Enhancing scarce-resource language translation through pivot combinations

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 1361–1365, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Enhancing scarce-resource language translation through pivot combinations Marta R. Costa-juss`a Barcelona Media Innovation Center Av. Diagonal, 177 08018 Barcelona marta.ruiz@barcelonamedia.org Carlos Henr´ıquez TALP-UPC Jordi Girona, s/n 08034 Barcelona carlos.henriquez@upc.edu Rafael E. Banchs Institute for Infocomm Research 1 Fusionopolis Way 21-01 Singapore 138632 rembanchs@i2r.a-star.sg.edu Abstract Chinese and Spanish are the most spoken lan- guages in the world. However, there is not much research done in machine translation for this language pair. We experiment with the parallel Chinese-Spanish corpus (United Na- tions) to explore alternatives of SMT strate- gies which consist on using a pivot language. Particularly, two well-known alternatives are shown for pivoting: the cascade system and the pseudo-corpus. As Pivot language we use English, Arabic and French. Results show that English is the best pivot language between Chinese and Spanish. As a new strategy, we propose to perform a combination of the pivot strategies which is capable to highly outper- form the direct translation strategy. 1 Introduction Although they are very distant languages, Chinese and Spanish are very close to each other in the rank- ing of the most spoken languages in the world1. Nevertheless, when interested in bilingual resources between these two languages they become far apart again. Similarly, the related amount of work we have found within the computational linguistic com- munity, can be reduced to a very small set of refer- ences.The most popular research event recently per- formed was the 2008 IWSLT evaluation campaign 2. This evaluation organized two Chinese-to-Spanish tracks. One of them was focused on direct trans- lation and the other one on pivot translation through 1www.ethnologue.org/ethno docs/distribution.asp?by=size 2http://mastarpj.nict.go.jp/IWSLT2008/ English. Best translation results were obtained by far in the pivot task. The best system in the pivot task (Wang et al., 2008) compared two different approaches: The first one, training two translation models on the Chinese-English corpus and English- Spanish corpus, and then building a pivot translation model for Chinese-Spanish translation using English as a pivot language as proposed in (Wu and Wang, 2007); the second one obtained better results and it was based on a cascade approach. The idea here is to translate from Chinese into English and then from English to Spanish, which means performing two translations. Besides the research mentioned above, which directly addressed the Chinese-Spanish lan- guage pair, we may also find in the literature another approach similar to Wu’s (2007) authored by Cohn and Lapata (2007). Basically, they also used sev- eral intermediate pivot language to create source-to- target phrases that are lately interpolate with a direct system build with a source-to-target parallel corpus. Apart from the BTEC3 corpus available through the IWSLT4 competition and Holy Bible datasets de- scribed in (Paul, 2008) and (Banchs and Li, 2008), respectively, there is a recent release of a six lan- guage parallel corpus (including both Chinese and Spanish) from United Nations (UN) for research purposes (Rafalovith and Dale, …

Content truncated to 500 words