Distance from Unimodality for the Assessment of Opinion Polarization

Area of Science:

• Computational linguistics
• Social computing
• Natural language processing

Background:

• Current methods for analyzing commonsense knowledge and opinions often overlook polarized viewpoints.
• The standard approach of using a 50% positive classification threshold can misrepresent data with divided opinions, impacting analyses of subjectivity, sentiment, and toxicity.

Purpose of the Study:

• To introduce a novel metric, distance from unimodality (DFU), for quantifying opinion polarization.
• To demonstrate the effectiveness of DFU in analyzing real-world datasets, including social media and toxicity annotations.

Main Methods:

• Developed the distance from unimodality (DFU) measure to estimate opinion distribution polarization.
• Applied DFU to analyze tweet sentiment data over nine months during the pandemic and crowd-annotated toxicity data.
• Utilized DFU as an objective function for training predictive models.

Main Results:

• DFU correlates well with human judgment in assessing opinion polarization.
• Identified specific days with polarized sentiment in tweets, varying by US state.
• Found that polarized opinions on toxicity are more common among annotators from different countries.
• Demonstrated DFU's utility in training models to predict future opinion polarization.

Conclusions:

• DFU offers a more nuanced approach to understanding opinion dynamics than traditional methods.
• The findings highlight the importance of considering opinion polarization in social media and toxicity analysis.
• DFU can be leveraged for predictive modeling of opinion polarization.