[raams95]

Summary

BleuSampleEvaluator returned bleu_score[0] from Bleu.compute_score(). That list holds the cumulative scores [BLEU-1, ..., BLEU-n], so BleuScore was always unigram precision and n did nothing. With the default n=4, "mat the on sat cat the" scored 1.0 against "the cat sat on the mat".

Two smaller problems in the same path:

• eff="longest" is documented and accepted by the constructor, but Bleu._single_reflen() had no branch for it, so it raised AssertionError: unsupported reflen option longest.
• Bleu.cook_append() built the n-gram counts with the default n=4 instead of self.n, so n > 4 raised IndexError.

Fix

• bleu_sample_evaluator.py: return bleu_score[self.n - 1].
• bleu/bleu.py: pass n=self.n to cook_refs / cook_test, and add a longest branch (max(reflens)) to _single_reflen.

Spot check against a hand calculation: "the cat sat on a mat" vs "the cat sat on the mat" now gives 0.5373 for n=4, which is (5/6 · 3/5 · 2/4 · 1/3)^(1/4).

Tests

I added test/cpu_only/test_bleu_sample_evaluator.py:

• default run() gives 1.0 for an identical sentence and ~0 for a shuffled one

• n=1 still ignores word order (unigram precision)

• identical text scores 1.0 for n = 1, 2, 4 and 6

• every documented eff option (shortest, average, longest) works

• Before the fix: 3 failed, 6 passed (shuffled sentence scored 1.0, n=6 IndexError, eff="longest" AssertionError)

• After the fix: 9 passed

pytest -m "cpu or api" test/cpu_only/test_bleu_sample_evaluator.py -q
9 passed

Fixes #566