telunyang/rocling2022_ner

ROCLING2022 shared task - Chinese Healthcare Named Entity Recognition - NER

crowNER

團隊成員: 張智星教授、戚得郁、張秋霞、楊德倫

套件安裝

• Step 1: PyTorch (GPU 版) 參考連結

pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html

• Step 2: SimpleTransformers 參考連結

pip install simpletransformers

• Step 3: matplotlib

pip install matplotlib

• Optional: opencc 與 flask

pip install opencc flask

程式檔案簡介

• app.py: Flask 的 app.py
• s2t.py: 簡轉繁體 (透過 OpenCC)
• train.py: 訓練模型
• predict.py: 人工評估成效
• checkGPU.py: 確認電腦環境是否擁有 GPU
• plot.py: 基本 plot 輸出/檢視
• example.py: 測試語法用的程式檔
• check.py: 檢查關鍵字的數量
• convert.py: 將夥伴的整理好的資料，轉換成訓練資料
• web_scraper_kingnet.py: 取得 KingNet 網站的衛教資訊

NER 的 tokens

labels = [
    "O",
    "B-BODY","I-BODY",
    "B-CHEM","I-CHEM",
    "B-DISE","I-DISE",
    "B-DRUG","I-DRUG",
    "B-EXAM","I-EXAM",
    "B-INST","I-INST",
    "B-SUPP","I-SUPP",
    "B-SYMP","I-SYMP",
    "B-TIME","I-TIME",
    "B-TREAT","I-TREAT"
]

預訓練模型選擇

• HuggingFace
  • bert-base-chinese
  • hfl/chinese-macbert-base
  • ckiplab/albert-base-chinese-ner
  • hfl/rbt6

顯示卡型號

• 桌機 (主要訓練用)
  • GPU: 1080 Gaming 8G
• 筆電 (測試訓練程式用)
  • GPU: 2070 Max-Q Design

語料

• train.json、test.json
• ccks2017.json、ccks2018.json (由 秋霞、得郁 提供)

訓練設定

• batch size: 64
• epochs: 30

成果

• ACL Anthology: CrowNER at Rocling 2022 Shared Task: NER using MacBERT and Adversarial Training