Disambiguating Polysemous Prepositions and Functional Markers

29 concepts28 papers1 proposal

Consider the word with in the following sentences:

  1. He cut the bread with a knife.
  2. He ate the bread with a friend.
  3. He ate the bread with enthusiasm.

To a human, the distinction is instantaneous. In the first sentence, with introduces an instrument; in the second, an association (comitative); in the third, a manner. Yet the word with itself never changes. It carries virtually no independent lexical meaning. Instead, it acts as a grammatical chameleon, projecting a relational bridge between a governor (the verb) and a dependent (the noun phrase).

This phenomenon—prepositional and functional word polysemy—presents a profound challenge for Natural Language Processing (NLP). Because functional words are among the most frequent tokens in any language and can signal dozens of distinct semantic relations, resolving their precise meaning is essential for downstream tasks like machine translation, semantic role labeling, and knowledge base construction.

The Early Challenge of Functional Ambiguity

The realization that functional words complicate automated inference dates back to the foundational eras of AI. Early systems struggled when encountering words whose meanings were entirely context-dependent. Granger, in designing the FOUL-UP system, noted that highly polysemous functional words make it exceptionally difficult to determine the precise semantic link between different parts of a sentence without structured templates of world knowledge (FOUL-UP: A Program that Figures Out Meanings of Words from Context).

To transform static text into structured knowledge, early computational pipelines built dedicated modules to parse these relational markers. For example, Ravin developed a "Disambiguation Module" specifically to interpret dictionary definitions containing the phrase "to VERB with NP," categorizing the elusive preposition into five distinct semantic relations to map how verbs and objects relate (Disambiguating and Interpreting Verb Definitions).

Cross-Linguistic Manifestations: Particles and Case Markers

This relational ambiguity is not unique to English prepositions. Across different language families, the same underlying problem manifests as polysemous case markers, adpositions, and grammatical particles.

In Japanese, the particle no serves as a highly versatile marker connecting two nouns ($N_1$ no $N_2$). As shown by Kurohashi et al., no can denote possession, location, composition, or complex functional roles, requiring external lexical resources (like dictionary definitions of the nouns) to decode the underlying relationship. Furthermore, predicting core Japanese case particles like ga, wo, and ni is notoriously difficult because a single particle can denote multiple distinct semantic roles such as time, location, or indirect objects (Suzuki et al.).

Similarly, in context-heavy languages like Chinese, multi-functional words frequently create ambiguity by shifting between temporal indicators and expressions of desire or necessity (Ge et al.). Whether analyzing an English preposition or a Hindi case marker, the core computational hurdle remains identical: a single linguistic token mediates vastly different conceptual relationships depending on its environment.

From Feature Engineering to Geometric Subspaces

How do modern computational models resolve this ambiguity? The evolution of Preposition Sense Disambiguation (PSD) strategies reflects a broader shift in NLP from manual feature engineering to geometric representation learning.

Local Context and Joint Modeling

Early statistical approaches attempted to isolate the exact linguistic features that drive performance. Ye et al. demonstrated that local word collocations and synonym chains from resources like WordNet perform the heavy lifting in distinguishing fine-grained preposition senses. To capture broader structural context, Hovy et al. introduced linguistically motivated features derived from phrase structures, significantly outperforming simple bag-of-words baselines.

A key insight emerged when researchers stopped treating preposition disambiguation as an isolated task. Because a preposition's sense is fundamentally tied to the role the entire phrase plays in a clause, Dahlmeier et al. proposed a joint probabilistic model. By simultaneously optimizing Word Sense Disambiguation (WSD) for the preposition and Semantic Role Labeling (SRL) for the prepositional phrase, the authors proved that these two tasks are mutually informative, yielding substantial performance gains when solved together.

Unsupervised and Cross-Lingual Supervision

A persistent bottleneck in functional word disambiguation is the scarcity of human-annotated data. To bypass this, Hovy et al. pioneered unsupervised PSD by leveraging sparsity-inducing training techniques, such as MAP-EM with an $L_0$ norm, to discover distinct senses without explicit labels.

Alternatively, Gonen et al. looked to machine translation for a "free" source of supervision. Their core intuition was that while an English preposition like at or by is highly ambiguous, it often translates into entirely different words in target languages depending on its meaning (e.g., spatial vs. temporal). By pre-training a neural network to predict these cross-lingual translations, the model naturally learns a highly nuanced, context-aware representation of the preposition's sense.

The Geometry of Interplay

With the rise of continuous word representations, researchers began exploring how the interplay between words is encoded geometrically. Gong et al. demonstrated that the relationship between the words preceding and succeeding a preposition can be captured as a geometric subspace within word embeddings. By focusing on the mathematical interaction between the governor and the dependent, this approach completely bypasses the need for complex, error-prone dependency parsers, achieving state-of-the-art disambiguation purely through embedding geometry.

Formalizing Meaning: The SNACS Framework

To move beyond ad-hoc sense inventories, the field required a unified, cross-linguistic framework capable of formalizing functional polysemy. This led to the development of the Semantic Network of Adposition and Case Supersenses (SNACS) framework.

Introduced by Hwang et al., the breakthrough of SNACS lies in its dual-layer annotation strategy. It explicitly separates two distinct facets of a functional word's meaning:

  • Lexical Function: The inherent, dictionary-like semantic constraint of the word itself.
  • Scene Role: The actual semantic role the word fulfills within the specific context of the event or scene.

For example, in the phrase "he is working for a paycheck," the lexical function of for implies a beneficiary, but the scene role in this context is a commodity or reward. By disentangling these two layers, SNACS allows models to capture both the grammatical constraints of the language and the universal semantic reality of the scene.

Because SNACS maps language-specific grammar to universal semantic categories, it has become the standard for cross-lingual functional analysis. The framework has been successfully adapted to model the extreme polysemy of Hindi case markers and adpositions (Arora et al.), as well as Japanese particles (joshi), providing a robust computational bridge for comparing how different languages encode structural meaning (Taguchi et al.).

Key Foundational Work

FOUL-UP: A Program that Figures Out Meanings of Words from Context (1977) — contributed 1 concept elements to this direction.
Disambiguating and Interpreting Verb Definitions (1990) — 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_prepositional_polysemy_2b16d596,
  title = {prepositional polysemy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/prepositional-polysemy/rdr_2b16d596d24c4b47ad04974d9fe73538},
  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_prepositional_polysemy_2b16d596_v1,
  title = {prepositional polysemy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/prepositional-polysemy/rdr_2b16d596d24c4b47ad04974d9fe73538/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