120 results about "Distributed representation" patented technology

The invention provides a Mongolian-Chinese machine translation method for enhancing semantic feature information based on a Transformer model. The method comprises the following steps: firstly, starting from the language characteristics of Mongolian, finding out the characteristics of the additional components of the Mongolian in terms of stem, affixes and lattices, and merging the language characteristics into the training of a model; secondly, distributed representation for measuring the similarity between the two words is taken as a research background, and the influence of depth, density and semantic coincidence degree on the concept semantic similarity is comprehensively analyzed; in the translation process, a Transformer model is adopted, and the Transformer model is a multi-layer encoder which performs position encoding by using a trigonometric function and is constructed on the basis of an enhanced multi-head attention mechanism. A decoder architecture, which completely dependson the mechanism of attention to draw the global dependency between the input and the output, eliminates recursion and convolution.

A spatial and temporal memory system (STMS) processes input data to detect whether spatial patterns and / or temporal sequences of spatial patterns exist within the data, and to make predictions about future data. The data processed by the STMS may be retrieved from, for example, one or more database fields and is encoded into a distributed representation format using a coding scheme. The performance of the STMS in predicting future data is evaluated for the coding scheme used to process the data as performance data. The selection and prioritization of STMS experiments to perform may be based on the performance data for an experiment. The best fields, encodings, and time aggregations for generating predictions can be determined by an automated search and evaluation of multiple STMS systems.

The invention relates to an event trigger word extraction method, in particular to an event trigger word extraction method based on a document level attention mechanism, comprising the following steps: (1) preprocessing training corpus; (2) performing word vector training by using PubMed database corpus; (3) constructing a distributed representation way of a sample; (4) constructing a characteristic representation way based on BiLSTM-Attention; (5) adopting CRF learning, and acquiring an optimal sequence labeling result of the current document sequence; and (6) extracting event trigger words.The method provided by the invention has the advantages that firstly a BIO tag labeling way is adopted, and recognition including multi-word trigger word recognition is realized; secondly a corresponding simple word and characteristic distributed representation way is constructed for a trigger word recognition task; and thirdly, a BiLSTM-Attention model is proposed, a distributed representation structure specific to the currently input document level information is realized by introducing an Attention mechanism, and trigger word recognition effect is improved.

The invention discloses an information retrieval-based question and answer system and method for knowledge graph energization, which integrally improve the question and answer effect of the system, expand the user consultation range and improve the question feedback accuracy. According to the technical scheme, the system comprises a knowledge map database for storing domain knowledge map information; a word segmentation and part-of-speech tagging module which segments the user questions and tags the part-of-speech of the user questions; an entity identification and link module which identifiesentities in the user questions and links the entities to nodes in the knowledge graph database; an intention understanding module which obtains an intention understanding result of the user problem based on the entity link result and the distributed representation vector; a retrieval module which retrieves a plurality of corresponding question and answer pairs as roughing results according to theinformation in the user questions based on the retrieval data source; a sorting module which is used for resorting the roughing results by utilizing the distributed representation vectors of the entities; and a semantic matching module which scores the reordering result by using the distributed representation vector of the entity and finally outputs an answer.

An approach is provided for automatically generating and processing concept vectors by extracting concept sequences from one or more content sources and generating a first concept vector for a first concept by supplying the concept sequences as inputs to a vector learning component, such that the first concept vector comprises information interrelating the first concept to other concepts in the concept sequences which is inferred from the concept sequences.

The invention discloses a relation extraction method in combination with clause-level remote supervision and semi-supervised ensemble learning. The method is specifically implemented by the following steps of 1, aligning a relation triple in a knowledge base to a corpus library through remote supervision, and establishing a relation instance set; 2, removing noise data in the relation instance set by using syntactic analysis-based clause identification; 3, extracting morphological features of relation instances, converting the morphological features into distributed representation vectors, and establishing a feature data set; and 4, selecting all positive example data and a small part of negative example data in the feature data set to form a labeled data set, forming an unlabelled data set by the rest of negative example data after label removal, and training a relation classifier by using a semi-supervised ensemble learning algorithm. According to the method, the relation extraction is carried out in combination with the clause identification, the remote supervision and the semi-supervised ensemble learning; and the method has wide application prospects in the fields of automatic question-answering system establishment, massive information processing, knowledge base automatic establishment, search engines, specific text mining and the like.

A method and apparatus are provided for automatically analyzing candidate concepts extracted from a first source text against a reference concept set comprising a plurality of concepts by obtaining a vector representation for each of the concepts in the first concept set and the reference concept set and performing a natural language processing (NLP) analysis comparison of the candidate concepts to the reference concept set to determine a similarity measure corresponding to each candidate concept and validating one or more of the candidate concepts based on the similarity measure for each candidate concept meeting a minimum similarity threshold requirement.

Provided are methods and system for recognizing characters such as mathematical expressions or chemical formulas. An example method comprises the steps of receiving and processing an image by a pre-processing module to obtain one or more candidate regions, extracting features of each of the candidate regions by a feature extracting module such as a convolutional neural network (CNN), encoding the features into a distributive representation for each of the candidate regions separately using an encoding module such as a first long short-term memory (LSTM) based neural network, decoding the distributive representation into output representations using a decoding module such as a second LSTM-based recurrent neural network, and combining the output representations into an output expression, which is outputted in a computer-readable format or a markup language.

The invention relates to a biological event extracting method, in particular to a biological event extracting method based on a candidate event element attention mechanism. The method includes the following steps that (1), a training corpus is pre-processed; (2), a PubMed database corpus is used for word vector training; (3), a distributed expression mode of sequences is established; (4), a feature expression mode based on BiLST-Attention is established; (5), an optimal sequence labeling result of a current document sequence is acquired through CRF learning; (6), biological events are extracted. The method has the following advantages that event element recognition aiming at trigger words is carried out in a sequence labeling mode, and multiple event elements can be recognized for one trigger word; an Attention layer aiming at candidate event elements is established, and compared with a traditional sequence labeling method, attention weights of the candidate event elements can be increased; the steps of generating events are simplified.

The invention discloses a deep neural network-based text consistency analysis method. After a section of text is input, a distributed method is adopted to translate each word in sentences into vectorsto form distributed sentence matrices, the words repeatedly appearing in the adjacent sentences are counted, and repeated information of the adjacent sentences is added through a manner of expandingdimensionality of the matrices; a convolutional neural network is utilized to learn distributed representations of the sentences, and important features of logic, semantics, syntax and the like in thesentences are extracted to form sentence vectors; and similarity degrees between the adjacent sentence vectors are calculated to add context association contents, finally, a neural network is continuously trained, and probability of text consistency is output. The method is characterized in that complex artificial feature extraction operations do not need to be carried out, external resources arealso not relied on, and compared with existing consistency analysis technology, the method provided by the invention has a great improvement in an accuracy rate, and has a better practical value.

Coordinate data is encoded as a distributed representation for processing and analysis by a machine-intelligence system such as a hierarchical temporal memory system. Input coordinates represented in coordinate space having at least one dimension are obtained. The input coordinates change over time. A corresponding region around each of the input coordinates in the coordinate space is determined. A subset of coordinates within the corresponding region for each of the input coordinates is selected. A distributed representation for each of the input coordinates reflecting the selected subset of coordinates for each of the input coordinates is generated. The distributed representation may be provided to one or more processing nodes for detection of temporal sequences and spatial patterns. Based on discrepancies between predicted coordinate data and actual coordinate data, anomalies may be detected.

The invention provides a semantic matching method for question and answer texts. The method can effectively solve problems in related technologies. For example, in the question and answer text semantic matching technology based on a deep learning model in the related technology, only context local semantic feature information can be provided, background global feature information and syntactic feature information of a question and answer text are lacked, feature unification is caused, and semantic matching information of the question and answer text cannot be completely embodied. The inventionprovides a question and answer text semantic matching method based on multi-level features and deep learning, word and syntactic structure distributed representation is carried out on words and syntactic information of the question and answer text; and the context local feature information and the syntactic structure feature information of the question and answer text are extracted by using the recurrent neural network, and then the background global feature information is extracted by using an attention mechanism, so that the feature information of the question and answer text is richer, andthe semantic matching accuracy of the question and answer text is improved.

Performing semantic analysis on a user-generated text string includes training a neural network model with a plurality of known text strings to obtain a first distributed vector representation of the known text strings and a second distributed vector representation of a plurality of words in the known text strings, computing a relevance matrix of the first and second distributed representations based on a cosine distance between each of the plurality of words and the plurality of known text strings, and performing a latent dirichlet allocation (LDA) operation using the relevance matrix as an input to obtain a distribution of topics associated with the plurality of known text strings.

The invention relates to the technical fields including natural language processing, information retrieval, medical data mining and the like, in particular to a similar case recommendation system based on word and phrase distributed representation, and a corresponding method applied to an Internet inquiry platform. The system comprises a data module, a recommendation module, an evaluation module and an on-line module, wherein the data module comprises a data acquisition submodule, a data storage submodule, a data preprocessing submodule, a word segmentation submodule and a word vector training submodule; the recommendation module comprises a decision submodule, a semantic similarity algorithm submodule and a recommendation sorting submodule; the on-line module comprises a recommendation submodule and a feedback submodule; the data module transmits effective data to the recommendation module, the recommendation module receives data from the data module and the index of the evaluation module, recommends an associated case and transmits a recommendation result to the on-line module; and the on-line module transmits the recommendation result to a user, and meanwhile, the user returns the feedback of the recommendation result to the on-line module.

A spatial and temporal memory system (STMS) processes input data to detect whether spatial patterns and / or temporal sequences of spatial patterns exist within the data, and to make predictions about future data. The data processed by the STMS may be retrieved from, for example, one or more database fields and is encoded into a distributed representation format using a coding scheme. The performance of the STMS in predicting future data is evaluated for the coding scheme used to process the data as performance data. The selection and prioritization of STMS experiments to perform may be based on the performance data for an experiment. The best fields, encodings, and time aggregations for generating predictions can be determined by an automated search and evaluation of multiple STMS systems.

The invention proposes a state transition and neural network-based Chinese chunk parsing method. The method comprises the steps of converting a chunk parsing task into a serialized tagging task; tagging a sentence by using a state transition-based framework; scoring a transition operation to be carried out in each state by using a forward neural network in the tagging process; and taking a distributed representation characteristic of words and part-of-speech tagging learned by utilizing a two-way long short-term memory neural network model as an additional information characteristic of a tagging model, thereby improving the accuracy of chunk parsing. Compared with other Chinese chunk parsing technologies, the Chinese chunk parsing method has the advantages that characteristics of chunk levels can be more flexibly added by using the state transition-based framework, combination modes among the characteristics can be automatically learned by using the neural network, the useful additional information characteristic is introduced by utilizing the two-way long short-term memory neural network model, and the combination of the state transition-based framework, the neural network and the two-way long short-term memory neural network model effectively improves the accuracy of chunk parsing.

A method enables identification of a similarity level between a user-provided data item and a data item within a set of data documents. The method includes a representation generator determining, for each term in an enumeration of terms, occurrence information. The representation generator generates, for each term, a sparse distributed representation (SDR) using the occurrence information. The method includes receiving, by a filtering module, a filtering criterion. The method includes generating, by the representation generator, for the filtering criterion, at least one SDR. The method includes generating, by the representation generator, for a first of a plurality of streamed documents received from a data source, a compound SDR. The method includes determining, by a similarity engine executing on the second computing device, a distance between the filtering criterion SDR and the generated compound SDR. The method includes acting on the first streamed document, based upon the determined distance.

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.

A large-scale image semantic retrieval method is provided. The method comprises: using an unsupervised deep learning model to train the network to acquire feature vectors of pictures, and comprehensively considering a semantic relationship between the text descriptions of the pictures to achieve large-scale picture retrieval; for processing the feature vectors of the pictures, using a generated countermeasure network composed of a 4-6th layer discriminant network and a 4-6th generation network to extract the features of the pictures; for processing the texts of the pictures, using a distributed expression method of word vectors to obtain the picture vectors, and using the word nesting to describe the semantic information of the picture; using a clustering method to cluster the retrieved pictures, and displaying only one of a certain class of commodities to the user through clustering, so that the time for the user to look for commodities can be reduced; obtaining picture text description vectors through the trained word vectors; connecting the text vector and the picture vector together as the feature representation of the picture; and clustering the pictures through the k-means++.

The invention discloses an answer selection method for a question and answer system. A hierarchical length adaptive neural network structure is adopted to generate sentence distributed representationof questions and answers. The method aims to extract high-quality sentence features by adopting different neural feature extractors according to the lengths of input sentences. The method comprises the following steps: firstly, generating a word distributed representation for each word in an input statement by connecting a fixed word insert and a fine-tuning word insert; then, using a feature extractor based on BiLSTM for short sentences, and using a feature extractor based on Transformer for long sentences; and finally, generating a sentence vector representation for measuring the correlationbetween the question and the candidate answer by using an attention pooling layer considering the interaction between the question and the answer sentence. Experimental results show that the answer selection model based on the length adaptive neural network provided by the invention can be greatly improved in the aspects of MAP and MRR compared with the most advanced baseline.

The invention discloses a question classification method in a computer question and answer system, which is used for classifying and predicting questions of a user based on context data of the user. The method mainly comprises the following steps: 1, constructing context environment information influencing user intention according to an intelligent question and answer system application field; 2, obtaining the situation information data of the user by means of user portrait construction, log analysis and sensor reading; thirdly3, designing a problem intention prediction network based on an attention mechanism is designed, embedding the situation information is embedded into a problem, forming the problem distributed representation considering the situation is formed, and substituting the problem is substituted into the model for prediction; and 4, performing the model training and prediction. The method solves the problems that the user intention identification only depends on the problem of natural language dialogue, the problem of incomplete expression of the user is possible, and the expression meanings of the questions under different backgrounds are different, and the intention identification accuracy is improved.

The invention discloses an image description method based on a distributed word vector CNN-RNN network, and the method is characterized by comprising the following steps of 1), generating a distributed representation word vector; 2) generating a distribution representation label; 3) generating the distributed representation semantic tags; 4) carrying out the network design; and 5) generating the descriptive statements of the image. The method is introduced into an original CNN-RNN model, so that a more accurate result can be generated, and the CNN subnet provides richer semantic content for the RNN subnet, so that the whole CNN-RNN network model can still maintain the advantage of structuralization. According to the method, the low-dimensional dense distribution representation can be easily embedded into a large number of words to form a complete semantic space, the visual content can be better mapped into the semantic space, and the visual content can be more accurately summarized andthe vector space can be more fully utilized to supervise the CNN optimization direction based on the supervision signal of the distribution representation word vector design.

The invention discloses a knowledge map distributed representation method capable of encoding relationship semantic diversity structure, From the point of view of graph structure diversity in knowledge map, mining and summarizing the subgraph structure which contains abundant semantic information, and connecting the subgraph structure with the related semantic information, such as reflexivity, symmetry, transitivity and so on; The method of the invention can better encode the diversity of the graph structure, Therefore, it can better capture the semantic information in the knowledge map, achieve better completion and prediction effect, achieve better performance in the link prediction task, and has a strong scalability, which has a good practical value for the application of the semantic diversity structure of encoding relations in the distributed representation of the knowledge map.

The invention discloses an equipment resource configuration optimization method based on knowledge graph driving. The method comprises the steps that a knowledge graph of manufacturing resources is organized and generated, and machining process knowledge modeling and workshop machining equipment modeling are involved; the formed knowledge graph is subjected to a distributed representation learningand resource prediction matching method, and the relation between the knowledge graph and mining manufacturing resource implicit knowledge is further represented; when a processing order is issued, acandidate equipment set is constructed in combination with a fuzzy hierarchy method based on knowledge graph driving, and further the candidate equipment set is evaluated and screened through a community load model to complete reconstruction of the manufacturing unit; and finally, configuration optimization of an order task is completed by utilizing an optimization algorithm of the resource configuration mathematical model, an available processing equipment set link is formed, and a production process is guided. According to the method provided by the invention, the data processing capabilityis enhanced to a great extent, the equipment utilization rate and the equipment processing flexibility are improved, and meanwhile, the optimal configuration cost of manufacturing resources requiredby a processing task is also reduced through integrated knowledge reuse.

The invention relates to computer engineering and can be used for developing new-architecture multiprocessor multithreaded computers. The aim of the invention is to produce a novel method for organizing a computer, devoid of the disadvantageous feature of existing multithreaded computers, i.e., overhead costs due to the reload of thread descriptors. The inventive method encompasses using a distributed presentation which does not require loading the thread descriptors in the computer multi-level virtual memory, whereby providing, together with current synchronizing hardware, the uniform representation of all independent activities in the form of threads, the multi-program control of which is associated with a priority pull-down with an accuracy of individual instructions and is totally carried out by means of hardware.

The invention provides a target recognition method based on brain-like cross-media intelligence for an unmanned autonomous system. The method comprises the following steps: 1. acquiring video and audio data of a target scene collected by an unmanned autonomous system, preprocessing the video and audio data, extracting language characters from the video and audio data, adopting distributed representation of the language characters, and obtaining a word vector; 2, combining that spatio-temporal context information of the object to carry out significant calculation on the preprocessed video-audiodata and the word vector; 3, training a multimodal fusion neural network according to that significance calculation result and the spatio-temporal context information of the object, and extracting attribute semantic information of the audio-visual data and the word vector; 4, according to that spatio-temporal context information of the object, the attribute semantic information and the prior knowledge of the target knowledge map, adopting cognitive calculation and Bayesian inference to obtain the target semantic information in the target scene. The invention can effectively improve the intelligent target recognition effect of the unmanned autonomous equipment.

The invention discloses an omission recovery method for short text understanding. The model is mainly divided into an embedding layer, a coding layer and a decoding layer. The embedding layer is usedfor obtaining distributed representation of discrete words; the coding layer is used for mining features of the text; and the decoding layer generates a result after completion omission by using the features extracted by the coding layer. The method has the beneficial effects that the given model assumes that words in the sentence are possibly omitted, the structure information of the to-be-recovered sequence is fully considered when the model is trained and predicted, and the problem of 'sick sentences' caused by seq2seq is effectively solved. Besides, in the aspect of short text understanding, the model fuses a cross attention mechanism and a self-attention mechanism, and compared with a seq2seq model, more and deeper text features can be extracted.