Why limit abstractness?

First we must understand what motivates concern over abstractness.

Limiting possible analyses

One reason to limit the divergence between underlying and surface forms is to constrain the theory of phonology, to prevent it from making wrong claims about how languages work. With no constraint on abstractness, every conceivable derivation from underlying to surface form would in principle be allowed by the theory. Just as the theory of phonology seeks to constrain the concept of “possible rule,” so that an imaginable rule such as {s, p, q, r} ! {m, l, t, v} / _ {s, k, ə, m} (unattested in any human language) can be ruled out on formal grounds, so too might we wish to rule out a derivation from underlying /qøɬijʌ/ to surface [gəˈraʒ] as too abstract. Since a goal of linguistic theory has been to restrict the class of theoretically possible languages to just the type that is actually observed, limiting abstractness in a well-defined way limits the number of possible languages.

Another reason for concern over abstractness is that it makes a particular claim about human cognition, that the mentally stored units of language can include things that the speaker has not actually heard, but arrives at by inference based on a line of indirect evidence. Since first language acquisition does not proceed by conscious reasoning, it cannot be taken for granted that everyday academic reasoning skills are automatically available to children.

Mental reality and language acquisition. This second consideration, whether abstractness (of some particular degree) is part of human cognitive capacity, is the most important question arising in this debate: this is a fundamental consideration for a theory such as generative grammar that seeks a model of language in the mind. Because the details of specific languages are not built into children at birth but must be induced from the ambient linguistic data aided by general cognitive capacity and whatever language faculty is universally available to all humans (i.e. the theory of grammar), a basic concern regarding the psychological reality of grammatical constructs – for phonology, rules, and underlying forms – is whether they can be learned from the primary language data.

The role of a universal grammatical component is to make the job of language acquisition easier, by uncompromisingly removing certain kinds of imaginable descriptions from consideration. Distinctive features are one way of making this job easier, since they limit the ways of analyzing data. Universal constraints on abstractness might similarly help a child trying to arrive at underlying representation for a language, and there have been a number of proposals as to the relationship between the underlying and surface forms. Attractive as it might seem to propose formal constraints on the theory of grammar to prohibit English from having /qøɬijʌ/ be the underlying form of [gəˈraʒ] garage, we will not actually assume that this is a matter for the formal theory of grammar; rather, it is a consequence of how a phonology is learned, thus the question of abstractness is outside the domain of grammatical theory.

Faced with a word pronounced [dɔg], a child learning English has no reason to assume that its underlying form is anything other than /dɔg/. But faced with the word atom [ˈæɾəm] and the related word atomic [əˈt h ɔmɪk], the child needs to arrive at an underlying representation for the root on which these two words are based, such that rules of English phonology can apply to derive the phonetic variants [ˈæɾəm] and [əˈt h ɔmɪk]: an appropriate representation would be [ætɔm]. It is in the face of such a specific motivation for an abstract underlying form that we would assume the underlying form isn’t simply the surface form. The solution to the so-called problem of abstractness which will be adopted here is, simply, that abstractness per se is not a problem: what really requires investigation is the kind of evidence that properly motivates a phonological analysis

Abstractness and phonemic representations. One particular degree of abstractness is widely accepted as self-evident, needing no further justification, namely that underlying representations do not contain allophonic variants of phonemes. It is generally assumed that English [stɔp], [t h ɔp] are underlyingly /stɔp/, /tɔp/, without aspiration, because there is (by assumption) no underlying aspiration in English. Similarly, we know that the underlying form of [hɪɾɪŋ] hitting is /hıtıŋ/, not only because the flap is an allophone in English, but also because of the related word [hɪt] hit where the [t] is directly pronounced. Thus, it is commonly assumed that underlying forms are at least as abstract as phonemic representations, with all allophonically predictable features eliminated.

This assumption can lead to problems. What is the medial consonant in the underlying form of a word like [waɾr̩ ] water? Assuming that the flap is not a phoneme in English (there are no minimal or near-minimal pairs contrasting [t] or [d] vs. [ɾ]), this forces us to say that it must be something other than [ɾ]. The word is spelled with t, but spelling is not relevant to underlying representations. Children acquire words without knowing how to spell, and most languages of the world are unwritten yet underlying representations must be acquired for all human languages. Spelling is also unreliable, and could lead us to the unjustified conclusion that the underlying vowels of [tuw] too, to, two, [θruw] through, [duw] due, and [druw] drew are all different.

Since [waɾr̩ ] is not composed of a root plus suffix, we cannot look at related forms to reveal the underlying consonant (as we can in wad-er versus wait-er, both [wejɾr̩ ]). Any number of hypotheses could be set forth – /waɾr̩ /, /watr̩ /, /wadr̩ /, /waðr̩ /, /waβr̩ /, /waγr̩ /, and so on. Hypotheses like /waβr̩ / and /waγr̩ / can be rejected on the grounds that they are pointlessly abstract, containing segments which do not occur phonetically in English, and there is no reason to believe that they exist underlyingly. Nothing is gained by positing such underlying representations, thus nothing justifies these hypotheses. Two facts argue decisively against hypothetical /waβr̩ /, /waγr̩ /, and their ilk. First, there is no evidence for a rule in English effecting the change /γ/ ! [ɾ] or /β/ ! [ɾ] and addition of such a rule, required to convert the underlying form into the surface form, rules against such an analysis since there exist analyses which at least do not force the inclusion of otherwise unmotivated rules. Second, a specific choice between /waβr̩ / and /waγr̩ /, or /waʔr̩ / and innumerable other possibilities which also lack an underlying flap, is totally arbitrary and leaves the language analyst – student and child alike – with the unresolvable puzzle “why this underlying form and not some other?”, which can only be resolved by fiat.

The hypothesis /waðr̩ / is less abstract since it is composed only of observed segments of English; it is, however, factually wrong, because it would be impossible to craft rules for English to turn /ð/ into a flap in this context (consider father, bother, weather which indicate that there cannot be a rule changing /ð/ into a flap in some context). Only three hypotheses remain viable: /waɾr̩ /, /watr̩ /, and /wadr̩ /. None of these hypotheses posits surface nonexistent segments, and given the rules of English – Flapping, specifically – any of these underlying representations would result in the correct surface form.

There is no standard answer to the question of the underlying form of water, but certain arguments can be marshalled to support different positions. We initially rejected the theory that the underlying form might be /waɾr̩ / because it posits what we assumed to be a nonexistent underlying segment in the language, but we should reconsider that decision, to at least explain our argument for rejecting an underlying flap. Hypothesizing /waɾr̩ / necessitates another phoneme in the inventory of English underlying segments, violating an analytic economy principle which says that you should select a parsimonious underlying inventory for a language. This perhaps reflects the basic principle of scientific reasoning that simpler, more economical solutions are better than complicated solutions that posit unnecessary machinery. But no concrete linguistic arguments indicate that elimination of phonemes is an actual goal of phonological acquisition. Economy of the underlying inventory cannot be judged in a theoretical vacuum, and in at least one contemporary theory, Optimality Theory, it is impossible to state generalizations about underlying representations, so it is impossible to say that English has no underlying flap.

A somewhat stronger argument against allowing an underlying flap is that the surface distribution of [ɾ] is restricted. It only appears between vocoids (vowels and glides), and only if the following vowel is unstressed, which is precisely the context where /t, d/ actively are changed into the flap [ɾ] (hit [hɪt] ~ hitting [hɪɾɪŋ]; hide [hajd] ~ hiding [hajɾɪŋ]). We can explain the lack of words in English like *[hiɾ], *[ɾuwl], *[æfɾr̩ ], and *[əɾǽk], if we assume that the flap [ɾ] is not in the inventory of underlying segments of English, and only derives from /t/ or /d/ by this specific rule. This argument recognizes the importance of capturing major generalizations about language, which is the central concern of linguistics: it says that it would be too much of a coincidence if, in assuming underlying /ɾ/ in water, we failed to note that underlying flap only appears in a very few contexts.

This argument is founded on the presumption that distribution of segments in underlying forms cannot be restricted: otherwise we would simply state a restriction on where underlying flaps appear and let the underlying form of [waɾr̩ ] be fully concrete. Some theories do not have conditions on underlying forms (Optimality Theory), others do. Something like conditions on underlying forms seems inevitable, since for example there cannot be any words in English of the form sCiVCi, hence *slil, *sneen, *spup, *skuck; yet, it is uncertain what status such conditions have in the theory of grammar. The assumption that all regularities about a language must be captured in the grammar has been a fundamental assumption for many theories of phonology, but has also been challenged (see Hale and Rice 2006), so we cannot take it for granted that the grammar is solely responsible for explaining the distribution of the flap in English.

Still, even if we decide that the underlying form doesn’t have a flap, that leaves open the choice between /t/ and /d/, which is purely arbitrary. The choice might be made by appealing to markedness, insofar as [t] is a less marked, i.e. crosslinguistically common, segment than [d]. Whether this reasoning is correct remains to be determined empirically.