The Morphology-Syntax Interface in LFG

Realizational or Paradigm Function Morphology (PAF) has been argued to provide an ideal theory of morphology for lexicalist theories of syntax (Stump 2001). Spencer and Sadler (2001) take up this argument by critiquing current notions of the morphology-syntax interface within LFG. They see a major weakness within standard approaches (e.g., Bresnan 2001, Börjars, Vincent and Chapman 1996) in that LFG essentially operates with morpheme-based notions. Morphemes are represented as sublexical bits of the grammar and can be associated with lexical entries. In contrast, PAF rejects the idea that morphemes are form-content pairs. Rather, inflectional morphology like the English -s is seen as an exponent of abstract morphological feature sets. One remedy proposed by Spencer and Sadler (2001), loosely based on Frank and Zaenen (2000), is to posit a distinction between morphological features such as m-tense (encoded at m-structure) and syntactic features such as s-tense (encoded at f-structure).

We agree that a more complex view of the morphology-syntax interface is needed, but prefer an architectural design which places more of the burden on a separate morphological component, thus avoiding the projection of morphological features from f-structure (m-structure is projected from f-structure). In this paper we formalize more precisely the syntax/morphology interface of the LFG architecture in terms of a complex relation M, so that we can clearly distinguish the role of syntactic properties and operations from those that properly belong to morphology. In particular, the relation M must be understood as the composition of two further relations, D and R. The D relation maps sets of feature pairs such as NUM:SG or TENSE:PRES into sets of feature descriptions (and vice versa). Feature descriptions are just the kind of f-descriptions we are familiar with within standard LFG, e.g., (^ NUM) = sg or (^ SUBJ NUM) = sg (depending on the particular syntactic analysis). The R relation can be seen as the realization mapping of traditional morphology.It contains triples of the form < word, category, set-of-dnames>, indicating that the given collection of d-names in a particular category is realized as the given word.

In the paper, we explore the internal structure of the morphological component by working through examples from English, German, Latin and Sanskrit. We show that under this archictecture, the morphology/syntax interface in LFG is formed by the mapping from sets of d-names (such as NUM:SG) to sets of f-descriptions, which can be dealt with by the syntactic component in the usual way. The mapping does not necessarily need to be one-to-one (and in fact most of the time is not). We hope that this result will to serve focus and clarify some of the recent issues raised with respect to the nature of morphology/syntax in LFG.

Our discussion furthermore highlights a mathematically simple approach to morphological realization, the regular relations described by Karttunen et al. (1992), Beesley and Karttunen (2003), and others. Regular (or finite-state) morphology is also attractive from computational and perhaps even psycholinguistic points of view since, unlike Paradigm Function Morphology, it is naturally and symmetrically interpretable in processing models for both generation and recognition of word-forms.