Sentiment Analysis Vader Normalization of results

LaraNeuLaraNeu Member Posts: 4 Learner I
Hi RapidMiner Community,

I have a question about the results of my sentiment analysis I conducted for online reviews from Airbnb that I want to compare with ratings given from the reviewers. As a result of the Extract Sentiment operator using Vader I get unstandardized values that I can hardly compare with other values like ratings (am I correct).
In another post in the forum by @mschmitz I found two formulas to normalize the sentiment scores and get results between -1 and +1:
- Sentiment Score/Total tokens
- Sentiment Score/(Total tokens - uncovered toked)

I used both formulas for my data and obviously got two different results. Can someone help me with the interpretation of the results for both formulas? What is the 'right' formula? What is the best way to standardize my data to be able to compare it with the ratings?

Thanks a lot for your help! Really appreciate it!

Answers

  • mschmitzmschmitz Administrator, Moderator, Employee, RapidMiner Certified Analyst, RapidMiner Certified Expert, University Professor Posts: 3,233 RM Data Scientist
    i think there is no real 'correct'. The question is how to normalize the score correctly. I think all three approaches (incl. don't normalize at all) are valid ones.
    Why don't you just calculate the correlation to the stars or a metric like information gain to see whats best?

    ~Martin
    - Head of Data Science Services at RapidMiner -
    Dortmund, Germany
Sign In or Register to comment.