Desktop version

Home arrow Language & Literature

The Operational (IP) model

From a classical perspective, what was most wrong about the IA model was the attempt to associate meaning with units of form which, at least in the case of flectional patterns, could not be assigned meanings in isolation. This foundational issue was not revisited by the architects of the IA model until later. Instead, their initial responses explored modest revisions of assumptions about the disassembly of forms.[1] The decomposition of meanings raised parallel issues, but these were largely passed over. In his discussion of Harris’s criteria for grouping morphs into morphemes, Hockett (1947) declares:

The first of them, involving meaning, is obviously the most difficult to handle... The second and third requirements are purely distributional, and more easily subject to analysis and modification. (Hockett 1947:327f.)

The definition of “a morpheme unit” as “a group of one or more alternants which have the same meaning” (Harris 1942:171) is carried over in the notion of “recurrent partials with constant meaning” (Hockett 1947:322). Yet ‘sameness of meaning’ was only ever formulated operationally, via the ‘paired utterance test’ (Harris 1951:32). Hence the ‘units of meaning’ in an IA model depended on a purely operational notion of ‘sameness’ or ‘constancy’ of meaning.

Procedures of morphotactic analysis were articulated more explicitly. But this explicitness only served to highlight the biases built into the IA model. By reducing Bloomfield’s diverse ‘arrangements’ to a single property of linear order, the Post-Bloomfieldians had, whether by accident or design, arrived at a model that was optimized for the analysis of agglutinative patterns. Deviations from an agglutinative ideal could only be described indirectly, by constructing an abstract agglutinative arrangement that is subsequently converted to an observable surface form, as Hockett (1987) later acknowledges:

We were providing for alternations by devising an “agglutinative analog” of the language and formulating rules that would convert expressions in that analog into the shapes in which they are actually uttered. Of course, even such an agglutinative analog, with its accompanying conversion rules, could be interpreted merely as a descriptive device. But it was not in general taken that way; instead it was taken as a direct reflection of reality. (Hockett 1987: 83)

The highly abstract conception of grammatical structure implied by this practice (what Hockett (1987:184) termed ‘the transducer fallacy’) was again not reconsidered until much later. It was the problems that derived from the treatment of ‘units of form’ as “sequences of phonemes” (Harris 1942:170) that were of most immediate concern. By the early 1950s, Hockett had come to realize that the initial solutions, which coerced form variation into types of special morphs, merely subverted the nature of the original IA model:

Could we not modify our definition of ‘morph’ in such a way as to allow subtractives and replacives in those circumstances where they seem so clearly convenient? Of course we can do so. But such action seems to be equivalent—perhaps rather unexpectedly—to removing the keystone of the whole IA arch; the model begins to collapse... When we pick up the pieces and try to fit them together, again—without restoring the keystone—we find that we are no longer dealing with anything that looks like IA; we have a new model on our hands. (Hockett 1954:394)

Hockett terms the new model ‘item and process’ (IP), in recognition of its debt to the process-based view of Sapir (1921), which Bloomfield and his successors had regarded with suspicion. To clarify how this model differs from the IA model, it is useful to begin with the Bloomfieldian notion of a linguistic form as “a phonetic form which has a meaning” (Bloomfield 1933:138). A form in this sense has two components, a ‘meaning, which can be represented by a set B of grammatical and semantic properties, and a ‘phonetic form’ X. As a minimal linguistic form, a morpheme {m} is a pair (B, X), where B is a minimal property set and X is a single morph. In an IA model, non-minimal morphological units are built up from adjacent morphemes by combining the feature sets and concatenating the morphs of the morphemes. An IP model retains a set of basic stem or root morphemes but represents affixal morphemes by processes. Like morphemes, a process has two components, a grammatical function f that maps one property set B onto an augmented set B' and an operation o that maps a phonetic form X onto a modified form X'.[2] In an IP model, non-minimal morphological units are built up by applying a process P to a morphemic pair (B, X) to define a new ‘output’ pair (f (B), o(X)).

The derivation of hoofs and men in Figure 2.5 illustrates this process-based perspective. Since processes, like segmental morphemes, have constant meaning and variable form, the functionf adds the feature ‘plu’ to the property set of hoofs and men. However, whereas /hufz/ is formed by suffixation of /z/, /mm/ is defined by an ablauting operation that substitutes ? for ж in /msn/.

These analyses highlight the central insight that affixal and nonaffixal alternations can be treated more naturally as types of processes than as types of segments. In the case of ablaut alternations, an operation (whether stated in terms

IP analyses of affixal and ablauting plurals in English

Figure 2.5 IP analyses of affixal and ablauting plurals in English

of segment replacement or phonological feature change) can replace /ж/ by Izl in Imsnl. Including this operation among those available to the plural process avoids the artificiality of introducing a ‘replacive’ morph to trigger replacement. Similar analyses apply to ‘subtraction’ and other process morphs, all of which can be described directly as operations on items, rather than as special, operation- inducing items. From an IP perspective, affixal patterns reflect simple types of operations which tend to preserve their inputs. For example, the regular ending /z/ can be encapsulated in the operation ‘/X/ ^ /Xz/’. The morphotactic effect is the same in these cases, but the new model drops the pretense that ‘replacement’ or ‘subtraction’ are special kinds of item that must therefore somehow be brought into a linear arrangement with other items.

In other respects, the new model represents a modest revision of the IA model. Like IA approaches, IP models assume a lexicon consisting of an inventory of minimal elements, stems or roots, from which complex forms are derived. The derivation of complex forms involves adding grammatical properties to those associated with the root, in parallel with the modification of the root form.[3] Although no longer uniformly item-based, the resulting IP model remains firmly morpheme-based. Whereas the IA model associates all meaning with segmental morphemes, the IP model associates lexical meaning with root morphemes, and grammatical meaning with morphemic processes.

As in the IA model, the morphophonemic level plays a vital role mediating between morphemes and surface forms. Morphologically conditioned allomorphy is not eliminated, but relocated, expressed by variation in the operations associated with a process. The selectional relation between a noun and plural exponent is expressed in the IP model by the choice of the operation that either suffixes /z/, ablauts the stem vowel, or else effects some other change.

Hockett (1954) also suggests that the IA strategies for dealing with violations of total accountability remain available in an IP model, should the analyst wish to invoke them. Yet despite the close correspondence between the models, it is not entirely clear that these strategies are directly transferable:

Empty root-alternants [and] portmanteau root-alternants... are definable and allowable as in IA, should there be any need for them. Zero alternant roots and zero markers of processes are likewise allowable, under similar limitations. (Hockett 1954:396)

To begin with, the shift to a process-based perspective greatly weakens the internal motivation for zero morphs. There is no obvious need for zero exponents in models that can contain processes that apply no operation.[4] For example, the ‘zero marking’ of plural sheep can be described by a process that applies no operation to the noun stem (though a plural process is still required to add plural features to the underspecified properties of the stem).[5]

Empty morphs raise more fundamental difficulties. An ‘empty’ strategy that modifies forms while preserving features can be formulated as a process with a feature-preserving function and a form-altering operation. Yet the interaction of processes in an IP model is largely governed by the ‘feeding’ and ‘bleeding’ relations determined by the successive addition of features to an initially underspecified stem entry. Thus the interaction of feature-preserving processes must be regulated in some other way. The applicability of such processes can be restricted if the feature-preserving function is specified forfeaturesthatmustbepresent (orabsent) in the input, like the ‘constraining equations’ of LFG (Kaplan and Bresnan 1982) and repeated application is barred in some way. However, this extension amounts to introducing a class of realization rules into an IP model, since standard exponence rules preserve the features and (potentially) modify the form of their inputs. Hence the ‘realizational processes’ required to accommodate empty morphs within an IP model are no less artificial than process morphs are within IA models.

To a large extent, the IP model reassigns the complexity of an IA analysis from the morphotactic structure of individual forms onto the derivational structure defined by the successive application of processes to a root entry. This shift in the direction of a realization-based perspective is reflected in Hockett’s suggestion that the material present in a form need not be either morphemic or ‘empty’, but may serve merely as a marker of a process that has applied:

Some of the phonemic material in a derived form may be, not part of any underlying form, but rather a representation or marker of the process. (Hockett 1954:396)

The recognition of non-morphemic ‘markers’ again subverts the conception of morphological analysis as the association of minimal ‘units of meaning’ with minimal ‘units of form’. In some cases a process may add (or modify) a single feature, and/or introduce a single marker. But there is no principled grounds within this scheme of analysis for regarding a biunique relation between features and exponents as normative or canonical.[6]

In sum, the introduction of processes eliminates some of the artifactual problems created by the IA model, by allowing a description to incorporate non-segmental phonetic alternations that have been co-opted to express morphological contrasts. In particular, by interposing processes between properties and forms, the IP model avoids the artificiality of treating process morphs as ‘change-inducing’ items. The resulting model is, however, less uniform than a pure IA model. Items no longer occupy the central role that they have within a system based on procedures of segmentation and classification, but they are not eliminated altogether in favour of processes. Instead, roots constitute a residual class of morphemic items in the new model, even as exponents are recast as ‘markers’ of processes. In some cases, the appeal to processes shifts the problems posed by feature-form mismatches from the morphotactic to the derivational structure of a form. Process-based analyses simplify the morphotactic structure of elements like plural sheep, or singular nouns in English in general, since they avoid the need to introduce zeros bearing number features. However, the derivational structure of sheep is no simpler, since a form-preserving process must still apply to derive the plural, and a separate process must apply to derive the singular. A model that constructs inflected forms ‘incrementally’ from stems—whether by concatenating morphs or by applying processes—has no straightforward way to declare that singular number is realized by the basic noun stem in English. The fundamental obstacle to expressing this kind of generalization is not the role of items in the IA model but the role of morphemes; the problem is intrinsic to morphemic analysis. The challenges posed by cumulative and overlapping exponence derive similarly from the nature of the relation between features and forms, not from assumptions about the representation of forms or form variation.

  • [1] As shown later by Roark and Sproat (2007: §3), IA and IP models are in fact computationallyequivalent and can both be implemented by finite state transducers.
  • [2] The separation between processes and operations anticipates a similar distinction between rulesand operations in models of Montague Grammar (Bach 1979, 1980).
  • [3] In the terms of Stump (2001), both IA and IP are ‘incremental’ models that increment theproperties of underspecified roots in the course of constructing words.
  • [4] Equivalently, processes can apply inert ‘identity’ operations which preserve the input form,thereby achieving the same effect as the ‘identity function default’ of Paradigm Function Morphology(Stump 20or).
  • [5] Requiring that every process must contain at least some operation captures the condition that “nomorpheme has zero as its only alternant” (Bloch Г947:402).
  • [6] The notion that processes must represent a simple property by a single exponent thus has no placein later approaches that adopt an IP perspective, such as Articulated Morphology (Steele Г995) or earlymodels of HPSG (Pollard and Sag Г987).
 
Source
< Prev   CONTENTS   Source   Next >

Related topics