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