Stephen R. Anderson
Yale University

[Prepared for Cambridge Encyclopedia
of Language and Linguistics

As opposed to phonetics, which deals with the properties of sounds from a language-independent point of view, phonology constitutes the study of the sound structure of units (morphemes, words, phrases, utterances) within individual languages.  Its goal is to elucidate the system of distinctions in sound which differentiate such units within a particular language, and the range of realizations of a given unit’s sound structure as a function of the shape of other units in its context.  These two goals – the study of invariants of sound structure and of the variation shown by these elements in combination – are obviously closely related, but attention has tended to shift between them over time.
Late nineteenth and early twentieth century study of sound structure focused on the details of sound production.  As these studies (in both articulatory and acoustic phonetics) became more sophisticated, however, it was increasingly apparent that the resulting explosion of data about sound properties was obscuring, rather than enhancing scholars’ understanding of the way sound is organized for linguistic purposes.  Much that is measurable in the speech signal is predictable, internal to the system of a given language, even though exactly comparable properties may serve to distinguish items from one another in a different language.  
Vowels in English, for example, are relatively longer before certain consonants than before others, but the difference in the vowels of e.g. cod and cot is entirely predictable from this principle alone. By contrast, an exactly parallel difference between the vowels of kaade ‘dip’ and kade ‘envious’ in Finnish serves as the sole difference between these words.  A focus on phonetic features alone fails to reveal the role played by sound properties within a language.
The result of this insight was the development within various theories of structuralism of attempts (Anderson, 1985) to define the phoneme, a presumed minimal unit of contrast within the sound system of a single language.  While there is considerable diversity among these views, it is fair to say that by and large, they focused on the elucidation of the contrastive properties of elements of surface phonetic form to the exclusion of other aspects of sound structure.
The Development of Modern Phonology
Post-structuralist theories fall broadly within the tradition of Generative Phonology, associated in its origins with (Chomsky & Halle, 1968).  The distinguishing character of this view was its attention not simply to surface contrasts but also to patterns of alternation in shape, and its positing of an abstract underlying representation (where contrasts among elements are characterized) that is related to surface phonetic form by a system of rewriting rules.  Each of these rules represents a single generalization about the realization of phonological elements (e.g., “Vowels are long before voiced obstruents”).  Much of the theoretical discussion in the 1960s and early 1970s concerned the role of an explicit formalism for these rules.
The rules were presumed to apply in a sequence, with each applying to the result of all previous rules.  As a consequence, some of the generalizations represented by individual rules may only be valid at an abstract level and not true of all surface forms to the extent subsequent changes obscure the conditioning factors of a rule or its effects, leading to the opacity of the rule in question.  For example, in many varieties of American English the medial consonants of words like ladder and latter are both pronounced as the same voiced flap [D].  The vowels of the initial syllables of such words continue to differ in length, however, reflecting the abstract difference in voicing between /d/ and /t/, even though that difference is obscured by the (subsequent) application of a rule of flapping which renders the vowel length rule opaque.  Much attention was paid in this period to the theories of rule ordering necessary to describe such phenomena.
In the years immediately following the publication of Chomsky & Halle 1968, a number of scholars reacted strongly to the perceived abstractness of the underlying phonological representations to which it appeared to lead.  Various proposals intended to restrain this aspect of the theory appeared, some of them based on the idea that if the rules themselves could be constrained so as to permit only highly ‘natural’ ones, drawn from some substantively constrained universal set, the underlying representations would thereby be forced to be closer to surface forms.  Others proposed to constrain the relation between phonological and phonetic representation directly (again, often in the name of ‘naturalness’).
In general, these attempts to limit the power of phonological systems by fiat ran into apparent counter-examples that deprived them of their appeal.  Other developments in phonological theorizing shifted scholars’ attention away from this issue, while also leading (as somewhat unintentional by-products) to a general reduction in the degree of abstractness of representation.  Some of these elaborations and re-orientations of the program of generative phonology are sketched below.
Autosegmental Phonology.  The bulk of research during the ‘classical’ period of generative phonology was concerned with segmental phenomena (although the main goal of Chomsky & Halle 1968 was an account of English stress).  In the early 1970s, attempts to describe the phonology of tonal systems led to important changes in assumptions about representations, and a concurrent shift of attention on the part of phonologists.
The classical theory had assumed that phonological (and phonetic) representations were given in the form of a simple matrix, where each row represented a phonological distinctive feature and the columns represented successive segments.  Such a representation is based on the assumption that there is a one-to-one relation between the specifications for any given feature and those for all other features, since each column contains exactly one specification for each feature.  Tonal phenomena, however, made it clear that features need not be synchronized in this way: a given feature specification might take as its scope either more or less than a single ‘segment’.  This led to the development of ‘autosegmental’ representations, in which feature specifications were linked by lines of association (subject to specific constraints) rather than all being aligned into segments.  The extension of this insight to other phenomena, and its consolidation, essentially displaced the earlier concerns of rule notation and ordering in phonologists’ attention.
Metrical Phonology. A similar development took place in the analysis of stress and the study of the syllable.  The analysis in Chomsky & Halle 1968 treated stress as simply one more phonological feature, with a value assigned to some (but not all) of the segments in the representation of a word.  This account was forced to attribute a number of basic properties to the feature [Stress], however, that had no obvious correlates in the behavior of other features.
It became possible to rationalize these properties by viewing stress not as a segmental feature, but as a relational property of the organization of syllables into larger structures.  This, in turn, required the recognition of syllables as significant structural units: a notion that was explicitly rejected in the earlier theory in favor of an attempt to re-formulate all apparent syllable-based generalizations in terms of segmental structure alone.  The organization of segments into syllables, and these in turn into larger units called feet, which themselves are organized into phonological words (and phrases, etc.), allows for the elimination of the anomalous character of segmentalized stress.  The study, within metrical phonology, of these units, their internal organization and their relation to one another completed the enrichment of the notion of phonological representation begun within autosegmental phonology.
Feature Geometry.  A standard theme of classical generative phonology was that of natural classes of phonological segments, groups of segments that function in some parallel fashion in phonological rules to the exclusion of others.  It was originally hoped that the analysis of segments into distinctive features would provide the solution to this issue: segments sharing a feature (or set of features) were thereby characterized as similar to one another, and thus predicted to behave in the same way in rules.
It soon became apparent, however, that feature analysis by itself does not exhaust this matter.  When nasal consonants assimilate in place of articulation to a following obstruent, for instance, each individual place is specified by a distinct feature (or set of features), and the overall unity of the process is not expressed.  Nothing in the notation, that is, makes it clear that a rule assimilating labiality, coronality, and velarity is more coherent in some sense than one assimilating labiality, voicing and nasality.  
The response to this problem was a program to treat the features themselves as organized into a hierarchy, such that all place of articulation features (for example), and no others, are daughters of a unitary node [Place].  On that approach, place assimilation could be viewed as a unitary association of the [Place] node, while no such single unit corresponds to the hypothetical alternative.  Attention focused on such problems of the internal geometry of the feature system generally assumed that the way to approach them was to assume that the theory of rules should be limited to a very simple set of re-associations and deletions within the autosegmental structure of an utterance, and that a single, universal feature hierarchy could be specified on the basis of which all observed ‘natural’ rules (and no ‘un-natural’ ones) could be formulated.  Arguments for and against specific proposals about such a hierarchy have drawn considerable attention, though it is perhaps notable that the theoretical assumptions underlying the program have been much less discussed.
Lexical Phonology.  In classical generative phonology, the interface between word structure and sound structure is quite simple.  Morphological elements are combined into words in the syntax, these elements are provided with phonological (underlying) forms, and the resulting syntactically organized labeled, bracketed structure serves as the input to the phonology.  At least some of the phonological rules were assumed to apply according to the principle of the cycle, based on this structure, in a uniform way.  To the extent morphological elements display different phonological properties in their combinations with others, this was represented as differences within an inventory of boundary elements separating them from adjacent items.
Originating from the apparent generalization that elements with the same phonological behavior (hence, associated with the same boundary type) tend to appear adjacent to one another, the theory of Lexical Phonology proposed a substantial revision to this architecture.  Instead of constructing the entire representation once and for all, and then submitting it to the phonology for realization, this view proposed that the lexicon of morphological elements is divided into multiple ‘strata’ or levels.  Basic roots can combine with elements of the first stratum; after each such morphological addition, the resulting form is subject to adjustment by the rules of a corresponding level of the phonology, and the output is then eligible to serve as the input to further morphological elaboration.  At some point, addition of elements from the first stratum is replaced by use of the morphology and phonology of the next, and from then on, no further elements from the initial stratum can be added.  This process continues (perhaps vacuously) through all of the strata of the lexicon, yielding a potential surface word.  All of the words in a given syntactic structure are then subject to adjustment by another set of post-lexical phonological processes.
There are a number of further points that characterize this view, including proposed differences in the properties of lexical and post-lexical rules, and the relations between rules on one level and those on the others.  The central point for a broader theory of grammar, however, is probably the replacement of a syntax-based (but purely phonological) notion of cyclic rule application by a repeated cycle of morphological addition and phonological adjustment.  This results, for example, in the possibility that a phonologically derived property (on one cycle) can be relevant to conditioning a morphological operation (on a following cycle), a possibility that has been shown to quite real.
Optimality Theory.  In the early 1990s, a much more radical challenge to the classical model was presented by the development of Optimality Theory (“OT”),  a view of phonology based on a system of ranked, violable constraints on surface shape as opposed to a system of ordered rules deriving the phonetics from an underlying phonological representation.  These constraints govern (in the standard formulation) a one-step relation between underlying and surface representations, with no intermediate stages of the sort produced in a rule-based description.  The constraints can be divided into general classes: (a) Markedness constraints, which express universally preferred configurations; and (b) Faithfulness constraints, requiring that contrasts present in the phonological representation be preserved in the surface form.  In general, these are in conflict, and the ranking of the constraints governs the resolution of those conflicts, in conformity with general principles of grammar.
Initially, OT seemed to offer its greatest promise in the analysis of stress, syllable structure, and related phenomena, but subsequent development has encompassed a full range of segmental and other facts.  Descriptions in constraint-based terms are at least superficially very different from those couched in terms of traditional rules, and theoretical discussion in phonology since their introduction has been largely dominated by comparisons of the two frameworks.
Current Approaches to Phonology
The central issues in phonology in the first decade of the twenty-first century concern the comparative merits of OT and rule-based descriptions.  On the one hand, constraint-based formulations seem much better equipped to describe global properties of phonological systems.  It was noted in work from the classical period of generative phonology that multiple distinct processes in an individual language may all have the effect of ensuring (or avoiding) a single characteristic property of surface form, but no satisfactory account of the unity displayed by these  “conspiracies” was ever achieved.  OT, in contrast, provides a very direct description of such facts.  
In some ways the surface constraint approach goes beyond anything available in principle to the rule-based theory.  For example, when languages accommodate loan words to the surface patterns of other words of the language, the adjustments needed to achieve this may include changes that do not correspond to any rule of the phonology of native forms.  Constraints accomplish this directly and without further stipulation, while a system of rules may have to be arbitrarily extended to account for loanword adaptation.
On the other hand, some of the same issues that rule-based phonology dealt with (and at least largely resolved) have re-surfaced as serious challenges to the architecture of grammar generally assumed in constraint based theories.  Most important among these is the problem of opaque generalizations.  The standard model of OT assumes that its constraints apply directly to surface forms, and govern a single-stage mapping between these and underlying phonological representations, and so has no place for generalizations that crucially apply to any sort of intermediate level.  Nonetheless, a number of compelling examples of such phenomena have been demonstrated, and some sort of accommodation of these facts must be provided by an adequate phonological theory.
Some responses to this challenge have attempted to maintain the standard OT model by introducing new sorts of constraints.  Mechanisms such as “Output-Output” constraints or “Sympathy Theory,” however, have not generally succeeded in dealing with all of the relevant phenomena, and have been shown to produce new difficulties of their own.
One approach that seems promising is that of “Stratal OT,” an architecture that grafts a constraint-based account onto the standard model of Lexical Phonology.  The result is a framework in which the phonological mapping at each stage is a one-step process governed by a constraint system.  Since the model is built on a cyclic interaction of phonology and morphology, however, it also provides for multiple successive stages in the overall derivation, thus accommodating opacity to the extent it can be related to morphological structure (as in the best established examples).
Examples also seem to exist in which the specific changes through which a language achieves conformity with a general constraint on surface forms do not follow directly from the content of the constraint (together with other interacting generalizations).  In such a case, something like a re-writing rule might be necessary, as a supplement to the constraint system – a notion which is clearly antithetical to the basic philosophy of OT.
A quite different problem concerns the very nature of universals of phonological structure.  Phonological theorizing has generally accepted the premise that generalizations which are true of phonological systems in general result from the cognitive organization of the human language faculty, and thus must be incorporated in some way into the architecture of phonological theory.  Recently, however, it has been argued that at least some such typological regularities result not from the content of a Universal Grammar constraining synchronic systems, but rather from the mechanisms of diachronic change that have produced the systems we observe.  To the extent this is true, it requires investigators to examine closely the arguments for incorporating any particular regularity into phonological theory per se as opposed to seeking its basis elsewhere.
While there have of course been other trends not covered above, it seems fair to say that the bulk of theoretical discussion in phonology from the 1960s to the present has been devoted to the elaboration and refinement of the generative program of Chomsky & Halle 1968.  The most recent developments in that tradition, involving the wholesale replacement of rules by constraints as the mechanism for expressing regularities of a language’s sound pattern, have shown great promise but cannot yet be considered wholly consolidated.  Apparently, some appropriate synthesis of the classical and OT models remains to be found, and it is that search that dominates discussion today.
— Stephen R. Anderson
Works Cited and Suggestions for Further Reading
Anderson, Stephen R. 1985. Phonology in the Twentieth Century: Theories of Rules and Theories of Representations. Chicago: University of Chicago Press.  (Development of phonological theory from its origins through the classical period of generative phonology)
Chomsky, Noam, and Morris Halle (1968). The Sound Pattern of English. New York: Harper & Row.
Gussenhovern, Carlos, and Haike Jacobs (2005). Understanding Phonology (2nd edition). New York: Oxford University Press. (Lucid elementary introduction to current phonology)
Kager, René (1999).  Optimality Theory: A Textbook.  Cambridge: Cambridge University Press.  (Clear introduction to the main ideas of Optimality Theory in phonology and their implementation)
Kenstowicz, Michael (1994). Phonology in Generative Grammar.  Cambridge, MA: Blackwell.  (Comprehensive description of the principal themes in phonology up to the introduction of Optimality Theory)