Shortcuts

Translation Edit Rate (TER)

Module Interface

class torchmetrics.TranslationEditRate(normalize=False, no_punctuation=False, lowercase=True, asian_support=False, return_sentence_level_score=False, **kwargs)[source]

Calculate Translation edit rate (TER) of machine translated text with one or more references.

This implementation follows the one from SacreBleu_ter, which is a near-exact reimplementation of the Tercom algorithm, produces identical results on all “sane” outputs.

As input to forward and update the metric accepts the following input:

  • preds (Sequence): An iterable of hypothesis corpus

  • target (Sequence): An iterable of iterables of reference corpus

As output of forward and compute the metric returns the following output:

  • ter (Tensor): if return_sentence_level_score=True return a corpus-level translation edit rate with a list of sentence-level translation_edit_rate, else return a corpus-level translation edit rate

Parameters
  • normalize (bool) – An indication whether a general tokenization to be applied.

  • no_punctuation (bool) – An indication whteher a punctuation to be removed from the sentences.

  • lowercase (bool) – An indication whether to enable case-insesitivity.

  • asian_support (bool) – An indication whether asian characters to be processed.

  • return_sentence_level_score (bool) – An indication whether a sentence-level TER to be returned.

  • kwargs (Any) – Additional keyword arguments, see Advanced metric settings for more info.

Example

>>> preds = ['the cat is on the mat']
>>> target = [['there is a cat on the mat', 'a cat is on the mat']]
>>> ter = TranslationEditRate()
>>> ter(preds, target)
tensor(0.1538)

Initializes internal Module state, shared by both nn.Module and ScriptModule.

Functional Interface

torchmetrics.functional.translation_edit_rate(preds, target, normalize=False, no_punctuation=False, lowercase=True, asian_support=False, return_sentence_level_score=False)[source]

Calculate Translation edit rate (TER) of machine translated text with one or more references. This implementation follows the implmenetaions from https://github.com/mjpost/sacrebleu/blob/master/sacrebleu/metrics/ter.py. The sacrebleu implmenetation is a near-exact reimplementation of the Tercom algorithm, produces identical results on all “sane” outputs.

Parameters
  • preds (Union[str, Sequence[str]]) – An iterable of hypothesis corpus.

  • target (Sequence[Union[str, Sequence[str]]]) – An iterable of iterables of reference corpus.

  • normalize (bool) – An indication whether a general tokenization to be applied.

  • no_punctuation (bool) – An indication whteher a punctuation to be removed from the sentences.

  • lowercase (bool) – An indication whether to enable case-insesitivity.

  • asian_support (bool) – An indication whether asian characters to be processed.

  • return_sentence_level_score (bool) – An indication whether a sentence-level TER to be returned.

Return type

Union[Tensor, Tuple[Tensor, List[Tensor]]]

Returns

A corpus-level translation edit rate (TER). (Optionally) A list of sentence-level translation_edit_rate (TER) if return_sentence_level_score=True.

Example

>>> preds = ['the cat is on the mat']
>>> target = [['there is a cat on the mat', 'a cat is on the mat']]
>>> translation_edit_rate(preds, target)
tensor(0.1538)

References

[1] A Study of Translation Edit Rate with Targeted Human Annotation by Mathew Snover, Bonnie Dorr, Richard Schwartz, Linnea Micciulla and John Makhoul TER