Author: Valerio Basile

S. Mahed Mousavi from University of Trento will present his research from his PhD (first part) and more recent projects (second part) including the CORE research program.

Abstract
Conversational AI systems can now produce fluent, human-like text and are increasingly used in everyday and sensitive settings. This talk argues that fluency should not be mistaken for contextual understanding, and that current AI systems are not trained to track who they are talking to, what has happened before, or what is socially appropriate in a given situation. 

The talk presents two lines of work addressing this problem. The first focuses on deploying ConvAI models for longitudinal dialogues, i.e. a sparse sequence of personal dialogue sessions. It presents methods for collecting multi-session conversations, building and updating a model of the individual user over time, and evaluating AI models in a standardized way. These efforts resulted in the first registered Randomized Controlled Trial (RCT) with a ConvAI system in the mental health domain. The second line of work asks why AI systems struggle to understand context in the first place. It shows that many benchmarks used to claim reasoning abilities in AI are themselves flawed, and traces the problem to how these systems are trained, showing evidence that the current training paradigm is not sustainable, and that loss optimization is not a reliable signal of learning.

Where: Sala Conferenze, 3rd Floor
When: 22/09/2026, 14:00

Links:

• Slides
• Video

Seyed Mahed Mousavi from University of Trento will give a seminar to our Department.

Abstract
Conversational AI systems can now produce fluent, human-like text and are increasingly used in everyday and sensitive settings. This talk argues that fluency should not be mistaken for contextual understanding, and that current AI systems are not trained to track who they are talking to, what has happened before, or what is socially appropriate in a given situation. 

The talk presents two lines of work addressing this problem. The first focuses on deploying ConvAI models for longitudinal dialogues, i.e. a sparse sequence of personal dialogue sessions. It presents methods for collecting multi-session conversations, building and updating a model of the individual user over time, and evaluating AI models in a standardized way. These efforts resulted in the first registered Randomized Controlled Trial (RCT) with a ConvAI system in the mental health domain. The second line of work asks why AI systems struggle to understand context in the first place. It shows that many benchmarks used to claim reasoning abilities in AI are themselves flawed, and traces the problem to how these systems are trained, showing evidence that the current training paradigm is not sustainable, and that loss optimization is not a reliable signal of learning.

When: Tuesday, 22nd of September, 14:00
Where: Sala Conferenze, 3rd floor

Gavin Abercrombie will give a talk while visiting our group on his recent research.

Abstract
In this talk, I investigate the question of whether we should, ethically, automate counterspeech. I discuss criteria for the ethical assessment of automating counterspeech, look at the legal and philosophical justifications, both for and against doing so, and discuss how the applicability of these ethical considerations depends on automated counterspeech’s autonomy, transparency, and intended aims. I conclude by drawing together implications for both practice and future research.

Bio: Gavin Abercrombie is Assistant Professor at Heriot-Watt University, Scotland. He was a co-organiser of the Workshop on CounterSpeech for Online Harms (2023) and the ACL Tutorial on Counterspeech against Hate and Misinformation, and is a co-author of a chapter on counterspeech in the Bloomsbury Handbook of Dangerous Speech (Forthcoming).

When: Wednesday, 9th of September, 11:00
Where: Sala Conferenze, 3rd floor

Slides PDF

CCC Seminar by Anna Koufakou, visiting professor from Florida Gulf University, United States

Abstract

The evolution of Natural Language Processing (NLP) technologies, with the advent of Large Language Models (LLMs) and Generative AI, has significantly expanded the scope and effectiveness of NLP-driven tools across applications. Among these, Automated Emotion Recognition (AER) represents a particularly challenging yet promising area of research. Unlike Sentiment Analysis, which usually categorizes sentiment as positive or negative, AER aims to identify specific emotional states, such as anger, joy, or sadness, that can vary widely in expression and meaning. This talk explores the current landscape of relevant text corpora and resources, including our recent efforts towards a unifying benchmark. We will also discuss recent practices and key challenges in this evolving field.

When:  29/05/2025, h 10.00

Where:  Sala Conferenze – 3rd floor

Juan Manuel Pérez will give a talk during his visiting week to the Content-centered Computing group.

In recent years, hate speech has gained great relevance in social networks and other virtual media because of its intensity and its relationship with violent acts against members of protected groups. Due to the incommensurable amount of content generated by users, great effort has been made in the research and development of automatic tools to aid the analysis and moderation of this speech, at least in its most threatening forms.

One of the limitations of current approaches to automatic hate speech detection is the lack of context. Most studies and resources are performed on data without context; that is, isolated messages without any type of conversational context or the topic being discussed. This restricts the available information to define if a post on a social network is hateful or not.

In this talk, I will comment on some experiments we have performed to assess the impact of context in hate speech detection. With this in mind, we built a contextualized dataset for hate speech detection based on user responses to news posts from media outlets on Twitter. This corpus was collected in the Rioplatense dialectal variety of Spanish and focuses on hate speech associated with the COVID-19 pandemic.

For the two proposed tasks using this novel corpus (binary detection; and granular detection, where the system has to predict the attacked characteristics), the classification experiments using state-of-the-art techniques show evidence that adding contextual information improves hate speech detection performance.

When: September 29, 2022, at 11:00

Where: Conference room 3rd floor (Sala Seminari)

Tolúlọpẹ́ Ògúnrẹ̀mí will present her work as a PhD student at Stanford University.

Recognising code-switched speech (alternating between two or more languages or varieties of language across sentences in conversation) is an important technical and social issue essential for modern society. The majority current speech recognisers are trained monolingually and therefore do not perform well on such utterances. The use of Deep Neural Network (DNN) architectures to train models allow for shared representations and provide an opportunity to level them to better handle code-switching. In the two studies contained in this work, we show multilingual fine-tuning of self-supervised speech representations can handle code-switching in a zero-resource scenario and through analysis of the latent representations, that code-switching is encoded in the model. We find that monolingual data is enough for character-level decoding in the code-switched scenario and that representations are not similar to word vectors.

When: 4/7/2022

Where: Sala conferenze on the 3° floor

Antonio Lieto

Inventing novel knowledge to solve problems is a crucial, creative, mechanism employed by humans, to extend their range of action. In this talk, I will show how commonsense reasoning plays a crucial role in this respect. In particular, I will present a cognitively inspired reasoning framework for knowledge invention and creative problem solving exploiting TCL: a probabilistic non-monotonic extension of a Description Logic (DL) of typicality able to combine prototypical (commonsense) descriptions of concepts in a human-like fashion. The proposed approach has been tested in a variety of fields and applications. I will present the obtained results, the lessons learned, and the road ahead of this research path.

See this page https://www.antoniolieto.net/tcl_logic.html for a list of the main papers and applications