In most Natural Language Processing (NLP) pipelines, a character is treated as an atomic unit—an arbitrary integer ID in a vocabulary. While this works for alphabetic scripts, it ignores the rich internal logic of logographic scripts like Chinese or Korean. A Chinese character is not a random arrangement of lines; it is a structured composition of radicals and strokes that carry semantic and phonetic clues.
Research into sub-character structural decomposition seeks to "break the atom" of the character. By decomposing logographic symbols into their constituent parts, models can learn the underlying morphology of the script, allowing them to understand rare words, correct spelling errors, and bridge historical linguistic gaps.
Radicals as Semantic Building Blocks
The most common level of decomposition is the radical. In Chinese, radicals often anchor the meaning of a character. For example, characters related to "wood" (timber, forest, pole) often share the "木" radical. Shi et al. argue that treating characters as indivisible units loses this internal information; by training "radical embeddings," models can capture granular semantic meanings that are otherwise invisible.
This structural logic is not unique to Chinese. Kim et al. demonstrate that Korean Hangeul can be decomposed into "atomic" components based on its generative principles, significantly improving a model's resilience to typos and out-of-vocabulary (OOV) words. Guo frames these radicals as functional equivalents to affixes in alphabetic languages, serving as the primary mechanism for categorization and meaning construction.
Fine-Grained Modeling with Stroke Sequences
While radicals provide semantic categories, they do not capture the full visual or morphological identity of a character. To reach a deeper level of granularity, researchers decompose characters into stroke sequences.
Cao and Zhou et al. propose using stroke $n$-grams to learn embeddings. This approach identifies semantic relationships between characters that share structural components even if they share no common radicals. By looking at the "DNA" of the character—the individual brush strokes—models can generalize meanings across a vast vocabulary.
This sequential view also enables cross-lingual transfer. Liu et al. introduced StrokeNet, which maps Chinese stroke sequences to Latin letters. This transformation allows Chinese text to be processed by standard sequence-to-sequence architectures designed for alphabetic languages, effectively breaking the "representation bottleneck" in neural machine translation.
Capturing Spatial and Hierarchical Layouts
Characters are more than just a "bag of strokes"; their meaning is often derived from the spatial arrangement of their components. A radical on the left might mean something different than the same radical on the bottom.
To capture this, Lyu et al. and Huang et al. utilize hierarchical structural trees. These trees represent not just the components, but their spatial relationships (e.g., left-right, top-bottom, or internal-external). Similarly, Wang et al. treat characters as graphs of strokes, using Graph Neural Networks to encode both the spatial layout and the temporal order of writing. This makes embeddings more "semantically aware" by explicitly modeling how a character is constructed.
Robustness and Visual Similarity
Sub-character decomposition is particularly powerful for tasks where visual appearance is critical, such as Chinese Spelling Check (CSC) and Optical Character Recognition (OCR).
Traditional models struggle when users intentionally perturb characters to bypass filters. Cui et al. found that analyzing sub-units—both visual radicals and phonetic components—is essential for detecting toxic content that uses visual or phonetic substitutions. For spelling correction, Ji et al. and Liu et al. show that decomposing symbols into structural description sequences allows for a more granular comparison of visual similarity than simple pixel-based methods.
By moving from atomic character IDs to structured decompositions, these models gain a "human-like" ability to recognize that two characters are related because they look similar or share a common root, even if the model has never seen that specific character before.
Key Foundational Work
International Conference on Computational Linguistics COLING 1969: Preprint No. 10 (1969) — contributed 1 concept elements to this direction.
Using Structural Information for Identifying Similar Chinese Characters (2008) — contributed 1 concept elements to this direction.
Structural Patterns of Chinese Characters (1969) — contributed 1 concept elements to this direction.