Kanji Radical Match AI

I worked in a team of four for my Foundations of Artificial Intelligence course to build a neural network that maps English words to the kanji radicals most associated with their meaning. I created a custom dataset from different sources, the architecture of the neural network, and training and testing it.

Dataset

There was no ready-made English to radical mapping, we needed to create one. We chained English to kanji, kanji to English, and kanji to radical to assemble an English to radical training set. The sources included pre-made JSON files which were originally derived from the EDRDG’s KRADFILE.

Model development and architecture

The model developed was a feedforward neural network that learns to map an English word that’s represented as a vocabulary-sized tensor to a probability distribution over all kanji radicals. It takes an English word and returns a probability for every radical on a 0.0 to 1.0 scale, then ranks the top 10 radicals it associates with the word. This is a multi-label model since one word can have a greater-than-zero probability with multiple radicals. In this project, one word is simply a one-hot encoding of the vocabulary, and the target is a set of radical probabilities. There’s no sequence or 2D grid, so a feedforward neural network was the most realistic choice.

Results

The initial model, V3, generally predicted lower probabilities. After adjusting the architecture, the final model, V4, predicted higher probabilities where confidence dropped sharply with radicals that were not associated. From these results, the model was able to show connections between English words and radicals that had similar meanings.

Interactive app

The project has a console GUI, and from there, it accepts English input and plots the chosen radicals against their probabilities in a Matplotlib bar chart.

The live demo displays a chart, and all predictions were precomputed from the trained model. It runs instantly with no Python nor server.