🚨 How much exactly are two languages similar?
…and how can language models use that information?
We propose first actually usable language similarity score: qWALS. Language models can use it to get better results even with ZERO data in target language❗️
www.mdpi.com/2504-4990/8/...
Official transcription is スクウォドフスカ, but let’s not be too finicky about this. Katakana for names is arbitrary. 😉
Polish here. Actually, it should be メリ・スクオドフスカ・クリ 🤷🏻♂️😂
With friends from Samurai Labs we have just released a new study about two types of abusers on the Internet and how to handle them. The paper is free to download!
#CyberBullying #Hate #OnlineSafety
authors.elsevier.com/c/1line15hYd...
⚡️今年も北海道NLPミーティングを開催します．自然言語処理関連の研究をさており，北海道周辺に活動し，または北海道に心を残した方がぜひ投稿をご検討ください．❤️
You can publish in journals for free/with subscription and then just release your preedited version on arxiv and you’ll have best of both worlds. 💁🏻♂️
Why do they want this?
You have a choice. You can freely choose to send all of your research directly to arxiv and not use any peer-review venue ever again. Maybe start by asking - why won’t you do it yourself?
What’s the world's largest wooden structure on Wednesday? 🤔
Not to brag too much, but this is one of the most innovative and fun little studies we've done recently.
"Token and part-of-speech fusion for pretraining of transformers with application in automatic cyberbullying detection"
www.sciencedirect.com/science/arti...
“no gaffes, no policy missteps” - dear god, imagine if this is the actual bar for political relations these days.
Let’s delve into that. 😄
With pleasure! I might even be able to find the necessary number of people.
A very good research! I wonder how this would replicate for other languages.
Hi Brandon, good to see you hear as well. 🙂
Reposting from @acm-sigchi.bsky.social : 📢 We’re thrilled to announce that registration for #CHI2025 is officially OPEN! 🎉
🌍 Early registration ends March 31, 2025 (EOD AOE) – don’t miss out!
💻 In-person & virtual participation options are available.
🔗 Register now: buff.ly/4gh539Y
Georgia on My Mind
Dear, @bsky.app , please, implement bookmarks right away.
“Many of the instances that are currently called multi-agent systems might as well be described as modular single-agent systems, or as multi-expert systems.” This basically describes the whole agent-based research field.
最初は自己PRですみませんが，PhD学生がいい研究をしたのでシェアをせざるを得ない🤩
bsky.app/profile/ptas...
日本で自然言語処理をしている方，ぜひリストにご参加ください．
bsky.app/profile/did:...
👍
👍
So - Mori was wrong in his "The Uncanny Valley" paper. It seems he was imagining things from his own perspective, rather than taking into account the progress of technology and society.
But, on the bright side - your next Siri will be more naturally emotional. ☺️
And if you think of it, it's actually quite obvious. Technology is always developing, so by the time there actually are life-like humanoid robots to use at home, we will be well acquainted with them not to feel uncanny with them (although spoiler - its pure sci-fi AI-robot-hype).
This also confirms my own theory.
It shows that when people obtain a new technology, (1) they already know the technology a bit and (2) prefer to smoothly adopt it, focusing on its overall utility rather than minor imperfections.
Thus - the theory of Uncanny Valley is incorrect.
Whats more, you can just train a voice model from scratch only on artificial data - it will be worse, but still perfectly usable (!).
This result for Japanese language holds up for basically all age groups, genders, nationalities, even for people who don't know Japanese (!).
You do not need to collect huge high quality datasets for each emotion to produce a good emotional voice. Just use a raw voice model and fine-tune it for specific emotion(s) on small emotional data - even artificial data (!). The emotions will still be perfectly understandable.
🚨 The Uncanny Valley doesn't exist. 🚨
Mujahid Khalifah in his PhD research just showed that Japanese emotional speech models created from artificial voice datasets match original data in emotional expression.
What are the implications? 🧵
doi.org/10.1109/ACCE...
