Ambirig at SemEval-2026 Task 5: Distributional Ordinal Modelling for Ambiguous Word Senses in Narrative Contexts

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

Soumyajit Roy's Ambirig system addresses the SemEval-2026 Task 5, which challenges traditional Word Sense Disambiguation (WSD) by requiring systems to predict human-perceived plausibility scores for ambiguous word senses in short narratives, moving beyond categorical decisions. The research hypothesizes that standard classification objectives are inadequate for capturing the ordinal nature of human uncertainty in this domain. While experiments with complex auxiliary tasks like Siamese networks, Task-Adaptive Pretraining (TAPT), and Natural Language Inference (NLI) transfer learning proved ineffective in low-resource settings, the proposed solution is a streamlined architecture. This system utilizes DeBERTa-v3-base with a GlossBERT-style Cross-Encoder, optimized using Earth Mover's Distance (EMD) loss. By framing the problem as ordinal regression over a probability distribution and enhancing inputs with prototypical examples, Ambirig achieved 73% accuracy and a Spearman correlation of 0.593, establishing a robust baseline that surpasses more complex, parameter-heavy methods.

Key takeaway

For NLP Engineers developing systems that require nuanced understanding of word sense plausibility, this research suggests a shift from categorical classification. If you are tackling tasks like SemEval-2026 Task 5, consider implementing an ordinal regression approach with Earth Mover's Distance (EMD) loss, utilizing models such as DeBERTa-v3-base. This method, enriched with prototypical examples, offers a robust baseline (73% accuracy, 0.593 Spearman correlation) that outperforms more complex architectures, particularly in low-resource scenarios, optimizing for both performance and computational efficiency.

Key insights

Ordinal regression with EMD loss on DeBERTa-v3-base effectively models human-perceived word sense plausibility, outperforming complex methods in low-resource WSD.

Principles

• Human language understanding is graded.
• Ordinal uncertainty needs specific models.
• Simpler models can beat complex ones.

Method

Utilizes DeBERTa-v3-base with a GlossBERT-style Cross-Encoder. Optimizes with Earth Mover's Distance (EMD) loss, framing the problem as ordinal regression over a probability distribution, and enriches inputs with prototypical examples.

In practice

• Use EMD loss for graded predictions.
• Employ DeBERTa-v3-base for WSD.
• Enrich inputs with prototypical examples.

Topics

• Word Sense Disambiguation
• SemEval-2026 Task 5
• Ordinal Regression
• Earth Mover's Distance
• DeBERTa-v3-base
• Cross-Encoder

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• DiffoR: A Unified Continuous Generative Framework for Universal Ordinal Regression — DiffoR redefines Ordinal Regression as a continuous generative task using diffusion models to overcome quantization and capture soft semantic transitions.
• SemEval-2026 Task 4: Narrative Story Similarity and Narrative Representation Learning — The NSNRL task operationalizes narrative similarity for evaluation and identifies areas for automated system improvement.
• The Origins of Stochasticity: Comprehensive Investigations on Uncertainty Quantification for Large Language Models — A new taxonomy and evaluation framework reveal that consensus-based methods excel in quantifying LLM uncertainty, which decreases with model scale.
• World Properties without World Models: Recovering Spatial and Temporal Structure from Co-occurrence Statistics in Static Word Embeddings — Recent research demonstrates that static co-occurrence-based word embeddings, such as GloVe and Word2Vec, can recover significant spatial and temporal structure, challenging the interpretation that su…
• Quantifying and Predicting Disagreement in Graded Human Ratings — Annotation disagreement in graded human ratings can be predicted from textual features, especially for inappropriate language.
• LaViSA: A Language and Vision Structural Ambiguity Benchmark — VLMs struggle with visual disambiguation of structural ambiguity, especially subtle semantic distinctions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.