Great research and thanks for opening source code! We implement Viterbi decoding algorithms in this pull request, which can find the output that maximizes P(A,Y|X) / |Y|^{beta}. The speed of Viterbi decoding is slightly slower than Lookahead. The performance of Viterbi decoding is between Lookahead and beam search.
@inproceedings{shao2022viterbi,
author = {Chenze Shao and Zhengrui Ma and Yang Feng},
title = {Viterbi Decoding of Directed Acyclic Transformer for Non-Autoregressive Machine Translation},
booktitle = {Findings of EMNLP 2022},
year = {2022}
}
Great research and thanks for opening source code! We implement Viterbi decoding algorithms in this pull request, which can find the output that maximizes P(A,Y|X) / |Y|^{beta}. The speed of Viterbi decoding is slightly slower than Lookahead. The performance of Viterbi decoding is between Lookahead and beam search.