Ideal Windows Size in Pk Evaluation Metric

Question

I am very new to nlp. I am doing a text segmentation task and for evaluating my model I need to calculate Pk and Windiff scores. My question is what is the ideal value for window size (k) for Pk score because different window sizes give different results. I am using this function nltk.metrics.segmentation.pk. Thanks.

Answer 1

I am also new to NLP, but according to some research papers:

Window size k should follow the next expresion:

This applies for WinDiff and Pk scores.

Reference: https://books.google.es/books?id=lus_EAAAQBAJ