Dr. Lucia Specia - Quality Estimation of Machine Translation - 4th July 2011

Video Category: 
Research Seminar Talk

Quality Estimation for Machine Translation

Abstract: Although significant progress has been observed in the overall quality of machine translation (MT) systems in recent years, it is well known that the quality of a given MT system can vary considerably across translation segments. The notion of quality and the level of “acceptability” of translations depend on the intended use of the translated segments, ranging from post-editing by fluent speakers of both source and target languages to gisting by speakers of the target language only. The need for models to assess the quality of translated segments in the various possible scenarios involving the use of automatic translations is becoming more and more evident and a number of efforts have been made recently in this direction.

The most obvious uses of such models include filtering out bad quality translations from a translation post-editing workflow (estimating translation post-editing effort) and informing readers who cannot understand the source text (estimating translation adequacy). In this talk I will present some of my recent work on modelling the problem of quality estimation for these and other applications. I will present the outcomes of a number of experiments with different language pairs and text genres/domains which demonstrate that it is possible to predict different aspects of quality using a relatively small number of training examples and standard supervised machine learning techniques.

Bio: Lucia Specia is senior lecturer at the Research Institute in Information and Language Processing, University of Wolverhampton, UK. She is part of the Research Group in Computational Linguistics, headed by Prof. Ruslan Mitkov. Before coming to Wolverhampton, she worked as research engineer with the Cross-Language Technology Group, at the Xerox Research Centre Europe, France. She received a PhD from the University of São Paulo, Brazil, in 2007, where she was supervised by Prof. Maria das Graças Volpe Nunes and Dr. Mark Stevenson, from the Natural Language Processing Group, Sheffield University, UK. Lucia Specia is currently the principal investigator of a project on estimating the adequacy of Arabic-English machine translation, supported by DARPA, under GALE scheme, as part of the IBM team.