Topic Coherence Computation: Division by 45?

[mona-timmermann]

Why are they dividing by 45 for topic coherence based on normalised PMI? It says in the paper but the computation in the code looks different to me.

[jfcann]

Hi mona-timmermann, the reason for the 45 is that there are 45 ways of picking 2 distinct words from a list of 10 words. Equivalently, there are 45 (i, j) summation indices used in the TC equation above. You divide by 45 so that you have the average PMI.

[yuyangstatistics]

If we run the 'eval' mode, then the log file will show counter = 55. I think this is due to a tiny error in the get_topic_coherence() function: top_10 = list(beta[k].argsort()[-11:][::-1]). It should instead be top_10 = list(beta[k].argsort()[-10:][::-1]). After we change it, the counter will equal to 45.