Extractive Question-Answering for the Thai language

This work is built based on thai2transformers.

The environment is described in requirement.txt

Datasets

Three extractive question answeirng datasets for the Thai language.

 - xquad-th (the split version on https://huggingface.co/datasets/zhufy/xquad_split)
 - thaiqa (https://huggingface.co/datasets/thaiqa_squad)
 - iapp_wiki_qa_squad (https://huggingface.co/datasets/iapp_wiki_qa_squad)

Combine the training data from the three datasets for training. The script for combining the dataset is

./utils/combine_datasets.py

Models

Include both mono- and multi- lingual language models that can work for the Thai language.

Monolingual language model

WangchenBERTa: we use the wangchanberta-base-att-spm-uncased released by the VISTEC-depa AI Research Institute of Thailand on HuggingFace.

Multilingual language models

• Multilingual BERT: we use the bert-base-multilingual-cased released on HuggingFace.
• XLM-Roberta: we use the xlm-roberta-base released on HuggingFace.

Fine-tune Models

Run the script bash run_train.sh to fine-tune the model.

Monolingual Language Model

The parameters can be set are as follows,

python train_combined.py \
  --model_name wangchanberta-base-att-spm-uncased\
  --dataset_name iapp_thaiqa_xquad\
  --output_dir ~/saved/thai/combined/wangchanberta-base-att-spm-uncased \
  --log_dir ./logs/combined-wangchanberta\
  --lowercase \
  --pad_on_right \
  --fp16 \
  --num_train_epochs 7\
  --batch_size 50\
  --gpu 7 \
  --seed 42 \

train_base.py Train baseline models with single dataset

train_combined.py Train baseline model with the combined dataset

Evaluate Modles

Run the script bash run_eval.sh to evaluate the model.

The parameters can be set are as follows,

python eval_combined.py \
  --model_name bert-base-multilingual-cased\
  --eval_base \
  --dataset_name iapp_wiki_qa_squad\
  --output_dir ~/saved/thai/combined/bert-base-multilingual-cased\
  --log_dir ./test/\
  --lowercase \
  --pad_on_right \
  --fp16 \
  --num_train_epochs 7\
  --batch_size 50 \
  --seed 42 \
  --gpu 5\

eval.py Evaluate the base models

eval_combined.py Evaluate the models trained with the combined dataset

eval_huggingface Evaluate a model released on HuggingFace