Daoud Clarke

AI/ML Engineer at and part-time #NLProc researcher.

[Daoud Clarke @daarkecloud]

A few thoughts on "Prompt Repetition Improves Non-Reasoning LLMs" by @yanivle, @matankal and @ymatias

Clearly we shouldn't need to repeat prompts to get this improvement in performance, so is there a better way?

https://t.co/XqBXD9I6al

[Daoud Clarke @daarkecloud]

So glad to hear @ToniaAntoniazzi is supporting the @gbsumudflotilla "I’ve written to the Foreign Secretary expressing concern...

"The letter urges the UK government to ensure the safety of those on board and to defend their right to engage in non-violent humanitarian action."

[mwmbl @mwmblorg]

Why the world needs a non-profit search engine? https://t.co/ElMphpsnDK written by @daarkecloud

[Daoud Clarke @daarkecloud]

I feel like it's harder to get phrase vectors from tools like FastText etc than it should be in 2021. Gensim does an ok job of joining words together to get phrases as a preprocessing step, but something about it doesn't feel right.

[Silvia Chiappa @csilviavr]

A position is still available. For informal queries, please reach out to me by email.
Retweets appreciated!

[Daoud Clarke @daarkecloud]

@alephic2 @swarat @gregd_nlp This is really cool. I love that you can learn so much from a single sentence using the template.

[Daoud Clarke @daarkecloud]

@jonty I had stick insects when I was a teenager. This problem doesn't get any better! I ended up releasing them into the wild I think...

[Daoud Clarke @daarkecloud]

Biases in our models can cause problems for NLG in e.g. chatbots or translation. Here's a way to mitigate such biases.

From the paper: "What was Shanice known for?"
Before: "Fighting people?"
After: "She’s a professor at MIT and she was a
professor at NYU."

[Emily Sheng @ewsheng]

Excited to finally share our work “Towards Controllable Biases in Language Generation” (https://t.co/Y7TbcSOsbX), to appear in Findings of #emnlp2020, and done with @kaiwei_chang, Prem Natarajan, and @VioletNPeng :)

[Naomi Saphra @nsaphra]

"In service to their sociopolitical agenda, they established the statistician as an authority figure, a numerical referee who is by nature impartial, they claimed, since statistical analysis is just unbiased number-crunching." https://t.co/1379yof6PN

[Daoud Clarke @daarkecloud]

@bothness Yes, would also be interesting to see combinations of "Islamophobia" with the two parties...

[Daoud Clarke @daarkecloud]

@jonty I thought it was a law of Twitter that the best tweets have tpyos

[Daoud Clarke @daarkecloud]

@bothness Nice! Does it ever end? It's a bit addictive, so I'd like to know how far I have to go before I get a final score...

[David R. MacIver @DRMacIver]

A software development war story about what can happen when you keep solving the problem in front of you rather than taking a step back and going "Wait, absolutely none of this is necessary" https://t.co/YsvFJc8Cgl

[John Platt @johnplattml]

Interesting paper by @jurafsky et al. that estimates statistical power of #nlproc experiments: https://t.co/ZUPxs8zfwl Some benchmark datasets are too small to be useful.

[Daoud Clarke @daarkecloud]

@MikaelKoutero @kdnuggets Thanks! You can have VSCode. Anything's better than Jupyter notebooks 😜

[Daoud Clarke @daarkecloud]

@dirk_hovy Intellectual laziness. Coming from a physics degree I was astounded that this was normal. It's also not just #NLProc but ML and data science in general.

Minor point: your student probably mean standard *error* rather than standard deviation.

[Daoud Clarke @daarkecloud]

@emnlp2020 what time of day/timezone will presentations be in? so that I can prepare in advance for possible late nights/early mornings