🧑‍🔬 Johannes Leveling

Johannes Leveling works in natural language processing, with papers here on machine translation, multilingual language technology and the pretraining of large language models. A recurring concern is how system choices and evaluation settings shape what can be measured, whether through shared benchmarking campaigns or through tightly controlled training experiments that compare languages and model components.

The most cited of these papers is the Findings of the 2014 Workshop on Statistical Machine Translation. It reports the ninth round of an annual campaign that compares translation systems and the methods used to judge them. That year the campaign covered news translation, quality estimation and a newly added medical translation task. The news task paired English with several languages, including Hindi as a distant, low-resource case. Its test sets were rebuilt so that every sentence came from text originally written in one of the two languages of the pair, which reduces artifacts of translated input. The paper was written with a large group of organisers, among them Ondřej Bojar and Philipp Koehn.

The newer work turns to large language models. Tokenizer Choice For LLM Training: Negligible or Crucial? examines whether the tokenizer behind a decoder-only model affects its quality and training cost. The study varies the algorithm, software library, language coverage and vocabulary size, and trains a separate model for each tokenizer on matched English-only and five-language corpora. Co-authors on this paper include Klaudia Thellmann. Leveling is currently at the Fraunhofer Institute for Intelligent Analysis and Information Systems in Sankt Augustin, part of the Fraunhofer-Gesellschaft. Earlier bylines point to Dublin City University and, before that, FernUniversität in Hagen.

Johannes Leveling has 2 papers in our library, in 2014–2024, most of them on Machine translation, Language models & pretraining, Multilingual & cross-lingual NLP; often with Barry Haddow, Christian Buck, Christian Federmann.