Natural Language Processing (1/2)

Friday 21 Nov 2025, 14:15 → 15:45 Europe/Berlin

DO: JvF25/3-303 | BN: b-it/1.047

NLP and How It Was Solved by LLMs. Or Was IT? by Akbar Karimi

Natural language processing has seen major milestones in its history, the last of which was the development of large language models. When they came around, it seemed that NLP was solved as they were capable of tackling many NLP tasks of the time. But as they were improving, it turned out that they were not the panacea we were expecting. On the contrary, we had opened a Pandora’s box of many problems yet to come, including hallucination, misinformation, biases, privacy concerns, and most important of all, the existential risks of AI. In this lecture, we will briefly discuss what challenges NLP can address now and how these challenges have evolved over time. 

LLMs Get It Wrong (But Humans Do Too): Evaluating Hallucinations across Languages by Patrícia Schmidtova

What counts as a hallucination, and who gets to decide? I’ll examine how definitions, automatic metrics, and human judgments interact when evaluating hallucinations in summarization, and discuss the main sources of annotator disagreement that arise in span-level annotation. I’ll then connect these observations to the challenges of cross-lingual summarization, where the limits of current LLMs become especially clear.