It is fun to watch Artificial Intelligence fail. Their logical innocence (stupidity?) can be hilarious. One of my first encounters with AI was with ELIZA — the 1966 NLP program — using the doctor package on emacs. I would spend hours making it say stupid things, and while we’ve come a long way with NLP since Eliza, computers still haven’t lost their innocence — just ask Siri if they know Eliza.
In the examples above gray is a neutral sentiment, blue and its deeper shades are for negative sentiment, and pink and it's brighter shades are used to show positive sentiment.
While it can handle negation and two word phrases it is no match for sarcasm, unless of course you use the right emoji: 😝
But human sentiments aren’t limited to tokenized words or emojis, they have layers of meaning. Testing the package with songs and poetry really reveals the depth of the problem we’re trying to solve for. It is a tall order for any sentiment analysis technique.
Here are a few lines from the poem Figures in a Landscape by Doppler by Ranjit Hoskote. Notice how it only catches the positive sentiment at the end:
Seeing AI fumble, and sometimes even fail, is the how we learn more about it. It helps us decide the direction of our work and look at new ways of solving these problems. Stay tuned for our next experiment where we compare different approaches to sentiment analysis.
The GIFs above were made using our testing tool Sentimental, do give it a try.