Vera Lee-Schoenfeld
Phonology Seminar, Term Paper
December 5, 2002
1
Third Tone Prominence*
1. Introduction
The four tones of Mandarin Chinese are involved in various sandhi processes. The most well-known and
thoroughly studied of these phenomena is traditionally assumed to turn a third tone into a second tone
before another third tone. Despite all the attention third tone sandhi has received in the literature (see e.g.
Yip 2000 and Chen 2000), Optimality-Theoretic (OT) accounts of the phenomenon are sparse and not
always convincing. In this paper, I follow Yip’s (2000) initiative of working out an OT account that is
driven in part by de Lacy’s (1999, 2002a) tone prominence constraints. These constraints favor higher
tones in prominent and lower tones in non-prominent positions.
In the remainder of this introductory section, I will provide some background information about the
Mandarin tone system, the third tone sandhi phenomenon, and de Lacy’s tone prominence constraints.
Then, in section 2, I will lay the groundwork for my analysis. The OT part of the proposal will be fully
fleshed out in section 3. The last two sections are devoted to residual issues and the conclusion.
1.1 Mandarin tones and third tone sandhi
The four contrastive tones of standard (i.e. Beijing) Mandarin are listed in (1). T1-T4 represent the four
tonal categories, the numbers in square brackets indicate pitch levels, and the letters stand for high (H),
mid (M), and low (L) tone segments.
(1) Mandarin tone system (Chen 2000)
T1 high level [55] H
T2 rising [35] MH
T3 dipping [214] MLH
T4 falling [51] HL
T1 is the only level tone; the other three are contour tones. The three pitch levels of T3 form a complex
contour. In connected speech, the citation form of these tones is often altered. Besides various
coarticulation effects resulting in pitch level changes that are too slight to be perceptible to the unaided
ear, there are a few tone sandhi (TS) processes, dissimilatory or assimilatory, which seem to result in
* This paper is an extension of previous work I did on third tone sandhi in focus positions. The main goal is to build on
the crucial parts of the earlier paper, gain a deeper understanding of the phenomenon, present new insights, and solve some of the
problems my OT account faced. My thanks to Jenn Smith and Armin Mester for consulting with me and encouraging me to stay
with this ongoing project.
2
categorical tone changes. The most well-known of these processes is T3 sandhi, which, according to
Chen and other researchers (e.g. Lin 1993), can be broken down into two parts, as shown in (2).
(2) T3 Sandhi
a. T3 [214] ‡ T2 [35] / __T3 [241]
b. T3 [214] ‡ T3 [21] / elsewhere
(2a) is the prototypical T3 sandhi rule which is traditionally represented as a T3 changing to a T2 before
another T3, and (2b) is what Chen calls the “half T3 sandhi.” This latter process turns the citation contour
[214] into the low truncated form [21] when T3 occurs anywhere but before another T3. Utterance-
finally, the final rise of [214] is optionally preserved. Thus, Chen lists the following allotones of T3.
(3) Allotones of T3
allotones
[214] [21] [35]
before [214] – – +
utterance-final + + –
sandhi
contexts
elsewhere – + –
In section 2 of this paper, I will take issue with the [35] representation of the T3 sandhi allotone. I will
also add another environment in which the full underlying [214] contour of T3 is realized. Generally,
however, I agree with the concept of a T3 allotone system.
1.2 Tone Prominence
Yip (1980, 2000), who takes T3 to be underlyingly specified as L only, represents T3 sandhi as a
dissimilatory process that inserts an H tone segment (or “toneme”) between two Ls. The resulting LH is
then perceived as MH, i.e. a T2. This is shown in (4)
(4) T3‡ T2 / __ T3
L ‡ MH/ __ L (or: L.L ‡ MH.L)
In order to make her “H-insertion” representation of T3 sandhi compatible with de Lacy’s (1999, 2002a)
tone prominence constraints, which are given in (5), Yip concludes that T3 sandhi must apply in binary
constituents which are left-headed. Insertion of H on the non-head of the constituent would violate de
Lacy’s *NONHD/H, which bans high tones from foot non-head positions.
(5) Tone prominence hierarchies
*HD/L >> *HD/M >> *HD/H
*NONHD/H >> *NONHD/M >> *NONHD/L
3
De Lacy’s constraints1, which Yip (2000) classifies as negative positional markedness constraints,
combine a tone prominence scale (H > M > L) with a structural prominence scale that has the Designated
Terminal Element (DTE) as its relevant unit of structure. The DTE or head of a foot is more prominent
than its non-head. The given rankings are universally fixed. In the next two sections, I follow Yip in
proposing an account of T3 sandhi that is compatible with de Lacy’s tone prominence constraints. My
analysis differs from Yip’s in terms of the empirical generalizations on which I base my proposal. While
both Yip and Chen argue that T3 changes to T2 under contrastive stress, I maintain that the full
underlying contour, namely [214] gets preserved.
2. T3 sandhi as toneme deletion
In this section, I present and defend my version of Chen’s (2000) T3 allotone system and explain how it
leads to a positional faithfulness (Beckman 1997) account. Furthermore, I will compare my proposal with
Yip’s (2000) and question the nature of the domains in which TS applies.
2.1 T3 allotones
Since the low truncated [21] allotone of T3 can nicely be derived from the full underlying contour [214]
by deletion of the final toneme, I propose that a T3 preceding another T3 (traditionally represented as
[35]) be derived by a deletion process as well. In order to derive a rising tone from [214], all that needs to
happen is deletion of the low middle toneme. The result is [24], which presumably does not differ much
from [35] perceptually. In fact, even if there is a slight audible difference, the proposal of representing a
post-sandhi T3 as [24] instead of [35] finds support in the literature. As reported by Chen (2000),
Zee (1980) shows that underlyingly T3 preceding another T3 has an overall lower F0
(fundamental frequency) than the corresponding T2 in the same environment. This would seem
to suggest that T3 Sandhi does not entail a paradigmatic substitution of T3 by T2, but rather only
turns T3 into some allotone, say T3', in the sandhi context, and hence maintains the categorical
distinction between T2 and T3' (pp. 26-27).
1 In his 1999 manuscript, de Lacy uses DTE (D) and non-DTE (-D) instead of Head (HD) and non-Head (NONHD). The
notation given here is from his 2002 publication and has also been used by Yip (2000). Unlike the 1999 version, however, the
2002 hierarchies do not include the lowest-ranking constraints, *HD/H and *NONHD/L. He explains that the lack of constraints
against high-toned heads and low-toned non-heads ensures that these types are the least marked of all.
In his dissertation (2002b), de Lacy formalizes these constraints as stringent and freely rankable: *DFt£H
(*DFt{H,M,L}), *DFt£M (*DFt{M,L}), *DFt£L (*DFt{L}) and *-DFt≥L (*-DFt{L,M,H}), *-DFt≥M (*-DFt{M,H}), *-DFt≥H (*-DFt{H}).
For the purposes of this paper, the distinction between the stringent and the non-stringent version of the constraints is not crucial.
The non-stringent constraints relevant for my analysis in sections 2 and 3 are *HD/L and *NONHD/H. The corresponding
stringent constraints are *DFt£L (*DFt{L}), where M and H do not incur a violation, and *-DFt≥H (*-DFt{H}), where M and L do
not incur a violation.
4
Another discrepancy between Chen’s and my allotone systems is the environment in which the underlying
contour of T3 occurs. According to the empirical generalizations resulting from my own work with a
native-speaking informant2, the [214] contour of T3 is preserved under contrastive stress. In fact, even
when the informant was asked not to contrastively stress anything, the word that seemed to receive
neutral sentence stress also failed to undergo T3 sandhi, i.e. did not change to a T2. Since regular
sentence stress in Chinese is, as Chen puts it, “phonetically elusive,” I will not distinguish between
contrastive and neutral sentence stress. The finding that the full contour of T3 gets preserved under stress
contradicts the standard assumption that low tones are the least salient and thus incompatible with
prominent positions (de Lacy 1999) and has not been noted in the literature before. While I support the
consensus that an underlying T3 in focus position is realized with a high final rise, just like a T2, I
disagree on the disappearance of the low T3 dip. My informant prolonged and thereby exaggerated the
full T3 contour, including both the dip and the rise. Considering that [214] contains not only an L but
also an M and even an H segment, the stressability of T3 apparent in my data is not all that surprising. It
could be that a stressed T3 is generally transcribed as [35] because the exaggerated high rise immediately
gets associated with a T2. To sum up, the T3 allotone system I propose is given in (6).
(6) T3 [MLH] allotones:
(i) under stress3: MLH
(ii) before T3: M[L]H ‡ MH
(iii) elsewhere: ML[H] ‡ ML
As will be shown in section 3, the “under stress” allotone (6i), i.e. the allotone that resists toneme
deletion, can be given a classic positional faithfulness account, where MAX-s(T) protects the stressed
contour from toneme deletion. As for the other two allotones (6ii-iii), besides the slight pitch-level
difference in the representation of MH ([35] vs. [24]), my toneme deletion system makes the same
predictions as traditionally assumed. Now, however, both the prototypical T3 sandhi variant in (ii) and
the elsewhere (or half T3 sandhi) variant in (iii) are neatly derivable by one simple move. Combining
both parts of Chen’s T3 sandhi (i.e. including the “half” rule), T3 sandhi as a whole can be formalized as
follows.
(7) T3 sandhi
MLH.MLH ‡ MH.ML
2 I am referring to a recoded set of data I collected for a previous project. My informant was a native Beijing Mandarin
speaking graduate student in the Linguistics Program at Boston University. In order to make sure my transcriptions of the tonal
patterns are accurate, I consulted Florence Woo, a native speaker of Cantonese, who is also fluent in Mandarin.
3 Since I do not want to deal with the optionally preserved H in utterance-final position, I am not listing this
environment here. As the end of an utterance often seems to coincide with sentence stress (see Chen 2000), the two
environments at least partially overlap.
5
The first T3 occurs before another T3 and is therefore realized as MH. The second T3 is in an elsewhere
environment and consequently surfaces as ML. Comparing the traditionally assumed with the newly
proposed T3 sandhi rule in (7), they both serve the purpose of dissimilation and contour reduction. The
rule in (7), however, has the advantage of allowing for a more elegant OT account. Since a language in
which tone is contrastive probably protects underlying tone categories with a high-ranking IDENT(tone)
constraint, a change from T3 [214] to T2 [35] is unlikely. In contrast, a change that deletes a tone
segment but preserves the categorical identity of the underlying tone merely requires a low-ranking
MAX(tone segment) constraint. As for the deletion of either L or H, section 3 will demonstrate how de
Lacy’s (1999, 2002a) tone prominence constraints make the right predictions. Yip’s (2000) T3 sandhi
analysis is based on a toneme insertion system that also allows for an efficient OT account. Her proposal
is the focus of the following subsection.
2.2 The H-insertion alternative
As mentioned in the introduction, Yip (2000) posits L as the underlying representation of T3. Her version
of the T3 sandhi rule was given in (4) and is repeated here as (8)
(8) (L.L) ‡ (MH.L)4
Referring to Chen’s (2000) work, Yip claims that T3 sandhi takes place within prosodic constituents that
are preferentially binary and also influenced by syntax. Within these binary constituents, the left syllable
must be the head because the creation of a rising tone, i.e. the insertion of an H segment, on the non-head
of the constituent would violate de Lacy’s *NONHD/H. The constraint ranking Yip needs to account for
her T3 sandhi representation is shown in (9).
(9) T3 sandhi as H-insertion
OCP >> DEP-H >> *NONHD/H, MAX/L
*FOCUS/L >> BIN
In order for H-insertion to be allowed, the OCP, which drives the dissimilation, must outrank DEP-H.
Since a positional faithfulness account disallows insertion or deletion involving elements in prominent,
i.e. head, positions, Yip argues against it, of course. As for contrastively stressed T3s, Yip, like Chen and
many other researchers (see e.g. Chang 1992), claims that a change to T2 occurs, even if the TS domain
boundaries have to be redrawn in such a way that the binarity requirement (BIN) of the constituents and
their alignment with the underlying syntactic structure are violated. This is illustrated in (10). The
4 Again, technically H-insertion leads to (LH.L), not (MH.L). Yip explains that there is no audible difference between
the two rising contours.
6
utterance in (a) shows canonical T3 sandhi application, whereas the same utterance in (b) contains a
contrastively stressed (underlined) element.
(10) a. only buy stocks
zhi mai gu-piao
L L L HM U.R.
(MH L) (L HM) Sandhi
b. zhi mai gu-piao
L L L HM U.R.
(L) (MH L) (HM) Sandhi
In order for H to get inserted on the contrastively stressed syllable, i.e. in order for mai to be in head
position, the domain boundaries get redrawn. This results in two monosyllabic constituents and the
break-up of the compound gu-piao. According to Yip, all this is caused by the undominated constraint
*FOCUS/L, which she considers a type of *HD/L constraint. Chen at least partly supports this view in that
he argues for an inviolable constraint he calls IP-BOUND which says that TS domains cannot cross
Intonation Phrase boundaries. Special intonation or focus thus overrides the principles according to
which TS domains are built5. This issue will be addressed again in subsection 2.3.
My empirical generalizations force me to argue against Yip’s H-insertion account. Since I maintain that
the full T3 contour surfaces under stress, an underlying L would require both H and M-insertion.
Assuming that the underlying form is MLH thus leads to a simpler, more economical solution.
Furthermore, my positional faithfulness approach does not need to redraw the TS domains in the case of
contrastive stress. If, as Chen claims, there is evidence for the existence of TS domains that is not based
on TS application, then it is desirable to avoid (as much as possible) redrawing the domains in order to fit
the tonal patterns. The crucial elements my account shares with Yip’s are left-headed TS domains and the
appeal to de Lacy’s tone prominence constraints. Since my representation of T3 sandhi involves the
deletion of L on the first T3, i.e. the head, and the deletion of H on the second T3, i.e. the non-head, I
need to appeal to both *Hd/L and *NonHd/H. As the tableaux in section 3 will demonstrate, the
interaction of the positional faithfulness constraint Max-s(T), the tone prominence constraints, and
general Max(T) determines the correct tonal outputs without domain changes. The tonal pattern my
account predicts for Yip’s data in (10) is given in (11).
(11) a. only buy stocks
zhi mai gu-piao
MLH MLH MLH HM U.R.
(MH ML) (ML HM) Sandhi
5 Unlike Yip, however, Chen argues for a high-ranking constraint (“No Straddling,” see subsection 2.3) that ensures the
unity of immediate syntactic constituents such as compounds.
7
b. zhi mai gu-piao
MLH MLH MLH HM U.R.
(MH MLH) (ML HM) Sandhi
The first constituent in (11a) shows the canonical T3 sandhi application, following the rule I give in (7).
The second constituent contains a T3 followed by a non-T3, here a T4 (realized as HL or HM). This is
one of the environments where the elsewhere allotone ML shows up. In (11b), the contrastively stressed
syllable is protected by MAX-s(T), while the first syllable behaves just like it would in the canonical T3
sandhi pattern; it loses its L tone segment. The resulting MH contour is in conflict with Yip’s
transcription of this syllable in (10). Which output is correct is an empirical question, of course6.
2.3 Sandhi domains
One of the biggest and as of yet unsolved mysteries of the T3 sandhi phenomenon is the nature of the
domains within which it applies. Although Yip (2000) states that these domains are preferably binary,
based on syntax, and left-headed, she does not explain how they come to be left-headed. She does not
provide evidence for the association of stress with the heads of these constituents. What does a head
represent when there is no stress involved. The answer could be a kind of mark for a rhythmic pattern
that aids in the pronunciation of long strings of syllables. Yip also does not address the question of
headedness in domains that are bigger than binary. There are many attested utterances consisting of
strings of T3s, where all but the last T3 are realized as the rising allotone. Do domains like that have
multiple heads? In order for Yip’s and my analyses to work, the answer must be yes. De Lacy’s (1999,
2002a) tone-prominence constraints require that every T3 which undergoes H insertion and every T3
which undergoes L deletion be a head. Chen (2000), whose account is couched in a derivational version
of OT, and for whom the existence of heads is not crucial, does not have this problem. He does, however,
have to deal with finding evidence for the existence of TS domains independent of T3 sandhi application.
He maintains that Mandarin TS domains, since they can be both sub and supralexical, are neither stress-
feet nor phonological phrases. In fact, they do not fit into the conventional prosodic hierarchy at all.
Chen thus proposes what he calls the Minimal Rhythmic Unit (MRU) as the relevant domain. He states:
Clearly, the MRU stands apart from the conventional prosodic hierarchy; it is basically a device
to group syllables of a wide variety of grammatical ranks and status into rhythmic units, as
determined largely by constituency or tree configuration. The MRU constitutes a prosodic unit
sui generis; it is off-scale, hors-séries, so to speak. For this reason, I have settled on MRU as a
neutral term vis-à-vis the conventional prosodic hierarchy (pp. 377-378).
6 According to my Cantonese-speaking informant Florence Woo, it seems more natural to pronounce a slightly rising
than a low falling tone before the fully realized T3 contour. This is an argument for the tonal pattern my analysis, rather than
Yip’s analysis, predicts. More fieldwork is needed to confirm the validity of this argument.
8
As for the way MRUs are built, Chen adopts Shih’s (1986) Foot Formation Rules (FFR), as listed in (12),
and encodes them in the constraints given in (13).
(12) Foot Formation Rules:
a. Immediate Constituency: Link immediate constituents into disyllabic feet.
b. Duple Meter: Scanning from left to right, string together unpaired syllables into
binary feet, unless they branch to the opposite direction.
c. Super-foot construction: Join any leftover monosyllable to a neighboring
binary foot according to the direction o