CHAPTER 2 LITERATURE REVIEW
2.1. Models of Speech Production
2.1.4. Connectionist Model
國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
14
substituted.
On account of the various error types in speech, Stemberger supposed that the
“noise” among the levels always leads to diverse errors during the cascading and feedback activation, such as phonological, semantic, or inflectional types, especially when we want to find a processing route to generate the production of mixed errors with more simplicity.
2.1.4. Connectionist Model (Dell, 1986, 1992)
Dell’s (1986) lexical process model of speech production is known as a connectionist model. The different point from serial-ordering model is that connectionist model allows the activation weight sent back to the former nodes which have been processed. It is depicted as the “feedback” of activation among the levels interacting with each other. The general process of the connectionist model is depicted in figure 2-5 below as well as the diverse representation levels dealing with external signals in figure 2-6:
Figure 2-5. Lexical network in the spreading activation production model.
Like discrete stage models, non-discrete spreading activation models distinguish semantic-conceptual units, lemma units, and form units. The units in this model are
log dog cat rat mat
Semantic Features
Lexical Nodes
Phonological Segment Nodes
/d/ /k/ /r/ /m/ /o/ /a/ /g/ /t/
/l/
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
15
organized into a network in which the connections allow for a bidirectional spreading of activation between units at adjacent levels. For example, the lexical word unit can pass activation downward to phonological level and backward to semantic level. As supposed in Dell and O’Seaghdha’s work (1991), lexical access involves the following six steps:
(1) The semantic units receive external inputs.
(2) Activation spreads throughout the lexical network in a non-discrete fashion.
(3) The unit which receives the most activation is selected.
(4) When the word unit is ready for phonological encoding, it is given a triggering jolt of activation, whether it’s a multi-word utterance or a single-word utterance.
(5) Activation continues to spread in lexical network, and the appropriate phonological units become significantly activated.
(6) The most activated phonological units are selected and linked backward to the slots of the word-shape frame (morphological structure) to implement sounding on those words, and so as to the frame of syntactical level.
Dell argued that the interactive model lies in its ability to account for a range of speech errors, particularly the effects of the similarity of the target and error in utterances. For example, Fay & Cutler (1977) proposed the fact that the all substituting and replaced words are very often from the same grammatical class, and it has been regarded as evidence that these substitutions are lexical errors rather than segmental errors. However, Dell believed interactive perspective among linguistic levels provides better explanation for those form-related (phonological-related) errors, which also imply the influence from grammatical and semantic levels.
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
16
Phonological word substitution can be attributed to a target being replaced by a phonologically related word, not just segments exchange. For example, as in figure 2-5, the word cat is activated, and, in turn, /k/, /æ/, and /t/ are appropriate units to be activated. Meanwhile, these active phonological nodes pass activation back to the lexical level to activate the word nodes relevant to them, such as mat or rat. Once the mat gets the most activation, the pure phonological error will be induced (for the rhyme sharing in phonological units); once rat gets the most activation, mixed error (with both phonological and semantic relation) will be generated. Therefore, mixed errors could be attributed to an interactive influence from both semantics and phonology in lexical access, which could make the whole model explain mixed errors not as costly as other discrete models do.
Figure 2-6. A simplified network in which external signals are received in sequence
by nodes at four representational levels (Dell & O’Seaghdha, 1991)Figure 2-6 shows a simplified network consisting of a single column of semantic, word, phonological, and phonetic nodes, and the different representations can also be interactive to induce the influence on the accessing efficiency.
Adapted from Levelt’s serial model, there are also linguistic levels for nodes to process information from diverse sources, which might be from concept, word, sound
External Signals
Semantic Word Phonolog
-ical
Phonetic articulatory
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
17
levels or other linguistic layers. The difference is that Levelt’s model is unidirectional-oriented, while Dell’s is bi-directional interaction (stemming from the tradition of Stemberger, 1985). As to the main concern of mixed speech errors in this study, interactive processing model could give speech errors better explanation on many sound-meaning errors because the incorrect activation could occur in more than one layer during lexical process. In current study, we assumed that connectionist model might have better account for the lexical access and the speech errors in a wider view, but we need more empirical evidence from the experiments. On the other hand, if concepts competition (from literal and visual information) occurs in lexical process, whether the mixed errors would be induced more is mainly concerned in this study.
2.2 Linguistic Effect in Previous Studies
In order to make an objective judgment on the lexical access model that could well-account the speech errors, we have to know the internal structure among these errors, especially for the linguistic effects on the semantic-lexical level, lexical-phonological level, and phonological-phonetic level between the target words and errors. In this section, some linguistic effects will be reviewed, such as initialness effect, phonological similarity effect, rhyme effect, syllable structure effect, tone effect, phonotactic regularity effect, and Stroop effect (originally used as psycholinguistic term).
2.2.1 Initialness Effect
Most syllables have an initial (onset). Some languages restrict onsets to be only a single consonant, while others allow multi-consonant onsets according to various language specific rules. Chinese is the language allows both null and occupied initial in syllable. However, the status of prenucleus glide, attributed to initial, rime or
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
18
independence, still keeps undetermined in Chinese. One thing is for sure, initial consonants are more likely to slip than non-initial ones (Dell & Juliano & Govindjee 1993). Take an initial metathesis error as example:
(1) tonal phonology fonal phonology
The initial [t] was substituted by the initial [f] of the following syllable in an anticipatory manner. Initials tend to slip more than other phonological structures, which can also be seen in the study found in Boomer & Laver (1968), Mackay (1970), Dell & Reich (1981), Dell (1986), Shattuck-Hufnagel (1986), Stemberger & Treiman (1986), Stemberger (1989), and Dell & Juliano & Govindjee (1993).
Dell (1986) indicated that the initial consonants may activate stronger connection weight from higher levels and thus tend to induce slips. Dell et al. (1993) noticed that there was approximately 80% of consonant movement error involving syllable onset, such as the error in (1), while there was still 40-50% for non-movement errors. The further point they proposed as typical explanation for initialness effect has been that initial consonants of words are structurally salient and distinct in the phonological frame. The frame corresponds to Shattuck-Hufnagel’s (1987) word onset and MacKay’s (1972) syllable onset, which are more detachable than the other structures of a syllable. This effect are assumed by Dell et al (1993) to reveal the manner of phonological structure as they are used in the retrieving the sounds of words.
2.2.2 Rhyme Effect
Rhyme effect intends to examine if interacting units of lexical errors share their rhymes, including CV, VC and CVC. It reveals the hierarchical structure of syllables.
According to Syllable Structure Hypothesis (Hockett, 1967), a syllable could be divided into two subgroups—consonant clusters formed one while the vowel and the final consonant was the other. Thus VC sequences were frequently preserved in
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
19
phonological errors in previous researches (e.g. Nooteboom 1969, Mackay 1970, Shattuck-Hufnagel 1979, Rapp & Samuel 2002).
In order to test whether the hypothesis is also applicable in Mandarin, two types need to be approved— a) onsets should be more likely to err than codas, and b) VC sequences, as a unit, tend to induce more errors than CV sequences do. If, in this study, the onset induces less errors or VC intents to induce more speech errors, there could be a rhyme effect.
2.2.3 Phonological Similarity Effect
Similarity effect means that phonetically-similar segments tend to interchange with one another. (e.g. Nooteboom 1969, Mackay 1970, Fromkin 1971, Fay & Cutler 1977, Shattuck-Hufnagel 1979, Levitt & Healy 1985, Stemberger 1985 & 1989, Wan 2007). Fay and Cutler (1977) asserted that mental lexicon was phonologically-arranged based on a distinctive feature system. During the process of lexical selection, the phonologically-similar item in the neighborhood of the intended unit is apt to be selected and thus generated lexical errors. It can account for why the exchanging error in (2) occurred.
(2) reading list leading rist
The two initial consonants [l] and [r] share many except for one phonological features:
[±lateral]. Fay and Cutler (1977) observed 156 malapropisms and found nearly 25 percent differed in only one feature. Therefore, in the process of word activation, the significance of phonological similarity effect illustrates the cognitive status as segments. Wan (1999) examined speech errors and found feature values can alter or be substituted. Wan agreed with Fromkin’s study in which features have cognitive status in speech production, but only the segment has function of segments primary
‧
2.2.4 Syllable Structure Effect
.Syllable Structure contains two levels of meaning: syllabic constituent interacting and syllable pattern interacting. The first level indicates the errors within syllable, especially single-segment substitution. It is also called Syllabic Similarity Phenomenon, which indicates that interacting segments occupy the same syllabic constituent in the syllable, i.e. onset substitutes for onset, nucleus replaces nucleus and coda substitutes for codas. Mackay (1970) and Wan (2007) found that such kind of replacement errors occupy 98% and 99.22% respectively. This phenomenon can also be seen in what Boomer & Laver (1968), Fromkin (1971), and Shattuck-Hufnagel (1979) have found.
In addition, the second level indicates the interacting syllable patterns are usually the same among substitution errors, i.e. the CVC pattern of errors interact with the CVC target, and CV pattern interacts with the target in CV pattern. Such similarity of syllable structures between interacting words is described in Dell’s (1988) model, as shown in figure 2-7.
Figure 2-7. Syllable Constituent and Structural Hierarchy (Dell, 1988)
In Dell’s proposal, each word node in the lexical network connects to a word shape ‘header’ node with the CVC or CV pattern, and it links to Phoneme Category to
2 Psycholinguists proposed that segment could be an advanced unit for planning, such as Fromkin (1971). Current phonologists, based on OT approach, supposed that phonological feature should be a basic unit in lexical process.
CVC CV
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
21
specify the syllable constituent, i.e. Ci (initial/ onset), V and Cf (final/ coda). Finally, each category node connects further to all possible phonemes (/b/, /d/…). Under this framework, the number of activated phonemes is defined by the word-shape header, called ‘categorically trigged selection,’ which means the nodes with the same syllable pattern are activated, and the highest one will be selected in the end.
2.2.5 Tone Effect
The status of tone in Chinese has disputed whether the mental status of tone exists in lexicon or is pure phonological information.
With the view to tone independency, Packard (1986) observed that Chinese aphasics with a brain lesion in the left hemisphere experienced a deficit in producing tones that was qualitatively and quantitatively similar to the deficit in producing consonants. He argued that “tones, like consonants, are listed in the lexicon as unit phonemes … [and] the tonal phonemes, like segments, are composed of bundles of distinctive features” (p.220). Besides, Chen (1999) and Chen & Dell (2003) followed Packard and argued that tone is lexically distinctive, but tones were just like suprasegmentals in the underlying forms and became associated with rhymes or vowels at the phonetic level. They observed the tone status by means of implicit priming task. They found that the units of syllable-alone and tonal syllable would induce priming effects, while bare tone unit seldom plays the role of priming in implicit priming task. They proposed that the tone may be possible to be separable in lexical process, but it cannot be served as an independent unit to have priming or speech planning. Their explanation assumes that tones behave like segments in Mandarin Chinese because of the priming effect; they also assumed the tone to be less free because it alone induced little priming effect. Chen (1999) further stipulated that tones in Chinese were part of the phonological frame because it acts like the stress in
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
22
Dutch and English.
On the other hand, Wan & Jaeger (1998), Wan (2007) seek for the evidence form speech errors in Chinese and supported that tones are represented lexical underlyingly and ought to be part of the phonological organization of the lexicon. Wan & Jaeger (1998) observed the speech errors and pointed out that when one lexical item is substituted for another, its tone (level or contour tone) remains unitary; when its item is deleted, its tone is also deleted, further proposed that if two words have been blended into a single phonological form during lexical insertion, one of the two underlying tones, usually that of the selected rhyme, would be spoken in the errors.
Besides, Wan (2007) and H. Chen (2008) disagree with J. Y. Chen’s (1999) and J. Y.
Chen & Dell’s (2003) arguments, because the speech errors sharing the same tones with the targets outnumber those with the rhyme effect. (Wan & Jaeger: Tone effect 57.01%; Rhyme Effect 16.51%). The evidence showed that tone should not be associated with rhyme since the error rates of this pattern were not salient.
2.2.6 Phonotactic Regularity Effect
Phonological speech errors seldom create sound sequences that are illegitimate in language. Stemberger (1983) provided one of the rare counterexamples, as in (3):
(3) dorm → dlorm
Error (3) showed that this neuroticism error is a violation of phonological constraint of English prohibiting /dl/ initials, which is very rare in speech errors. By means of observation of error collection, Meringer and Mayer (1895) initially note this effect, and it was termed the “first law” of speech errors by Wells (1951).
Stemberger (1983) noticed there were less than 1% phonological errors which were violations of phonotactic constraint. Production theories claim much of the effect in terms of phonological frame. Dell (1993) indicated that frames are specified in such a
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
23
manner which impossible sound sequences are proscribed. For example, in Dell’s (1988) model, the phonological frame enforced the regularity effect by controlling which sequences of categorized slots were made available. It was thought that there is no available frame that will allow an illegitimate sequence, such as /dlorm/ in English, to be encoded in production. The studies of Dell, Stemberger, as former mentioned, and Fromkin (1971) all took this effect as solid evidence for the active use of phonological rules or frames in encoding phonological sequences of words.
2.2.7 Stroop Effect
This effect was named after John Ridley Stroop who first delivered the effect in English in 1935 in an article entitled Studies of interference in serial verbal reactions, and the Stroop effect refers to the color term reading overrides and is quicker than color naming, which was conducted by measuring the reaction time between visual representation and naming.
In psychology, the Stroop effect is a demonstration of the reaction time in naming task. When the name of color (e.g., "blue," "green," or "red") is printed in a color not refer to its name (e.g., the word "red" printed in blue ink instead of red), naming the color of the word is longer, and naming would result in more errors and hesitations than naming the stimulus which matches to its denoted color. Stroop’s experiment included the following three main parts:
(1) The Effect of Interfering Color Stimuli upon Reading Names of Colors Serially: all the color term were printed in black to measure the effect of the black color as an interfering stimulus.
(2) The Effect of Interfering Word Stimuli upon Naming Colors Serially
The name of the color were appeared in different ink other than it referred to.
(3) The Effects of Practice upon Interference
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
24
The visual stimulus was appeared in squares of a given color.
The difference among the time span for reading the words printed in various colors, the same words printed in black, as well as the same colors printed in squares were measured to observe the interference of conflicting word stimuli upon naming colors. The result showed that the interference of conflicting word stimuli upon the time for naming colors caused an increase of reaction time for naming visual colors.
In psychology, Stroop effect set up that the association process in naming visual colors is radically different from that in reading printed words. However, for linguistics, by means of Stroop technique, we would like to observe further whether the time course is different when the colors are well-controlled in sound similarity, and we could also analyze the phonological relationship between the target and the induced error. Phonological similarity, reaction time, and speech errors will be mainly concerned in experiment of current study.
2.3 The Planning Unit in Sound Production
The issue on advanced planning unit will also be reviewed in this section because speech errors can be evidence of the processing quality and precision from syntactic level to phonological level.
Research in speech production has long been focused on Indo-European languages. The WEAVER++ theory (Rowlof, 1997) asserts that the compositional morphemes of a word need to be spelt out before phonological encoding of the word can proceed. However, the postulation about the word encoding was not satisfied to stay valid in Chinese. Chen & Dell (2002, 2003) conducted an implicit priming task (Meyer 1990, 1991; Roelof & Meyer 1998), and noticed it seems not to be the case.
Their results showed that: (1) syllable-alone unit can induce some priming with different tone, whereas tone alone cannot induce such effect; (2) “syllable + tone”
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
25
(tonal syllable) is a possible unit in word production and orthography and the word’s morpheme do not seem to play a role in language planning; (3) tone could be a metrical frame for the segmental syllable to attach to in Chinese; (4) syllable-onset cannot induce priming effect either, so it couldn’t be regarded as a planning unit in underlying.
2.4 Literature Summary and Research Hypothesis
Researchers have investigated sounds activation in production mechanisms, as proposed by the former serial and interactive model, by means of error-inducing experiments (Dell 1986 & 1990 & 1992, Schwartz 1994, Chen et al. 2002, Martin et al. 2004, Schwartz 2006) and speech errors in a naturalistic setting (Fromkin 1971, Buckingham 1980, Stemberger 1984, Bloch 1986, Schwartz 1994). Both production models should be compared to identify the role of the sound, and how the elements are activated in these models.
Theoretical frameworks of serial-ordering and connectionist have debated on whether the error sounds are produced from a sequential level of processing or the consequence of levels interacting altogether. As to serial approach, the sound error was dealt with the interference within the chain of linguistic components, including semantic, syntactic, morphological or phonological generators (Fromkin 1971). In order to explain how the retrieved lexicon is proceeded to the next level to retrieve the sound, which influenced Levelt (1989 & 1999) to propose lemma and lexeme levels in explaining the phonological units are encoded falsely to the retrieve a similar lexicon (at lemma level), thus the errors are produced. Either target sound or error sound is uttered; serial-ordering model prohibits the possibility of backward activation in lexical processing.
On the other hand, interactive approach proposes bi-directional activation: not
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
26
only spreading but also feedback during lexical activation. Speech error seems to be the noise between lexeme and lemma nodes, or we may say it is also a false connection (Stemberger 1985), on the way to the target lexicon which receives enough activation falsely. The phonetic output in the interactive mechanism is the product of mutual interaction among syntactical, lexical and phonological levels, which in serial-ordering model is merely the product of single source, route, or Fromkin said the false encoding of morphophonemic rule results in error, or the Levelt said it is the false retrieving from lemma level to lexeme level. In Dell’s
only spreading but also feedback during lexical activation. Speech error seems to be the noise between lexeme and lemma nodes, or we may say it is also a false connection (Stemberger 1985), on the way to the target lexicon which receives enough activation falsely. The phonetic output in the interactive mechanism is the product of mutual interaction among syntactical, lexical and phonological levels, which in serial-ordering model is merely the product of single source, route, or Fromkin said the false encoding of morphophonemic rule results in error, or the Levelt said it is the false retrieving from lemma level to lexeme level. In Dell’s