Electronic device and method for text processing

Patent metadata

Field	Value
Publication number	US-10860798-B2
Application number	US-201716080670-A
Country	US
Kind code	B2
Filing date	Mar 21, 2017
Priority date	Mar 22, 2016
Publication date	Dec 8, 2020
Grant date	Dec 8, 2020

Abstract

An electronic device and method for text processing, the electronic device comprises a processor (100), and the processor is configured to: determine a correlation between a first text vector and a second text vector, wherein the first text vector and the second text vector are multi-dimensional, real number vectors generated on the basis of a same text, respectively; obtain, according to the correlation, a third text vector representing the text, wherein a vector space in which the third text vector is located is correlated to vector spaces in which first text and second text vectors are located. The electronic device and method of the present invention can be used to create a text-feature representation model which represents text features by combining a plurality of view angles, thereby improving the performance of natural language processing.

First claim

The invention claimed is:

1. An electronic device for text processing, comprising:
a processor configured to:
determine, for each of a plurality of texts, correlation between a first text vector and a second text vector which are multi-dimensional real number vectors generated based on the same text respectively and are based on a first word feature representation model and a second word feature representation model respectively, wherein the first word feature representation model and the second word feature representation model are obtained based on different word feature representation training mechanisms;
obtain a third text vector based on the correlation for each of the plurality of texts, to represent the corresponding text, wherein a vector space where each third text vector is located is related to vector spaces where the corresponding first text vector and the second text vector are located,
determine, for each of a plurality of texts, correlation between a fourth text vector and a fifth text vector which are multi-dimensional real number vectors generated based on the same text respectively and are based on a fourth word feature representation model and a fifth word feature representation model respectively, wherein the fourth word feature representation model and the fifth word feature representation model are obtained based on different word feature representation training mechanisms; and
obtain a sixth text vector based on the correlation for each of the plurality of texts, to represent the corresponding text, wherein a vector space where each sixth text vector is located is related to vector spaces where the corresponding fourth text vector and the fifth text vector are located; and

a memory configured to store the third text vectors of the plurality of texts to establish a first multi-view text feature representation model and to store the sixth text vectors of the plurality of texts to establish a second multi-view text feature representation model,
wherein the processor is further configured to:
fuse the first multi-view text feature representation model and the second multi-view text feature representation model to form a third multi-view text feature representation model;
receive one or more words as a text object;
map each of the one or more words of the text object into a corresponding word vector based on the third multi-view text feature representation module;
perform at least one of a slot filling, a statement classification, and a translation based on the one or more word vectors mapped from the one or more words of the text object based on the third multi-view text feature representation module; and
present a result of the at least one of a slot filling, a statement classification, and a translation.

2. The electronic device according to claim 1 , wherein the text corresponds to a word.

3. The electronic device according to claim 1 , wherein the text corresponds to at least one of a phrase constituted by a plurality of words and a sentence constituted by a plurality of phrases.

4. The electronic device according to claim 1 , wherein the word feature representation training mechanism comprises at least one of a Word2Vec mechanism, a GloVe mechanism and a C&W mechanism.

5. The electronic device according to claim 1 , wherein the first word feature representation model and the second word feature representation model are obtained based on different training corpuses.

6. The electronic device according to claim 1 , wherein the processor is further configured to determine the correlation between the first text vector and the second text vector based on Canonical Correlation Analysis, and regulate parameters of the Canonical Correlation Analysis with an object of making the correlation satisfy a predetermined condition.

7. The electronic device according to claim 1 , wherein the processor is further configured to process the first text vector and the second text vector using a neural network to obtain a variable of the first text vector and a variable of the second text vector, determine the correlation based on the variable of the first text vector and the variable of the second text vector, and regulate parameters of the neural network with an object of making the correlation satisfy a predetermined condition.

8. The electronic device according to claim 7 , wherein the processor is further configured to process the variable of the first text vector and the variable of the second text vector using an auto-encoder to reconstruct the first text vector and the second text vector, and regulate parameters of the auto-encoder and the neural network with an object of further making an error between the reconstructed first text vector and the first text vector and an error between the reconstructed second text vector and the second text vector satisfy a predetermine condition, to determine the correlation.

9. The electronic device according to claim 1 , wherein the processor is further configured to, for each of the plurality of texts, determine correlation between a first text vector and a second text vector corresponding to the text further based on the correlation regarding other texts.

10. A method for text processing, comprising:
determining, for each of a plurality of texts, correlation between a first text vector and a second text vector which are multi-dimensional real number vectors generated based on the same text respectively and are based on a first word feature representation model and a second word feature representation model respectively, wherein the first word feature representation model and the second word feature representation model are obtained based on different word feature representation training mechanisms;
obtaining a third text vector based on the correlation for each of the plurality of texts, to represent the corresponding text, wherein a vector space where each third text vector is located is related to vector spaces where the corresponding first text vector and the second text vector are located;
determining, for each of a plurality of texts, correlation between a fourth text vector and a fifth text vector which are multi-dimensional real number vectors generated based on the same text respectively and are based on a fourth word feature representation model and a fifth word feature representation model respectively, wherein the fourth word feature representation model and the fifth word feature representation model are obtained based on different word feature representation training mechanisms;
obtaining a sixth text vector based on the correlation for each of the plurality of texts, to represent the corresponding text, wherein a vector space where each sixth text vector is located is related to vector spaces where the corresponding fourth text vector and the fifth text vector are located;
storing the third text vectors of the plurality of texts to establish a first multi-view text feature representation model and the sixth text vectors of the plurality of texts to establish a second multi-view text feature representation model;
fusing the first multi-view text feature representation model and the second multi-view text feature representation model to form a third multi-view text feature representation model;
receiving one or more words as a text object;
mapping each of the one or more words of the text object into a corresponding word vector based on the third multi-view text feature representation module;
performing at least one of a slot filling, a statement classification, and a translation based on the one or more word vectors mapped from the one or more words of the text object based on the third multi-view text feature representation module; and
presenti

Assignees

• Sony Corp

Inventors

• Wu Youzheng
• Qi Jun

Classifications

• G06N3/047

  Probabilistic or stochastic networks · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/0499

  Feedforward networks · CPC title