Doktorandské kolokvium KAI - Marek Šuppa (31.3.2025)
v pondelok 31.3.2025 o 13:10 hod. v miestnosti I/9
Prednášajúci: Marek Šuppa
Názov: One Year Later: Did LLMs Solve Multilingual NLP or (Simply) Redefine the Problem?
Termín: 31.3.2025, 13:10 hod., I/9
Abstrakt:
The continued proliferation and advancement of Large Language Models (LLMs) over the past year, marked by releases like Claude 3.7 Sonnet, Gemini 2.5 Pro, GPT-4o, Grok 3 and Llama 3, have further solidified their transformative impact on Natural Language Processing. Tasks once considered challenging are frequently handled with remarkable proficiency. Concurrently, a significant focus has emerged on enhancing reasoning capabilities, exemplified by models such as OpenAI's o3 series and DeepSeek-R1. Despite these strides, the initial observation persists: impressive performance, often benchmarked in English, raises questions about equitable capabilities across the linguistic spectrum, particularly for less-resourced languages.
This presentation revisits the state-of-the-art, specifically asking whether LLMs have truly 'solved' multilingual NLP or if their primary effect has been to redefine the challenges. We will examine the tangible progress observed for languages like Slovak, where new models and dedicated resources have yielded notable improvements, suggesting a brighter outlook, potentially including Slovak-specific models. Our recent findings evaluating these newer LLMs on (not only) Slovak tasks will be presented, illustrating both advancements and persistent gaps. Ultimately, we argue that while LLM capabilities are undeniable, the multilingual NLP problem is far from solved; rather, it has been reshaped, demanding new evaluation paradigms, focus on lower-resource languages, and nuanced approaches beyond simple capability scaling.