🧑‍🔬 Stefan Riezler

Stefan Riezler works in computational linguistics, mainly on statistical approaches to syntax and grammar formalisms and to machine translation. Their papers here pair rich language-processing systems, such as a hand-built grammar or a phrase-based translation system, with discriminative log-linear models. Two questions recur: how such systems can be trained and made robust on real text, and how differences between their outputs can be measured in a trustworthy way.

The most cited of these papers is Parsing the Wall Street Journal using a Lexical-Functional Grammar and Discriminative Estimation Techniques (2002). It took a stochastic parser driven by a detailed English Lexical-Functional Grammar and applied it to the Penn Wall Street Journal treebank. Rather than converting the treebank into the grammar's own formalism, the work used its shallow brackets and labels to narrow down the parser's candidate analyses. When no complete parse was found, it covered a sentence with as few well-formed chunks as possible. It also capped processing effort, so that long inputs were still handled within reasonable time.

The next most cited work, and the more recent of the two, is On Some Pitfalls in Automatic Evaluation and Significance Testing for MT (2005). Here the focus moves from parsing to evaluation. The paper built a reranker over the candidate translations of a phrase-based system trained on Europarl and tested whether small differences between its variants could be detected reliably. It compared the metrics BLEU, NIST and a dependency-based F-score, and it set the bootstrap test against approximate randomization.

Riezler is currently at Heidelberg University. Earlier bylines include the University of Stuttgart, the Palo Alto Research Center and Google. The parsing paper was written with Ronald M. Kaplan, Tracy Holloway King and Mark Johnson.

Stefan Riezler has 2 papers in our library, in 2002–2005, most of them on Machine translation, Evaluation & benchmarks, Syntax & grammar formalisms; often with Mark Johnson, Ronald M. Kaplan, Tracy Holloway King.