Module pysimt.metrics
Expand source code
from .metric import Metric
from .multibleu import BLEUScorer
from .sacrebleu import SACREBLEUScorer
from .meteor import METEORScorer
from .cer import CERScorer
from .wer import WERScorer
from .simnmt import AVPScorer, AVLScorer, CWMScorer
"""These metrics can be used in early stopping."""
# Generation related metrics
beam_metrics = ["BLEU", "SACREBLEU", "METEOR", "CER", "WER"]
metric_info = {
'BLEU': 'max',
'SACREBLEU': 'max',
'METEOR': 'max',
'LOSS': 'min',
'ACC': 'max',
'RECALL': 'max',
'PRECISION': 'max',
'F1': 'max',
'CER': 'min',
'WER': 'max',
# simultaneous translation
'AVP': 'min', # Average proportion (Cho and Esipova, 2016)
'AVL': 'min', # Average Lagging (Ma et al., 2019 (STACL))
'DAL': 'min', # Differentiable AL (not implemented)
'CW': 'min', # Consecutive wait (Gu et al., 2017) [Not Implemented]
}
Sub-modules
pysimt.metrics.cer
-
Character error rate (CER).
pysimt.metrics.meteor
-
METEOR metric.
pysimt.metrics.metric
-
Base Metric class to be derived from.
pysimt.metrics.multibleu
-
Tokenized BLEU through sacreBLEU API.
pysimt.metrics.sacrebleu
-
Detokenized BLEU i.e. sacreBLEU.
pysimt.metrics.simnmt
-
Simultaneous MT latency metrics.
pysimt.metrics.wer
-
Word error rate (WER).