‘usually at this moment’
In this error the glide [w] is perseverated from the onset of the first word and added into the coda slot after the vowel in the target word [ts51]; since [] can only occur in open syllables, the vowel is realized as the correct variant for this context [ts__w], which is [o].
Besides the substitution and addition errors, there are 7 omission errors in our corpus in which a consonant before or after the vowel is omitted, causing the vowel to shift to a different alternant. These errors include 4 instances of the [ε]-[ə] alternation, as illustrated in (26) above, where the omission of the prenuclear [j] causes [ε] to revert to [ə]. Other alterna-tions produced by omission include one each of [ε]-[], [e]-[], and [ɔ]-[];
an example is given in (35).
(35) I: xej55-t hjɑw35 phaj35 →
Heiqiao brand
E: x 55-t hjɑw35 phaj35
‘Heiqiao brand (a brand of sausages)’
In this error, the glide [j] was deleted from the coda location in the target word [xej55] in dissimilation from the other palatal glides in the utter-ance. This produced the sequence *[xe], which is illegal since [e] only occurs before the palatal glide. Thus the vowel shifted to the correct variant [], which occurs in the context [x__$].
In addition to the fairly straightforward errors discussed above, there were 5 errors which involved the metathesis of a vowel and glide, causing a different vowel to surface. This included 3 cases of the [ε]-[e] alterna-tion, as illustrated in (36), and two cases of the [o]-[ɔ] alternation.
(36) I: mej35 li21-mɑw51 not polite E: m 35 li21 mɑw51
‘not polite’
In this error the sequence of underlying vowels in the target word is /(mid vowel) i/. In the target utterance, the /i/ is realized as the postnuclear glide
jε
[j], and the mid vowel is realized as the correct variant before the palatal glide, [e]. In the error word, the two vowels have metathesized, causing the sequence /i (mid vowel)/; in this case the /i/ is realized as the prenu-clear glide [j], and the vowel takes on the correct allophone for the sequence [j__$], which is [ε]. Note that this is a case where Lin would agree that the [ε] is derived from the mid vowel, since it is in an open syllable. This example is exactly analogous to the 2 cases where /(mid vowel) u/ [ow]
metathesizes to produce [wɔ].
Finally, there were 4 other errors which involved combinations of addition, substitution, omission, and metathesis of consonants, and which caused mid vowels to shift to different realizations. This included one each of the following four alternations: [ɔ]-[ə], [ɔ]-[e], [o]-[e], and [o]-[ɔ]. Two examples are given below.
(37) a. I: swɔ35-i21 → so
E: sej35
‘so’
b. I: tsaj51 thu35-su55-kwan21 tən21 →
in library wait
E: tsaj51 thu35-su55-kwan21 t 21
‘waiting in the library’
The error in Example (37a) is called a ‘telescoping’ error, since elements from sequential syllables are omitted and the syllables are collapsed into one. In this case the [w] in the target word [swɔ] is omitted, and the [i] from the second syllable is resyllabified with the vowel [ɔ] of the first syllable.
The vowel [i] now becomes a postnuclear glide [j], but the sequence *[sɔj]
is illegal, so the [ɔ] shifts to the correct allophone for this sequence, which is [e]. In (37b), it appears that the [w] from the source word [kwan] has been perseverated and added into the prenuclear glide slot of the target word [tən]; furthermore the [n] in the target word has been omitted in dissimi-lation from the [n] in the source word. This leaves the sequence *[twə], and thus the [ə] is realized as the correct allophone for the sequence [w__$], which is [ɔ].
A glance at Table 5 shows that every one of the six mid-vowel phones alternates with every other one of these phones in at least one error, and that overall there are 60 errors showing these alternations. This is very strong evidence that these 6 phones should be grouped into a single phoneme.
Moving on to Question 3 above, the status of [ε], these data make the answer wɔ
very clear. In these mid vowel errors there are 20 cases where the vowel [ε] alternates with other mid vowel phones, and [ε] enters into alterna-tions with all 5 of the other phones. In 15 of these cases, the [ε] occurred in the environment [j__n], which is the environment in which Lin argues the [ε] is derived from /a/. In this way [ε] behaves like any other mid vowel phone, and shows no behavior which would set it apart from the other phones. Although there was one example discussed in Section 4.3 above where [ε] alternated with [ɑ], this is not convincing evidence that [ε] should be considered a member of the /a/ phoneme, since there is much more evidence for its association with the other mid vowels. We conclude that in the dialect of Mandarin under study here, Taiwan Mandarin, the phone [ε] is a member of the mid vowel phoneme.
The grouping of these six allophones into a single phoneme allows us to account for some data which is otherwise difficult to explain. In our data we have 14 errors which we classified as ‘non-contextual’, that is, where there was no obvious source for the error in the utterance. Non-contextual errors are typically quite rare, since most phonological errors are caused by the substitution or addition of some other segment in the planned string, the exchange or metathesis of two planned segments, or omission in dissimilation from another segment in the string. However, if we adopt the hypothesis that these six phones derive from the same underlying phono-logical category, then 10 of these non-contextual errors involving mid vowels become contextual, since the source could be said to be a different allophone of this phoneme which does occur in the utterance. Two examples are given in (38).
(38) a. I: khaj55 w 21-t wan35- jɑw51 →
open my joke
E: khaj55 w 21-t w n35- jɑw51
‘teasing me’
b. I: i51 ti55 jεn21 lej51 → one drop eye tear E: i51 ti55 jεn21 li51
‘a tear drop’
In (38a), the error looks at first pass like a non-contextual substitution of [ə] for [a]. However, if we consider the source to be one or both of the other mid vowel phones in the preceding word, then this would be a contextual perseveration/substitution of either [ɔ] or [] for the [a] in [wan]; since
ɔ
ɔ ə
the only allophone of the mid vowel phoneme allowed in the environment [w__n] is [ə], the vowel is realized as this variant. Similarly, in (38b), the error appears to be the non-contextual omission of the [e] in [lej], which causes the glide [j] to be realized as its vocalic variant [i]. However, if we assume that the [e] has been deleted in dissimilation from the preceding mid vowel [ε], then this becomes a contextual error. Thus our analysis helps account for a large number of the otherwise non-contextual errors in our data in a principled way.
The only remaining question is whether or not we can determine a basic variant of the mid vowel phoneme. Using the logic we applied to the low vowels, the central variant should be [] since it is the only of these phones that can occur alone as a full syllable and in open syllables with no pre-nuclear glide conditioning factors. However, this analysis would produce a somewhat asymmetrical phoneme inventory: /i, y, u, , a/ (see Maddieson (1984: Chapter 8)). Lin (1989) posits [ə] as the basic variant because it is the phonetic center of the category, and both R. Cheng (1966) and Wu (1994) also select [ə] as the basic variant. Furthermore, of all the mid vowels, [ə] occurs following the broadest range of onset consonants, although it must always be in a syllable closed by [n]. Choosing [ə] as the basic variant of the phoneme provides us with the more symmetrical phoneme system /i, y, u, ə, a/; furthermore /ə/ as a phoneme is far more common in the world’s languages than // (Maddieson (1984: 252, 257)).
For all these reasons we agree that this phoneme category is best charac-terized as underlyingly /ə/.
5. SUMMARY AND DISCUSSION
5.1. The Vowel Phoneme System of Taiwan Mandarin
In this paper we examined four hypotheses regarding the underlying vowel categories of Mandarin, and then looked at phonetic, distributional, and especially psycholinguistic evidence regarding the system of TM. We found that none of the previous theories could completely account for the findings in our study, although we concur with R. Cheng (1966) and Lin (1989) that the system consists of five underlying vowel categories. Our first dif-ference from all four accounts is that in this dialect of Mandarin, there are not two differing apical vowels [ι, ι]; in TM there is only a high central unrounded vowel [], and there is no retroflex vowel. We provided evidence that this vowel [] is a conditioned variant of /i/, which occurs only after
coronal fricatives and affricates: specifically, [i] and [] cannot substitute for each other in errors, and these two phones alternate with each other systematically in errors involving a change in the contiguous environment.
We agree with the analysis of R. Cheng and Wu who group the apical phones with the /i/ phoneme; we disagree with C. Cheng who considers the apical phones to constitute a phoneme in their own right, and with Lin who derives the apical phones by rule. Second, we find no evidence that [y] is derived from a sequence of /ui/, as proposed by Wu, since [y] can freely substi-tute for both [u] and [i] in vowel substitution errors. Third, we find a great deal of evidence that [ε] is a variant of the mid vowel phoneme only, and not an allophone of the low vowel phoneme. The phone [ε] alternates with all of the other 5 mid vowel phones when the contiguous environ-ment changes in 20 errors, whereas there is only one error that suggests an affiliation between [ε] and the low vowel phoneme. Thus our analysis disagrees with all four theories on the status of this phone. We agree with all the theories that [a, ɑ] are variants of the same phoneme, and concur with all but C. Cheng that [a] is the basic variant. Finally, we presented evidence supporting the hypothesis that glides are derived from under-lying high vowel phonemes. Thus the vowel system which we propose is the following:
(39) The vowel system of TM (Wan and Jaeger) /i/ → [i, , j]
/y/ → [y, ɥ]
/u/ → [u, w]
/ə/ → [e, ε, ə, ɔ, o, ]
/a/ → [a, ɑ]
If indeed this is the correct system for TM, then we should be able to show that only these five vowels (including all their respective allophones) can substitute for each other in errors, and that no two phones which we have classified into the same phoneme can substitute for each other. In Table 4 above we presented the number of errors from our data in which each vowel phone substituted for the other phones. In (40) below we have collapsed these numbers into the vowel phoneme categories which we are arguing to be the underlying system of TM; this figure shows the number of times any of the phones from each phoneme substituted for any of the phones of any other phoneme. These figures account for all of our data (N = 56, including the vowel/glide substitutions); this table shows that there were no cases where two phones which we consider to belong to the same phoneme substituted for each other.
In Sections 3 and 4.2 above we gave several examples of the high vowels substituting for each other (Examples (11), (14), (16), (17), (18), (19)).
Below are examples of substitutions across other phonemes; in (41a) the mid vowel phoneme /ə/ [] is anticipated and substituted for the high vowel phoneme /u/ [u]; in (41b) the high vowel phoneme /i/ [i] is anticipated and substituted for the low vowel phoneme /a/ [a]; and in (41c) the low vowel phoneme /a/ [ɑ] is perseverated and substituted for the mid vowel phoneme /ə/ [ə].
(41) a. I: u35-kwɔ21 t han21-t hu55 wən51-thi35 t xwa51 →
if cause problem
E: u35-kwɔ21 t han21-t h 55 wən51-thi35 t xwa51
‘If (the product) caused a problem . . .’
b. I: tha55 li35-kaj55 xɑw35-t jow21 l → he leave very-long
E: thi55 li35-kaj55 xɑw35-t jow21 l
‘He has been gone for a long time’
c. I: wɔ21 k ŋ55-k ŋ55 kən55 ej35 wɔ55 →
I just with who say
E: wɔ21 k ŋ55-k ŋ55 kan55 ej35 wɔ55
‘Who did I speak with?’
Thus all of our vowel substitution errors support this proposed 5-vowel system.
ɑ ɑ
ɑ ɑ
(40)
y u ə a
i 19 12 2 01
y 07
u 2 01
ə 12
5.2. A Psycholinguistic Model of Phonological Representation and Processing
We would like to end this paper by exploring a possible psycholinguistic account of the representation of vowels in lexical representations in Taiwan Mandarin, and the processes by which they arrive at their correct surface forms in speech production planning. In order to do so we must first present a partial speech production planning model, which illustrates how the phono-logical representations of words and their surface manifestations are related in processing. This model has been developed based partly on our research, and partly on the previous models of Fromkin (1973) Garrett (1988), and Levelt (1989).
In this model, the circled categories in the left-hand column indicate the representation of information in long-term memory. This includes the
‘Lemma Lexicon’, which represents information about the semantic and syntactic properties of lexical items, and the ‘Form Lexicon’, which contains underlying phonological representations (as well as information regarding morphological and orthographic structure of lexical items). The ‘Syntax’
component includes morphosyntactic rules or algorithms, and syntactic tem-plates, while the ‘Phonology’ component includes knowledge about the phonological units, sequences, and alternations allowed in the language. The right-hand column includes processing components (in rectangular boxes), which represent actual work being done in short-term memory while a particular utterance is being compiled for speaking, and output/input representations (in italics) which are essentially on-line representations of the output from one processing component, which is then input into the next component for further processing. The final representation in this model, the ‘Phonetic Level Representation’ is the representation which will be formatted by stored motor programs for actual speaking.
The portions of this model which are relevant to the present discussion are the representations in the ‘Form Lexicon’ and ‘Phonology’, as well as the ‘Phonological Processes’ component. After lexical items are selected and inserted into the syntactic string, their phonological forms are activated;
the result of these processes is a string with all the abstract phonological representations for each word or morpheme concatenated into the correct order, i.e., the ‘Positional Level Representation’. This representation is input into the ‘Phonological Processes’ component. This component acts on the abstract phonological input in various stages, in levels similar to those posited in non-linear (e.g., autosegmental) phonological theories (Wan and Jaeger 1998). However, our error data show that the first thing to occur is any kind of error in which an abstract segment (or segmental node) is
misplaced, added, or omitted in the abstract string. This is because the remaining phonological processes, i.e., both context-dependent rules and feature fill-in redundancy rules which are stored in the ‘Phonology’ com-ponent in long term memory (see below), all operate on the erroneous string as if it were the intended linear string of phonemes. Thus the phonolog-ical rules of the language ‘clean up’ any messes made by erroneous placement of segments, by causing the phonemes to be realized in their correct surface forms given the order in which they occur after the errors are made. Let us now put some specific phonological content to this general outline.
Figure 1. A speech production planning model.
5.3. The Featural Structure of the Vowel System of Mandarin We assume that the phonological information in lexical representations is stored in terms of some sort of abstract or underspecified phonological units, similar to the analysis of Lin (1989: 58). In Lin’s model, the representa-tion of the vowels of Mandarin can be accounted for in terms of the underspecified feature grid shown in (42).
(42) i y u ə a
High + + + –
Low +
Back – – +
Round – + +
Lin also posits the following set of redundancy rules for assignment of the surface phonetic forms for each phoneme:
(43) [ ] → [–high] [ ] → [–round]
[ ] → [+back] [ ] → [–low]
Given our somewhat different analysis of the vowel system of TM, we will need to make some changes in this underlying feature specification in order to account for our findings. First, Lin has specified /i/ as [–back]
underlyingly. This was done in her system because the default value for the feature [back] supplied by the redundancy rules is [+back], so the [–back] feature needs to be specified in the underlying representation.8 However, in our system this would cause us to have a context-dependent rule which would change the feature [–back] to [+back] after the coronal fricative and affricates, in order to derive the surface form []. (Note that this designation of the central vowel [] as [+back] does not mean that it is literally a back vowel; we are using the usual convention of marking central and back vowels [+back] as opposed to front vowels which are [–back].) Thus we propose that in TM, the phoneme /i/ is unspecified for the feature [back]; the [–] value for this feature will be filled in by a context-dependent rule to designate [i], and the [+] feature will be filled in by the redundancy rules to designate []. The underlying representation of this phoneme will reflect the fact that the phoneme can be realized as both a [+back] and [–back] variant, which makes the representation more in line with that of /ə/, which is also unspecified for [back] and can have both
[+back] and [–back] surface alternates. Furthermore, a designation on the feature [back] is not necessary in order to distinguish /i/ from all the other vowel phonemes in underlying form, since it is the only category designated as [+high, –round]. Finally, as we will show below, this will allow us to have context sensitive rules which only add feature values, not change feature values.
Second, following the same logic as above, we find that the phoneme /u/ does not need to be specified for the feature [back] (see also note 8).
This phoneme is differentiated from all other phonemes without this spec-ification, and the value [+back] will be added by redundancy rules.
Third, our designation of [ε] as being part of the mid vowel phoneme solves a problem in Lin’s system. Since in her system [ε] is an allophone of /a/ as well as /ə/, she needed to have some provision for changing the [+low] specification to [–low] for those surface variants which derive from /a/. In our model all instances of [ε] derive from the mid vowel phoneme, and are thus designated only as [–high] in underlying form, and therefore no feature designations need to be changed to derive the surface form.
With these three differences in mind, then, we would argue that the vowels of TM can be accounted for in terms of the following underspecified feature grid.9
(44) i y u ə a
High + + + –
Low +
Back –
Round – + +
Besides specifying each phoneme so that it is distinct from every other phoneme underlyingly, this underspecified feature grid also captures the fact that the /y/ phoneme is the most marked phoneme in TM, since it has three features specified underlyingly whereas all the other phonemes have only one or two. It is also true that /y/ is a “marked” vowel universally, in the sense that it is much less common in the world’s languages than /i, u/, and has the marked combination of [–back, +round] which is disfa-vored universally (Maddieson (1984: Chapter 8)). In line with it being a marked vowel universally, it is the least frequent of the vowel phonemes in TM, and occurs in the most restricted set of environments, as shown in Table 2 above.
In order to predict the correct surface forms from these various under-specified input forms, we need to have both context-dependent rules and context-independent redundancy rules, as discussed above. These rules will be part of the ‘Phonology’ in the processing model, that is, the set of rules (or constraints) stored in long term memory which specify what a possible and legal output can be in this language. In (43) above we pre-sented Lin’s four redundancy rules, and we agree that these are the correct rules for our new analysis. However, we find that in order to account for all the surface forms of vowels, one further feature is needed. In Wan (1999), the feature [tense] was explored as the possible necessary fifth feature.
However, because this feature has little consistent phonetic content (Jaeger (1983), Ladefoged (2001: 80)), it was not found to be adequately explana-tory. But a look at the pairs of vowels which need to be distinguished from each other by some feature other than the four listed in Lin’s model gives a strong clue as to what this feature actually is phonetically: [e-ε], [ə-], [o-ɔ] and [a-ɑ]. These pairs each represent two vowels with the same basic backness, rounding, and height, but in every case the second of the two vowels is somewhat retracted from the first; in some cases the tongue is pulled directly back, but in others it is pulled back and downward.
However, because this feature has little consistent phonetic content (Jaeger (1983), Ladefoged (2001: 80)), it was not found to be adequately explana-tory. But a look at the pairs of vowels which need to be distinguished from each other by some feature other than the four listed in Lin’s model gives a strong clue as to what this feature actually is phonetically: [e-ε], [ə-], [o-ɔ] and [a-ɑ]. These pairs each represent two vowels with the same basic backness, rounding, and height, but in every case the second of the two vowels is somewhat retracted from the first; in some cases the tongue is pulled directly back, but in others it is pulled back and downward.