Extended Tree Transducers in Natural Language Processing

Tree transducers are finite-state devices computing relations on trees. Their study was initiated by Thatcher (1970) and Rounds (1970), who established the classical top-down tree transducers that process the input tree from the root towards the leaves. Shortly afterwards, Baker (1973) introduced the bottom-up tree transducers that process the input tree from the leaves towards the root in analogy to the top-down and bottom-up tree automata (Thatcher, 1973). Due to applications in syntax-directed semantics (Fülöp and Vogler, 1998), tree transducers were extensively studied in the following years as detailed in (Gécseg and Steinby, 1984) and (Gécseg and Steinby, 1997). Notable extensions to the original top-down tree transducers include the top-down tree transducers with regular look-ahead by Engelfriet (1977), the attributed tree transducers of Fülöp (1981), and the macro tree transducers by Courcelle and Franchi-Zannettacci (1982) and Engelfriet and Vogler (1985). In statistical machine translation (Koehn, 2009), syntax-based models (Chiang, 2010) [i.e., models that translate from or to syntax trees] have recently seen a lot of progress. It was identified already by Eisner (2003) that the classical linear top-down and bottom-up tree transducers cannot properly handle phenomena (such as rotation) that occur during the translation between natural languages. This result was presented for synchronous contextfree grammars [SCFG] (Chiang, 2006), which is a formalism similar in spirit to (and essentially equally expressive as) the syntax-directed translation schemata by Aho and Ullman (1969), which were later refined to the more general bimorphism approach by Arnold and Dauchet (1982). Instead Eisner (2003) proposes synchronous tree