CoLa: Chinese Character Decomposition with Compositional Latent Components

TL;DR

CoLa is a novel deep latent variable model that learns to decompose Chinese characters into compositional components without predefined schemes, enabling effective zero-shot recognition and cross-dataset generalization.

Contribution

It introduces a learning-to-learn approach for Chinese character decomposition, surpassing prior methods that relied on human-defined schemes, thus improving zero-shot recognition capabilities.

Findings

Outperforms previous methods in zero-shot CCR.

Learned components reflect character structure interpretably.

Generalizes to historical oracle bone characters.

Abstract

Humans can decompose Chinese characters into compositional components and recombine them to recognize unseen characters. This reflects two cognitive principles: Compositionality, the idea that complex concepts are built on simpler parts; and Learning-to-learn, the ability to learn strategies for decomposing and recombining components to form new concepts. These principles provide inductive biases that support efficient generalization. They are critical to Chinese character recognition (CCR) in solving the zero-shot problem, which results from the common long-tail distribution of Chinese character datasets. Existing methods have made substantial progress in modeling compositionality via predefined radical or stroke decomposition. However, they often ignore the learning-to-learn capability, limiting their ability to generalize beyond human-defined schemes. Inspired by these principles, we propose a deep latent variable model that learns Compositional Latent components of Chinese characters (CoLa) without relying on human-defined decomposition schemes. Recognition and matching can be performed by comparing compositional latent components in the latent space, enabling zero-shot character recognition. The experiments illustrate that CoLa outperforms previous methods in both character the radical zero-shot CCR. Visualization indicates that the learned components can reflect the structure of characters in an interpretable way. Moreover, despite being trained on historical documents, CoLa can analyze components of oracle bone characters, highlighting its cross-dataset generalization ability.