US 12,743,589 B1Granted

Retrieval and edit-based machine translation

AssigneeAmazon Technologies, Inc.

InventorsZhou Yu, Yi Liu, Simi Wang, Junnosuke Kuroda, Girish Subramanian, Yonggang Deng

Life of this patent

Estimating the term…

Loading patent…

Abstract

Devices and techniques are generally described for machine translation. In various examples, first text in a first spoken language may be received. A first candidate translation of the first text in a second spoken language may be determined based at least in part on the first text from among a candidate pool of non-translated text associated with a first domain. Combined data representing the first text and the first candidate translation may be generated. A first machine learning model may generate second text by modifying at least one word of the first candidate translation. The second text representing a machine translation of the first text may be output.

Claims20

Dependent claims

A computer-implemented method of language translation, comprising:

determining first text comprising at least one word composed in a first spoken language;
generating first embedding data representing the first text;
selecting second embedding data stored in non-transitory computer-readable memory based on a similarity between the first embedding data and the second embedding data, the second embedding data representing second text comprising at least one word composed in a second spoken language, wherein the second text is a candidate translation of the first text in the second spoken language;
generating first combined data by combining the first text and the second text;
inputting the first combined data into a transformer model trained to edit the candidate translation in the second text using the first text;
generating, by the transformer model using the first combined data, third text by altering at least one word of the second text, wherein the third text is a corrected translation of the first text in the second spoken language;
and
outputting the third text.

5 more elements

Dependencies are inferred from the claim text.