Researchers Martin Feldkircher, Márton Kardos and Petr Koráb propose three recently new techniques to refresh the NLP toolbox for 2026. Seeded topic modeling lets users steer models with free‑text prompts, while LLM‑generated summaries sharpen topic coherence. They also showcase automatic topic naming via generative language models on a central‑bank speech corpus.

Traditional statistical pipelines such as FASTopic have reduced junk topics, yet neural models remain unstable and opaque. By encoding documents and seed phrases with sentence‑transformers, the authors’ KeyNMF factorization builds a keyword matrix that is re‑weighted according to document relevance scores. The result: focused, stable topics without heavy preprocessing.

Applying the pipeline to 279 ECB press‑conference statements demonstrates economic‑meaningful clusters such as euro‑zone expansion and inflation outlook. Because the approach blends probabilistic ML efficiency with LLM interpretability, banks can gain transparent insights while avoiding the cost of full‑scale language‑model runs. Future work will test scalability on larger policy corpora.