MediumNLPnlp-304

BLEU-2 with Clipping

Problem

Compute the clipped BLEU-2 score between candidate and reference token lists. Unigram and bigram precisions use multiplicity clipping against the reference, and the score is the geometric mean exp((log p1 + log p2) / 2).

Return 0.0 when either side lacks unigrams or bigrams, or when either precision is 0. Brevity is not penalized here.

Hint

Clip each distinct n-gram count by how often it appears in the reference.

Test cases

case 1

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

expected: 1

case 2

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

expected: 0.5773502691896257

case 3

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

expected: 0

case 4

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

expected: 1

case 5

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

expected: 0

Starter codePython

import math
def bleu2_clipped(candidate, reference):
    # Your code here
    pass

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