🧑‍🔬 Miles Osborne

Miles Osborne works on statistical machine translation, with a particular interest in how translation systems are built, tuned and judged. Their papers here deal with phrase-based models, word alignment, reordering and the automatic metrics used to measure output quality. They also cover small-corpus and spoken-language settings alongside the large text collections more usual in the field.

The most cited of these is the Edinburgh System Description for the 2005 IWSLT Speech Translation Evaluation. It describes a phrase-based log-linear system entered in a limited-domain speech task, translating from Arabic, Japanese, Korean and Chinese into English and from English into Chinese. Among its additions was a lexicalised model of how neighbouring phrases change order, and the model weights were tuned directly against an error measure. The paper also observed that alignment heuristics suited to large corpora did not always carry over when training data was scarce. The companion paper, Re-evaluating the Role of Bleu in Machine Translation Research from 2006, asks whether Bleu tracks human judgments of quality. It argues that the metric barely constrains n-gram order or which reference a match comes from, so many very different, even ungrammatical, translations can receive identical scores. Counterexamples from evaluation campaigns support the argument. Together the two papers move from building systems toward questioning the evaluation practices used to tune and compare them.

Osborne is based at the University of Edinburgh, with earlier bylines at the University of York and the University of Groningen. Co-authors on these machine translation papers include Philipp Koehn and Chris Callison-Burch, as well as Amittai Axelrod and David Talbot.

Miles Osborne has 2 papers in our library, in 2005–2006, most of them on Machine translation, Evaluation & benchmarks, Speech & spoken language; often with Chris Callison-Burch, Philipp Koehn, Amittai Axelrod.