Patent · US Active

System and method for training a neural machine translation model

US11875141B2 · kind B2 · utility

0Cited by

10References

12Claims

0Family size

Assignee

Infosys Limited · IN

Inventors

Kamalkumar Rathinasamy · Ayanambakkam, IN
Amanpreet Singh · Bremen, DE
Balaguru Sivasambagupta · Bengaluru, IN
Prajna Prasad Neerchal · Kasaragod, IN
Vani Sivasankaran · Ayanambakkam, IN

Key dates

Filing date	Nov 22, 2021
Grant date	Jan 16, 2024
Priority date	—
Expiry date	Mar 23, 2042

Classification

Technology area (CPC G)Physics
CPC primaryG06F40/40
WIPO fieldComputer technology
WIPO sectorElectrical engineering

Abstract

The system and method for training a neural machine translation (NMT) model is disclosed wherein training data in terms of source statements and equivalent targets statements may be received. The source statements and equivalent targets statements may be encoded using source and target vocabulary respectively. A source-target map containing relation between tokens is created. The source statements and equivalent target statements is split into multiple fragments using fragments generator based on the source-target map. Such generated multiple fragments are used to train NMT model. Whenever the trained NMT model receives a source codes as input, the source codes are transformed to target codes.

Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.