Author: Samuele D'Avenia

CCC Seminar by Alessandro Corona Mendozza, predoc researcher at the center for language Technology (Copenhagen University) and visiting at the University of Turin.

Abstract:
We investigate belief-like representations in decoder-only autoregressive LLMs using linear controlled probes on residual stream activations and single attention heads. Following Herrmann and Levinstein’s (2025) criteria (Accuracy, Use, Coherence, and Uniformity) we find that large models exhibit strong truth sensitivity (Accuracy), and steering activations along probe directions reliably changes downstream behavior (Use). Coherence, measured via calibrated probes and cross-dataset probing, is moderate across models, while training on diverse data yields domain-consistent truth directions (Uniformity). The results are particularly encouraging at the head level and align with some standard philosophical accounts of belief, e.g., minimal functionalism, supporting the view that LLMs can maintain propositional attitudes under such theoretical frameworks.

Short bio:
Alessandro Corona Mendozza is a predoc researcher working at the intersection of LLM interpretability, AI epistemology and philosophy of mind/language. He is currently assisting in research for an eye-tracking project at the center for language technology (Copenhagen University) and University of Turin (visiting). 

When: 18/02/2026, h 14:00
Where: Sala Conferenze, 3rd Floor

CCC Seminar by Silvia Casola, postdoc researcher at the MaiNLP group of the Ludwig Maximilian University of Munich and Munich Center for Machine Learning.

Abstract:
Automatic evaluation in NLP often assumes a single ground truth, such as a reference or a gold label. However, language is inherently variable: multiple outputs can be valid, annotators frequently disagree, and metric behaviours can differ across languages. In this talk, I will present three case studies showing how evaluation can fail and how it can be improved under such variation. Focusing on NLG, I will show that metrics can be highly sensitive to the choice of reference, leading to large changes in system rankings. I will then examine classification evaluation under annotator disagreement and present an approach for accounting for systematic disagreement. Finally, I will discuss recent work on steering multilingual neural metrics to improve their correlation with humans.
Starting from these failure modes, the talk shows how studying and modeling variation in references, annotations, and languages can improve the stability and reliability of automatic evaluation.

Short Bio:
I am a postdoc researcher at the MaiNLP group of the Ludwig Maximilian University of Munich and Munich Center for Machine Learning, supervised by Barbara Plank. Previously, I was a postdoc at the University of Turin, where I worked on perspective-aware NLP. I completed my PhD at the University of Padua and Fondazione Bruno Kessler, working on NLG. During my Ph.D., I visited UPF and interned with Spotify and Huawei Research. My research interests lie in NLP, focusing on Natural Language Generation and evaluation.

When: 31/03/2026, afternoon
Where: Sala Conferenze, 3rd Floor

CCC Seminary by Marcos Gonçalves, Full Professor of Computer Science at the Federal University of Minas Gerais (UFMG), who will present his work INCT TILDIAR: Data-Centric and Sustainable Paths Beyond the ‘Law of More’ in NLP.

Abstract:
This talk introduces the INCT TILDIAR, a national Brazilian research network dedicated to responsible and sustainable Artificial Intelligence, and presents the main research directions developed within the institute. We illustrate how our work challenges the prevailing “Law of More” in NLP by emphasizing data-centric and efficiency-driven approachesto AI.

As a concrete example, the talk briefly summarizes research on instance selection and data engineering, showing that substantial reductions in training data and energy consumption can be achieved while maintaining model effectiveness. The overarching message is that sustainable NLP is possible by rethinking how data is selected and used, rather than relying solely on ever-larger models.

Short bio:
Marcos André Gonçalves is a Full Professor of Computer Science at the Federal University of Minas Gerais (UFMG). He holds a PhD in Computer Science from Virginia Tech, with prior degrees from UFC and UNICAMP, and has completed postdoctoral research at UFMGand Politecnico di Torino. His research focuses on Information Retrieval, Machine Learning, and Natural Language Processing, with over 400 peer-reviewed publications, an h-index of 61, and more than 15,000 citations.

He has received multiple national awards, including CAPES Thesis Awards (2024 – Advisor; 2020 – Honorable Mention), and several Best Paper awards. Prof. Gonçalves has served as General Chair of ACM/IEEE JCDL 2018, is a Senior Program Committee member of leading conferences (SIGIR, ACL, CIKM, WSDM, RecSys, ECIR), and serves on the editorial boards of TACL and the Journal of the Brazilian Computer Society. He is also the Coordinator of the Brazilian National Institute for Responsible AI in NLP (INCT TILDIAR).

When: 02/02/2026, h 15:00
Where: Sala Riunioni, 1st Floor

CCC Seminary by Celso França, Ph.D. student at UFMG, who will present his work, xCoRetriev: A Retrieval-Centric Paradigm for Extreme and Multi-Class Text Classification.

Abstract:
We address Extreme Multi-Label Text Classification (XMTC) and Multi-Class Text Classification (MCTC) under a unified paradigm that reframes classification as a ranking and retrieval problem over large, noisy, and skewed label spaces. In this talk, we synthesize our recent SIGIR 2025 paper and our best paper of SBBD 2025 to demonstrate how retrieval-based formulations can jointly improve scalability, effectiveness, and label semantics across both XMTC and MCTC settings. Our core proposal is xCoRetriev, a dynamic two-stage retrieval and fusion pipeline designed to tackle the main challenges of label space volume, extreme skewness, and label quality by effectively combining dense and sparse representations. We further discuss recent attempts to enhance xCoRetriev’s effectiveness through Dimension Importance Estimation (DIMES) strategies and learned sparse representations trained via masked language modeling (MLM). While these approaches show promise in emphasizing discriminative signals and improving tail-label sensitivity, our analysis highlights their current limitations. Across multiple large-scale datasets, our results demonstrate consistent gains in propensity-scored metrics, improved robustness to noisy and weakly supervised label spaces through RAG-enhanced labels, and strong scalability at both training and inference time. Overall, this work advocates for a retrieval-centric view of large-scale text classification, bridging XMTC and MCTC through ranking, fusion, and importance-aware representations.

When: 02/02/2026, h 14:00
Where: Sala Riunioni, 1st Floor

CCC Seminar by Prof Gülşen Eryiğit, faculty of the Artificial Intelligence and Data Engineering Department at Istanbul Technical University, coordinator of ITU Natural Language Processing Group, and the director of ITU TÖMER (Turkish Teaching Application and Research Center).

Abstract
Natural language processing (NLP) has made remarkable progress in recent years, yet many of its most celebrated models and methods remain shaped by the structure and resources of Indo-European languages. This seminar examines the unique linguistic and computational challenges presented by agglutinative and morphologically rich languages such as Turkish, where high morphological variability, extensive affixation, and productive derivation complicate core NLP tasks including tokenization, morphology analysis, syntactic parsing, semantic representation, and downstream applications like information extraction and text understanding. The talk also offers a brief overview of the broader research agenda and key publications in this area.

Short Bio
Gülşen Eryiğit is a faculty of the Artificial Intelligence and Data Engineering Department at Istanbul Technical University, the coordinator of ITU Natural Language Processing Group, and the director of ITU TÖMER (Turkish Teaching Application and Research Center). She is the top-cited researcher in Turkey in NLP and listed in Stanford/Elsevier’s Top 2% Scientist Rankings career-long impact category. She received her master’s and PhD degrees from ITU Computer Engineering Department in 2002 and 2007. She has worked as a referee and author in many prestigious journals and conferences on NLP. She has worked as a coordinator or researcher in many scientific projects funded by EU, Tubitak, and the Ministry of Industry and Technology, and as a consultant in several industrial R&D projects funded by EU and Tubitak TEYDEB. She also acts as a project evaluator and observer for these funding agencies. She owns two issued patents. She is the person who realized the first software export from the ITU Technology transfer office. She is actively a management committee member in the European Cost Action UNIDIVE where she co-chairs the work package on “Multilingual and Cross-Lingual Language Technology”. She prioritizes the development of local technologies and has been actively involved in R&D projects on language technologies with various institutional firms in Turkey through university-industry collaboration.

When:  20/02/26 , h 11.00

Where:  Sala Conferenze – 3rd floor

Kênia Carolina Gonçalves will present her doctoral research on the cross-platform analysis and modeling of harmful social media content targeting children and adolescents. The talk addresses the growing exposure of young users to potentially harmful online content and the specific vulnerabilities they face due to ongoing emotional, social, and cognitive development. It will provide a comprehensive characterization of content directed at children across major social media platforms, with a focus on identifying and understanding potential harms. The presentation will introduce a proposed taxonomy for classifying harmful content aimed at children and adolescents, grounded in empirical analyses of cross-platform datasets. Finally, it will explore the use of unsupervised machine learning techniques based on dataset similarity to support the automatic detection of harmful content, with the goal of enabling early alerts for immediate risks and identifying more subtle and insidious forms of harm beyond explicit manifestations.

Where: Sala Conferenze (3rd Floor)
When: 16/01/26 11:30