Synchronous TFG for speech translation

The use of synchronous TAG for Machine Translation has been described by Abeille et al. [1990] and has resulted in several implementations [Prigent, 1994] [Egedi et al., 1994], mainly developed using the XTAG system [Paroubek et al., 1992]. While we subscribe to the general arguments in favour of the use of TAG+ for Machine Translation, it appears that speech translation could constitute an ideal application of these ideas [Harbusch & Poller, 1994]. lt is actually easier to select specific areas where speech translation is both feasible and of practical impact (see e.g. the CST AR, VERBMOBIL and SRI Speech Translation projects). In this paper, we report the implementation of a minimal speech translation prototype based on synchronous TAG+ (more exactly, synchronous Tree Furcating Granunars or STFG), which has been developed as a direct extension of our TFG parser [Cavazza, 1998]. Sheiber & Schabes [1990] originally coined the tenn "synchronous TAG". They described synchronous derivation of semantic structures from tree operations carried on lexicalised trees. A synchronous TAG is thus a pair of two elementary trees, one representing the source language and the other a logical formula, which is also represented as a variant of TAG (and is lexicalised as weil). However, the term of synchronous TAG, when used for machine translation, actually subsumes different approaches and deserves some clarification. The initial presentation of TAG for Machine Translation by Abeille et al. (1990] referred to synchronous TAG, though in fact it directly mapped lexicalised trees to one another, without making recourse to the "semantic" trees described by Shieber & Schabes [1990]. In that sense, it could be considered as a transfer formalism or a structural correspondence system [Kaplan et al„ 1989]. Further implementations by Prigent [1994] within the XTAG system [Paroubek et al„ 1992] have been based on an extended transfer paradigm, mapping between derivation trees in the source and target languages, thus introducing an intermediate representation. On the other hand, direct mapping between lexicalised trees has also been adopted in the STAG project [Egedi et al„ 1994] [Egedi & Palmer, 1994]. We would like to suggest, adopting a terminology from Prigent [1994], that approaches based on the direct mapping between lexicalised trees should be renamed iso-synchronous. This would clearly indicate that the synchronous trees on which adjunction (resp. substitution) operations are carried are of the same kind. Our own implementation follows the isosynchronous approach, and is based on paired elementary trees in the TFG formalism [Cavazza, 1998].