Using Categorial Grammar to Label Translation Rules

Proceedings of the 7th Workshop on Statistical Machine Translation, pages 222–231, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics Using Categorial Grammar to Label Translation Rules Jonathan Weese and Chris Callison-Burch and Adam Lopez Human Language Technology Center of Excellence Johns Hopkins University Abstract Adding syntactic labels to synchronous context-free translation rules can improve performance, but labeling with phrase struc- ture constituents, as in GHKM (Galley et al., 2004), excludes potentially useful translation rules. SAMT (Zollmann and Venugopal, 2006) introduces heuristics to create new non-constituent labels, but these heuristics introduce many complex labels and tend to add rarely-applicable rules to the translation grammar. We introduce a labeling scheme based on categorial grammar, which allows syntactic labeling of many rules with a mini- mal, well-motivated label set. We show that our labeling scheme performs comparably to SAMT on an Urdu–English translation task, yet the label set is an order of magnitude smaller, and translation is twice as fast. 1 Introduction The Hiero model of Chiang (2007) popularized the usage of synchronous context-free grammars (SCFGs) for machine translation. SCFGs model translation as a process of isomorphic syntactic derivation in the source and target language. But the Hiero model is formally, not linguistically syntactic. Its derivation trees use only a single non-terminal la- bel X, carrying no linguistic information. Consider Rule 1. X → ⟨ maison ; house ⟩ (1) We can add syntactic information to the SCFG rules by parsing the parallel training data and pro- jecting parse tree labels onto the spans they yield and their translations. For example, if house was parsed as a noun, we could rewrite Rule 1 as N → ⟨ maison ; house ⟩ But we quickly run into trouble: how should we label a rule that translates pour l’´etablissement de into for the establishment of? There is no phrase structure constituent that corresponds to this English fragment. This raises a model design question: what label do we assign to spans that are natural trans- lations of each other, but have no natural labeling under a syntactic parse? One possibility would be to discard such translations from our model as im- plausible. However, such non-compositional trans- lations are important in translation (Fox, 2002), and they have been repeatedly shown to improve trans- lation performance (Koehn et al., 2003; DeNeefe et al., 2007). Syntax-Augmented Machine Translation (SAMT; Zollmann and Venugopal, 2006) solves this prob- lem with heuristics that create new labels from the phrase structure parse: it labels for the establish- ment of as IN+NP+IN to show that it is the con- catenation of a noun phrase with a preposition on either side. While descriptive, this label is unsatis- fying as a concise description of linguistic function, fitting uneasily alongside more natural labels in the phrase structure formalism. SAMT introduces many thousands of such labels, most of which are seen very few times. While these heuristics are effective (Zollmann et al., 2008), they inflate grammar size, hamper effective parameter estimation due to feature sparsity, and …

Content truncated to 500 words