System and method for machine translation of text

Patent metadata

Field	Value
Publication number	US-2020394431-A1
Application number	US-201916542392-A
Country	US
Kind code	A1
Filing date	Aug 16, 2019
Priority date	Jun 13, 2019
Publication date	Dec 17, 2020
Grant date	—

Abstract

A method and system for machine translation of text is disclosed. The method includes processing an image comprising a text to generate a pattern associated with the text based on a trained Convolution Neural Network (CNN). The method further includes mapping the pattern to a word in a mapping table and at least one text attribute, based on a classifier network. The method further includes initiating an Optical Character Recognition (OCR) conversion for the pattern, when at least one of the mapping between at least one of the pattern and at least one word in the mapping table and the mapping between the pattern and the at least one text attribute is below a predefined threshold. The method further includes performing incremental learning for the trained CNN and the classifier network based on the OCR conversion.

First claim

What is claimed is:

1 . A method of machine translation of text, the method comprising:
processing, by a translation device, an image comprising a text to generate a pattern associated with the text based on a trained Convolution Neural Network (CNN), wherein the text is in a source language; mapping, by the translation device, the pattern to a word in a mapping table and at least one text attribute, based on a classifier network, wherein the word is in a target language; initiating, by the translation device, an Optical Character Recognition (OCR) conversion for the pattern, when at least one of the mapping between at least one of the pattern and at least one word in the mapping table and the mapping between the pattern and the at least one text attribute is below a predefined threshold; and performing, by the translation device, incremental learning for the trained CNN and the classifier network based on the OCR conversion.

2 . The method of claim 1 , wherein the at least one text attribute comprises at least one of font, style, or color, and wherein the style is associated with a text being in bold, italics, strikethrough, or underline.

3 . The method of claim 1 , further comprising identifying the text from the image, wherein the image comprises a plurality of words.

4 . The method of claim 3 , wherein the identifying comprises:
identifying column and paragraph separation within the image based on difference in pixel intensity; and for each paragraph in each column, identifying each word based on difference in pixel intensity of an area designated by space between two adjacent words.

5 . The method of claim 4 further comprising extracting words sequentially from each paragraph of each column in response to the identifying of each word.

6 . The method of claim 5 further comprising creating sentences in response to sequential extraction of words from each paragraph of each column, using a Long Short Term Memory (LSTM).

7 . The method of claim 6 further comprising rendering the sentences created using the LSTM to a user.

8 . The method of claim 7 further comprising:
receiving feedback from the user on the sentences rendered; and
performing the incremental learning based on the feedback.

9 . The method of claim 7 further comprising generating a notification for the user, wherein the notification comprises at least one of warning of incorrect decoding and inability to support an attribute of the at least one attribute.

10 . A translation device for machine translation of text, the translation device comprising:
a processor; and a memory communicatively coupled to the processor, wherein the memory stores processor instructions, which, on execution, causes the processor to:
process an image comprising a text to generate a pattern associated with the text based on a trained Convolution Neural Network (CNN), wherein the text is in a source language;
map the pattern to a word in a mapping table and at least one text attribute, based on a classifier network, wherein the word is in a target language;
initiate an Optical Character Recognition (OCR) conversion for the pattern, when at least one of the mapping between at least one of the pattern and at least one word in the mapping table and the mapping between the pattern and the at least one text attribute is below a predefined threshold; and
perform incremental learning for the trained CNN and the classifier network based on the OCR conversion.

11 . The translation device of claim 10 , wherein the at least one text attribute comprises at least one of font, style, or color, and wherein the style is associated with a text being in bold, italics, strikethrough, or underline.

12 . The translation device of claim 11 , wherein the processor instructions further cause the processor to identify the text from the image, wherein the image comprises a plurality of words.

13 . The translation device of claim 12 , wherein to identify the processor instructions further cause the processor to:
identify column and paragraph separation within the image based on difference in pixel intensity; and for each paragraph in each column, identify each word based on difference in pixel intensity of an area designated by space between two adjacent words.

14 . The translation device of claim 13 , wherein the processor instructions further cause the processor to extract words sequentially from each paragraph of each column in response to the identifying of each word.

15 . The translation device of claim 14 , wherein the processor instructions further cause the processor to create sentences in response to sequential extraction of words from each paragraph of each column, using a Long Short Term Memory (LSTM).

16 . The translation device of claim 15 , wherein the processor instructions further cause the processor to render the sentences created using the LSTM to a user.

17 . The translation device of claim 16 , wherein the processor instructions further cause the processor to:
receive feedback from the user on the sentences rendered; and perform the incremental learning based on the feedback.

18 . The translation device of claim 17 , wherein the processor instructions further cause the processor to generate a notification for the user, wherein the notification comprises at least one of warning of incorrect decoding and inability to support an attribute of the at least one attribute.

19 . A non-transitory computer-readable storage medium having stored thereon, a set of computer-executable instructions causing a computer comprising one or more processors to perform steps comprising:
processing, by a translation device, an image comprising a text to generate a pattern associated with the text based on a trained Convolution Neural Network (CNN), wherein the text is in a source language; mapping, by the translation device, the pattern to a word in a mapping table and at least one text attribute, based on a classifier network, wherein the word is in a target language; initiating, by the translation device, an Optical Character Recognition (OCR) conversion for the pattern, when at least one of the mapping between at least one of the pattern and at least one word in the mapping table and the mapping between the pattern and the at least one text attribute is below a predefined threshold; and performing, by the translation device, incremental learning for the trained CNN and the classifier network based on the OCR conversion.

Assignees

• Wipro Ltd

Inventors

• Ramachandra Iyer Manjunath

Classifications

• G06F40/44 Primary

  Statistical methods, e.g. probability models · CPC title

• G06V30/1916

  Validation; Performance evaluation · CPC title

• G06V30/19173

  Classification techniques · CPC title

• G06V10/82

  using neural networks · CPC title

• G06N3/045

  Combinations of networks · CPC title