Ensemble Triangulation for Statistical Machine Translation

International Joint Conference on Natural Language Processing, pages 252–260, Nagoya, Japan, 14-18 October 2013. Ensemble Triangulation for Statistical Machine Translation∗ Majid Razmara and Anoop Sarkar School of Computing Science Simon Fraser University Burnaby, BC, Canada {razmara,anoop}@sfu.ca Abstract State-of-the-art statistical machine transla- tion systems rely heavily on training data and insufficient training data usually re- sults in poor translation quality. One so- lution to alleviate this problem is triangu- lation. Triangulation uses a third language as a pivot through which another source- target translation system can be built. In this paper, we dynamically create multi- ple such triangulated systems and combine them using a novel approach called ensem- ble decoding. Experimental results of this approach show significant improvements in the BLEU score over the direct source- target system. Our approach also outper- forms a strong linear mixture baseline. 1 Introduction The objective of current statistical machine trans- lation (SMT) systems is to build cheap and rapid corpus-based SMT systems without involving hu- man translation expertise. Such SMT systems rely heavily on their training data. State-of-the- art SMT systems automatically extract transla- tion rules (e.g. phrase pairs), learn segmentation models, re-ordering models, etc. and find tuning weights solely from data and hence they rely heav- ily on high quality training data. There are many language pairs for which there is no parallel data or the available data is not sufficiently large to build a reliable SMT system. For example, there is no Chinese-Farsi parallel text, although there ex- ists sufficient parallel data between these two lan- guages and English. For SMT, an important re- search direction is to improve the quality of trans- lation when there is no, insufficient or poor-quality parallel data between a pair of languages. ∗This research was partially supported by an NSERC, Canada (RGPIN: 264905) grant and a Google Faculty Award to the second author. One approach that has been recently proposed is triangulation. Triangulation is the process of translating from a source language to a target lan- guage via an intermediate language (aka pivot, or bridge). This is very useful specifically for low- resource languages as SMT systems built using small parallel corpora perform poorly due to data sparsity. In addition, ambiguities in translating from one language into another may disappear if a translation into some other language is available. One obvious benefit of triangulation is to in- crease the coverage of the model on the input text. In other words, we can reduce the number of out-of-vocabulary words (OOVs), which are a ma- jor cause of poor quality translations, using other paths to the target language. This can be espe- cially helpful when the model is built using a small amount of parallel data. Figure 1 shows how triangulation can be useful in reducing the number of OOVs when translating from French to English through three pivot lan- guages: Spanish (es), German (de) and Italian (it). The solid lines show the number of OOVs for a direct MT system with regard to a multi-language parallel test …

Content truncated to 500 words