Recurrence and redundancy
From a classical WP perspective, the entire Bloomfieldian notion of ‘compactness’ rests on a conflation of recurrent and redundant structure. A classical description of an inflectional system will exhibit a network of recurrent patterns. The same stem or stems will tend to recur in the inflected forms that make up the paradigm of an item. The distribution of exponents will also exhibit systematic patterns across the paradigms of distinct items. Yet these recurrent patterns are only redundant if they are predictable from their components or from other aspects of the system. An atomistic description that disassembles an inflectional system into inventories of basic stems and exponents will eliminate any recurrent morphological structure. But it will not reduce redundancy unless the decomposition satisfies each of the preconditions in (4.1).
(4.1) Preconditions for lossless decomposition
a. The parts identified as recurrent must be genuinely identical.
b. Properties of the wholes can be reconstituted from their parts.
c. The original wholes can be recovered from their simple parts.
Neither of the first two preconditions is met by any decompositional model. As discussed in Chapter 8.3, the Bloomfieldian model of “scientific compactness” operates with imprecise notions of orthographic and phonemic ‘identity' that obscure sub-phonemic contrasts that are systematically produced and discriminated by speakers. The distributional properties of whole forms are also not determinable from the distribution of their parts. In fact, the relationship between these distributions is grammatically significant, exerting an influence on everything from the processing strategies applied to the wholes to the productivity of their parts (Hay 2001; Hay and Baayen 2002, 2006).
It is, however, the third precondition in (4.1) that is most directly relevant to the status of paradigms. The reassembly of forms from their minimal parts is exactly what is not in principle possible in an inflection class system. The need to assign items to inflection classes only arises when the co-occurrence of base and exponent is not determined by substantive properties of the whole or those of its parts. An inflection class system is thus defined by distributional patterns over bases and exponents that cannot be reduced to independent selectional properties. The qualification ‘independent’ is of course critical here, since any unit can be represented as ‘parts’ that explicitly encode their affiliation with a larger unit. To a large extent, this strategy merely provides a distributed representation of the original unit. In contemporary morphemic and realizational approaches, these types of distributed representations tend to be expressed in terms of ‘indexical’ features of various kinds. The proliferation of diacritic features has a subtly subversive effect on morphological analysis. Rather than exhibiting the structure and organization of a morphological system, analyses formalize a process of reassembling units from parts that were explicitly encoded to mark the forms from which they were initially obtained. Even when this procedure is successfully implemented, it is unclear what it achieves.