Emotional States vs. Emotional Words in Social Media

Abstract

A number of social media studies have equated people's emotional states with the frequency with which they use affectively positive and negative words in their posts. We explore how such word frequencies relate to a ground truth measure of both positive and negative emotion for 515 Facebook users and 448 Twitter users. We find statistically significant but very weak (ρ in the 0.1 to 0.2 range) correlations between positive and negative emotion-related words from the Linguistic Inquiry Word Count (LIWC) dictionary and a well-validated scale of trait emotionality called the Positive and Negative Affect Schedule (PANAS). We test this for tweets and Facebook status updates, focus on different time slices around the completion of the survey, and consider participants who report expressing emotions frequently on social media. With rare exception, this pattern of low correlation persists, suggesting that for the typical user, dictionary-based sentiment analysis tools may not be sufficient to infer how they truly feel.