Detecting public discourse about COVID-19

Xu, Huimin; Erk, Katrin

Permalink

https://hdl.handle.net/2142/117364 Copy

Description

Title

Detecting public discourse about COVID-19

Author(s)

• Xu, Huimin
• Erk, Katrin

Issue Date

2023-03-13

Keyword(s)

• COVID-19
• NLI task
• Public discourse
• Lockdown policy and RoBERTa model

Abstract

The discussions about COVID-19 vary from government policy, treatment effec-tiveness, to origin of the virus. These controversial topics have increased the need for understanding the public’s opinions about COVID-19. Detecting the mutual relationship between opinions (entailment, contradiction, neutral) has important implications for governments to ease anxiety and take action. In this project, we fine-tune RoBERTa models based on 10,320 relationships between 637 COVID-19 related claims, reaching 82% accuracy. By comparing the difference between model prediction and human annotation, we found that models have huge room to improve in knowledge similarity and implicit negation. We also apply the trained models to another lockdown dataset on Twitter, in order to help us understand the adaptability and limitation of our COVID-19 models.

Publisher

iSchools

Series/Report Name or Number

iConference 2023 Proceedings

Type of Resource

• Other
• text

Language

eng

Handle URL

https://hdl.handle.net/2142/117364

Copyright and License Information

Copyright 2023 is held by Xu, Huimin, Erk, Katrin. Copyright permissions, when appropriate, must be obtained directly from the author.

Owning Collections