Suman Deb Roy
Jul 10, 2017 · 1 min read

I am not certain about sentiment analysis performance either. If you observe the dataset with two classes (yelp or amazon polarity), the accuracy is above 90. But if you notice the sentiment datasets with 5-classes (amazon full or yelp full), the accuracy again suffers 4-fold degradation, and errors are >37. Binary classification is of course easier, and the drastic error on increasing number of classes is what bothers me. Yeah, curious to see how LSTMs do

Suman Deb Roy

Written by

Algorithms, startups and media