🧑‍🔬 Dragos Stefan Munteanu

Dragos Stefan Munteanu works on statistical machine translation, focusing on the training material such systems need and on what happens when they meet text unlike the text they learned from. A recurring concern in Munteanu's papers here is the gap between the parallel corpora that exist, which are scarce and drawn largely from parliamentary or United Nations proceedings, and the genres, such as news, that people actually want translated.

The most cited of these papers, Improving Machine Translation Performance by Exploiting Non-Parallel Corpora (2005), written with Daniel Marcu, tackles the shortage of parallel text. It searches comparable collections, for example multilingual news feeds, for sentence pairs that genuinely translate each other, so that they can serve as extra training material. A later paper, Measuring Machine Translation Errors in New Domains (2013), moves from supplying more data to explaining failure. It proposes a method for working out why output quality falls so steeply when a system built on parliamentary text is applied to a different domain.

The direction of the work therefore runs from enlarging the pool of usable bilingual data toward diagnosing domain mismatch. Between 2002 and 2013 the papers carry an affiliation with the University of Southern California. Co-authors on the papers here include Marcu, Hal Daumé III and Marine Carpuat.

Dragos Stefan Munteanu has 2 papers in our library, in 2005–2013, most of them on Machine translation; often with Daniel Marcu, Hal Daumé III, Marine Carpuat.