USC Graduate Phonology ✳︎ Fall 2019 ✳︎ Smith

The rules of the grammar operate in a mechanical fashion; one may think of them as instructions that might be given to a mindless robot, incapable of exercising any judgment or imagination in their application. Any ambiguity or inexplicitness in the statement of rules must in principle be eliminated, since the receiver of the instructions is assumed to be incapable of using intelligence to fill in gaps or to correct errors."

– Chomsky & Halle (1968: 60)

The Sound Pattern of English (SPE)

• SPE: Chomsky & Halle’s (1968) book The Sound Pattern of English.
  • Proposes a framework that generates mappings via rule-based derivations.
• Algorithmic computation: the central goal of SPE is to develop fully explicit algorithms which generate all and only the surface forms of a language.
  • A fully explicit algorithm is a description of a process that can be carried out in a finite amount of time on a computational device.
    • See the quote at the top of this handout.
    • If the algorithm could be implemented in a modern programming language, it meets this requirement.
  • When multiple algorithms are possible, formal simplicity dictates the best one.

Anatomy of an SPE-style rule

• Rewrite rules take the form X ➔ Y / A ＿ B
  • Transforms AXB into AYB
  • “X becomes Y when it’s between A and B”
• Rules have the following parts.
  • Target / focus (X)
  • Structural change (Y)
  • Context (A＿B)
  • Structural description (AXB)
    • Describes the structure necessary for the rule to apply
• The environment may contain sounds that aren’t adjacent to the focus.
  • For example: X ➔ Y / A B C ＿ D E F
  • Transforms ABCXDEF into ABCYDEF
• Here are some examples, including some with features and other conventions you may have seen (#, ]_X, Ø, C, V).

• An important skill is being able to translate a rule into prose (and back again). Here are some tips and pitfalls for stating rules in prose:
  • Give rules helpful and descriptive names.
  • Be as unambiguous as possible.
    • The best way to test this is to show the prose to others and see if they can reconstruct the rule.
    • If the prose can be translated into notation and back again without loss, you’ve achieved a perfectly unambiguous description of the rule.
    • Sometimes a rule can’t fruitfully be translated into prose, and needs to be paraphrased. Make this clear if you need to do it.
  • “Precedes” implies “immediately precedes”, but is still possibly ambiguous. Same for “before”, “after”, etc.
  • “Between A and B” implies A_B, but is still possibly ambiguous.
  • It’s a good idea to use phonetic terms instead of features. For example, “An obstruent becomes voiceless at the end of a word” instead of “A [–sonorant] segment becomes [–voice] at the end of a word”. Features vary quite a bit.
  • “Left” and “right” can be usefully unambiguous.
• Properties of the focus can be written under the underscore, but this is relatively uncommon.

Rule ordering

• Rules are ordered and applied in sequence.
  • input ➔ rule 1 ➔ rule 2 … rule k ➔ output
    • The output of rule 1 serves as the input to rule 2.
  • Rules are extrinsically ordered.
    • Extrinsic ordering means you can’t determine the order of rules without looking at the mappings of the language.
    • This means that children need to learn the ordering of rules.
  • There are proposals for instrinsic rule ordering, determined by characteristics of the rules themselves, or by characteristics of the rules and URs.
  • How else could rules apply?
• Order matters – but not always.
  • Sometimes changing the order of two rules changes the output of the grammar, sometimes it doesn’t.
  • Let’s consider what (if any) order of rules is necessary to account for the mappings below, and write out a derivation.
    • Mappings: /bɛt-d/ ➔ [bɛtəd], /ʤʌmp-d/ ➔ [ʤʌmpt], /fɪl-d/ ➔ [fɪld]
    • Epenthesis Rule: Ø ➔ ə / [+coronal] ＿ [+coronal]
    • Devoicing Rule: [–sonorant] ➔ [–voice] / [–voice] ＿
  • There are five types of rule interaction. We’ll discuss these at length later, so it’s okay if you don’t remember them yet.
    • Feeding: X feeds Y if Y is applicable to some form only because the form has undergone X. (X creates an input for Y)
    • Bleeding: X bleeds Y if Y is not applicable to some form because the form has undergone X. (X destroys an input for Y)
    • Counterfeeding: X is counterfed by Y if Y could’ve fed X, but X is ordered first, so it doesn’t. (Y applies too late to feed X)
    • Counterbleeding: X is counterbled by Y if Y could’ve bled X, but X is ordered first, so it doesn’t. (Y applies too late to bleed X)
    • …or order doesn’t matter.

The evaluation procedure in SPE

• SPE suggested a metric for the evaluation of a system of rules, arguing that grammars with fewer symbols better capture linguistic generalizations.
  • Fact: There’s usually (always?) more than one analysis that works for the observed data.
  • Problem: How does the language learner choose one?
    • Conjecture: The language learner favors rules that are simple, and analyses with fewer rules; SPE defines “simpler” as “having fewer symbols.”
  • Corollary 1: Rules with fewer symbols should be more widespread among the world’s languages.
  • Corollary 2: A theory that can capture widespread rules or sets of rules with fewer symbols than it needs to capture rare rules or sets of rules is more likely to be a correct model of the learner.
  • Note that the evaluation procedure is for comparing analyses with the same empirical coverage.
    • A helpful aphorism attributed to Einstein: “Everything should be made as simple as possible, but not simpler.”
• Here is the evaluation procedure in the words of Chomsky and Halle (1968):
  • “The “value” of a sequence of rules is the reciprocal of the number of symbols in its minimal representation.” (p. 334)
  • “[…] the number of symbols in a rule is inversely related to the degree of linguistically significant generalization achieved in the rule.” (p. 335)
• When rules resemble each other in specific ways, this resemblance is a significant linguistic generalization which needs to be captured.
• Let’s brainstorm some other measures we could use to compare rule-based systems.
• To help capture generalizations and minimize the number of symbols, SPE used both phonological features and abbreviatory conventions to combine rules.

Phonological features

• In SPE, URs are built of distinctive features, and this is what the grammar manipulates.
• /kæt/ is really just an abbreviation for a sequence of three feature bundles. We write /kæt/ to preserve our sanity, but we really mean:

• There is quite a bit of variation in how features are used (see the reading Zsiga 2013 for some examples).
  • We’ll assume that URs are composed of a series of feature bundles.
  • We’ll assume the feature set is universal and the same in every language.
  • We’ll assume that features are binary: they are always + or –.
  • We’ll also assume that every sound is specified for most features.
  • But not every sound can be specified for every feature.
    • For example, most consonants don’t have features relating to tongue height like [high] or [low].
    • This is often represented as [0feature].
    • Rules cannot target [0feature] sounds –- just [+feature] or [–feature].
  • Disclaimers
    • There are proposals with non-binary (“privative”) features.
      • Especially for major place features, as discussed below.
    • There are proposals in which feature specifications are as minimal as possible, and all redundant/predictable features are added by rules.
      • For example, in a language without voiceless sonorants, sonorants don’t need a [voice] feature if a rule makes [+sonorant] sounds [+voice].
    • There are arguments against feature being universal (more on this below).
• Features allow us to combine related-looking rules by referring to natural classes.
  • A natural class is group of sounds in an inventory which share one or more distinctive features, to the exclusion of all other sounds in the inventory.
  • These two rules of English…
    • n ➔ n̪ / ＿ θ
    • l ➔ l̪ / ＿ θ
  • …can be combined into one rule:
    • [+coronal] ➔ [+dental] / _ [+dental]
    • “A coronal becomes dental before a dental.”
    • This rule predicts that any [+coronal] sound will become [+dental] before any [+dental] sound.¹
• In SPE-style rules, e.g. [+coronal] ➔ [+dental]…
  • Features in the focus mean: find a sound with the required features, in the required environment.
  • Feature in the change mean: make the specifed change.
  • The rule [+coronal] ➔ [+dental] / _ [+dental] does the following:
    • 1. find a sound that is [+coronal]
    • 2. if the [+coronal] sound is followed by a sound that is [+dental]
    • 3. change the [dental] feature of the [+coronal] sound to [+dental]

How do we decide what the features should be?

• Phonologists have developed a more-or-less consensus set of features, based on the processes seen in languages.
• This is based on the assumption that features are innate and universal.
• Typical desiderata:
  • If two sounds are contrastive in some language, they must differ in at least one feature.
    • E.g., if tap [ɾ] and trill [r] are contrastive in Spanish, we need a feature to distinguish them, like [trill].
  • If a group of sounds is targeted by a rule, they should be describable with a set of features.
    • E.g., if we see a rule like {A, B, C, D} ➔ … or … / {A, B, C, D},
    • especially in more than one language, then {A,B,C,D} should be a natural class.
    • E.g., if lots of languages nasalize a vowel before a consonant with nasal airflow (like [m, n, ŋ]), then [+nasal] seems like a good feature.
    • E.g., the feature [+strident] picks out the set of sounds that trigger epenthesis in the English plural [s, z, ʧ, ʤ, ʃ, ʒ]. Without [strident], these sounds wouldn’t form a natural class.
  • A feature should correspond to some phonetic property.
    • Most are articulatory, but they can be acoustic too.
• A different view is that feature systems are neither innate nor universal.
  • The language learner discovers his/her language’s feature set during the course of language learning via exposure to ambient language patterns.
  • As a result, a sound shared by two languages could have different features in each language.
  • A typological survey in Mielke (2007) finds that existing feature systems can only account for about 70% of rule-targetting classes in natural languages.
  • There are many “unnatural” processes that refer to disjointed sets of sounds.
    • For example, consider this pattern from Kolami (Mielke 2011 citing Emenau 1961).²
      • Allomorph [–l] after / ʈ ɖ n̪ i iː e eː a aː /
      • Allomorph [–ul] after / p t̪ k d̪ g s z m ŋ j /

Phonetics review

• The rtMRI chart via SPAN
• Voicing (left vs. right in the call)
  • voiced, voiceless
  • creaky voice
  • breathy voice
  • aspiration
• Place (columns)
  • labial
  • labiodental
  • dental
  • alveolar
  • postalveolar
  • palatal
  • retroflex
  • velar
  • uvular
  • pharyngeal
  • glottal
• Manner (rows)
  • stop (a.k.a. plosive)
  • nasal (a.k.a. nasal stop)
  • fricative
  • affricate
  • approximant (glide, liquid)
  • lateral (lateral approximant)

Commonly used features worth knowing

• You do not need to memorize these, but they are the features most commonly used.
• Voicing features
  • [voice]: voiced sounds are +
  • [spread glottis]: aspirated sounds are +
  • [constricted glottis]: glottalized/creaky voice and ejectives are +
• Major place features
  • [labial]: lips are active articulator (e.g., bilabial, labiodental)
  • [coronal]: tongue tip/blade/front is active articulator, passive articulator is teeth, alveolar ridge, or palate
    • Dental, alveolar, postalveolar, alveopalatal, palatal
    • Further broken down with [distributed] and [anterior], two features which are only specified for coronals.
    • [distributed]
      • sounds made with tongue blade or front are +
      • sounds made with the tongue tip are –
    • [anterior]
      • sounds made at the alveolar ridge or fronter are +
      • sounds made further back are –
  • [dorsal]: tongue body/back is active articulator (e.g., palatal, velar, uvular)
    • Note: palatal is both [+dorsal] and [+coronal]
  • [pharyngeal]: tongue root is active articulator
  • [laryngeal]: “glottal” sounds, e.g. [h] and [ʔ]
  • Major place features are often unary.
    • A coronal sound simply has a [coronal] feature, and non-coronals don’t have a [coronal] feature.
    • This means rules can’t target [–coronal] as a natural class.
• Manner features
  • [syllabic]: vowels and syllabic consonants are +
  • [consonantal]: vowels and glides are –
  • [sonorant]
    • Obstruents (non-nasal stops, fricatives, affricates) are –
    • Sonorants (nasals, approximants, vowels) are +
  • [continuant]
    • Fricatives and approximants are +
    • Nasals are [–continuant]!
  • [nasal]: nasals are +
  • [lateral]: laterals are +
  • [delayed release]: affricates are +
    • In some charts, fricatives are also +
  • [strident]: certain fricatives and affricates are +
    • Fricatives and affricates with loud, high-pitched noise
    • [f, v, s, z, ʧ, ʤ, ʃ, ʒ, χ, ᴚ] are +
    • [ɸ, β, θ, ð, x, ɣ] are –
• Vowel features
  • [high]
  • [low]
  • [front]
  • [back]
  • [round]
  • [ATR], [tense]
    • [i e u o] are [+ATR] or [+tense]
    • [ɪ ɛ ʊ ɔ] are [+ATR] or [+tense]
    • Low vowels depend on the feature system.
  • Most feature systems use vowel features to distinguish between different [+dorsal] consonants, in addition to using them for vowels.

Individual feature charts

• The class materials contain a few different feature charts, if you’re curious to see how they differ.
  • Elisabeth Zsiga’s (2013) Phonetics and Phonology feature chart.
    • Place features and some other features are unary: they aren’t [+] or [–].
  • Bruce Hayes' (2009) Introductory Phonology feature chart.
    • All features are binary.
    • Has an [approximant] feature, and specifies fewer sounds for every feature. For example, [strident] is only used for coronals, and [delayed release] is only used for obstruents.
  • Jason Riggle’s (2008) feature chart.
    • All features are binary.
    • Has a [tense] feature AND an [ATR] feature. The former is used to distinguish between vowels and dorsal consonants, and the latter is used for vowels only.
  • A feature chart for SPE, typeset and commented by Ewan Dunbar.
    • This chart also has a very useful key to translate the IPA used in SPE to IPA you’re familiar with.
• To be explicit, we’ll use the Hayes' chart for our next assignment, since it has good coverage and the [approximant] feature is useful, and I have an Excel spreadsheet.
  • Most of the time, our rules will operate the same regardless of the chart we use.
  • In your write-up, be very explicit about the features you assume.
• Here is a very helpful summary for the manner features.

Abbreviatory conventions

• Rules can also be combined using abbreviatory conventions. Here are four of the most common:

• Each of the rules above is expanded into sub-rules in pre-defined ways. Here are some made-up examples, some of which are a bit strange.
  • C ➔ Ø / ＿ {C, #}
    • First, try C ➔ Ø / ＿ C.
    • Second, try C ➔ Ø / ＿ #.
  • C ➔ Ø / ＿ (C)C
    • First, try C ➔ Ø / ＿ CC.
    • If the first sub-rule didn’t apply, try C ➔ Ø / ＿ C.
  • C ➔ Ø / ＿ C⁵₀
    • First, try C ➔ Ø / CCCCC ＿
    • If it didn’t apply, try C ➔ Ø / CCCC ＿
    • If it didn’t apply, try C ➔ Ø / CCC ＿
    • If it didn’t apply, try C ➔ Ø / CC ＿
    • If it didn’t apply, try C ➔ Ø / C ＿
    • If it didn’t apply, try C ➔ Ø / ＿
  • C ➔ [αnasal] / ＿ [αnasal]
    • Apply these two sub-rules simultenously:
      • C ➔ [+nasal] / ＿ [+nasal]
      • C ➔ [–nasal] / ＿ [–nasal]
  • C ➔ [αnasal] / ＿ [–αnasal]
    • Apply these sub-rules simultenously:
      • C ➔ [+nasal] / ＿ [–nasal]
      • C ➔ [–nasal] / ＿ [+nasal]
  • C ➔ [αnasal, βvoice] / ＿ [αnasal, –βvoice]
    • Apply these sub-rules simultenously:
      • ɑ+, β+: C ➔ [+nasal, +voice] / ＿ [+nasal, –voice]
      • ɑ–, β+: C ➔ [–nasal, +voice] / ＿ [–nasal, –voice]
      • ɑ+, β–: C ➔ [+nasal, –voice] / ＿ [+nasal, +voice]
      • ɑ–, β–: C ➔ [–nasal, –voice] / ＿ [–nasal, +voice]
• Let’s see abbreviatory conventions in action using the following example from SPE (p. 339, adapted slightly). Consider a fictional language in which:
  • / i / and / j / are actualized as [ j ] before / p r j a /.
  • / w / and / u / are actualized as [ u ] before / p r j a /.
  • Let’s assume this fictional language has the same consonants as English and five vowels: / i e a o u /.
• At first glance, it looks like we need 8 rules, but we can use braces, features, and trickery to reduce this down to just one rule. Solution at the bottom.

• Although it may seem needlessly complex, being able to combine these rules is (probably) a good thing: whenever we have different rules that seem to be doing the same thing, we want to be able to combine them.

A real example from SPE

• The fictional example is far from the most complex rule in SPE.
  • Rules like the vowel tensing rule below have many sub-rules, capturing a complicated set of requirements for application.
  • Without abbreviatory conventions, SPE would have many, many more rules for vowel tensing.
  • Since rule simplicity plays a pivotal role in the evaluation of grammars, the effect of abbreviatory conventions in reducing rules’ structure is a good result.

• The top part of this rule expands into the sub-rules:

  • Sub-rule #1: V ➔ [+tense] / ＿ V
  • Sub-rule #2a: [+low, +stress] ➔ [+tense] / _ #
  • Sub-rule #2b: [–low, +stress] ➔ [+tense] / _ #
  • Sub-rule #2c: [–low, –stress] ➔ [+tense] / _ #

• Let’s think about the predictions of this rule for the vowels below.

• After tensing applies, another rule reduces unstressed lax vowels to [ə].

Some potential issues

• The evaluation metric is blind to phonetic naturalness.
  • Many sound patterns can be accounted for by appealing to principles of production and perception.
  • …but the evaluation metric only counts symbols.
• As a theory of possible phonological alternations, SPE is possibly too powerful.
  • Consider the prediction from above: Rules with fewer symbols should be more widespread among the world’s languages.
  • It’s quite easy to write two rules with the same number of symbols, where the process captured by one rule is very common and the processed by the other rule is very rare.
    • For example: word-final devoicing (very common across the world’s languages) and word-final voicing (very rare across the world’s languages).
  • In fact, nearly any phonological process can be generated using rules.
• There are sometimes related-looking rules that can’t be combined using abbreviatory conventions.
  • When this happens, we’re failing to capture a generalization.

A recap of some of the defining characteristics of SPE

• The following are major characteristics of the SPE research program, many of which set it apart from earlier work.³
  • Generative grammar, URs and SRs:
    • SPE is a model of mappings between URs and SRs.
  • Fully explicit algorithms:
    • The goal is to generate all and only the mappings of a language.
    • Formal simplicity dictates the best analysis, and is achieved using features and abbreviatory conventions.
  • The SRs of a language are generated via a sequence of phonological rules.
    • input ➔ rule 1 ➔ rule 2 … rule k ➔ output
    • A priority for the phonologist is to determine – to the greatest extent possible – the ordering of rules.
  • Features:
    • In SPE, representations are linear sequences of feature bundles.
    • They contain nothing more – not even syllables.
    • Later work in phonology proposed richer representations, containing syllable structure, autosegmental features, metrical feet, moras, etc.
  • More abstractness:
    • In SPE, there are fewer limitations on proposing underlying segments.
    • Analyses may contain URs that differ substantially from SRs, even containing segments that never appear on the surface, giving more abstract URs.
    • For example, to account for stress, SPE proposes a distinction between final /ss/ (in caress) and final /s/ (in Paris) (p. 151).
• These differences are why Chomsky and Halle (1968) distance themselves from earlier work and explicitly avoid the terms phoneme and allophone (p. 11, p. 65). Earlier work in the Structuralist program…
  • … uses unordered “taxonomic” rules such as “Phoneme A is realized with variant B in context X––Y”.
    • In SPE, rules are ordered, apply in sequence, and are defined as algorithms.
  • … has rigid requirements for positing phonemes, such as the requirement that every phonetic representation has a unique phonemic representation, and the requirement that the set of phonemes be as small as possible.
    • In SPE, the goal is to construct a model that generates the correct surface forms, and a UR is justified so long as it explains some data.
  • From this point on, we’ll rarely see the terms phoneme and allophone.
    • We’ll often talk about sounds in URs and sounds in SRs.
    • You should feel free to use phoneme and allophone, since it’s clear that you don’t intend the terms in the pre-1950s sense.

Recap of major ideas

• The Sound Pattern of English
• explicit algorithm
• rules and rule ordering
• distinctive features
  • natural class
  • desiderata for features
• rule formalism
  • target / focus
  • structural change
  • environment
  • structural description
• the evaluation procedure in SPE
• abbreviatory conventions (), {}, C_n, Greek letter variables
• abstract URs

Practice problems

• Imagine a language that has the inventory shown below.
  • Nine of the circled sets of sounds are natural classes.
  • Describe each class in terms of distinctive features, using only the features defined in this handout.
  • Use as few features as possible to uniquely identify each set relative to the whole inventory.
  • Indicate the set that is not a natural class.
  • [ɻ] is a retroflex approximant.

• Translate the following rules from SPE-style notation to English.
  • [+syl, +high] ➔ [−voice] / [−voice, +cont] ＿ [−voice, +cont]
  • [+dorsal, −cont, −son, +voice] ➔ [+cont] / [+syl, +back] ＿ [+syl, +back]
  • [−son, +cont] ➔ [−cont] / ＿ #
  • [+voice, −cont, −son] ➔ Ø / [+nasal] ＿ #
  • [+syl, +low] ➔ [−low, +round] / ＿ C₀ [+syl, +high, −back]
• Translate the following rules from English to SPE-style notation.
  • Nasals delete preceding fricatives, l, and r.
  • In a sequence of two (non-nasal) stops, the first becomes a fricative.
  • Voiceless stops and fricatives become voiced between sonorants.
  • [ɻ] becomes [l] preceding [l].
  • Mid vowels agree with the vowel of the preceding syllable in ATR value.
• Write a system of rules for /t/, capturing the distribution in the diagram, along with the generalizations below. Make the rules as minimal as possible.
  • /p/ and /k/ also aspirate before stressed vowels.
  • /d/ also becomes a tap between vowels.

Solution

• We can create two rules using braces.

• There’s still some similarities between the two rules not being captured.
• Braces, features, and some other conventions reveal the similarities:

• Adding braces gives us one rule:

• We could also obtain a single rule using lowercase Greek letters.