Abstract
Few-shot stressor extraction from temporally shifting social-media text is challenged by unstable span boundaries and by a mismatch between structural confidence and task-semantic validity. We propose VFR-MISE, an asymmetric dual-evidence verification–fallback framework. VFR-MISE retains the temporal few-shot and knowledge-inheritance paradigm of MISE, strengthens candidate generation with DeBERTa-v3-base, and constructs two candidate-level signals: a guideline-conditioned verifier for semantic evidence and BIOES emission aggregation for structural evidence. Semantic evidence governs candidate admission; structural evidence is consulted only in the rejected region for constrained recovery, after which deterministic span merging produces the final output. In controlled reruns using identical data splits, training budgets, and Exact-span evaluation, replacing RoBERTa-base with DeBERTa-v3-base increases cross-shot F1 from 0.6841 to 0.7236 (+3.95 percentage points). With the DeBERTa-v3 backbone fixed, the complete VFR-MISE reaches 0.7248. These results suggest that stronger candidate representations and asymmetric dual-evidence decision refinement provide complementary benefits, improving the precision–recall trade-off while retaining an auditable decision process.

This publication has 28 references indexed in Scilit: