We take a major diversion from 'Vowels and Consonants' to look at The Blackwell Companion to Phonology.

Since the purpose of these blogs posts is partly to communicate some linguistics to a interested but non-specialist audience, that introductory book was a good place to be. Alas, the other purpose was to provide incentives to get through my reading list, and that book is sufficiently introductory that writing each summary post took longer than reading the chapter! Now, we move onto something a little more dense, and I will need my note-taking...

(Apologies for any Atlantic spelling that follows - the book is in American English, and I appear to be randomly alternating between copying out quotes verbatim, and replacing the spelling with my own...)

Underlying representations

What is the point of phonology?

Orthography (developing a writing system)
Description of language
Mental encoding (figuring out what goes on in speakers' brains)

All of these assume some abstraction from real (messy, continuous, interconnected) speech, which they call the underlying representation.

Differences in UR

Is UR an artifact of the description, or a real mental thing?

This is somewhat analagous to the debate as to whether the universe can coincidentally be described very well with maths, or whether the mathematical models are in some sense real in their own right - perhaps the Platonic Ideals to which the physical world aspires. A helpful difference is that with sufficient advances in neuroscience, phonologists may be able to answer the question with measurements, whereas physicists have to resort to philosophy. Although then we are left with measurements of the brain, and the question of whether physical biology is the same as a person's thoughts, and philosophy pops up again...

How is it related to morphological form and phonetics?
Does it encode morphosyntactic content?

A given theory can model:

Diachronic or synchronic phenomena (changes in time or just at one point)
Dialectal or style-dependent variation
Corpus data or speaker intuitions
Child production
Poetry or language games (intentional manipulation of phonology)

Phonemes

Introduced in the 1870s, one view was that the phoneme was an abstraction of a part of speech.

The other was that it was a unit of contrast - to distinguish words from one another, like 'pit' and 'bit'.

With only knowledge of a sequence of phonemes, a native speaker will be able to produce a phonetically detailed word that is recognisably the same word on each repetition. This core phonological representation is why varying pronunciations from men, women, and children from Scotland, London and New York are considered 'the same word'.

Indeterminacy in phonemic representations

It is not always straightforward to go from speech to phonemic representation.

Fuzzy phonemes

Indeterminacy can happen when the 'surface' segment could be either one or two underlying segments, and it is not clear which.
E.g.

bird - a vowel, or a sequence vowel-r?
'ng' - a velar nasal, or /ng/?
Tense diphthongizing vowels
Schwa (ə) - an allophone of /ʌ/, or a sound in its own right?

In American English, words like 'better' and 'data' may have /ɾ/ instead of /t/ in the middle. This doesn't change the meaning - but native speakers know they are saying 'bedder' as opposed to 'better' (unlike many other allophones which native speakers don't notice).

So the flap is a 'fuzzy' phoneme which "shares some but not all of the properites of more robust phonemes".

Neutralization

Some sounds that contrast in certain places don't contrast in others.

E.g. English voicing: try saying "cats and dogs". Notice that it will come out as "cats and dogz" - the voicing of the sequence of consontants in the coda has to agree. The difference between /s/ and /z/ is neutralized in the coda.

'Prague School' phonology posits 'archiphonemes': an underlying phoneme in these situations which leaves certain features unspecified, and which therefore takes them from surrounding cues as necessary.

E.g. Japanese /s/ and /ʃ/ - they contrast everywhere except before /e/ and /i/: only 'se' and 'shi' are possible, not 'she' or 'si'.

An alternative to archiphonemes is 'abstract phonemes'. The abstract phoneme /s/ is underlying, and simply has an allophone

[ʃ] that appears before /i/ or /j/. (This may be familiar from kana.)

E.g. Basque palatal sonorants

/l/ and /n/ become palatalized in some Basque dialects following /i/ or /j/, before a vowel. If /j/, then the /j/ was absorbed into the new segment.

If we assume ordered rules - first palatalization, then /j/ deletion - this is straightforward to explain. If we then apply these rules morpheme-internally (as well as when part-words combine to give the necessary conditions), then we can get rid of /ʎ/ and /ɲ/ as phonemes.

But should the morpheme-internal ones get a 'free ride' on the inter-morpheme rules? In this case, there is evidence that this more abstract phoneme is correct - but we have to choose between a larger phoneme inventory and a more abstract phonemic representation.

UR in morphophonemic representations

'American structuralist' phonologists use morphological alternations to figure out the underlying representation (e.g. atom-atomic implies that /ɾ/ really is /t/ underlyingly). Others think that only phonological information should be considered.

With a morphophonemic level, rules can generate the surface forms, whereas without one, a different UR has to be listed for each environment.

UR in Generative Phonology

Chomsky and Halle's 'Generative Phonology' assumes only a morphophonemic level. The phonological rules operate on the output of syntactic structures, which are morpheme-based, not word-based; it must be morphemes, not words, which are lexically encoded in this view.

Maximising grammatical generalization

URs not only specify lexically constrastive features, they also try to maximise grammatical generalization: leave out any feature that is predictable from the phonological content, map to all observed surface forms of the morpheme, express regularities in the distribution of sounds in the language overall.

Underspecification in underlying representation

Features that are non-contrastive in a language are omitted from the underlying representation.

Taking this a step further, features that are non-constrastive in a certain context can also be omitted from the UR. This is called underspecification.

This is pretty similar to the archiphoneme from earlier, when underspecification occurs in a neutralization environment.

Radical underspecification claims that for binary features, one is 'marked', and specified in the UR. The other does not need to be included in the UR, and is filled in by default. This theory does have its advantages, but on the other hand it gets even more abstract - some segments may be completely lacking in defined features!

URs and novel word formation

Invented words form plurals, past tenses etc in a way consistant with phonological rules being applied. However, it could be that each word has allomorphs stored with it, and new words have allomorphs generated by analogy to existing words. Alternatively, phonological rules could be applied, but not to abstract URs; there could a mixture of rules and stored forms.

Indeterminancy in morphophonemic representations

Indeterminancy in UR selection

Which morpheme reflects the underlying form, if any? This question is less important if your theory doesn't think that the UR is a mental construct, but just a helpful description: you can list all the alternatives and leave it there.

Abstractness in URs

What is the limit on how abstract an UR can be?

Again, this problem is greater for Generative Phonology - morphemes vary more than segments in their surface forms.

Kiparsky's Alternation Condition disallows absolute neutralization - i.e. all phonological derivations of an underlying form results in the neutralization of a contrastive element. This means that you cannot introduce 2 different underlying forms just to explain why a rule applies in some places and not others.

Some people think this is too restrictive, since there are sometimes advantages to breaking this condition. Kenstowicz, for example, argues that you should judge validity based on what produces the simplest grammar, and produces the broadest generalization in characterising sound patterns.

An opaque pattern is one where the environment that gives rise to a change in surface form is itself not present in the surface form, and so it is not clear where the change came from.

Abstract URs can explain seeming exceptions to e.g. stress rules, since the form which the stress rules apply to need not be the surface form.

However, do we want this? Do speakers treat them as predictable, or memorised exceptions? "Productivity may be implemented as analogical patterns without the need for abstract, morphophonemic URs".

Indeterminacy in word relatedness

How do you choose if two words contain the same morpheme? In 'walk, walked', it's fairly obvious, but what about "long, length"? Just because two words are historically - diachronically - the same, that doesn't make them the same synchronically - now, in the mind of a speaker who has never heard Old English.

Another problem is that as your vocabulary expands, you would therefore need to change the UR - which could then have "ripple effects possibly extending throughout the rule system". (E.g. when you learn 'righteous', presumably some years later than 'right'.)

Summary

"The problem of the indeterminancy of URs remains largely unresolved today."

Underlying representations in Optimality Theory

Generative Phonology has rules which are input-orientated: they are triggered by certain inputs.

Optimality Theory (OT), by constrast, has constraints on surface form.

OT claims that URs are entirely unconstrainted - "Richness of the Base" - any possible phonological structure can potentially be a UR in any language. Each morpheme has a unique UR, like in Generative Phonology.

'Lexicon Optimization' forces the UR to be that which is most harmonious with the surface form - i.e. that which gives the fewest constraint violations for equally ranked constraints. (See my OT post.)

OT 'allows for the possibility of eliminating URs altogether'.

Given the emphasis on surface constraints, there is more phonetic specification of non-contrastive details, both in URs and in constraints.

Exemplar Phonology

Generative Phonology's position that phonological representations had psychological reality went unchallenged for many years.

Newer evidence suggests that phonetic detail is stored in the long-term memory of words, and that frequency of occurance can have an effect on the pronunciation of individual words.

Exemplar Phonology originates from psychological theories of categorisation. Rather than phonetic detail arising from an abstract lexical representation, "the abstract elements... are formed on the basis of statistical patterning of phonetic detail as experienced by the speaker/hearer". Higher level phonological structures (features, phonemes, syllables) are therefore not universal, but vary between individuals and even words.

Articulatory Phonology

Despite phonetic attributes being associated with phonological features, Generative Phonology does not equate them with actual articulatory or acoustic parameters.

Articulatory Phonology uses articulatory gestures as the atoms of phonological encoding. Gestures represent the actions of the lips, tongue and jaw and are co-ordinated in 'ensembles'. "Segments have no direct representation in this model, but may be emergent from stable and recurring gesture ensembles."

Like Exemplar Phonology, it does not recognise "explicit, distinct levels of phonological representation", and doesn't try to model morphophonemic alternation.

Unlike Exemplar Phonology, each phonological forms has "a distinct and singular representation, not a cluster of individual instances of spoken words".

Ramblings on linguistics

The Blackwell Companion to Phonology, Volume I: Chapter 1: Underlying Representations