In alphabetic languages, words are linear sequences of letters. In logographic languages like Chinese and Japanese, characters are two-dimensional compositions of smaller semantic units called radicals or sub-characters.
If a machine learning model treats each character as an isolated, atomic token, it misses out on the rich internal semantics shared across related characters. However, if the model simply breaks a character down into a flat list or a "bag" of its constituent radicals, it encounters a fundamental flaw: spatial arrangement dictates meaning.
In logographic scripts, the exact same set of sub-components can form entirely different characters depending on how they are positioned. To build accurate representations, models must explicitly encode the 2D spatial layout and hierarchical topology of these components.
The Failure of Flat Decompositions
Early sub-character models attempted to capture internal character structure by decomposing tokens into sequences of radicals. However, as noted by Shi et al. in Radical Embedding: Delving Deeper to Chinese Radicals and Yu et al. in Joint Embeddings of Chinese Words, Characters, and Fine-grained Subcharacter Components, this flattening process discards critical geometric relationships.
Consider two characters that share the exact same sub-components. If one character arranges them horizontally (left-to-right) and the other stacks them vertically (top-to-bottom), they represent completely different words with distinct meanings. A model that treats components as an unordered set or a simple linear sequence becomes blind to this distinction, leading to semantic ambiguity.
Similarly, in Japanese kanji modeling, Nguyen et al. observed in Sub-character Neural Language Modelling in Japanese that flattening the internal hierarchy of characters strips away valuable structural information, limiting the language model's ability to leverage sub-character compositionality effectively.
Preserving Geometry via Structural Operators
To preserve 2D geometry without relying purely on raw pixel inputs, recent frameworks introduce explicit spatial layout features into the encoding process.
One elegant solution is the use of structural operators. Lyu et al. demonstrated in Glyph Enhanced Chinese Character Pre-Training for Lexical Sememe Prediction that interleaving spatial operators (such as tokens indicating "vertical stacking" or "horizontal alignment") directly into the radical sequence allows models to distinguish between characters with identical components. The spatial layout acts as a primary feature, providing the geometric context needed to resolve structural anagrams.
Flat Sequence (Ambiguous): [Radical A] + [Radical B]
Layout-Aware Sequence: [Vertical Stack] + [Radical A] + [Radical B]
Taking this a step further, Zhang et al. developed a framework in HierCode: A Lightweight Hierarchical Codebook for Zero-shot Chinese Text Recognition that explicitly maps out layouts and radicals as distinct, multi-hot encoded features. By treating the spatial configuration as an independent coordinate of the character's identity, the model retains the discriminative structural information necessary for high-fidelity text recognition.
Modeling Characters as Formation Trees
Logographic characters are not just flat grids; they are often hierarchically nested. A compound character might be composed of two sides, where the left side is a single radical and the right side is itself a complex sub-character made of two vertically stacked components.
To capture this recursive nature, Hong et al. introduced "formation trees" in Improving Chinese character representation with formation tree. Instead of a sequence, a character is modeled as a tree where internal nodes represent spatial operations (e.g., left-to-right, surround) and leaf nodes represent basic radicals. This tree-like compositionality allows encoders to learn the systematic, hierarchical logic of how characters are built, mirroring the actual historical and cognitive rules of the script.
Why Spatial Awareness Enables Zero-Shot Generalization
The ultimate value of spatial layout encoding becomes clear when dealing with rare, low-frequency, or entirely unseen characters—a notorious "long-tail" problem in Optical Character Recognition (OCR) and Natural Language Processing (NLP).
When a model encounters a character it has never seen during training, a standard token-based model fails completely. However, if the model understands the finite vocabulary of radicals and the geometric rules of how they combine, it can reason about the unseen token's identity and meaning.
As Diao et al. showed in RZCR: Zero-shot Character Recognition via Radical-based Reasoning, combining visual feature extraction with a structural layout reasoning engine allows a system to accurately recognize rare historical characters by decoding their spatial composition.
Furthermore, this structural awareness directly benefits downstream semantic tasks. Zheng et al. demonstrated in AHVE-CNER: Aligned Hanzi Visual Encoding Enhance Chinese Named Entity Recognition with Multi-Information that explicitly aligning visual character images with their internal radical layout structures significantly boosts performance in Named Entity Recognition (NER). By mastering the spatial topology inside the character, the model gains a finer, more robust grasp of the language's underlying semantics.
Key Foundational Work
Joint Embeddings of Chinese Words, Characters, and Fine-grained Subcharacter Components (2017) — contributed 3 concept elements to this direction.
Radical Embedding: Delving Deeper to Chinese Radicals (2015) — contributed 1 concept elements to this direction.
Sub-character Neural Language Modelling in Japanese (2017) — contributed 1 concept elements to this direction.