File size: 7,037 Bytes
50fa40e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3158261
50fa40e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bf53319
50fa40e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3158261
50fa40e
3158261
 
50fa40e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e651f46
 
cc9d424
 
e651f46
73454cf
 
e651f46
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
---
language: 
- multilingual
- en 
- ar 
- bg 
- de 
- el 
- es 
- fr 
- hi 
- ru 
- sw 
- th 
- tr 
- ur 
- vi 
- zh
license: apache-2.0
tags:
- zero-shot-classification
- text-classification
- nli
- pytorch
metrics:
- accuracy
datasets:
- multi_nli
- xnli
pipeline_tag: zero-shot-classification
widget:
- text: "Angela Merkel ist eine Politikerin in Deutschland und Vorsitzende der CDU"
  candidate_labels: "politics, economy, entertainment, environment"
---
# Multilingual ernie-m-large-mnli-xnli
## Model description
This multilingual model can perform natural language inference (NLI) on 100 languages and is therefore also suitable for multilingual 
zero-shot classification. The underlying model was pre-trained by Baidu, based on Meta's RoBERTa (pre-trained on the 
[CC100 multilingual dataset](https://huggingface.co/datasets/cc100). It was then fine-tuned on the [XNLI dataset](https://huggingface.co/datasets/xnli), 
which contains hypothesis-premise pairs from 15 languages, as well as the English [MNLI dataset](https://huggingface.co/datasets/multi_nli).
The model was introduced by Baidu in [this paper](https://arxiv.org/pdf/2012.15674.pdf). The model outperforms RoBERTa models of equal size. 

If you are looking for a much faster (but less performant) model, you can 
try [multilingual-MiniLMv2-L6-mnli-xnli](https://huggingface.co/MoritzLaurer/multilingual-MiniLMv2-L6-mnli-xnli).
If you are looking for a base-sized model with a good mix of performance and speed, 
you can try [mDeBERTa-v3-base-mnli-xnli](https://huggingface.co/MoritzLaurer/mDeBERTa-v3-base-mnli-xnli)

### How to use the model
#### Simple zero-shot classification pipeline
```python
from transformers import pipeline
classifier = pipeline("zero-shot-classification", model="MoritzLaurer/ernie-m-large-mnli-xnli")

sequence_to_classify = "Angela Merkel ist eine Politikerin in Deutschland und Vorsitzende der CDU"
candidate_labels = ["politics", "economy", "entertainment", "environment"]
output = classifier(sequence_to_classify, candidate_labels, multi_label=False)
print(output)
```
#### NLI use-case
```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

model_name = "MoritzLaurer/ernie-m-large-mnli-xnli"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name).to(device)

premise = "Angela Merkel ist eine Politikerin in Deutschland und Vorsitzende der CDU"
hypothesis = "Emmanuel Macron is the President of France"

input = tokenizer(premise, hypothesis, truncation=True, return_tensors="pt")
output = model(input["input_ids"].to(device))  # device = "cuda:0" or "cpu"
prediction = torch.softmax(output["logits"][0], -1).tolist()
label_names = ["entailment", "neutral", "contradiction"]
prediction = {name: round(float(pred) * 100, 1) for pred, name in zip(prediction, label_names)}
print(prediction)
```

### Training data
This model was trained on the XNLI development dataset and the MNLI train dataset. 
The XNLI development set consists of 2490 professionally translated texts from English 
to 14 other languages (37350 texts in total) (see [this paper](https://arxiv.org/pdf/1809.05053.pdf)). 
Note that the XNLI contains a training set of 15 machine translated versions of the MNLI dataset for 15 languages, 
but due to quality issues with these machine translations, this model was only trained 
on the professional translations from the XNLI development set and the original English 
MNLI training set (392 702 texts). Not using machine translated texts can avoid overfitting the 
model to the 15 languages; avoids catastrophic forgetting of the other 85 languages ernie-m 
was pre-trained on; and significantly reduces training costs. 

### Training procedure
ernie-m-large-mnli-xnli was trained using the Hugging Face trainer with the following hyperparameters.
```
training_args = TrainingArguments(
    num_train_epochs=3,              # total number of training epochs
    learning_rate=3e-05,
    per_device_train_batch_size=16,   # batch size per device during training
    gradient_accumulation_steps=2,
    per_device_eval_batch_size=16,    # batch size for evaluation
    warmup_ratio=0.1,                # number of warmup steps for learning rate scheduler
    weight_decay=0.01,               # strength of weight decay
    fp16=True,
)
```
### Eval results
The model was evaluated on the XNLI test set on 15 languages (5010 texts per language, 75150 in total). 
Note that multilingual NLI models are capable of classifying NLI texts without receiving NLI training 
data in the specific language (cross-lingual transfer). This means that the model is also able of 
doing NLI on the other 85 languages mDeBERTa was training on, but performance is most likely lower 
than for those languages available in XNLI.

Also note that if other multilingual models on the model hub claim performance of around 90% on languages 
other than English, the authors have most likely made a mistake during testing since non of the latest papers 
shows a multilingual average performance of more than a few points above 80% on XNLI 
(see [here](https://arxiv.org/pdf/2111.09543.pdf) or [here](https://arxiv.org/pdf/1911.02116.pdf)). 

|Datasets|avg_xnli|mnli_m|mnli_mm|ar|bg|de|el|en|es|fr|hi|ru|sw|th|tr|ur|vi|zh|
| :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
|Accuracy|0.822|0.881|0.878|0.818|0.853|0.84|0.837|0.882|0.855|0.849|0.799|0.83|0.751|0.809|0.818|0.76|0.826|0.799|
|Inference text/sec (A100, batch=120)|1415.0|783.0|774.0|1487.0|1396.0|1430.0|1206.0|1623.0|1482.0|1291.0|1302.0|1366.0|1484.0|1500.0|1609.0|1344.0|1403.0|1302.0|


## Limitations and bias
Please consult the original ernie-m paper and literature on different NLI datasets for potential biases. 

## Citation
If you use this model, please cite: Laurer, Moritz, 
Wouter van Atteveldt, Andreu Salleras Casas, and Kasper Welbers. 2022. 
‘Less Annotating, More Classifying – Addressing the Data Scarcity Issue of Supervised Machine 
Learning with Deep Transfer Learning and BERT - NLI’. Preprint, June. Open Science Framework. https://osf.io/74b8k.

## Ideas for cooperation or questions?
If you have questions or ideas for cooperation, contact me at m{dot}laurer{at}vu{dot}nl 
or [LinkedIn](https://www.linkedin.com/in/moritz-laurer/)


## Debugging and issues
The ernie-m architecture is only supported with transformers==4.27 or higher 
(which is not yet released and causes an error in the inference widget as of 03.03.23). 
In order to run the model before the release of 4.27, you need to install transformers from source with: `pip install git+https://github.com/huggingface/transformers`  
as well as the sentencepiece tokenizer with: `pip install sentencepiece`  
After the release, you can run: `pip install transformers[sentencepiece]>=4.27`