Search code examples
bert-language-modeltopic-modelingperplexity

How to calculate perplexity of BERTopic?

Is there a way to calculate the perplexity of BERTopic? I am unable to find any such thing in the BERTopic library and in other places.

Solution

  • I managed to figure it out how to get the log perplexity, and then convert it back

    import numpy as np
model = BERTopic(top_n_words =15,
                   calculate_probabilities=True)

topics, probs = model.fit_transform(docs) # docs = dataset
log_perplexity = -1 * np.mean(np.log(np.sum(probs, axis=1)))
perplexity = np.exp(log_perplexity)