Pivot-based triangulation for low-resource languages

Pivot-based Triangulation for Low-Resource Languages Rohit Dholakia rdholaki@cs.sfu.ca Anoop Sarkar anoop@cs.sfu.ca School of Computing Science, Simon Fraser University, Burnaby, V5A 1S6, Canada Abstract This paper conducts a comprehensive study on the use of triangulation for four very low- resource languages: Mawukakan and Maninkakan, Haitian Kreyol and Malagasy. To the best of our knowledge, ours is the first effective translation system for the first two of these languages. We improve translation quality by adding data using pivot languages and exper- imentally compare previously proposed triangulation design options. Furthermore, since the low-resource language pair and pivot language pair data typically come from very different domains, we use insights from domain adaptation to tune the weighted mixture of direct and pivot based phrase pairs to improve translation quality. 1 Introduction Triangulation for phrase-based statistical machine translation (SMT) [Utiyama and Isahara, 2007, Cohn and Lapata, 2007, Wu and Wang, 2007] refers to the use of a pivot language when translating from a source language to a target language. Previous research into triangulation for machine translation either used Europarl [Cohn and Lapata, 2007, Utiyama and Isahara, 2007, Huck and Ney, 2012] or languages with large corpora in same domain [Gispert and Mario, 2006] or as- sumed presence of languages that are very closely related [Nakov and Ng, 2012, Wang et al., 2012]. However, low resource languages are quite different when it comes to the kind and size of parallel data that is available. This paper considers machine translation into English from four diverse low-resource languages: Mawukakan and Maninkakan, which are West African languages; Haitian Kreyol, in the domain of short messages sent in the aftermath of the Haiti earthquake in 2010; and Malagasy, an Austronesian language from Madagascar. This is the first comprehensive study of triangulation for these four languages, and to our best knowledge, Mawukakan and Maninkakan have not been studied before in the SMT literature. Al-Onaizan & Simard (Eds.) Proceedings of AMTA 2014, vol. 1: MT Researchers Vancouver, BC © The Authors 315 Faced with a low-resource language pair, several questions arise when trying to use the approach of triangulation: • [Utiyama and Isahara, 2007] use a different way of computing lexical scores from [Cohn and Lapata, 2007]. Which one is better suited for triangulation in a resource-poor scenario? • In [Utiyama and Isahara, 2007, Cohn and Lapata, 2007, Wang et al., 2012, Wu and Wang, 2007] many different feature functions are provided for the log- linear model over triangulated phrase pairs. We conduct extensive experiments to show which features should be used for real world low-resource languages based on the data settings for each language pair. • In [Cohn and Lapata, 2007] a mixture of the direct system and the triangulated system is shown to work better. However, they used uniform weights. In [Wang et al., 2012] a few different weights were selected heuristically while in [Wu and Wang, 2007] 0.9 is assumed for the baseline. We provide an algorithm that combines grid search for learning the mixture weights and minimum error …

Content truncated to 500 words