NamBert-for-csc

Official model for the paper "Unveiling the Impact of Multimodal Features on Chinese Spelling Correction: From Analysis to Design".

Github: https://github.com/iioSnail/NamBert

The sentence-level performance of the model in SIGHAN datasets is as follows:

	Detect-Acc	Detect-Precision	Detect-Recall	Detect-F1	Correct-Acc	Correct-Precision	Correct-Recall	Correct-F1
Sighan2013	82.70	87.72	82.39	84.97	81.60	86.51	81.26	83.80
Sighan2014	79.76	69.03	75.00	71.89	79.10	67.79	73.65	70.60
Sighan2015	86.18	77.52	85.40	81.27	85.73	76.68	84.47	80.39

Usage

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("iioSnail/NamBert-for-csc", trust_remote_code=True)
model = AutoModel.from_pretrained("iioSnail/NamBert-for-csc", trust_remote_code=True)

inputs = tokenizer("我喜换吃平果，逆呢？", return_tensors='pt')
logits = model(**inputs).logits

target_ids = logits.argmax(-1)
target_ids = tokenizer.restore_ids(target_ids, inputs['input_ids'])

print(''.join(tokenizer.convert_ids_to_tokens(target_ids[0, 1:-1])))

Or

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("iioSnail/NamBert-for-csc", trust_remote_code=True)
model = AutoModel.from_pretrained("iioSnail/NamBert-for-csc", trust_remote_code=True)

model = model.to(device)
model = model.eval()
model.set_tokenizer(tokenizer)

model.predict("我是炼习时长两念半的个人练习生菜徐坤")
model.predict(["我是炼习时长两念半的个人练习生菜徐坤", "喜欢场跳rap篮球！！"])