Bleu

evaluatio.metrics.bleu¶

BLEU metrics

Evaluatio does not implement BLEU natively, but instead relies on sacrebleu³. Evaluation complements sacrebleu by providing statistical comparison tools, which are not included in sacrebleu itself. This module contains those functions.

This module provides a paired bootstrap significance test for comparing two machine translation systems using the BLEU metric. It follows the method introduced by Koehn (2004), in which corpus-level BLEU is recomputed on each bootstrap resample rather than aggregating per-sentence scores directly.

Sufficient statistics (clipped n-gram counts, total n-gram counts, and lengths required for the brevity penalty) are precomputed per sentence using sacrebleu, ensuring that tokenisation and scoring are fully compatible with the sacrebleu reference implementation. The resampling itself is performed in Rust for efficiency.

The confidence interval function works the same way as the paired bootstrap and the required statistics are precomputed using sacrebleu.

Note

BLEU is a corpus-level metric and does not decompose meaningfully at the sentence level. The sufficient statistics approach used here preserves corpus-level correctness while avoiding redundant tokenisation on each resample.

Tokenisation is handled by sacrebleu using the 13a tokeniser by default, consistent with WMT evaluation practice. Scores produced by this module are directly comparable to sacrebleu corpus-level BLEU scores computed with the same settings.

For sentence-decomposable metrics such as chrF or COMET, use :func:evaluatio.hypothesis.paired_bootstrap_test directly with per-sentence scores.

References

1. Papineni, K., et al. (2002). BLEU: a method for automatic evaluation of machine translation. ACL.

2. Koehn, P. (2004). Statistical significance tests for machine translation evaluation. Proceedings of EMNLP 2004, 388-395.

3. Post, M. (2018). A call for clarity in reporting BLEU scores. Proceedings of the Third Conference on Machine Translation, 186-191.

Functions¶

bleu_bootstrap_test¶

bleu_bootstrap_test(references: Iterable[Iterable[str]], hyp1: Iterable[str], hyp2: Iterable[str], iterations: int, effective_order: bool=True) -> float

Perform a paired bootstrap significance test comparing two MT systems using corpus-level BLEU.

Sufficient statistics are precomputed per sentence using sacrebleu, then passed to the Rust resampling backend. On each bootstrap iteration a pseudo-test-set is drawn by sampling sentences with replacement, and corpus-level BLEU is recomputed for both systems from the accumulated sufficient statistics. The p-value is the proportion of iterations in which the worse system appears to outperform the better, using the (count + 1) / (iterations + 1) correction.

Parameters

• references : iterable of iterable of str
  Reference translations. The outer iterable has one entry per sentence, and the inner iterable contains one or more reference strings for that sentence. May be a generator; it is fully materialised before processing.

• hyp1 : iterable of str
  Hypothesis strings for the first system, one per sentence. Must be the same length as references.

• hyp2 : iterable of str
  Hypothesis strings for the second system, one per sentence. Must be the same length as references.

• iterations : int
  Number of bootstrap resamples. Values of 5000 to 10000 give stable p-value estimates for most purposes.

• effective_order : bool
  If True (default), scales the n-gram order to the maximum order for which counts are non-zero. Recommended for sentence-level sufficient statistic computation to avoid zero precision on short segments. Set to False to match standard corpus-level BLEU behaviour exactly.

Returns

• float
  Two-sided p-value in the range (0, 1]. A value below 0.05 indicates that the observed BLEU difference is unlikely under the null hypothesis that the two systems are equivalent.

Raises

• ValueError
  If references, hyp1, and hyp2 are not all the same length, or if any is empty.

Note

The minimum possible p-value is 1 / (iterations + 1). With iterations=9999 this is 0.0001.

Tokenisation uses sacrebleu’s 13a tokeniser by default, consistent with WMT evaluation practice. BLEU scores computed internally are directly comparable to sacrebleu corpus-level scores produced with the same tokeniser and effective_order setting.

Examples

>>> references = [["the cat sat on the mat"], ["the dog ate the bone"]]
>>> hyp1 = ["the cat sat on the mat", "the dog ate the bone"]
>>> hyp2 = ["a cat sat on a mat", "a dog ate a bone"]
>>> p = bleu_bootstrap_test(references, hyp1, hyp2, iterations=9999)
>>> print(f"p = {p:.4f}")
p = 0.0231

bleu_ci¶

bleu_ci(references: Iterable[Iterable[str]], hypotheses: Iterable[str], iterations: int, alpha: float, effective_order: bool=True) -> ConfidenceInterval

Estimate a confidence interval for corpus-level BLEU using bootstrap resampling.

This function computes a percentile bootstrap confidence interval for the BLEU score by repeatedly resampling sentence-level sufficient statistics with replacement and recomputing corpus-level BLEU for each resample.

The returned interval reflects uncertainty due to sampling variation in the evaluation dataset.

Parameters

• references : Iterable[Iterable[str]]
  Reference translations. Each element corresponds to a sample and should be an iterable of reference strings (to support multiple references per hypothesis).

• hypotheses : Iterable[str]
  Model predictions (hypotheses). Must be aligned with references such that each hypothesis corresponds to the same-indexed reference set.

• iterations : int
  Number of bootstrap resampling iterations. Larger values yield more stable estimates but increase computation time.

• alpha : float
  Significance level for the confidence interval. For example, alpha=0.05 corresponds to a 95% confidence interval.

• effective_order : bool
  Whether to enable effective n-gram order when computing BLEU. This is passed directly to sacrebleu.BLEU and is recommended for shorter sequences. Default is True.

Returns

• ConfidenceInterval
  Estimated confidence interval for the corpus level word error rate.

Note

• BLEU is not well-defined at the sentence level; this method operates on corpus-level BLEU computed from resampled sentence-level sufficient statistics.

• The interval is computed using the percentile bootstrap method.

• The resampling procedure preserves alignment between hypotheses and their corresponding references.

• Small differences in BLEU may not be practically meaningful, even if confidence intervals do not overlap.

References

1. Papineni, K., et al. (2002). BLEU: a method for automatic evaluation of machine translation. ACL.