nju-websoft / SpanQualifier

SpanQualifier

This is a pytorch version of SpanQualifier, your can find the implementation based on MindSpore 2.0 in here

Code for Spans, Not Tokens: A Span-Centric Model for Multi-Span Reading Comprehension (CIKM 2023)

Installation

pip install -r requirements.txt

The Usage of SpanQualifier

To train the SpanQualifier model, we can use the following command:

python run_SpanQualifier.py \
 --model_path $Pre_Train_Language_Model_PATH$ \
 --data_path_train $TRAIN_FILE_PATH$ \
 --data_path_dev $DEV_FILE_PATH$ \
 --data_path_test $TEST_FILE_PATH$ 

After training, the program will evaluate SpanQualifier and will report the scores of Exact Match F1 (EM) and Partial Match F1 (PM) on the dev and test sets.

The predictions on the dev and test sets will be saved in /results/Dataset_Name/Model_Name/HyperParameters/dev.json and /results/Dataset_Name/Model_Name/HyperParameters/test.json, where the file structure is as follows:

[
    {
        "The id of question": [
          "Answer span 1",
          "Answer span 2",
          ...
        ]
    },
    {
        ...
    }
]

The checkpoint of SpanQualifier will be saved in /outputs/Dataset_Name/Model_Name/HyperParameters/pytorch_model.bin.