In the Positive Negative Sentiment Analysis, Would it make sense mathematically to instead of keeping a score of the positive frequencies and negative frequencies of a word, calculate the difference between them? That way each word would have a positivity 'heat' in which a very high value would indicate a very positive word and vice-versa. How this approach would change the model performance?
Let's take an example, consider two words A & B. A's positive/negative values are +1/0 and B's are +0.5/-0.5. Their difference would appear equivalent (diffAB = 1). When in fact they are quite different sentiments.
If you wanted to compute a single metric, you could do something like "polarity" i.e. take the squared sum of the values which would give you a positive value showing how "polar" the word is. In the case above, the word B would have a lower score because it's positive/negative values are less extreme.
When designing metrics, test a few scenarios to see if it matches your needs.