Variable Morphotactics in Information-based Morphology1

out shared properties both for morphotactics and shape, one would still have to specify all combinations manually. Fortunately, the concept of online type construction (Koenig & Jurafsky 1994) provides the desirable solution, since it turns otherwise static type hierarchies into a generative device, essentially making them dynamic. In order to address similar problems within the lexicon (i.e. systematic alternation traditionally captured by means of lexical rules), Koenig & Jurafsky (1994) revise the status of lexical inheritance type hierarchies: instead of describing lexeme categories directly, these categories are obtained by means of a closure operation on a type underspecified hierarchical lexicon, partitioned into orthogonal dimensions. According to their definition, a well-formed lexical category is obtained by systematic intersection (conjunction) of leaf types. Assuming that leaf types within one dimension are disjoint while dimensions are conjunctive, the full set of well-formed categories is obtained by combining, under unification, each leaf type from one dimension with each leaf type from all other dimensions. Their approach can be straightforwardly applied to the problem at hand: instead of regarding (leaf) types in our rule type hierarchy as well-formed rules by themselves, we shall assign them the status of underspecified rule schemata or partial description of rules. The set of well-formed realisation rules is then obtained by pairwise combination of leaf rule types from the dimensions of MORPHOTACTICS and EXPONENCE. As captured in Figure 3, rule types in the EXPONENCE dimension pair morphosyntactic properties to be expressed with exponents, or, more precisely, description of their phonological shape. While some of these rule types specify grammatical function (either subj or obj), the majority of rule types actually has this piece of information underspecified (cf. the paradigms in Table 3). The two types in theMORPHOTACTICS dimension, by contrast, abstract out the systematic relation between position class and grammatical function. As indicated by the dashed lines in Figure 23, pairwise combination of leaves in the EXPONENCE dimension with those in the MORPHOTACTICS dimension yields the set of fully expanded realisation rules. If an EXPONENCE rule type is already constrained as to grammatical function (e.g. third singular m-wa class a or m), it will only combine with the appropriate type in the MORPHOTACTICS dimension, the other combination being ruled out by unification failure, effectively fixing position class information. For the majority of Swahili subject and object markers, where grammatical function is underspecified for EXPONENCE rule types, pairwise combination with MORPHOTACTICS rule types will yield two maximally specific types: one that constrains the exponent to position class 2 and disambiguates its function to subj, and another that constrains it to position class 5, disambiguating grammatical function to obj. In essence, intersection of EXPONENCE types with either type from the MORPHOTACTICS dimension will model positional disambiguation, whereas underspecification within the EXPONENCE hierarchy will capture parallelism of exponence.