BERT Chinese QA

Prerequisites

• Python >= 3.8
• pip install -r requirements.txt to install other dependencies

Data

A context file context.json should be provided, containing all the relevant context with following structure,

[
    "context1",
    "context2",
    ...
]

Two files train.json and valid.json should be provided for training. The JSON file should contain a list of data sample with structure as follow,

[
    {
        "id": "9f0860355c55e113d615cc88b3fa9420",
        "question": "伊利諾大學是哪一個國家的大學?",
        "paragraphs": [
            6037,
            4555,
            5641,
            1029
        ],
        "relevant": 4555,
        "answer": {
        "text": "美國",
        "start": 15
        }
    },
    ...
]

For inference, a file test.json should be provided, and will only contains question and relevant context index.

[
    {
        "id": "5e7a923dd6e4ccb8730eb95230e0c908",
        "question": "卡利創立的網際網路檔案館要到什麼時後才開放存取？",
        "paragraphs": [
            8912,
            7873,
            8209,
            7497
        ]
    },
    ...
]

Training

There are two phases of the task, context selection and question answering. Two phases need to be trained seperately.

Context Selection

python3 cs_train.py cs_config.json

You can also modify cs_config.json for different training configuration.

Question Answering

python3 qa_train.py qa_config.json

You can also modify qa_config.json for different training configuration.

Logging

tensorboard can be used for logging illustration.

tensorboard --logdir [logging directory]

The default logging directory is [task]_output/runs/[current datetime]_[hostname].

Testing

Pretrained Checkpoints

Pretrained checkpoints are provided(hfl/chinese-macbert-large is used for both context selection and question answering) and can be obtained by running the download script,

bash download.sh

Two-phased Inference

To inference testing data with both context selection and question answering model, run

python3 inference.py inference_config.json

QA-only Inference

To inference testing data with only the question answering model(concatenating all the relevant contexts as one), run

python3 inference-qa-only.py inference_config.json