Resolving Excessive Granularity in Word Sense Inventories

40 concepts35 papers1 proposal

When we think of word meanings, we often imagine clear-cut boundaries: a financial bank versus a river bank. However, traditional lexical databases and automated systems often slice meanings far more finely. For instance, a single verb like make can have dozens of micro-definitions in a dictionary, distinguishing between subtle contextual nuances that are nearly imperceptible in everyday use.

This phenomenon is known as sense explosion or lexical over-specification. While capturing every semantic shade seems ideal for completeness, it creates a severe semantic granularity mismatch for practical natural language processing (NLP). When a system is forced to choose between nearly identical definitions, accuracy plummets, human annotators disagree, and downstream applications suffer from unnecessary noise.

To build robust language systems, researchers have shifted away from hyper-specific sense inventories toward methods that dynamically group, merge, or restructure word meanings into coarse-grained, functional categories.


Why Algorithms and Dictionaries Over-Split Meanings

The problem of excessive granularity stems from two distinct sources: human lexicography and unsupervised machine learning.

In human-curated resources like WordNet, lexicographers aim for exhaustive coverage. However, as Sun et al. observe, this high level of detail results in overlapping semantic representations that reduce the precision of computational models. Frequent words are particularly prone to this; as Kallini et al. demonstrate, light verbs explode into dozens of overlapping dictionary definitions, making consistent classification nearly impossible for both humans and machines.

When we turn to unsupervised machine learning to automatically induce word senses from raw text, the issue is often magnified. Without explicit constraints on semantic granularity, clustering algorithms tend to over-split concepts based on minor contextual, syntactic, or stylistic variations rather than true semantic shifts.

[Raw Text Contexts] ──> [Unsupervised Clustering] ──> [Over-Split Senses]
"apple pie"             ───────┐
"apple orchard"         ───────┼─> Cluster 1 (Fruit/Agriculture)
"crisp apple"           ───────┘
"eat an apple"          ─────────> Cluster 2 (Ingestion Context)  <-- Redundant Split!

As Logacheva et al. point out, unsupervised sense induction frequently fragments a single coherent concept into multiple clusters based on functional differences. Non-parametric Bayesian models like AdaGram face similar hurdles; Bartunov et al. show that when hyperparameters favor high sensitivity, models generate redundant prototypes for the same underlying meaning. Similarly, Chang et al. note that Bayesian non-parametric models regularly over-segment word meanings compared to human intuition, creating a barrier when aligning induced clusters with standard lexical resources.


The Downstream Cost of Over-Specification

Intuition might suggest that more detail is always better for an AI system. In practice, forcing a model to resolve micro-distinctions introduces a massive amount of classification noise that degrades performance on real-world tasks.

Information Retrieval

In Information Retrieval (IR), one might expect that resolving a query word to its exact WordNet sense would yield highly precise document matching. However, Uzuner et al. discovered that the noise introduced by imperfect, fine-grained Word Sense Disambiguation (WSD) actually degrades IR performance compared to standard keyword searches. Because the micro-distinctions are functionally irrelevant for matching documents, forcing a system to choose between them creates unnecessary points of failure.

Sentiment Analysis and Text Classification

Downstream tasks rarely require dictionary-level precision. For example, in sentiment analysis, a system needs to know if a word is being used objectively or subjectively, not its exact dictionary sub-sense. Akkaya et al. demonstrated that treating subjectivity as a coarse-grained grouping of dictionary senses makes the disambiguation task far more feasible and significantly boosts the accuracy of sentiment systems.

When models are evaluated on hyper-specific inventories, it becomes difficult to tell if a model's poor performance reflects a lack of linguistic understanding or simply the arbitrary difficulty of the classification schema (Tedeschi et al.).


The Solution: Finding the Semantic Middle Ground

To resolve this granularity mismatch, researchers have developed frameworks to map hyper-specific senses into broader, more robust categories. This process relies on three main strategies: training on semantic classes, leveraging systematic linguistic patterns, and structural reorganization.

1. Training on Semantic Classes and Basic Level Concepts

Instead of forcing a model to predict one of a hundred micro-senses, systems can target mid-level abstractions. Izquierdo et al. introduced the use of Basic Level Concepts (BLCs)—automatically derived mid-level abstractions from WordNet hierarchy.

By training classifiers on these broader semantic classes rather than individual word senses, the authors bypassed the knowledge acquisition bottleneck (Izquierdo et al.). This coarse-graining allows systems to automatically harvest cleaner, monosemous training examples from domain texts, leading to higher accuracy and more stable performance (Izquierdo et al.).

Hyper-Fine Senses (WordNet)        Mid-Level Abstraction (BLC)       Downstream Task
[chicken: animal, bird]      ───┐
[chicken: poultry, food]     ───┼─>   [ANIMAL / FOOD CLASS]   ───>   High-Accuracy WSD
[chicken: cowardice slang]   ───┘

2. Exploiting Systematic Polysemy

Word meanings do not drift at random; they follow predictable, regular patterns across a language. For example, the relationship between an animal and its meat (chicken, turkey) or an institution and its building (school, bank) is highly systematic.

Noriko Tomuro leveraged this predictability in Tree-Cut and a Lexicon Based on Systematic Polysemy, using an information-theoretic "tree-cut" algorithm to automatically group fine-grained WordNet senses into broader categories. By capturing these regularities, the model eliminates micro-distinctions that are functionally irrelevant to downstream applications, reducing computational complexity without sacrificing meaningful semantic diversity.

3. Structural Reorganization

Sometimes, the solution is not to blindly merge senses, but to fix the underlying structural logic of the database. A common culprit behind over-granularity is specialization polysemy, where a word simultaneously refers to a general category and a specific member of that category.

Freihat et al. developed a semi-automatic method to identify these instances in WordNet. Instead of flattening the senses and losing information, they reorganized the hierarchy to make these implicit relationships explicit. This structural fix resolves algorithmic confusion while actually increasing the total structured knowledge within the database.


Conclusion

The shift toward coarse-grained sense representations works because it aligns the granularity of the sense inventory with the actual discriminative capacity of both human annotators and machine learning models. By collapsing hyper-specific definitions into broader semantic classes or supersenses, researchers eliminate the noisy, low-confidence boundaries that cause automated systems to fail. The result is a more interpretable, stable, and practical framework for machine understanding.

Key Foundational Work

Word Sense Disambiguation for Information Retrieval (1999) — contributed 3 concept elements to this direction.
A Method for Word Sense Disambiguation of Unrestricted Text (1999) — contributed 1 concept elements to this direction.
Sense Classification of Verbal Polysemy based-on Bilingual Class/Class Association (1996) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_semantic_granularity_mismatch_ca5ece94,
  title = {semantic granularity mismatch},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-granularity-mismatch/rdr_ca5ece94e32049e6a746b08866ccb2cf},
  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_semantic_granularity_mismatch_ca5ece94_v1,
  title = {semantic granularity mismatch},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/semantic-granularity-mismatch/rdr_ca5ece94e32049e6a746b08866ccb2cf/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake