39 results about "Treebank" patented technology

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.

A method or module for creating an Language Neutral Syntax (LNS) representation of a sentence from a language particular syntax representation such as found in the Penn Treebank for use by different applications. The method or module includes a node generator configured to create hierarchical and dependent nodes using phrasal and constituent nodes of the language particular syntax. A node dependency generator is configured to create an unordered hierarchical dependency structure for the hierarchical and dependent nodes using a semantic relation between the hierarchical and dependent nodes derived from the language particular syntax.

The invention discloses bilingual unsupervised syntactic analysis method and system. The method comprises the following steps: 1 respectively building random syntactic analysis treebanks on a bilingual corpus source side and a target side; 2 individually training a monolingual unsupervised syntactic analysis model by computing the probability of a monolingual unsupervised syntactic analysis tree on the random syntactic analysis treebanks; 3 carrying out bilingual syntactic analysis on the monolingual unsupervised syntactic analysis model by computing the relaxation isomorphism similarity and a bilingual syntactic analysis algorithm, obtaining a bilingual syntactic analysis treebank which meets a relaxation isomorphism bilingual syntactic analysis target, so as to replace the random syntactic analysis treebanks; 4 repeating the steps 1 to 3 until the bilingual unsupervised syntactic analysis model is astringed; therefore, the better monolingual unsupervised syntactic analysis model is obtained, so as to be applied to all downstream application requiring syntactic analysis.

The invention discloses a dependency structure treebank acquisition method and system. The method comprises the steps of calling a first treebank and converting phrase structures in the first treebank into dependency structures by adopting a conversion tool of the first treebank; converting phrase structures of flat structures in the first treebank into dependency structures by utilizing a syntactic analyzer; and performing dependency relationship conversion on the dependency structures in the first treebank by utilizing a dependency relationship mapping model obtained by training to obtain a dependency structure treebank of a second treebank type. According to the method and the system, the treebank after the conversion can be combined with the original dependency structure treebank, so that the treebank scale is increased and the performance of the syntactic analyzer is improved.

The present invention provides a method and system directed to predicting implicit rhetorical relations between two spans of text, e.g., in a large annotated corpus, such as the Penn Discourse Treebank (“PDTB”), Rhetorical Structure Theory corpus, and the Discourse Graph Bank, and particularly directed to determining a rhetorical relation in the absence of an explicit discourse marker. Surface level features may be used to capture pragmatic information encoded in the absent marker. In one manner a simplified feature set based only on raw text and semantic dependencies is used to improve performance for all relations. By using surface level features to predict implicit rhetorical relations for the large annotated corpus the invention approaches a theoretical maximum performance, suggesting that more data will not necessarily improve performance based on these and similarly situated features.

The invention belongs to the field of processing for natural languages of computational linguistics, and discloses a method for processing unknown words in Chinese-language dependency tree banks. The method includes steps of A, searching all synonyms of the unknown words by the aid of synonym forests; B, computing character pattern similarity degrees among the unknown words and all the synonyms of the unknown words according to character pattern features of Chinese characters; C, extracting mapped words and information quantities of word classes of the mapped words when the character pattern similarity degrees among the unknown words and the multiple synonyms are high, and improving character pattern similarity degree computation models; D, extracting the words with the maximum character pattern similarity degrees as the optimal mapped words of the unknown words and using the extracted words as explanation for the unknown words in the tree banks. The method has the advantages that unit pairs (word classes, word classes) in dependency syntactic analysis can be recovered to unit pairs (word classes, words) or unit pairs (words, word classes) on the premise that the scales of the tree banks are no longer expanded, accordingly, the information granularity can be refined, the problem of data sparseness can be solved, and the dependency syntactic analysis performance can be improved.

The invention relates to a method for establishing a Vietnamese dependency tree bank based on an improved Nivre algorithm, and belongs to the technical field of natural language processing. The method comprises the steps of firstly, establishing an initial training corpus, an expansion corpus and a test corpus; secondly training two dependency parsing weak learners S1 and S2 based on the improved Nivre algorithm by utilizing the established initial training corpus to serve as two fully redundant views; thirdly, performing dependency parsing on the expansion corpus by utilizing the two trained weak learners S1 and S2 and building a Vietnamese dependency tree bank model; and finally, performing dependency parsing testing on the test corpus and finally establishing the Vietnamese dependency tree bank. According to the method, the powerful support can be provided for upper applications of syntactic analysis, machine translation, information acquisition and the like of a Vietnamese language; the process of manually marking a dependency relation of Vietnamese sentences can be effectively avoided, so that the time of manpower and material resources is saved; and a large amount of unmarked Vietnamese sentence level corpora can be effectively utilized for improving the accuracy of dependency parsing.

The invention relates to a phrase tree to dependency tree transformation method capable of combining Vietnamese grammatical features, and belongs to the technical field of natural language processing. The phrase tree to dependency tree transformation method comprises the following steps: firstly, constructing a Vietnamese phrase tree library; utilizing a center subnode filter table which combines the Vietnamese grammatical features and a dependency relationship annotator to finish the phrase tree to dependency tree transformation in the Vietnamese phrase tree library to obtain a first-level Vietnamese dependency tree library; according to the corpus of the manually annotated first-level Vietnamese dependency tree library, training to obtain a MSTParser model, utilizing the MSTParser model to carry out the expansion of the first-level Vietnamese dependency tree library to obtain an expanded second-level Vietnamese dependency tree library; and utilizing a dependency relationship corrector to correct the corpus of the expanded second-level Vietnamese dependency tree library to obtain a final three-level Vietnamese dependency tree library. The method avoids a process that the Vietnamese dependency tree library is manually collected and annotated, saves manpower and time for constructing the tree library, and obviously improves accuracy.

The invention relates to a dispute focus detection method and device based on a deep learning mixed model, belonging to the field of natural language processing. The method includes the following steps: ①Constructing a tree bank of disputed focus; ②Completing the data labeling and obtaining a data set; ③Obtaining a complete and trainable data set; ④Preprocessing the data set obtained in step S3 for Chinese data; ⑤Using BERT‑ The wwm model obtains the text word vector matrix; ⑥Use the LSTM network model to extract the global semantic features of the text; use various convolution kernels of the TextCNN model to extract the local semantic features of different granularities of the text; average the probability results of the two models, set The threshold is predicted, and the output probability exceeds the threshold of controversy. Aiming at the problem that a single model cannot capture and utilize multi-level semantic features at the same time, the present invention provides a mixed model method for predicting the focus of disputes, which greatly improves the prediction accuracy.

The invention relates to an automatic tree bank conversion method and system based on pattern embedding, which is designed to obtain an accurate supervised conversion model. The present invention is based on the automatic tree bank conversion method of pattern embedding, and determines the word w i and the word w j pattern; the word w i and the word w j The pattern of is transformed into the corresponding pattern embedding vector; the word w in the source tree i , word w j , the smallest common ancestor node w a The dependency labels corresponding to the three are respectively transformed into dependency embedding vectors; the pattern embedding vector and the three dependency embedding vectors are spliced ​​together as the word w in the source tree i and the word w j The representation vector of the structural information of the cyclic neural network, the top-level output of the recurrent neural network is spliced ​​with the representation vector respectively, and used as the input of the perceptron MLP; the word w is obtained by biaffine calculation i and the word w j The target end depends on the arc-score value; the invention makes full use of the source-end syntax tree to describe the corresponding laws of the two labeling specifications, and finally completes the high-quality tree bank conversion.

The invention relates to an automatic tree bank conversion method and system based on a tree-shaped recurrent neural network, which is designed to obtain an accurate supervised conversion model. The automatic tree bank conversion method based on the tree-shaped recurrent neural network of the present invention includes: based on the bidirectional tree-shaped recurrent neural network TreeLSTM, the word w is obtained i , word w i , word w a The hidden layer output vector of the hidden layer output vector is spliced ​​together as the word w i and the word w j In the source tree, the representation vector output vector of the top layer of the recurrent neural network BiSeqLSTM is concatenated with the representation vector respectively, and used as the input of the perceptron MLP, the perceptron extracts syntactic information; the word w is calculated using biaffine i and the word w j The target side of the depends on the arc minutes value. The invention makes full use of the source-end syntax tree to describe the corresponding rules of the two labeling norms and provides necessary data support for establishing a high-quality tree supervised transformation model.

The invention provides an event relationship graph generation method and apparatus. The method comprises the steps of splitting a manuscript into statements according to preset punctuation marks; extracting characters in the split statements and screening out the statements containing the characters as standby statements, wherein the characters include at least one of a personal name, a role and a personal pronoun; performing syntactic analysis on the standby statements by utilizing a pre-obtained syntactic analysis tree bank to generate an associative relationship among the characters; and generating an event relationship graph by utilizing the characters and the associative relationship. By utilizing the method, events occurring among the characters can be visually viewed, so that a user can understand substances of the events for a relatively short time and the reading time is shortened.