Microsoft Research Treelet Translation System: IWSLT Evaluation

Microsoft Research Treelet Translation System: IWSLT Evaluation Arul Menezes and Chris Quirk Microsoft Research One Microsoft Way Redmond, WA 98052 {arulm,chrisq}@microsoft.com Abstract The Microsoft Research translation system is a syntactically informed phrasal SMT system that uses a phrase translation model based on dependency treelets and a global reordering model based on the source dependency tree. These models are combined with several other knowledge sources in a log- linear manner. The weights of the individual components in the log-linear model are set by an automatic parameter-tuning method. We give a brief overview of the components of the system and discuss its performance at IWSLT in two tracks: Japanese to English (supplied data and tools), and English to Chinese (supplied data and tools). 1. Introduction The dependency treelet translation system developed at MSR is a statistical MT system that takes advantage of linguistic tools, namely a source language dependency parser, as well as a word alignment component. [1] To train a translation system, we require a sentence- aligned parallel corpus. First the source side is parsed to obtain dependency trees. Next the corpus is word-aligned, and the source dependencies are projected onto the target sentences using the word alignments. From the aligned dependency corpus we extract all treelet translation pairs, and train an order model and a bi-lexical dependency model. To translate, we parse the input sentence, and employ a decoder to find a combination and ordering of treelet translation pairs that cover the source tree and are optimal according to a set of models. In a now-common generalization of the classic noisy-channel framework, we use a log-linear combination of models [2], as in below:      = ∑ ∈F T ST FΛ S f f f translation (,) λ argmax ) (,, (3) Such an approach toward translation scoring has proven very effective in practice, as it allows a translation system to incorporate information from a variety of probabilistic or non- probabilistic sources. The weights Λ = { λf }are selected by discriminatively training against held out data. 2. System Details A brief word on notation: s and t represent source and target lexical nodes; S and T represent source and target trees; s and t represent source and target treelets (connected subgraphs of the dependency tree). Where the intent is clear, we will disregard the structure of these elements and consider these structures to be sets of lexical items: the expression ∀t∈ T refers to all the lexical items in the target language tree T. Similarly, |T| refers to the count of lexical items in T. We use subscripts to indicate selected words: Tn represents the nth lexical item in an in-order traversal of T. 2.1. Training We use the broad coverage dependency parser NLPWIN [3] to obtain source language dependency trees, and we use GIZA++ [4] to produce word alignments. The GIZA++ training regimen and parameters are tuned to optimize BLEU [5] scores on held-out data. Using the word alignments, we follow a set of dependency …

Content truncated to 500 words