Traditional natural language processing treats word meanings as a set of discrete entries in a dictionary. In this view, a word like break has a fixed list of senses, and the task of a system is to pick the single "best" one for a given context. However, this rigid approach fails to capture the inherent fluidity of human language. Research in lexical semantics is shifting toward models that represent meaning as a continuous, graded, and context-dependent space rather than a collection of mutually exclusive categories.
The Problem with Discrete Inventories
The reliance on static sense inventories, such as WordNet, creates several bottlenecks. As Dligach and Palmer argue, rigid dictionaries fail to account for the multi-faceted nature of word meanings, leading to information loss. When forced to choose one sense from a list, both humans and machines struggle with instances that fall between or across semantic boundaries.
This difficulty is reflected in low inter-annotator agreement for traditional Word Sense Disambiguation (WSD) tasks. Bevilacqua et al. note that these artificial boundaries do not reflect how language is actually used. Similarly, Rumshisky suggests that standard inventories lack "membership flexibility," preventing models from understanding the "fuzzy" boundaries that native speakers naturally navigate.
Measuring Meaning on a Spectrum
If word senses are not discrete buckets, how should we represent them? One approach is to treat meaning as a matter of degree. Erk et al. demonstrated through the WSsim and Usim experiments that human annotators naturally employ graded judgments. Instead of picking one sense, humans can rate how applicable multiple senses are to a single usage, capturing typicality effects and borderline cases.
To formalize this, researchers have turned to Word Usage Graphs (WUGs). Instead of mapping a word to a predefined sense, these graphs map the "semantic proximity" between different occurrences of a word. Schlechtweg et al. use a Weighted Stochastic Block Model to infer sense structures from these graphs, providing a principled way to handle graded meanings. This approach has been scaled by Schlechtweg and Tahmasebi to track how these fluid meanings evolve over time across multiple languages.
From Classification to Generation
The shift toward continuity is also visible in how models are built. Rather than treating WSD as a classification problem (picking a label), newer methods treat it as a representation or generation problem.
- Generative Glossing: Bevilacqua et al. introduced Generationary, a model that generates context-specific definitions (glosses) from scratch. By moving away from a "pick-from-a-list" paradigm, the model can describe nuanced meanings that might not exist in a static dictionary.
- Contextualized Embeddings: Hettiarachchi and Ranasinghe show that modern transformer-based embeddings can capture subtle, graded shifts in meaning influenced by surrounding text, outperforming methods that rely on discrete sense assignments.
- Probabilistic Frameworks: Bernardy et al. propose a compositional Bayesian semantics where sentence interpretation is treated as probabilistic inference, allowing the system to handle semantic vagueness and degrees of truth.
Evidence Across Modalities and Languages
The limitations of discrete senses are not confined to text. In the realm of computer vision, Raganato et al. highlight how static inventories struggle with the continuous semantic space required for Visual WSD, where a word must be linked to an image.
Cross-lingual research provides further evidence. Muramoto et al. mapped "Word Translation Networks" to show that the space between languages is a continuous web of cultural and semantic overlaps that bilingual dictionaries often miss. By analyzing the polysemous verb "break," Petersen and Potts demonstrate that Large Language Models (LLMs) naturally capture these nuanced, usage-based distinctions, suggesting that the future of lexical semantics lies in these fluid, high-dimensional representations rather than fixed sense lists.
Key Foundational Work
Investigations on Word Senses and Word Usages (2009) — contributed 2 concept elements to this direction.
Novel Semantic Features for Verb Sense Disambiguation (2008) — contributed 1 concept elements to this direction.
Semantic Visualization and Meaning Computation (2008) — contributed 1 concept elements to this direction.