Abstract
Visual Word Sense Disambiguation (VWSD) is the task of identifying the correct image for a target word based on its linguistic context. While current state-of-the-art models utilize dictionary definitions (glosses) to aid disambiguation, they often suffer from semantic bias and fail to capture the complex, overlapping nature of word meanings. This paper introduces the Quantum Inference Model for Visual Word Sense Disambiguation (Q-VWSD). By representing ambiguous word definitions as quantum superpositions rather than independent classical probabilities, the model leverages "interference effects" to dynamically adjust semantic understanding. Our results demonstrate that Q-VWSD significantly outperforms classical inference methods, particularly when handling polysemous words and noisy definitions generated by Large Language Models (LLMs).
Introduction
In natural language, words are frequently polysemous—possessing multiple distinct meanings. In the VWSD task, a model must decide, for instance, whether "Andromeda" refers to a galaxy, a mythological princess, or a flowering shrub based on the surrounding text.
Classical models typically address this by calculating the probability of an image matching each definition and summing these probabilities. However, this approach follows the "law of total probability," which assumes definitions are independent variables. This often leads to semantic bias; if a dictionary is skewed toward one meaning, the model may ignore the correct context. We argue that human language is inherently uncertain and that quantum probability provides a more robust framework for capturing the "interference" between different word senses.
The Q-VWSD Framework
The proposed Q-VWSD model shifts the paradigm from classical to quantum probability through several key innovations:
- Representation as Superposition: Instead of weighting definitions classically, multiple glosses are encoded into a single quantum superposition state ($|G\rangle$). This allows the model to maintain a holistic representation of all possible meanings simultaneously.
- Quantum Inference Circuit (Q-VWSD_QC): The model utilizes a quantum circuit where context and glosses are encoded as quantum states. When a candidate image is introduced as an observation, the state "collapses," and the resulting probability determines the image's rank.
- The Interference Term: Mathematically, the quantum approach introduces an extra term ($Int$) that classical models lack. This term represents the interference between different meanings, allowing the model to dynamically refine its judgment based on the specific image being evaluated.
To ensure practical utility given current hardware limitations, we also developed a Quantum-Inspired Heuristic (Q-VWSD_QI). This version runs on classical GPUs using complex-valued vectors to simulate quantum interference, making the model scalable for large-scale datasets.
Experimental Results
We evaluated Q-VWSD using the SemEval-2023 VWSD Task dataset, comprising nearly 13,000 samples. Glosses were sourced from WordNet 3.1, GPT-3.5-turbo, and GPT-4.0 to test the model's robustness against varying levels of definition quality.
- Superior Accuracy: Q-VWSD consistently outperformed standard models like CLIP and FLAVA. Using CLIP-Large, the quantum-inspired model achieved a Hits@1 of 89.63%, surpassing the best classical inference configuration (88.62%).
- Robustness to Polysemy: The model's strength was most evident in cases with high polysemy (10+ meanings). While classical models often treat additional definitions as noise, Q-VWSD maintained high accuracy by effectively "filtering" meanings through interference.
- Mitigating LLM Bias: LLMs often generate biased or "noisy" definitions. Q-VWSD proved more robust than classical models in these scenarios, as the superposition approach is less prone to making arbitrary decisions based on skewed dictionary data.
Conclusion and Future Work
Q-VWSD demonstrates that the logic of quantum mechanics provides a superior framework for modeling the ambiguity of human language. By violating the classical law of total probability, the model more closely mimics human intuition. Future research will focus on deeper modeling of the "phase" in the interference term to capture nuances such as sarcasm and emotion, as well as scaling the true quantum circuit version as quantum hardware continues to evolve.
Summary updated