"""TER (Translation Edit Rate) metric implementation for DashAI."""
import evaluate
import numpy as np
from DashAI.back.dataloaders.classes.dashai_dataset import DashAIDataset
from DashAI.back.metrics.translation_metric import TranslationMetric, prepare_to_metric
[docs]
class Ter(TranslationMetric):
"""A class for calculating TER scores between source and target sentences.
TER (Translation Edit Rate, also called Translation Error Rate) is a metric
to quantify the edit operations requiered to match a reference translation.
References
----------
[1] https://huggingface.co/spaces/evaluate-metric/ter
"""
@staticmethod
def score(source_sentences: DashAIDataset, target_sentences: np.ndarray):
"""Calculate the TER score between source and target sentences.
Parameters
----------
source_sentences : DashAIDataset
Sentences in the original language.
target_sentences : ndarray
Sentences in the target language.
Returns
-------
float
The calculated score.
"""
metric = evaluate.load("ter")
source_sentences, target_sentences = prepare_to_metric(
source_sentences, target_sentences
)
return metric.compute(
references=source_sentences, predictions=target_sentences
)["score"]