«Formal Phonology* David Odden OSU Abstract Two problematic trends have dominated modern phonological theorizing: over-reliance on machinery of ...»
To summarize the issue of privativity, the first question is whether privative feature theory is possible, that is, is logically consistent with what it known. Assimilation of [-voice], [-hi], [-nasal] etc. appears to show that universal privativity is actually not possible. An analytical tack to harmonizing such rules with privativity is to assume intermediate nodes as in (6) and (8), so resolution of the privativity question must defer to resolution of the intermediate-node question. If empty intermediate nodes are allowed, a privative analysis is possible. If they are not, then some principle must prevent them (and that principle must be justified). The best option for such a principle seems to be a ban on formally-terminal organizing nodes, which requires adding a further representational distinction between “organizing node” (e.g. Laryngeal, Place) and “feature” (e.g. Coronal, Nasal) which is orthogonal to the formal terminal / nonterminal property. Adding such a distinction then requires justification. Saying “doing so allows a principle to be added to ban terminal Laryngeal without banning terminal Coronal” is not proper justification – it is a circular appeal to the presumed correctness of the very principle whose possibility is in question.
FP demands not just that the claim be shown to be possible, it must be shown to be necessary.
The reason why it matters whether features are privative or binary is that the notion of a “variable ranging over values” is meaningless with privative representations. If Privative Theory is correct, neither Value-Variable nor Identical-Value theory can be right, since the notion “value” which these theories depend on is meaningless in Privative Theory. However, a concept which is similar to Identical-Value Justifying a specific principle that appeals to “representational potential” would also require prior demonstration of the claim that rules or meta-grammatical principles can refer to “representational potential”.
theory, call it Identical-Presence theory, is definable on privative representations. Rather than comparing the plus-or-minus values of a feature Fk in segments Si, Sn, one could compare the existence of Fk in Si, Sn, so that the condition “[=Fk]... [=Fk]” is satisfied iff Fk exists in both Si, Sn or is lacking in both Si, Sn.
A third theory, call it Identical-Presence-or-Value theory is definable on mixed privative and binary representations, where “[=Fk]... [=Fk]” is satisfied iff Fk exists in both Si, Sn or is lacking in both Si, Sn when Fk is a privative feature; and the condition is satisfied for binary-valued Fk iff Fk has the value “-” in both Si, Sn or the value “+” in both Si, Sn.
What these theories have in common is the most general concept, call it Feature-Identity theory, which says that comparisons are made between instances of the same feature, and both theories state that rules can require segments to be “the same” with respect to a feature, abstracting away from the exact nature of the feature qua entity or attribute-value pair. In short, the correct concept would not be Identical-Value, Identical-Presence or Identical-Presence-or-Value theory, it would be that simplest justified concept, Feature-Identity theory, which is that rules refer to the sameness of features, not saying whether features are privative or binary. By the logic of FP, the concept of variable would not be so specific as to be framed in terms of values, presence or a combination therein – it would abstract away from that orthogonal question. Privativity is a separate question.
At this point, it should be clear that the correctness of Privativity has no impact on the correctness of Feature-Identity theory, but it and the facts of phonological systems are logically incompatible with Value-Variable theory, on the assumption that no features are immune to symmetrical assimilation.
6.2. Representations or variables?
A partially independent question bearing on variable behavior is whether the effect can also be explained on the basis of representations. Under certain assumptions, feature-value references might be accounted for representationally, without variables. This refers in part to the potential relevance of the OCP in rule operation, and in part to the theory of assimilation as spreading.
The potential for capturing identical-feature reference via autosegmental representations derives from the fact that a surface same-valued sequence potentially (but not obligatorily) has a unique characteristic. Non-autosegmental representations allow one representation of a different-valued sequence and one similar representation of a same-valued sequence in (10a). Autosegmental representations allow one representation of a different-valued sequence and two representations of a same-valued sequence in (10b), one of them being quite different in nature.
(10) a. [X, +F] [X, -F] Segmental non-identical sequence [X, +F] [X, +F] Segmental identical sequence b. [X] [X] Autosegmental non-identical sequence
[+F] If an identical-value sequence is represented as one specification with a multi-segmental domain, the fact of value-identity could be referred to indirectly, via the fact that such sequences are single features with multiple linkages. The literature on the OCP makes it clear, though, that there is no guarantee that a multilinked representation prevails over the identical-value sequence representation. The fact that a rule needs
to identify adjacent identical feature values is often simply taken to be sufficient proof that an identicalvalue sequence representation must be compressed into a multi-linked representation.
The standard autosegmental theory of rules of the form X → [αF] / __ [αF] is that they are spreading rules of the form:
(11) X X
F The autosegmental convention of not mentioning a feature value and inferring “whatever value” is nothing but a covert variable, saying exactly what Feature-Identity theory says (though not Value-Variable theory, where rules can have distinctive value variables).
As observed in Odden (1988) and Reiss (2003), there is more to the facts of identity reference than assimilatory spreading plus reference to multiple-linkage. A wide range of references to “value for Fi” exists, for example syncope in Syrian Arabic is blocked when the surrounding consonants have (certain) same feature values, a phenomenon termed “antigemination” by McCarthy (1986). Likewise, there are “Anti-antigemination” rules such as Koya syncope, where vowels are deleted just in case they are flanked by homorganic consonants, which apply only when doing so creates a violation of the OCP. The proffered explanation for Antigemination is that the OCP prevents the rule from applying, when doing so would create a violation of the OCP. In the case of Anti-antigemination rules, output-violation of the OCP would be the crucial triggering factor. There are rules of Geminate Epenthesis, where a vowel is inserted between homorganic consonants, such as in Lithuanian. Such sequences cannot be identified as “homorganic” via the representational property of multiple linkage, since multiple linkage would prevent epenthesis in the first place – see Hayes (1985), Schein & Steriade (1986). The representational theory must apparently admit two formal mechanisms to derive identity-references: the fact of multiple linkage, and the fact that the OCP is violated, which is not a strictly representational property. It is clear that a purely representational account of identity references is untenable, since there is no representational property that signals “identical sequence” in Antigemination, Anti-Antigemination and Geminate-Epenthesis effects.
Since a pure representational account of identity is untenable, the question is whether a pure OCPbased account might work (if not, then a theory of feature variables would be necessitated). Numerous issues underlie an OCP-based analysis, such as what exactly the OCP says. In the formulation of McCarthy (1986), “at the melodic level, adjacent identical elements are prohibited”, but there is no straightforward interpretation of “element” and “identical” especially degree of identity required in this context, as discussed in Odden (1988). There is no general answer to the question of when this principle holds (it is clearly not universal for all languages, rules and representational objects). Yip (1988) pursues the idea
that all sorts of phonological effects can be attributed to the OCP, and in light of the current view that many putative universals are “violable” in a language-specific way, it would be reasonable to consider a unified OCP-based account, one which does not depend (solely, therefore at all) on a multi-linked representation. Simply saying that identity references can be referred to the OCP does not eliminate the need for some grammatical mechanism to express identify references: a complex mechanism attributed to UG is still a complex mechanism, and attributing a mechanism to UG does not immunize the proposal from the need for critical scrutiny.
An OCP-based account must be parametrized in terms of the representational units that it holds of (tone, voicing, place of articulation etc.), and in that respect there is no meaningful difference between a rule which mentions the feature properties of a feature or node (two features must not be the same), versus specifying a feature or node as the argument of the OCP and applying a rule on the basis of an OCP evaluation. There is a real difference between presuming a theory of blocking and triggering, along the lines of Yip (1988), versus a theory such as that of Reiss (2003) which employs explicit identity and nonidentity conditions on standard rule formulations, but the issue at stake there is what the proper method is for formalizing “triggering” and “blocking” conditions, which is orthogonal to the question of how identity references are made in grammars. OCP conditions could be imposed on inputs or outputs, and could serve as blocking or triggering conditions, and we would derive the range of known identity effect. The Antigemination effect derives from blocking outputs that violate the OCP; the anti-Antigemination effect (Koya) where syncope only applies between homorganic consonants is where creating an OCP violation in the output is what triggers the rule; homorganic epenthesis and deletion derive from input triggering by the OCP (the rule only applies if the input violates the OCP); homorganic integrity, i.e. the failure of epenthesis in homorganic clusters is the case of input OCP blockage.
The relevant question is whether “[=Fi]... [=Fi]” and “OCP(Fi)” are anything other than notational variants. Since the condition “OCP” is false (violated) on Fi iff [αFi] is identical to [βFi], and the requirement “[=Fi]... [=Fi]” is true (satisfied) iff [αFi] is identical to [βFi], then “OCP(Fi)” is nothing more than the opposite of [=Fi]. Therefore, a choice between Feature-Identity variables and an OCP condition on a rule is a false dichotomy – they are the same thing. Needless to say, Feature-Identity/OCP theory is different from Value-Variable theory, since in the latter theory a condition may also compare the plus or minus values of different features, which cannot be done in a theory where the central concept is “sameness w.r.t. a given feature”.
6.3. The independence of variables Analysis of alternative theories has led to the conclusion that there is one real issue at stake, namely whether variable behavior always involves comparison on a feature-to-feature basis, as claimed by Feature-Identity theory, or are there also rules of the type [αFm] → X / ___ [αFk] as claimed by ValueVariable theory. This is in part an empirical question, but a precursor to inspecting the facts is scrutinizing the logic behind establishing the existence of such a rule.
Chomsky & Halle (1968) raise the question whether variables should be formally restricted,30 concluding on the basis of Viennese German that rules matching same value of different features do exist.
They point to the fact that in Viennese German, the words in Standard German vier [fiɐ] ‘four’ and für [fyɐ] ‘for’ are both pronounced [fi:r], and the words in Standard German viele [fi:ləM] ‘many’ and fühle [fy:ləM] are both pronounced [fy:ləM]. They postulate the following rule to account for these facts.
This is not to imply that look-ahead output conditions are actually required to express the full range of identity conditions. Rather, the point is to use conceptual vocabulary that is familiar from discussions of the OCP, to make it clear that Identical-Value theory and OCP theory say the same thing, as long as the scope of the OCP is restricted in a manner befitting the facts.
FP would say “emphatically not”, and would focus on the fact that a better theory of variables does not require any restriction.