Hi! Here you can find the trained models for bunpo-check.com, as well as libraries for generating the data and carrying out the model training.

• Permutations library: permut8r
• Trained models available for download

• Model training code
• Website JS/HTML/CSS
• Functions for interpreting output

See requirements.txt

This is a permutation library I wrote for creating random permutations of correct strings, for use as pre-training data.

What does it do?

# Original string from Wikipedia
"その使用は1世紀に遡ることができ、5世紀中葉から現代に至るまでの変遷がわかる。"

# Output from permut8r
"その使用は1世紀に遡ることができ、5世紀中葉から現代に至るまでの変遷がわかる。,011111111111111111111111110000000000000000000000"
"その使用1世紀に遡ることができ、5世紀中た葉から現代に至るまでの変遷業がわかる。,012211111111112211111112211000000000000000000000"

A string is converted to tokens via the Huggingface tokenizer. Then, random permutations are applied to it from the following:

• Deleting tokens or characters
• Swapping tokens around
• Swapping kanji for other kanji with the same reading
• Swapping particles for other random particles
• Inserting random kanji or particles

Each token is assigned a label of 0 (unimportant eg. PAD tokens), 1 (correct grammar) or 2 (error). The correct string is saved with a label sequence of only 0/1. The permuted string contains some new errors, and these are encoded with a label of 2 for the affected tokens.

This allowed me to build a ~263M dataset of labelled sentences which reached a decent F0.5 score of 45.0 on a test set, which went a looong way to getting a good performance here.

Will upload soon!

To use the trained model, you will need to use the Huggingface transformers library.

Step 1: download the model from my Google drive and save the folder to your working directory.

Step 2: run check.py to check your sentence.

python check.py "⽂法ーCHECKは⼈⼯知能により⽂法が正しいか確かめられるサイトです。"