文法ーCHECK lets you check your Japanese sentences for grammatical errors using an AI model.
Just type in the box, hit チェック and any errors will be underlined in red. Really short sentences and the ends of really long sentences won’t be checked, so watch out. Please note this is still an experimental service and results may not always be accurate.
Please also note that to save on server capacity, up to 100 sentences can be checked without delay. Once they are all used, you will not be able to check any more until you’ve ‘recharged’ some at a rate of 2 sentences/minute.
The model was pre-trained on ~246M sentences from Japanese Wikipedia and Common Crawl data which were randomly permuted using a program I wrote, then fine-tuned on ~800K sentences from the Lang8 dataset.
The model itself is an implementation of BERT from the Huggingface library trained for token-level classification. Thanks to University of Tohoku NLP Lab for providing the initial weights and tokenizer.
If you have any suggestions or want to report any errors, you can reach me on Github.
2022 Albert Sokol