Hello internet! This is Feelsbot. Feelsbot tries to understand how humans are feeling by reading their tweets. Feelsbot can analyze Tweets of any geographical location or any public Twitter profile.
Feelsbot uses artificial intelligence to analyze emotional sentiment in Tweets. Play around with Feelsbot and see if it surprises in any way.
How Feelsbot Works
Feelsbot uses a machine learning model from IBM to analyze emotions in tweets. Using this model, Feelsbot put tweets into five categories: joy, sadness, anger, fear, and disgust. This model only works in English, limiting the number of tweets Feelsbot can analyze. Each tweet receives a score, called a confidence score, of how strongly it matches one of those categories. Feelsbot puts tweets that have a confidence score higher than 65% into each category. Once the tweets have been categorized, the joy meter is calculated as the percentage of joyful tweets versus every other emotion.
When using the map, Feelsbot uses Twitter’s API to fetch the last 100 tweets that are geotagged near the location entered. If there are not many recent geotagged tweets near that location, Twitter fetches tweets of users whose profile locations are nearby. When analyzing tweets by a specific Twitter account, Feelsbot fetches the last 150 tweets by that account. Twitter profiles need to be public for Feelsbot to work.
Inspiration for Creating Feelsbot
There are a few reasons why I decided to create Feelsbot. First off, I was inspired by how easy it is to hop in and use some of the machine learning APIs from IBM and Google. Additionally, I wanted to create a software project to help people better understand how AI works.
Feelsbot highlights a couple of important things about AI and natural language processing.
- Many machine learning models have a rather simplistic view of emotion, and can’t pick up on subtle things, such as sarcasm.
- Sometimes figures of speech can throw off the model, for example, “I’m so happy I could cry” is detected as sad with a 65% confidence.
Although Feelsbot is not perfect, it’s more accurate than you might expect! I am often surprised by how well it can categorize more ambiguous tweets. Try it out to see for yourself.