Mike0307/multilingual-e5-language-detection

Overview

This model supports the detection of 45 languages, and it's fine-tuned using multilingual-e5-base model on the common-language dataset.
The overall accuracy is 98.37%, and more evaluation results are shown the below.

Download the model

from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained('Mike0307/multilingual-e5-language-detection')
model = AutoModelForSequenceClassification.from_pretrained('Mike0307/multilingual-e5-language-detection', num_labels=45)

Example of language detection

import torch

languages = [
    "Arabic", "Basque", "Breton", "Catalan", "Chinese_China", "Chinese_Hongkong", 
    "Chinese_Taiwan", "Chuvash", "Czech", "Dhivehi", "Dutch", "English", 
    "Esperanto", "Estonian", "French", "Frisian", "Georgian", "German", "Greek", 
    "Hakha_Chin", "Indonesian", "Interlingua", "Italian", "Japanese", "Kabyle", 
    "Kinyarwanda", "Kyrgyz", "Latvian", "Maltese", "Mongolian", "Persian", "Polish", 
    "Portuguese", "Romanian", "Romansh_Sursilvan", "Russian", "Sakha", "Slovenian", 
    "Spanish", "Swedish", "Tamil", "Tatar", "Turkish", "Ukranian", "Welsh"
]

def predict(text, model, tokenizer, device = torch.device('cpu')):
    model.to(device)
    model.eval()
    tokenized = tokenizer(text, padding='max_length', truncation=True, max_length=128, return_tensors="pt")
    input_ids = tokenized['input_ids']
    attention_mask = tokenized['attention_mask']
    with torch.no_grad():
        input_ids = input_ids.to(device)
        attention_mask = attention_mask.to(device)
        outputs = model(input_ids=input_ids, attention_mask=attention_mask)
    logits = outputs.logits
    probabilities = torch.nn.functional.softmax(logits, dim=1)
    return probabilities

def get_topk(probabilities, languages, k=3):
    topk_prob, topk_indices = torch.topk(probabilities, k)
    topk_prob = topk_prob.cpu().numpy()[0].tolist()
    topk_indices = topk_indices.cpu().numpy()[0].tolist()
    topk_labels = [languages[index] for index in topk_indices]
    return topk_prob, topk_labels

text = "你的測試句子"
probabilities = predict(text, model, tokenizer)
topk_prob, topk_labels = get_topk(probabilities, languages)
print(topk_prob, topk_labels)

# [0.999620258808, 0.00025940246996469, 2.7690215574693e-05]
# ['Chinese_Taiwan', 'Chinese_Hongkong', 'Chinese_China']

Evaluation Results

The test datasets refers to the common_language test datasets.

index	language	precision	recall	f1-score	support
0	Arabic	1.00	1.00	1.00	151
1	Basque	0.99	1.00	1.00	111
2	Breton	1.00	0.90	0.95	252
3	Catalan	0.96	0.99	0.97	96
4	Chinese_China	0.98	1.00	0.99	100
5	Chinese_Hongkong	0.97	0.87	0.92	115
6	Chinese_Taiwan	0.92	0.98	0.95	170
7	Chuvash	0.98	1.00	0.99	137
8	Czech	0.98	1.00	0.99	128
9	Dhivehi	1.00	1.00	1.00	111
10	Dutch	0.99	1.00	0.99	144
11	English	0.96	1.00	0.98	98
12	Esperanto	0.98	0.98	0.98	107
13	Estonian	1.00	0.99	0.99	93
14	French	0.95	1.00	0.98	106
15	Frisian	1.00	0.98	0.99	117
16	Georgian	1.00	1.00	1.00	110
17	German	1.00	1.00	1.00	101
18	Greek	1.00	1.00	1.00	153
19	Hakha_Chin	0.99	1.00	0.99	202
20	Indonesian	0.99	0.99	0.99	150
21	Interlingua	0.96	0.97	0.96	182
22	Italian	0.99	0.94	0.96	100
23	Japanese	1.00	1.00	1.00	144
24	Kabyle	1.00	0.96	0.98	156
25	Kinyarwanda	0.97	1.00	0.99	103
26	Kyrgyz	0.98	1.00	0.99	129
27	Latvian	0.98	0.98	0.98	171
28	Maltese	0.99	0.98	0.98	152
29	Mongolian	1.00	1.00	1.00	112
30	Persian	1.00	1.00	1.00	123
31	Polish	0.91	0.99	0.95	128
32	Portuguese	0.94	0.99	0.96	124
33	Romanian	1.00	1.00	1.00	152
34	Romansh_Sursilvan	0.99	0.95	0.97	106
35	Russian	0.99	0.99	0.99	100
36	Sakha	0.99	1.00	1.00	105
37	Slovenian	0.99	1.00	1.00	166
38	Spanish	0.96	0.95	0.95	94
39	Swedish	0.99	1.00	0.99	190
40	Tamil	1.00	1.00	1.00	135
41	Tatar	1.00	0.96	0.98	173
42	Turkish	1.00	1.00	1.00	137
43	Ukranian	0.99	1.00	1.00	126
44	Welsh	0.98	1.00	0.99	103
	macro avg	0.98	0.99	0.98	5963
	weighted avg	0.98	0.98	0.98	5963
	overall accuracy			0.9837	5963