79 results about "Word target" patented technology

The invention provides a text reinforcement learning method and device, electronic equipment and a computer storage medium, relates to the technical field of text reinforcement learning, and is applied to a text generation model. The method comprises the following steps: determining one or more target vocabularies in a statement to be analyzed, and generating a replacement word set according to the target vocabularies; replacing the target vocabulary in the to-be-analyzed statement with a current candidate replacement word to obtain a replacement statement; calculating one or more difference index values between the replacement statement and the to-be-analyzed statement; wherein the difference index value is used for quantifying the influence of the replacement of the target vocabulary on the original sentence from different dimensions; based on the difference index value, determining an importance weight of the target vocabulary in the to-be-analyzed statement; wherein the importance weight is used for redistributing the reward of the single character in the sentence. Vocabulary importance is quantified through disturbance analysis and multi-dimensional evaluation, and then fine-grained distribution of rewards is achieved.

The invention discloses a speculation decoding method and device based on a large language model, equipment and a medium, and relates to the technical field of large models, and the method comprises the steps: determining the output probability distribution of a plurality of lexical element positions, and determining a confidence score based on the output probability distribution; determining a target control parameter according to the target state signal, performing aggregation calculation on the confidence coefficient vector to obtain an aggregation confidence coefficient, and determining a target draft length through the aggregation confidence coefficient and the target control parameter; determining a verification tolerance threshold based on the target control parameter and the confidence score, generating a draft sequence based on the target draft length by using a target deep learning model, and verifying each lexical element in the draft sequence based on a preset large language model and the verification tolerance threshold; and adding the successfully verified target lexical units to the end of the currently generated sequence to obtain a target sequence, determining the target sequence as the currently generated sequence, and skipping again to perform the next round of speculative decoding. According to the invention, the efficiency of speculative decoding can be improved.

The embodiment of the invention provides an observation information generation method, device and equipment. According to the scheme, the method comprises the steps that in the process that a generative large language model generates reply information for a target request, a probability distribution vector of a target lexical element is collected; then, based on the probability distribution vector, determining recommendation lexical element information with high probability; acquiring sampling parameter information corresponding to the target request; afterwards, the recommended lexical element information, the sampling parameter information and the target lexical element are stored in an associated mode, and quality observation information is obtained so as to be used for analyzing reasons for quality abnormity of the reply information.

The invention discloses a text data classification method and device, electronic equipment and a nonvolatile storage medium. The method comprises the steps of obtaining target text data; adopting a first model to match the target text data with vocabularies in a target word bank, and determining a first score corresponding to the target text data according to a matching result, a second model is adopted, the first semantic feature corresponding to the target text data is matched with a target semantic feature library, a second score corresponding to the target text data is determined according to the matching result, and the operand and complexity of the second model are larger than those of the first model; and determining a classification result of the target text data according to the first score and the second score. According to the method and the device, the technical problem of poor text data classification effect caused by limitation of a single model on semantic comprehension and classification decision in related technologies is solved.

The invention relates to the technical field of artificial intelligence, is suitable for the financial field and the medical health field, and provides a verbal skill generation method and device, equipment and a storage medium, and the method comprises the steps: obtaining a user input text of which a topic deviates from a business topic, and extracting a target word group from the user input text; determining a first response template text from a plurality of preset response template texts associated with different business themes; generating a second response template text based on the target word group and the first response template text; analyzing the user input text and the second response template by using a preset language model to generate a plurality of target lexical elements for responding to the user input text, and obtaining probability data corresponding to each target lexical element; and based on the multiple target lexical elements and the probability data corresponding to each target lexical element, generating a target response verbal skill used for guiding the topic to return to the business theme. The method can help an artificial intelligence customer service to naturally guide the dialogue to return to the business theme.

This application relates to the field of artificial intelligence security technology and discloses a method, apparatus, device, and medium for security optimization of a large language model, comprising: acquiring an attack task dataset; wherein the attack task dataset includes at least one attack task type and multiple malicious instruction data under the attack task type; generating a general word-level adversarial suffix corresponding to the attack task type based on the malicious instruction data, and generating a target word-level adversarial suffix corresponding to the malicious instruction data based on the general word-level adversarial suffix; performing a first fine-tuning on the large language model based on the target word-level adversarial suffix to obtain a first fine-tuned model; and performing a second fine-tuning on the first fine-tuned model based on the acquired target semantic-level adversarial hints to obtain a second fine-tuned model. This application can improve the ability of a large language model to resist various known and unknown forms of jailbreak attacks.

This application provides a text processing method, apparatus, device, and storage medium. The method involves segmenting the text to obtain several words to be grouped; calculating a first semantic similarity between the words to be grouped; grouping the words to be grouped into word combinations based on the first semantic similarity, with each word combination containing words to be grouped with similar semantics; searching for words in the word combinations that do not match a preset correct word as replacement words; replacing the replacement words with the preset correct word to obtain a target word combination; and using the target word combination to correct the word combinations in the text to obtain the target text. This application, based on the first semantic similarity, automatically performs consistency checks and processing on the text in the front-end system, improving the efficiency and accuracy of the front-end system's text consistency checks, thereby saving manpower and time costs.

Embodiments of the present application provide a cross-model collaborative reasoning method and device, equipment and a storage medium, relating to the technical field of intelligent computing equipment. The input prompt is sent to the first model for reasoning to generate a probability value corresponding to each word element at the current time step. The highest probability value is determined from the probability value, and the marginal benefit of the current time step is determined according to the highest probability value. The current benefit threshold corresponding to the current time step is obtained, and the decision variable is determined according to the marginal benefit and the current benefit threshold. The target word of the current time step is obtained based on the decision variable, and the output sequence corresponding to the input prompt is obtained according to the target word sequence before the current time step and the target word. The on-demand switching of the double model is realized by the dynamic determination of the marginal benefit and the benefit threshold. The first model is used to realize low-energy-consumption reasoning at the time step with high reasoning confidence. The second model is used to ensure accuracy at the key time step with low reasoning confidence. The reasoning accuracy is ensured while reducing the reasoning energy consumption.

The application discloses a semantic recognition model based on NPL and a training method thereof, relates to the technical field of semantic recognition model training, and aims to solve the problem of low recognition accuracy of a semantic recognition model. The current whole sentence is searched, multiple segmented sentences are screened out, words in the segmented sentences are recognized, the part-of-speech information of the words is collected, the part-of-speech weight under the corresponding part-of-speech information is called, the number of repeated occurrences of each word in the current whole sentence is counted and the repeated word frequency is calculated, the processing level of each word is calculated by comprehensively considering the repeated word frequency and the part-of-speech weight, the target words are screened out and recognized, the semantics of the target words are acquired and the number of semantics is counted under the current interactive window, the interactive duration of the target words is analyzed by calling historical data, the number of word reference sources is set as a reference, the number of semantics of the target words and the interactive duration are comprehensively considered to adjust the number of word reference sources of each target word, and the semantic analysis accuracy of each target word in the interactive sentence is improved.

The application discloses a cross-media model error attribution and auxiliary correction method based on causal decoupling, which comprises the following steps: image segmentation is performed on an input image, and the input image is divided into a plurality of sub-regions; a target function is solved according to the sub-regions, the minimum key region which supports the generation of a target word element to the greatest extent and removes the generated ability is sorted according to the contribution to the generation decision, and an attribution saliency map is generated; an influence score is calculated according to the change of the generation probability of the target word element in the process of inserting the sub-regions in the ordered subset; the ordered subset, the attribution saliency map and the influence score are evaluated from the dimensions of fidelity, positioning ability and error correction guiding ability, and an evaluation result is obtained; the evaluation result is obtained without relying on the internal gradient, attention weight or activation map of the model, the dependence on the internal structure of the model is reduced, the regional attribution of an arbitrary word element set is realized, and the relative dependence of the generation process on visual evidence and language prior is quantified.

Embodiments of the present specification disclose a key information extraction and live content processing method, device and equipment. The scheme comprises: obtaining original text to be extracted key information; determining an attention weight vector for the original text, and correspondingly generating a first target word according to the attention weight vector; judging whether the first target word is included in the original text; if yes, obtaining the attention weight of the first target word when generating the first target word according to the attention weight vector, generating penalty data for the first target word according to the attention weight of the first target word, and continuing to generate a second target word according to the penalty data, wherein the penalty data is used to reduce the possibility that the second target word generated continues to be the same as the first target word; and generating key information according to the first target word and the second target word.

The invention provides a code generation method and device, electronic equipment and a storage medium. The method comprises the following steps: analyzing at least one piece of key information in demand information corresponding to a natural language search text; in order to determine the importance degree of the key information, respectively generating second probability distribution obtained according to semantic features of the natural language search text and first probability distribution corresponding to a preset lexical element library obtained according to mask semantic features, since the mask semantic feature is the semantic feature corresponding to the natural language search text after hiding the key information, the influence degree of the key information on the probability distribution of the generated lexical elements can be determined according to the difference distribution between the first probability distribution and the second probability distribution. According to the method, the difference value distribution and the second probability distribution are obtained, the target lexical element sequence is determined according to the difference value distribution and the second probability distribution, the accuracy of the target lexical element sequence is improved, finally, the target code is generated according to the target lexical element sequence, and the matching degree of the target code and the demand information corresponding to the natural language search text is improved.

Systems or techniques that facilitate systems and methods for providing explainability of natural language processing are provided. In various embodiments, a system can access a plain text clinical sentence. In various aspects, the system can generate, via execution of a first machine learning model, an assertion status classification label for a word of interest in the plain text clinical sentence. In various instances, the system can extract, from a hidden attention layer of the first machine learning model, word-wise attention scores corresponding to the plain text clinical sentence and render, on an electronic display, both the assertion status classification label and a graphical representation of the word-wise attention scores. In various cases, the system can determine, via execution of a second machine learning model, a reliability score for the assertion status classification label, based on the word-wise attention scores, and can render the reliability score on the electronic display.

The invention discloses a semantic sentiment analysis method, an intelligent cabin, a medium and a vehicle. According to the scheme, user voice is split, and global semantic information is put into a token sequence for coding; a target lexical element semantic code of each independent neural network is screened from the target lexical element semantic codes by using a gating network, and is processed into an emotion semantic feature; newly adding a target neural network, and generating shared semantic features in combination with the gating network; therefore, shared semantic features, global semantic codes and own emotion semantic features are fused in each independent neural network, when a specific emotion category is analyzed, local emotion semantic features are focused to serve as a detail basis, and overall emotion deviation can be grasped by means of the global semantic codes, so that the emotion classification accuracy is improved. In addition, semantic support common with all emotion categories can be obtained by sharing semantic features, and deep mining is performed on local and overall association, so that judgment of each emotion category can give consideration to detail expression and global emotion deviation, and the emotion analysis ability of the intelligent cockpit is improved.

The application discloses a kind of fusion knowledge enhancement and the method and system of original meaning prediction of confrontation training, it is related to natural language processing technical field, obtain target word and its corresponding original dictionary explanation;Original dictionary explanation is carried out semantic integrity evaluation using semantic perception selective enhancement intelligent agent;Target word, original dictionary explanation and supplementary context are structured as structured input sequence;Structured input sequence is sent into original meaning encoder, semantic features are extracted, and original meaning prediction vector is obtained;Using the method based on fast gradient, dynamic confrontation disturbance is applied to word embedding layer during model training process, clean sample loss and confrontation sample loss are optimized jointly to update model parameters;The original meaning label set corresponding to target word is output.The application introduces two big mechanisms of selective knowledge enhancement and confrontation training, and systematically solves the problem that decision boundary is weak due to static dictionary information bottleneck and original meaning long tail distribution in traditional original meaning prediction.

The application provides an intention recognition method and device, electronic equipment and storage medium. The intention recognition method comprises: performing word segmentation processing on a target sentence in current session text from a customer to obtain at least one target word; determining whether there is a word with the same character as a plurality of keywords in a preset keyword table in the at least one target word, and if so, determining the word with the same character as a candidate word; calculating the similarity value between at least part of the keywords in the preset keyword table and the candidate word to obtain at least one similarity value; when there is a similarity value greater than or equal to a first preset threshold in the at least one similarity value, determining the context information of the target sentence; and performing intention recognition on the target sentence according to the context information to obtain a target intention label. The technical scheme of the application can realize the real-time of intention recognition, and improve the accuracy of the recognized intention label.

Embodiments of the present specification relate to a text desensitization method based on differential privacy, a program product and a computing device. The method comprises: obtaining a cluster set, any target word in any cluster has a preset replacement probability distribution, the replacement probability distribution is determined based on an exponential mechanism, and is used to describe the probability of replacing the target word with each word in the cluster; performing named entity recognition on the input text to be desensitized to determine the entity type to which each word belongs, which includes an entity type preset as a sensitive word type; for any target sensitive word belonging to the sensitive word type, it is judged whether it exists in any cluster of the cluster set, when the target sensitive word exists in a first target cluster, a target replacement word is obtained by sampling according to the replacement probability distribution of the target sensitive word in the first target cluster; and the target sensitive word in the text to be desensitized is replaced with the target replacement word to obtain a desensitized text.

The invention discloses a task processing method and device based on a task knowledge base, equipment and a storage medium, and relates to the technical field of semantic communication, and the method comprises the steps: converting an index sequence of a target information source semantic representation vector into a first lexical element sequence of a large language model based on an information source semantic representation knowledge base; on the basis of a channel semantic representation knowledge base, converting the channel state information into a second lexical element sequence of the large language model; splicing the structured task demand prompt information, the first lexical element sequence and the second lexical element sequence into target prompt information, and inputting the target prompt information into a large language model for reasoning to obtain a target lexical element sequence; and based on the property of the target task, processing the target lexical element sequence to obtain a task processing result. Through the mode, a bidirectional conversion mechanism between the multi-modal task / channel environment semantics and the lexical elements of the large language model is constructed, so that the large language model can understand and process diversified task data from a dynamic channel.

The invention discloses a model reasoning control method and device, a medium and a product, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining a probability distribution sequence of lexical elements in a pre-training language model reasoning process, the probability distribution sequence comprises current probability distribution of the lexical elements in the current reasoning step and historical probability distribution of the lexical elements in the historical reasoning step; extracting current probability distribution and historical probability distribution, identifying a current ranking of a preset target lexical element in a current step and a historical ranking of the preset target lexical element in a historical step, and generating a ranking sequence according to the current ranking and the historical ranking; according to the method and the device, the ranking sequence is ranked, an abortion reasoning step of the pre-training language model is positioned according to the ranking sequence and a preset reasoning abortion threshold value, and a reasoning process of the pre-training language model is controlled based on the abortion reasoning step, so that the technical problems of excessive thinking and calculation waste caused by incapability of accurately identifying a convergence critical point in related technologies are solved; the technical effects of reducing redundant thinking and reducing calculation overhead are achieved.

This invention relates to the field of natural language processing technology, providing a word recommendation method, apparatus, electronic device, and storage medium. The method first obtains the definition to be queried; then, based on a reverse dictionary model, selects target words corresponding to the definition from a candidate word list; finally, based on the target words, determines the recommendation result. This method utilizes a reverse dictionary model obtained through multi-task learning to predict words and their parts of speech, considering the parts of speech rather than solely relying on the quality of definitions in the dictionary and the quality of the user-input definition. This ensures the accuracy and quality of the recommendation results, thereby improving the user experience. The introduction of parts of speech helps reduce the prediction space, constrains the recommendation results, makes the results more reliable, reduces interference from easily confused words, and avoids situations where the recommended results differ significantly from the user-input definition.

The application discloses a kind of classroom teaching auxiliary systems for classical Chinese knowledge explanation, comprising: text acquisition module, obtains the full-text of target argument type classical Chinese and course knowledge point mapping table;Text preprocessing module, full-text is segmented and segmented to text;Text marking module is used to mark target word from full-text, and the sentence belonging is marked as target sentence;Text feature coding module generates the feature vector of target word and the context concatenation feature vector of target sentence;Text analysis module, the feature vector of target word and the context concatenation feature vector of target sentence are matched with course knowledge point mapping table respectively to output the explanation and usage of target word and the explanation of target sentence;Text information output module, output the explanation and usage of target word and the explanation of target sentence.

The invention provides a method for reasoning based on a language model, the method is executed by a server, and the server is deployed with a target language model; the target language model comprises a target vocabulary and a target weight matrix, the target vocabulary is obtained by multiplying each embedded vector in an original vocabulary of the original language model by a preset rotation matrix, and a corresponding relationship between the embedded vectors and the input data is stored in the target vocabulary; the target weight matrix is obtained by multiplying the inverse matrix of the rotation matrix by the original weight matrix of the original language model. In the reasoning process, receiving input data corresponding to an input text from a client, and then obtaining a first intermediate result corresponding to the input data; the first intermediate result comprises a plurality of embedded vectors; processing the first intermediate result according to the target weight matrix to obtain a second intermediate result; and obtaining a reasoning result of the target language model according to the second intermediate result, and returning the reasoning result to the client.

The invention relates to the technical field of natural language processing, and discloses a synonym replacement-based text data enhancement method, which comprises the following steps of: retrieving sentences containing vocabularies from a corpus and sampling; performing word segmentation and coding on the sentences obtained by sampling, and extracting context representation vectors of vocabularies by using a pre-training language model; clustering the context representation vectors of the same vocabulary in different contexts to obtain a central semantic representation vector; storing the vocabularies and the corresponding central semantic representation vectors into a vector database; for a target vocabulary in the sentence to be processed, searching similar candidate vocabularies in the vector database by using the corresponding context representation vector; performing integral gradient calculation to obtain candidate word scores; and sorting the candidate words according to the candidate word scores to generate an enhanced sentence. According to the method, the polysemy vector representation of the vocabularies is stored by utilizing the vector database, the candidate vocabularies are sorted in combination with the integral gradient, and diversified training data are generated, so that the model training efficiency and the data utilization effect are improved.

The present disclosure provides a question and answer method, device, equipment and storage medium, which can be applied to the field of artificial intelligence and finance. The method comprises: determining a question text vector; determining N target word vectors; determining answer information according to the N target word vectors; wherein, the N target word vectors are determined by repeatedly performing the following operations until the N target word vectors are obtained: under the condition that 1 < n ≤ N, determining an n th first candidate word vector set from a preset word vector set according to the question text vector and the first n-1 target word vectors in the preset word vector set; randomly dividing the n th first candidate word vector set to obtain a plurality of n th first candidate word vector subsets; determining an n th second candidate word vector set from the n th first candidate word vector subsets according to the n th first candidate word vector set, the question text vector and the first n-1 target word vectors in the preset word vector set; and determining an n th target word vector according to the n th second candidate word vector set.

The embodiment of the invention provides a model training method, a word vector generation method and device, equipment and a medium, and is applied to the technical field of computers. The model training method comprises the following steps: acquiring a plurality of sample vocabularies; obtaining a grayscale image of each sample vocabulary; for the grayscale image of each sample vocabulary, determining a vocabulary vector of the sample vocabulary according to the grayscale value of each pixel point in the grayscale image; and performing model training according to each sample vocabulary and the vocabulary vector of each sample vocabulary, and generating a vocabulary vector generation model for predicting the vocabulary vectors corresponding to adjacent vocabularies. When word vectors are generated on the basis of a word vector generation model, the word vector generation method comprises the following steps: acquiring words to be processed; and inputting the to-be-processed vocabulary into the vocabulary vector generation model, and obtaining a target vocabulary vector output by the vocabulary vector generation model. Based on the model training method and the use of the vocabulary vector model, the effect of improving the word vector learning efficiency is achieved.

The invention belongs to the technical field of artificial intelligence, and relates to an artificial intelligence-based data processing method, which comprises the following steps of: processing a target image and a description text based on a large visual language model to generate lexical elements; if a modal prejudice quantification result obtained by quantifying the lexical elements meets attention intervention conditions, calculating an original attention matrix of the lexical elements, and performing weight adjustment on the original attention matrix based on the modal prejudice quantification result to obtain a specified attention matrix; calculating the original attention matrix to obtain first conditional probability distribution, and calculating the specified attention matrix to obtain second conditional probability distribution; fusing the first conditional probability distribution and the second conditional probability distribution to obtain target conditional probability distribution; and decoding the target conditional probability distribution to generate a target lexical element, and generating and outputting a target text based on the target lexical element. The method can be applied to image analysis scenes in the financial science and technology field and the medical field, and the accuracy of the generated target text is improved.

This invention proposes an intelligent enterprise accounting voucher generation and compliance review system, comprising: a text processing module, including a target lexicon generation unit, which can extract several negative words, several modifier texts paired with each negative word, and several correction words from preprocessed enterprise accounting text data and generate a target lexicon; and a current text recognition module, which is used to calculate the part-of-speech relevance C of the current text based on the target lexicon, and to identify the current text as a negative word or a corrected word based on the part-of-speech relevance C and the context of the current text, so as to perform a correction operation based on the recognition result, thereby solving the problem that the existing system has insufficient ability to analyze the scope of negative words and the consistency of context before and after correction, which leads to the existing system easily misclassifying negative statements and ignoring the associated impact of correction operations, thus causing voucher generation errors or compliance review loopholes.

The invention provides a position coding method, device and equipment for a structured sequence. The method comprises the following steps: acquiring a target lexical element in a structured input sequence of a target language model and position information of the target lexical element in the structured input sequence; obtaining a multi-dimensional structure coordinate of the target lexical element according to the position information of the target lexical element in the structured input sequence; according to the multi-dimensional structure coordinates, rotating each structure dimension of the target lexical element to obtain a multi-dimensional rotating field of the target lexical element; and obtaining a position coding result of the target lexical element in the target language model according to the multi-dimensional rotation field of the target lexical element. According to the scheme, the multi-dimensional structure information can be fused to perform position coding on the input sequence, and the processing performance of the structured task of the target language model is improved.

The embodiment of the invention provides a method, device and equipment for generating reply information. According to the scheme, the method comprises the steps that key vectors and value vectors of all lexical elements in lexical element sequences corresponding to all text blocks are calculated in advance in an off-line mode; in the online stage, target lexical elements with high semantic importance in target text blocks are recognized by adopting a lexical element screening model aiming at the target text blocks which are recalled through retrieval and contained in cue words, and then on one hand, key vectors and value vectors of the target lexical elements can be calculated online, and on the other hand, the key vectors and the value vectors of the target lexical elements can be calculated online; on the other hand, obtaining a pre-calculation key vector and a pre-calculation value vector of a non-target lexical element generated offline; and the large language model can generate reply information aiming at the user query information based on the key vector and the value vector of the target lexical element generated online and the pre-calculation key vector and the pre-calculation value vector of the non-target lexical element generated offline.