New Experiment: Lab-Made Lemko?

I will be conducting an experiment this month to see if machines can be made to translate into Lemko better than Google Translate or humans.

Hypothesis

A machine can be configured to translate from English into the endangered Slavic language of Lemko and achieve quality scores higher than those of Google Translate’s Ukrainian service, but not yet higher than those of humans.

Predictions

• My English to Lemko rule-based machine translation (RBMT) engine will achieve a bilingual evaluation understudy (BLEU) score of 15 against a clean bilingual corpus.
• The above engine will achieve a BLEU score that is a third higher (e.g. 20) when coupled with an improvised dictionary-based machine translation (DBMT) created from Lemko-Polish unit-test assertion pairs.
• Google Translate’s English to Ukrainian translation service will achieve a BLEU score of 10 against the above corpus.
• I, a human, will achieve a higher BLEU score than all the above machines against the above corpus.

The experiments will be conducted over the next week or so, for subsequent publication.