Asymmetric Semantic Composition in Nominal Compounds

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 226 papers32 conceptsfrozen 2026-10-04

When we encounter a phrase like "apple juice," our brains perform a swift, invisible act of logic. We know it is a type of juice (the category) and that it is made from apples (the specification). In linguistics, these two-word constructions are known as nominal compounds, consisting of a head (the primary noun) and a modifier.

While simple vector addition—adding the embedding of "apple" to "juice"—is a common baseline in NLP, research shows that semantic composition is rarely a symmetric 50/50 split. Instead, the meaning of a compound is built through an asymmetric hierarchy where the head and modifier play fundamentally different roles.

The Head Noun as a Semantic Anchor

The most consistent finding across decades of research is that the head noun (usually the final word in English compounds) acts as the structural and semantic anchor. Barbara Rosario and Marti A. Hearst demonstrated this in the medical domain, finding that the head noun's category is the most critical factor in determining the implicit logic of a compound—for instance, whether a "treatment" is defined by its method or its target (Classifying the Semantic Relations in Noun Compounds via a Domain-Specific Lexical Hierarchy).

This "anchoring" effect is so strong that models often generalize better when the head noun remains constant. Giulia Rambelli and colleagues found that Large Language Models (LLMs) are more adept at interpreting novel compounds if they have already seen the head noun in other contexts, suggesting the head provides a stable template for possible relationships (Can Large Language Models Interpret Noun-Noun Compounds?). Similarly, Gabriel Stanovsky showed that in proper noun compounds like "Oxford vaccine," the common noun "vaccine" dictates the set of possible semantic predicates (Semantic Interpretation of Proper Noun Compounds).

The Modifier as the Driver of Transparency

If the head provides the category, the modifier often determines the "flavor" and "literality" of the phrase. Paradoxically, while the head is the structural anchor, humans often perceive the modifier as the more influential component when judging how "literal" or "related" a phrase is.

Shima Asaadi and Saif M. Mohammad discovered that the semantic relatedness of a bigram is often driven more by the modifier; humans perceive "apple pie" as more closely related to the attributes of an "apple" than to the general category of "pie" (Big BiRD: A Large, Fine-Grained, Bigram Relatedness Dataset). This is echoed in work by Stefan Müller and Stephen Roller, who found that the semantic transparency of the modifier is a significantly stronger predictor of whether a German compound is literal or idiomatic (Exploring Vector Space Models to Predict the Compositionality of German Noun-Noun Compounds).

Modeling Asymmetric Weighting

Because head and modifier contribute differently, simple additive models (where $Compound = Word_1 + Word_2$) often fail to capture the nuances of language. Researchers have moved toward asymmetric constituent weighting to better reflect this reality.

Corina Dima introduced a "masking" approach where words learn different transformation filters depending on whether they are acting as a modifier or a head (Reverse-engineering Language: A Study on the Semantic Compositionality of German Compounds). This acknowledges that the same word (e.g., "fire") contributes different semantic features when it is the head ("forest fire") versus the modifier ("fire drill").

Modern transformer models like Mistral and GPT-4 appear to have internalized these asymmetric transformations. Aishik Nagar and colleagues found that these models apply distinct mathematical transformations to heads and modifiers, allowing them to reconstruct compound meanings with high precision (How do Transformer Embeddings Represent Compositions?).

Why Asymmetry Matters for Idioms

Understanding this asymmetry is crucial for identifying when a phrase stops being literal and becomes an idiom. When a phrase like "glass ceiling" evolves, the semantic shift is not distributed evenly.

Michaela Socolof and Jackie Chi Kit Cheung found that in non-literal expressions, the dependent words (modifiers) typically stray much further from their canonical meanings than the head words do (Characterizing Idioms: Conventionality and Contingency). This suggests that idioms are "anchored" by a relatively stable head noun, while the modifier undergoes a radical semantic transformation. Filip Miletic confirmed this diachronically, showing that the semantic evolution of the modifier is a far stronger predictor of a phrase's eventual idiomaticity than the evolution of the head (Modeling the Evolution of English Noun Compounds).

Structural Biases in Neural Models

Despite the progress in LLMs, neural architectures often exhibit "structural biases" that conflict with human linguistic intuition. Julia Rozanova and Andre Freitas observed a "mean-pooling bias" where models over-focus on modifiers, causing the base concept to lose its identity when qualified by adjectives (Montague semantics and modifier consistency measurement).

Furthermore, Emmy Liu and Graham Neubig found that models often prioritize the "right-hand child" in a syntax tree simply due to sequence order, rather than a balanced semantic understanding (Are representations built from the ground up?). These findings suggest that while models are "doing the math" of composition, they may still lack the nuanced, asymmetric "anchoring" that characterizes human conceptual reasoning.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 32 papers
  2. Version 1 · 2026-10-04 · 26 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_head_noun_semantic_dominance_56b6bf73,
  title = {head-noun semantic dominance},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/head-noun-semantic-dominance/rdr_56b6bf73ca4241b58183dc6b72d5d942},
  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_asymmetric_semantic_composition_in_nomin_56b6bf73_v1,
  title = {Asymmetric Semantic Composition in Nominal Compounds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/asymmetric-semantic-composition-in-nominal-compounds/rdr_56b6bf73ca4241b58183dc6b72d5d942/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}