Sentiment Analysis Offers a Better Way to Conduct Polls

Screenshot from estorm.org displaying Twitter sentiment in terms of political orientation.

The majority of the polls for the 2016 Presidential Election turned out to be mistaken. One possible reason for this is that the majority of polls are based on survey questions.

However, a group of engineering researchers may have found a better polling method by using the election to get an idea of popular opinion without resorting to survey questions.

“Sentiment analysis” is a software which combines information and emotion. It was developed by engineering researchers to determine how a person feels based on what they say through verbal or written communication.

The researchers started with over 250 million tweets posted from around the world between June 5^th and October 30^th of last year. They then weeded out all non-political tweets by using a specific set of keywords. The resulting data set consisted of more than 1.6 million political tweets.

The remaining tweets were then processed through the sentiment analysis software. A score between 0 and 1 was assigned to each tweet, where 0 signified the most negative sentiment, 1 signified the most positive sentiment and 0.5 indicated neutral sentiment. All of the scores were then collected in a database that calculated an area’s political orientation, in real-time, based on the tweets published from that county or state. This means the database was continuously updated with new tweets.

In order to measure the accuracy of the model, the team compared their sentiment analysis software results to the New York Times Upshot election forecast website. The results were very similar for state-by-state analysis, with the software matching up with the major events which happened during the election season.

While the final election result signifies an upset for forecasts and predictions, the sentiment analysis software is not limited to predicting electoral outcomes. The engineering researchers behind this development believe their tool could be deployed in situations involving crowd-sourced opinions on the Internet to more accurately reflect feelings.

Furthermore, it could be incorporated with voice-enabled assistants in order to determine what the user wants by using how the user says what they want in addition to what they actually say.

You can explore the results of using sentiment analysis on the presidential race by visiting this website.

For more natural language processing news, find out how surfing the web improves machine learning.