Forward this message to friends and colleagues who want to stay up to date with research and news
from the Allen Institute for AI.  Subscribe here | Archive

AI2 NEWSLETTER | February 2021
In this edition: Teaching AI agents by playing hide-and-seek, TLDRs for academic papers, unveiling the hidden biases in question answering models with cleverly worded questions, revisiting the ultimate trajectory of rejected papers, and more.

Learning through interaction and play


A new paper from the AI2 PRIOR team recently accepted to the 2021 International Conference on Learning Representations (ICLR) introduces a novel paradigm for training AI agents: learning through interaction and play.

Instead of training the standard way with huge static collections of manually labeled images, we introduce AI agents that can learn visual representations by playing Cache, a variant of hide-and-seek, where one AI agent hides objects for another agent to find. We show that these representations can achieve surprisingly high performance, rivaling approaches developed using traditional datasets. (More in this article on the AI2 Blog from lead author Luca Weihs.)

Single-sentence paper summaries


Semantic Scholar recently launched a new feature that puts automatically generated single-sentence paper summaries right on the search results page, allowing you to quickly locate the right papers and spend your time reading what matters to you.
TLDRs (Too Long; Didn't Read) are super-short summaries of the main objective and results of a scientific paper generated using expert background knowledge and the latest GPT-3 style NLP techniques. This new feature is available in beta for nearly 10 million papers and counting in the computer science domain in Semantic Scholar.

See some examples and read testimonials from authors at tldr.semanticscholar.org.

More about this exciting new capability of Semantic Scholar:

"UnQovering" Stereotypical Biases in Question Answering Models


The web text used to train modern NLP systems is more or less unfiltered, and it contains all of the stereotypical associations and implicit biases one might expect to find in any huge and aimless collection of human writing.

In new work that recently appeared in Findings of EMNLP 2020, we ask the question:
To what extent are stereotypical biases present in question-answering models?

Read our blog post to learn more, and check out the UnQover demo to experience it for yourself.

What Open Data Tells Us About Rejected ICLR Submissions


Senior Research Engineer Mark Neumann writes:
"Having published a few papers in NLP venues such as ACL, I’ve often found myself frustrated with the reviewing process. It sometimes turns out to be noisy, with low-quality reviews, which is an unsatisfying result for what can often be a lot of hard work put into a paper. So I wondered: Is there actually a difference between papers that are accepted vs rejected? Was there a way to quantify that?"  Read more →

Featured AI2er: Chelsea Haupt


We interview Semantic Scholar Product Manager Chelsea Haupt about what she's been up to lately, what she's looking forward to building with her team, and more on the AI2 Blog.

A guide to language model sampling in AllenNLP


Learn about how Stochastic Beam Search can add ~creativity~ to your generated text, in this blog post from AI2 Intern on the AllenNLP team Jackson Stokes.

Using AI to Extract a Knowledge Base of COVID-19 Mechanisms


Can we leverage AI to help researchers navigate the eclectic and ever-growing landscape of scientific literature around COVID-19? What if there was a KB containing diverse and structured information on causal relations, methods, objectives, and activities — coming from any scientific area? Check out this blog post by AI2 researcher Tom Hope and try out the COVID-19 Mechanism Knowledge Base yourself. 


Congratulations to AI2's top interns of 2020!


We're thrilled to announce the 2020 winners of the AI2 Outstanding Intern of the Year award: Sarah Wiegreffe, Sean MacAvaney, and Unnat Jain. Thank you for your excellent contributions to AI2!
Twitter
LinkedIn
Instagram
Website
Copyright © 2021 The Allen Institute for Artificial Intelligence, All rights reserved.
For media inquiries please contact press@allenai.org
AI2 Newsletter Archive