Zoom in to a "hot" area and click a circle to see that individual Tweet, which contains instances of danger such as robberies, assaults, battery, etc. You can also search for a particular place as well as rotate the map.
First, SafeTweet detects whether or not a Tweet contains a dangerous incident.
For example, these two Tweets contain a dangerous incident:
while these two Tweets do not contain danger:
To detect danger, first I queried 1000 Tweets using the Twitter API. These Tweets were hand-labelled for instances of danger and then used to train a deep learning model with Keras.
Furthermore, I used Snorkel AI weak supervision to generate labels for the model using danger keywords, keyword inflections, and Vader sentiment analysis. I chose to use weak supervision in order to clean and correlelate the Tweet data, since it is unlabelled and often messy.
Finally, the model was fine tuned with Google's BERT Natural Language Processing Transformer and achieved an accuracy of 93%.
Using the Twitter API, over 50,000 Tweets were queried from police departments of major cities, including @NYPDnews, @SeattlePD, @SFPD, and many more. The location of the incidents were extracted from the Tweets using the Google Geocoding API, and each Tweet was plotted at the location of the incident using the MapBox API.
The website is hosted with Google Firebase.