Notice
Recent Posts
Recent Comments
Link
관리 메뉴

one by one ◼◻◼◻

[NLP논문리뷰]How Much Knowledge Can You Pack Into the Parameters of a Language Model? 본문

논문리뷰

[NLP논문리뷰]How Much Knowledge Can You Pack Into the Parameters of a Language Model?

JihyunLee 2021. 9. 12. 23:43

제목 : HowMuchKnowledge Can You Pack Into the Parameters of a Language Model?

저자 : Adam Roberts, Colin Raffel, Noam Shazeer

발행년도 : 2020

paper : https://arxiv.org/abs/2002.08910

code : https://github.com/google-research/google-research/tree/master/t5_closed_book_qa

 

Review

QA모델에 대한 논문 중 google에서 출간한 논문.

기존의 QA task는 passage를 보여주고 이에 대한 답을 찾아가는 방식이었다. 그러나 이 논문은 passage를 보여주지 않고 질문 부터 한다.

즉 다양한 task에 대해 학습한 T5에게 "루즈벨트 대통령이 태어난 해는?" 이라는 질문을 한다면, T5는 위키피디아의 지식 없이도 1882이라고 답을 하는지 실험해 본 논문이다.

논문에서는 Passage를 보여주고 Passage안에서 답을 찾도록 한 방식을 Open book test, passage를 보여주지 않고 답을 찾도록 한 방식을 Closed book test라고 비유했다!

Experimetns

실험에는 Natural Question, WebQuestions, TriviaQA 데이터를 사용하였고, passage는 사용하지 않고, 질문-답만 사용해서 학습시켰다.

Result

여기서 T5.1.1 은 unlabeled data only로 학습된 T5를 의미한다. 놀랍게도, passage를 보여주지 않았음에도 SOTA의 성능을 달성한 Task들이 있었으며, 다른 데이터 셋에 대한 성능도 SOTA에서 크게 뒤쳐지지 않았다.

 

Comments