VEST

From cdec Decoder

VEST stands for Viterbi Envelope Semiring Training, and it's what cdec calls its implementation of the minimum error training algorithm (MERT; Och 2002). Specifically, it is an implementation of MERT where the line-search inference is carried out using the Inside algorithm using a particular semiring.

Supported loss functions

• BLEU (three definitions of the brevity penalty: the original, average, and closest)
• TER
• (BLEU - TER) / 2
• AER (Alignment error rate, for supervised training of alignment models)