Characterizing the performance of minimum Bayes risk (MBR) decoding for neural machine translation (NMT) when using various candidate generation methods by examining BLEU score, length bias, and token frequency bias. [CS287: Deep Learning for Natural Language Processing, Final Project, Fall 2021]
transformers
beam-search
neural-machine-translation
nmt
huggingface-transformers
minimum-bayes-risk-decoding
mbr-decoding
length-bias
token-frequency-bias
-
Updated
Dec 11, 2021 - Jupyter Notebook