國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
9
adaptation patterns. It may be questioned that the selection of certain TM characters in these categories involves semantic factors and thus do not serve as valid tokens.
Nonetheless, they are still counted in generalizing the patterns in this research since they still bear a phonetic similarity to the English source words to a certain degree, indicating that the phonetically similar parts have undergone certain phonological processes. Furthermore, the academic value of loanwords with semantic correlation still remains particularly for analyses that lay the focus on the retention/deletion of consonants only, without making reference to the featural change of the target segments.
1.3 The focal points and a bipartite processing model
Authors advocating the need for an independent loanword grammar from the native phonology have done analyses of loanword adaptation patterns by postulating fairly intricate rankings, oftentimes accompanied by loanword-specific constraints, within the production grammar which can neither be derived from the native language nor be motivated by the foreign language. One question that naturally arises, however, is that if these rankings for loanword adaptation are something that cannot be learned due to the lack of L1 stimuli that are structurally similar to L2, it becomes a puzzle why the adaptation patterns from the collected data have been systematic and shown consistency across L1 speakers, particularly in determining the retention/deletion of a consonant. On the other hand, it will be discussed in more detail in the upcoming chapter that any biased stance towards a perception/production account may find it deficient in dealing with loanwords from L2 into L1. In response to the problems, it might well be argued that the retention/deletion patterns of L2 consonants in L1 that appear to involve unlearnable rankings are mostly, if not all, an effect of perception grammar. That is, the retention/deletion and segmental change of an L2 consonant in
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
10
the L1 adaptation output are determined as early as in the level of perception, rather than a result solely from production grammar.
The proposal of a perception-production model dates back to Silverman (1992), and the plausibility of the proposed analyses are reinforced by subsequent papers with the advent of OT. A representative work along this line is Kenstowicz (2003b), where he sketches two distinct rankings for perception and production with the same constraint set, successfully explaining Gbéto’s (1999) study of loanwords from French, Portuguese, and English to Fon without the need to propose loanword-specific constraints (e.g. Davidson and Noyer 1997, Yip 2006). In his argument, Dep-V outranks Max-C in the perception mapping, and thereby the French input [.post.]
(poste) is interpreted as /.pos./, which serves as the input of the production mapping.
The input /.pos./ is then subject to the production grammar with the reversed dominance Max-C >> Dep-V, and the output [.po.su.] is derived under this ranking.
Nonetheless, despite the seemingly persuasive resolution without redundantly positing any loanword-specific constraint, one may still question this treatment by raising the questions below. First, the markedness constraint *stop/obstruent_# may successfully rule out the candidate with an sC cluster, [.post.], but it fails to eliminate the other alternative, [.pot.], as the perceptual output, which is not considered to be a potential winner in his tableau of perception mapping. Second, if it is true that phonotactics functions in both perception and production, as first put forth by Silverman (1992) and lent support to by a handful of others in this vein, there seems to be no reason why the perceptual output is solely governed by *stop/obstruent_# in Fon, a language that allows only open syllables, not by *obstruent/_#, which forbids any obstruent consonant in coda? That is, if *obstruent/_# is considered in his grammar, the only winner from the perception level would be [.po.], rather than
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
11
[.pos.]. Last but not least, a simple inquiry from a functional point of view is what motivates the distinct rankings between Dep-V and Max-C in perception and production. Specifically, in the perceptual phase, what is the point in saying that one’s unwillingness to insert a vowel outweighs his willingness to preserve a consonant, if auditory perception involves sensory information of which the retention/deletion cannot be determined by logical reasoning?
Unlike papers that fiddle with faithfulness constraints in both levels of perception and production, this dissertation employs Boersma and others’ cue constraints (CUE, Boersma 1997, 1998, 2000, 2006, 2007ab, 2008; Escudero 2005; Boersma and Hamann 2009) to interact with structural constraints (STRUCT), i.e. constraints regulating L1 phonotactics. Cue constraints are a constraint type that should be defined independently from markedness and faithfulness constraints, and act as the main ingredients in the perception grammar. The cue constraints formulate the listener’s knowledge of perceptual cues, i.e. the relation between the auditory forms and their correspondent surface form in phonology, on the basis of either the auditory information provided by the external context, or the internal acoustic cues of the involved segment. As perception, least controversially, involves the interface between phonetics and phonology, i.e. a process with direct competition between perceptual cues (cue constraints) and language-specific phonology (L1 structural constraints), this dissertation intends to build an OT-architecture for perception that is devoid of faithfulness constraints. For example, conventional OT ranks Max-C over Dep-V to ensure retention of a consonant through an epenthetic vowel, whereas the same effect may be generalized by ranking *[C]/ / over *[ ]/+vocalic/8 in the present dissertation, saying that the degree one perceives the existence of a consonant induces his
8 The cue constraint that functions the same in Boersma’s related works is *[ ]/–cons/. We revise it so as to exclude glides, a semivowel that also has the feature value [–cons].
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
12
hallucination of a non-existent vowel. With cue constraints applied to a language-specific perception grammar, it gives more direct explanations for the perceptual processing of the raw auditory information and events.
Another idea at the core of this dissertation is the modeling of a perceptual constraint grammar that well accounts for the phenomenon of lexical variation (Zuraw 2010), as observed in our loanword corpus. Given perception is the mapping from raw sensory data to more abstract mental representations (Boersma 2007), it should be variable by nature, since a variety of factors may influence the perceptual interpretation of a given input, such as background noise and individual differences in perception and production. Rather than free variation, a more widespread variation in loanword adaptation is lexical variation. Lexical variation in loanwords refers to the situation where an element, e.g. a segment, in an L2 input may undergo a certain phonological process (e.g. deletion) to conform to the L1 phonotactics, while the identical element in another L2 input in the same or similar phonetic context may undergo a different phonological process (e.g. vowel insertion). For example, the [ ] in English Norman ([.no .m n.]) is deleted as 諾曼 ([.nwo.man.]9) in TM, while that in Hormone ([.ho .mon.]), with a similar context, is retained via schwa insertion in TM as 荷爾蒙 ([.x . .mo .]). In free variation, on the other hand, a single foreign word has two or more adapted forms in the recipient language. For example, the L2 name Truman ([.t u.m n.]) is adapted to 杜魯門 ([.tu.lu.m n.]) to refer to an American president, but meanwhile it is interpreted as 楚門 ([.t u.m n.]) in the TM transliteration of the Hollywood film Truman Show. In this dissertation, we are concerned particularly about the lexical variation patterns in the retention/deletion and segmental change in voicing and aspiration of English consonants in their TM
9 English stresses and Mandarin tones are beyond the research scope of this dissertation, hence ignored in the phonetic transcriptions throughout the dissertation to save space and meanwhile avoid distraction.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
13
nativized form, since, as will be discussed in more detail later, the alternation in other features and tones in a large number of adaptation forms are not purely phonological but results greatly from semantic consideration of the TM characters.
Either deemed to be a minor issue or simply left aside in prior literature on loanword adaptation, lexical variation in effect reflects the nature of uncertainty in perceptual interpretation due to background noise and/or individual differences in speech perception and production. In this current research, therefore, it is argued that the mentioned adaptation patterns of English consonants in TM adaptation are something that is determined early in perception due to the lack of absoluteness in this level. As the observed lexical variations in the adaptation patterns reveal a strong correlation with the perceptual salience of the target consonant, we are in need of an appropriate framework to model the inevitable variation patterns that take place in the perceptual stage. Among the constraint-based theories that are developed to formulate language variations, Boersma’s (1997, 1998) and Boersma and Hayes’ (2001) stochastic evaluation is privileged in two facets. First, it averts multiple grammars and holds a single ranking parsimoniously by seeing constraints as values on a linear scale of strictness. Whether and to what extent variation may occur rely on the closeness of two constraints, the dominance between which is crucial in determining the output form. Under this hypothesis, positing multiple grammars to account for variations is rendered unnecessary. Second, in combination with cue constraints in perception, the
“quantification” of constraints well indicates the “weight” of cue constraints that functionally reflect perceptual salience. Moreover, the measure of the overlapping area induced by the closeness of two constraints succeeds in embodying the probabilistic distribution of two variable forms, e.g. how likely an L2 consonant is retained or deleted in its L1 realization. The application of stochastic evaluation to the OT grammar in perception is termed stochastic perception grammar (SPG)
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
14
throughout this dissertation.
What seems to be a minor concern in this dissertation is the content of the production grammar. In a sequential perception-production model, the output (surface form) of perception grammar becomes the underlying representation (UR) that is stored in the adapter’s short-term memory, which in turn serves as the input to the production grammar. The UR is the winner from the evaluation of the perception grammar. Based on the elaboration just given on the perception level, the “survival”
or “sacrifice” of an L2 consonant has been settled in the UR. Compared with the perception grammar, the constituents of the L1 production grammar are considered to be more complex. Presumably, at least three types of constraints/factors in addition to the native structural constraints should be in play. Like other languages studied in literature, first, faithfulness constraints require that the output form be identical to the input (UR). Secondly, a series of articulatory constraints that make reference to the speaker’s preferences in both manner and place of articulation exerts a certain influence on the evaluation process.
Finally, different from other alphabetic languages, what remains to be done in production is the selection of TM characters in the surface form. Chinese is a logographic system and nearly each character (a syllable) acts as a minimal unit in meaning (a morpheme). The selection of TM characters to realize a UR of English origin thus depends largely on the category in semantics where it belongs. Let us provide a couple of examples for this. The English given name Spike ([.spa k.]) is adapted as 史派克 ([. .p ai.k .]) in TM, where the first character 史 ([. !.]) is also a renowned family name in China. In the sequential convention of an English full name, the given name goes first, and then the family name, while the opposite is true in TM. Not surprisingly, in TM adaptations of English person’s names, the first character is usually a TM family name so as to correspond to the English source
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
15
semantically. By analogy, when the English source is not a person’s name, the [s] in a similar context is mostly mapped to TM 斯 ([.s!.]), which is phonetically closer to its English origin but semantically less relevant, if any, to people’s names10. For example, the English word Sparta ([.sp .t .]), a city-state in ancient Greece, is adapted to TM 斯巴達 ([.s .pa.ta]). In another example, the English word hacker ([.hæ.k .]) is adapted to 駭客 ([.xai.k .], ‘frightening guest’) in TM to reserve as much as possible the word meaning of the source. However, the name of a character in the Harry Potter series Hagrid ([.hæ.( d.]) is transliterated into 海 格 ([.xai.k .], literally ‘ocean style’), where the chosen characters are irrelevant to the source word in meaning.
In light of the foregoing discussion, we sketch a bipartite model where an English source word is processed into TM, as shown in (8). Conceptually, this model bears a fundamental resemblance to the widely recognized two-phase mechanism that traces back to Silverman (1992). What majorly makes the bipartite model distinct from those in relevant studies, however, is the constraints that make up the perception and production grammars.
10 TM 斯 ([.s!.]) is a Chinese surname, but extremely rare.
‧
(8) A bipartite processing model for English loanwords in TM adaptation
Perception Production
/underlying representation/ /underlying representation/
perception grammar (SPG) production grammar
[auditory form] [articulatory form]
L2 lexicon L1 lexicon
To elaborate on the operation of this model, let us give an illustration of how a TM adapter interprets the English person’s name Spencer. The first adapter hears the English auditory form [.sp n.s .]. And this raw sensory data serves as the input to the perception grammar. Through the stochastic evaluation of the cue constraints and TM-specific structural constraints in a particular ranking (SPG), the output is very likely to be /.s!.pin.s ./11, which is stored temporarily as the abstract mental representation in the short-term memory, namely the underlying representation (UR).
The UR subsequently acts as the input to the production grammar and undergoes the evaluation of faithfulness and articulatory constraints, under the government of TM structural constraints. They function to preserve as much as possible the input information in the output and meanwhile ensure the phonological legality of it. What
11 We are unable to determine on the precise phonetic transcription of the UR, since the alternations in feature rely greatly on the listening individuals. For example, the second syllable in the output can be /.pin./, /.pan./, /.pjen./, etc. However, it is held in this dissertation that the retention/deletion of a segment from the auditory input, such as retention of the first onset [s] and deletion of the postvocalic [ ], is comparatively more fixed in the UR and consistent across all L1 adapters, due to the universal perceptual cues to the noise, pitch, silence, transition and duration of the segments and the single phonological system the L1 adapters share.
STRUCT
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
17
plays a major role here is the consideration of semantic factors—as mentioned above, English Spencer is a person’s name, and the adapter chooses the character 史 ([. !.]) as the initial character in place of 斯 or 思 (both are pronounced as [.s!.]), despite the less phonetic closeness of the former. The articulated form is thus 史賓塞 ([. !.pin.s .]). Over time, when it is extensively recognized by the TM-speaking community after it is frequently used in print media or addressed in public, it enters the TM vocabulary. TM speakers thereafter use the nativized form from their mental lexicon, and no adaptation for this word happens ever after.
Of particular interest here is the construction of SPG that governs TM perception of English consonants, which is assumed to be the phase that crucially determines the retention/deletion and segmental change of an L2 segment. What this dissertation features, in comparison with other computer-based works in stochastic-OT, however, is that we figure out the mathematical axioms with which we are able to work out the precise ranking values of the constraints yielding binary variation (two variants), in accordance with the generalized patterned distribution. Within this a theoretical breakthrough is that we take a further step and explicate the mathematical operations that logically fit multiple variation (more than three variants) into the fashion of OT. It turns out to be the case that stochastic OT is no more confined to binary variation, but applicable to phonological processes that involve multiple variable outputs.