Table 6:
Pairwise-BLEU (pBLEU) (Shen et al., 2019) for candidate translations generated from different number of experts. BLEU from the doc-reranker taking different sets of candidate translations. We obtain different experts by training the document transformer (Zhang et al., 2018) with backtranslation with different random initialization. The size of the candidate pool is 50. The experts for the human proposal baseline are the reference translations.
Proposal#ExpertspBLEUBLEU
human 21.40 –

Doc-transformer 70.41 53.63
59.09 54.70
53.54 55.21
Proposal#ExpertspBLEUBLEU
human 21.40 –

Doc-transformer 70.41 53.63
59.09 54.70
53.54 55.21
Close Modal