Presentation - Sentiment Analysis of Tweets by Topic and Region AKA Pulse

Overview of our goals

• Track public perception of social goods and ills
• Look at how this varies over time/space

Why our solution is good for the problem

• Twitter is fairly ubiquitous
• Hashtags are handy for finding the key topic of a tweet
• Tweets are relatively available as a data source
• Training data for twitter sentiment analysis already exists

Overview of our solution

• Train NN on dataset of tweets
• Use that to detect +/- opinions in tweets
• Create visual representation of this on a map
• Useful for policymakers, and for evaluation interventions & whether they worked

How our solution works

• We tried Naive Bayes, Support Vector Machines, Random Forest Classifiers, and Neural Networks. We raised accuracy by running a gridsearch to determine hyperparameters. Neural Networks worked best, with SVMs a close second.
• Neural Network for Sentiment Analysis
• NN will classify tweets that are input, with ~80% accuracy (so some noise will be in data)
• Use GIS software and location data to generate heatmap of positive/negative opinions?)
• Process is not fully automated yet - we can't pipe straight into mapping software from Python

Examples of our solution

• Show what the perception of Hackathons is, as a demo
• Show something actually useful - pick something with social utility
• Show something funny (National Sandwich Day?)

Extensions we'd like to do if we had more time

• Across space/time
• Compare different hashtags
• Automate the map creation after python program run
• Normalise for population

Built With

• keras
• python