SemTechLab at SemEval-2026 Task 5: Context-Aware Homonym Disambiguation via Span-Specific Interaction Features

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

Summary

The SemTechLab system, named HINTS, was submitted to SemEval-2026 Task 5, which focuses on rating the plausibility of word senses in ambiguous sentences within a narrative context. HINTS employs a hybrid Transformer architecture built upon nli-mpnet-base-v2. Distinct from conventional Cross-Encoders that rely solely on the "[CLS]" token, HINTS extracts span-specific embeddings for the target homonym from both the narrative context and its sense definition. It then computes interaction features, including concatenation, difference, and element-wise product, between these spans to explicitly model semantic alignment. The model was trained using Kullback-Leibler Divergence to predict the full distribution of human ratings. For the official submission, scores were rounded to integers (1–5), while continuous float scores were used for subsequent analysis. On the test set, HINTS achieved a Spearman correlation of 0.603 and an accuracy of 75.8% with its best configuration, utilizing local homonym features.

Key takeaway

For NLP Engineers developing word sense disambiguation systems, consider adopting span-specific interaction features rather than relying solely on "[CLS]" token embeddings. Your models, particularly those based on Transformer architectures like nli-mpnet-base-v2, can achieve higher semantic alignment accuracy by explicitly computing feature interactions (concatenation, difference, product) between context and sense spans. This approach yielded a 0.603 Spearman correlation and 75.8% accuracy in SemEval-2026 Task 5, suggesting a robust method for improving plausibility ratings.

Key insights

Span-specific interaction features enhance homonym disambiguation by explicitly modeling semantic alignment between context and sense definitions.

Principles

• Explicitly model semantic alignment for word sense disambiguation.
• Span-specific embeddings can outperform "[CLS]" token representations.
• Continuous scores offer improved metric sensitivity for evaluation.

Method

HINTS extracts span-specific embeddings for target homonyms from narrative context and sense definitions using nli-mpnet-base-v2. It computes concatenation, difference, and element-wise product interaction features, then trains with Kullback-Leibler Divergence.

In practice

• Implement span-specific embeddings for ambiguous words.
• Combine concatenation, difference, and product features.
• Utilize KL Divergence for predicting rating distributions.

Topics

• Homonym Disambiguation
• Word Sense Disambiguation
• Transformer Architectures
• Span Embeddings
• Semantic Alignment
• SemEval-2026

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

Related on AIssential

• IdioLink: Retrieving Meaning Beyond Words Across Idiomatic and Literal Expressions — Idiom understanding in retrieval requires semantic abstraction beyond lexical overlap.
• Adapting Multilingual Embedding Models to Turkish via Cross-Lingual Tokenizer Surgery and Offline Distillation — Efficient adaptation of multilingual embedding models to specific languages is achievable through tokenizer surgery and offline distillation.
• Sense2vec with spaCy and Gensim — Sense2vec enhances word2vec by learning context-sensitive word vectors for detailed text analytics.
• DySem: Uncovering Dynamic Semantic Components via Multilingual Consensus for Calculating Semantic Textual Similarity — DySem uses multilingual consensus to find dynamic, sample-specific semantic dimensions for improved STS in LLMs.
• MAAM: Anchor-Preserving Compression and Contextual Calibration for Chinese Discriminatory Language Detection — MAAM uses anchor preservation and contextual calibration for robust, compact Chinese discriminatory language detection.
• Multi-Granularity Reasoning for Natural Language Inference — MGRN enhances NLI by explicitly modeling hierarchical semantic interactions across transformer layers.

Open in AIssential →

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