Albert Xu

PhD Student @USC, @nlp_usc. Prev. @UCBerkeley, @BerkeleyNLP

[Albert Xu @albertxu__]

We built an AI system for solving crossword puzzles that outperforms the world’s best humans.

It works by combining deep learning with traditional probabilistic search and inference methods.

ACL ‘22 Paper: https://t.co/RR8MTjOn4f
Demo: https://t.co/zYkQir3fx5

👇[1/8]

[Jacob Bremerman @JLBrem]

me(俺), me(僕), or me(私)? What makes translation especially fascinating is the difference between types of nuance captured in one language (but not another). In our #EMNLP2022 paper, we investigate this in MT. @CSatUSC @cutelabname_nlp @nlp_usc

[Bill Yuchen Lin @billyuchenlin]

Multi-task prompted LMs (e.g., T0) are powerful in unsupervised task generalization. Can we further enhance them via *retrieval augmentation*? To this end, we proposed a simple yet effective method named ReCross in #NeurIPS2022! ✨ [1/6]

📜: https://t.co/7urLFQjhKC

[Albert Xu @albertxu__]

@arXiv_Daily Interesting paper! But this is a different Albert Xu (I am not the author)

[Albert Xu @albertxu__]

@pcastr Thank you! Just read the post: it’s a great summary of why crosswords require different ideas in combination. e.g., it turns out that extractive QA models don’t do well alone since you can’t find many answers in corpora. Saw your puzzle before, joked that wr should feed it in :)

[Albert Xu @albertxu__]

We built an AI system for solving crossword puzzles that outperforms the world’s best humans.

It works by combining deep learning with traditional probabilistic search and inference methods.

ACL ‘22 Paper: https://t.co/RR8MTjOn4f
Demo: https://t.co/zYkQir3fx5

👇[1/8]

[Albert Xu @albertxu__]

@dat1randomG Thanks for the kind words! The demo is using cached states, so it runs however fast you want. In practice, solving a puzzle is on the order of a minute or so.

[Albert Xu @albertxu__]

#ACL2022 Paper: https://t.co/RR8MTjOn4f
Demo: https://t.co/zYkQir3fx5
Code: https://t.co/Foh6Rswt4Y
[8/8]

[Albert Xu @albertxu__]

See our paper for more details, especially on the remaining challenges left in crossword puzzles!

With @Eric_Wallace_, @NickATomlin, @albertxu__, @KevinYa33964384, @eshaan_pathak, Matt Ginsberg, and Dan Klein. [7/8]

[Albert Xu @albertxu__]

@janewakefield @BerkeleyNLP Revising the above, you can directly reach out to our advisor Dan ([email protected]). He'd be happy to chat today or tomorrow!

[Albert Xu @albertxu__]

@janewakefield @BerkeleyNLP Could you DM me with your email address? Would be happy to put you in touch with the group!

[Albert Xu @albertxu__]

@user2462120195 Ouch, I don’t think I want to disclose.

[Albert Xu @albertxu__]

We won the biggest crossword tournament this weekend! Learned so many fun tricks about crosswords while working on this project. Stay tuned for a paper!

[Berkeley Engineering @Cal_Engineer]

AI systems can now solve crosswords better than humans. The Berkeley Crossword Solver from @BerkeleyNLP helped propel Dr.Fill, an AI program created by @mattlginsberg, to 1st at #ACPT2021, the premier crossword tournament, solving the playoff puzzle in 49 seconds with no errors.