Author: Samuele D'Avenia

Guillermo Villate Castillo, a visiting PhD student from the University of Deusto, will present his ongoing work on understanding the sources of annotator disagreement in toxicity annotation and distinguishing genuine differences in perspective from other factors.

Abstract
Annotator disagreement is increasingly recognized as a valuable source of information in subjective natural language processing tasks, motivating a shift from consensus-based annotation toward perspectivist approaches that preserve diverse human judgments. However, treating disagreement as inherently meaningful perspectival variation risks conflating genuine differences in interpretation with disagreement arising from linguistic ambiguity, missing context, epistemic uncertainty, and annotation-related factors. In this work, we challenge the assumption that annotator disagreement constitutes a homogeneous phenomenon and propose a data-centric framework for systematically identifying, characterizing, and disentangling its underlying sources in toxicity annotation. Our approach moves beyond conventional categorical and scalar toxicity labels by introducing a fine-grained, evidence-grounded annotation framework that captures multiple plausible judgments, their supporting textual evidence, interpretative mechanisms, and propositional content. Additionally, we collect annotator-level information, including demographic characteristics, lived experiences, and moral identity, alongside detailed annotation traces to investigate the relationship between individual differences and toxicity judgments. We propose an experimental methodology to examine three research questions: (1) whether richer annotation representations reveal disagreement structures that remain hidden under conventional labeling schemes; (2) how different sources of disagreement, individually and jointly, contribute to observed variation in toxicity judgments; and (3) whether residual disagreement, after accounting for identifiable data- and task-related factors, is associated with annotator characteristics. By distinguishing potentially reducible sources of disagreement from genuine perspectival variation, this work aims to provide a better understanding of human disagreement and inform the development of more reliable, interpretable, and perspective-aware toxicity annotation practices as well as a better decision-making.

When: 16/09/2026, 11:30
Where: Sala Riunioni, 1st floor

Arabella Sinclair, Assistant Professor at UCL (London) and at the University of Aberdeen will present her work who will present her work on large language models (LLMs), language processing, and cognition.

Abstract
Both humans and Large Language Models (LMs) generate predictions about upcoming words and structures based on recent context. In dialogue, speakers continuously choose how to realise their communicative intentions, with these choices shaped by multiple, sometimes competing pressures, including production costs borne by the speaker and comprehension costs incurred by the listener. Production costs reflect the effort involved in planning and generating an utterance, while comprehension costs reflect how easily an utterance can be processed and interpreted. These costs are influenced by a range of factors, including priming, utterance length, informational content, and contextual predictability.
In this talk, I explore how LMs can serve both as tools for studying human interaction and as components of broader cognitive models of language processing. First, I present evidence for parallels between humans and LMs in primed comprehension facilitation, showing how prior exposure to linguistic structures influences subsequent processing. Second, I discuss ongoing work investigating analogous effects in a controlled production setting. Finally, I introduce a new procedure for constructing contextual alternative sets that enables probabilistic pragmatic models of language production to be instantiated and evaluated at scale in open-ended communicative settings. This framework also provides a principled basis for comparing competing notions of communicative cost.
Taken together, these findings shed light on the mechanisms underlying LMs’ in-context learning behaviour while also assessing their potential as models of human linguistic processing and communicative decision-making.

When: 11/09/2026, 10:30
Where: Sala Conferenze, 3rd Floor

Bernardo Magnini, Senior Researcher at Fondazione Bruno Kessler (FBK) in Trento, will present a series of results from the eCREAM (enabling Clinical Research in Emergency and Acute care Medicine) Horizon project.

Abstract
This seminar provides an overview of the use of Large Language Models (LLMs) within the eCREAM (enabling Clinical Research in Emergency and Acute care Medicine) Horizon project. I will first introduce a unique corpus of clinical notes collected from the emergency departments of several Italian hospitals and show how these data have been used to train small LLMs for a range of clinical tasks. I will then focus on Case Report Form (CRF) filling, a key task addressed in eCREAM, where manually annotated clinical notes are leveraged to extract clinically relevant information. I will present results showing that substantial performance improvements can be achieved by fine-tuning models with additional training data automatically extracted from unannotated clinical notes. Finally, I will discuss ongoing work on the automatic interpretation of patient–doctor interactions in emergency departments, highlighting current challenges and outlining future research directions.

Where: Sala Conferenze, 3rd Floor
When: 11/09/2026, 9:30

Samuele D’Avenia will provide a brief overview on statistical methods commonly used in NLP, with a focus on common errors.

Abstract
Statistical significance testing is widely used in Natural Language Processing, yet its application is often incomplete or misleading.
This talk provides a high-level overview of common statistical tools employed in NLP research and highlights several recurring issues that can compromise the interpretation of experimental results.

We will mainly focus on hypothesis testing, highlighting the multiple comparison and the large sample issues, while also discussing current practices.

Additionally, we will provide a brief overview of the current practices within the Italian NLP community.

When: Friday, 10th of July, 11:30
Where: Sala Conferenze, 3rd floor

Giulio Corradi (Privacy Network), and Marta Marchiori Manerba (University of Turin) will present one of their latest work.

Abstract:
In this talk we present Advocacy Radar, a platform that allows civil society practitioners, journalists, and policy researchers to monitor the state of digital rights advocacy in Italy and Europe from an interactive interface.
Advocacy Radar adapts NLP methods to collect and process heterogeneous public sources – including regulatory authorities, NGOs, and technology media – spanning both Italian and multilingual documents.
Analyses are presented through a dynamic multi-view dashboard. A human-in-the-loop correction through the interface allows domain experts to refine model outputs and incrementally improve models over time.
In this seminar, we describe the system architecture, its main dashboard components, and representative use cases for different user profiles.

Remark: The project is developed in partnership with Privacy Network, an association committed to safeguarding digital rights in Italy and beyond.
The work is ongoing; a version of this paper has been submitted to CLiC-it 2026.
We welcome informal feedback and suggestions from the CCC audience: your expertise and perspective would be super precious in shaping the next stages of development.

Short Bio:
Giulio Corradi is a Management Engineer based in Milan, working as a consultant at SDG Group. He holds an MSc in Management Engineering from the University of Bergamo, with a thesis on algorithmic accountability under the EU AI Act, developed in collaboration with the Department of Law.

His interests sit at the intersection of data-driven analysis, NLP, and business processes, with a particular focus on the social and political implications of technology. Currently He is a Research Officer at Privacy Network, one of Italy’s leading digital rights organizations, he researches automated surveillance systems and their impact on fundamental rights and democratic processes. He also writes for Scomodo, contributing to public discourse on technology, power, and society.

Marta Marchiori Manerba is a Postdoc at the Computer Science Department of the University of Turin, where she works on perspectivist approaches to dialogue modeling. She holds a Ph.D. in AI & Society from the University of Pisa, with a dissertation focused on fairness auditing through explainability particularly in the context of abusive language detection.

When: Friday, 19th of June, 11:30
Where: Sala Riunioni, 1st floor

The seminar is presented by Dr Dilara Torunoğlu Selamet, Lecturer in the Department of Computer Engineering at Istanbul Technical University (ITU), Türkiye.

Abstract:
Word Sense Disambiguation (WSD) is a fundamental Natural Language Processing (NLP) task that aims to identify the intended meaning of an ambiguous word in context. While substantial progress has been achieved for high-resource languages, WSD remains challenging for morphologically rich and low-resource languages such as Turkish due to the scarcity of large-scale annotated datasets.
In this talk, I will present my PhD research on Turkish Word Sense Disambiguation, which combines data-centric and model-centric approaches. First, I will introduce DodoMe, a large-scale gamified crowdsourcing platform developed to collect sense-annotated Turkish sentences. The resulting dataset contains more than 158,000 annotations covering 30 highly ambiguous Turkish words and represents one of the largest publicly available WSD resources for Turkish.
Second, I will discuss a systematic comparison of modern approaches to WSD, including contextual embedding-based methods, prompting-based inference with Large Language Models (LLMs), and instruction-based fine-tuning of open-source LLMs. The results demonstrate that recent LLMs substantially outperform traditional embedding-based approaches and that instruction tuning can further improve performance when sufficient high-quality annotated data is available.
Finally, I will discuss the broader implications of combining human computation, crowdsourcing, and large language models for developing semantic resources and language technologies for under-resourced languages.

Short Bio:
Dilara Torunoğlu Selamet is a Lecturer in the Department of Computer Engineering at Istanbul Technical University (ITU), Türkiye, and recently defended her PhD in Computer Engineering. She is a member of the ITU Natural Language Processing Research Group, and her research focuses on Natural Language Processing, lexical semantics, word sense disambiguation, meaning representation, and multilingual language technologies.
Her doctoral research investigates Turkish Word Sense Disambiguation through the integration of large-scale crowdsourced datasets and Large Language Models. She is the creator of DodoMe, a gamified crowdsourcing platform designed for collecting semantic annotations in Turkish. Her work explores contextual embeddings, prompting strategies, and instruction-based fine-tuning approaches for semantic disambiguation tasks.
Dilara is actively involved in international collaborations through the UniDive COST Action, contributing to multilingual and multimodal language technology initiatives. She serves as a language leader and coordinator in the AdMiRe (Advancing Multimodal Idiomaticity Representation) shared task series, which focuses on multilingual and multimodal idiomaticity understanding across dozens of languages.
Her recent work includes contributions to the AdMiRe shared tasks at EACL 2026 and LREC-COLING 2026, as well as research on Turkish Word Sense Disambiguation, Abstract Meaning Representation (AMR), Uniform Meaning Representation (UMR), and multilingual idiomaticity understanding. She has co-authored large-scale international publications involving researchers from more than 30 languages and actively contributes to the development of multilingual benchmarks, linguistic resources, and evaluation campaigns for Natural Language Processing.

When: 10/06/2026, h 11:00 am
Where: Sala Conferenze, third floor

CCC Seminar by Luigi Sauro, Professor of Computer Science at Università degli Studi di Napoli “Federico II”.

Abstract (Italian): Supponete di avere un insieme finito di meccanismi stocastici (arms), ognuno dei quali restituisce ad ogni interazione un reward sulla base di una distribuzione a voi sconosciuta. Lo scopo del Fixed-Budget Best Arm Identification è quello di identificare, sulla base di un numero prefissato di possibili interazioni sequenziali, l’arm con expected reward massimo. Questo problema induce un tipico dilemma exploration vs exploitation che si riscontra in molti contesti applicativi (clinical trials, wireless network selection, recommender systems, A/B testing). In questo seminario illustrerò una nuova strategia di interazione che si applica sotto le ipotesi che gli arm siano governati da distribuzioni sub-gaussiane e che fa uso di tecniche di non-linear convex optimization. Mostrerò inoltre un upper bound teorico che evidenzia il perché questa strategia raffina lo stato dell’arte ed un’analisi sperimentale che supporta la sua efficacia.

Bio: https://www.docenti.unina.it/teacher/4c55494749534155524f5352414c475537345432394638333941/profile/references

When: June 5th, 15:00
Where: Sala Riunioni, Primo Piano

Seminar by Beatrice Savoldi, researcher at Fondazione Bruno Kessler (FBK) within the MT research unit.

Abstract
Societal gender asymmetries, inequalities and stereotypes can be embedded in our communication practices and perpetuated in language technologies, including multilingual and Machine Translation (MT) systems used as scale. In this presentation, we will delve into the current landscape of MT and multilingual gender bias, as well as current proposals towards more inclusive language. We will discuss the challenges and opportunities — both theoretical, technical but also linguistic — in fostering a more equitable technology for different groups of users.

Bio
Beatrice Savoldi is a researcher at Fondazione Bruno Kessler (FBK) within the MT research unit, where she mainly works on gender inclusive and human-centered translation technologies.
Beatrice carried out a joint international PhD at the University of Trento and Augsburg with a dissertation on gender bias in Machine Translation, which was awarded the 2023 best thesis Research Prize from the Augsburg University Foundation.
Her research interests broadly encompass ethical and social considerations of (multilingual) language technologies.

When: 29/05/2026 11.00
Where: Aula 3.06 Thin Client (3rd floor) – Via Sant’Ottavio 54, Torino.

Samuele D’Avenia and Eliana di Palma will be presenting one of their latest work.

Abstract:
Offensive language detection systems often rely on majority-aggregated annotations, overlooking the diversity of perspectives that shape how different communities perceive harm. In this contribution, we introduce HurtLens, a perspectivist corpus of hurtful language leveraging four disaggregated datasets which are automatically enriched through HurtLex lemmas, a multilingual resource of offensive and derogatory terms. Using mixed-effects modeling, we investigate how annotators’ sociodemographic backgrounds, the presence of specific types of offensive language (through Hurtlex categories) and their interaction influence offensiveness ratings. Our analysis reveals that offensiveness ratings are influenced both by annotators’ sociodemographic characteristics (particularly when considering them in intersection) and by the presence of specific types of offensive language. Additionally, we identify significant interaction effects showing that different demographic groups vary in their sensitivity to texts containing particular types of offensive language.

When: Friday, 15/05 at 11:30
Where: Sala Conferenze, 3rd Floor

CCC Seminar by Jussara M. Almeida, Full Professor of Computer Science at the University of Minas Gerais, Brazil.

Abstract:
This talk examines how information spreads, clusters, and targets specific audiences across social media platforms. Drawing on three empirical studies, I analyze (i) the structure and discourse of misogynistic communities in Brazil, (ii) content directed at children on Instagram, and (iii) the emergence and alignment of political narratives in large-scale Telegram conversations during the 2024 U.S. elections. Together, these findings highlight relevant patterns underlying information diffusion across platforms and raise important questions about exposure, influence, and the amplification of targeted content in online ecosystems.

Short bio:
Jussara M Almeida is Full Professor of Computer Science at the University of Minas Gerais, Brazil, where she currently leads the Laboratory of Social Computing at the Department of Computer Science. She holds a PhD from the University of Wisconsin-Madison (US) and is a former affiliate member of the Brazilian Academy of Sciences. Her main research interests are social computing, user behavior analysis, as well as performance analysis and modeling of large-scale distributed systems.

When: 14/04/2026, 15:00-16:00
Where: Sala Riunioni, 3rd Floor