2.1 Two Processing Hypotheses: Modularity versus Interaction
2.1.3 Modularity and Interaction Hypothesis in Lexical Ambiguity Resolution 15
In psycholinguistics, how the language processing system is organized is a
central issue. To find out the operation of language processing, one needs to figure out
a crucial question: whether different levels or subsystems in language processes are
modular or interactive in nature, i.e., whether they operate autonomously or they
interact with one another.
Lexical ambiguity in sentence processing is a useful test bench for investigating
modularity and interaction hypothesis in language processing (Ahrens, 2001, 2002; Li,
Shu, Yip, Zhang, & Tang, 2002; Lupker, 2007; Zhang, Wu, & Yip, 2006). The
question to be probed is whether sentential context, or contextual information from
the discourse level, can influence the access of a word’s meaning at the lexical level.
Since ambiguous words have more than one meaning, researchers are concerned
about which meanings of an ambiguous word are accessed or at what stages they are
accessed in sentence processing. If processing is modular, different levels of language
system should operate independently, and thus sentential context (a higher-level
subsystem) may not influence lexical access (a lower-level subsystem). On the other
hand, if processing is interactive, different levels of language system operate
simultaneously, and hence sentential context may have an effect on lexical access. For
the issue of lexical ambiguity resolution in language processing, two basic theoretical
16
models associated with modularity and interaction hypotheses are context-
independent model
and context-dependent model respectively. Over the past fewdecades, the research question about which hypothesis is a more viable model to
account for lexical processing has long been in debates.
Within the modularity hypothesis, the context-independent model argues that all
meanings of an ambiguous word will be accessed momentarily regardless of the
influence of sentential context. This model assumes that sentence processing is a
modular, bottom-up process in which non-lexical or contextual information does not
penetrate lexical access. The context-independent models could be further divided
into two versions, the exhaustive/multiple access model and the ordered access model,
depending on whether or not the factor of frequency is taken into account.
The exhaustive or multiple access model postulates that all meanings of an
ambiguous word are accessed immediately and automatically and that sentential
context can only help to select the appropriate meaning from the output of the lexical
module at a later selection stage. Previous research in this position has shown that as
soon as an ambiguous word was encountered immediately, when lexical selection had
not occurred, all associated meanings of ambiguity were initially accessed (Ahrens,
1998, 2001, 2002, 2006; Conrad, 1974; Kintsch & Mross, 1985; Lucas, 1987; Onifer
& Swinney, 1981; Seidenberg, Tanenhaus, Leiman, & Bienkowski, 1982; Swinney,
1979; Swinney & Love, 1996; Tanenhaus, Leiman, & Seidenberg, 1979; Till, Mross,
& Kintsch, 1988).
In lexical processing, several stages are involved (Harley, 1996; Lucas, 1999):
First, recognizing a word begins when the representation of the sensory input makes
initial contact with the lexicon. This is the initial contact stage or pre-access stage.
Once lexical entries match the contact representation, they become ‘activated’. Hence,
access stage occurs when the perceptual representation is used to activate lexical
items and the semantic level of information would be activated as well. Third, in the
selection stage, sensory input continues to accumulate until one lexical entry is
chosen. Finally, the integration stage that then follows is the start of comprehension
process, where the semantic and syntactic properties of the word are integrated into
the higher-level sentence representation. According to Harley (1996), the first two
phases (pre-access and access processes) could be referred to as pre-lexical stages,
and the last two phases could be classified as post-access or post-lexical stages (see
Table 2.2). According to the exhaustive or multiple access model, context can only
interfere with lexical activation at a post-access stage, but not at the earlier stages.
Table 2.2. Stages of lexical processing
Pre-lexical stage Post-lexical (post-access) stage
Initial contact(pre-access) stage Access stage Selection stage Integration stage
18
On the other hand, the ordered access model attempts to take into account the
effects of meaning dominance (Hogaboam & Perfetti, 1975; Simpson & Burgess,
1985). According to this account, lexical meanings are activated based on frequency
ranking, with more frequent meanings being accessed faster (Forster & Bednall,
1976). Under this model, when an ambiguous word appears, its most frequent
meaning would be always retrieved. This meaning would be tested against the context,
and if it is determined to be appropriate, processing stops with no activation of other
meanings. If it is not appropriate, then the next most frequent meaning is retrieved,
and so on, until an appropriate match is found. Therefore, frequency bias is viewed as
a default procedure, and the initial retrieval or access of the meanings is presumed to
occur independently of context. Different from the exhaustive/multiple access model,
the ordered access model maintains that the alternative meanings of an ambiguous
word are accessed in order of dominance or relative frequency of the individual
meanings of that word. Nonetheless, both of these two models hold the position that
the preceding sentential context has no influence on access to lexical meanings at an
initial stage. It affects the output of the lexical module only during the post-lexical
stage.
By contrast, the context-independent model within the interactive hypothesis
argues that the contextually appropriate meaning of an ambiguous word can be
selectively accessed early on if sentence context provides a semantic bias to the
appropriate meaning. This model assumes that sentence processing is an interactive
process in which information can flow in both bottom-up and top-down ways and that
lexical access and sentential context can mutually influence each another at a very
early stage. There are also two versions of the context-independent model: one is the
(strictly) selective access model and the other one is the reordered access model,
which also depends on whether frequency is taken into account as a factor.
The (strictly) selective access model suggests that sentential context can be used to
access the appropriate meaning of an ambiguous word at an early stage, so context
plays a pre-lexical role in the process of meaning access (Glucksberg, Kreutz, & Rho,
1986; Li, 1998; Li & Yip, 1996, 1998; Oden & Spira, 1983; Paul, Kellas, Martin, &
Clark, 1992; Simpson, 1981; Simpson & Krueger, 1991; Tabossi, 1988; Tabossi et al.,
1987; Vu, Kellas, Metcalf, & Herman, 2000; Vu, Kellas, & Paul, 1998). Studies in this
position have shown that only the contextually appropriate meaning of the ambiguous
words was activated no matter whether the context bias towards the primary or
secondary meaning of the ambiguity.
The reordered access model, on the other hand, incorporates meaning dominance
into the interactive view (Duffy, Morris, & Rayner, 1988; Guo et al., 2007; Li et al.,
2002; Tabossi & Zardon, 1993). This model is supported by a number of evidence
20
from eye movement studies, in which subjects read ambiguous words embedded in
sentences (Duffy et al., 1988; Rayner & Duffy, 1986; Rayner & Frazier, 1989; Rayner
& Morris, 1991; Sereno, 1995). According to these findings, gaze duration on
ambiguous words was affected by both context and relative meaning frequency of
ambiguity. Two types of ambiguity are distinguished in their studies: one is balanced
(or equi-biased) ambiguous words, whose meanings are relatively equal in frequency,
and the other one is unbalanced (or biased, polarized) ambiguous words, for which
one meaning is more frequent than others. When the context is biased toward the
subordinate meaning of ambiguity, balanced ambiguous words do not cause
processing difficulty, but unbalanced ambiguous words show longer gaze durations. It
seems that, for unbalanced ambiguous words, the context can boost the activation of
the subordinate meaning so as to make it compete with the primary meaning for
access. In this sense, the access is thus “reordered” in terms of the relative meaning
frequency (Duffy et al., 1988). That is, the frequency-based priority of meanings may
be altered by strongly biasing context toward the subordinate sense. Based on this
reordered access model, the alternative meanings of an ambiguous word are accessed
in order of relative frequency, and sentential context is also assumed to play a role in
activating the contextually appropriate meaning at an early stage. In other words, both
meaning dominance (frequency) and context can act in parallel and lead to immediate
automatic activation. Previous research in this view such as Tabossi and Zardon (1993)
found that there was “a frequency and context effect”, which indicated that the
contextually inappropriate meaning was activated only when its frequency of meaning
was high. That is, only the primary meaning was activated when the sentential bias
was towards the primary meaning, but both the primary and secondary meaning were
activated when the sentential bias was towards the secondary meaning. Simpson
(1984, 1994) suggested that the results of the studies on lexical ambiguity processing
could be better explained by having a more flexible model in which all meanings are
activated, but the degree of activation would be sensitive to the relative frequency of
the meanings and the context in which the ambiguous word occurs.
At first glance, the ordered access and reordered access models appear to make a
similar prediction that if the secondary meaning is facilitated, the primary meaning (or
the most frequent meaning) will also be accessed. However, the critical difference is
whether the context effect acts at very early stages of meaning activation. In the
reordered access model, both context and frequency have immediate effects on lexical
access; in the ordered access model, however, the consideration of frequency for
meaning retrieval is prior to the effect of context. Ahrens (1998, 2002) also pointed
out a solution to differentiate between these two models by examining ambiguous
words with more than two meanings as stimulus items. For example, the ordered
22
access model (within the modular view) would predict that, if the fourth meaning of
an ambiguous word is accessed, all more dominant meanings (i.e., from the primary
to the third meanings) should be also activated. In the reordered access model (within
the interactive view), on the other hand, sentential context can influence the access of
the fourth meaning of that word when the context is biased to that meaning, and the
most dominant meaning is also accessed, but not for the secondary and third
meanings.
To summarize, under the context-independent and context-dependent models,
there are four versions of models to account for lexical ambiguity resolution, in terms
of sentence context and meaning dominance (or relative frequency). The classification
of these accounts is summarized in Table 2.3.
Table 2.3. Different types of models of lexical ambiguity resolution Processing
2.2 Methodological Issues: Past and Present
Based on the above discussion, it is revealed that some studies indicated that
context can influence lexical access at an early stage, but others do not. Either the
context-independent model within the modularity view or the context-dependent
model within the interactive view has a lot of empirical studies in support of them.
Although these conflicting results make it difficult to figure out which account of
lexical ambiguity resolution is more persuasive, it is shown that these previous
experiments in fact differed along a number of parameters (Ahrens, 2002, 2006; Guo
et al., 2007; Lucas, 1999; Lupker, 2007; Simpson, 1994). Inconsistent findings from
previous works have plagued this area due to their inconsistent methodologies. When
the experimental tasks operated over different modalities, the obtained results of
course would reflect different aspects of processing. For example, if one experiment
sets the conditions for a post-access decision or selection stage to occurr, the results
probably would not show modular processing. As pointed out by Ahrens (1998, 2002),
two central aspects for modular processing are automaticity and immediacy. If
automaticity and immediacy are compromised in any experimental methodology,
modular processing may be affected and thus cannot be considered as a satisfactory
test of the modularity hypothesis. Therefore, variations in methodology have a
significant impact on the results, which may support either the modular view or the
24
interactive view. Compared with the conditions in interactive processing, those for
modular processing to occur are stricter since modularity assumes processing is
immediate and automatic. As shown in Table 2.3, the critical difference between the
modular and interactive hypothesis is whether there is a context effect at an initial
stage. As a result, to test between modularity and interaction hypothesis, the
experimental design should allow for the processing stage of immediate access to be
detected. In the current research, the experimental design hence would be set for
immediate and automatic access to occur.
In what follows, discrepancies over three main aspects of methodology in
previous studies are discussed below, including (1) language, (2) experimental design
and (3) different types of lexical ambiguity. Additionally, the rationale for what
methodology is adopted, and how each variable among these parameters is controlled
and manipulated in the present study will also be explained:
(1) Language
The languages being studied on lexical ambiguity resolution include English (e.g.
Oden & Spira, 1983; Onifer & Swinney, 1981; Seidenberg et al., 1982; Simpson,
1981; Simpson & Krueger, 1991; Swinney, 1979; Tanenhaus et al., 1979; Vu et al.,
2000; Vu et al., 1998), Italian (e.g. Tabossi, 1988; Tabossi et al., 1987; Tabossi &
Zardon, 1993), Cantonese (e.g. Li, 1998; Li et al., 2002; Li & Yip, 1996, 1998) and
Mandarin Chinese (e.g. Ahrens, 1998, 2001, 2002, 2006). In those studies on English,
as indicated earlier, some of them supported the modular view, but others did not. In
Italian studies, Tabossi and colleagues found their findings in support of the
interactive view. In Chinese studies, there is also no agreement among the studies in
Cantonese and Mandarin Chinese. So far, the Cantonese studies had results in favor of
the interactive view, while studies on Mandarin found evidence for the modularity
hypothesis.
Compared with the studies on Indo-European languages like English and Italian,
the number of studies on Chinese is still limited. Chinese differs significantly from
most Indo-European languages in its phonological, lexical, syntactic, and discourse
structures, which provide unique properties for psycholinguistic investigations of
lexical and sentence processing from a cross-linguistic perspective (Li, 1998). For
example, Chinese is a language that heavily relies on contextual information to arrive
at linguistic interpretation (Ahrens, 1998, 2002). It would be notable to see how
ambiguity in Chinese is resolved in contextual situations. In the present study,
therefore, Chinese (Mandarin) is the target language to be examined on the resolution
of lexical ambiguity in sentence comprehension.
26
(2) Experimental design
a. Modality of sentence presentation
There are two modalities presenting the ambiguity and the sentential context (in
which the ambiguity is embedded) for comprehension: one is auditory modality and
the other one is visual modality. According to Ahrens (1998, 2002), these two
modalities have distinct properties, which may yield different processing behaviors:
First, different areas in the brain deal with different types of incoming stimuli, and these different stimuli may be handled with different time courses from one another. Second, speech production and comprehension is learned earlier than reading. Third, speech is the basis for learning how to read. Fourth, reading is carried out at a slower speed than listening. (Ahrens, 2002, p.14)
In previous research, inconsistent findings exist in studies with either auditory or
visual presentation. In those with auditory presentation, some supported modular
position (e.g. Ahrens, 1998, 2001, 2002, 2006; Oden & Spira, 1983; Onifer &
Swinney, 1981; Seidenberg et al., 1982; Simpson, 1981; Swinney, 1979; Tanenhaus et
al., 1979) and others did not (e.g. Li, 1998; Li & Yip, 1996, 1998; Tabossi, 1988;
Tabossi et al., 1987). In those with visual presentation, some results favored the
modularity hypothesis (e.g. Kintsch & Mross, 1985; Till et al., 1988) but others found
interactive processing (e.g. Paul et al., 1992; Simpson & Krueger, 1991).
Since listening speed is faster than reading speed, auditory modality offer a more
appropriate test bench for the initial access stage of lexical processing. In order to
examine automatic and immediate access of lexical ambiguity in natural speech
comprehension, the auditory modality of presentation is hence used in the present
study.
b. Experimental paradigms
To study the issue of lexical ambiguity resolution, priming method is the most
frequently used experimental paradigm.3 This on-line method is advantageous for its
sensitivity to the time course of lexical access processes and for direct measures of
meaning activation (Lucas, 1999; Simpson, 1994). In priming studies, a related target
word follows the ambiguous word (as the prime word) that is embedded in a sentence
and subjects are asked to respond to a visual target on the screen either by making a
lexical decision or by naming it. For example, subjects may read or hear the sentence
including the ambiguous word bug: “for several weeks following the exterminator’s
visit they did not find a single bug anywhere in the apartment” (taken from Onifer &
Swinney, 1981, p.234). The visual targets which appear immediately after the
ambiguity include two related targets (associated with two meanings of bug) insect
3 Priming is an experimental technique by which a stimulus is used to sensitize the subjects to a later presentation of the semantically related stimulus. For example, shown the word “doctor” first, a subject would later recognize a related word such as “nurse” more rapidly. The word “doctor” is referred to as the prime and “nurse” as the target or probe. The appearance of the word “doctor” is said to prime “nurse”. A semantic priming effect refers to an increase in the speed with which a word is recognized, which results from having seen or heard a word that is closely associated with it.
Exposure to the prime is represented as activating (or bringing into prominence) a range of associated words. These words then become easier to identify since they are already foregrounded in the mind. The activation process is often examined by the reaction times or how long it takes to
28
and spy (as in the experimental condition) and unrelated targets such as custom and
tan
(as in the control condition). When subjects respond to a visual target on thescreen either by making a lexical decision or by naming it, their response times or
reaction times to a particular word could be measured, which reveals how much time
it takes for the subjects to recognize or access a lexical word. The priming effects,
measured by the differences on reaction times to the target in two (experimental vs.
control) conditions, indicate the speed of access to a particular meaning of the
ambiguity. For instance, researchers then could examine whether only the
contextually appropriate target insect is accessed (relative to the reaction times to the
unrelated control target custom) or both meanings of targets insect and spy are
activated. In this way, by employing the priming task, whether or not the initial lexical
processing depends on contextual information could be answered. If access to the
lexical representation of the prime is a modular/context-independent process, targets
related to all meanings of the prime should be activated, relative to their control
targets. On the other hand, if lexical access is an interactive/context-dependent
process, only the target related to the contextually biased meaning should be
facilitated.
According to different modalities, there are two main types of the priming
paradigm used to explore the resolution of lexical ambiguity: single-modal (visual)
priming and cross-modal priming: In single-modal (visual) priming studies, subjects
are asked to read a written sentence, either presented each sentence at one time (e.g.
Simpson & Krueger, 1991) or presented word by word with rapid serial visual
presentation (RSVP) in which each word was presented one at a time and each one
replaces the previous one on the computer screen (e.g. Kintsch & Mross, 1985; Till et
al., 1988), or presented with unfolding procedure in which each word appears to the
right of the preceding word (Paul et al., 1992; Vu et al., 2000; Vu et al., 1998). As
indicated in the studies with visual modality, evidence has been found for either
modular or interactive view in those studies with visual priming paradigm.
In the cross-modal priming paradigm, on the other hand, sentence presentation is
with auditory modality. Subjects are asked to listen to a sentence containing an
with auditory modality. Subjects are asked to listen to a sentence containing an