Word Lattices for Multi-Source Translation

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 719–727, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Word Lattices for Multi-Source Translation Josh Schroeder, Trevor Cohn, and Philipp Koehn School of Informatics University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB Scotland, United Kingdom {jschroe1, tcohn, pkoehn}@inf.ed.ac.uk Abstract Multi-source statistical machine transla- tion is the process of generating a single translation from multiple inputs. Previous work has focused primarily on selecting from potential outputs of separate transla- tion systems, and solely on multi-parallel corpora and test sets. We demonstrate how multi-source translation can be adapted for multiple monolingual inputs. We also ex- amine different approaches to dealing with multiple sources, including consensus de- coding, and we present a novel method of input combination to generate lattices for multi-source translation within a single translation model. 1 Introduction Multi-source statistical machine translation was first formally defined by Och and Ney (2001) as the process of translating multiple meaning- equivalent source language texts into a single tar- get language. Multi-source translation is of par- ticular use when translating a document that has already been translated into several languages, ei- ther by humans or machines, and needs to be fur- ther translated into other target languages. This situation occurs often in large multi-lingual organ- isations such as the United Nations and the Euro- pean Parliament, which must translate their pro- ceedings into the languages of the member in- stitutions. It is also common in multi-national companies, which need to translate product and marketing documentation for their different mar- kets. Clearly, any existing translations for a docu- ment can help automatic translation into other lan- guages. These different versions of the input can resolve deficiencies and ambiguities (e.g., syntac- tic and semantic ambiguity) present in a single in- put, resulting in higher quality translation output. In this paper, we present three models of multi- source translation, with increasing degrees of so- phistication, which we compare empirically on a number of different corpora. We generalize the definition of multi-source translation to include any translation case with multiple inputs and a sin- gle output, allowing for, e.g., multiple paraphrased inputs in a single language. Our methods include simple output selection, which treats the multi- source translation task as many independent trans- lation steps followed by selection of one of their outputs (Och and Ney, 2001), and output combina- tion, which uses consensus decoding to construct a string from n-gram fragments of the translation outputs (Bangalore et al., 2001). We also present a novel method, input combination, in which we compile the input texts into a compact lattice, over which we perform a single decoding pass. We show that as we add additional inputs, the simplest output selection method performs quite poorly rel- ative to a single input translation system, while the latter two methods are able to make better use of the additional inputs. The paper is structured as follows. §2 presents the three methods for multi-source …

Content truncated to 500 words