Subjectivity in Sentiment Analysis: Addressing Bias in Word Lists

Abstract

Since the presidency of Franklin Roosevelt, the first 100 days of an administration has been used as a measuring stick to estimate the ability of a new president to govern. This is still true today. The first 100 days of the Trump administration invoked strong sentiment both for and against his policies. However, was the sentiment generally positive or negative or neutral? Using different sentiment analysis algorithms and Trump’s favorite social media platform, Twitter, we scraped over 181,000 English language tweets between January 20th, 2017 and April 29th, 2017 to get an idea of Twitter user sentiment regarding the new Commander-in-Chief during his first 100 days.

While our results reveal an interesting snapshot of the heightened emotions of the first 100 days of this presidency, they also raised some concerns regarding the bias inherent in the sentiment analysis process. More specifically, in the different dictionaries used to determine which words are “positive” and which words are “negative" issues of bias regarding race, gender, sexuality, and religion emerge. Therefore, it's important to "look underneath the hood," even when using a vetted dictionary, to examine the assumptions made, tweak the dictionary, and make transparent any assumptions left in the lexicon. We have parsed a further 16K tweets from the weekend of the Charlottesville protests to show what happens both before and after dictionary is tailored to an event focused on issues that are source of bias.