«Formal Phonology* David Odden OSU Abstract Two problematic trends have dominated modern phonological theorizing: over-reliance on machinery of ...»
One of the reasons why “independent evidence” has become important in evaluating phonological theories is that theorizing has tended to be an ad libitum enterprise where argumentation often reduces to saying “Here’s my position, show me the counterexample”. Extragrammatical evidence has become a necessary arbiter in theorizing, since there are few counterexamples (the inventory of theoretical concepts has become vast and there are many degrees of freedom in constructing analyses), and theoretical concepts tend to be only thinly attached to perceptible reality. Coupled with confusion over what the domain of phonology is (phonology is not the same as “language sound systems”, nor is it “what speakers know or can do w.r.t. sounds in language, or not even in language”), this has meant that important questions cannot be resolved except in a partisan manner that reflects arbitrary assumptions. Formal Phonology is a metatheoretical framework that guides the formation of theoretical concepts, thus limiting the potential for creating complex systems of unrelatable constructs that are in competition.
6. The construction of a hierarchy of phonological concepts The essential characteristic of the bottom-up approach of Formal Phonology is that theoretical claims are to be built on an existing solid foundation. This section lays out the methodological issues involved in answering one question, namely what is the proper account of feature-variable behavior, initiated in section 4. I phrase the question in terms of “feature-variable behavior”, not just feature-variables themselves, since the very question whether variables are needed is itself a matter to be addressed. I give the benefit of the doubt to the SPE theory of value variables, so “feature-variable behavior” will refer to “facts that would be formally treated using Greek letter variables”. Since, historically speaking, phonological research has not rigorously pursued the bottom-up formal approach, this section illustrates how construction, evaluation and repair of theories can take place “in the middle” of a conceptual hierarchy.
Certain concepts can be assumed as already established. For example, the claim that grammars contain rules that map representations onto representations is taken to be beyond reasonable doubt. As clarified in fn. 10, by “rule”, I do not necessarily mean an ordered, same-type representational mapping of derivational rule-based phonology, but in order to simplify the discussion, I only discuss how the conceptual analysis works in a derivational rule-based theory, leaving exploration of analogous constraint-based analysis as an exercise for the reader. I also take it to be established that representations include “features”, and rules are stated as operations on features.
Other unresolved questions impinge on theories of feature-variable behavior. Those issues and their logical relations must be identified, since they bear on how feature-variable facts should be modeled. The most important question is whether features are binary attribute-value pairs, or privative present / notpresent distinctions. A formal theory based on the assumption that features are binary attribute-value pairs could not be correct if they are actually privative present / not-present distinctions. On the other hand, it may be possible to articulate a theory requiring only minor changes in the structure of the theory to accomodate a determination that features are binary, or are privative (or a mix of the two): in fact, such a theory is articulated below.
6.1. Privativity In the development of theories of feature underspecification, it has been proposed that at least some features are single-valued, “privative” in the terminology of Trubetzkoy. An argument for the privativity of [voice] is advanced in Mester & Ito (1989), to resolve a contradiction between Radical versus Contrastive Underspecification theories. The two most-compelling arguments for privativity are transparency and I ignore pseudo-privativity, where each binary-theory feature metatheoretically maps to two privative-theory mutually-exclusive attributes such as “voiced” and “voiceless”, “oral” and “nasal”, and so on, which translates value-attribute pairs into two single lexical items. I also assume the standard principle of privative analysis that rules cannot refer to the lack of a specification. If this principle is abandoned, the resulting theory would be a notational variant of binary feature theory, where “F” maps to [+F] and “lacking F” maps to [-F].
asymmetry. The asymmetry argument is the claim that one value, usually [-F], is never referred to in rules, and segments that are [-F] fail to undergo rules that they might be expected to undergo (or, are maximally succeptible to being specifically targeted). Tendency towards asymmetrical behavior is not a valid basis for deriving a theoretical construct in FP, especially when the supposed asymmetry is only marginally valid from an empirical perspective – there are no hard-core asymmetries where e.g. labial is never the sole target for place of articulation assimilation (Hume & Tserdanelis 2002); both [-high] and [+high] do actually spread in vowel harmony; [-round] spreads in Hungarian vowel harmony, etc.
The transparency argument is more credible, since it depends on a strong formal claim, that no element can intervene (by definition, on the same tier) between Si and Sk which are rule-adjacent. Lyman’s Law in Japanese voices an initial obstruent, provided that no voiced obstruent follows within the stem.
The scan for a blocking voiced obstruent would be thwarted by a voiceless obstruent between the voiced obstruent on the right and the target if voiceless obstruents are [-voice]. Since voicing is contrastive in Japanese obstruents, they cannot be underspecified for voicing, according to Contrastive Underspecification (which Mester and Ito argue for). The combination of transparency evidence for nonspecification of [-voice] plus the principles of Contrastive Underspecification can only be resolved, argue Mester and Ito, if voice is universally assumed to be a privative feature. This at least renders the hypothesis of privativity worthy of further consideration, hence a matter relevant to the theory of feature-variable behavior.
It is uncontroversial that languages (for example Hungarian) have voicing assimilation rules whereby voiced obstruents become voiceless before voiceless obstruents, and voiceless obstruents become voiced before voiced obstruents, standardly stated in SPE notation as (5).
(5) [-son] → [αvoice] / ___ [-son,αvoice] This formulation presupposes that [t] is [-voice] and [d] is [+voice], which contradicts the presupposition that features are privative. If segments are unorganized sets of (unvalued) attributes, regressive voicing assimilation cannot be formalized. Because Hungarian exists, then either features are (at least) binaryvalued, or else segments are not unorganized sets of attributes. Since Clements (1985), there has been reasonable evidence for imposing some form of organization on features, which opens the possibility that the distinction [t] / [d] is represented as in (6).
(6) root root
= Laryngeal Laryngeal Underlying /gt/ would surface as [kt] by spreading the Laryngeal node of /t/. Since /t/ dominates nothing, when its Laryngeal node replaces that of /g/, both segments would be phonetically voiceless unaspirated, following a widely-assumed interpretive convention on representations. The same kind of formalization This option implies that all laryngeal features assimilate, not just voicing. Symmetrical assimilation of voicing but not glottalization or aspiration would challenge this rule, but might also spur the postulation of an additional intervening node between Laryngeal and voice. This point is discussed below.
[nasal] Depending on the actual existence of rules apparently spreading both values of all features, the correctness of Privative theory may depend on whether such intermediate nodes exist. As a case in point, Sagey (1986) postulates that a non-branching Soft Palate node intervenes between [nasal] and the Supralaryngeal node, thus Y in (8) is “Soft Palate”.
Lombardi (1991), who argues for the privative theory of voicing, actually rejects this kind of alternative for voicelessness, and presumes (p. 38) that “the proper representation [of voiceless obstruents] is one with no Laryngeal node”. We will refer to this theory as “constrained privativity”, since it limits grammatical theory to including a subset of the representations that would otherwise be possible. Such a representation is automatically possible given the general formal principles of theories of feature geometry, so to enforce this desideratum, a theoretical concept must be added to the theory. The motivation for the claim is that “(t)here is no phonological contrast between a representation with a bare Laryngeal node and no Laryngeal node at all. Therefore any theory should allow only one type of representation”. The simplest formal system allows a language to distinguish (9a) and (9b), but no such language had been identified. Lombardi’s claim is that there is a grammatical principle which renders (9a) impossible.
(9) a. root b. root Laryngeal [T] [t] As discussed in section 4, Formal Phonology rejects “economy of things” in theory-evaluation, embracing instead “economy of concepts”. FP has no methodological commitment to dealing with the claim that languages with such a contrast are lacking: the proper concern would be over the conceptual systems admitting or prohibiting representations, and a formal prohibition of such a difference is otiose in lieu of positive evidence for a prohibition. Indeed, if features are privative, such a distinction could be invaluable in harmonizing the model to the facts of language (symmetrical assimilation of features is not unknown), and one could reasonably posit representation (9a) for a voiceless consonant which causes devoicing.
Stronger evidence than the lack of compelling examples of such languages is required to justify adding a complicating principle to the theory.
Unfortunately, Lombardi does not identify a specific theoretical principle to yield this restriction.
An FP analysis of the question requires not just a phenomenological statement “this is how it should be done”, it requires a specific concept which has that result. Ordinary application of simple feature-structure building rules would allow both representations in (9), so to rule one of them out (and prohibit a contrast within a language), an additional theoretical concept must be added to force a choice of representationtypes. Being generous to the proposal, that concept would likely be a node-pruning convention to the effect that non-terminal node Ni not dominating a feature is prohibited, and is eliminated if derivationally created. Hence one complication necessitated to implement constrained privativity is the inclusion of an otherwise unnecessary pruning convention (a “concept of action”).
A three-way contrast in Turkish stops w.r.t. final voicing is analyzed in this way (Inkelas 1994).
In order to implement this convention, it also becomes necessary to distinguish typically terminal nodes i.e. “features” from pre-terminal nodes such as Laryngeal and Place. Under any theory of feature organization, a distinction can be computed from the principle that says “Laryngeal dominates [voice], [s.g.], [c.g]”, but apart from rules stating what Laryngeal etc. dominate (where the notion “preterminal” is implicit), a formal distinction between terminal and pre-terminal nodes has been unnecessary. In order to implement the desideratum of not having empty Laryngeal nodes, a permanent representational distinction between “nodes” and “features” must be added to the theory, and a principle prohibiting empty nodes must be added.
This distinction cannot be read off of unalloyed representations. In (6), the actually-terminal node of [d] is the feature [voice], and the terminal node of [t] is the Laryngeal node: only the latter is prohibited.
The distinction cannot be computed by referring to representational potential, namely nodes that may dominate material (subject to pruning) versus ones that may not. In feature-geometric work since Sagey (1986), the features [anterior] and [distributed] have been dominated by Coronal, but it is standardly assumed that Coronal can itself be a terminal node (when a language has a single coronal series, thus no specifications of [anterior] and [distributed]). Of course it is sometimes assumed that Coronal itself and possibly Place are unspecified when there are no [anterior], [distributed] distinctions e.g. Avery & Rice (1989), Paradis & Prunet (1989), but this assumption is contradicted by the wide-spread assumption that laryngeal consonants are the ones that are unspecified for place, e.g. Steriade (1987). The specific claim of Avery & Rice does not contradict the theory of placeless laryngeals, because they do not assume the convention against empty “nodes” which Lombardi’s account requires.
It would be possible to reject claims of possible coronal underspecification and the theory that [anterior] and [distributed] are dominated by Coronal in order to preserve the node-pruning convention at issue (the central idea necessary to express constrained privativity), but that convention must then shoulder a substantial extra burden of proof and re-conceptualization, since the idea of constrained privativity leads to the denial of a number of other claims that seem empirically plausible. Taken in isolation, the claim that languages do not make contrasts like those in (9) seems innocuous, but a logical analysis of the consequences of the claim reveal numerous countervailing considerations from domains not necessarily connected to voicing and, as FP insists, claims about the nature of phonological grammar cannot be considered in isolation from the totality of knowledge about phonology.