Encoding and adaptive, scalable accessing of distributed models
AssigneeGoogle Inc.
InventorsFranz Josef Och, Jeffrey Dean, Thorsten Brants, Alexander Mark Franz, Jay Ponte, Peng Xu, Sha-Mayn Teh, Jeffrey Chin, Ignacio E. Thayer, Anton Carver, Daniel Rosart, John S. Hawkins, Karel Driesen
Life of this patent
Estimating the term…
Abstract
Systems, methods, and apparatus for accessing distributed models in automated machine processing, including using large language models in machine translation, speech recognition and other applications.
Claims21
A method comprising:
- receiving a first segment of text in a source language;
- obtaining a set of candidate translations of the first segment of text in a target language, each candidate translation in the set of candidate translations comprising one or more tokens in the target language;
- for each candidate translation in the set of candidate translations:
- obtaining, from a translation cache, first language model data about the tokens in the candidate translation, and
- computing an intermediate translation score using the first language model data;
- removing, based on the intermediate translation scores, one or more candidate translations from the set of candidate translations to obtain a reduced set of candidate translations;
- in response to obtaining the reduced set of candidate translations, for each of one or more candidate translations in the reduced set of candidate translations:
- obtaining, from one or more language model servers, second language model data about the tokens in the candidate translation, and
- computing a final translation score using the intermediate translation score and the second language model data;
- and
- selecting the candidate translation with a best final translation score.
A system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
- receiving a first segment of text in a source language;
- obtaining a set of candidate translations of the first segment of text in a target language, each candidate translation in the set of candidate translations comprising one or more tokens in the target language;
- for each candidate translation in the set of candidate translations:
- obtaining, from a translation cache, first language model data about the tokens in the candidate translation, and
- computing an intermediate translation score using the first language model data;
- removing, based on the intermediate translation scores, one or more candidate translations from the set of candidate translations to obtain a reduced set of candidate translations;
- in response to obtaining the reduced set of candidate translations, for each of one or more candidate translations in the reduced set of candidate translations:
- obtaining, from one or more language model servers, second language model data about the tokens in the candidate translation, and
- computing a final translation score using the intermediate translation score and the second language model data;
- and
- selecting the candidate translation with a best final translation score.
One or more non-transitory computer storage media storing instructions that are operable, when executed by one or more computers, to cause the one or more computers to perform operations comprising:
- receiving a first segment of text in a source language;
- obtaining a set of candidate translations of the first segment of text in a target language, each candidate translation in the set of candidate translations comprising one or more tokens in the target language;
- for each candidate translation in the set of candidate translations:
- obtaining, from a translation cache, first language model data about the tokens in the candidate translation, and
- computing an intermediate translation score using the first language model data;
- removing, based on the intermediate translation scores, one or more candidate translations from the set of candidate translations to obtain a reduced set of candidate translations;
- in response to obtaining the reduced set of candidate translations, for each of one or more candidate translations in the reduced set of candidate translations:
- obtaining, from one or more language model servers, second language model data about the tokens in the candidate translation, and
- computing a final translation score using the intermediate translation score and the second language model data;
- and
- selecting the candidate translation with a best final translation score.
Dependencies are inferred from the claim text.