Towards Efficient HPSG Generation for German, a Non-Configurational Language

In this paper, we propose a rule-based method to improve efficiency in bottom-up chart generation with GG, an open-source reversible large-scale HPSG for German. Following an indepth analysis of efficiency problems in the baseline system, we show that costly combinatorial explosion in brute force bottom-up search can be largely avoided using information already contained implicitly in the input semantics: either (i) information is globally present, but needs to be made locally available to a particular elementary predication, or (ii) semantic configurations in the input have a clear translation to syntactic constraints, provided some knowledge of the grammar. We propose several performance features targeting inflection and extraction, as well as more language-specific features, relating to verb movement and discontinuous complex predicates. In a series of experiments on three different test suites we show that 7 out of 8 features are consistently effective in reducing generation times, both in isolation and in combination. Combining all efficiency measures, we observe a speedup factor of 4.5 for our less complex test suites, increasing to almost 28 for the more complex one: the fact that performance benefits drastically increase with input length suggests that our method scales up well in the sense that it effectively heads off the problem with exponential growth. The present approach of using a generator-internal transfer grammar has the added advantage that it locates performance-related issues close to the grammar, thereby keeping the external semantic interface as general as possible. TITLE AND ABSTRACT IN GERMAN Effiziente HPSG-Generierung fur das Deutsche Wir stellen eine regelbasierte Methode vor, zur automatischen Anreicherung der semantischen Eingabe einer reversiblen HPSG des Deutschen, die es erlaubt, teure uninformierte Suche bei der Bottom-Up-Chart-Generierung weitgehend zu vermeiden, indem (i) globale Information, die implizit in der Eingabe vorhanden ist, explizit und lokal verfugbar gemacht wird, und (ii) syntaktische Constraints aus semantischen Konfigurationen abgeleitet werden. Wir schlagen Performanzfeatures fur verschiedene Phanomene vor, wie Flexion, Extraktion, Verbbewegung und diskontuierliche komplexe Pradikate. Unsere Experimente zeigen erhebliche Effizienzsteigerungen (Faktor 4.5–Faktor 27.8), deren Zunahme mit steigender Eingabekomplexitat korreliert, was die gute Skalierbarkeit unserer Methode belegt. Der generator-interne Transferansatz zeichnet sich weiterhin dadurch aus, das Performanzaspekte grammatik-nah behandelt werden, wodurch die externe Semantikschnittstelle so allgemein wie moglich bleibt.