Don't roll! How to reproduce a paper with high quality?

Hello everyone , My name is Aaron.

Recently, I have been frequently asked about , How to reproduce a paper with high quality , Let me simply write down some of my thoughts , At present I am doing NLP Some things in the field , Also encountered some problems of reproducing the paper , I think the following points need attention .

First , The premise of reproducing the algorithm is to understand the algorithm , This requires an in-depth understanding of the paper itself , Understand what problem the author proposes this algorithm to solve ？ Compared with previous algorithms in the same field, what are the improvements ？ Why these improvements are needed ？ And how to achieve these improvements ？ ad locum , In addition to a deep understanding of the thesis itself , On the core part of the thesis reference We should also have a more in-depth reading .

Then you can reproduce the code , According to my own understanding of the paper , Iterative recurrence algorithm , The results are compared with those given by the author . If the results cannot be reproduced , Please go back to the previous step , Find some of the tricks Continue iteration . Still, if you can't , Please email the author with your test results .

Now you want to reproduce a classic thesis quickly ？

For example, use 3 Days time to reproduce an article NLP baseline The paper ？

I highly recommend you to listen to this class today , I'm going to learn by myself , Benefited greatly .

zhi

straight

bo

seeding

yu

pre

gao

Sue

6 month 29 Japan -6 month 30 Japan 、NLP Baseline Paper Experts Albert, Live sharing —— be the speaker Word2Vec as well as ELMO,BERT.

Sweep code payment 0.1 Yuan booking live broadcast

Courseware will be provided after the live broadcast & Code data sets

One 、Word2Vec as well as ELMO,BERT extensive reading （6 month 29 Japan 20:00）

1. The development and direction of naturallanguageprocessing

2. Technological evolution path

3.Word2vec Background and significance of the thesis

4.ELMO Background significance

5.BERT Background significance

6. Learning path recommendation

Two 、Word2Vec as well as ELMO,BERT intensive reading （6 month 30 Japan 20:00）

1. Overview of the pre training language model

2.ELMo Language model

3.BERT Language model

4. Learning path recommendation

Sweep code payment 0.1 Yuan booking live broadcast

Courseware will be provided after the live broadcast & Code data sets