HardNLPnlp-172

Topic Coherence (Lite)

Problem

Compute a lite topic coherence. For each topic, average the PMI log((c12 * N) / (c1 * c2)) over all word pairs that have positive unigram and pair counts, treating pair counts as symmetric.

Return the average of the per-topic means, or 0.0 when no pair qualifies.

Hint

Look up the pair count in both orders before computing the pointwise mutual information.

Test cases

case 1

in: [[["cat","dog"]],{"cat":2,"dog":2},{"cat dog":1},4]

expected: 0

case 2

in: [[["cat","dog"]],{"cat":2,"dog":2},{"cat dog":2},4]

expected: 0.6931471805599453

case 3

in: [[["cat","dog"],["sun","moon"]],{"cat":2,"dog":2,"sun":1,"moon":1},{"cat dog":2,"sun moon":2},4]

expected: 1.3862943611198906

case 4

in: [[["cat","dog"]],{"cat":0,"dog":2},{"cat dog":2},4]

expected: 0

case 5

in: [[],{},{},4]

expected: 0

Starter codePython

import math
def topic_coherence_lite(topics, word_counts, pair_counts, n_docs):
    # Your code here
    pass

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