Research map / Label-efficient and robust learning / Recognition and long-tail
Quantum Interference Model for Semantic Biases of Glosses in Word Sense Disambiguation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a quantum-interference WSD model that represents multiple BabelNet glosses of a sense as a quantum superposition state and uses the interference term as a confidence signal, reaching competitive F1 on English all-words WSD and cross-lingual XL-WSD.
Problem
Single-vector word sense representations, built from glosses across dictionaries or languages, ignore semantic biases between glosses of the same sense, losing information for word sense disambiguation.
Approach
Formalize each sense as a superposition of gloss state vectors (via BERT embeddings normalized to unit norm) and compute the match probability with the target word using a quantum interference model, where the cross-gloss interference term acts as a confidence signal.
Key results
- Our-Base reaches 80.1 F1 on the concatenated ALL test set, versus 79.0 for BEM and 77.0 for GlossBERT under the same framework.
- Our-Large reaches 81.6 F1 on ALL, the best among the compared systems in the paper's table, though the paper says gains over prior work are not significant due to different settings.
- Ablation: replacing multi-view glosses with a single representation drops ALL from 80.1 to 79.0, and removing the interference term drops it to 77.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add multi-view word-sense representations (ablation) | Our-Base · SE2 | F1-score: 79.4 -> 80.8 | Helped |
| Add multi-view word-sense representations (ablation) | Our-Base · SE3 | F1-score: 77.4 -> 78.5 | Helped |
| Add multi-view word-sense representations (ablation) | Our-Base · SE13 | F1-score: 79.7 -> 80.0 | Helped |
| Add multi-view word-sense representations (ablation) | Our-Base · SE15 | F1-score: 81.7 -> 82.9 | Helped |
| Add multi-view word-sense representations (ablation) | Our-Base · ALL | F1-score: 79.0 -> 80.1 | Helped |
| Add the quantum interference term (ablation) | Our-Base · SE2 | F1-score: 77.8 -> 80.8 | Helped |
| Add the quantum interference term (ablation) | Our-Base · SE3 | F1-score: 75.5 -> 78.5 | Helped |
| Add the quantum interference term (ablation) | Our-Base · SE13 | F1-score: 76.5 -> 80.0 | Helped |
| Add the quantum interference term (ablation) | Our-Base · SE15 | F1-score: 80.0 -> 82.9 | Helped |
| Add the quantum interference term (ablation) | Our-Base · ALL | F1-score: 77.1 -> 80.1 | Helped |
| Use multi-view word-sense representations and the quantum interference model | Our-Base · SE15 | F1-score: 82.9 vs 83.2 | Hurt |
| Use English glosses instead of glosses in the target language | Our model · Cross-lingual WSD datasets | F1-score | Mixed |
Surprising findings
- The model underperforms ARES on SE15 and on the Adv. set, which the authors attribute to dominance of high-frequency senses where a single representation suffices.
- On cross-lingual XL-WSD, the model uses English glosses for non-English contexts, and performs worse on niche or non-alphabetic languages.
Limitations
- Gains over prior work are not significant because the comparisons use different experimental settings and external resources.
- Cross-lingual evaluation substitutes English glosses for other languages, which hurts performance on non-alphabetic and niche languages.
- Multi-view representation can hurt performance on datasets dominated by high-frequency senses.
Experimental setup
Datasets: SemCor SemEval-2007 Senseval-2 Senseval-3 SemEval-2013 SemEval-2015 XL-WSD BabelNet glosses
Benchmarks: SE2 SE3 SE13 SE15 ALL (concatenated test sets) XL-WSD
Metrics: F1
Baselines: GlossBERT EWISE BEM SensEmBERT ARES QWSD XLMR-Base ESR Z-Reweighting
Builds on: BERT GlossBERT BEM
Closest papers by meaning
- Quantum-inspired Non-homologous Representation Constraint Mechanism for Long-tail Senses of Word Sense Disambiguation (AAAI 2025 · similarity 0.90)
- Quantum-Inspired Representation for Long-Tail Senses of Word Sense Disambiguation (AAAI 2023 · similarity 0.90)
- QPEN: Quantum Projection and Quantum Entanglement Enhanced Network for Cross-Lingual Aspect-Based Sentiment Analysis (AAAI 2024 · similarity 0.82)
- Retrieval-based Disentangled Representation Learning with Natural Language Supervision (ICLR 2024 · similarity 0.81)
- DOCS: Quantifying Weight Similarity for Deeper Insights into Large Language Models (ICLR 2025 · similarity 0.80)
- Inv-Entropy: A Fully Probabilistic Framework for Uncertainty Quantification in Language Models (NeurIPS 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.