Ordered Neurons - Integrating Tree Structures into Recurrent Neural Networks​

ICLR 2019 (Best Paper Award) paper - slide

Posted by Jexus on March 26, 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

Ordered Neurons: Integrating Tree Structures into Recurrent Neural Networks​

Paper Link

TL;DR

利用嚴格遞增及嚴格遞減的條件去限制 LSTM 的 forget gate 及 input gate,讓 model 表現變好,並且在 training 完之後可以根據 forget gate 遺忘程度的多寡來找出藏在句子中的 Tree Structure,做到Unsupervised Syntactic Parsing。

此 paper 於 ICLR 2019 獲 Best Paper Award。

Slide:

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

關於這篇提出的嚴格遞增及嚴格遞減的 gate 條件,是否真的能讓 model 表現變好,有討論於此:
「如何评价ICLR 2019 best paper: Ordered Neurons ?」 https://www.zhihu.com/question/323190069

很可能是前人實驗 report 了較低的分數造成。

更新:我的完整講解可至 https://youtu.be/YIuBHB9Ejok 觀看。