Events

Please register here if you wish to attend.​

Time:  Thursday 24 September (15:00-17:00)

Location: S.KS203 

Abstract

When a problem is hard, people rarely solve it alone. In fact, we do the opposite and form teams, discuss, and challenge one another. Multi-agent debate offers a similar proxy setup in which language models (as agents) can do the same. Instead of querying a single model once, several agents propose answers, argue over possible solutions for a few rounds, and then settle on a decision. It sounds appealing, yet whether it actually helps remains an open problem. This talk presents what we have learned by studying multi-agent debate systematically, using MALLM, an open-source framework that decomposes a debate into four components: who the agents are, how they respond, how they exchange messages, and how they decide. In MALLM, different setups can be systematically compared and analyzed. We derive four studies from MALLM: how agents should settle on an answer (voting versus consensus); what happens when debates run long ("problem drift" that stems from how agents talk); whether giving agents personalities helps; and how debate behaves when the setting turns adversarial or an agent reasons from a false premise. The latter allows us to probe deception, persuasion, and misinformation as early safety signals. How agents interact matters as much as which model we use, and "more" rounds, stronger personas, and longer debates do not translate into performance gains. Studied carefully, these systems reveal social failure modes we can measure and begin to fix if we want to improve our current systems.

Bio

PD Dr. Terry Ruas is a research group leader at the University of Göttingen and the State and University Library (Germany), at the Scientific Information Analytics Chair (Prof. Bela Gipp, GippLab). His research focuses on natural language processing (NLP) and artificial intelligence (AI), investigating how AI usage influences human norms in writing, reading, and problem-solving.

PD Dr. Ruas obtained his Habilitation (Venia Legendi) at the University of Göttingen, a Ph.D. in Computer Science at the University of Michigan (USA), a master’s in Information Engineering at the Federal University of ABC (Brazil), and a Bachelor’s in Computer Science and Science & Technology at the same institution. He also worked as a visiting researcher at the National Institute of Informatics (Japan) at the Aizawa Laboratory. In the industry, he worked at IBM (Brazil) for six years in various positions, including software product manager, team leader, and (certified) IT specialist. His research interests primarily lie at the intersection of NLP, AI, and data science. More specifically, he is passionate about research challenges that require extracting and applying semantic features from textual data to solve significant problems such as paraphrase generation and detection, responsible AI, scientometrics, media bias, text generation, and summarization.

Please register here before 18/05 if you wish to attend.​

Time: Thursday May 21 (15:00-17:00)

Location: S.KS. 203

Abstract (English version below)

Tijdens haar praatje zal ze het kader schetsen van een PhD in DH binnen 5 jaar pijlsnelle technologische verandering aan de hand van onderzoek naar het paraplubegrip Artificiële Intelligentie in Digital Humanities en NLP. Hoe wordt de term tegenwoordig gebruikt in onderzoek? Zien we een uitholling van het technische jargon? Daarnaast worden enkele lopende projecten met een NLP-component toegelicht, waaronder RedressHub, het Navez-project en BelHisFirm, die stuk voor stuk andere datatypes en (historische) talen verwerken met verschillende socio-historische onderzoeksdoeleinden. De nadruk zal hierbij worden gelegd op procesmatige ervaringen, gebruikte modellen en resultaten bij de ontwikkeling van de respectievelijke pipelines, maar ook op de zin en de onzin van het methodologische perspectief in DH. We eindigen met een probabilistische blik op de toekomst.
---
In this talk, Tess Dejaeghere will outline what it means to pursue a PhD in Digital Humanities during five years of rapid technological change, drawing on research into the umbrella concept of Artificial Intelligence in Digital Humanities and NLP (Natural Language Processing). How is the term currently used in research? Are we witnessing a dilution of technical jargon?In addition, several ongoing projects with an NLP component will be discussed, including RedressHub, the Navez project, and BelHisFirm. These projects process different data types and (historical) languages for a variety of socio-historical research purposes. Particular attention will be paid to practical experiences, models used, and results obtained during the development of the respective pipelines, as well as to the value and limitations of methodological perspectives in Digital Humanities. The talk concludes with a probabilistic outlook on the future.

Bio (English version below)

Tess Dejaeghere is doctoraatsstudente Digital Humanities aan de UGent, en actief bij de onderzoeksgroepen Ghent Center for Digital Humanities en LT3 (Language Translation and Technology Team). Ze behaalde respectievelijk een MA in tolken (VUB), Digital Humanities (KUL) en Digital Text Analysis (UA), en spitst zich nu toe op de ontwikkeling van tekstanalytische methodes voor de humane wetenschappen.
---
Tess Dejaeghere is a PhD student in Digital Humanities at Ghent University and is affiliated with the Ghent Center for Digital Humanities and LT3 (Language Translation and Technology Team). She holds MA degrees in Interpreting (VUB), Digital Humanities (KUL), and Digital Text Analysis (UA). Her current research focuses on the development of text-analytical methods for the humanities.