Anchor Graph: Global Reordering Contexts for Statistical Machine Translation

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 501–512, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Anchor Graph: Global Reordering Contexts for Statistical Machine Translation Hendra Setiawan ∗ IBM Research 1101 Kitchawan Road NY 10598, USA Bowen Zhou IBM Research 1101 Kitchawan Road NY 10598, USA Bing Xiang ∗ Thomson Reuters 3 Times Square NY 10036, USA Abstract Reordering poses one of the greatest chal- lenges in Statistical Machine Translation re- search as the key contextual information may well be beyond the confine of translation units. We present the “Anchor Graph” (AG) model where we use a graph structure to model global contextual information that is crucial for reordering. The key ingredient of our AG model is the edges that capture the relation- ship between the reordering around a set of selected translation units, which we refer to as anchors. As the edges link anchors that may span multiple translation units at decoding time, our AG model effectively encodes global contextual information that is previously ab- sent. We integrate our proposed model into a state-of-the-art translation system and demon- strate the efficacy of our proposal in a large- scale Chinese-to-English translation task. 1 Introduction Reordering remains one of the greatest challenges in Statistical Machine Translation (SMT) research as the key contextual information may span across mul- tiple translation units.1 Unfortunately, previous ap- proaches fall short in capturing such cross-unit con- textual information that could be critical in reorder- ing. For example, state-of-the-art translation mod- els, such as Hiero (Chiang, 2005) or Moses (Koehn et al., 2007), are good at capturing local reordering within the confine of a translation unit, but their for- mulation is approximately a simple unigram model ∗ This work was done when the authors were with IBM. 1We define translation units as phrases in phrase-based SMT or as translation rules in syntax-based SMT. over derivation (a sequence of the application of translation units) with some aid from target language models. Moving to a higher order formulation (say to a bigram model) is highly impractical for several reasons: 1) it has to deal with a severe sparsity issue as the size of the unigram model is already huge; and 2) it has to deal with a spurious ambiguity issue which allows multiple derivations of a sentence pair to have radically different model scores. In this paper, we develop “Anchor Graph” (AG) where we use a graph structure to capture global contexts that are crucial for translation. To circum- vent the sparsity issue, we design our model to rely only on contexts from a set of selected translation units, particularly those that appear frequently with important reordering patterns. We refer to the units in this special set as anchors where they act as ver- tices in the graph. To address the spurious ambigu- ity issue, we insist on computing the model score for every anchors in the derivation, including those that appear inside larger translation units, as such our AG …

Content truncated to 500 words