Rating Plausibility of Word Senses in Ambiguous Sentences Using Multi-Architecture Analysis

AbstractWord sense disambiguation in narrative contexts requires systems to reason about subtle semantic relationships between candidate senses and discourse context. This paper addresses SemEval 2026 Task 5, which reformulates WSD as a graded plausibility prediction problem on a 1–5 Likert scale using the AmbiStory dataset. We present two complementary approaches: (1) a DeBERTa-v3-Large encoder with attention-weighted pooling and ordinal regression, achieving a Spearman correlation of 0.718, and (2) a rank-based ensemble combining FLAN-T5 and RoBERTa, achieving 0.692. Through ablation studies, we show that explicitly modeling ordinal structure improves performance over standard regression by 17.3%. We further analyze the strengths of each approach, showing that fine-tuned encoders capture fine-grained semantic relationships, while ensemble methods provide robustness through complementary modeling biases. Our results provide a detailed empirical analysis of design choices for graded plausibility prediction in narrative understanding.