Experimental evaluation demonstrates improved stressor extraction in social media text, indicating refined precision-recall trade-offs through dual-evidence verification.
Key Points
Develop an asymmetric dual-evidence verification–fallback framework (VFR-MISE) to resolve unstable boundaries and semantic mismatches in few-shot stressor extraction from social media text.
Built a candidate generation pipeline using DeBERTa-v3-base, coupled with a guideline-conditioned semantic verifier and BIOES emission-based fallback recovery for high-confidence rejected candidates.
Evaluated candidate extraction and end-to-end framework performance across 20 random seeds against RoBERTa baselines and profiled inference latency with a lightweight MiniLM verifier.
The complete DeBERTa-based system achieved an Exact-span F1 of 0.7248 versus 0.6977 for the RoBERTa baseline (+2.70 percentage points; exact sign-flip test, p = 0.0000458).
The full VFR-MISE framework increased the mean cross-shot Exact-span F1 by 0.33 percentage points over the A0 baseline.
Deploying MiniLM in place of the DeBERTa verifier decreased end-to-end inference latency by 26.28% at a batch size of 16.