Weighting of words in lexicon based sentiment analysis

I have a a question regarding my current project, i am trying to do a lexicon based sentiment analysis on my data, where i calculate the sentiment score as following:

$$Score = \frac{\sum_{i}{word_i}}{\mid words \mid}$$

So according to the score the word will be classified in either negative or positive. But i have also calculated for every word in the article its salience and frequency and would like to know if its possible to use them in my sentiment analysis formula above.

 words| salience| frequence
happy  0.5       2


Yes, you can. Not quite sure what else to add. Your formula can then look like:

$$Score = \sum_{i}{f(salience_i, frequency_i, sentiment_i)}$$

Where $$f$$ is a function that weighs your sentiment score with the salience and frequency. Up to you to define how.

• What if you don't know which $$f$$ to use?

Now, bear with me, this isn't something I've tried per se, but this could be an interesting approach. You could use a recurrent neural network and your input could be the salience, frequency, and sentiment score for each word. Not only will your RNN "create" (ideally) the best $$f$$ for your particular problem, but it will also use the sequential information of the words, which may even improve your results.

• The thing is i want to multiply the salience with the sentiment, but iam not sure about multiplying the result with the frequency, imo that wouldnt mke that much sense as its already relevant in calculating the salience. Do you know any papers regarding this matter? – voltage May 7 '20 at 21:54
• *I mean adding the values, not multiplying them – voltage May 7 '20 at 22:16
• To make sure we are on the same page, can you post the formula you use to compute the salience. Anyway, I edited my answer to provide an alternative way to do this using RNNs – Valentin Calomme May 8 '20 at 7:33