Resolving Polysemy via Maximum Sense Similarity

23 concepts17 papers1 proposal

When humans evaluate the similarity between two words, we do so by dynamically isolating their most relevant meanings. If asked how similar "crane" is to "stork," we instantly think of birds. If asked how similar "crane" is to "hoist," we switch contexts to heavy machinery.

For computational models, however, words have traditionally been represented as monolithic entities—either as single nodes in a taxonomy or as static vectors in a distributional space. When a word is polysemous (possessing multiple meanings), a monolithic representation forces an unnatural averaging of its distinct senses. The vector for "crane" gets pulled simultaneously toward biology and construction, degrading its similarity scores with both "stork" and "hoist."

To solve this, a long-standing research direction shifts the comparison from the word level to the sense level. By breaking words down into their constituent meanings and evaluating the cross-product of those meanings, models can dynamically align the correct contexts.

The Core Framework: Pairwise Sense Evaluation

The intuition behind sense-level similarity is straightforward. Instead of treating a word $W$ as a single point, we treat it as a set of discrete senses $S = \{s_1, s_2, \dots, s_n\}$, which can be derived from dictionary definitions, ontological nodes (like WordNet synsets), or distinct embedding clusters.

To find the similarity between two words, $W_1$ and $W_2$, the system computes a matrix of pairwise similarity scores between every sense of the first word and every sense of the second word. The core challenge then becomes an aggregation problem: how do we compress this matrix of pairwise scores into a single metric that accurately reflects human judgment?

Word 1: "Bank" ─── Senses: [ Financial, Riverbank ]
                                │          │
                                ╳  (Pairwise Comparison)
                                │          │
Word 2: "River" ── Senses: [ Stream, Reservoir ]

The Maximum Sense Heuristic

The most dominant and intuitive strategy for aggregating sense pairs is the maximum sense similarity heuristic. In a foundational paper on taxonomy-based metrics, Philip Resnik in Using Information Content to Evaluate Semantic Similarity in a Taxonomy established that the similarity between two polysemous terms can be effectively modeled by simply selecting the maximum similarity score found among all possible pairings of their respective senses.

$$\text{Sim}(W_1, W_2) = \max_{i, j} \text{Sim}(s_{1i}, s_{2j})$$

This maximum-selection strategy acts as an implicit, greedy form of joint sense disambiguation. In the absence of explicit sentence context, the model assumes that the most relevant meaning of a word is the one that shares the strongest semantic connection with the paired word.

This heuristic has proven remarkably robust across different eras of natural language processing:

Applications in Relational and Contextual Tasks

Because the maximum sense heuristic effectively bypasses the need for a separate, explicit word sense disambiguation step, it has been widely deployed to resolve ambiguity in complex relational tasks.

Hypernymy Detection

Detecting "is-a" relationships often fails when words have rare or domain-specific meanings. To circumvent this, Wenpeng Yin 0001 et al. developed an approach in Term Definitions Help Hypernymy Detection that cross-references dictionary definitions. By calculating a relationship score for every possible sense-pair combination and selecting the highest probability outcome, the system integrates disambiguation directly into the classification task.

Thematic Fit Evaluation

Standard models struggle to determine if a noun fits a verb's semantic role (e.g., judging the plausibility of "mash banana" vs. "mash milk") because they rely on a single averaged prototype for the verb. Clayton Greenberg et al. resolved this in Improving unsupervised vector-space thematic fit evaluation via role-filler prototype clustering by clustering role-fillers into distinct sense-clusters. Evaluating a candidate noun against these clusters and selecting the maximum similarity score ensures that a valid fit in one specific context is not penalized by a lack of fit in unrelated senses.

Pun Localization

Humor often relies on homographic puns—words spelled the same way but carrying dual meanings within the same sentence. In Fermi at SemEval-2017 Task 7: Detection and Interpretation of Homographic puns in English Language, Vijayasaradhi Indurthi leveraged maximum semantic similarity between sentence word pairs. The system operates on the intuition that a punning word will exhibit a highly specific, maximum-similarity link to an explicit "anchor" word elsewhere in the sentence context, allowing the system to pinpoint the exact location of the pun.

Beyond the Max: Weighted and Average Profiles

While the greedy maximum heuristic is highly effective, it operates on an assumption that secondary meanings have zero influence on human semantic judgment. However, cognitive evidence suggests that humans experience a semantic dominance bias, where the relative frequency and familiarity of all senses of a word subtly color our interpretation.

To capture this nuance, an alternative branch of research explores softer aggregation strategies that weight the entire sense profile:

Aggregation Strategies:
├── Greedy Max:   [ Score 1, Score 2, Score 3 ] ──> Select Highest Only
└── Weighted Sum: [ Score 1, Score 2, Score 3 ] ──> Adjust by Sense Dominance & Frequency

For instance, to detect metaphorical verbs without explicit context, Eirini Florou in Detecting Metaphor by Contextual Analogy calculated an average similarity score across all potential senses. This distributed approach provides robustness against extreme polysemy by ensuring the model maintains a holistic view of the word's semantic neighborhood.

Taking this further, Yangyang Wu et al. introduced an adaptive weighting scheme in Chinese-English Bilingual Word Semantic Similarity Based on Chinese WordNet. Instead of picking a single maximum, their model computes a weighted sum across multiple sense pairs, dynamically adjusting the weights based on the total number of possible combinations to better mirror human common sense.

Similarly, the authors of SensEmbed: Learning Sense Embeddings for Word and Relational Similarity, Ignacio Iacobacci et al., demonstrated that while the maximum similarity between any two senses is a powerful baseline, human judgments are best modeled by aggregating scores using a weighted average that explicitly accounts for the relative dominance and frequency of each individual sense.

Topological and Cross-Lingual Variations

The mechanism of sense selection also adapts to the underlying geometry of the semantic space. The choice of whether to use a strict maximum or a broader neighborhood aggregation often depends on the nature of the concepts being compared.

In Semantic Similarity Computation for Abstract and Concrete Nouns Using Network-based Distributional Semantic Models, Elias Iosif et al. discovered a fundamental split between noun types: concrete nouns (like car) are best modeled using the maximum sense approach, whereas abstract nouns (like justice) are better captured through attributional similarity aggregated across their entire semantic neighborhoods.

To formalize these neighborhoods geometrically, Georgia Athanasopoulou et al. introduced Low-Dimensional Manifold Distributional Semantic Models in Low-Dimensional Manifold Distributional Semantic Models. By treating the global lexical space as a collection of overlapping, low-dimensional manifolds, their system implements a sense-selection mechanism by finding the minimum distance (maximum similarity) across shared sub-manifolds, effectively isolating local semantic contexts.

Finally, this paradigm serves as a powerful bridge for cross-lingual applications. Heyan Huang et al. demonstrated in An English-Chinese Cross-lingual Word Semantic Similarity Measure Exploring Attributes and Relations that by decomposing bilingual words into language-independent "sememes" (fundamental units of meaning), a system can evaluate all cross-lingual sense combinations. Selecting the maximum similarity score among these pairings successfully identifies the intended meaning across language barriers, proving that the intuition of maximum sense alignment holds true regardless of the underlying language.

Key Foundational Work

Using Information Content to Evaluate Semantic Similarity in a Taxonomy (1995) — contributed 1 concept elements to this direction.
An Automatic Clustering of Articles Using Dictionary Definitions (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_maximum_sense_similarity_fe544210,
  title = {maximum sense similarity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/maximum-sense-similarity/rdr_fe544210382341a1977244223b8170b0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_maximum_sense_similarity_fe544210_v1,
  title = {maximum sense similarity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/maximum-sense-similarity/rdr_fe544210382341a1977244223b8170b0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake