15 results about "Text compression" patented technology

The invention relates to the technical field of natural language processing, and provides a multi-granularity text compression method and system for large-model multi-round dialogues, and the method comprises the steps: carrying out the semantic coding of historical question and answer pairs and a current dialogue question through a pre-training model, obtaining a semantic vector, generating a historical dialogue semantic representation through a self-attention mechanism, and carrying out the semantic representation of the historical dialogue. A semantic vector of a current dialogue problem is combined, a correlation enhancement vector is obtained through a target attention mechanism, weighted fusion is carried out on the correlation enhancement vector and historical dialogue semantic representation, a fusion vector is obtained, a feature vector is generated through a maximum pooling and average pooling combination strategy, clustering is carried out, positive and negative sample pairs are constructed, and semantic similarity loss calculation is carried out. Iteratively training the classification model; and based on the trained classification model, screening question and answer pairs related to the current dialogue question, and generating a structured abstract. Redundant content is remarkably reduced while key semantic information is reserved, and efficient context representation adaptive to large model reasoning is formed.

The invention belongs to the technical field of voice communication, and particularly relates to a voice communication method, device and equipment based on three-group decomposition. The method comprises the following steps: performing three-way parallel feature extraction on an input voice signal to obtain a text feature, a rhythm feature and a tone feature; performing text compression, rhythm compression and timbre compression on the text features, the rhythm features and the timbre features to obtain a text packet, a rhythm packet and a timbre packet; transmitting the text packet, the rhythm packet and the tone packet according to the transmission priority; and respectively decompressing the text packet, the rhythm packet and the tone packet, synthesizing and playing. According to the voice communication system, voice is explicitly decomposed into three independent dimensions of language content, rhythm expression and speaker identity, compression strategies are optimized respectively, and the normal form that all information is entangled and coded by a traditional codec and an existing semantic compression method is broken through. While the bit rate is low, the quality level close to that of natural speech can be obtained.

The invention relates to the technical field of voice semantics, and discloses a cue word determination method and device based on compression and rearrangement, equipment and a medium, and the method comprises the steps: carrying out semantic analysis and blocking processing on an original cue text, and generating an original semantic unit; calculating a correlation score based on the task description information; distributing a text compression ratio according to the correlation score; and rearranging the target compressed semantic units to generate a target prompt word sequence. Through the above mode, redundant information is eliminated through intelligent compression ratio distribution and fine-grained semantic compression based on correlation scores, so that the length of the input sequence is greatly reduced. The position weighting function and the correlation score are introduced to perform sequence rearrangement on the semantic unit, the calculation complexity during large language model reasoning is reduced, the method can be applied to the fields of financial science and technology, medical treatment, health, old-age care and the like, and the efficiency of compressing redundant information and generating cue words by an intelligent interaction system is improved.

The invention discloses a defense method and device for a context injection attack of a large language model, and relates to the technical field of artificial intelligence network security protection. The method comprises the following steps of: performing non-uniform sampling on an input context, and extracting a head system instruction area, a tail user query area and a plurality of representative text fragments in the middle; by calculating a text compression ratio feature value and an N-gram repetition rate feature value of the middle fragment, capturing structural anomaly in a non-semantic understanding mode; calculating a semantic similarity characteristic value of the tail query and the middle segment, and carrying out intention corresponding detection; a total risk score is calculated, and hierarchical non-intrusive intervention from adding a security instruction at the end of the context to applying a dynamic Logit bias in a model decoding phase is performed according to different risk levels. According to the method, the safety detection and defense of the millisecond-level delay of the ultra-long context are realized, the internal weight of the model does not need to be modified, the universality is high, and the false alarm rate is low.

The invention discloses a text compression and recognition method and device based on a visual modality and a medium, and the method comprises the steps: obtaining to-be-processed original text data, rendering the original text data into a text image, and preprocessing the text image; inputting the preprocessed text image into a depth visual encoder to extract image features through a multi-layer neural network architecture, and compressing the number of visual marks by adopting a mark compression mechanism to generate a compressed visual mark sequence; and inputting the compressed visual mark sequence into a hybrid expert decoder to activate a preset number of expert networks according to the compressed visual mark sequence, and decoding the compressed visual mark sequence into a corresponding text mark sequence to realize reconstruction of the original text data.

The present application relates to a kind of training methods of clinical auxiliary diagnosis model, belong to medical information processing technical field, solve the problem of poor coordination of size model in prior art affects the accuracy of diagnosis decision.Method includes: obtaining the medical text of multiple patients and constructs sample set;Text compression model and medical diagnosis large model are constructed, and the pre-training of text compression model and medical diagnosis large model constructed respectively is carried out based on sample set;The parameter of medical diagnosis large model is fixed, and the compressed text output by text compression model is input to medical large model, and text compression model is optimized based on the diagnosis task loss of medical diagnosis large model;The parameter of text compression model is fixed, and the compressed text output by text compression model is input to medical large model, and medical diagnosis large model is fine-tuned based on the diagnosis task loss of medical diagnosis large model, and the trained clinical auxiliary diagnosis model is obtained.The accuracy of diagnosis in complex clinical scenario is improved.

This invention discloses an Lz4 text compression method based on a sliding dictionary, primarily addressing the problems of slow compression speed and difficulty in hardware implementation in current text compression methods. This invention establishes a sliding dictionary, a hash table, and a register for the number of unsuccessful matches. The text to be compressed is read into the sliding dictionary, hash addresses are calculated to store string position information, the sliding dictionary is used to find duplicate strings, and the information of duplicate strings is stored to compress the text information, finally outputting the compressed text stream. The parallel processing of hash address calculation is easily implemented in a pipelined hardware manner. The high probability of finding duplicate strings through pre-prediction reduces the number of hash address calculations. The compression ratio of the text to be compressed is determined by the number of unsuccessful string matches, and the input speed of the text to be compressed is adjusted accordingly, significantly improving the compression speed.

The invention discloses a digital watermark steganography method and system based on optical text compression, and the method comprises the steps: 1, carrying out the load generation and optical coding, carrying out the optical rendering of a steganography text or image, inputting a visual semantic encoder, and obtaining a semantic vector set; step 2, performing semantic vector discretization and robustness enhancement; step 3, carrying out deep steganography embedding; 4, carrying out steganography extraction and error correction recovery; and 5, performing semantic vector reconstruction and optical decoding content generation. According to the method, cross-modal long text steganography based on visual semantic compression is realized for the first time, the limitation of a traditional steganography method in the aspects of capacity, recoverability and aggressiveness resistance is broken through, and the method can be widely applied to the fields of copyright identification, secret communication, digital traceability, media asset security and the like and has important engineering value and application prospect.

The embodiment of the application discloses a long text compression method and device based on a language model, wherein the method comprises the following steps: performing sentence division on a first initial long text by a regular method to obtain a first sentence sequence, and performing sentence division on a second initial long text by the regular method to obtain a second sentence sequence; obtaining a third sentence sequence from the first sentence sequence according to a Top K algorithm, and obtaining a fourth sentence sequence from the second sentence sequence according to the Top K algorithm; determining a sixth sentence sequence with the maximum similarity to the third sentence sequence from a fifth sentence sequence, and determining an eighth sentence sequence with the maximum similarity to the fourth sentence sequence from a seventh sentence sequence; splicing the third sentence sequence and the eighth sentence sequence to obtain a ninth sentence sequence, and splicing the fourth sentence sequence and the sixth sentence sequence to obtain a tenth sentence sequence. The application can improve the accuracy of long text compression.

The invention relates to a training method of a clinical auxiliary diagnosis model, belongs to the technical field of medical information processing, and solves the problem that the diagnosis decision accuracy is influenced by poor collaboration of large and small models in the prior art. The method comprises the following steps: acquiring medical text construction sample sets of a plurality of patients; constructing a text compression model and a medical diagnosis large model, and pre-training the constructed text compression model and medical diagnosis large model based on the sample set; fixing parameters of the large medical diagnosis model, taking a compressed text output by the text compression model as input of the large medical diagnosis model, and optimizing the text compression model based on diagnosis task loss of the large medical diagnosis model; and fixing parameters of the text compression model, taking a compressed text output by the text compression model as input of the large medical diagnosis model, and performing fine adjustment on the large medical diagnosis model based on diagnosis task loss of the large medical diagnosis model to obtain a trained clinical auxiliary diagnosis model. And the diagnosis accuracy in a complex clinical scene is improved.

This invention relates to an optimization method for ultra-large text transmission based on dynamic block compression and routing, belonging to the field of ultra-large text transmission technology. The method includes: constructing a semantic relationship graph from the ultra-large text through semantic analysis; dividing the ultra-large text into logical segments based on semantic relevance; dynamically adjusting the size of the block data based on the access frequency of the logical segments; matching the optimal compression algorithm to the block data according to the data characteristics of each block data; and dynamically predicting the transmission network congestion status based on real-time monitoring of the current network status. If there is no congestion, the compressed data blocks are transmitted to the target node one by one according to the original path; if there is congestion, the optimal path is replanned, the SDN route is updated, and the compressed data blocks are transmitted to the target node according to the updated path. This invention improves the efficiency of ultra-large text transmission, optimizes text compression performance, and ensures the stability and efficiency of data transmission.

The application discloses a defense method and device for large language model context injection attack, and relates to the technical field of artificial intelligence network security protection. The method comprises the following steps: non-uniform sampling is performed on input context, a head system instruction area, a tail user query area and a plurality of representative text segments in the middle are extracted; a text compression characteristic value and an N-gram repetition rate characteristic value of the middle segment are calculated to capture structural abnormalities in a non-semantic understanding manner; a semantic similarity characteristic value of the tail query and the middle segment is calculated to perform intention echo detection; a total risk score is calculated, and according to different risk levels, a graded non-intrusive intervention is performed from appending a safety instruction at the end of the context to applying a dynamic Logit bias in the model decoding stage. The application realizes millisecond-level delay safety detection and defense for an ultra-long context, and does not need to modify the internal weight of the model, so that the generality is high and the false alarm rate is low.

The application belongs to the technical scheme of human-computer interaction interface optimization, and particularly relates to a self-adaptive area Chinese text compression display method and system. Step one: dynamic display parameters are collected, and the collected parameters are preprocessed; step two: the input English text is subjected to semantic structure analysis through a Chinese semantic analysis module; step three: based on the display parameters preprocessed in step one and the Chinese text subjected to semantic structure analysis in step two, the text compression level is calculated based on a space density evaluation matrix, and a compression strategy is implemented; step four: the original text data and character rendering are stored through a reversible compression marking system. The application is used to solve the core problems existing in the field of intelligent terminal text display, such as semantic integrity damage, cross-platform adaptation error, unreliable data restoration and lagging behind of technology evolution.

The application discloses a multi-modal corpus storage method and device, computer equipment and medium. The compression method respectively performs compression processing on the text modal data and audio-video modal data of the multi-modal corpus, adopts a suitable compression algorithm according to the characteristics of different modal data, and performs similarity calculation on the text compression features, audio-video compression features and corresponding modal features of the newly added multi-modal corpus, so that the semantic redundancy of the newly added multi-modal corpus and the historical compressed corpus can be identified, the fusion compression process is triggered to eliminate cross-modal redundancy and avoid repeated storage of the same semantic information. Furthermore, the corpus features of the corpus not triggering the fusion process are independently stored, and the corpus features triggering the fusion process are fused to generate multi-modal joint features, so that the amount of data stored is further reduced. The corpus features are migrated to the corresponding data storage layer in combination with the access frequency and response time, thereby reducing the storage cost.

The application provides a summary generation system based on multi-modal key information analysis, comprising the following modules: a text encoder module for obtaining encoding information of a text; an image encoder module for obtaining feature information at different angles of an image; a selection gate mechanism module for mining key information of text and image modalities, comprising a multi-modal selection gate network and a visual selection gate network; a multi-head selection attention module for mining key object information in the image; a multi-modal feature fusion module for obtaining multi-modal context representation; and a decoder module for generating a summary sentence; when generating the summary, the system introduces object information of the image, and simultaneously shares the multi-modal key information extraction process of text compression to the encoding end; by mining key information of the text and the image, the application models the mutual dependence relationship between the object and the text sequence, highlights the key object information of the image, and thus generates a high-quality text summary.