akhtet/mDeBERTa-v3-base-myanmar-xnli

Model Card for mDeBERTa-v3-base-myXNLI

mDeBERTa-v3-base-myXNLI is a transformer model for text classification English and Myanmar (Burmese).

It is based on multilingual DeBERTa v3 model and fine-tuned using myXNLI dataset on the Natural Language Inference task in English and Myanmar.

Thus it is useful for Natural Language Inference and related tasks such as Zero-shot Text Classification on both English and Myanmar data.

Model Details

• Model type: Transformer Encoder
• Language(s) (NLP): Fine-tuned for Myanmar (Burmese) and English
• License: MIT
• Finetuned from model: mDeBERTa v3 base
• Paper : Myanmar XNLI: Building a Dataset and Exploring Low-resource Approaches to Natural Language Inference with Myanmar
• Demo : A demo of Zero-shot Text Classification in Myanmar can be found on this page.

Bias, Risks, and Limitations

Please refer to the paper on MyXNLI, as well as the papers for the foundation models: DeBERTa and DeBERTaV3.

How to Get Started with the Model

Use the code below to get started with the model for zero-shot classification task.

from transformers import pipeline

classifier = pipeline(task="zero-shot-classification", model="akhtet/mDeBERTa-v3-base-myXNLI", framework="pt")

output = classifier("မြန်မာ့စီးပွားရေးမှာ ရွှေ နဲ့ ဒေါ်လာက အရေးပါသလို ဒေါ်လာစျေးပေါ်မူတည်ပြီး အခြားစားသောက်ကုန်ပစ္စည်းတွေကလည်း လိုက်ပါပြောင်းလဲလေ့ ရှိပါတယ်။",
    candidate_labels=["commerce", "fashion", "music", "politics", "sports"],
)

print (output)
# output
# {'sequence': 'မြန်မာ့စီးပွားရေးမှာ ရွှေ နဲ့ ဒေါ်လာက အရေးပါသလို ဒေါ်လာစျေးပေါ်မူတည်ပြီး အခြားစားသောက်ကုန်ပစ္စည်းတွေကလည်း လိုက်ပါပြောင်းလဲလေ့ ရှိပါတယ်။',
# 'labels': ['commerce', 'politics', 'fashion', 'music', 'sports'],
# 'scores': [0.8995707631111145, 0.048580411821603775, 0.035297513008117676, 0.009092549793422222, 0.007458842825144529]}

Fore more details on zero-shot classification, please refer to HuggingFace documentation https://huggingface.co/tasks/zero-shot-classification

Training Details

The model is fine-tuned on myXNLI dataset https://huggingface.co/datasets/akhtet/myXNLI. The English portion of myXNLI is from XNLI dataset.

From this dataset, 4 different copies training data from myXNLI were concatenated, each with sentence pairs in en-en, en-my, my-en and my-my combinations.

Training on cross-matched language data as above improved the NLI accuracy over training separately in each language. This approach was inspired by another model https://huggingface.co/joeddav/xlm-roberta-large-xnli

The model was fine-tuned using this combined dataset for a single epoch.

Evaluation

This model has been evaluted on myXNLI testset for Myanmar accuracy. We also provide the accuracy for English using XNLI testset.

Model	English accuracy	Myanmar accuracy
mDeBERTa-v3-base-myXNLI	88.02	80.99

Citation

[More Information Needed]

Model Card Contact

Aung Kyaw Htet