Chinese Question Answering

Task Description

• Input: Context list + Question
• Output: Answer

Example

Context

"鼓是一種打擊樂器，...，最早的鼓出現於西元前六千年的兩河文明"
"盧克萊修生於共和國末期，...，被古典主義文學視為經典"
"視網膜又稱視衣，...，約3mm2大的橢圓。"

Question

"最早的鼓可以追溯至什麼古文明?"

Answer

"兩河文明"

Data Format

context.json

list of short paragraphs

[
  "鼓是一種打擊樂器，也是一種通訊工具，非洲某些部落...的兩河文明。",
  "這次出售的贖罪券很特別，是全大赦贖罪券，可以贖買...購買，可見其盛況。",
  "處在千年古都的西安交大校園少不了和歷史千絲萬縷的...之一，可謂人傑地靈。",
  ...
]

questions

• id: question ID
• question: question text
• paragraphs: list of paragraph IDs, up to 7 entries (0-based)
• relevant: ID of the relevant context (0-based) (* absent in private.json)
• answers: list of answers to the question (* absent in private.json)
  • text: the answer text
  • start: answer span start position in the relevant context

Example

{
  "id": "ab39567999fd376480ac3076904e598e",
  "question": "舍本和誰的數據能推算出連星的恆星的質量？",
  "paragraphs": [5234, 6952, 8264, 836, 92, 2018],
  "relevant": 836,
  "answers": [
    {
      "text": "斯特魯維",
      "start": 108
    }
  ]
}

Instructions

Download Model

bash download.sh

Testing

bash run.sh

Training

python train.py [path/to/context.json] [path/to/train.json] [path/to/save_model]