Compound score diverges for long sequences

[VincentGurgul]

The compound score has a serious flaw – it diverges for long sequences. Example:

polarity_scores('bad good')
>>> {'neg': 0.547, 'neu', 0.0, 'pos': 0.453, 'compound': -0.1531}
polarity_scores('bad good bad good bad good bad good bad good bad good bad good bad good bad good bad')
>>> {'neg': 0.547, 'neu', 0.0, 'pos': 0.453, 'compound': -0.8979}

It seems, the 'neg' and 'pos' scores are averages, whereas the 'compound' score is some sort of a sum. Thus, the compound score always takes on extreme values for long sequences, like Reddit posts or news articles.

This is particularly unfortunate, since a lot of beginners will blindly use the compound score without noticing this and get discouraged by the poor results. I suggest replacing the current implementation of the compound score with compound = pos – neg or completely removing it.

[VincentGurgul]

It's regrettable that a package that is part of NLTK has such a serious bug at all, and even more disheartening that it isn't being addressed...