@sho_yokoi_ Profile picture

Sho Yokoi

@sho_yokoi_

Researcher at #NLProc team @NlpTohoku & at #statistics team @RIKEN_AIP_EN | interested in geometry of word embedding space and optimal transport | Ja:@sho_yokoi

Similar User
Hiroki Ouchi photo

@blankeyelephant

YANS photo

@yans_official

Tohoku NLP Group photo

@tohoku_nlp

NLPコロキウム photo

@nlp_colloquium

Reina Akama photo

@jqk09a

Takuma Udagawa photo

@futsaludy

Sho Takase photo

@shot4410

Goro Kobayashi photo

@goro_koba

Ryoko TOKUHISA / 徳久良子 photo

@rtokuhisa

Esin Durmus photo

@esindurmusnlp

梶原智之 photo

@moguranosenshi

Yuji NARAKI photo

@yuji_research

hiroshi matsuda photo

@hmtd223

Kaori Abe photo

@KaoriAbe11

Shun Kiyono photo

@shunkiyono

Sho Yokoi Reposted

This is a good piece of work (finally online): arxiv.org/abs/2411.00680 Word embeddings typically assume uniform word frequencies, but in reality, they follow Zipf’s law. Simply applying PCA whitening weighted by empirical word frequencies can improve task performance.


Sho Yokoi Reposted

Zipfian Whitening ift.tt/lBr1gDA


Sho Yokoi Reposted

Interested in word frequency / geometry of language representations / contrastive learning? Check out our paper “Zipfian Whitening” at #NerIPS2024! See you all in Vancouver! 🇨🇦

Our paper “Zipfian Whitening” got accepted to #NeurIPS2024🎉 Joint work w/ amazing @levelfour_ @hiroto_kurita (@tohoku_nlp) & @hshimodaira! A simple idea—consider word frequency when taking expectations—yields rich empirical/theoretical insights into language representation🔬

Tweet Image 1
Tweet Image 2
Tweet Image 3
Tweet Image 4


Our paper “Zipfian Whitening” got accepted to #NeurIPS2024🎉 Joint work w/ amazing @levelfour_ @hiroto_kurita (@tohoku_nlp) & @hshimodaira! A simple idea—consider word frequency when taking expectations—yields rich empirical/theoretical insights into language representation🔬

Tweet Image 1
Tweet Image 2
Tweet Image 3
Tweet Image 4

Sho Yokoi Reposted

[CL] Analyzing Feed-Forward Blocks in Transformers through the Lens of Attention Map G Kobayashi, T Kuribayashi, S Yokoi, K Inui [Tohoku University & MBZUAI] (2023) arxiv.org/abs/2302.00456 - Transformers have become ubiquitous in NLP tasks, so interpreting their internals is…

Tweet Image 1
Tweet Image 2
Tweet Image 3
Tweet Image 4

Sho Yokoi Reposted

Happy to share our paper on LM’s Feed-forward network (FF) analysis has been accepted as an #ICLR2024 spotlight! 💡FF boosts attention between words forming compound nouns, named entities, etc. 💡FF and LayerNorm cancel out each other’s effects 📄arxiv.org/abs/2302.00456

Tweet Image 1

Sho Yokoi Reposted

[ARR Co-CTO here] Quick correction - our policy is based on your arXiv submission time, which we acknowledge will sometimes mean the paper appears on arXiv inside the anonymity window (see the FAQ on the ARR author page for details: aclrollingreview.org/authors#faq)


Sho Yokoi Reposted

The Tohoku NLP Lab will present nine works at #EMNLP2023. See you in Singapore! #NLProc nlp.ecei.tohoku.ac.jp

Tweet Image 1
Tweet Image 2

Sho Yokoi Reposted

It's now on arXiv! arxiv.org/abs/2310.15921

Tweet Image 1

Thrilled to announce my first paper, “Contrastive Learning for Sentence Encoder Induces Word Weighting by Information-Theoretic Quantities” got accepted at #EMNLP2023 findings! Our work connects sentence encoders with information theory. Will upload the paper to arXiv soon! 💨

Tweet Image 1


Sho Yokoi Reposted

Thrilled to announce my first paper, “Contrastive Learning for Sentence Encoder Induces Word Weighting by Information-Theoretic Quantities” got accepted at #EMNLP2023 findings! Our work connects sentence encoders with information theory. Will upload the paper to arXiv soon! 💨

Tweet Image 1

Sho Yokoi Reposted

Three full papers and four findings papers have been accepted to #EMNLP2023 ! #NLProc nlp.ecei.tohoku.ac.jp/news-release/9…


Sho Yokoi Reposted

Cool idea! We actually observed that BERT tries to turn off attention by shrinking the value vector of [CLS] [SEP] punctuations and paying unnecessary attention there: arxiv.org/abs/2004.10102 I'd like to see his simpler trick will make Transformer better.

Tweet Image 1

I hit a bug in the Attention formula that’s been overlooked for 8+ years. All Transformer models (GPT, LLaMA, etc) are affected. Researchers isolated the bug last month – but they missed a simple solution… Why LLM designers should stop using Softmax 👇 evanmiller.org/attention-is-o…



Sho Yokoi Reposted

Introducing OTAlign, optimal transport (OT) based monolingual word alignment. We established a connection between the family of OT and monolingual word alignment concerning the null alignment ratio. (1/3) 📄arxiv.org/abs/2306.04116 w/ @levelfour_ and @sho_yokoi_ #ACL2023NLP


Sho Yokoi Reposted

How the “prediction head” works in LMs? Our #ACL2023 Findings paper shows this module at the end of LMs adjusts word frequency in prediction 📊 (bonus) This property can easily be used to promote more diverse outputs of LMs! 📜 arxiv.org/abs/2305.18294

Tweet Image 1

Sho Yokoi Reposted

Monolingual alignment, anyone? Our #ACL2023NLP paper "Unbalanced Optimal Transport for Unbalanced Word Alignment" w/ @levelfour_ and @sho_yokoi shows the family of optimal transport are natural and powerful tools for this problem. More information coming soon. #NLProc


Sho Yokoi Reposted

Hey! We've uploaded two pre-prints! One focuses on the intersection of vision, language, and cognitive science. arxiv.org/abs/2302.00667 The other examines the role of feedforward in Transformers. arxiv.org/abs/2302.00456 These are still **in progress**. Stay tuned for updates!

Tweet Image 1
Tweet Image 2

The separation of “(recognizing) textual entailment” and “natural language inference” in the peer-review field selection form for ACL 2023 is disturbing.


Sho Yokoi Reposted

[CL] Subspace-based Set Operations on a Pre-trained Word Embedding Space Y Ishibashi, S Yokoi, K Sudoh, S Nakamura [NAIST & Tohoku University] (2022) arxiv.org/abs/2210.13034 #MachineLearning #ML #AI #NLP #NLProc

Tweet Image 1
Tweet Image 2

Sho Yokoi Reposted

One paper entitled "Word Tour: One-dimensional Word Embeddings via the Traveling Salesman Problem" has been accepted to #NAACL2022 (main conference, short paper) 🎉

Tweet Image 1

Sho Yokoi Reposted

Here is the slide of my today's talk at #NeurIPS Meetup Japan 2021. slideshare.net/SatoshiHara3/e…


Loading...

Something went wrong.


Something went wrong.