Mixed domain vs. multi-domain statistical machine translation

Mixed-Domain vs. Multi-Domain Statistical Machine Translation Matthias Huck mhuck@inf.ed.ac.uk Alexandra Birch a.birch@ed.ac.uk Barry Haddow bhaddow@inf.ed.ac.uk School of Informatics, University of Edinburgh, 10 Crichton Street, Edinburgh EH8 9AB, UK Abstract Domain adaptation boosts translation quality on in-domain data, but translation quality for domain adapted systems on out-of-domain data tends to suffer. Users of web-based translation services expect high quality translation across a wide range of diverse domains, and what makes the task even more difficult is that no domain label is provided with the translation request. In this paper we present an approach to domain adaptation which results in large-scale, general purpose machine translation systems. First, we tune our translation models to multiple individ- ual domains. Then, by means of source-side domain classification, we are able to predict the domain of individual input sentences and thereby select the appropriate domain-specific model parameters. We call this approach multi-domain translation. We develop state-of-the-art, domain-adapted translation engines for three broadly-defined do- mains: TED talks, Europarl, and News. Our results suggest that multi-domain translation performs better than a mixed-domain approach, which deploys a system that has been tuned on a development set composed of samples from many domains. 1 Introduction Domain adaptation is a common approach to significantly improve machine translation quality on input documents from a given domain. Domain adaptation techniques for statistical ma- chine translation (SMT) have been extensively studied and are well established (Federico and Bertoldi, 2012). In practice, machine translation systems are often engineered to perform well on the domain of one specific application. Most research on domain adaptation assumes that any prospective input data originates from a single domain, and the characteristics of this do- main are known beforehand, e.g. by means of existing samples from the same domain which can be employed for training and tuning. The adaptation task is then defined as utilizing a small amount of in-domain training resources effectively in order to learn system parameters that are more appropriate for translating in-domain input. The in-domain training resources constitute a minor fraction of the overall training data only, the majority of which has a domain mismatch with the designated application. The downside of systems that have been highly tweaked towards the characteristics of a single domain is a diminished translation quality on out-of-domain data (Haddow and Koehn, 2012). Online translation systems, on the other hand, are usually designed for open-domain scenarios where the domain of the input text is not predefined. Being able to take advantage of the benefits of domain adaptation while not having to compromise quality on out-of-domain data would be desirable for online systems. Proceedings of MT Summit XV, vol.1: MT Researchers' Track Miami, Oct 30 - Nov 3, 2015 | p. 240 A viable utilization of domain adaptation approaches in open-domain online translation systems comes in two components: • A number of different parameter sets, each tuned to optimize translation quality on texts from a specific domain. • A text classifier that predicts the domain of foreign-language input data prior to decoding. As …

Content truncated to 500 words