Model Description

• Language model: XLM-RoBERTa
• Fine-tune: MRCQuestionAnswering
• Language: Vietnamese, Englsih
• Downstream-task: Extractive QA
• Dataset (combine English and Vietnamese):
  • Squad 2.0
  • mailong25
  • UIT-ViQuAD
  • MultiLingual Question Answering

This model is intended to be used for QA in the Vietnamese language so the valid set is Vietnamese only (but English works fine). The evaluation result below using 10% of the Vietnamese dataset.

Model	EM	F1
base	76.43	84.16
large	77.32	85.46

MRCQuestionAnswering using XLM-RoBERTa as a pre-trained language model. By default, XLM-RoBERTa will split word in to sub-words. But in my implementation, I re-combine sub-words representation (after encoded by BERT layer) into word representation using sum strategy.

Using pre-trained model

• Hugging Face pipeline style (NOT using sum features strategy).

from transformers import pipeline
# model_checkpoint = "nguyenvulebinh/vi-mrc-large"
model_checkpoint = "nguyenvulebinh/vi-mrc-base"
nlp = pipeline('question-answering', model=model_checkpoint,
                   tokenizer=model_checkpoint)
QA_input = {
  'question': "Bình là chuyên gia về gì ?",
  'context': "Bình Nguyễn là một người đam mê với lĩnh vực xử lý ngôn ngữ tự nhiên . Anh nhận chứng chỉ Google Developer Expert năm 2020"
}
res = nlp(QA_input)
print('pipeline: {}'.format(res))
#{'score': 0.5782045125961304, 'start': 45, 'end': 68, 'answer': 'xử lý ngôn ngữ tự nhiên'}

• More accurate infer process (Using sum features strategy)

from infer import tokenize_function, data_collator, extract_answer
from model.mrc_model import MRCQuestionAnswering
from transformers import AutoTokenizer

# model_checkpoint = "nguyenvulebinh/vi-mrc-large"
model_checkpoint = "nguyenvulebinh/vi-mrc-base"
tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)
model = MRCQuestionAnswering.from_pretrained(model_checkpoint)

QA_input = {
  'question': "Bình được công nhận với danh hiệu gì ?",
  'context': "Bình Nguyễn là một người đam mê với lĩnh vực xử lý ngôn ngữ tự nhiên . Anh nhận chứng chỉ Google Developer Expert năm 2020"
}

inputs = [tokenize_function(*QA_input)]
inputs_ids = data_collator(inputs)
outputs = model(**inputs_ids)
answer = extract_answer(inputs, outputs, tokenizer)

print(answer)
# answer: Google Developer Expert. Score start: 0.9926977753639221, Score end: 0.9909810423851013

Training model

In data-bin/raw folder already exist some sample data files for the training process. Do following steps:

• Create environment by using file requirements.txt

• Clean data

python squad_to_mrc.py
python train_valid_split.py

• Train model

python main.py

• Test model

python infer.py