Coarse Grained Word Sense Disambiguation

This direction is split into other directions: semantic granularity mismatch. Its last version stays readable and citable here.

Version 1 of 183 papers113 conceptsfrozen 2026-10-04

In the quest to make machines understand human language, researchers have long grappled with Word Sense Disambiguation (WSD)—the task of determining which "sense" of a word is being used in a given context. For decades, the gold standard for this task has been WordNet, a massive lexical database that defines thousands of subtle distinctions between word meanings.

However, a recurring theme across 67 research papers suggests a fundamental problem: WordNet and similar dictionaries are often too detailed for practical use. This "excessive granularity" creates a performance ceiling that neither human experts nor state-of-the-art AI can easily break. The solution emerging from the field is a shift toward Coarse-Grained Sense Inventories, which group hyper-specific "micro-senses" into broader, more distinct semantic categories.

The Granularity Bottleneck

The primary obstacle in traditional WSD is that the distinctions between dictionary definitions are often metaphysical rather than functional. As Roberto Navigli notes in Valido: A Visual Tool for Validating Sense Annotations, high granularity leads to low inter-annotator agreement. If two linguists cannot agree on which of five nearly identical definitions of "call" applies to a sentence, a machine learning model has little hope of learning a stable pattern.

This over-specification creates several issues:

  1. Data Sparsity: When a word has 20 fine-grained senses, the available training data is fragmented. Models struggle to find enough examples for each "micro-sense" to build a robust representation.
  2. Artificial Complexity: Many downstream tasks, such as machine translation or sentiment analysis, do not need to know the difference between two subtle nuances of a verb. They only need to know if the word is being used in a "financial" vs. "geographic" context.
  3. Evaluation Noise: In A Method for Word Sense Disambiguation of Unrestricted Text, Rada Mihalcea points out that systems are often penalized for choosing a "semantically reasonable" sense that simply doesn't match a specific, arbitrary gold-standard tag.

Clustering as a Semantic Zoom Lens

To overcome this, researchers have developed various methods to "zoom out" and cluster these fine-grained senses into more manageable groups.

One influential approach involves Topic Signatures. Eneko Agirre and Olatz Ansa demonstrated in Enriching very large ontologies using the WWW that by mining the web for words co-occurring with specific senses, they could create "signatures" that naturally cluster related meanings. Similarly, Noriko Tomuro used "tree-cut" algorithms in Tree-Cut and a Lexicon Based on Systematic Polysemy to identify predictable patterns where words share related meanings, such as an animal and the food derived from it (e.g., "chicken").

By grouping these senses, researchers create a "middle ground" that preserves enough detail for understanding while removing the noise that confuses algorithms. Cecily Jill Duffield and Jena D. Hwang showed in Criteria for the Manual Grouping of Verb Senses that adopting these intermediate groupings can improve machine learning performance by as much as 16%.

Psycholinguistic and Practical Validation

The push for coarse-grained senses isn't just a computational shortcut; it appears to reflect how the human brain actually works. In Choosing Sense Distinctions for WSD: Psycholinguistic Evidence, Susan Windisch Brown provides evidence that humans process closely related senses almost identically. This suggests that the hyper-fine distinctions in traditional dictionaries are often "metaphysical" rather than cognitive.

This shift has led to the creation of new benchmarks like the Coarse Sense Inventory (CSI). As described in CSI: A Coarse Sense Inventory for 85% Word Sense Disambiguation, mapping WordNet's thousands of senses to just 45 domain-based labels (like FOOD, PERSON, or FINANCE) allowed neural models to shatter the long-standing 80% accuracy ceiling, reaching an F1 score of 85.9%.

The Role of Modern Language Models

Recent work has re-evaluated the capabilities of Pretrained Language Models (PLMs) like BERT through this lens. In Analyzing Homonymy Disambiguation Capabilities of Pretrained Language Models, researchers found that while BERT might struggle with the "picky" senses of WordNet, it distinguishes unrelated homonyms with up to 95% accuracy without any specific training.

This suggests that the perceived limitations of AI in understanding word senses may be a byproduct of the evaluation frameworks we use. When we move toward coarse-grained classification—as seen in Oscar Sainz and German Rigau's work on zero-shot WSD—we find that modern models already possess a sophisticated latent understanding of broad semantic categories.

Conclusion

The evolution of word sense research is a move from dictionary-matching to semantic-mapping. By recognizing that "more detail" does not always equal "better understanding," the field has found that clustering fine-grained senses into coarse-grained categories creates more robust, interpretable, and high-performing systems. This transition aligns computational targets with both human cognitive patterns and the practical needs of real-world language applications.

Papers in this version

Showing 60 of 83 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_coarse_grained_word_sense_disambiguation_55867cd9,
  title = {Coarse Grained Word Sense Disambiguation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/coarse-grained-word-sense-disambiguation/rdr_55867cd96c2e4f92b5956819b8187f58},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_coarse_grained_word_sense_disambiguation_55867cd9_v1,
  title = {Coarse Grained Word Sense Disambiguation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/coarse-grained-word-sense-disambiguation/rdr_55867cd96c2e4f92b5956819b8187f58/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}