Suggested value for length_penalty

Research
1

Based on this discussion I’m considering setting the length_penalty value that Argos Translate uses with CTranslate to 0.2 and trying to decide if this makes sense. “Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation” seems to recommend 0.2, and based on this forum post I wasn’t going to try to use coverage penalty like the paper recommended.

CTranslate currently defaults to a value of 0, is there a reason for this / are there potential drawbacks to increasing the length penalty?

Thanks!

2

The paper mentions:

We found that the optimal α and β vary slightly for different models.

So we can’t set a default value that works for all models and you should tune these parameters on your development set.

1 Like