BLEU Score Lite

Problem

Compute a BLEU-lite score from clipped unigram and bigram precisions, using only unigrams when the candidate has a single token. The score is the brevity penalty times the geometric mean of the precisions, where BP = 1 when the candidate is at least as long as the reference, else exp(1 - ref_len / cand_len). Return 0.0 if the candidate is empty or any precision is zero.

Hint

Clip each candidate n-gram count at its reference count before dividing.

Test cases

case 1

in: ["the cat sat on the mat","the cat sat on the mat"]

expected: 1

case 2

in: ["the cat sat","the cat"]

expected: 0.5773502691896257

case 3

in: ["a b c d","a b c"]

expected: 0.7071067811865475

case 4

in: ["x y","a b"]

expected: 0

Starter codePython

import math
def bleu_lite(candidate, reference):
    # Your code here
    pass

4 test cases · run your code against them in the editor