iamdenay/roberta-azerbaijani

Roberta base model trained on Azerbaijani subset of OSCAR corpus as a part of research on application of text augentation for low-resource languages. It was developed to enhance text classification tasks in Azerbaijani, a low-resource language in the NLP domain. The model was trained using the Azerbaijani subset of the OSCAR corpus and further fine-tuned on a labeled news dataset.

Training Data

The model was pre-trained on the Azerbaijani subset of the OSCAR corpus, and fine-tuned on approximately 3 million sentences from Azertag News Agency covering diverse topics such as politics, economy, culture, sports, technology, and health.

Citation

@article{ziyaden2024augmentation,
    title        = {Text data augmentation and pre-trained Language Model for enhancing text classification of low-resource languages},
    author       = {Ziyaden, Atabay and Yelenov, Amir and Hajiyev, Fuad and Rustamov, Samir and Pak, Alexandr},
    year         = 2024,
    journal      = {PeerJ Computer Science},
    doi          = {10.7717/peerj-cs.1974},
    url          = {https://doi.org/10.7717/peerj-cs.1974}
}

Usage

from transformers import AutoTokenizer, AutoModelWithLMHead

tokenizer = AutoTokenizer.from_pretrained("iamdenay/roberta-azerbaijani")

model = AutoModelWithLMHead.from_pretrained("iamdenay/roberta-azerbaijani")

from transformers import pipeline
model_mask = pipeline('fill-mask', model='iamdenay/roberta-azerbaijani')
model_mask("Le tweet <mask>.")

Output

[{'sequence': 'azərtac xəbər verir ki',
  'score': 0.9791,
  'token': 1053,
  'token_str': 'verir'},
 {'sequence': 'azərtac xəbər verib ki',
  'score': 0.0044,
  'token': 2313,
  'token_str': 'verib'},
 ... ]

Limitations

• Language Specificity: The model is trained exclusively on Azerbaijani and may not generalize well to other languages.
• Data Bias: The fine-tuning data is sourced from news articles, which may contain biases or specific journalistic styles.
• Agglutinative Language Challenges: Azerbaijani's agglutinative nature can lead to sparsity in the word space due to numerous morphological variations.

Ethical Considerations

• Content Sensitivity: The dataset may include sensitive topics. Users should ensure compliance with ethical standards when deploying the model.
• Bias and Fairness: Be aware of potential biases in the training data that could affect model predictions.

Config

attention_probs_dropout_prob:0.1
bos_token_id:0
classifier_dropout:null
eos_token_id:2
gradient_checkpointing:false
hidden_act:"gelu"
hidden_dropout_prob:0.1
hidden_size:768
initializer_range:0.02
intermediate_size:3072
layer_norm_eps:1e-12
max_position_embeddings:514
model_type:"roberta"
num_attention_heads:12
num_hidden_layers:6
pad_token_id:1
position_embedding_type:"absolute"
torch_dtype:"float32"
transformers_version:"4.10.0"
type_vocab_size:1
use_cache:true
vocab_size:52000