Date: 10 March 2026

Time: 11 AM (GMT)

Join Online: https://app.huawei.com/wmeeting/join/99929826/eUCUm7VD5Hca9IkxlqmDguXqr04EHhPwl

Abstract
Personalisation in language models is often operationalised by assigning a persona or giving stylistic instructions via prompting. Yet, successful adaptation requires models to adjust both to who they represent and who they are interacting with, often from sparse or implicit signals. In this talk, I will address personalisation across two tasks: persona-driven dialogue and literary machine translation.

In the dialogue task, presenting joint work with Daniela Occhipinti and Marco Guerini, I unpack the “who they are interacting with” aspect. I will discuss how models balance their own persona with that of their interlocutor, and how access (or lack thereof) to partner information affects stylistic alignment and speaker recognisability. In the translation task, I address the “who they represent” aspect in conditions where personal stylistic choices might be subtle. In joint work with Daniel Scalena, Gabriele Sarti, Arianna Bisazza, and Elisabetta Fersini, we study how models can reproduce a specific translator’s style in literary translation from only a handful of examples, using inference-time steering rather than explicit prompts.

In closing, I will be happy to discuss with the audience broader style and personalisation issues in the current LLM-dominated research landscape.

Bio
Malvina Nissim (https://malvinanissim.github.io) is full professor of Computational Linguistics & Society at the University of Groningen, The Netherlands, where she is a member of the GroNLP (Groningen Natural Language Processing) group. Before settling in Groningen over 10 years ago, she lived, studied, and worked in Italy (where she grew up), Scotland, and Germany. She works on multiple aspects of Natural Language Processing, with particular attention to language itself and the interaction of Language Technology and society. Relatedly, she is the coordinator of the Sectorplan‘s theme Humane AI at the University of Groningen, and is a member of the ACL Ethics Committee. As a complement to her research activities she is keen to engage in science communication and outreach, including to school children, especially towards an aware and responsible use of language technologies. She very much enjoys public speaking, and she was lucky to get the chance to give a TEDx talk and to be invited by the European Council to give a plenary lecture on gender bias in Artificial Intelligence on International Women’s Day 2021. As a university professor she teaches quite a bit and was elected as the 2016 University of Groningen Lecturer of the Year. Malvina is also a member of the Scientific Advisory Board of the Jantina Tammes School for Digital Society, Technology, and AI at the University of Groningen, and a member of the board of AILC, the Italian Association for Computational Linguistics. To survive a full time job she loves to run near the water, to knit, and to spend time with her two teenagers.

【Edinburgh Bayes Coffee House Talk】Style and Interaction in Large Language Model Personalisation / Huawei-Edinburgh Joint Lab by blogadmin is licensed under a Creative Commons Attribution CC BY 3.0