The ItaliaNLP Lab @ Istituto di Linguistica Computazionale “Antonio Zampolli” (ILC) - CNR
Pisa, Tuscany
Joined November 2015
- Tweets238
- Following2.3K
- Followers1.8K
- Likes1.2K
🔥 @lucadini_ and @workerplacemint won the Best Student Paper Award at @CLiC_it_conf for their paper “From Weights to Representations: How Pre-Training Data Order Shapes Language Models”! 🎉🎉 @AILC_NLP #NLPRoc
Last presentation from our group: Cristiano Ciaccio and Asya Zanollo presenting “Reading Between the Clues: A Fine-Grained Analysis of Italian Crossword Solving Systems” (with @AlessioMiaschi, @gsarti_, Dell’Orletta F. and @MalvinaNissim)! 🔥 @CLiC_it_conf @AILC_NLP #NLPRoc
Last day of @CLiC_it_conf! Massimiliano Orsini is presenting “PUNder-IT: Assessing Language Models’ Sensitivity to Italian Puns” (with Dominique Brunato)! 🔥 @AILC_NLP #NLPRoc
And now it’s time of @lucadini_ with “From Weights to Representations: How Pre-Training Data Order Shapes Language Models” (with @workerplacemint, Brunato D. and Dell’Orletta, F.)! 🔥 @CLiC_it_conf @AILC_NLP #NLPRoc
🎉 Great news! We got 6 papers (5 regular + 1 research communication) accepted at CLiC-it 2026! Looking forward to presenting them this year in Palermo! #CLiCit2026 #NLProc
@CLiC_it_conf
@AILC_NLP
[1/4]🚨 Happy to share our #CoNLL2026 paper presented by Lucia Domenichelli at #ACL2026!
"Linguistic Profiling of Transformer Embedding Geometry", @workerplacemint, Dominique Brunato, Felice Dell'Orletta
#NLP #LLMs #Interpretability
aclanthology.org/2026.conll-…
[3/4] A few findings:
• BERT maintains more isotropic, higher-dimensional representations than GPT-2.
• GPT-2 develops a distinct cluster of sentence-initial tokens.
• Content words and syntactically central predicates occupy richer geometric subspaces.
🧶 1/3
We’re excited to share our second paper presented at #LREC2026 in Palma: “Controllable Sentence Simplification in Italian: Fine-Tuning Large Language Models on Automatically Generated Resources” by @mpapucci_, Giulia Venturi and Felice Dell'Orletta
🧶 2/3
In this work, we introduce IMPaCTS: a new multi-level sentence simplification dataset for Italian, automatically generated to support controllable text simplification research.
😊 Dataset on HF: huggingface.co/datasets/mpap…
🔧 Dataset on GitHub: github.com/michelepapucci/im…
🧶 3/3
We show that models fine-tuned on IMPaCTS outperform instruction-tuned zero-shot LLMs on controlled sentence simplification tasks, becoming proficient in both simplifying texts and in adhering at the specified target readability level.
Paper: tinyurl.com/s37f4nth
🚀 Excited to share our latest work at READIxTSAR workshop at #LREC2026
"Lexical Conditioning of Model's Distribution through Uncertainty-gated Soft-Mixing of Probabilities" by @mpapucci_ Giulia Venturi and Felice Dell'Orletta.
In this work we present UGLD, a decoding method that nudges LLMs toward or against a predefined vocabulary at inference time. No fine-tuning needed and without loss of fluency.
📄 Paper: minifyurl.in/xGNf
🔧 GitHub: github.com/michelepapucci/ug…
🐍 Install UGLD: pip install ugld
[1/4] How does the order of pretraining data impact Neural Language Models?
We study this in our 🎉new paper🎉:
"On the impact of pretraining data ordering in transformer encoder- and decoder-only language models"
Published in Knowledge-Based Systems
🔗sciencedirect.com/science/ar…
[3/4] We examine effects on:
• Perplexity dynamics
• Linguistic probing
• Downstream tasks performance
• Embedding geometry
Data ordering shapes both learning dynamics and representations.
[4/4] 🧑💻 Code & experiments:
github.com/lucadinidue/curri…
Happy to discuss, feedback welcome! 🙌
