Decomposing Word Meaning into Atomic Semantic Primitives

This direction is split into other directions: atomic semantic decomposition. Its last version stays readable and citable here.

Version 1 of 138 papers46 conceptsfrozen 2026-10-04

In traditional Natural Language Processing, words are often treated as the fundamental building blocks of meaning. Whether represented as discrete symbols or high-dimensional vectors, a word like "apple" is typically handled as an indivisible unit. However, this "word-as-atom" approach struggles with the inherent fluidity of language: words can have multiple meanings (polysemy), rare words lack sufficient data to learn good representations (sparsity), and models often fail to understand how simple concepts combine into complex ones (compositionality).

A long-standing research direction proposes a different perspective: words are not atoms, but molecules. Just as matter is composed of a finite set of chemical elements, lexical meaning can be decomposed into a structured set of indivisible units called sememes. By shifting the focus from the word to these atomic semantic primitives, researchers have developed more robust, interpretable, and computationally efficient ways for machines to "understand" language.

The Intuition of Semantic DNA

The core idea, pioneered in early work like Yorick Wilks’s Preference Semantics, is that meaning is built from a small set of universal primitives Understanding Without Proofs. For example, the word "man" might be decomposed into sememes like human, male, and adult.

This decomposition provides a "semantic DNA" for every word in a language. As Zhiyuan Liu, Maosong Sun, and Yankai Lin argue, representing language through a finite set of these units creates a powerful regularization bottleneck Sememe Knowledge Representation. Instead of a model having to learn the meaning of 100,000 distinct words from scratch, it only needs to learn the behavior of a few thousand sememes and how they combine.

Solving the Sparsity and Polysemy Problems

One of the most immediate benefits of this approach is handling rare or unseen words. In standard models, a rare word has a "noisy" vector because it appears in few contexts. However, Yunfang Wu and Xueqiang Lv demonstrate that by aggregating the embeddings of a word's constituent sememes, we can construct high-quality representations even for words the model has never seen before Improving Word Vector with Prior Knowledge in Semantic Dictionary. Because the sememes themselves (like food or tool) appear frequently across the entire corpus, they provide a stable foundation for generalizing meaning.

Similarly, sememes offer a surgical tool for resolving polysemy. A word like "bank" is ambiguous, but its sememe sets are distinct: one set contains institution and money, while the other contains land and water. By using attention mechanisms to select the most relevant sememes based on context, researchers like Ruobing Xie and Zhiyuan Liu have significantly improved word similarity and analogy tasks Improved Word Representation Learning with Sememes.

From Symbolic Knowledge to Neural Architectures

Much of this research relies on HowNet, a specialized lexical knowledge base that manually annotates over 100,000 word senses with sememes. To make this accessible for modern AI, Chenghao Yang and Fanchao Qi released OpenHowNet, allowing neural models to plug into this structured semantic alphabet.

Recent work has moved beyond simply using sememes as external features, instead embedding the logic of decomposition directly into model architectures. Chen Huang and colleagues developed the Compositionality-Aware Transformer (CAT), which forces the model to decompose contextual representations into discrete primitives and then recompose them Towards Equipping Transformer with the Ability of Systematic Compositionality. This "structured bottleneck" helps models understand novel word combinations—like distinguishing "water body" from "body water"—which standard Transformers often conflate.

Explainability and Logical Consistency

Finally, semantic decomposition provides a path toward more interpretable AI. When a model calculates the similarity between two sentences, it is often a "black box." However, by decomposing embeddings into explainable features like semantic roles or negation, as explored by Anette Frank and Juri Opitz, we can see exactly which facets of meaning (e.g., the actor vs. the action) contribute to the similarity score SBERT studies Meaning Representations.

This granular view also allows models to check for "semantic rationality." By evaluating whether the sememes of a subject and a verb are logically compatible (e.g., can the object stone perform the action eat?), models like the one proposed by Xu Sun can detect subtle nonsense that grammatically correct but semantically flawed sentences often contain Evaluating Semantic Rationality of a Sentence.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_decomposing_word_meaning_into_atomic_sem_af164b7b,
  title = {Decomposing Word Meaning into Atomic Semantic Primitives},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decomposing-word-meaning-into-atomic-semantic-primitives/rdr_af164b7b0454479dbfe8b5ffbbcf3951},
  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_decomposing_word_meaning_into_atomic_sem_af164b7b_v1,
  title = {Decomposing Word Meaning into Atomic Semantic Primitives},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decomposing-word-meaning-into-atomic-semantic-primitives/rdr_af164b7b0454479dbfe8b5ffbbcf3951/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}