statements

• 数据说明（./data）

  • BoolQ
    • train/dev.jsonl: BoolQ format data.
    • train/dev.transformers_format.json: the transformer format data of original BoolQ data.
    • train/dev.statement.transformers_format.csv: the statement & transformer format data of original BoolQ data.

• question-to-statement

  • ./question-to-statement
  • 参考：
    • https://github.com/SunnyWay/question_to_statement
  • 启动脚本：
    • python q2s.py path_to_input_file path_to_output_file path_to_stanfordCoreNLP
      • input_file:
        • BoolQ format data.
    • 环境依赖：
      • stanfordcorenlp

• statement-to-question

  • ./statement-to-question
  • 参考：
    • https://github.com/JacobHA/Sentence-to-Question
  • 启动脚本：
    • python s2q.py path_to_input_file path_to_output_file
    • input_file:
      • transformers format data.
    • 环境依赖：
      • nltk

• Classification Script

  • run_glue.py

  • 运行命令：

    • # Finetune on MNLI dataset
      export TASK_NAME=MNLI
      
      python run_glue.py \
        --model_name_or_path bert-large-uncased \
        --task_name $TASK_NAME \
        --do_train \
        --do_eval \
        --max_seq_length 128 \
        --per_device_train_batch_size 32 \
        --learning_rate 2e-5 \
        --num_train_epochs 3.0 \
        --save_steps 10000 \
        --output_dir /tmp/$TASK_NAME/
      

    • # Finetune on BoolQ dataset
      python run_glue.py \
        --model_name_or_path $path_to_model_finetuned_on_MNLI \
        --do_train \
        --do_eval \
        --fp16 \
        --train_file $path_to_train_file \
        --validation_file $path_to_dev_file \
        --max_seq_length 256 \
        --per_device_train_batch_size 12 \
        --per_device_eval_batch_size 12 \
        --learning_rate 2e-5 \
        --num_train_epochs 3.0 \
        --save_steps 10000 \
        --output_dir $path_to_output_path \
        --overwrite_output_dir
      # Note: $path_to_model_finetuned_on_MNLI只加载BERT的编码层参数
      

    • # Prediction on statement data
      python run_glue.py \
        --model_name_or_path  $path_to_model_finetuned_on_BoolQ\
        --do_eval \
        --train_file $none_file \
        --validation_file $path_to_eval_file \
        --overwrite_cache \
        --max_seq_length 256 \
        --per_device_eval_batch_size 32 \
        --learning_rate 2e-5 \
        --num_train_epochs 3.0 \
        --output_dir $path_to_output_path
      

  • 环境依赖

    • datasets
    • tranformers