cooperative search and rescue strategy is generated by defining formalized languages and lexical methods for cooperative control of the ground robots; each ground robot executes a search and rescue task in the corresponding sub-region according to the heterogeneous robot cooperative search and rescue strategy, when it is detected that the resource use condition or the network condition meets the preset condition, a calculation unloading decision is started, part of the search and rescue tasks are unloaded to other calculation nodes according to the decision result, and the search and rescue action continues to be executed. And the search and rescue task is completed.

The invention discloses a marketing mobile terminal dynamic behavior compliance monitoring method and system, belongs to the technical field of information transmission monitoring, and aims to solve the problems that existing monitoring is insufficient in non-text information processing, redundant in sensitive word judgment, incomplete in violation expression variant coverage and low in data reuse rate. The method comprises the steps that dynamic behavior information is captured, a composite retrieval code retrieval history library is constructed, and whether sensitive word recognition is conducted or not is judged; if not, generating a standardized word segmentation sequence; comparing the segmented words to determine matched segmented words and unmatched segmented words, calling a historical synonym set of the matched segmented words, positioning core semantics of the unmatched segmented words, screening evading expressions to generate a synonym set of the unmatched segmented words, and constructing a synonym data set; calling a matched word segmentation historical judgment result, establishing a synonym sequence calculation probability, and judging a sensitive word in combination with a significant proportion; executing a processing flow and updating the history library and the word segmentation library; according to the invention, efficient and accurate compliance monitoring is realized, the data reuse rate is improved, and the effectiveness of a monitoring closed loop is guaranteed.

The invention discloses an English vocabulary intelligent learning path planning system and method based on reinforcement learning, and relates to the technical field of education learning, and the system comprises a learning path preliminary planning module, a learning equipment feedback module, a learning effect analysis module and a learning path updating module. According to the English vocabulary intelligent learning path planning system and method based on reinforcement learning, the basic vocabulary test result, the planned learning time and the target of the learner are obtained, and personalized planning of the learning path is achieved in combination with dynamic lexicon matching and learning starting point setting. A reinforcement learning model is adopted to dynamically divide vocabulary tasks, the vocabulary difficulty gradient, the use frequency and a multi-modal exercise mode are fused into a stage task, the problems that in a traditional method, a learning path is rigid, and task allocation and user ability are disjointed are solved, and therefore the learning efficiency is improved while the learning target achievement rate is guaranteed. And the pertinence and the time resource utilization rate of vocabulary practice are obviously improved.

The invention discloses a word embedding similarity-based lexicon dynamic expansion and ESG expression quantitative evaluation method, and relates to the technical field of natural language processing and enterprise social responsibility evaluation. Comprising the steps of preprocessing obtained enterprise annual report unstructured text data; vectorizing the preprocessed data set, and dividing the vectorized data set into training set data and test set data; performing training set data word screening and classification by utilizing a cosine similarity method so as to construct E, S and G word libraries; calculating the similarity between the test set data and the E, S and G word libraries to obtain the ESG score of the enterprise; and constructing an enterprise ESG performance radar map according to the ESG score, and displaying the ESG performance of the enterprise. According to the method, the ESG performance of the enterprise is evaluated by applying the unstructured data, so that a method system for evaluating the ESG performance is greatly enriched, and a brand new analysis view angle is provided for predicting the comprehensive strength of the enterprise.

The invention relates to a livelihood public opinion big data-based territorial space planning implementation monitoring method, which comprises the following steps of: constructing a multi-level space unit and a territorial space information lexicon according to territorial space multi-source data, and matching livelihood public opinion big data in the territorial space information lexicon to obtain a spatial position; matching the obtained spatial position with a multi-level spatial unit to obtain a spatial unit corresponding to the spatial position; identifying an emotion label for each record in the livelihood public opinion big data, and calculating a livelihood emotion index planned and implemented in the spatial range of each hierarchical spatial unit according to the emotion label; and processing the livelihood public opinion big data to obtain a territorial space livelihood keyword library, updating the territorial space livelihood keyword library, and generating each level of space unit planning and implementing livelihood public opinion keywords in combination with the territorial space information word library and the updated territorial space livelihood keyword library. The method can dynamically monitor subjective feelings of the public on planning implementation conditions and key problems of social livelihood attention.

The invention provides an urban flood disaster risk assessment method based on crowd sensing, a medium and equipment, and the method comprises the steps: collecting flood disaster text data on a social platform, carrying out the preprocessing, and constructing a flood disaster text data set; extracting flood disaster common words from the flood disaster text data set, classifying the flood disaster common words according to subject terms, and constructing a disaster descriptive lexicon based on a co-occurrence network; designing a crowdsourcing questionnaire based on the disaster descriptive word bank, issuing the crowdsourcing questionnaire to a target user group, and collecting flood disaster feedback data and geographic information fed back by the target user group; performing spatial distribution on the collected flood disaster feedback data according to geographic information; and carrying out risk assessment by combining two types of indexes which respectively represent the flood disaster situation and the flood disaster influence area. According to the invention, real-time monitoring and accurate assessment of urban flood disasters are realized, and the accuracy and real-time performance of flood disaster risk assessment are improved.

The invention discloses a work order auditing management method and system based on a large language model and a medium. The method comprises the steps of receiving input and correspondingly matching a work order auditing rule; based on the matched work order auditing rule, analyzing and processing the input through a large language model NLP, and converting the input into a natural language rule; carrying out synonym and antonym conversion on input white words, and carrying out equal rule conversion on converted words based on a corpus word library; performing rule text and code conversion on the converted natural language rule; and carding the processed rules into SQL statements which can be identified by the system according to the fields and the judger rule base. Updating the rule knowledge base; and writing the processed SQL statements which can be identified by the program into a rule knowledge base. According to the method, non-technical personnel can adjust and modify the work order information verification rule by using a natural language by using a large language model technology, so that the intelligent level of a work order system is enhanced, the development cost of the system is reduced, the auditing efficiency of a work order process system is improved, and the method has relatively good practicability.

The present invention discloses a method and device for evaluating a multi-modal large model for urban governance based on multi-level texts, comprising the following steps: obtaining the true label text of a sample picture and the output text of the multi-modal large model for urban governance; using an urban management thesaurus to segment the text into independent words and calculating the similarity score of lexical units; using the urban management thesaurus to create a bag-of-words model to convert the text data into a word frequency matrix and calculating the low-level word frequency similarity score; using a trained language model to extract semantic features with context in the text and converting them into vector representations, and calculating the semantic similarity score; using a general large language model with a formatted request with business guiding words and evaluation requirements to obtain the high-level semantic similarity score corresponding to the text; and combining the similarity scores at each level to calculate the comprehensive evaluation score of the multi-modal large model. The present invention fully considers the similarity measurement in the multi-level space of texts, improving the accuracy of evaluating the text output of the urban governance large model.

The invention relates to the field of large model security testing, and particularly provides a large model output content security testing method and device, and the method comprises the following steps: S1, preparing and managing a test set, a sensitive word library and a regular expression which are required by testing; s2, reading a test set, and obtaining a large model output result according to the test set and the large model interface information; s3, judging whether the output content of the large model is safe or not according to the sensitive lexicon and the regular expression; s4, extracting semantic risk features according to the output content of the large model by using the large model and the oriented Prompt, and automatically storing the semantic risk features after confidence verification; and S5, storing the information result of each request in a file. Compared with the prior art, the test time can be shortened, and the evaluation efficiency can be improved; and the security of the output content of the large model can be effectively evaluated by using a method for dynamically constructing the sensitive word bank by using the output result of the large model.