Handling Linear Precedence Constraints by Unification

Linear precedence (LP) rules are widely used for stating word order principles. They have been adopted as constraints by HPSG but no encoding in the formalism has been provided. Since they only order siblings, they are not quite adequate, at least not for German. We propose a notion of LP constraints that applies to linguistically motivated branching domains such as head domains. We show a type-based encoding in an HPSG-style formalism that supports processing. The encoding can be achieved by a compilation step. INTRODUCTION Most contemporary grammar models employed in computational linguistics separate statements about dominance from those that determine linear precedence. The approaches for encoding linear precedence (LP) statements differ along several dimensions. Depending on the underlying grammatical theory, different criteria are employed in formulating ordering statements. Ordering constraints may be expressed by referring to the category, grammatical function, discourse rôle, and many other syntactic, semantic, morphological or phonological features. Depending on the grammar formalism, different languages are used for stating the constraints on permissible linearizations. LP rules, first proposed by Gazdar and Pullum (1982) for GPSG, are used, in different guises, by several contemporary grammar formalisms. In Functional Unification Grammar (Kay 1985) and implemented versions of Lexical Functional Grammar, pattern languages with the power of regular expressions have been utilized. Depending on the grammar model, LP statements apply within different ordering domains. In most Research for this paper was mainly carried out in the project LILOG supported by IBM Germany. Some of the research was performed in the project DISCO which is funded by the German Federal Ministry for Research and Technology under Grant-No.: ITW 9002. We wish to thank our colleagues in Saarbrücken, three anonymous referees and especially Mark Hepple for their valuable comments and suggestions. frameworks, such as GPSG and HPSG, the ordering domains are local trees. Initial trees constitute the ordering domain in ID/LP TAGS (Joshi 1987). In current LFG (Kaplan & Zaenen 1988), functional precedence rules apply to functional domains. Reape (1989) constructs word order domains by means of a special union operation on embedded tree domains. It remains an open question which choices along these dimensions will turn out to be most adequate for the description of word order in natural language. In this paper we do not attempt to resolve the linguistic issue of the most adequate universal treatment of word order. However we will present a method for integrating word order constraints in a typed feature unification formalism without adding new formal devices. Although some proposals for the interaction between feature unification and LP constraints have been published (e.g. Seiffert 1991), no encoding has yet been shown that integrates LP constraints in the linguistic type system of a typed feature unification formalism. Linguistic processing with a head-driven phrase structure grammar (HPSG) containing LP constraints has not yet been described in the literature. Since no implemented NL system has been demonstrated so far that handles partially free word order of German and many other languages in a satisfactory way, we have made an attempt to utilize the formal apparatus of HPSG for a new approach to processing with LP constraints. However, our method is not bound to the formalism of HPSG. In this paper we will demonstrate how LP constraints can be incorporated into the linguistic type system of HPSG through the use of parametrized types. Neither additional operations nor any special provisions for linear precedence in the processing algorithm are required. LP constraints are applied through regular unification whenever the head combines with a complement or adjunct. Although we use certain LP-relevant features in our examples, our aproach does not hinge on the selection of specific linguistic criteria for constraining linear order. Since there is no conclusive evidence to the contrary, we assume the simplest constraint language for formulating LP statements, i.e., binary LP constraints. For computational purposes such constraints are compiled into the type definitions for grammatical categories. With respect to the ordering domain, our LP constraints differ from the LP constraints commonly assumed in HPSG (Pollard & Sag 1987) in that they apply to nonsibling constituents in head domains. While LP constraints control the order of nodes that are not siblings, information is accumulated in trees in such a way that it is always possible to detect a violation of an LP constraint locally by checking sibling nodes. This modification is necessary for the proper treatment of German word order. It is also needed by all grammar models that are on the one hand confined to binary branching structures such as nearly all versions of categorial grammar but that would, on the other hand, benefit from a notion of LP constraints. Our approach has been tested with small sets of LP constraints. The grammar was written and run in STUF, the typed unification formalism used in the project LILOG. LINGUISTIC MOTIVATION This section presents the linguistic motivation for our approach. LP statements in GPSG (Gazdar et al. 1985) constrain the possibility of linearizing immediate dominance (ID) rules. By taking the right-hand sides of ID rules as their domain, they allow only the ordering of sibling constituents. Consequently, grammars must be designed in such a way that all constituents which are to be ordered by LP constraints must be dominated by one node in the tree, so that "flat" phrase structures result, as illustrated in figure 1. Vmax V0 sollte should Vmax NP[nom] der Kurier the courier ADV nachher later NP[dat] einem Spion a spy NP[acc] den Brief the letter V0 zustecken slip The courier was later supposed to slip a spy the letter. Figure 1 Uszkoreit (1986) argues that such flat structures are not well suited for the description of languages such as German and Dutch. The main reason1 is socalled complex fronting, i.e., the fronting of a non-finite verb together with some of its complements and adjuncts as it is shown in (1). Since it is a well established fact that only one constituent can be 1Further reasons are discussed in Uszkoreit (1991b). fronted, the flat structure can account for the German examples in (1), but not for the ones in (2). (1) sollte der Kurier nachher einem Spion den Brief zustecken zustecken sollte der Kurier nachher einem Spion den Brief den Brief sollte der Kurier nachher einem Spion zustecken einem Spion sollte der Kurier nachher den Brief zustecken nachher sollte der Kurier einem Spion den Brief zustecken der Kurier sollte nachher einem Spion den Brief zustecken (2) den Brief zustecken sollte der Kurier nachher einem Spion einem Spion den Brief zustecken sollte der Kurier nachher nachher einem Spion den Brief zustecken sollte der Kurier In the hierarchical tree structure in figure 2, the boxed constituents can be fronted, accounting for the examples in (1) and (2).