Thread by @MarkNeumannnn

I have a lot of respect for Quoc - his research is extremely influential, creative and far reaching. However, both papers referenced here are cited in the ELMo paper. So this cannot be a complaint of academic attribution.(1/7)

https://twitter.com/quocleix/status/1098417363064508416

So what caused this difference in credit assignment? Why are the ELMo/BERT/OpenAI GPT more credited in current research?

The difference is, code, pre-trained models and open source software. (2/7)

You can use ELMo in TF OR Pytorch. You can download 5 different pre-trained models here: allennlp.org/elmo. You can learn how to apply it to your own use case here: github.com/allenai/allenn…
And you can reproduce the original training here: github.com/allenai/bilm-tf (3/7)

The same is true of BERT and OpenAI GPT (mainly due to @Thom_Wolf !).

It's definitely possible this is true of "Semi-Supervised Sequence Learning" too! There is code here:
github.com/tensorflow/mod…

(4/7)

However, this doesn't look like the original implementation, as the author of the code is not present on the paper, so perhaps it is a re-implementation? Happy to be corrected here. Perhaps the code was released at the time and I just can't find it online now.
(5/7)

I tried to install it and run the tests with the single "tensorflow>=1.3" requirement. The tests don't pass. I can't run any of the 3 main executable files. Which version of python does it use? I don't know. I tried with python 3.5. How do I download a pre-trained model?
(6/7)

So really, the lesson here is that being first is not the sole contributor to academic attribution. If you want people to cite and build on your research
1) release your code
2) release your models
3) make it silly easy to install
3) make it silly easy to use and extend.
(7/7)