40 results about "Language modelling" patented technology

A system (and method) for off-line detection of textual topical changes includes at least one central processing unit (CPU), at least one memory coupled to the at least one CPU, a network connectable to the at least one CPU, and a database, stored on the at least one memory, containing a plurality of textual data set of topics. The CPU executes first and second processes in first and second directions, respectively, for extracting a segment having a predetermined size from a text, computing likelihood scores of a text in the segment for each topic, computing likelihood ratios, comparing them to a threshold, and defining whether there is a change point at the current last word in a window.

A system and method for speech recognition includes generating a set of likely hypotheses in recognizing speech, rescoring the likely hypotheses by using semantic content by employing semantic structured language models, and scoring parse trees to identify a best sentence according to the sentence's parse tree by employing the semantic structured language models to clarify the recognized speech.

A computer system for language modeling may collect training data from one or more information sources, generate a spoken corpus containing text of transcribed speech, and generate a typed corpus containing typed text. The computer system may derive feature vectors from the spoken corpus, analyze the typed corpus to determine feature vectors representing items of typed text, and generate an unspeakable corpus by filtering the typed corpus to remove each item of typed text represented by a feature vector that is within a similarity threshold of a feature vector derived from the spoken corpus. The computer system may derive feature vectors from the unspeakable corpus and train a classifier to perform discriminative data selection for language modeling based on the feature vectors derived from the spoken corpus and the feature vectors derived from the unspeakable corpus.

A system for extracting a key phrase from a document includes a neural key phrase extraction model (“BLING-KPE”) having a first layer to extract a word sequence from the document, a second layer to represent each word in the word sequence by ELMo embedding, position embedding, and visual features, and a third layer to concatenate the ELMo embedding, the position embedding, and the visual features to produce hybrid word embeddings. A convolutional transformer models the hybrid word embeddings to n-gram embeddings, and a feedforward layer converts the n-gram embeddings into a probability distribution over a set of n-grams and calculates a key phrase score of each n-gram. The neural key phrase extraction model is trained on annotated data based on a labeled loss function to compute cross entropy loss of the key phrase score of each n-gram as compared with a label from the annotated dataset.

The features relate to artificial intelligence directed detection of user behavior based on complex analysis of user event data including language modeling to generate distributed representations of user behavior. Further features are described for reducing the amount of data needed to represent relationships between events such as transaction events received from card readers or point of sale systems. Machine learning features for dynamically determining an optimal set of attributes to use as the language model as well as for comparing current event data to historical event data are also included.

The invention provides a fault management method based on a Modelica model. The method comprises the following specific steps of (1) fault analysis; (2) fault classification; (3) fault modelling; (4) fault extraction; and (5) fault injection for realizing modification of parameters of a fault component in the Modelica model, wherein the fault injection includes parameter type fault injection and variable type fault injection. Through the steps, the fault management based on the Modelica model is realized, the flexible modelling of the fault component and a unified fault management interface are provided, and a unified fault injection mode is adopted to be convenient for application of subsequent search methods related to fault management. The method has the characteristics of simple Modelica language modelling, visual display, strong applicability and the like, so that the method can well adapt to a complex system with the stronger electromechanical liquid coupling, the unified and reliable fault management method is provided, the construction period of subsequent fault diagnosis platforms is reduced, and the method has popularization and application value.

The invention discloses a joint simulation system based on Modelica and a construction method thereof. The method comprises the steps of: establishing a subsystem model in each modeling software; using various different modeling software and languages, for example, a C language model is established in Visual Studio, a Simulink model is established in Simulink, an AMESim model is established in AMESim, a Fortran language model is established in Visual Studio, and the like; compiling each subsystem model to generate a dynamic link library, wherein the dynamic link library comprises a model simulation interface and a model solver; packaging each sub-calculation model into a subsystem Modelica model by utilizing an external function mechanism of Modelica; and finally, writing a master controlmodel by using a Modelica language, defining a sampling step length of each subsystem module, and scheduling each subsystem by using a solver of a master control end to realize data synchronization ofthe system. According to the technology, multi-modeling software and multi-language modeling can be supported, system simulation is divided into system scheduling and subsystem calculation, all subsystems are placed in independent processes to be solved, and the solving speed of the system is increased.

The invention belongs to the technical field of computer software and discloses a method for converting an xUML4MC model into an MSVL program, and a computer program. The method for converting the xUML4MC model into the MSVL program comprises the steps of establishing a class diagram model and activity diagram model description system; converting a class diagram and an activity diagram into an object-oriented abstract grammar tree; preprocessing and converting the object-oriented abstract grammar tree into a process-oriented abstract grammar tree; and converting the process-oriented abstract grammar tree into the MSVL program. The method is suitable for model detection system modeling; a UML-based visual modeling mode is intuitive in modeling and easy to master, and is more convenient andaccurate compared with a text language-based modeling mode; and the MSVL program generated by conversion can be directly used for model detection. The unification of software design modeling and modeldetection system modeling is realized; and the popularization of a model detection technology in the industrial circles is facilitated.

The invention discloses an encoder-decoder framework pre-training method for neural machine translation. The encoder-decoder framework pre-training method comprises the steps of: constructing a largenumber of multi-language document-level monolingual corpora, and adding a special identifier in front of each sentence to represent the language type of the sentence; processing sentence pairs to obtain training data; training monolingual data of different languages to obtain converged pre-training model parameters; constructing parallel corpora, and initializing parameters of a neural machine translation model by using the pre-training model parameters; finely adjusting model parameters of the initialized neural machine translation model through parallel corpora to finish a training process;and in a decoding stage, encoding a source language sentence by using an encoder of the trained neural machine translation model, and decoding by using a decoder to generate a target language sentence. According to the encoder-decoder framework pre-training method, the model has language modeling capability and language generation capability, the pre-training model is applied to the neural machinetranslation model, the convergence rate of the model can be increased, and the robustness of the model is improved.

The present invention relates to a method for statistical language modeling and speech recognition providing a predetermined number of words in a predetermined order, providing a training corpus comprising a predetermined number of sequences of words, wherein each of the sequences consist of the provided predetermined number of words in the predetermined order followed by at least one additional word, providing word candidates and for each word candidate calculating on the basis of the training corpus the probability that the word candidate follows the provided predetermined number of words, determining at least one word candidate for which the calculated probability exceeds a predetermined threshold, wherein the probabilities of the word candidate are calculated based on smoothed maximum-likelihood probabilities calculated for the sequences of words of the training corpus, wherein the smoothed maximum-likelihood probabilities are larger than or equal to a predetermined real positive number that is less than or equal to the inverse of the predetermined number of sequences of words of the training corpus.

Content generation leverages an unsupervised, generative pre-trained language model (e.g., a generative-AI). In this approach, a model derived by applying to given content relevant competitive content and one or more optimization targets is received. Based on optimization criteria encoded as embedding signals in the model, a determination is made regarding whether a template suitable for use as an input to the generative-AI exists in a set of templates. If so, the model embedding signals are merged into the template, or the template itself is transformed using the embedding signals, in either case creating a modified template. If, however, no template suitable as the input exists, the model and other information are input to a natural language processor to generate a generative-AI input. Either the modified template or the generative-AI input, as the case may be, is then applied through the generative-AI to generate an output competitively-optimized with respect to the optimization targets.

The present invention relates to an improved artificial neural network for predicting one or more next items in a sequence of items based on an input sequence item. The improved artificial neural network has greatly reduced memory requirements, making it suitable for use on electronic devices such as mobile phones and tablets. The invention includes an electronic device on which the improved artificial neural network operates, and methods of predicting the one or more next items in the sequence using the improved artificial neural network.

The invention refers to a natural language processing system configured for receiving an input sequence c_i of input words representing a first sequence of words in a natural language of a first text and generating an output sequence of output words representing a second sequence of words in a natural language of a second text and modeled by a multinominal topic model, wherein the multinominal topic model is extended by an incorporation of full contextual information around each word v_i, wherein both preceding words v_<i and following words v_>i around each word v_i are captured by using a bi-directional language modelling and a feed-forward fashion, wherein position dependent forward hidden layers {right arrow over (h)}_i and backward hidden layers for each word v_i are computed.

The invention relates to a target tracking method based on a language building mode twin network, and the method comprises the steps: S1, obtaining a video containing the continuous motion of a target, and making a training data set according to the video; s2, training a twin neural network according to the training data set; s3, keeping parameters of the twin neural network unchanged, and training a target position extraction network; s4, performing joint training on the trained twin neural network and the trained target position extraction network to obtain a language building mode twin network; and S5, acquiring a real-time image of the target to be tracked, and performing real-time tracking on the target to be tracked by using the language modeling twin network. Expert experience knowledge does not need to be integrated into the algorithm, and the implementation process is simpler and more convenient. In addition, the system is high in expansibility and can be connected to a more general intelligent system.

The invention discloses a semantic meaning understanding method fusing semantic information, and belongs to the technical field of language information processing. The problem that an existing language modeling method is high in complexity and cannot give consideration to the effect is solved. The method comprises the following steps of: firstly, processing a language by taking each word as a unitaccording to two paths, wherein the left path is a word encoder + RNN + word decoder, and the output of the left path is recorded as wl; the right path comprises a synonym encoder+RNN + synonym decoder+a word decoder + sigmoid, and the output of the right path is recorded as wr; and then fusing the outputs of the two paths. The method is mainly for language meaning understanding.

The invention discloses a streaming speech recognition system and method based on a non-autoregression model. The method comprises the following steps: S11, extracting an acoustic feature sequence; s12, generating an acoustic coding state sequence; s13, generating an acoustic coding state sequence; s14, CTC output probability distribution and connection time sequence loss are calculated; s15, performing alignment by using a viterbi algorithm; s16, inputting section by section and calculating joint cross entropy loss; s17, calculating a gradient according to the joint loss of the joint time sequence loss and the joint cross entropy loss, and carrying out back propagation; s18, circularly executing the steps S12 to S17 until the training is completed; the system comprises an acoustic feature sequence extraction module, a streaming acoustic encoder, a CTC linear transformation layer and a non-autoregressive decoder which are sequentially connected with one another. According to the invention, non-autoregressive decoding is carried out on the input audio segments segment by segment, so that the streaming reasoning speed is improved. And the loss of the language modeling capability is avoided.

The present invention relates to an improved artificial neural network for predicting one or more next items in a sequence of items based on an input sequence item. The artificial neural network is implemented on an electronic device comprising a processor, and at least one input interface configured to receive one or more input sequence items, wherein the processor is configured to implement theartificial neural network and generate one or more predicted next items in a sequence of items using the artificial neural network by providing an input sequence item received at the at least one input interface and a side input as inputs to the artificial neural network, wherein the side input is configured to maintain a record of input sequence items received at the input interface.

The invention provides an SIMSCRIPT language-oriented discrete event simulation graphical modeling method, which comprises the following steps of: adding entities, routines, events and other primitives into a canvas serving as a primitive bearing container in a dragging manner according to entity flow graphs, activity cycle graphs and other modeling technologies, and representing an interaction relationship among the entities, the routines and the events through connecting lines, according to the method, primitives can be drawn and managed in the canvases, the canvases can be divided into a plurality of canvases according to the calling relation and the hierarchical relation, all the canvases can be stored as engineering files of specific formats, and the engineering files can automatically generate SIMSCRIPT simulation codes according to mapping rules. The discrete event simulation program is built based on the SIMSCRIPT simulation language and a graphical dragging mode, the problem that SIMSCRIPT grammar and manual code writing need to be familiar with when SIMSCRIPT language modeling is adopted is solved, the modeling process can be clear and visual, model reuse is simpler, thelearning cost is lower, user groups are wider, and communication between domain experts and modeling personnel is facilitated.

The invention provides a Latin Zhuang language vector representation and calculation method. A representation and calculation model composed of a Latin Zhuang language preprocessing module, a Latin Zhuang language modeling module, a Latin Zhuang language vector representation normalization module and a Latin Zhuang language vector representation and calculation module which are sequentially connected is included. After a text written by the Latin Zhuang language is processed through the representation and calculation model, the vector representation of the Latin Zhuang language text can be finally obtained. The method solves the problem of vector representation and calculation of the Latin Zhuang language, and can be used for automatic processing and application of the Latin Zhuang language.

The invention discloses a language modeling system structure search method for translation tasks, and the method comprises the following steps: obtaining and processing training data through the Internet, and modeling and training a network structure representation space; carrying out normalization operation on structure parameter values of meta-structure topology and operation in the training process; optimizing structure parameters and model parameters of the used model, and adjusting and optimizing a network structure and target parameters; further obtaining a discretized final structure according to the weight difference of different topologies and operations obtained after tuning, wherein the search result comprises the topological structure of the meta-structure and the operation used between the nodes; and circularly unfolding the searched meta-structures by using a connection mode between the meta-structures to obtain an integral model, performing parameter tuning on the model again by using training data, and finally training until convergence. The method greatly improves the possibility that the optimal solution of the model structure falls into the representation space of the search structure, thereby improving the effectiveness of the network structure search method.