Applied Scientist at Amazon Alexa. Opinions my own :)
Seattle, WA
Joined August 2015
- Tweets150
- Following490
- Followers115
- Likes401
Excited to share that our work on Failure point isolation in Large scale conversational assistants got accepted to NAACL as an oral talk (Industry Track). #alexa @AmazonScience
amazon.science/publications/…
Rakesh Chada retweeted
Check out the 51-language dataset for natural language understanding we released today amazon.science/blog/amazon-r… @AmazonScience
We’re giving you a CHANCE TO WIN these one-of-a-kind GeForce builds.
To celebrate The Matrix Resurrections we partnered with @wbpictures and challenged modders to create their own custom PCs inspired by The Matrix 👀
Entering is easy:
🟢 RT this post
🟢 Comment #TheMatrix
Rakesh Chada retweeted
Amazon’s 23 papers at EMNLP 2021 (@emnlpmeeting) - Natural-language understanding and question answering are areas of focus, with additional topics ranging from self-learning to text summarization. amazon.science/blog/amazons-…
Session Live now. I am in gather town poster room 2 (#1968). @emnlpmeeting #EMNLP2021
How to build fast and cost-efficient few-shot question answering systems? Please join us tomorrow at #EMNLP2021 virtual poster session (8:30-10:30 am PST/12:30-2:30pm AST) to learn and discuss more. @emnlpmeeting . Paper: arxiv.org/abs/2109.01951
How to build fast and cost-efficient few-shot question answering systems? Please join us tomorrow at #EMNLP2021 virtual poster session (8:30-10:30 am PST/12:30-2:30pm AST) to learn and discuss more. @emnlpmeeting . Paper: arxiv.org/abs/2109.01951
Rakesh Chada retweeted
Lucky to have @naoya_in working with us at @stonybrooknlp. He has been a postdoc in the @AI_SBU during this pandemic year and has contributed to bunch of interesting projects. Can't wait for the others in his pipeline to come out. He is fantastic. Hire him!
How can we get QA explanations that are concise? Ones that are sufficient, minimal and comprehensible! Check out SuQA which is to appear in #EMNLP2021.
arxiv.org/abs/2109.06853 (code coming soon)
Joint work w/ @harsh3vedi, Steven Sinha, @b_niranjan, @inuikentaro
Thread👇
We experiment with several input-output designs to learn what works best. More details regarding this in the paper (amazon.science/publications/…). Thanks for reading 😀 I'm looking forward to seeing this applied to other settings.
The generic design of the framework allows us to test it in a multitude of settings. We evaluate BART-Large, T5-base and also test in a multilingual setting (TydiQA). We see consistent improvements in all settings.
1
After only a few hundred training updates using 16 training examples, this framework gets ~55.5 F1 on SQuAD using a modest sized BART-base model.
1
A key feature of this framework is that it only requires modifications to the fine-tuning framework instead of the pre-training (in contrast to most other works such as Splinter (arxiv.org/abs/2101.00438)). This significantly improves the training time and reduces costs.
1
We show that the standard span selection objective does poorly in a few-shot setting and that aligning the fine-tuning with pre-training framework shows significant improvements
1
We propose a simple & cost-effective fine-tuning method that requires just modifying the input-outputs and using the same pre-training denoising objective for the QA task.
1
Delighted to share that our few-shot question answering (QA) framework that leverages pre-trained text-to-text models got into #emnlp2021 main conference. Details in the 🧵 below:
@emnlpmeeting #AmazonScience
amazon.science/publications/…
Happy to share that our work on Error Detection in Large-scale Natural Language Understanding systems using Transformer models got accepted at #ACL2021NLP Findings. #AmazonScience @AmazonScience
amazon.science/publications/…
Great post with a neat underlying message that you can take a not 100% accurate system (NMT) and still add tons of value to an use-case (meeting scheduling).
This quoted post is unavailable.
1
Rakesh Chada retweeted
This is mind blowing.
With GPT-3, I built a layout generator where you just describe any layout you want, and it generates the JSX code for you.
W H A T
Thanks @ErikaVaris for the detailed thread. Please do tune-in to the Live Q&A tomorrow 15:30-00:00 GMT/8:30-17:00 PDT if you're interested. virtual.acl2020.org/workshop…
This Post is from an account that no longer exists. Learn more