Is Language Visual? An Experiment with Chinese Characters

An experiment explores if Chinese is fundamentally visual by training a language model on character images instead of traditional text tokens. The visual model demonstrates a "hot-start" effect, learning much faster initially by leveraging the visual similarity of characters sharing common radicals. However, both the visual and text-based models eventually converge to the same final accuracy, as linguistic co-occurrence patterns ultimately dominate performance. This visual approach proves most advantageous in low-data settings and for applications like interpreting damaged historical documents.