Cross-corpora argument analysis using textual entailment

Algis Petlin, Sue-Ellen Duffy, Ankita Gupta, Brendan O'Connor

Keywords: argument mining, corpus analysis, textual entailment

Submission Type: Non-Archival

Abstract: We present an entailment-based clustering method for conducting cross-corpora argument analysis. Unlike traditional clustering based on semantic similarity, our approach uses natural language inference to cluster textual propositions based on entailment relationships. We then apply our method to 2.34 billion Reddit discussions across three vaccines, namely COVID-19, HPV, and MMR. Our results demonstrate that the entailment-based clustering method better preserves distinct argumentative relations, compared to traditional hierarchical clustering, and uncovers reasoning patterns involved in vaccine hesitancy across multiple different vaccines.

Submission Number: 26