0

I am working on a dataset of airline customer complaints. Since it is "complaints" the general consensus is all the sentence are "negative" sentiment. So I am think of an approach to quantize the negativity score.

For example:

Less Negative review:

 "the cabin did not have enough leg space but the food was decent" - Score: 0.3

High Negative Review:

"complete service was horrible, I will not recommend them ever" - Score: 0.8

Any suggestions on existing approaches?

P.S I am not looking for an exact answer, any suggestions on the direction or approaches would be great.

data_person
  • 4,194
  • 7
  • 40
  • 75
  • you have to define what is your criteria for "less negative review" and "high negative review". you can extract positive and negative words from the comments and then give weight them. for example "horrible" has more weight and then you can increase the score for that review. – vb_rises May 24 '19 at 19:37
  • @Vishal interesting. Something like a weighted value of individual word polarity scores? – data_person May 24 '19 at 19:39
  • Yes. Weighted value. And it is not for all the words, may be for top common negative words only. – vb_rises May 24 '19 at 19:56

1 Answers1

0

As @Vishal suggests in the comments, the simplest approach to add more sophistication would be to implement a lexicon based sentiment analysis scoring model where you custom score the polarity of each word in the lexicon to be higher and lower. You should also include some bigram and trigram lexicon phrases for more accuracy.

Public sentiment analysis APIs from Google, Azure, IBM, etc. do provide a scale of sentiment too.

Adnan S
  • 1,852
  • 1
  • 14
  • 19