Who's Who of AI

MilaNLP Lab

78 trust researcher @milanlp.bsky.social · 498 followers
The Milan Natural Language Processing Group #NLProc #AI milanlproc.github.io
What they're sharing

Articles & links

For today's reading group, @marlutz.bsky.social presented "State media control influences large language models" by Waight et al. (2026) Paper: www.nature.com/articles/s41... #NLProc

State media control influences large language models | Nature nature.com
AI Weekly's analysis →
  • Chinese state-media content appears in typical LLM training sets at roughly 41 times the rate of Chinese-language Wikipedia.
  • Across 37 countries, models prompted in the local language produce more regime-favorable responses in countries with lower press freedom.
  • A pretraining experiment with just 6,400 state-scripted documents pushed an open-weight model to pro-government responses nearly 80 percent of the time.
Read full analysis →
View on Bluesky · ♥ 6 ↻ 3 ↩ 0 · 7 from the directory shared this · 13d ago

For today's reading group, @esradonmez.bsky.social presented "RLHF May Not Reflect Genuine Preferences" by Ghafouri et al. (2026). Interesting thoughts on whether annotations are actually real preferences! Paper: arxiv.org/abs/2604.03238 #NLProc #RLHF

RLHF May Not Reflect Genuine Preferences arxiv.org
AI Weekly's analysis →
  • A new arxiv preprint argues RLHF annotator responses may not represent genuine preferences at all, but responses constructed on the spot.
  • Filtering high-inconsistency annotators in two RLHF datasets flipped majority harm classifications for 18.6% of prompts.
  • The same filtering shifted mean ratings by more than 13 points on a 100-point scale, suggesting systematic rather than random noise.
Read full analysis →
View on Bluesky · ♥ 6 ↻ 1 ↩ 1 · 2 from the directory shared this · 6d ago

For today's reading group, @pia-p.bsky.social presented "When Personalization Meets Reality: A Multi-Faceted Analysis of Personalized Preference Learning" by Yijiang River Dong et al. (2025) Paper: aclanthology.org/2025.finding... #NLProc

aclanthology.org
View on Bluesky · ♥ 8 ↻ 2 ↩ 0 · 2 from the directory shared this · 34d ago

#MemoryMonday #NLProc 'Entropy-based Attention Regularization Frees Unintended Bias Mitigation from Lists' by Attanasio et al. redefines bias reduction in #AI, sans prior term knowledge. #2022Publication aclanthology.org/2022.finding...

Entropy-based Attention Regularization Frees Unintended Bias Mitigation from Lists aclanthology.org
AI Weekly's analysis →
  • EAR adds an objective that penalizes tokens with low self-attention entropy, discouraging BERT from overfitting to specific training terms.
  • Across three benchmark corpora in English and Italian, EAR matches or exceeds state-of-the-art for hate-speech classification and bias metrics.
  • Because the method needs no list, it also surfaces the terms most likely to induce bias as a diagnostic byproduct.
Read full analysis →
View on Bluesky · ♥ 4 ↻ 1 ↩ 0 · 2 from the directory shared this · 2d ago

#TBT #NLProc Hessenthaler et al.'s 2022 work delves into AI's link with fairness & energy reduction in English NLP models, challenging bias reduction theories. #AI #NLP #sustainability aclanthology.org/2022.emnlp-m...

Bridging Fairness and Environmental Sustainability in Natural Language Processing aclanthology.org
AI Weekly's analysis →
  • An EMNLP 2022 paper reports that knowledge distillation, a common efficiency technique, can actually decrease model fairness rather than preserve it.
  • The case study evaluates distilled models on natural language inference and semantic similarity, with gender bias measured via the Word Embedding Association Test.
  • The authors argue fairness and environmental sustainability are studied in isolation, and that an exclusive focus on one can hinder the other.
Read full analysis →
View on Bluesky · ♥ 7 ↻ 1 ↩ 0 · 2 from the directory shared this · 6d ago

#MemoryMonday #NLProc 'BERTective: Language Models and Contextual Information for Deception Detection' by Fornaciari, T. et al. (2021) explores AI's ability to detect deceit through context. www.aclweb.org/anthology/20...

BERTective: Language Models and Contextual Information for Deception Detection - ACL Anthology aclweb.org
AI Weekly's analysis →
  • Fornaciari, Bianchi, Poesio and Hovy combine BERT with attention over surrounding text to identify deceptive statements in Italian dialogues.
  • Only context near the target utterance helps, and only when it comes from the same speaker rather than from an interlocutor's questions.
  • The authors report a new state of the art on the task and release the dataset and code for reproducibility on GitHub.
Read full analysis →
View on Bluesky · ♥ 1 ↻ 1 ↩ 0 · 2 from the directory shared this · 9d ago

#TBT #NLProc 'Exploring challenges in Zero-shot Cross-lingual Hate Speech Detection, @deboranozza.bsky.social (2021) reveals how current models may inaccurately label non-hateful, language-specific interjections as hate speech signals.' aclanthology.org/2021.acl-sho...

Exposing the limits of Zero-shot Cross-lingual Hate Speech Detection aclanthology.org
AI Weekly's analysis →
  • Debora Nozza's 2021 ACL-IJCNLP short paper tests transferring an English hate speech model to Italian and Spanish without target-language labels.
  • Post-hoc explanations show the model misreads non-hateful, language-specific taboo interjections as signals of hate speech.
  • The paper concludes zero-shot cross-lingual models cannot be used as they are and need to be carefully designed.
Read full analysis →
View on Bluesky · ♥ 2 ↻ 1 ↩ 0 · 2 from the directory shared this · 13d ago

#MemoryMonday #NLProc 'MilaNLP @ WASSA: Does BERT Feel Sad When You Cry?' by Fornaciari et al. (2021) indicates that emotion and empathy are not related tasks for prediction. aclanthology.org/2021.wassa-1...

MilaNLP @ WASSA: Does BERT Feel Sad When You Cry? aclanthology.org
AI Weekly's analysis →
  • MilaNLP at Bocconi submitted to the WASSA 2021 shared task on essay-level emotion classification of reactions to English news stories.
  • Adding empathy as an auxiliary task and demographic attributes as input both gave worse performance than the plain single-task model.
  • The authors conclude emotion and empathy are not related tasks for prediction, despite the submission remaining competitive at the competition.
Read full analysis →
View on Bluesky · ♥ 1 ↻ 1 ↩ 0 · 2 from the directory shared this · 16d ago

For today's reading group, Urs Zaberer presented "Toward universal steering and monitoring of AI models" by Beaglehole et al. (2026) Paper: www.science.org/doi/10.1126/... #NLProc

science.org
View on Bluesky · ♥ 4 ↻ 1 ↩ 0 · 2 from the directory shared this · 20d ago

#TBT #NLProc 'Universal Joy: A Data Set and Results for Classifying Emotions Across Languages' by Lamprinidis et al. (2021) explores how AI research affects our planet. Tech can be green too! #SustainableTech aclanthology.org/2021.wassa-1.7

Universal Joy A Data Set and Results for Classifying Emotions Across Languages aclanthology.org
AI Weekly's analysis →
  • Universal Joy assembles over 530,000 anonymized public Facebook posts across 18 languages, each labeled with one of five emotions.
  • Using multilingual BERT, the authors report that emotions can be inferred both within and across languages, with typologically similar languages helping each other.
  • Zero-shot transfer to low-resource languages is reported as promising, suggesting cross-lingual emotion models can extend beyond the training-language set.
Read full analysis →
View on Bluesky · ♥ 2 ↻ 1 ↩ 0 · 2 from the directory shared this · 20d ago

#MemoryModay #NLProc 'Measuring Hurtful Sentence Completion in Language Models' by @debora_nozza et al. introduces HONEST, a new metric for harmful stereotypes. Language models use hurtful words 4.3% of the time.

HONEST: Measuring Hurtful Sentence Completion in Language Models aclanthology.org
AI Weekly's analysis →
  • The HONEST benchmark found language models produced a hurtful word in roughly 4.3% of sentence completions across the study.
  • When prompts targeted women, 9% of completions referenced sexual promiscuity; when they targeted men, 4% referenced homosexuality.
  • The template- and lexicon-based methodology was applied across six languages: English, Italian, French, Portuguese, Romanian, and Spanish.
Read full analysis →
View on Bluesky · ♥ 1 ↻ 1 ↩ 0 · 2 from the directory shared this · 30d ago
Their own posts

Recent commentary

@taniseceron.bsky.social is presenting her work about political content in pre-training and post-training data at the AI & Society conference. #AIandSociety #NLProc

View on Bluesky · ♥ 17 ↻ 5 ↩ 0 · 26d ago

🎉 Excited to welcome Lorena Calvo Bartolomé to our lab! Her expertise in NLP and signal processing will bring fresh perspectives to our research. She will be working on TOLD, a project exploring voice-based data collection as a richer alternative to written annotation. #NLProc

View on Bluesky · ♥ 11 ↻ 5 ↩ 0 · 35d ago

We had the pleasure of hosting @tresiwald.bsky.social and Alireza Salemi at our latest seminar. Andreas spoke about the reliability of language models through computational argumentation, while Alireza presented his work on personalizing LLMs. Thank you both for the inspiring talks! #NLProc

View on Bluesky · ♥ 9 ↻ 3 ↩ 0 · 37d ago

We were honored to welcome Giulia Barbareschi to our lab for an inspiring conversation on inclusive technology and what NLP should learn from it. Thank you for sharing your insights and vision! #InclusiveTech #Accessibility #NLProc

View on Bluesky · ♥ 7 ↻ 3 ↩ 0 · 15d ago

Last Friday, we had the pleasure of hosting Clement Jonathan Mazet-Sonilhac and Monika Kaczorowska for an insightful discussion on AI adoption and its impact on the quality of financial services. #NLProc #AI

View on Bluesky · ♥ 6 ↻ 1 ↩ 0 · 28d ago
Their network

In MilaNLP Lab's orbit

Center = MilaNLP Lab. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.