Lately there has been some interest in generators that derive strings from feature structures within unification based formalisms (e.g. Shieber 1988 for PATR II, and also references cited there, Wedekind 1988 for an LFG generator, and Dymetman & Isabelle 1988 for generation with DCG's). In the first section I will show that the generators that have been proposed all face some problems. Some analyses in the spirit of Unification Categorial Grammar (Zeevat et al. 1987) and HPSG (Pollard & Sag 1987) that account for Dutch crossing dependencies cannot be handled by top-down generators. On the other hand, Shieber's bottom up generator for PATR II requires grammars to be semantically monotonic. I will assume that this requirement is too strong for general usage. As an alternative I will define BUG, a bottom up generator that uses a top-down oracle , as is common in directed bottom up parse strategies (for an overview cf. Kay 1980, see also the BUP parser in Matsumoto et al. 1983, and a similar parser in chapter 6.3 of Pereira & Shieber 1987). I require that all grammar rules have exactly one semantic head. Furthermore a restriction on semantic heads will be defined. In the third section I will compare this restriction with Shieber's requirement on semantic monotonicity. It will be shown that the requirement on semantic heads allows some useful analyses not available in Shieber's approach.
BUG is part of an experimental machine translation system
for translating international news items of Teletext. The
analysis/generation module of the system uses a Prolog
version of PATR II (similar to Hirsch 1987).
For this reason I will define BUG within this Prolog based
PATR II environment, but I assume that the approach is
valid for unification based formalisms in general.