Definition Based Semantic Similarity

72 concepts61 papers1 proposal

How do we know that "automobile" and "car" are deeply related concepts? For humans, the answer lies in their definitions: both describe a wheeled motor vehicle used for transportation. A dictionary definition, or gloss, acts as a dense, human-curated summary of a concept's semantic footprint.

In natural language processing, exact string matching breaks down when vocabulary diverges, while deep neural embeddings can be data-hungry and difficult to interpret. Definition-based semantic similarity offers an elegant alternative: it leverages the text of dictionary definitions themselves to compute conceptual proximity. The core intuition is that if two words or senses share a underlying meaning, the vocabulary used to define them will naturally overlap.

Quantifying Meaning via Gloss Overlap

The most direct way to implement this intuition is to measure the literal overlap of content words between different definitions. When we strip away common stopwords and lemmatize the remaining words to their roots, the overlapping vocabulary isolates the core semantic attributes of a term.

"Automobile" gloss: A wheeled motor vehicle used for transportation.
                     |       |     |       |        |
"Car"        gloss: A wheeled vehicle powered by a motor.

This mechanism is highly effective for aligning disparate linguistic resources. For example, Daniel M. Bikel demonstrates in Automatic WordNet Mapping Using Word Sense Disambiguation that calculating the overlap of content words between source dictionary definitions and target ontology glosses allows for the automatic mapping of WordNets across different languages.

While traditional information retrieval long favored rigid structural hierarchies (such as hypernymy trees), Hui Fang shows in A Re-examination of Query Expansion Using Lexical Resources that definition-based overlap provides a significantly richer semantic signal for query expansion. Descriptive text captures contextual nuances that categorical links in a database omit.

This overlap strategy can also be adapted for real-time evaluation. In automated grading systems, Nihal V. Nayak et al. show in V for Vocab: An Intelligent Flashcard Application how calculating a synonym-overlap ratio between a student's short answer and a reference definition enables accurate, conceptual grading that ignores surface-level phrasing variations.

Resolving Ambiguity and Homographs

A major challenge in lexical semantics is polysemy—words that share a spelling but have entirely distinct meanings (e.g., "bank" as a financial institution versus a riverbank). Definition-based similarity provides a clean mechanism for Word Sense Disambiguation (WSD) by treating each distinct definition as a unique coordinate in semantic space.

Instead of treating a word as a single ambiguous entity, the authors of these techniques represent senses via meta-features derived from definition comparisons. In Dictionary Definitions based Homograph Identification using a Generative Hierarchical Model, Anagha Kulkarni et al. utilize cohesiveness scores and overlap counts across multiple online dictionaries to isolate homographs. To prevent nearly identical definitions from skewing the model, they employ a post-pruning maximum similarity technique that ignores redundant entries and focuses on informative secondary matches.

Target Word: "Bank"
  |- Sense 1 Gloss: "A financial institution..."  --> High overlap with "money", "deposit"
  |- Sense 2 Gloss: "The slope beside a river..." --> High overlap with "water", "earth"

In unsupervised environments where labeled training data is unavailable, definition similarity can anchor words to a standardized sense inventory. As Wanxiang Che et al. demonstrate in HIT-CIR: An Unsupervised WSD System Based on Domain Most Frequent Sense Estimation, a system can estimate the most frequent sense of a word within a specific domain by measuring the overlap between the definitions of a target word's senses and the definitions of its contextual neighbor words.

Capturing Latent Attributes and Soft Matching

Because definitions describe the attributes and behaviors of a concept, analyzing gloss text allows models to extract latent semantic dimensions, such as sentiment or subjectivity. For instance, Yaw Gyamfi et al. show in Integrating Knowledge for Subjectivity Sense Labeling that comparing the gloss of an unknown target sense against the glosses of pre-established subjective and objective seed sets reveals fine-grained emotional nuances that structural database links completely miss.

Finally, these definition-based distances can be injected back into traditional string-matching algorithms to make them robust to vocabulary variation. This creates a "soft matching" capability:

By turning human-readable definitions into structured mathematical comparisons, definition-based semantic similarity provides an interpretable, robust framework for understanding text without relying on massive, opaque corpus training.

Key Foundational Work

Lexical Disambiguation using Simulated Annealing (1992) — contributed 1 concept elements to this direction.
Lexical Disambiguation Using Constraint Handling In Prolog (CHIP) (1993) — 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_definition_based_semantic_similarity_ed425086,
  title = {definition-based semantic similarity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/definition-based-semantic-similarity/rdr_ed425086a9404f26ab001a4895e9377c},
  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_definition_based_semantic_similarity_ed425086_v1,
  title = {definition-based semantic similarity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/definition-based-semantic-similarity/rdr_ed425086a9404f26ab001a4895e9377c/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