Download non-linear phonology and more Schemes and Mind Maps Phonetics and Phonology in PDF only on Docsity!
Feature geometry
We have seen that tone and segmental duration m ust be autosegmentalized in order to account for a num ber of phonological phenomena. In this chapter, we will continue this theme. We will replace the matrix comprising the remaining distinctive features with a hierarchical configuration of these features. We will begin by considering a num ber of arguments for replacing the unstructured list of laryngeal, m anner and place features with a feature tree. The main advantage of this ‘autosegmentalized’ representation is that features or particular groups of features can spread to neighbouring segments, which greatly improves the description of assimilation processes. Moreover, the assumption made in Chapter (^6) that not all segments are specified for all features appears to have interesting consequences in the new model. Some of these will be discussed in this chapter, while others will be discussed in Chapter 13, which continues this topic.
Assimilations frequently show a non-arbitrary relationship between the structural description and the structural change. Time and again, rules appear to transfer a specific feature or group of features from one segment to a neighbouring segment. However, the representation we introduced in Chapter 7 is incapable of expressing this relationship. To illustrate the problem, consider once more the Dutch rule of which voices obstruents before [b,d]. Rule might be said to be a natural rule in the sense that voiced segments cause preceding segments to be voiced. It would be very strange if Dutch had a rule which caused segments to
become voiceless before [b,d]. The problem with the SPE notation, however, is that
it is in fact as easy to write a rule or constraint that describes such an implausible process (cf. (2)) as it is to write a perfectly ordinary rule like ( 1 ). That is, the notation represents a theory which cannot distinguish between the impossible and the commonplace.
1 This is not to suggest that dissimilation processes are in general impossible. What is implausible is that a dissimilation rule like (2) would ever be needed to describe a phonological process.
186 Feature geometry
1 REGRESSIVE VOICING
—son —cont
—son —cont
What would be an elegant way of expressing the fact that assimilations involve the adoption of a feature by one segment from another? First, we might think of a copying operation. This would produce a non-arbitrary result in that only features and feature values that are present in the context can be copied. However, as we have seen in Chapter 10, phonologies show a pervasive dislike of sequences of identical features (the OCP), and it would obviously be a bad idea to design a theory of rules whose essence it is to violate the OCP. A better solution would be to find a
representation which would allow segments to share the same feature(s), much in
the way successive TBUs can be associated with the same tone.
There is a second striking fact about assimilations. Commonly, the same groups of
features appear to be transferred from one segment to the next. For example, many languages have processes that transfer the features specifying a consonant’s place of articulation to a preceding nasal consonant. Since the features in the matrix are unstructured, it is impossible to express that the feature groupings we find in the processes of so many languages are in fact natural groups of features. That is, in a theory without a feature tree, an unnatural feature group like [l a b ia l , ± son] can be as easily expressed as a natural one like [ ± voice, ± spread, ±constr]. The solution to both problems is known as feature geometry. The problem of the grouping of features is solved by representing segments as trees, in which the nodes represent features and feature groups. The assimilation problem is solved by assuming that a single node m aybe part of more than one tree. We will first introduce the feature tree. Then, in section 12.5, we turn to the ways in which one feature tree (i.e. segment) can interact with an adjacent feature tree.
As was observed above, assimilations of place involve the transfer of a collection of features. Many languages have rules assimilating nasal consonants to the place of articulation of the following consonant. For instance, Hindi, which has the underlying nasals [m,n,r|J, assimilates the place of articulation of nasal consonants to that of a following consonant within the word. Word-internally, sequences like *[nk], *[q,t] or *[mt] are therefore ill-formed. As a result of n a s a l a s s im il a t io n , a prefix like [sam] ‘together’ shows alternations of the type illustrated in (3) (Ohala 1983). Clearly, the place features m ust be accessible as a group in order to express this
188 Feature geometry
must be as in ( 6 b). Further, on the basis of processes like that illustrated in (5) we would seem to need a representation like that in ( 6 c).
6 a.^ b.^ c.
+cons —son —voice —spread —constr —nas —lat —cont COR +ant —distr
+cons —son —voice —spread —constr —nas —lat —cont COR +ant —distr
+cons —son —voice —spread —constr —nas —lat —cont COR +ant —distr
What this means, of course, is that the segment is not a single constituent, but has internal structure. Such constituency is conventionally represented by means of a tree diagram. The development of feature trees (or feature geometry) is mainly due to Clements (1985), Sagey (1986) and McCarthy (1988).
12.4. The node dominating the place features is called the pl a c e node. Within this constituent there are four articulator nodes corresponding to the univalent features [l a b ia l ], [c o r o n a l ], [d o r s a l ], [r a d ic a l ]. Each of these four nodes dominates subconstituents corresponding to their relevant features. Thus, the [l a b ia l ] node dominates [±round] and [±distr], the [c o r o n a l ] node dominates [±ant] and [ ± distr] and the [d o r s a l ] node dominates [±back], [ ±high] and [ ±low]. This representation of the pl a c e node is given in (7). With the help of this representation, processes can refer to the constituent [l a b ia l , distr, round], but not to the nonconstituent [high, ant], for instance. We will assume that the position of [±tense] will be under [r a d ], assuming it is equivalent to [ ± ATR].
(7) PLACE
LABIAL COR DORSAL RAD
distr round ant distr back high low tense
Q
- Which constituent is transferred from a consonant to a preceding [t,d,n,l ] by English CORONAL ASSIMILATION?
- Which constituent is transferred from a consonant to a preceding nasal by Hindi n a s a l ASSIMILATION?
Building a tree 189
What other natural feature classes are there besides place features? Lass (1976: 145ff.) has argued that the laryngeal segments [h,?] have no supralaryngeal specification at all. He based his case on the frequently observed phenomenon that obstruents sometimes appear to lose all their supralaryngeal information. Consider the London English data in ( 8 ), which are representative of this kind of process. In certain contexts, preglottalized plosives are pronounced as [?]. (The variation between preglottalized plosives and glottal stops is stylistic, i.e. both pronunciations may be heard from the same speaker.)
pi?ktfo or pi?tJo^ picture
mii?t wiljom or mii? wiljom meet William
m?k næ?ks or ni? næ?ks knick-knacks
no?t næu or^ no? næu not now
kii?p smailir) or kii? smailir) keep smiling
In New World Spanish, [s] has been lost in coda position. That is, the plural of [klase] ‘class’, which in Peninsular Spanish is [klases], is [klase]. (The [e] arises from a separate process which is not relevant here.) However, the [s] has not left without a trace: an [h]-like offglide can be heard after the plural in careful speech. Similarly, voiceless sonorants have arisen word-internally after the loss of the oral articulation of [s], as shown in (9).
(9) Stage I Stage II
mismo miimo or mimmo ‘same’
fosforo fooforo ‘match’
Both processes could be described if, following Lass, we assume that laryngeal and supralaryngeal features are separate constituents. Under that assumption, both the English and the Spanish data would receive the same description: deletion of the supralaryngeal constituent. In the case of English, this would leave [ + constr, —voice, —spread] behind, i.e. [?], and in the case of Spanish it would leave [—constr, —voice, —spread] behind, i.e. a voiceless interval at that point in the word. Thus, such orphaned laryngeal features generally end up as [?] and [h], respectively. To express Lass’s proposal in our feature tree, we need to assume that segments consist of two main parts: the laryngeal and the supralaryngeal constituent. The dominating node is referred to as the r o o t node. If we want to ensure that segments that have only one of these constituents can be specified for the major class features ([± cons] and [± so n ]), these features m ust be present in either of these constituents. Following McCarthy (1988), we assume they make up the root node. This decision can safely be made, since no processes have been reported that transfer [ ±cons] or [ ± son] from one segment to the next. This means that our tree now looks like (10). We will treat the feature [ ± approx] as a manner feature.
Building a tree 191
(12) Underlying
in-hiakja oan-higrg
Northern
inhiakja ognhigra
Southern
fçjekja ooçjerâ
‘to h o o k in ’ ‘to liste n to ’
When Frisian [h] is transformed to [g], it acquires place and m anner features from another segment, but it does not itself lose any place or m anner features, because it did not have any. Segments other than [h,fi,?] do of course have m anner and place features. The location of manner features in the pl a c e /s u pr a l a r constituent in ( 10 ) predicts that segments can lose these features, while acquiring new ones from a neighbouring segment. Such a transfer should be possible independently of the laryngeal features. This can indeed be shown for the lateral consonants of Klamath. In addition to a voiced [1], Klamath has a globalized [I5] and a voiceless [1]. When appearing after [n], the lateral consonant transfers its lateral articulation to the preceding segment, but without its laryngeal features, which remain behind as [?] and [h], as shown in (13). This suggests that there is a constituent in the feature tree that contains m anner but not the laryngeal features. Since the new segment is not [+ nas], this Klamath process suggests that the pl a c b /s u pr a l a r node of (10) contains [±lat] (because it is transferred from the second to the first segment) as well as [±nas] (because it is deleted from the second segment).
(13) [nl’l - > [1?] [nl] -> [lh]
While it is clear that the m anner features are inside the supralaryngeal constituent, it is also clear that the m anner features cannot be under the pl a c e node. In Hindi, as in many other languages, it is just the consonant’s place node that is transferred to a preceding nasal, not its [—nas] specification. From the node labelled pl a c e /s u pr a l a r in (10), we must therefore split off the pl a c e node. Since there have been no reports of
processes that transfer [ ± nas], [± cont], [± lat], [± strid ], [ ± approx], en bloc, from
one segment to the next, these features are not grouped in a m anner constituent, but rather form separate terminal nodes dominated by s u pr a l a r , as shown in (14).
(14) ROOT cons son
LAR
spread voice constr nas
LABIAL COR DORSAL RAD
dist ro u n d ant dist back high low tense
192 Feature geometry
Single m anner features can be shown to transfer independently to neighbouring
segments. Processes that transfer [+ n as] are widely attested. (We have seen
examples of this in American English and Frisian here and in earlier chapters.) Transfer of [± co n t] is much rarer. In American English, forms like [gibm] for
given, [hi:dn] for heathen, [bidnas] for business and [wAdnt] for w asn’t occur
(Bailey 1985: 63). No case has been made for the transfer of [± strid ], [± lat] or [± approx]. There are many proposals that modify the structure of (14); however, this structure is one that is widely used, and may be seen as a ‘consensus’ tree (Broe 1992).
The structure of the tree introduced in the previous section explains w hat must be
made accessible to phonological rules, i.e. what the natural feature classes are. In this section, we will deal with the way features or subconstituents of the segment tree spread or delink. In our representation, segments are arranged from left to right, which is the usual metaphor for the order from ‘early’ to ‘late’. Along this time axis, the segment trees are arranged much in the way the records in an old- fashioned jukebox are arranged: a row of parallel disks, where a disk represents a segment. Each disk defines a plane in which some instantiation of the feature tree is to be found. If you were to take out a disk to look at its right or left face, you would see some version of (14). This is known as an end view. How can a node of some segment be dominated by the appropriate node of an adjacent segment? Imagine two adjacent segments SI and S2. Now mentally draw a line from the [± n as] node of SI to the [± n as] node of S2, and a line from the s u pr a l a r node of SI to the s u pr a l a r node of S2. The lines you have just drawn are, of course, tiers. All corresponding nodes form tiers in this way, so that we can talk about the ‘nasal tier’, the ‘supralaryngeal tier’, the ‘root tier’, etc. Evidently, tiers are adjacent if no tier intervenes between them. This is the case with the node pl a c e and the immediately dominating s u pr a l a r y n g e a l node, for instance: no node intervenes between them. Two adjacent tiers define a plane.
Q Draw two feature trees in pencil, next to each other. Draw all tiers. Now erase all pencil lines that are ‘behind’ the planes nearest to you. As you will realize, your picture shows three dimensions.
We can now draw, in some plane, an association line from any node to the immediately dominating node in an adjacent segment. That is, from the [l a b ia l ] node of SI we can draw an association line to the pl a c e node of S 2 , and from the pl a c e node of SI to the s u pr a l a r y n g e a l node of S 2 , and so on. This spreading of nodes in the feature tree is entirely comparable with the spreading of tones, or with the association of one vowel to two moras.
194 Feature geometry
The rule in (16) thus reads as follows: the pl a c e node of any consonant spreads to (the supralaryngeal node of) a nasal on the left, while the pl a c e node of the latter is delinked. Observe that in our interpretation of Clements’s display, the upper and lower tier are always adjacent. The higher node label (here: s u pr a l a r ) is therefore redundant.
( 16 ) NASAL ASSIMILATION
SC:
I SUPRALAR
\
PLACE SD: j^+nasj jVconsJ
Similarly, English c o r o n a l a s s im il a t io n is given in (17): ‘Spread the coronal node (with all the features it dominates) to a noncontinuant coronal on the left.’
(17) CORONAL ASSIMILATION
SC:
SD:
T\
\
\
PLACE
CORONAL
contj
Q117 We sa w in Chapter 6 that distinctive features were phonetically non-abstract in the sense that they can be defined in terms of phonetic concepts. Would you say that natural feature classes are phonetically abstract?
Q118 English has an optional rule assimilating [t,d,n] to the place of articulation of following velar and labial plosives and nasals, as show n in the data below. Formulate this English pl a c e a s s imil a t io n in the display format. Gin 0rm buk thin book ôæt ôæk kAp that cup god gub b3i good boy ttn tern mailz ten miles ôæt ôæt naît that night
In a num ber of ways, the proposal to represent the phonological content of segments as a feature tree has led to more insightful formulations of phonological rules. As explained in Chapter 6 , two assumptions have been made here. One is that not all
Implications of underspecification 195
segments have all nodes. Thus, laryngeal segments have no supralaryngeal node, and the place node has only those articulator nodes that positively specify the place of articulation of the segment, which implies, for instance, that [t] does not have a labial, dorsal or radical node (Sagey 1986, McCarthy 1988). Second, it has been assumed that predictable information is absent in underlying representation, and is supplied by default rules which apply at the end of the derivation. Below, a num ber of advantages of such underspecification in feature trees are briefly outlined.
12.6. The decision to leave predictable features unspecified may have a simplifying effect on the formulation of rules that neutralize contrasts to some neutral, ‘default’ realization. Let us take (Low) German f in a l d e v o ic in g as an example. As we have seen in Chapter 6 , the expected laryngeal feature specification for obstruents is [—voice, —spread, —constr]. Let us assume that these values are filled in before the phonetic implementation rules start their work if no other value is specified, causing all obstruents without laryngeal specification to be voiceless, unglottalized and unaspirated, f in a l d e v o ic in g neutralizes the distinction between voiced and voiceless obstruents in the coda (cf. the uninflected and genitive forms of ‘variegated’, [bunt - buntas], and ‘union’, [bunt - bundas]). If a default rule were assumed for German which supplies the value [—voice] to obstruents in the coda, f in a l d e v o ic in g could consist of a rule delinking the laryngeal node in coda position. In (18), we formulate this version of the rule.
(18) f i n a l d e v o ic in g
SC:
ROOT
LARYNGEAL
SD: Coda
A separate question is whether voiceless obstruents in a language like German can be left unspecified for the feature [—voice] in the lexicon. Even though [±voice] is contrastive for obstruents in German, which has contrasts like [pain] ‘pain’ - [bain] ‘leg’, it would be possible to leave voiceless obstruents unspecified, while specifying voiced obstruents as [ + voice]. The voiceless obstruents then only acquire their [—voice] specification at the end of the derivation, or, if the feature needs to be referred to by some rule, as early as it is needed. This would be the assum ption made in the theory of Radical Underspecification (Kiparsky 1982, 1993, Archangeli 1988), which, as its name suggests, goes one step further than the theory of Contrastive Underspecification, which was briefly discussed in Chapter 6. In Radical Underspecification, only one value of a feature is ever lexically specified. A survey of the issues is provided by Roca (1994: ch. 2) and Steriade (1995). In the next chapter, we will discuss a num ber of further instances in which underspecification would appear to facilitate the application of rules.
Conclusion 197
Q121 In Brao, voiced, voiceless unaspirated, voiceless aspirated and preglottalized voiced plosives are contrastive in the onset, as illustrated by [dak] ‘walk (verb)’, [truu:] ‘fish’, [thun] ‘se a so n ’, [Tduir] ‘type of flute’ (the last may vary with an implosive [d]). In the coda, the only plosives that can occur are voiceless unaspirated, as illustrated by [pe;t] ‘drink (verb)’ (Keller 1993).
- Provide the full laryngeal feature specifications for the four types of initial plosive.
- Which features could be supplied by default rules?
- What underlying laryngeal feature specification would you assume for the coda plosives?
If the representation of a segment consists of a self-contained list of features, it is impossible to express that, frequently, sequences of segments share particular features, a phenomenon known as assimilation. Also, the fact that such assimilations often involve specific groups of features, such as the features specifying the place of articulation, cannot be expressed. If, by contrast, the features are arranged in a constituent structure, or feature tree, with each constituent defining a tier, it does become possible for two adjacent segments to share the same feature. Also, because the feature tree can group features under a single constituent node, it will be possible for groups of features to be transferred, or spread, to an adjacent segment.
