blue at SemEval-2026 Task 5: NarrBERT : Narrative-Aware BERT for Word Sense Disambiguation

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

Summary

Team blue's NarrBERT model, a BERT-based cross-encoder regression system, was developed for SemEval-2026 Task 5: AmbiStory, which requires rating the plausibility of word senses in ambiguous sentences. NarrBERT addresses this by encoding the full narrative context, including precontext, the ambiguous sentence, and its ending, alongside candidate sense definitions and examples. Crucially, it enables token-level interaction between narrative cues and sense meanings, a departure from bi-encoder sentence-level methods, to capture subtle disambiguation signals. The model achieved a Spearman rank correlation of 0.66 on the development set. On the official test set, it recorded a Spearman correlation of 0.4866 and an Accuracy-within-Standard-Deviation of 0.6613, significantly surpassing sentence-transformer bi-encoder baselines.

Key takeaway

For NLP Engineers developing advanced word sense disambiguation systems, consider adopting cross-encoder architectures like NarrBERT. Your models should integrate comprehensive narrative context and enable token-level interactions to capture subtle semantic cues, moving beyond simpler bi-encoder approaches. This strategy can yield more human-aligned plausibility ratings, crucial for tasks requiring nuanced contextual understanding.

Key insights

NarrBERT uses a BERT cross-encoder with token-level narrative context to rate word sense plausibility, outperforming bi-encoders.

Principles

• Fine-grain contextual modeling is vital for human-like sense plausibility.
• Token-level interaction improves narrative word sense disambiguation.
• Cross-encoders can outperform bi-encoders for complex contextual tasks.

Method

NarrBERT employs a BERT-based cross-encoder regression model that encodes full narrative context (precontext, ambiguous sentence, ending) with sense definitions and examples, enabling token-level interaction.

In practice

• Integrate full narrative context for improved WSD.
• Prioritize cross-encoder architectures for fine-grained interaction.
• Evaluate WSD models using Spearman correlation and Accuracy-within-Standard-Deviation.

Topics

• Word Sense Disambiguation
• BERT
• Cross-encoder Models
• Narrative Understanding
• SemEval
• Contextual Reasoning

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• MiqraBERT: Regression-Based Sentence-BERT Finetuning for Biblical Hebrew Parallel Detection — MiqraBERT finetunes Sentence-BERT for Biblical Hebrew semantic parallel detection, significantly improving narrative textual reuse identification over lexical methods.
• LinkBERT: Improving Language Model Training with Document Link — Incorporating document links during pretraining significantly boosts language models' multi-hop reasoning and knowledge acquisition.
• BERT in NLP: How an Encoder-Only Transformer Changed Language Understanding — BERT's encoder-only, bidirectional pretraining revolutionized NLP by enabling broad language understanding and easy task adaptation.
• Finding Video Games with Sense2Vec — Sense2Vec's contextualized phrase embeddings, trained on Reddit, effectively identify domain-specific entities like video games and their abbreviations.
• Omnilingual SONAR: Cross-Lingual and Cross-Modal Sentence Embeddings Bridging Massively Multilingual Text and Speech — Omnilingual SONAR creates a unified semantic space for 4,200+ languages and modalities, achieving state-of-the-art cross-lingual and cross-modal performance.
• 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.

Open in AIssential →

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