Synchronous Tree Adjoining Machine Translation

Proceedings of the 2009 Conference on Empirical Methods in Natural Language Processing, pages 727–736, Singapore, 6-7 August 2009. c⃝2009 ACL and AFNLP Synchronous Tree Adjoining Machine Translation Steve DeNeefe and Kevin Knight USC Information Sciences Institute 4676 Admiralty Way, Suite 1001 Marina del Rey, CA 90292 USA {sdeneefe,knight}@isi.edu Abstract Tree Adjoining Grammars have well-known advantages, but are typically considered too difficult for practical systems. We demon- strate that, when done right, adjoining im- proves translation quality without becoming computationally intractable. Using adjoining to model optionality allows general translation patterns to be learned without the clutter of endless variations of optional material. The appropriate modifiers can later be spliced in as needed. In this paper, we describe a novel method for learning a type of Synchronous Tree Ad- joining Grammar and associated probabilities from aligned tree/string training data. We in- troduce a method of converting these gram- mars to a weakly equivalent tree transducer for decoding. Finally, we show that adjoining results in an end-to-end improvement of +0.8 BLEU over a baseline statistical syntax-based MT model on a large-scale Arabic/English MT task. 1 Introduction Statistical MT has changed a lot in recent years. We have seen quick progress from manually crafted linguistic models to empirically learned statistical models, from word-based models to phrase-based models, and from string-based mod- els to tree-based models. Recently there is a swing back to incorporating more linguistic information again, but this time linguistic insight carefully guides the setup of empirically learned models. Shieber (2007) recently argued that proba- bilistic Synchronous Tree Adjoining Grammars (Shieber and Schabes, 1990) have the right com- bination of properties that satisfy both linguists and empirical MT practitioners. So far, though, most work in this area has been either more lin- guistic than statistical (Abeille et al., 1990) or statistically-based, but linguistically light (Nesson et al., 2006). Current tree-based models that integrate lin- guistics and statistics, such as GHKM (Galley et al., 2004), are not able to generalize well from a single phrase pair. For example, from the data in Figure 1, GHKM can learn rule (a) to translate nouns with two pre-modifiers, but does not gener- alize to learn translation rules (b) - (d) without the optional adjective or noun modifiers. Likewise, none of these rules allow extra material to be intro- duced, e.g. “Pakistan’s national defense minister”. In large enough training data sets, we see many examples of all the common patterns, but the rarer patterns have sparse statistics or poor coverage. NP JJ national NN defense NN minister wzyr AldfAE AlwTnY (a) NP JJ1 NN2 NN3 ↔ NN3 NN2 JJ1 (b) NP NN1 NN2 ↔ NN2 NN1 (c) NP JJ1 NN2 ↔ NN2 JJ1 (d) NP NN1 ↔ NN1 Figure 1: Rule (a) can be learned from this training example. Arguably, the more general rules (b) - (d) should also be learnable. To mitigate this problem, the parse trees used as training data for these systems can be binarized (Wang et al., 2007). Binarization allows rules with partial constituents …

Content truncated to 500 words