Synchronous binarization for machine translation

Abstract

Systems based on synchronous grammars and tree transducers promise to improve the quality of statistical machine translation output, but are often very computationally intensive. The complexity is exponential in the size of individual grammar rules due to arbitrary re-orderings between the two languages, and rules extracted from parallel corpora can be quite large. We devise a linear-time algorithm for factoring syntactic re-orderings by binarizing synchronous rules when possible and show that the resulting rule set significantly improves the speed and accuracy of a state-of-the-art syntax-based machine translation system.

Weight pushing and binarization for fixed-grammar parsing

“…Zhang et al (2006) introduced a synchronous binarization technique that improved decoding efficiency and accuracy by ensuring that rule binarization avoided gaps on both the source and target sides (for rules where this was possible). Their binarization was designed to share binarized pieces among rules, but their approach to distributing weight was the default (nondiffused) case found in this paper to be least efficient: The entire weight of the original rule is placed at the top binarized rule and all internal rules are assigned a probability of 1.0.…”

“…Increasing sharing reduces the amount of state that the parser must explore. Binarization has also been investigated in the context of parsing-based approaches to machine translation, where it has been shown that paying careful attention to the binarization scheme can produce much faster decoders (Zhang et al, 2006;Huang, 2007;DeNero et al, 2009).…”

Undirected Machine Translation with Discriminative Reinforcement Learning

“…For ease of presentation, and following synchronous-grammar based MT practice, we will henceforth restrict our focus to binary grammars (Zhang et al, 2006;Wang et al, 2007).…”

Lexicalized Syntactic Reordering Framework for Word Alignment and Machine Translation

“…In Table 2, dot column stands for artificial anchor points in SL sentence, Lw and Rw for previous word and successive word of the current one respectively, and P, LHS, Lw, RHS and Rw constitute the syntactic reordering features of our model. Notice that, inspired by [1] and [11], we assume SL parse trees are binarized before fed into the tree-to-string transformation algorithm. [1] suggests binary-branching ITG rules prune seemingly unlikely and arbitrary word permutations but yet, at the same time, accommodate most meaningful structural reversals during translation.…”