QA-for-Event-Extraction

A question answering method for Event Extraction on your own Chinese dataset (Also in English dataset).

Prerequisites

1. Install the packages.

   pip install pytorch_pretrained_bert
   pip install numpy torch
   

   Download BERT-base-chinese model from https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese.tar.gz

   Download bert-base-chinese-vocab from https://s3.amazonaws.com/models.huggingface.co/bert/bert-base-chinese-vocab.txt.

   Then put these file in

   ./bert-base-chinese
      --bert_config.json
      --pytorch_model.bin
      --vocab.txt
   

2. Prepare Your own dataset.

   sample.json

   [
     {
       "words": "在冲突中，一名年轻的激进示威者用尖刀刺伤警员。",
       "golden-event-mentions": [
         {
           "event_type": "冲突：冲击",
           "trigger": {
             "text": "刺伤",
             "start": 18,
             "end": 20
           },
           "arguments": [
             {
               "text": "警员",
               "start": 20,
               "end": 22,
               "entity_type": "人物",
               "role": "受害者"
             },
             {
               "text": "激进示威者",
               "start": 10,
               "end": 15,
               "entity_type": "人物",
               "role": "攻击者"
             }
           ]
         }
       ]
     },
   ]
   

   put the data into

   ./data/
      --train.json
      --dev.json
      --test.json
   

3. Add your trigger and argument category in the file ./const.py

4. Design question templates based on your event trigger and argument in ./question.csv

   for example

      ##事件类型,事件角色,设计的问题##
      冲突：冲击,攻击者,谁发起了这次攻击事件？
   

Usage

Run:

sudo python qa-trigger.py

• You can get Precision/Recall/F1 on event trigger extraction.
• Then you can get the "trigger_model.pt" in main directory.

sudo python qa-argument.py

• You can get Precision/Recall/F1 on event argument extraction.
• Then you can get the "argument_model.pt" in main directory.

Reference

• xinyadu's eeqa repository [github]
• Nlpcl-lab's bert_event_extraction repository [github]