Abstract
Word Sense Disambiguation (WSD) is a critical task in Natural Language Processing that involves identifying the correct sense of a word based on its context. This paper presents WSD-games, an unsupervised algorithm that models WSD as a non-cooperative evolutionary game. By treating words as players and their potential senses as strategies, the algorithm uses replicator dynamics to find a stable Nash equilibrium that maximizes textual coherence. Evaluation on the SemEval-2015 Task 13 benchmark shows that WSD-games achieves competitive results across multiple domains, particularly when using a modified Dice coefficient for word similarity.
Introduction
Determining the precise meaning of a word in context is a foundational challenge for NLP systems. While supervised models currently lead the state-of-the-art, they are often limited by the availability of labeled data and poor generalization across domains. Unsupervised methods offer a more scalable alternative but often struggle to capture the interdependent nature of word meanings within a sentence. WSD-games addresses this by conceptualizing a sentence as a dynamic system where the meaning of each word emerges from its interaction with others, aiming for a globally consistent labeling.
Methodology
The WSD-games algorithm frames disambiguation as a consistent labeling problem solved through Evolutionary Game Theory (EGT).
- Players and Strategies: Each word in a sentence is treated as a player in a game. The set of possible senses for that word (from WordNet 3.0) constitutes its available strategies.
- Graph Construction: A similarity graph is built where edges between words are weighted by co-occurrence statistics from the Google Web1T corpus, using measures such as Pointwise Mutual Information (PMI).
- Payoff Matrix: The "payoff" for a player choosing a specific sense is determined by its semantic relatedness to the senses chosen by neighboring players. This is calculated using the GlossVector measure.
- Replicator Dynamics: The algorithm employs an iterative process called replicator dynamics. Senses that are highly compatible with the surrounding context (providing higher payoffs) increase in probability, while less compatible senses are phased out. The process continues until the system reaches a Nash equilibrium, representing a coherent set of sense assignments for the entire text.
Results
The performance of WSD-games was evaluated on the SemEval-2015 Task 13 dataset, covering Mathematical, Biomedical, and General domains. The study compared three word-similarity measures: PMI, mDice (modified Dice), and $D^2$.
As shown in the table below, the mDice measure yielded the highest overall performance with an F1-score of 54.1%. The system demonstrated robust performance in the Biomedical (58.4% F1) and General (53.5% F1) domains. These results were comparable to contemporary state-of-the-art unsupervised systems, validating the game-theoretic approach.
Discussion and Conclusion
WSD-games demonstrates that word meaning can be effectively modeled as an emergent property of local interactions. By prioritizing "compatibility" over simple "importance" (as in PageRank-based methods), the algorithm better captures the semantic geometry of the text.
However, the system faces challenges in technical domains like Mathematics due to limited lexical coverage in WordNet. Future improvements could include integrating richer sense inventories like BabelNet, handling multi-word expressions, and utilizing semi-supervised initialization to boost the starting probability of likely senses. Despite these limitations, WSD-games provides a powerful, theoretically grounded framework for unsupervised semantic analysis.
Summary updated