Trouble Netizen

A simple chinese QA chatbot implement with pytorch and transformer trained by PTT data
簡易 ptt gossiping QA chatbot 使用 pytorch framework 和 transformers

Tutorial:
https://github.com/1tangerine1day/bert2bert_QA_chatbot/blob/master/QA_chatbot_step_by_step.ipynb

Demo for using:
https://github.com/1tangerine1day/bert2bert_QA_chatbot/blob/master/How_to_use.ipynb

Model:

picture from this
Instead of AraBERT, I use normal bert with pretrain model "bert-base-chinese"

Data:
https://github.com/zake7749/Gossiping-Chinese-Corpus

My trained weight (best.pt):
https://drive.google.com/file/d/1qc5uq5_Uag1XROoxbpruWZGzR3bhTyT-/view?usp=sharing

Training

• batch: 50
• learning rate: 1e-5
• optimizer: Adam
• epochs: 20
  

load model

from bert2bert import QA_chatbot

cahtbot = QA_chatbot()
cahtbot.load_model("best.pt")

predict

test case 1 (恩...健康 = 股價)

cahtbot.predict("加權指數狂跌小台卻狂拉")

'台灣人的健康就是這樣'

test case 2 (有點兇 XD)

cahtbot.predict("校園霸凌嚴重 立委兒子生殖器被踹到流血")

'，這種人真的很可憐，不要出來丟人現眼'

test case 3 (竟然有接到!)

cahtbot.predict("cosplay機師有什麼搞頭嗎？")

'我想看看'

test case 4 (某種雙關?)

cahtbot.predict("工程師月薪多少")

'，台灣的公司薪水不是一般人的一般人的薪'