Language Models are Unsupervised Multitask Learners (GPT-2)

OpenAI 2019 paper - slide

Posted by Jexus on March 4, 2019

Recently by the same author:


美國 EECS 博士班申請經驗分享 (ML/DL/NLP/Speech)

2021 Fall NLP/Speech PhD Application

You may find interesting:


My Work - Lifelong Language Knowledge Distillation

EMNLP 2020 long paper


My Work - Dual Inference for Improving Language Understanding and Generation

EMNLP 2020 findings paper

Language Models are Unsupervised Multitask Learners (GPT-2)

Paper Link

TL;DR:

用 Transformer-based (left-to-right) language model 在極大語料 (40GB) 上做訓練,並將所有 NLP 問題轉化成 language model 問題來解,是亂做一通,但也在一些 task (CoQA) 上有著還可以的表現,而另也有生成高品質文章的用途。

  • 可於此 blog 觀賞其所產生之優美獨角獸文章:https://openai.com/blog/better-language-models/

  • 可於此網站線上享用 GPT-2 的續寫功能:https://talktotransformer.com/

  • 可於此網站使用有 GPT-2 幫你續寫的文字編輯器:https://transformer.huggingface.co/

  • 可於此網站安裝 GPT-2 所搭建的 code 補全外掛程式:https://tabnine.com/

Slide:

Please wait a minute for the embedded frame to be displayed. Reading it on a computer screen is better.