Error Analysis in Online/Offline NMT


IWSLT De->EN

Offline systems

Samples from the test set formatted as:
 id)
 src: Source sequence.
 ref: Reference target sequence.
 Difficulty: AL score + the alignement between tokenized src-ref
 PA (beam): Hypothesis with Pervasive Attention using beam search.
 TF (beam): Hypothesis with Transformer using beam search.
 PA (greedy): Hypothesis with Pervasive Attention greeedily decoded.
 TF (greedy): Hypothesis with Transformer greedily decoded.
Removing BPE || Keeping the BPE tokenization

Online systems

Samples from the test set formatted as:
 id)
 src: Source sequence.
 ref: Reference target sequence.
 Difficulty: AL score.
 PA (offline): Hypothesis with Pervasive Attention greeedily decoded.
 TF (offline): Hypothesis with Transformer greedily decoded.
 PA (online): Hypothesis with Pervasive Attention trained above k=1. Decoding along wait-3.
 TF (online): Hypothesis with Transformer trained on k=7. Decoding along wait-3.
Removing BPE || Keeping the BPE tokenization

WMT De->EN

Removing BPE || Keeping the BPE tokenization