OpenNMT

Predict_score with target_prefix

Hello,

I was wondering if there is any option to get the predict_score of the portion translated by ctranslate2 excluding the known part (when we use the option target_prefix). As I’m prety sure it’s included in the calculation by default.

If there is none, is it something you believe could be achieved by a amateur programmer?

Best regards,
Samuel

Hi,

The returned score does not include the score of the target prefix. More precisely, we assign a score of 0 to each token coming from the target prefix.

2 Likes