USC Graduate Phonology ✳︎ Fall 2019 ✳︎ Smith

Recall: justifying constraints

• There are three ways to justify a constraint:
  • Formal: resembles other constraints in CON.
  • Functional: reflects articulation, perception, or memory.
    • Examples: Agree, Ident-Onset[Voice]
    • Potentially includes audience design
  • Typological: its cross-linguistic typological predictions are borne out.
• Today we’ll tackle functional justification.

Phonetics vs. phonology

• Widespread assumptions about the generative grammar (see e.g. Steriade 1997, who argues against this model, and Hayes 1999 who assumes it)
  • There is a phonological component of the model
    • Determines which sounds contrast and where
    • Manipulates phonological representations
  • There is a phonetic component of the model
    • Maps phonological representations onto articulatory instructions
  • These two models are separate, but the phonological model feeds the phonetic model.
    • The interaction is “one-way” (feedforward with no feedback) – phonology doesn’t consider physical factors directly, although phonological processes are possibly shaped (over time?) by functional considerations and phonetic “naturalness.”

Definitely phonological rules (at least probably)

• There are some processes that seem unambiguously phonological (see Kenstowicz & Kissberth 1978, Ch 2: The Nonphonetic Basis of Phonology for examples).
  • They refer to grammatical or lexical information.
    • For example, they refer to syntactic categories.
    • For example, they are sensitive to morpheme boundaries.
    • For example, they have morphological or lexical exceptions.
  • They aren’t phonetically natural, usually due to historical change.
    • Telescoping: multiple sound changes – each of which is natural – results in an phonetically unmotivated synchronic alternation.
    • Inversion: e.g., a phonetically natural rule of deletion is reanalyzed as a rule of epenthesis.
    • Paradigm leveling and analogy
  • These assume phonetic rules don’t refer to boundaries or lexical information, which now seems wrong, at least with respect to prosodic boundaries and item-specific phonetics.

Crazy rules

• Bach & Harms (1972) How do languages get crazy rules? addresses how this might happen.
  • This is a great paper to know and is a real classic.
• One of their examples is Japanese palatalization.
  • Here is the pattern:
    /t/ ta tʃi tsu
    /s/ sa ʃi su
    /d/ da dʒi (zu)
    /z/ za (dʒi) zu

• Bach & Harms say: “Under any feature system we can think of the rule exhibits interconnections that do not seem have any particular phonetic or phonological plausibility” ¹
  • Affrication before [u] seems especially unnatural.
• They suggest that this “crazy rule” arose via a series of simplifications and combinations of more plausible rules (see telescoping above).
• For each step, let’s write out the inventory.

• Why not three rules?
  • Palatalization, affrication, neutralization of voiced strIdents
  • All three processes target the same class of obstruents
  • The three processes aren’t independent: no dialect has affrication by itself.
  • The three rules would never be reordered with respect to another, and no rule ever comes between. (They act as a unit in terms of rule ordering.)
• Because of processes like these, the dream of a strictly innate universal constraint set probably won’t be fulfilled. For crazy processes, we probably need a way for learners to posit crazy constraints.

Phonetic and phonological “doublets”

• We’ve discussed what phonological processes can look like, now let’s return to phonetics.
• Both components – phonetic and phonological – can generate very similar looking processes.
  • Sometimes a phonetic process becomes phonologized, wherein the phonetic process becomes a phonological process.
  • Here are some examples from Hyman (2013) of processes that can either be phonetic or phonological, depending on the language.
  • If the triggering environment disappears due to a sound change, the phonetic conditioning is lost, and the process is unambigously phonological (maybe even resulting in contrast).

• How do we determine what’s phonetic and what’s phonological if it’s like one of the processes above?
  • This isn’t as easy as you might think…
  • Here are some characteristic differences from Hyman (2013), who cites: Cohn 1998, 2007; Keating 1996; Keyser and Stevens 2001; Kingston 2007; Pierrehumbert 1990; Stevens and Keyser 1989, etc.
  • Both phonology and phonetics seem to be language-specific to some degree.
  • Both phonology and phonetics seem to be subject to optionality to some degree.

• Two traditional diagnostics from Hyman (2013). A process has been phonologized if…
  • A phonetic effect is exaggerated beyond what can be considered universal (e.g. duration increases more than we’d expect)
  • A ‘categorical’ rule of phonology refers to it. (e.g. a clearly phonological process refers to vowel length, than that length must be phonological)

Revisiting the strict divide between phonology and phonetics

• If we assume that phonology can come from phonetics, an open question is how to connect them.
• Two papers today:
  • Hayes (1999): A feedforward model, except for the induction of phonological constraints. Learning (phonologized) constraints involves a tradeoff between increasing ease of articulation and minimizing complexity.
  • Steriade (2001): Complicated typological patterns in assimilation can be explained by appealing to ease of perception and confusability.
    • Both papers use maps of difficulty: one for articulation (Hayes) and one for perception (Steriade).

Hayes (1999): learning phonological constraints using phonetics

• Questions for discussion:
  • What are the main goals of the paper?
  • What factors influence the ease of voicing in obstruents?
    • Let’s go over them and discuss how and why.
    • What about sonorants?
  • Where does the map of articulatory difficulty come from?
  • How is the map used to induce constraints?
  • Under Hayes' model, to what extent is CON universal?
  • Does Hayes' model address the too-many-repairs problem?
  • What constraints (if any) have we talked about that couldn’t be induced under the algorithm presented here?

Steriade (2001): faithfulness constraints driven by perception

• Why do so many languages have /an+pa/ → [ampa] but not /an+pa/ → [anta], /ap+na/ → [apma]?
  • One possible answer à la Lombardi (1999): protect onsets
  • Steriade’s approach: in /an+pa/, /p/’s place is well cued (release burst, outgoing formant transition), while /n/’s isn’t.
  • Be faithful to the better-cued contrast.
  • Example of how different contexts for a stop have better and worse perceptual cues.

• Why this looks like a good idea: protecting onsets can’t account for how place and manner of articulation affects the directionality of assimilation.
  • The likelihood of a process being regressive depends on what the two sounds are, in both place and manner.

• Given the basic idea (preserve the better-cued constrast)
  • …speculate why apicals are progressive while non-apicals are regressive
  • …speculate why nasal+obstruent sequencess in apicals are regressive
• The implemented model: speakers have a “P-map”
  • implicit knowledge of perceptual distance ∆(X,Y) between any pair of sounds X, Y (potentially tagged for their contexts):

• We’ll consider a simplified example, where ∆(p/__V, t/__V) > ∆(n/__C, m/__C) (∆ for difference)
• Faithfulness constraints can refer to details of their target and their surface context:
  • not just Ident(place) but Ident(place):[–son]/ __ V, Ident(place):[+nas]/ __ C
  • not just Dep-V, but Dep-i, Dep-a, Dep-ə, Dep-V/s__t, Dep-V/t__r, …
• Constraints penalizing big changes should outrank constraints penalizing small changes:
  • Ident(place): [–son] / __V >> Ident(place):[+nas] / __ C
• Presumably these default rankings can be overturned by the learner in response to contradictory data, but they will be a drag on language change

• No matter where we rank the markedness constraint in relation to the faithfulness hierarchy, winner is (b) or (a):

Some things to ponder about the P-map

• Exactly what is being compared in ∆(X,Y), to give a faithfulness constraint its default ranking?
  • Output vs. input? That’s kind of funny because the input isn’t a pronounced form, so its perceptual properties are hypothetical.
  • Output vs. faithful output (candidate a in the above)?
  • Output vs. related output? E.g., [rat] vs. [rad-im].
    • Those are both real, pronounced forms, but it’s tricky because the target segments are in different contexts.
    • Do we measure ∆(d/V__V, t/V__#)?
• How well connected is the P-map?
  • Can ∆(X,Y) be measured for absolutely any X,Y? Or only for close-enough pairs?
  • Does ∆(X,Y) really act like a number, so that we can always compare ∆(X,Y) and ∆(Z,W)?
  • Or is the “greater than” relation sparser than that, so that some distances can’t be compared?

The “too-many-solutions” problem revisited

• Some markedness constraints have a variety of “solutions”
  • *NC̥, as we saw
  • OCP-labial in various Western Austronesian languages (Zuraw & Lu 2009)
  • *{ɪ,ʊ} in Romance metaphony (Walker 2005)
  • *InitialGeminate (Kennedy 2005)
• This is what we expect in OT
  • But some don’t — that’s the “too-many-solutions problem”:
    • *NC̥ never causes epenthesis
    • *CC deletes C₁, not C₂ in VC₁ C₂ V (Wilson 2000; Wilson 2001)
    • *[–son, +voice]# causes final devoicing, but not deletion, epenthesis, etc.
  • The lack of solutions for *[–son, +voice]# is predicted, if the P-map imposes a difficult-to-overturn ranking: Max-C, Dep-V >> Ident(voice)/__#
    • Deletion and epenthesis result in very perceptually salient differences, while devoicing doesn’t.

Recap

• Crazy rules
• Phonetics v. phonology: tough calls
• Universal CON: maybe not
• Hayes: Induction of markedness based on articulation
• Steriade: Perception-based faithfulness