Soft-Vote Ensembling for Counter-Causal News Extraction via Touch'e
October 5, 2026
The Touch'e system addresses counter-causal claims in news by ensembling three RoBERTa-large BILOU+CRF taggers. Instead of voting on spans, it uses soft-vote ensembling by averaging token-level scores to improve causality detection and extraction accuracy on the CCNC corpus.
HOW THIS AFFECTS YOU
●
researcherThis shows how token-level score averaging in ensembles can outperform discrete span voting for extraction tasks.