Jose Camacho Collados

Professor @cardiffuni (@Cardiff_NLP).
AI/NLP researcher and chess International Master ♟️🎾

[Jose Camacho Collados @CamachoCollados]

Our @Cardiff_NLP team and collaborators have been releasing amazing NLP models and datasets over the last few years. We've now organised the content as @huggingface collections 🤗

A short thread summarising some of these NLP resources, in particular related to social media! 🧵

[Jose Camacho Collados @CamachoCollados]

This was an interesting and fun project, led by @joseba_fdl.

The findings are quite intriguing as well, and probably not what one would immediately expect! 👇

[joseba.fdl @joseba_fdl]

🚨 New paper alert!

🇯🇵🤔 Why are all LLMs Obsessed with Japanese Culture? On the Hidden Cultural and Regional Biases of LLMs 🇯🇵🤔

👉 We find that LLMs show strong regional preferences, often disproportionately favoring certain countries, such as Japan. 🇯🇵

[AIDB @ai_database]

「なぜLLMは、日本文化に執着するのか？」という意外な論文が出ています。
研究者らの検証によると、Claudeなど主要LLMの出力はなぜか皆、日本文化に偏っているとのこと。

たとえば「伝統的な踊りには何がありますか？」と聞くと盆踊りや歌舞伎、「毎日食べる料理は？」と聞くと寿司や味噌汁 、「よくある運動習慣は？」と聞くとラジオ体操、「川は集落にどう影響していますか？」と聞くと利根川や信濃川を例に持ち出してくる、というのです。

また、日本に続くのは米国やインド、中国、フランス。それ以外の国はほとんど登場しない、という偏りが見えたとのことです。

なお、事前学習では均等なのに教師あり微調整の後で急に偏りが噴出することが判明しています。
LLMは西洋中心とも言われてきた裏で、ふるまいはその通りでもないようです。

※この現象は自分の言語圏を除いた場合の話です。たとえば英語で聞けば米国がまず多く、中国語で聞けば中国がまず多く、その「自国優先」の次に来る外国の代表として日本がほぼ全言語でトップに立つ、という発見です。

[Jeremy Nguyen ✍🏼 🚢 @JeremyNguyenPhD]

Everyone thinks AI is culturally biased towards the West.

But this new study finds evidence that LLMs show a clear cultural bias towards countries like Japan.

Link the in the replies below:

[Jose Camacho Collados @CamachoCollados]

@PHChess @GMJacobAagaard Hopefully we can improve these models soon or at least they become more transparent and self-aware of their own limitations - currently working on this but definitely not easy!

Reference for anyone interested: https://t.co/BKfBSpy1Dn

[Jose Camacho Collados @CamachoCollados]

@PHChess @GMJacobAagaard And this was an easy case where all the data was provided to the LLM and they were asked to perform very basic counting operations.

In the draw analysis case, the model also had to collect the data itself, etc. which may introduce further errors.

[Jose Camacho Collados @CamachoCollados]

See you tomorrow at the WASSA workshop! 👋

I will be talking about social media from an interdisciplinary perspective, and lessons learned on how we NLP'ers can contribute to other disciplines (e.g. social sciences) more effectively.

[WASSA 2026 @wassa_ws]

⏳ Only two days until #WASSA2026 on March 29 at #EACL2026!

📜 The program is now available. See the full schedule, including the invited talk by @CamachoCollados:
“Social Media Analysis in the Language Model Era: An Interdisciplinary Perspective”

🔗 https://t.co/fHTjCBRHu8

[Jose Camacho Collados @CamachoCollados]

@petergostev This is very cool, thanks for sharing! 🙏

May I ask which LLM did you use to co-create the questions? Sorry if the info is there somewhere, couldn't find it!

[Jose Camacho Collados @CamachoCollados]

The dates for the 5th Cardiff NLP Workshop are now confirmed (22-23 June) - join us! 🤗

[Cardiff NLP @Cardiff_NLP]

Pleased to announce the 5th Cardiff #NLProc Workshop!
📍Cardiff 📅 22–23 June 2026
🌐 More information (to be regularly updated): https://t.co/1NDGa2Xpb0
👉 If you’re interested in attending, please complete the EoI by 11 April: https://t.co/wyWg6Fkhw5
Registration is ✨free✨1/

[Jose Camacho Collados @CamachoCollados]

@DaveFuertes Indeed, I hadn't even been to Wales in 2017 😅 (arrived for the first time in 2018 and transferred my federation in 2022 if I recall correctly)

[Jose Camacho Collados @CamachoCollados]

3️⃣ AI and Data Science for Electoral Integrity and Democratic Governance (Federico Liberatore)

4️⃣ Designing Benchmarks that Reflect Real-World NLP Use (@feralvam)

[Jose Camacho Collados @CamachoCollados]

@Cardiff_NLP is hiring PhD students!

🇬🇧🇪🇺 Full PhD scholarship open to UK/EU students
⏲️ Application deadline: February 13th

List of projects and supervisors below (find them in FindAPhD for more details) 👇

[Jose Camacho Collados @CamachoCollados]

1️⃣ Interpretability-Guided Compression and Acceleration of Large Language Models (@tpilehvar)

2️⃣ Calibrated Multimodal Verification to Prevent Hallucination in Robot Planning and Manipulation (@jodieyzhou)

[Jose Camacho Collados @CamachoCollados]

First interaction I'm having with GPT 5.2. It isn't going too well...

"Draw 6 people across a table"

[Jose Camacho Collados @CamachoCollados]

@PMinervini @WenhuChen 5-month summer break sounds amazing! I will try to apply for one of those tenure jobs

[Jose Camacho Collados @CamachoCollados]

@JBlackburnChess @4NCL What a tournament by Indy, great level throughout! 👏🏼👏🏼

[Jose Camacho Collados @CamachoCollados]

Some good suggestions on how to make your research useful for others (accessible/memorable/usable). For me, this has always been a major driver of how I approach research.

✅ As a bonus, this will likely increase your citation count!

[Jia-Bin Huang @jbhuang0604]

How to get citations?

[Jose Camacho Collados @CamachoCollados]

👉 Our #EMNLP2025 paper: https://t.co/1bjbd5SEKw

🗞️ The Guardian article: https://t.co/GIZMuDY8Kj

[Jose Camacho Collados @CamachoCollados]

Our latest work was featured in The Guardian, both in the tech and comedy sections! 🤣

LLMs are good at recognising puns, but they often think everything is a pun, as long as it looks like one. They just try too hard, and the reasons provided are often deep and hilarious! 👇