Innovating Works

DASMT

Financiado
Domain Adaptation for Statistical Machine Translation
Rapid translation between European languages is a cornerstone of good governance in the EU, and of great academic and commercial interest. Statistical approaches to machine translation constitute the state-of-the-art. The basic kn... Rapid translation between European languages is a cornerstone of good governance in the EU, and of great academic and commercial interest. Statistical approaches to machine translation constitute the state-of-the-art. The basic knowledge source is a parallel corpus, texts and their translations. For domains where large parallel corpora are available, such as the proceedings of the European Parliament, a high level of translation quality is reached. However, in countless other domains where large parallel corpora are not available, such as medical literature or legal decisions, translation quality is unacceptably poor. Domain adaptation as a problem of statistical machine translation (SMT) is a relatively new research area, and there are no standard solutions. The literature contains inconsistent results and heuristics are widely used. We will solve the problem of domain adaptation for SMT on a larger scale than has been previously attempted, and base our results on standardized corpora and open source translation systems. We will solve two basic problems. The first problem is determining how to benefit from large out-of-domain parallel corpora in domain-specific translation systems. This is an unsolved problem. The second problem is mining and appropriately weighting knowledge available from in-domain texts which are not parallel. While there is initial promising work on mining, weighting is not well studied, an omission which we will correct. We will scale mining by first using Wikipedia, and then mining from the entire web. Our work will lead to a break-through in translation quality for the vast number of domains with less parallel text available, and have a direct impact on SMEs providing translation services. The academic impact of our work will be large because solutions to the challenge of domain adaptation apply to all natural language processing systems and in numerous other areas of artificial intelligence research based on machine learning approaches. ver más
30/11/2021
LMU MUENCHEN
1M€
Duración del proyecto: 80 meses Fecha Inicio: 2015-03-17
Fecha Fin: 2021-11-30

Línea de financiación: concedida

El organismo H2020 notifico la concesión del proyecto el día 2021-11-30
Línea de financiación objetivo El proyecto se financió a través de la siguiente ayuda:
ERC-StG-2014: ERC Starting Grant
Cerrada hace 10 años
Presupuesto El presupuesto total del proyecto asciende a 1M€
Líder del proyecto
LUDWIGMAXIMILIANSUNIVERSITAET MUENCHEN No se ha especificado una descripción o un objeto social para esta compañía.
Perfil tecnológico TRL 4-5