HardNLPnlp-029

Backoff Probability

Problem

Compute a Katz-style backoff probability for a trigram. context is [w1, w2] and token is w3.

If the key 'w1 w2 w3' is in tri_probs with a positive value, return it. Otherwise, if 'w2 w3' is in bi_probs with a positive value, return alpha times that value. Otherwise, if w3 is in uni_probs with a positive value, return alpha squared times that value. Return 0.0 when nothing matches.

Hint

Only back off when the higher-order probability is absent or not positive.

Test cases

case 1

in: [["a","b"],"c",{"a b c":0.2},{},{},0.4]

expected: 0.2

case 2

in: [["a","b"],"c",{},{"b c":0.5},{},0.4]

expected: 0.2

case 3

in: [["a","b"],"c",{},{},{"c":0.1},0.5]

expected: 0.025

case 4

in: [["a","b"],"c",{},{},{},0.5]

expected: 0

case 5

in: [["a","b"],"c",{"a b c":0},{"b c":0.5},{"c":0.9},0.5]

expected: 0.25

Starter codePython

def backoff_prob(context, token, tri_probs, bi_probs, uni_probs, alpha):
    # Your code here
    pass

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