462 results about "Sequence model" patented technology

An apparatus generates a statistical class sequence model called A class bi-multigram model from input training strings of discrete-valued units, where bigram dependencies are assumed between adjacent variable length sequences of maximum length N units, and where class labels are assigned to the sequences. The number of times all sequences of units occur are counted, as well as the number of times all pairs of sequences of units co-occur in the input training strings. An initial bigram probability distribution of all the pairs of sequences is computed as the number of times the two sequences co-occur, divided by the number of times the first sequence occurs in the input training string. Then, the input sequences are classified into a pre-specified desired number of classes. Further, an estimate of the bigram probability distribution of the sequences is calculated by using an EM algorithm to maximize the likelihood of the input training string computed with the input probability distributions. The above processes are then iteratively performed to generate statistical class sequence model.

The technology disclosed provides a so-called “pointer sentinel mixture architecture” for neural network sequence models that has the ability to either reproduce a token from a recent context or produce a token from a predefined vocabulary. In one implementation, a pointer sentinel-LSTM architecture achieves state of the art language modeling performance of 70.9 perplexity on the Penn Treebank dataset, while using far fewer parameters than a standard softmax LSTM.

The invention provides a Wi-Fi location fingerprint map building method and a system based on crowd-sourcing. The method comprises the steps as follows: acquiring relevant crowd-sourcing data; using a dead reckoning approach to calculate the relative motion trajectory of a user based on the crowd-sourcing data; identifying specific behaviors of the user including a plurality of behavior types through machine learning; building a behavior sequence model formed by the types of the specific behaviors in the relative motion trajectory and the relative spatial relationship between the specific behaviors; generating a point-line model of an indoor map, and matching the point-line model with the behavior sequence model through a hidden Markov model to obtain the indoor map coordinate information of the relative motion trajectory; and finally, building a Wi-Fi location fingerprint map based on the indoor location coordinate information and the crowd-sourcing data. The Wi-Fi location fingerprint map is built and updated automatically based on the crowd-sourcing data.

The invention discloses a text multi-label classification method based on semantic unit information, which comprises the following steps: establishing a semantic unit multi-label classification modelSU4MLC, taking a recurrent neural network sequence based on an attention mechanism to a sequence model as a baseline model for improvement, and improving the expression of the attention mechanism by improving a source end; Extracting semantic unit related information from the context representation of the source end of the baseline model by using hole convolution in deep learning to obtain semantic unit information; Combining the semantic unit information with the word level information by using a multi-layer mixed attention mechanism, and providing the combined information for a decoder; Anddecoding the tag sequence by using a decoder, thereby realizing text multi-tag classification based on semantic unit information. According to the method, the problems that an existing attention mechanism is easily influenced by noise and contributes to classification insufficiently can be solved, the contribution of the attention mechanism to text classification can be improved, and the text multi-label classification problem can be more efficiently solved.

The invention discloses a keyword extraction method based on a Seq2seq framework. The method comprises: creating a sequence model by utilizing a Seq2seq framework; introducing an attention mechanism,extracting features of keywords in the text; fusing a pointer network model and a Cover mechanism at a decoding end to improve the attention distribution of potential keywords; then using a softmax loss function to train a network model, and finally in a model prediction stage, using a Beam Search cluster search algorithm to generate a key word sequence with the maximum probability to serve as a key word result set to obtain appropriate key words. According to the method, deep semantics in the long text are well considered, the word distribution probability is calculated in combination with the context information context, the problem of repetition of low-frequency words and generative tasks is better solved, and the keyword extraction accuracy is improved.

The invention aims to provide a method and equipment for supplying a search result. Computer equipment acquires initial search results corresponding to a query sequence input by a user; a first sequencing model is used for screening out preferable search results from the initial search results; a second sequencing model is used for screening out the optimal search results from the preferable search results; and the optimal search results are supplied to the user. Compared with the prior art, the method and the equipment have the advantages that the search results are screened according to different levels through a plurality of sequencing models, so that the optimal search results are supplied to the user; and therefore, the precision and the efficiency of the search results are considered, and the search efficiency and the search effect are optimized. Moreover, the sequencing models are determined by a machine learning method; an upper layer model is generated by a submodel through the machine learning; and therefore, the arrangement of the sequencing models is optimized; and the instantaneity, the understandability and the controllability of the sequencing models of the systems are guaranteed.

The invention provides a text simplification method based on a word vector query model. Based on a sequence and a sequence model, when decoding is conducted, the correlation between the hidden state of a decoder and the word vectors of all vocabularies is obtained through the reference of an attention mechanism to serve as a measurement of the possibility of the words to be generated in the next step. The method includes the following steps that a text encoder is designed, and an original text is compressed; a text simplification decoding generator is designed, and the current hidden layer vector and the context vector at every moment are calculated circularly; the retrieval correlation of each word in a word list is obtained, the predicted words at the current moment are output, and a complete simplified text is obtained; a model for generating the simplified text is trained, and the log likelihood of the predicted words and actual target words is minimized; after training, the complete simplified text is generated. The method can improve the quality and accuracy of the generated text, greatly reduce the number of parameters of the existing method, and reduce the training time andthe memory usage.

The invention discloses a document classification method based on a hierarchical multi-attention network. The method comprises the following steps of utilizing a Bi-GRU sequence model for carrying outword-sentence and sentence-to-document modeling on the document; using Bi-GRU sequence model to encode each word, obtaining the context information in the sentence, and using the Soft attention to carry out the weight distribution on each word; for the process from the sentences to the document, introducing the CNN attention, and obtaining the local relevant characteristics between the sentencesin the window by utilizing a CNN model, so that the attention weight of each sentence is further obtained. Modeling can be carried out from words to sentences and from sentences to documents accordingto document characteristics, and the hierarchical structure of the documents is fully considered. Meanwhile, aiming at the word level and the sentence level, different attention mechanisms are respectively adopted to properly distribute the weights of the related contents, so that the document classification accuracy is improved.

[Object] An object is to provide an easy-to-use speech processing system attaining higher accuracy of speech recognition.

[Solution] Receiving a speech utterance, the speech processing system performs speech recognition and displays a text 158 of the recognition result. Further, the speech processing system translates the recognition result in accordance with settings to a text 176 of another language and displays and synthesizes speech of the translated result. Further, the speech processing system selects utterance candidates having high possibility to be uttered as the next utterance and having high translation and speech recognitions scores, using outputs of various sensors at the time of utterance, a pre-trained utterance sequence model and translation and speech recognition scores of utterance candidates, and recommends utterance candidates in the form of an utterance candidate recommendation list 192. A user can think of what to say next using the utterances in utterance candidate recommendation list 192 as hints.

The invention provides a Web application performance testing method based on a user frequent access sequence model. The method comprises the steps of (1) log data collection and preprocessing; (2) user access sequence mining; (3) testing model and corresponding testing scene making; (4) testing environment and monitoring tool setting; (5) performance testing and result acquiring. The method can be directly applied to practical Web application performance testing engineering. The testing method based on a user access frequent sequence model can improve the accuracy of Web application performance testing and effectively acquire the actual performance and performance bottlenecks of Web applications.

A computerized robust cell kinetic recognition method for moving cell detection from temporal image sequence receives an image sequence containing a current image. A dynamic spatial-temporal reference generation is performed to generate dynamic reference image output. A reference based object segmentation is performed to generate initial object segmentation output. An object matching and detection refinement is performed to generate kinetic recognition results output. The dynamic spatial-temporal reference generation step performs frame look ahead and the reference images contain a reference intensity image and at least one reference variation image.

The invention relates to a matching and classifying method and system for stock information, in particular to an automatic stock matching and classifying method and system based on news data. The automatic matching and classifying method is characterized by comprising the following steps: establishing a local data base; performing word segmentation and screening on historical news data, extracting key word sequences, constructing sequence models for individual share key word sequence correlation, calculating the correlation between the individual shares, and classifying stocks by combining a cluster analysis algorithm; and performing word segmentation and screening on real-time news data, extracting a real-time key word sequence, calculating sequences for real-time key word sequence correlation, and performing automatic matching with the sequence models for individual share key word sequence correlation. The automatic stock matching and classifying method and system adopt the stock key word sequence excavation technology based on the news data to achieve automatic classification of the stocks; the method is comprehensive, accurate, simple, convenient and feasible, and provides better investment reference for investors; and stocks with higher matching degree are given automatically aiming at breaking news events.

The invention discloses an intersection self-adaptation control method based on a car networking environment and belongs to the technical field of car networking. According to the method, the advantages of the car networking are made full use of, the real-time state information of cars is provided for a road side unit, firstly, abstract modeling is performed on a whole traffic network, and the dynamic priority of each traffic flow is calculated; then, an optimal phase position and phase sequence model is built according to specific characteristics of an intersection so that the optimal phase position sequence of the current intersection can be obtained, high-priority traffic flows can pass through the intersection preferentially, and meanwhile, it is guaranteed that the flow of the cars allowed to pass each time is maximum. According to the method, lots of buried sensors are not needed, and not only is city control system construction cost reduced, but also maintenance upgrading of a traffic control system is facilitated. The traffic flows and the states of the cars are accurately mastered in real time, and the situation that in the prior art, a traffic control system lags behind seriously, and obtained information is little and even wrong is greatly changed.

The invention discloses a neural network-based boundary combination named entity recognition method, which comprises the following steps of 1, extracting the entity boundary information based on a neural network model, and constructing a boundary recognition model; 2, implementing a boundary combination strategy, and combining the entity boundaries to obtain a candidate entity set; and step 3, constructing a neural network classifier, and screening the candidate entity set. According to the method disclosed by the present invention, by employing the boundary combination strategies and introducing the neural network techniques, the characteristic that the neural network automatically extracts the high-dimensional abstract features in a layered manner is fully exerted; by dividing the entityrecognition into three steps of boundary recognition, boundary combination and candidate entity recognition, the defects of a traditional sequence model are overcome, and the problem of feature sparseness generated by a traditional machine learning method is avoided to a certain extent, so that the performance of the nested named entity recognition is improved, and a very good effect is achieved.

The invention discloses a trajectory recovery method based on depth learning and Kalman filter correction, which comprises the following steps: S1, discretization of trajectory points; S2,loop NeuralNetwork and Trajectory Modeling; S3, trajectory recovery; S4, using spatio-temporal attention mechanism, obtains attention model based on sequence-to-sequence model; S5, combines Kalman filter and circulating neural network, introduces Kalman filter to optimize the mean square error, cooperatively trains the Kalman filter and the attention model, and obtains the final model. The invention providesa trajectory recovery method based on depth learning and Kalman filter correction, the transfer law between the trajectory points is modeled by the loop neural network. The attention mechanism in depth learning is used to help trajectory recovery. Finally, Kalman filter is introduced to model the movement of the object in time and space, which reduces the unexplainability and error of depth learning model, has stronger explainability, and reduces the error of trajectory recovery.

An optical defect detection method and a system thereof are disclosed. The detection method includes a process of detecting an image of an optical film by an optical detector. The image is converted into a clean detection image by conducting the following processes: uniforming the brightness, enhancing the contrast, filtering off the noise, smoothing the image and binarizing the image. A relative relation between a pixel and the surrounding pixels of the clean detection image is converted into a spatial relation sequence model. The spatial relation sequence model is compared to the different types of the defect sequence model, so that the defect type of the optical film is identified as a point defect, a lack of material defect or a ripple defect.

The invention discloses a text error correction method and device, and relates to the technical field of computers. A specific embodiment of the method comprises the steps of obtaining a to-be-corrected text according to a text error correction demand; and according to the error correction model, carrying out error correction on the to-be-corrected text, and outputting a standard text corresponding to the to-be-corrected text, the error correction model being a trained attention-based sequence to a sequence model. According to the embodiment, the text is corrected by using the trained attention-based sequence-to-sequence model, so that the calculation complexity is reduced, and the text correction accuracy is improved.

The invention relates to the technical field of passenger flow prediction, and discloses a rail transit space-time short-time passenger flow prediction method, device and equipment and a storage medium. The method comprises the steps of acquiring pull-in data and train timetable data of a historical time period, constructing an adjacency matrix according to the train timetable data; standardizingthe pull-in data and the adjacency matrix; adopting a graph convolutional neural network to extract spatial feature matrixes of the standardized pull-in data and the adjacency matrix; and extracting time features of the spatial feature matrix by adopting a sequence-to-sequence model based on a gating cycle unit and an attention mechanism so as to predict an outbound amount at the current moment. According to the method, the space-time relationship of large-scale passenger flow can be captured, high precision and high interpretability are achieved, the passenger flow distribution situation canbe mastered conveniently, and a basis is provided for passenger flow state analysis and early warning. Meanwhile, passenger flow organization is facilitated, transport capacity resources are reasonably allocated, congestion is relieved, and the service quality is improved.

An estimation method utilizing a pair of target-directional models 106 and 108 includes the steps 160 and 164 of decoding an input 142 utilizing the first and the second models 106 and 108, thereby producing k-best hypotheses 162 and 166 from each of the first and the second models 106 and 108; calculating a union of the k-best hypotheses, and re-scoring 168 each of the best hypotheses in the union utilizing the first and the second models; and selecting a hypothesis 144 with the highest score.

The embodiment of the invention provides a semantic key index creating method. The semantic key index creating method comprises the steps that properties of words of input statements and plying statements of all pairs of statements are analyzed, and semantic keys meeting preset property requirements in each statement are extracted; according to the semantic keys, all the statements in a dialogue corpus are clustered, and the statements of each category correspond to the same semantic key; all categories of statement training sequences in the dialogue corpus are utilized to obtain sequence models, and an encoding network capable of mapping statements into actual value vectors is obtained; the statements belonging to the same category are encoded by utilizing the encoding network, and actualvalue vector sets corresponding to the semantic keys are obtained; multiple actual value vectors are selected from the actual value vector sets corresponding to the semantic keys are selected to forma memory matrix, and semantic key indexes are established for the semantic keys and memory matrix keys. The embodiment of the invention also provides a semantic key index creating system. The statements generated in the embodiment of the invention have diversity and directionality.

The invention relates to a multi-target random fuzzy dynamic optimal energy flow modeling and solving method for a multi-energy coupling transmission and distribution network and belongs to the field of day-ahead scheduling plan research of electric power systems in an energy interconnection environment. The method comprises the following steps: basic data in a system scheduling period are obtained,; random fuzzy space-time sequence models for large-scale wind power, distributed power source and multi-energy loads are obtained via historical data mining; power and voltages of a power transmission network and all active distribution networks at joint nodes are used as share variables; multi-target SoS dynamic optimal energy flow models characterized by high economic performance, low carbon emission, renewable energy absorption, loss reduction and the like are built within static state security constraints; multi-energy source charge forecast can be realized through random fuzzy simulation; a Pareto solution set, an optimal compromise solution and an energy flow result can be obtained via adoption of an improved SoS layered optimizetion algorithm based on approximate dynamic programming and NSGA-11. The method can adapt to a development trend of energy interconnection, and comprehensive coordination optimization of day-ahead scheduling of transmission and distribution parties can be realized on the premise that requirements for static state safety and stabilization of systems can be satisfied.

The invention provides a seismic cycle analysis method, comprising the steps of: processing Hilbert-Huang transform for the reflection factor sequence and the time frequency analysis picture of a seismic sequence model of a seismic sequence model base; processing Hilbert-Huang transform for the seismic data of a target layer, finding a seismic sequence model of the target layer to judge the reflection character of the target layer. The invention utilizes Hilbert-Huang to analyze seismic section to attain accurate and reliable frequency spectrum and after decomposition display the effective information not displayed of the sections in the prior art, to provide data for reservoir predication.

The invention provides a locus prediction method utilizing a fuzzy locus sequence. A fuzzy locus sequence model is introduced, fuzzification processing on movement locus of a user is carried out, through a prediction method of a fuzzy time sequence, the movement locus of the user is predicted. Fuzzification processing on the locus of the user is carried out, a grid fuzzification method is designed, and the fuzzy locus sequence model is realized. The method is advantaged in that an off-group point processing mechanism is introduced, a locus prediction problem under the insufficient historical information condition can be processed, through introducing the off-group point detection mechanism and an off-group point prediction model, influence of off-group points on integral algorithm performance is reduced.