CHAPTER 5 DISCUSSION
5.2 Cohort and Merge models in Taiwan Mandarin
國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
carries the tone and occupies a relatively long period of time in the disyllabic words, it
lacks the advantage of being processed first. Hence, the vowel in the second syllable
is less influential than the vowel in the first syllable.
5.2 Cohort and Merge models in Taiwan Mandarin
According to the Cohort model (Marslen-Wilson & Zwitserlood, 1989; Tyler,
1984; Marslen-Wilson & Tyler, 1980, 1981), word initial input is of paramount
importance. Therefore, once the initial input is disrupted by the noise, the word can
hardly be recognized. This claim is not true on the basis of the results in this study.
The results of experiment 1 display that the accuracies of 1C for the high-frequency
and low-frequency words are 95.14% and 88.96%, respectively. This demonstrates
that even if the word initial information (the initial consonant) is replaced by the
hiccup noise, the words can still be successfully recognized in most of the cases.
Unlike the Cohort theory, the Merge model (Norris, McQueen, and Cutler, 2000)
proposed that it is the overall match between the acoustic input and the
representations that is the most crucial for spoken word recognition. This model
greatly reduces the importance of the initial input. However, the results of the study
are not fully compatible with the model. The results depict that it is the first vowel in
the disyllabic words that is the most important and the second vowel in the disyllabic
words that is the second important, but it is not stated in the model which segment is
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
the most important. Therefore, the test items whose first vowels are replaced by the
hiccup noise cause the lowest rate of successful recognition.
According to the results, the word initial information is not the most crucial in
the spoken word recognition of Taiwan Mandarin. The first vowel in the disyllabic
word is the most important. This is because the vowel occupies the longest period of
time in words and carries much important information which is very crucial for
spoken word recognition in Taiwan Mandarin. One of the important acoustic-phonetic
cues in vowel is tone. Tones are very important in Taiwan Mandarin since it can
distinguish the meaning of words. The results of the study are compatible with this
claim. The results of experiment 3 show that if the tones of the disyllabic words are
leveled to around 100Hz, subjects merely have 71.67 percent chance to recognize the
high-frequency words and 56.74 percent chance to recognize the low-frequency
words, which is much lower than the chance to recognize the words whose single
segment is replaced by the hiccup noise. The fact indicates that the whole tone of the
disyllabic words is more important than one single segment though some segments
also carry tone, such as the vowel and coda nasal.
According to the results in experiment 3, tone should be added to the processing
of spoken words in Merge. Nevertheless, it cannot be inferred only by the
experiments in this study whether Mandarin tones should be processed before the
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
segments or after the segments. Cutler and Chen (1997) asked the subjects to judge
whether the word and nonword in a pair differing only by the initial consonant, vowel,
or tone were the same or different. The results displayed that subjects’ responses to
the pair differing by tone were slower and more inaccurate than those differing by the
onset consonant and vowel. Therefore, they proposed that tone is processed slower
than segment. According to Cutler and Chen (1997), it can be proposed that tonal
level can be added to the Merge model after the phoneme level. The acoustic-phonetic
cues of tones can be processed in tonal level and sent to the lexical nodes by the
excitatory connections. The candidates activated in the lexical level not only need to
match the segmental information but the tonal information of the input as well. The
candidate having the best match to the acoustic input wins the lexical competition.
In addition, the results of experiment 1 can serve as the support for Merge. In
merge, the phoneme decision level is designed to resolve the issues regarding
phoneme decision making. The integration approach of Merge allows the prelexical
information to proceed independently of lexical processing. Both prelexical and
lexical processing information proceed to the phoneme decision level and then merge
together. In the Merge model, the prelexical processing activates some compatible
lexical candidates. At the same time, the prelexical processing also sends the
excitatory information to the phoneme decision level. The phoneme decision nodes
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
also keep accepting the facilitatory information from the lexical nodes and merge the
two inputs from different levels together. The merged information competes with each
other by inhibitory connections and decides which phonemes are actually present in
the input.
Both Merge and TRACE (McClelland & Elman, 1986) can account for why the
disruptive targets can still be recognized in experiment 1. However, TRACE would
overlook the disruptive segment because the interactive models run the stake of
hallucinating. Especially when the input is degraded or disruptive, the input
information tends to be abandoned. In TRACE, phoneme decision can mainly
depend on the lexical information from the lexical level. This is because top-down
activation can function as the distortion to the prelexical processing of the acoustic
input. The strong top-down feedback would override the disruptive segment and the
disruptive segment could be ignored. For example, the hiccup noise in /Na51 35/
would be overlooked because of the strong top-down feedback. In reality, subjects can
still notice the hiccup noise. In contrast, the hiccup noise would not be overlooked in
Merge. The prelexical phoneme nodes are independent of the lexical nodes; that is,
there is no top-down feedback from the lexical nodes to the prelexical level. The
prelexical nodes accept the hiccup noise and keep sending the prelexical processing
information of the following segments to the lexical nodes. The lexical nodes then
‧ 國
立 政 治 大 學
‧
N a tio na
l C h engchi U ni ve rs it y
activate the possible lexical candidates and send excitatory information to the
phoneme decision nodes. Therefore, although phoneme decision nodes cannot receive
the excitatory information of the segment replaced by the hiccup noise from the
prelexical nodes to decide what the disrupted segment is, they can still accept the
information from the lexical nodes and do the phoneme decision. Contrary to TRACE,
Merge can do the phoneme decision without overlooking the hiccup noise. Hence,
Merge is a better model than TRACE in this facet.