Large-scale Exact Decoding: The IMS-TTT submission to WMT14

Proceedings of the Ninth Workshop on Statistical Machine Translation, pages 163–170, Baltimore, Maryland USA, June 26–27, 2014. c⃝2014 Association for Computational Linguistics Large-scale Exact Decoding: The IMS-TTT submission to WMT14∗ Daniel Quernheim IMS University of Stuttgart daniel@ims.uni-stuttgart.de Fabienne Cap CIS Ludwig-Maximilian University of Munich cap@cis.uni-muenchen.de Abstract We present the IMS-TTT submission to WMT14, an experimental statistical tree- to-tree machine translation system based on the multi-bottom up tree transducer in- cluding rule extraction, tuning and decod- ing. Thanks to input parse forests and a “no pruning” strategy during decoding, the obtained translations are competitive. The drawbacks are a restricted coverage of 70% on test data, in part due to ex- act input parse tree matching, and a rela- tively high runtime. Advantages include easy redecoding with a different weight vector, since the full translation forests can be stored after the first decoding pass. 1 Introduction In this contribution, we present an implementation of a translation model that is based on ℓMBOT (the multi bottom-up tree transducer of Arnold and Dauchet (1982) and Lilin (1978)). Intuitively, an MBOT is a synchronous tree sequence substitution grammar (STSSG, Zhang et al. (2008a); Zhang et al. (2008b); Sun et al. (2009)) that has discon- tiguities only on the target side (Maletti, 2011). From an algorithmic point of view, this makes the MBOT more appealing than STSSG as demon- strated by Maletti (2010). Formally, MBOT is expressive enough to express all sensible trans- lations (Maletti, 2012)1. Figure 2 displays sam- ple rules of the MBOT variant, called ℓMBOT, ∗This work was supported by Deutsche Forschungsge- meinschaft grants Models of Morphosyntax for Statistical Machine Translation (Phase 2) and MA/4959/1–1. 1A translation is sensible if it is of linear size increase and can be computed by some (potentially copying) top-down tree transducer. that we use (in a graphical representation of the trees and the alignment). Recently, a shallow ver- sion of MBOT has been integrated into the popular Moses toolkit (Braune et al., 2013). Our imple- mentation is exact in the sense that it does abso- lutely no pruning during decoding and thus pre- serves all translation candidates, while having no mechanism to handle unknown structures. (We added dummy rules that leave unseen lexical ma- terial untranslated.) The coverage is thus limited, but still considerably high. Source-side and target- side syntax restrict the search space so that decod- ing stays tractable. Only the language model scor- ing is implemented as a separate reranker2. This has several advantages: (1) We can use input parse forests (Liu et al., 2009). (2) Not only is the out- put optimal with regard to the theoretical model, also the space of translation candidates can be ef- ficiently stored as a weighted regular tree gram- mar. The best translations can then be extracted using the k-best algorithm by Huang and Chiang (2005). Rule weights can be changed without the need for explicit redecoding, the parameters of the log-linear model can be changed, and even new features can be added. These properties are espe- cially …

Content truncated to 500 words