Chunk-Based EBMT

Chunk-Based EBMT Jae Dong Kim Ralf D. Brown Jaime G. Carbonell Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213 {jdkim,ralf,jgc}@cs.cmu.edu Abstract Corpus driven machine translation ap- proaches such as Phrase-Based Statistical Machine Translation and Example-Based Machine Translation have been successful by using word alignment to find transla- tion fragments for matched source parts in a bilingual training corpus. However, they still cannot properly deal with systematic translation for insertion or deletion words between two distant languages. In this work, we used syntactic chunks as trans- lation units to alleviate this problem, im- prove alignments and show improvement in BLEU for Korean to English and Chi- nese to English translation tasks. 1 Introduction In general, corpus-based machine translation sys- tems prefer longer units because they naturally convey local context and local reordering. This was achieved by phrases in Phrase-Based Statis- tical Machine Translation (Koehn et al., 2007; Vo- gel et al., 2003) and surface form matches in lex- ical Example-Based Machine Translation (Brown, 2005; Veale and Way, 1997). These systems use phrasal alignment to find translations of matched n-grams for an input sentence. However, because the alignment algorithms used in these systems purely depend on word alignment (Brown et al., 1993) they cannot ad- dress structural translations, other than hoping for structural parallelism between source and target. For example, these algorithms cannot reliably find ’an/the office’ as a translation of ’sa-moo-sil’ in c⃝ 2010 European Association for Machine Translation. Korean to English translation because Korean does not have articles. For this reason, we investigated the chunk as our translation unit. The chunk was pioneered by Ab- ney (1991). It is a continuous and non-recursive syntactic segment around a head, comparable to a morphologically complex word in synthetic lan- guages, and is not contained in any other chunks. A typical chunk consists of a single content word surrounded by related function words matching a template, though it could contain modifiers. We have observed several advantages by using chunks as basic translation units. First, we can to some degree systematically translate untrans- latable tokens (words, morphemes) that exist only in one side. For example, when we translate an English sentence into Korean, the word-to-word translation systems cannot produce a nominative case marker in Korean unless rules are given by human experts or the MT system “hallucinates” markers and uses the target language modeling to guess whether or not the case marker should in fact be present. Second, as chunks are n-gram phrases, they convey local reordering and context as well, though this advantage is also true of n-grams in phrasal translation. Third, the number of chunks may better match across languages than the num- ber of words, which may yield better alignment at the chunk level. Fourth, since the order of chunks is more flexible than the order of words within a chunk, it has more flexibility in re-ordering than arbitrary n-grams crossing syntactic chunk bound- aries. This is an important advantage when we translate from or into a relatively less strict …

Content truncated to 500 words