Selecting the Optimal Abstraction Level in Semantic Hierarchies

183 concepts141 papers

When organizing or learning from data, finding the right level of categorization is a delicate balancing act. Imagine trying to automatically label a cluster of words containing China, Japan, and Korea. Labeling them "East Asian Countries" is highly informative, whereas labeling them "Entities" or "Objects" is technically correct but far too vague to be useful. Conversely, labeling them "Sovereign states in East Asia with a combined population exceeding 1.6 billion" is overly narrow and fails to generalize to other similar groups.

This tension is what the authors of Knowledge-Based Automatic Topic Identification describe as finding a "Goldilocks" level of abstraction—identifying central concepts within a hierarchy that are neither too specific nor too general.

If a model selects concepts that are too specific, it suffers from severe data sparsity and fails to generalize to unseen instances, a challenge highlighted by Yuval Krymolowski et al. in Incorporating Knowledge in Natural Language Learning: A Case Study. Conversely, climbing too high up a taxonomic tree causes categories to lose their discriminative power. As demonstrated by Su Nam Kim et al. in Extending Sense Collocations in Interpreting Noun Compounds, model performance typically peaks at a moderate level of abstraction before sharply declining as distinct concepts collapse into generic entities.

The Fundamental Trade-Off

The core mechanism behind this research direction relies on balancing two competing forces: descriptive coverage (how many items a concept applies to) and specificity (how much unique information the concept conveys).

        [Entity]             <-- Too Broad (High Coverage, Low Specificity)
           |
    [Asian Country]          <-- "Goldilocks" Zone (Optimal Balance)
           |
 [East Asian Peninsula]      <-- Too Narrow (Low Coverage, High Specificity)

This trade-off manifests across a wide variety of domains:

Algorithmic Strategies for Finding the Middle Ground

To move away from manual tuning, researchers have developed mathematical and statistical frameworks to automatically pinpoint the optimal level of abstraction.

Information-Theoretic Compression (MDL)

One prominent strategy leverages the Minimum Description Length (MDL) principle. The intuition here is that the best abstraction level is the one that compresses the data most efficiently. The authors of On Conceptual Labeling of a Bag of Words utilize MDL to automatically balance descriptive coverage and conciseness when summarizing sets of items. Similarly, in Combining EM Training and the MDL Principle for an Automatic Verb Classification Incorporating Selectional Preferences, the authors combine Expectation-Maximization parameter learning with MDL to prune noun hierarchies, finding the exact level of abstraction that captures a verb's selectional preferences without overfitting to the training vocabulary.

Statistical Knee-Point Detection

Another approach treats taxonomic paths as curves where specificity and generalization compete. In Explainable Metaphor Identification Inspired by Conceptual Metaphor Theory, the authors apply a statistical knee-point algorithm to paths within a lexical hierarchy. This mathematically identifies the exact point where a concept transitions from being overly specific to overly broad, facilitating clear and explainable mappings between different semantic domains. This geometric optimization is further utilized in MetaPro Online: A Computational Metaphor Processing Online System to map metaphorical expressions to balanced cognitive domains.

Clustered Importance Sampling

In data selection for language models, David Grangier et al. demonstrate in Task-Adaptive Pretrained Language Models via Clustered-Importance Sampling that finding a balanced cluster resolution is critical. Broad clusters fail to capture domain-specific nuances required for specialist models, while overly fine clusters risk overfitting and memorizing specific examples rather than learning generalizable patterns.

Why It Works: Generalization Without Sparsity

Optimizing the level of abstraction works because real-world data is inherently hierarchical, but our data samples are finite. If a machine learning model treats every unique entity or word as an isolated feature, it runs into the curse of dimensionality and lexical sparsity. If it collapses everything into a single root category, it loses all predictive power.

By mapping sparse, concrete features to well-chosen intermediate concepts, models can generalize to entirely unseen data. As shown by Barbara Rosario et al. in Classifying the Semantic Relations in Noun Compounds via a Domain-Specific Lexical Hierarchy, an intermediate level of relation granularity allows a model to generalize across a diverse domain and successfully identify semantic relationships for technical terms it has never seen before. Finding this optimal abstraction level ensures that classification and representation systems remain both highly informative and statistically robust.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 136 papers
  2. Version 1 · 2026-10-04 · 161 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_optimal_abstraction_level_59178a50,
  title = {optimal abstraction level},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optimal-abstraction-level/rdr_59178a50fb6341dd99ec727897bff04e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_optimal_abstraction_level_59178a50_v2,
  title = {optimal abstraction level},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optimal-abstraction-level/rdr_59178a50fb6341dd99ec727897bff04e/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}

See something wrong? Report a mistake