Improved Reordering for Shallow-n Grammar based Hierarchical Phrase-based Translation

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 533–537, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Improved Reordering for Shallow-n Grammar based Hierarchical Phrase-based Translation Baskaran Sankaran and Anoop Sarkar School of Computing Science Simon Fraser University Burnaby BC. Canada {baskaran, anoop}@cs.sfu.ca Abstract Shallow-n grammars (de Gispert et al., 2010) were introduced to reduce over-generation in the Hiero translation model (Chiang, 2005) re- sulting in much faster decoding and restricting reordering to a desired level for specific lan- guage pairs. However, Shallow-n grammars require parameters which cannot be directly optimized using minimum error-rate tuning by the decoder. This paper introduces some novel improvements to the translation model for Shallow-n grammars. We introduce two rules: a BITG-style reordering glue rule and a simpler monotonic concatenation rule. We use separate features for the new rules in our log- linear model allowing the decoder to directly optimize the feature weights. We show this formulation of Shallow-n hierarchical phrase- based translation is comparable in translation quality to full Hiero-style decoding (without shallow rules) while at the same time being considerably faster. 1 Introduction Hierarchical phrase-based translation (Chiang, 2005; Chiang, 2007) extends the highly lexicalized models from phrase-based translation systems in order to model lexicalized reordering and discon- tiguous phrases. However, a major drawback in this approach, when compared to phrase-based systems, is the total number of rules that are learnt are several orders of magnitude larger than standard phrase tables, which leads to over-generation and search errors and contribute to much longer decoding times. Several approaches have been proposed to address these issues: from filtering the extracted synchronous grammar (Zollmann et al., 2008; He et al., 2009; Iglesias et al., 2009) to alternative Bayesian approaches for learning minimal gram- mars (Blunsom et al., 2008; Blunsom et al., 2009; Sankaran et al., 2011). The idea of Shallow-n gram- mars (de Gispert et al., 2010) takes an orthogonal direction for controlling the over-generation and search space in Hiero decoder by restricting the degree of nesting allowed for Hierarchical rules. We propose an novel statistical model for Shallow-n grammars which does not require addi- tional non-terminals for monotonic re-ordering and also eliminates hand-tuned parameters and instead introduces an automatically tunable alternative. We introduce a BITG-style (Saers et al., 2009) reorder- ing glue rule (§ 3) and a monotonic X-glue rule (§ 4). Our experiments show the resulting Shallow-n decoding is comparable in translation quality to full Hiero-style decoding while at the same time being considerably faster. All the experiments in this paper were done using Kriya (Sankaran et al., 2012) hierarchical phrase- based system which also supports decoding with Shallow-n grammars. We extended Kriya to addi- tionally support reordering glue rules as well. 2 Shallow-n Grammars Formally a Shallow-n grammar G is defined as a 5- tuple: G = (N, T, R, Rg, S), such that T is a set of finite terminals and N a set of finite non-terminals {X0, . . . , …

Content truncated to 500 words