MediumNLPnlp-289

Length Normalization Exponent Effect

Problem

Compare two beam hypotheses after length normalization with exponent alpha. Each hypothesis is a (score, length) pair, and a non-positive length is used as-is without normalization.

Normalized score is score / (length ** alpha) when length is positive. Return 'a' or 'b' for the higher normalized score, or 'tie'.

Hint

For negative log scores, dividing by a longer length makes the score less negative.

Test cases

case 1

in: [-2,10,-1.5,1,1]

expected: "a"

case 2

in: [-2,10,-1.5,1,0]

expected: "b"

case 3

in: [-1,2,-2,4,1]

expected: "tie"

case 4

in: [-0.5,2,-2,4,1]

expected: "a"

case 5

in: [-1,0,-2,4,1]

expected: "b"

Starter codePython

def length_norm_exponent_effect(score_a, length_a, score_b, length_b, alpha):
    # Your code here
    pass

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