MT model space: statistical versus compositional versus example-based machine translation

Abstract

We offer a perspective on EBMT from a statistical MT standpoint, by developing a three-dimensional MT model space based on three pairs of definitions:(1) logical versus statistical MT, (2) schema-based versus example-based MT, and (3) lexical versus compositional MT. Within this space we consider the interplay of three key ideas in the evolution of transfer, example-based, and statistical approaches to MT. We depict how all translation models face these issues in one way or another, regardless of the school of thought, and suggest where the real questions for the future may lie.

Hybrid data-driven models of machine translation

“…Marker-based EBMT is placed closer to the schema-based techniques than its earlier 'pure' counterparts and is also collocational (with some more compositional technlques than phrasebased SMT). Wu (2006) states that classical EBMT models make nontnvial use of a large library of examples at runtime rather than during training, essentially memorizing data rather than abstractmg away from it, thus agreeing with the definit~ons of Somers (1999Somers ( , 2003 and Turcato and Popowich (2003). SMT models, defined as making nontrivial use of mathematical statistics and probabihty, on the other hand, attempt to do just that, being more schema-based.…”

“…In response to the question presented in (RQl), we concluded, following the definitions presented by Wu (2006), that both modern EBMT and modern SMT methods constitute hybrid models of MT, as they borrow techniques from various disciplines. As SMT methods now make use of phrasal information (Koehn et al, 2003), they have become more EBMT-like, although they are still easily identifiable by their use of statistical modeling techniques in their distinct translation and language models.…”

“…The marker-based approach to EBMT extracts sets of examples, generalised templates and lexical alignments during a preprocessing stage, which has much in common with phrase-based SMT approaches (indicated by its position with regards to the z-ax~s in Figure 2.11). In addition, the use of weights, despite not qualifying as true probabilities, during its recombination stage also represents an SMT element, according to Wu's definition (Wu, 2006) (indicated by its position on the z-axis in Figure 2.11). However, the presence of generalised templates, together with a distinctive recombination stage and the lack of true probability models, make it more akin to EBMT that to SMT.…”

“…However, the presence of generalised templates, together with a distinctive recombination stage and the lack of true probability models, make it more akin to EBMT that to SMT. At the same time, as SMT models begin to make use of phrase alignments, which can also be considered as types of lexical collocation translation rules, the more EBMT-like they become (Wu, 2006) (cf. Figure 2.11).…”

“…Given the results from Chapter 3, and those presented in this Chapter, the implications for designers of data-driven MT systems are obvious: if available, incorporating sub-sentential resources from both SMT and EBMT into novel hybrid 'example-based SMT' and 'statistical EBMT' systems guarantee that systems will be derived that are capable of higher translation quality than either standalone translation variant. That is, while there is an obvious convergence between both paradigmatic variants (Way and Gough, 2005a;Wu, 2006), more gains are to be had from combining their relative strengths in novel hybrid systems.…”