53 results about "Sentence word" patented technology

Classification of sequences, such as the translation of natural language sentences, is carried out using an independence assumption. The independence assumption is an assumption that the probability of a correct translation of a source sentence word into a particular target sentence word is independent of the translation of other words in the sentence. Although this assumption is not a correct one, a high level of word translation accuracy is nonetheless achieved. In particular, discriminative training is used to develop models for each target vocabulary word based on a set of features of the corresponding source word in training sentences, with at least one of those features relating to the context of the source word. Each model comprises a weight vector for the corresponding target vocabulary word. The weights comprising the vectors are associated with respective ones of the features; each weight is a measure of the extent to which the presence of that feature for the source word makes it more probable that the target word in question is the correct one.

The invention discloses a method and a device for detecting text information in a video and electronic equipment. The method for checking the text information in the video comprises the following steps: extracting a target picture from a to-be-detected video, wherein the target picture comprises a key frame in the to-be-detected video; extracting text information from the target picture, and performing text sentence word segmentation on the text information to obtain sentences after word segmentation; further carrying out vector conversion on the sentences after word segmentation to obtain word vectors of segmented words in the sentences; and finally inputting the sentence subjected to word segmentation and the word vector obtained by conversion into a text classification model; carrying out semantic recognition through the text classification model, and outputting a semantic recognition result representing whether the text information contains characters with preset semantics or not so as to realize the illegal video character detection, namely, by extracting the key frame in the video and semantically identifying the character information, the problem that the illegal video character cannot be detected due to the change of a simple character expression mode is avoided, and the accuracy and the detection efficiency of the illegal video character detection are improved.

The invention relates to a neural machine translation method incorporating dependency relation, which is designed for obtaining more accurate neural translation model. The invention integrates the dependency relation neural machine translation method, analyzes the dependency tree of the source sentence, and determines the relevance information between the source sentence words and the words. Basedon the dependency relation information, the dependency relevance loss Delta dep is determined, and the overall loss of the sentence pair network is obtained. The invention adds a self-attention mechanism at the source end, and integrates the self-attention mechanism into dependency guidance.

The invention discloses a knowledge base construction method and device oriented to a program design field question-answering system. The method comprises the following steps that: according to the knowledge contents of an online evaluation system, preliminarily establishing an intelligent question-answering knowledge base; obtaining a user question, carrying out sentence word segmentation on thequestion, carrying out synonym replacement, and carrying out similarity calculation with contents in the knowledge base; if a user is satisfied with a current returning answer, according to question answering, synchronously updating an intelligent question-answering knowledge base, and applying a code similarity algorithm to detect data in a cache in real time; and otherwise, returning a suboptimum answer to the user until the user is satisfied. By use of the system, the problems in a traditional online after class question answering system that a time difference is in the presence, search accuracy is low, efficiency needs to be improved and the like, the information acquisition speed and accuracy of the user can be effectively improved, and meanwhile, the teaching effect of the online evaluation system is optimized.

Certain aspects of the present disclosure provide techniques for generating a replacement sentence with the same or similar meaning but a different sentiment than an input sentence. The method generally includes receiving a request for a replacement sentence and iteratively determining a next word of the replacement sentence word-by-word based on an input sentence. Iteratively determining the next word generally includes evaluating a set of words of the input sentence using a language model configured to output candidate sentences and evaluating the candidate sentences using a sentiment model configured to output sentiment scores for the candidates sentences. Iteratively determining the next word further includes calculating convex combinations for the candidate sentences and selecting an ending word of one of the candidate sentences as the next word of the replacement sentence. The method further includes transmitting the replacement sentence in response to the request for the replacement sentence.

The invention provides a method and system for generating a braille file. The method includes the steps that text of an object file is obtained so as to obtain text content information; based on the text content information, sentence word segmentation is conducted on the text to obtain sentence word segmentation information; based on the sentence word segmentation information, segmented-word Pinyin labeling is conducted on the text so as to obtain segmented-word Pinyin information; according to the segmented-word Pinyin information, braille conversion is conducted on the test so as to obtain braille character information; the method for generating the braille file further includes correction processing; the correction processing includes at least one of the following steps that in response to a user selection, ambiguity correction is conducted on the test so as to update the text content information, and the updated text content information is used for subsequent sentence word segmentation processing; in response to a user selection, sentence word segmentation correction is conducted on the text so as to update the sentence word segmentation information, and the updated sentence word segmentation information is used for subsequent segmented-word Pinyin labelling; in response to the user selection, segmented-word Pinyin correction is conducted on the text so as to update segmented-word Pinyin information, and the updated segmented-word Pinyin information is used for subsequent braille conversion.

The invention relates to a sentence semantic distance measurement method. The method comprises the following steps: firstly, carrying out word segmentation and stop word removal preprocessing on a sentence data set; selecting a word meaning similarity scheme, and setting a threshold value to execute normalization of synonymous words and synonymous words; then, calculating the vector space distanceof the two statements by combining smooth inverse frequency weighting and common component removal; measuring the word order distance of the two statements according to the out-of-order degree; calculating the semantic dependency distance of the two statements by combining the semantic dependency quintuple features; and finally, carrying out hybrid weighting calculation on the vector space distance, the word order distance and the semantic dependency distance. Measurement is from three dimensions of sentence vector representation, sentence word sequence and sentence component dependency, andfinally a final semantic distance is obtained in a weighted summation manner. A word level calculation means is utilized, and a sentence level operation idea is absorbed, and through introduction andcreative combination of a vector space distance, a word order distance and a semantic dependency distance, the semantic distance of the sentences is more comprehensively and reasonably measured.

The invention discloses a method for automatically mining a corresponding citing fragment and a citied literature original content fragment. The method comprises the following steps: extracting a sentence which cites references as a citing fragment; carrying out sentence segmentation and numbering on the references cited by the citing fragment; carrying out word segmentation on each sentence in the citing fragment and the references to form citing fragment word groups and reference sentence word groups, and calculating similarity between the sentence in the citing fragment and the sentence in the references; and according to the calculated similarity of the sentences, sorting the sentences, extracting the sentence, which has the highest similarity with the citing fragment, in the references, and taking the extracted sentence as the cited literature original content fragment corresponding to the citing fragment. By use of the method provided by the invention, corpus training does not need to be prepared in advance, calculation complexity is low, various similarity calculation methods can be flexibly realized, and high accuracy and a high recall rate are realized.

The invention requests to protect a mixed multi-feature sentence similarity calculation method and system, and a storage medium. The mixed multi-feature sentence similarity calculation method comprises the following steps: obtaining a test set and a training set for sentence similarity calculation, and obtaining a word vector corresponding to each word through a word vector model; calculating sentence word vector similarity through a computer by utilizing weighted sum to remove non-information noise from the word vectors based on a smooth inverse frequency algorithm; based on a word dependencytriple structure, respectively calculating the similarity between the test sentence and the sentence dependency syntax with the top ten screened similarities; and based on the sentence mixing similarity calculated by the two obtained sentence vectors, adjusting an optimization coefficient beta by adopting a P@N and MRR (mean sorting reciprocal) parameter determination method to obtain a sentencewith the maximum sentence similarity with the sentence in the training set. According to the mixed multi-feature sentence similarity calculation method, the characteristics of keywords, word vectors,syntactic structures and the like in sentences are considered, so that the deep meanings of the sentences are expressed more accurately, and the similarity of sentence contents is judged correctly.

The invention provides an intelligent dialogue system response method based on tree kernel computing and an intelligent dialogue system. The response method comprises the following steps: training Chinese words to obtain a word vector table; acquiring a question sentence A1 input by a user, and selecting one question sentence as A2 by the system in sequence; inputting A1 and A2 into a word segmentation tool to obtain sentence word sets C1 and C2; inputting C1 and C2 into a syntactic analyzer to obtain a syntactic tree T1T2; trimming the T1 and the T2; inputting the trimmed T1 and T2 into a tree kernel calculation model to obtain similarity scores of the T1 and T2; calculating to obtain similarity scores of A1 and A2 according to the similarity scores of T1 and T2; recording similarity scores of A1 and A2, and judging whether another question sentence exists or not as A2; and selecting the question sentence with the maximum similarity with the A1, and outputting the answer corresponding to the question sentence as a result. Compared with the prior art, the tree kernel computing-based intelligent dialogue system response method and the intelligent dialogue system are higher in accuracy.

The invention discloses an irony detection method based on an intra-sentence word pair relation and context user characteristics. A self-attention mechanism (also known as an internal attention mechanism) is used for analyzing word pairs with emotional polarity contradiction in a text. The method comprises the following steps: learning and fusing writing style characteristics and character characteristics of a user to obtain user embeddings as irony detection context information; and meanwhile, the sequence information of the text is coded in combination with an LSTM network. The method can better detect the irony expression, and can obtain good accuracy under the condition that the irony expression is obvious or hidden. Specifically, the invention discloses an irony detection model basedon an intra-sentence word pair relation and context user characteristics. The method provides a judgment basis for irony detection, and is beneficial to judgment of the anti-text without obvious contradiction word pairs. Therefore, the irony detection accuracy is improved from the two aspects.

The invention provides a method for selecting candidate answers based on depth learning, which comprises the following steps: step S1, inputting a question and a candidate answers, respectively resolving the question and the candidate answers into a question word sequence and a candidate answer word sequence; step S2, model that question word sequence and the candidate answer word sequence throughthe long-short time memory network to obtain the semantic representation of the question sentence and the semantic representation of the candidate answer; step S3, selecting that word vector of the word with the highest weight value in the question sentence word sequence to initialize the knowledge memory module; step S4, calculating the knowledge representation of the question according to the knowledge information stored in the knowledge memory module and the semantic representation of the question; Step S5, calculating the similarity between the knowledge representation of the question andthe semantic representation of the candidate answer, and selecting the candidate answer with the highest similarity to output. The invention introduces a knowledge memory module in the depth learningnetwork to improve the connection between the question sentence and the candidate answer, and to improve the quality of the answer selection, so as to be better applied to the community question answering website and the question answering system.

The invention discloses a bilingual sentence automatic alignment method and device, and the method comprises the steps of obtaining an article pair set, enabling each article pair to comprise a source language article S and a target language article T, dividing sentences of an article, and carrying out the statistics of the relative length of each sentence and the relative position of each sentence in the article; determining word similarity between sentences si in the source language article S and sentences tj in the target language article T by utilizing a word vector model; calculating the distance between the sentence in the source language article S and the sentence in the target language article T by utilizing the inter-sentence word similarity, the sentence relative length difference and the relative position difference of the sentence in the article, taking the relative length of the sentence as the information amount, minimizing the sum of the products of the distance and the information amount as an information transfer optimization model, and solving the model to establish an alignment relationship. According to the invention, alignment between sentences is converted into searching for an optimal transportation strategy, and under the condition that work is minimum, all information of a source language article is transferred into a target language article.

The invention discloses a Chinese word segmentation method based on depth learning and forgetting algorithm, which comprises the following steps: 1. scanning sentences word by word to obtain natural language; dividing the scanned natural language into word sequences by depth learning word segmentation method and inputting them into a first word library; 2. scanning sentences word by word to obtainnatural language; 3. 2. Scan sentence verbatim to obtain natural language. 3, the word sequence in the first thesaurus and the candidate words in the second thesaurus are fused to obtain the final word segmentation result. The fusing method is: if the word sequence in the second thesaurus corresponds to the word in depth learning, the word sequence is merged into the word; a single word in the second thesaurus, if it corresponds to a word in depth learning, is combined forward or backward to form a word. The word segmentation method of the invention can automatically detect domain knowledge,complete the function of finding new words in the unsupervised domain and improve the word segmentation effect by integrating the depth learning word segmentation method and the forgetting algorithm word segmentation method.

The invention discloses a calculating method for similarity of questions containing names in the question answering field. The similarities of a name part and a non-name part are calculated separately; considering the sentence structure, question similarity calculation is performed from the two aspects of sentence word orders and lengths, and finally according to the sentence semantic similarity and the structure similarity, weighing is performed to obtain the similarity of the whole sentence. The problem that when an existing sentence similarity algorithm is used for calculating sentences containing names, judgment conforming to human subjectivity cannot be obtained is solved. The method can more accurately work out the similarity of sentences containing names, and can be widely applied to the question answering field.

The invention provides a multi-round dialogue omission recovery method based on gated copying and masking. The method comprises the steps of obtaining original omission sentences and context text content of the original omission sentences; performing word segmentation on the text by using a word segmentation tool, and mapping a word sequence into a digital sequence by using a dictionary; using a pre-trained word vector file to represent words; based on a gating mechanism, fusing the multi-head self-attention information and a gating encoder of the Bi-GRU, and performing semantic encoding on the omitted sentence word vector sequence and the context word vector sequence; calculating soft mask features of the omitted sentence based on a soft mask mechanism; calculating probability distribution of the word list by using a mask decoder; calculating scores of the context words, and normalizing the scores by using a Softmax function to obtain context probability distribution; and adding the probability distribution of the word list and the context probability distribution by using a gating unit to obtain final omission word probability distribution, and selecting filling contents of omission sentences. The omission recovery result accuracy is improved.

The invention discloses a restatement generation method, system and equipment and a computer readable storage medium. The method comprises: based on a sequence comprising a replication mechanism to asequence framework, calculating word vectors of source sentence words and keyword words; a context vector and an attention vector, the mode selection of each target word in the target sentence is determined; if the mode is a duplication mode, the mode is a duplication mode;, if yes, calculating probability distribution of the keyword set through a softmax function according to the output of the decoder; selecting the keyword with the maximum probability as a predicted target word to be copied to a target sentence, if the keyword is in a write-in mode, calculating probability distribution of the whole dictionary through a softmax function according to the output of a decoder, and selecting the word with the maximum probability as the predicted target word to be written into the target sentence; sequentially generating target words to obtain a whole target sentence; the generated restatement keywords conforming to the user given keywords can be directly copied from the original sentences, the matching conforming to the user habits is generated from the dictionary, and the applicability is high.

The invention discloses an intelligent robot chat system and method based on similar-sentence searching. The system includes a receiving unit used for receiving a sentence input by a user, a similar-sentence calculation unit used for calculating a sentence which is in sentences included in a question and answer knowledge library and is closest to the sentence input by the user, an output unit usedfor outputting a corresponding answer in the question and answer knowledge library, and a question and answer knowledge library unit used for storing dialogues in various situations. According to thesystem, on the one hand, a method of information retrieval is adopted, an inverted index is established in the question and answer knowledge library, to-be-found vocabulary can be quickly located ininverted indexing, and index space is saved at the same time; and on the other hand, weight calculation and sentence word order distance comparison are carried out according to different part-of-speech of vocabulary, cases where words used in sentences are similar but meanings are different can be distinguished, thus the more accurate similar sentence is obtained, a more accurate searching resultis provided, and interactive experience of the user is enhanced.

The invention relates to an analyzing and typesetting method and device of a mobile terminal browser. The method comprises the following steps: (1) carrying out searching and grammatical analysis on the received network data flow character by character to generate a label; (2) generating and saving a document object model data structure according to the label; and (3) searching whether the label exists in the pre-stored event search table or not, if the label is found, judging whether the content corresponding to the label is a visible element or not, if the content corresponding to the label is a visible element, generating sentence words to be added into a webpage according to the event corresponding to the label, and displaying on a mobile terminal screen. By adopting the method and device, the content of the webpage can be analyzed and typeset simultaneously, and the characteristics of DOM (Document Object Model) data are combined, thus the invention has the advantages of saving the webpage return time, and the like.

The present invention translates the Chinese word composing the inputted Chinese sentence into the Japanese word consulting the external storing device of the Chinese knowledge data base, and translates the inputted Japanese words composing the Japanese sentence into the Chinese words. The Japanese words composing the Japanese sentence are endowed with word class consulting to the external storing device of the Japanese dictionary. The arrangement of the Japanese word composing the Japanese sentence is switched according to the word class endowed to the Japanese words in order to facilitate that the arrangement is suitable for the Japanese grammar or Chinese grammar. The conjunctions are inserted or deleted between the Japanese sentence word groups which are switched with arrangement by the word arrangement switching device.

The invention provides a sentiment analysis method and device fused with a sentiment dictionary, which can be applied to the field of artificial intelligence. The method comprises the following steps: inputting an obtained sentence in a source field and an obtained sentence in a target field into a pre-generated encoder to obtain a source field sentence word vector and a target field sentence word vector; inputting the source field word vector and the target field sentence word vector into a sentence vector encoder to respectively obtain source field sentence representation and target field sentence representation; and obtaining a sentiment classification result of the target domain by using the source domain sentence representation and the target domain sentence representation. According to the method, sentiment words are taken as prediction targets, corpora of a source domain and a target domain are jointly trained through a language model, an encoder capable of extracting sentiment analysis related features is generated, and then feature vectors of texts to be subjected to sentiment classification are generated by using a word encoder; therefore, the technical effect that the available classification result can be obtained by using less or no annotation data is achieved.

The invention discloses a question matching method in an electric power intelligent question-answering system. The method comprises the following steps: firstly, carrying out zero filling or truncation operation on a query question P input by a user and a question Q in a system knowledge base to enable a sentence sequence to be changed into a fixed length; generating a corresponding word vector inthe sentence sequence through a GloVe model; inputting the word vector into a BILSTM model to obtain a word vector with context meaning; on the basis, extracting word vector features by using a convolutional neural network to obtain feature vectors of sentences, and further extracting common features from the feature vectors; and finally, splicing the feature vectors of the question P and the question Q and the common feature vector, and inputting the spliced feature vectors into a full connection layer to obtain a similarity prediction result. According to the method, deep learning is applied to the field of electric power intelligent questions and answers, the context of sentence words can be effectively fused, user questions are accurately matched, and the accuracy of electric power intelligent questions and answers is improved.

The present invention relates to the technical field of fast input methods. The fast input method based on the scene lexicon in the translation process provided by the present invention is as follows: in the process of translating a sentence, first read the words in the sentence; then read the words from the scene lexicon Obtain the corresponding entry; inject the entry into the input method lexicon used by the user, so that after the user enters the original text or the corresponding pinyin, the corresponding appropriate translation vocabulary will be displayed at the front of the entry display column, realizing effective Targeted input; after the input of this sentence, exit the scene, clear the injected entry, and start the translation of the next sentence. On the basis of the existing input method, the present invention adopts the scheme of thesaurus and vocabulary injection, and temporarily and specifically modifies the display order (word frequency) of words during the translation input process. The invention adopts the method of combining the scene lexicon and the lexicon injection technology, and effectively improves the text input efficiency without changing the habit of the user's input method.

The invention discloses a semantic-fragment-oriented target sentiment analysis method, device and equipment and a medium, and is applied to the field of target sentiment analys.The method comprises the steps that a text sample is input into an embedding layer, and a sentence word vector, a context word vector and a target word vector are obtained; encoding the sentence word vector, the context word vector and the target word vector by using an attention encoding layer to obtain a corresponding hidden state representation; inputting the hidden state representation of the context and the target word into a multi-head attention layer to obtain a target context semantic feature representation; inputting the sentence hidden state representation into a structured self-attention layer to obtain sentence segmentation semantic feature representation; and inputting the target context semantic feature representation and the sentence segmentation semantic feature representation into a prediction layer to obtain a target word sentiment prediction result. Therefore, according to the method, the model focuses on each semantic segment, so that the noise on words is inhibited, and the accuracy of the target sentiment analysis task is improved.

The invention relates to a self-supervised public opinion comment viewpoint object classification method based on comparative learning, and belongs to the field of natural language processing. The method comprises the following steps: constructing a data set of microblog comment viewpoint object classification; k-means clustering is carried out on the basis of Word2Vec word vectors, a special self-attention mechanism is fused to obtain vector representation of comments, comment sentence representation is reconstructed, positive and negative examples of comment sentences are constructed through comment sentence word vector representation and sentence representation in the aspect of reconstruction, text features related to comment viewpoint objects are enhanced through a contrastive learning method, and the comment viewpoint objects are obtained. The unrelated distance between the sentences and the non-viewpoint objects is enlarged, so that the comment sentences are deduced and classified through the model. And finally, the comment text is classified into four case aspects: a certain institution, a party, a certain name and others, and support is provided for subsequent microblog comment abstracts.

The invention relates to a sentence sentiment classification method based on a deep representation technology and a three-way decision, and the method comprises the steps: (1) carrying out the sentence preprocessing, deleting punctuation marks in a sentence, and carrying out the Chinese word segmentation of the sentence; (2) inputting the word sequence after Chinese word segmentation into a deep learning training model to generate a sentence word vector matrix; and (3) inputting the word vector matrix generated in the deep learning training model into a three-way decision classifier so as to obtain an emotion classification result predicted by the model. According to the method, the judgment result of the new sentiment sentence can be predicted according to previous judgment, and the method is widely applied to sentiment classification of social media comments.

The invention discloses a relationship extraction method combining a neural network and feature calculation. The method comprises the following steps of 1, performing vector mapping on a text based ona random word vector; 2, extracting atomic features in the sentence, performing feature calculation on the atomic features to obtain composite features, and performing vector mapping on the compositefeatures; 3, performing convolution pooling operation on the word vector matrix through a neural network to extract features; 4, splicing a result after convolution pooling with the composite featurevector in the sentence; and step 5, performing full connection, and predicting a result on a Softmax layer. On the basis of making full use of sentence text complete information, structure and semantic information is obtained in combination with a feature calculation method. Simultaneous introduction of neural network technology, according to the method, the characteristic that a neural network automatically extracts high-dimensional abstract features in a layered mode is brought into full play, a result obtained after sentence word vectors are input into a convolution pooling layer is combined with composite feature vectors, the problem of feature sparseness caused by the limited number of words in sentences is avoided to a certain extent, and therefore the experimental performance of arelation extraction task is effectively improved.