Selectively tuning machine translation models for custom machine translations

A method, comprising:

• receiving, at a machine translation system, a request to translate an input text in a first language to a second language;
• identifying, by the machine translation system, a tuning data set for translating the input text, wherein the tuning data set comprises a plurality pairs of texts, wherein first texts in the plurality of pairs of texts are stated in the first language and wherein second texts in the plurality of pairs of texts are stated in the second language;
• tuning, by the machine translation system, a machine learning model that translates to the second language using only one or more second texts from one or more of the plurality of pairs of texts selected according to a comparison of the input text with first texts of the one or more pairs of texts in the tuning data set;
• and
• translating, by the machine translation system, the input text to the second language using the tuned machine learning model that translates to the second language.

One or more non-transitory, computer-readable storage media, storing program instructions that when executed on or across one or more computing devices cause the one or more computing devices to implement:

• receiving, at a machine translation system, a request to translate an input text in a first language to a second language;
• identifying, by the machine translation system, a tuning data set for translating the input text, wherein the tuning data set comprises a plurality pairs of texts, wherein first texts in the plurality of pairs of texts are stated in the first language and wherein second texts in the plurality of pairs of texts are stated in the second language;
• tuning, by the machine translation system, a machine learning model that translates from a pivot language to the second language using only one or more second texts from one or more of the plurality of pairs of texts selected according to a comparison of the input text with first texts of the one or more pairs of texts in the tuning data set;
• and
• applying, by the machine translation system, the tuned machine learning model to a version of the input text translated into the pivot language to generate a translated version of the input text in the second language.