Presentation on theme: "The Role of Features in Phonological Inventories Nick Clements Laboratoire de Phonétique et Phonologie (UMR 7018) CNRS / Paris."— Presentation transcript:
The Role of Features in Phonological Inventories Nick Clements Laboratoire de Phonétique et Phonologie (UMR 7018) CNRS / Paris Universitas
2 A word of thanks to those who have made this conference possible... For making this day possible: the UMR 7023 (Université de Paris 8) the Fédération Typologie et Universaux Linguistiques (TUL), CNRS FR 2559 For making the topic possible: Euclid of Alexandria (325 BC-265 BC), the most prominent mathematician of Antiquity. His treatise on geometry, the Elements, the only textbook of Antiquity still in use today, has influenced all branches of science.
3 "There is no royal road to geometry"
4 Organization of this talk Part 1. Two geometries of feature organisation Part 2. (Some) principles of inventory structure Part 3. Case study: the feature [+nasal]
5 A feature geometry tree Source: Avery and Rice (1989) Properties 1.Each node lies on a separate autosegmental tier (example) 2.Constituents correspond to sets of features that function as a unit in phonological rules of spreading, deletion, etc. (example) 3.The constituent structure is proposed as universal, not subject to parametric variation 4.Constituents usually correspond to independent articulators or regions of the vocal tract Root Laryngeal[continuant] [voice]Supralaryngeal Sonorant [nasal][lateral] Place Labial CoronalDorsal [round][distributed]
6 A feature geometry tree Source: Avery and Rice (1989) A node is justified if and only if a rule of some language targets it. Examples: Root Laryngeal [continuant] [voice]Supralaryngeal Sonorant [nasal][lateral] Place Labial Coronal Dorsal [round][distributed] -in Berber, all Root features of stops are geminated in the imperfect - in Korean, all Laryngeal features are deleted in the syllable coda - in Spanish, nasals assimilate to a following consonant in all Place features - in Basque roots, sibilants agree in all Coronal features
7 A feature geometry tree Source: Avery and Rice (1989) Constraint on rules Any rule may operate on one and only one node in the tree For example, a rule may target the Place node, spreading it to a neighboring segment, or deleting it. All dependent features spread or delete with it. Root Laryngeal[continuant] [voice]Supralaryngeal Sonorant [nasal][lateral] Place Labial CoronalDorsal [round][distributed]
8 A feature geometry tree Source: Avery and Rice (1989) 2 major predictions 1. Only sets of features forming constituents can be directly affected by phonological rules 2. These potential sets are the same from language to language Root Laryngeal[continuant] [voice]Supralaryngeal Sonorant [nasal][lateral] Place Labial CoronalDorsal [round][distributed]
9 A second type of feature tree
10 A feature hierarchy (adapted from Clements 2001) This tree displays the 14 commonest [+consonantal] sounds in the UPSID database (Maddieson & Precoda 1989) Upper-case letters represent types of sounds as characterized by the features that dominate them, e.g. P = any voiceless labial obstruent [sonorant] [posterior] [continuant] - [+consonantal] [labial] [dorsal] [strident] [voiced] [lateral] P B K G T D S TS M W N L R J _ _ _
11 A feature hierarchy tree Properties 1. Features are ranked in a hierarchy of preference, from top to bottom 2. Feature ranking is based on the crosslinguistic frequency of contrasts 3. Ranking is partial; six tentative ranks are shown above (separated by dashed lines) 4. Ranking is universal (strongest version) or probabilistic (weaker version) 5. Rooted feature paths define natural classes of sounds [+consonantal] [sonorant] [labial] [dorsal] [posterior] [strident] [voiced] [continuant] [lateral] P B K G T D S TS N L R J _ _ _ M W
12 A feature hierarchy tree Constraints on tree construction 1. The subset of features chosen by any given language must conform to the ranking 2. All nonterminal nodes must branch (thus there are no redundant feature specifications) [+consonantal] [sonorant] [labial] [dorsal] [posterior] [strident] [voiced] [continuant] [lateral] _ _ _ P B K G T D S TS M W N L R J
13 A feature hierarchy tree An ill-formed subtree: non- terminal nodes fail to branch. Such a tree would specify the redundant feature [+voice] for sonorants. [+consonantal] [sonorant] [labial] [dorsal] [posterior] [strident] [voiced] [continuant] [lateral] _ _ _ P B K G T D S TS M W N L R J
14 A feature hierarchy tree Predictions: 1. Inventories tend to employ higher-ranked features before employing lower- ranked features 2. Inventories tend to expand from top down (historical change, acquisition, etc.) 3. Inventories tend to contract from bottom up (historical change, aphasia, etc.) [+consonantal] [sonorant] [labial] [dorsal] [posterior] [strident] [voiced] [continuant] [lateral] _ _ _ P B K G T D S TS M W N L R J
15 A feature hierarchy tree Predictions (cont.): 4. If a given contrast appears within a marked category (+), it also appears in the corresponding unmarked category (-) ("priority of the unmarked") 5. For any constituent, the least marked sound of the unmarked branch is more frequent than the least marked sound of the marked branch [+consonantal] [sonorant] [labial] [dorsal] [posterior] [strident] [voiced] [continuant] [lateral] _ _ _ P B K G T D S TS M W N L R J UPSID frequency rank
16 The two models compared feature geometry tree feature hierarchy tree what does it model?segmentssegment inventories universal structure?yes phonetic motivation?mostly articulatorymostly auditory predicts rule types?yesno predicts inventory structure? noyes represents segmental markedness? yes (see below)
17 Segmental markedness can be directly represented in both theories Feature Geometry tree: only marked features are present (in some versions) -- as a result, markedness is directly reflected in the number of nodes. Example: the representation of [t] vs. [m] (following Avery and Rice 1989). Root = [t] Laryngeal Supralaryngeal Place Root = [m] Laryngeal Supralaryngeal Place Labial Sonorant [nasal] [t] = 3 nodes[m] = 6 nodes
18 Segmental markedness can be directly represented in both theories In the Feature Hierarchy tree, marked features are plus-specified (Clements 2001) As a result, markedness is directly reflected in the number of plus-specifications. Example: the representation of [t] vs. [m]. [+consonantal] [labial] [dorsal] [strident] [voiced] [lateral] P B K G T D S TS M W N L R J - [sonorant] [posterior] [continuant]
19 Summary In phonology, -- Feature geometry represents natural classes of features (those that function together as a unit), while the feature hierarchy represents a hierarchy of natural classes of sounds (sounds that function together as a unit) -- Rules are better expressed by the feature geometry, but Inventory structure is better represented by the feature heirarchy, as it is specifically designed for this purpose -- Markedness relations can be directly represented in both models
20 Part 2. (Some) principles of inventory structure
21 Observation: not just any set of consonants and vowels can make up a sound system -- a central finding of the earliest work in phonology was that sound systems are structured in terms of regular correlations defined in terms of features (see e.g. Trubetzkoy 1969/1939, Hockett 1955) -- the current project is to search for the principles that define inventory structure
22 The theory of inventories presented here is based on the following main components: -- The Feature Hierarchy -- Feature Economy -- Markedness Avoidance -- Enhancement Theory We now go on to consider the latter three, and their interactions.
23 Feature Economy: "maximize feature combinations" Languages tend to maximize the combinatory possibilities of features within their inventories of speech sounds; in other words, features present in one segment tend to be used to define others. The feature economy of any given system can be quantified in terms of a measure called its economy index. Given a system with S speech sounds and in which F features are required to characterize them, its economy index E can be given (to a first approximation) by the expression E = S/F The higher the value of E, the greater the economy. Feature economy can be defined as the tendency to maximize E. Martinet 1955; Clements 2003a, 2003b
24 Markedness Avoidance: "avoid marked feature values" Feature economy is counteracted by the tendency to avoid marked feature values Examples of widely avoided segment types: -- laryngealized vowels -- nasalized fricatives -- voiceless nasals The Marked Subset Principle: Within any class of sounds in which a given feature F is potentially distinctive, sounds bearing the marked value of F < ("are less frequent than") sounds bearing the unmarked or null value of F. --In other words, languages tend to avoid marked feature values, regardless of the class of sounds they occur in. Jakobson 1941, Greenberg 1966, Calabrese 1994, Clements 2005
25 A standard variety of English: 24 consonants p h t h tS h k h b ddZ g f TsS v DzZ m n N w l ry h [+voiced] and [+continuant] are used with near-maximal efficiency [+nasal] creates nasal stops at three of the four available places of articulation but there are a number of significant gaps, corresponding to marked feature combinations
26 Phonological enhancement: "enhance weak auditory contrasts" Enhancement can be achieved by introducing redundant features. Marked feature values are favored for this purpose when they serve an enhancing function. Example: [+rounded] (a marked feature value) frequently enhances the contrast between front and back high and mid vowels for instance, / i u / offers a better auditory contrast than / i ö / --- As a result, if a language has just two high vowels, it is most likely to have / i / and / u /. (Stevens & Keyser 1989, 2001, etc.)
27 Phonological enhancement: "enhance weak auditory contrasts" However, such cases counteract the Marked Subset Principle: In most contexts, the marked value [+rounded] is less frequent than its absence, as we expect (for example, rounded consonants are less frequent than plain consonants) It is just in the class of high and mid vowels that the expected trend is reversed: the marked value [+rounded] is more frequent than its unmarked counterpart [-rounded] (or [Ørounded]) Even in such cases of "frequency reversal", the value [+rounded] remains marked by most other criteria (e.g. spreading of marked values, neutralization to the unmarked value)
28 Summary: what the model predicts 1. For a marked feature value M and its unmarked counterpart value U, in any class of sounds, Economy: M = U "the number of segments bearing the marked value M is equal to the number of segments bearing the unmarked value U" example: voiced fricatives (a marked category) should be as numerous as voiceless fricatives (their unmarked counterparts)... recall English
29 A standard variety of English: 24 consonants p h t h tS h k h b ddZ g f TsS v DzZ m n N w l ry h voiced fricatives are just as numerous as voiceless fricatives (a Feature Economy effect)
30 Summary: what the model predicts 2. Markedness Avoidance: M < U "the number of segments bearing the marked value M is less than the number of segments bearing the unmarked value U" example: voiced fricatives (a marked category) should be less numerous than voiceless fricatives (their unmarked counterparts); this is true in German, for example This prediction contradicts the previous one. Both cannot be right in one and the same comparison. The expectation is that one will win out in some cases, and the other in others.
31 Summary: what the model predicts 3. Enhancement: M enh > U (M enh = a marked enhancing feature value) "the number of segments bearing a marked enhancing value M enh is greater than the number of segments bearing the unmarked value U" Again, this prediction contradicts the previous one. However, they are in complementary distribution: Enhancement applies to a special subset of cases in which Markedness Avoidance applies (those in which the marked feature value plays an enhancing role).
32 Summary: what the model predicts The interaction of these principles thus predicts the following: In any class of sounds and for any feature F, if M is an enhancing feature M enh, then M > U; otherwise, M U. _________________________ Note: It is assumed here that Enhancement always takes precedence over Feature Economy. This is true of all cases reported in Clements (2005).
33 Part 3. Case study: the feature [+nasal]
34 [+nasal] is marked all languages have oral sounds, but some lack nasal sounds oral sounds are more frequent than nasal sounds [+nasal] frequently spreads, while [-nasal] rarely if ever does [+nasal] is marked with respect to [-nasal] by nearly all criteria, for example: Some phonologists, following the original intuition of Trubetzkoy (1939), regard [nasal] as a (marked) privative feature.
35 Predictions 1) M U 2) M enh > U because either [+nasal] = [-nasal] (by Feature Economy) or [+nasal < [-nasal] (by Markedness Avoidance), because [+nasal] > [-nasal] in enhancing contexts
36 Data base Evidence is drawn from the expanded UPSID data base (Maddieson & Precoda 1989) contains 451 phoneme inventories (representing 6-7% of the world's languages) geographically and genetically balanced electronic database, facilitating rapid searches results can be independently verified by others Results are statistically evaluated for significance of trends Maddieson 1991, Clements 2003
37 Case 1: Nasality in obstruents [+nasal] is absent in obstruents, or rare (12.6% of the sample languages) if we analyze prenasalized sounds such as NT or NZ as nasalized obstruents. This result is as expected (M < U).
38 Case 2: Nasality in vowels Case 2: Nasality in vowels 100% of the sample languages have oral vowels, but only 22.6% have nasal vowels. This result is again as expected (M < U). But nasal vowels are rather well represented in the world's languages. Are there some languages in which nasal vowels outnumber oral vowels (contrary to the predictions)?
39 Nasal vowel systems (diphthongs excluded) type A: M U (nasal vowels oral vowels) i "â u u)i "â u u)i "â u u) e oe e) o o) e) o) a a) a a) a a) next commonest: 46unattested: 0commonest: 56
40 Distribution of nasal vowel systems Distribution of nasal vowel systems GroupGroup totalType AType BTotal A+B South American Niger-Congo North American Indo-European23448 Papuan39134 Sino-Tibetan21213 Khoisan4202 Austro-Tai14101 Caucasian7011 Dravidian6101 Nilo-Saharan23101 Totals
41 This result is as expected This result is as expected Type A systems (M < U) respect Markedness Avoidance, but violate Feature Economy Type B systems (M = U) respect Feature Economy, but violate Markedness Avoidance Type C systems (M > U) violate both Feature Economy and Markedness Avoidance N.b. The absence of languages having only nasal vowels is just the limit case of Type C.
42 Case 3: Nasality in sonorant consonants In the class of sonorant consonants overall, [+nasal] sounds are disfavored, but only by a small margin as shown below (the larger of the two figures is bolded): oral sonorants = 447 languages (unmarked) nasal sonorants = 435 languages (marked)
43 Case 3: Nasality in sonorant consonants However, if we carry out separate comparisons on continuants (rhotics, glides, other nonlateral approximants) and noncontinuants (nasal stops and laterals), a striking dissymmetry emerges: [-nasal][+nasal] A. sonorant continuants 433 w j R 5 w) jâ Râ B. sonorant noncontinuants 368 l ´ 435 m n ø...
44 Case 3: Nasality in sonorant consonants In class A, the unmarked [-nasal] value is overwhelmingly favored, as expected. -- But in class B we find a "frequency reversal" in which segments bearing marked [+nasal] are the more frequent [-nasal][+nasal] A. sonorant continuants 433 w j R 5 w) jâ Râ B. sonorant noncontinuants 368 l ´ 435 m n ø Why should this be so?
45 [+nasal] enhances [-continuant] in the class of [+sonorant] sounds [The release of a nasal consonant] "results in a significantly greater increase in the spectrum amplitude over a wide frequency range than does the release of a lateral consonant" (Stevens & Keyser 1989: 92) / nV / vs. / rV /makes a better contrast than / lV / vs. / rV / That is, the acoustic mark of the feature [-continuant] – an abrupt increase in amplitude at the consonant-vowel boundary – is more abrupt at the release of nasal stops than at the release of oral stops. The claim is that nasals offer a better contrast than laterals to [+sonorant, +continuant] sounds – that is, the class of rhotics, glides, etc. In this view, for example, for any vowel V,
46 Testing the claim Let us consider the distribution of nasals and laterals in "R- systems" – systems containing at least one an R-sound (such as / R/ or / r/ ) and just one noncontinuant sonorant series, either nasal or lateral, at the anterior coronal place of articulation (where all widely occur). Such systems fall into one of two types: type A: R N type B : R L The expectation is that type A systems should be more numerous If this explanation is correct, systems contrasting / nV / vs. / rV / should outnumber those contrasting / lV / vs. / rV /. This prediction can be tested.
47 Testing the claim The prediction is correct: Type A (R N): 54 Type B (R L ): 2 That is, in "R-systems", the noncontinuant series is almost invariably nasal rather than lateral ( 2 = , p<.0001) The two-term Type A system /N R/ is overwhelmingly preferred to the two-term Type B system /L R/, as is predicted by Enhancement Theory. _______________________________ In systems lacking R-sounds, N is also commoner than L, but by a smaller and statistically non-significant margin (Clements 2005).
48 Summary We find, then, that the model correctly predicts the distribution of [+nasal] and [-nasal] sounds across phoneme inventories. M Ueither [+nasal] = [-nasal] (by Feature Economy) or [+nasal < [-nasal] (by Markedness Avoidance), but M enh > U[+nasal] > [-nasal] in enhancing contexts. The marked value [+nasal] is less frequent than [-nasal] in most sound classes, as predicted by Marked Feature Avoidance, but it is more frequent than [-nasal] in systems in which continuant and noncontinuant sonorants contrast, as predicted by Phonological Enhancement.
49 Our current project... is to continue exploring the factors that condition sound inventories. Components of this research include studies of: inventory structure across the world's languages and within particular areas (Africa) phonetic bases of distinctive features within the context of quantal/enhancement theory (Stevens & Keyser) phonological acquisition, both normal and deviant (Portuguese, French,...) the internal structure of particular languages (Ikwere project, with S. Osu)
50 Some References Avery, Peter & Keren Rice "Segment structure and coronal underspecification," Phonology 1989, Clements, G.N "Representational Economy in Constraint-based Phonology." In T. Alan Hall (ed.), Distinctive Feature Theory, Berlin: Mouton de Gruyter. Clements, G.N "Feature Economy in Sound Systems", Phonology 20.3, Clements, G.N "The Role of Features in Speech Sound Inventories." To appear in Eric Raimy & Charles Cairns, eds., Contemporary Views on Architecture and Representations in Phonological Theory. Cambridge, MA: MIT Press. Halle, Morris The Sound Pattern of Russian: a Linguistic and Acoustical Investigation. The Hague: Mouton. Reprinted Hockett, C.F A Manual of Phonology. (Memoire 11 of IJAL 21.4, Part 1.) Baltimore: Waverley Press. Reprinted by the University of Chicago Press, 1974 Maddieson, Ian & Karen Precoda "Updating UPSID," UCLA Working Papers in Phonetics Martinet, André Economie des changements phonétiques : Traité de phonologie diachronique. Berne: Francke. Second revised edition Maisonneuve & Larose, Paris, Stevens, Kenneth N. & Samuel Jay Keyser "Primary features and their enhancement in consonants," Language 65.1, Trubetzkoy, N.S Grundzüge der Phonologie, Göttingen: Vandenhoeck & Ruprecht. English edition Principles of Phonology, tr. C.A.M. Baltaxe. Berkeley and Los Angeles: University of California, 1969.