Improve translation of short sentences

[eu9ene]

We have issues with translating shorter sentences and single words.

What might help:

• Better data cleaning to avoid removing short translations from the training datasets
• Try to find more datasets that contain short sentences

[gregtatum]

I wonder if we can load in dictionaries where it's literally one word to one word.

[gregtatum]

Or maybe even synthesize it with the alignment data.

[gregtatum]

This behavior is also visible with numbers. A good example is to do a list of numbers.

[gregtatum]

Verify the fix with:

https://bugzilla.mozilla.org/show_bug.cgi?id=1888972

[gregtatum]

Here is a word count distribution for the merged corpus sl-en: https://firefox-ci-tc.services.mozilla.com/tasks/groups/PPCzZRHaTT6Ys4BIhPGT5w

Generated via:

python3 pipeline/data/analyze.py --file_location https://firefox-ci-tc.services.mozilla.com/api/queue/v1/task/VK5zmxJRTLy0y0WBQ0DRJg/artifacts/public/build/corpus.en.zst --output data --dataset OPUS_corpus --language en

python3 pipeline/data/analyze.py --file_location https://firefox-ci-tc.services.mozilla.com/api/queue/v1/task/VK5zmxJRTLy0y0WBQ0DRJg/artifacts/public/build/corpus.sl.zst --output data --dataset OPUS_corpus --language sl

[marco-c]

So we have basically 0 sentences with 1 word?

[gregtatum]

I filed #878 which suggests augmenting with statistically synthesized single word translations.

[gregtatum]

I filed #879 which suggests harvesting short sentences from parallel datasets.

[gregtatum]

I filed #880 which suggests statistically synthesizing short sentence translations from monolingual data sources.