Coarse Grained Sense Inventories for Semantic Disambiguation

132 concepts94 papers1 proposal

Standard lexical databases like WordNet partition human language into hyper-specific, fine-grained definitions. While linguistically precise, this extreme granularity introduces a paradox: it harms both human annotation and machine learning performance. When definitions are separated by highly subtle, almost metaphysical boundaries, even human experts struggle to reach a consensus.

For computational models, this fragmentation results in severe data sparsity. If a word's usage is split across dozens of distinct synsets, the number of training examples per class plummets, making it nearly impossible for supervised learners to generalize. As Ng et al. demonstrate, fine-grained sense inventories impose a strict performance ceiling on Word Sense Disambiguation (WSD) systems.

To resolve this, a major research direction focuses on coarse-grained sense inventories—frameworks that collapse hyper-specific dictionary definitions into broader, functionally distinct semantic categories. This approach acts as a crucial middle ground between the narrow scope of Named Entity Recognition (NER) and the intractable complexity of full-scale fine-grained WSD.

Finding the Semantic Middle Ground

To build a coarse-grained inventory, researchers must find a way to group related meanings without losing the essential distinctions that prevent ambiguity. Several mechanisms exist to define these broader categories.

One prominent method is "supersense tagging," introduced by Ciaramita et al., which maps words to a fixed set of broad lexicographer classes (such as noun.food or verb.motion). Schneider et al. successfully applied this lightweight annotation strategy to Arabic, demonstrating that broad groupings provide extensive semantic coverage while remaining highly manageable for human annotators.

Alternatively, abstractions can be derived structurally or statistically:

  • Systematic Polysemy: Tomuro developed an automated "tree-cut" algorithm that leverages predictable patterns where words share related meanings—such as an animal and its meat (e.g., chicken)—to group fine-grained senses into robust conceptual boundaries.
  • Basic Level Concepts (BLCs): Izquierdo et al. utilized BLCs, which are automatically derived mid-level abstractions from WordNet, to provide a stable semantic "middle ground" that mitigates data sparseness.
Fine-Grained Senses (WordNet)       Mid-Level Abstraction             Coarse Category
[bank.n.01] (financial institution) ───► [Financial Entity] ─────────► FINANCE
[bank.n.02] (river bank)            ───► [Geological Feature] ───────► GEOGRAPHY
[bank.n.03] (building/container)    ───► [Financial Entity] ─────────► FINANCE

Why Coarse Granularity Works

Reducing semantic resolution improves model intelligence because of how data density and structural noise interact in machine learning models.

First, grouping fine-grained definitions into broader classes dramatically increases the number of training examples available per label. As shown by Izquierdo et al., class-based WSD allows models to achieve peak performance using only a fraction of the annotated data typically required. The model learns general patterns across related concepts rather than treating every subtle variation as an isolated class.

Second, it bypasses the inherent noise of dictionary definitions. Many fine-grained distinctions are artifacts of lexicography rather than functional differences in natural language. By treating WSD as a task of discriminating between semantic clusters rather than individual definitions—an approach formalized in the SANDWiCH framework by Guzman-Olivares et al.—models become significantly more resilient to the long-tail problem of rare word senses. This structural simplification allows frameworks like the Coarse Sense Inventory (CSI) to shatter traditional performance barriers, achieving F1 scores well over 85%.

Alignment with Pre-trained Language Models

The benefits of coarse-grained inventories extend naturally to modern deep learning architectures and Pre-trained Language Models (PLMs). When a model is pre-trained on massive text corpora, it inherently captures broad topical and contextual associations rather than hyper-specific dictionary mappings.

Sainz et al. demonstrated that reframing WSD as a zero-shot textual entailment task using broad domain labels aligns perfectly with the latent knowledge embedded within PLMs. Because coarse-grained categories reduce the complexity of the disambiguation space, language models can distinguish fundamentally different semantic concepts (such as homonyms) with remarkable accuracy without any task-specific fine-tuning, as verified by Tedeschi et al..

The Granularity Trade-off

Adopting a coarse-grained inventory involves a deliberate trade-off: it sacrifices the ability to capture highly specific semantic nuances in exchange for robust, scalable, and learnable classification. For instance, Mohammad et al. note that while using coarse thesaurus categories filters out statistical noise and provides a stable representational space, fine-grained distinctions are inevitably lost.

However, for the vast majority of downstream applications, this trade-off is highly favorable. Coarse-grained abstraction provides the exact level of semantic resolution that machines need to understand human language reliably, whether they are:

  • Improving translation selection by helping models distinguish polysemous words (Xiong et al.),
  • Structuring noisy, user-generated medical forum data (Gupta et al.), or
  • Simplifying sentiment analysis by grouping senses into functional categories like subjective or objective (Akkaya et al.).

Key Foundational Work

Word-Sense Disambiguation Using Statistical Models of Roget’s Categories Trained on Large Corpora (1992) — contributed 2 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_coarse_grained_sense_inventory_6e3004ad,
  title = {coarse-grained sense inventory},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/coarse-grained-sense-inventory/rdr_6e3004adba75479a992bdd2c20327852},
  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_sense_inventory_6e3004ad_v1,
  title = {coarse-grained sense inventory},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/coarse-grained-sense-inventory/rdr_6e3004adba75479a992bdd2c20327852/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