Selected publications

Balestrucci, P. F., Dušek, O., Anselma, L., & Mazzei, A.

Objective: This paper examines whether LLMs can rewrite task-oriented dialogues in styles associated with Baby Boomers, Generation X, Generation Y and Generation Z while preserving the original communicative goals. It introduces P-MultiWOZ, validates the resource through automatic and human evaluation, and compares its linguistic patterns with GeMoSC, a separate corpus of generation-annotated movie dialogue, to assess whether the generated styles are recognisable and grounded in observable language variation.

Balestrucci, P. F., Casola, S., Lo, S. M., Basile, V., & Mazzei, A.

Objective: This study investigates whether LLMs can generate contextually appropriate ironic replies to social-media posts. It fine-tunes models for ironic and non-ironic generation, analyses how their outputs relate to the original post and to human-written replies, and evaluates them at scale with human judges. It also explores whether models can reproduce irony associated with generational perspectives, highlighting both the potential and the limitations of perspective-conditioned generation.

Balestrucci, P. F., Di Nuovo, E., Sanguinetti, M., Anselma, L., Bernareggi, C., & Mazzei, A.

Objective: This article studies dialogue as an alternative to diagrams and tables for communicating information about finite-state automata, which can create substantial accessibility barriers for visually impaired students. It presents a rule-based educational dialogue system and evaluates it through an A/B study against standard tabular representations, measuring comprehension and user experience to determine when conversational, non-visual exploration can provide more effective access to formal structures.

Di Nuovo, E., Sanguinetti, M., Balestrucci, P. F., Anselma, L., Bernareggi, C., & Mazzei, A.

Objective: This paper introduces a corpus of real English dialogues between users and a task-oriented agent that verbally describes finite-state automata. The resource was created to support the development of accessible educational technologies for STEM content and includes interactions from both sighted and visually impaired participants, making it possible to examine differences in how the two groups communicate with the system. The work documents the data-collection methodology and annotation scheme, analyses the resulting dialogue patterns, and evaluates both supervised classification and the annotation capabilities of three large language models as a basis for future corpus expansion.

Oliverio, M., Balestrucci, P. F., Mazzei, A., & Basile, V.

Objective: The paper creates WebNLG-IT, the first complete Italian version of the WebNLG RDF-to-text corpus, by combining neural machine translation with a semi-automatic rule-based post-editing process. The resulting resource is used to fine-tune Italian and multilingual LLMs, compare RDF-to-text generation across Italian and English, and develop an Italian adaptation of a modular generation pipeline that verbalises semantic-web triples.

Oliverio, M., Balestrucci, P. F., Anselma, L., & Mazzei, A.

Objective: This work proposes AIML+, a framework that preserves the accessibility and authoring simplicity of AIML while extending it with frame-based natural-language understanding and LLM-supported response generation. The aim is to make educational dialogue systems more robust and scalable without abandoning explicit domain knowledge, predictable dialogue behaviour and the level of control needed when responses must remain faithful to structured educational content.