13.
LANGUAGE PRODUCTION
the earliest models of how we produce
LEXICALIZATION words and why we make word
Lexicalization is the process in speech substitutions. They observed that there
production whereby we turn the thoughts were two distinct types of whole word
underlying words into sounds: We substitution speech error: semantic
translate a semantic representation (the substitutions, such as examples (29) and
meaning) of a content word into its (30), and form-based substitutions, such
phonological representation of form (its as examples (31) and (32). Form-based
sound). word substitutions are sometimes called
phonologically related word substitution
There is widespread agreement that errors or malapropisms. (The word
lexicalization is a two-stage process, with “malapropism” originally came from a
the first stage being meaning-based, and character called Mrs. Malaprop in
the second phonologically based. When Sheridan’s play The Rivals, who was
we produce a word, we first go from the always using words incorrectly, such as
semantic level to an intermediate level of saying “reprehend” for “apprehend” and
individual words. Choosing the word is “epitaphs” for “epithets.” Note that while
called lexical selection. We then retrieve Mrs. Malaprop produced these
the phonological forms of these words in substitutions out of ignorance, the term is
a stage of phonological encoding. used slightly confusingly in
According to the best known lemma psycholinguistics to refer to errors where
theory (e.g., Levelt, 1989), each word is the speaker knows perfectly well what the
represented by a lemma. Lemmas are target should be.)
specified syntactically and semantically
but not phonologically. The stage of (29) fingers → toes (semantic)
specifying in a pre-phonological, abstract (30) husband → wife (semantic)
way the word that we are just about to say (31) equivalent →
is called lemma selection; the second equivocal(form based or
stage of specifying the actual concrete phonologically related)
phonological form of the word is called (32) historical → hysterical
lexeme or phonological form selection (form base, phonologically related)
(see Figure 13.3).
The important idea of Fay and
Cutler’s model is that phonological and
Two-stage model of lexicalization
semantic word substitutions happen as a
Conceptual representation result of mistakes in the word retrieval
process.
Lemma
In general, in the two-stage model
semantic and phonological substitutions
occur at different levels. The Fay and
Phonological word form
Cutler (1977) model predicts that semantic
and phonological processes should be
FIGURE 13.3 independent.
Experimental evidence
Evidence from speech errors
The earliest experimental evidence for
Fay and Cutler (1977) presented one of
the division of lexical access into two
13. LANGUAGE PRODUCTION
stages came from studies of the pictures which have superimposed
description of simple scenes (Kempen & distractor words that they have to ignore;
Huijbers, 1983). They analyzed the time naming times are longer when the picture
people take before they start speaking and the word are related. The distractors
when describing these scenes, and lead to the activation of semantic
argued that people do not start speaking competitors that slow down the selection
until the content to be expressed has of the lexical target. In the related word
been fully identified. The selection of translation task, semantically related
several lemmas for a multiword sentence words induce semantic interference;
can take place simultaneously. We cannot however, related pictures produce
produce the first word of an utterance facilitation (Bloem & La Heij, 2003). The
until we have accessed all the lemmas (at SOA is, however, critical; if the interfering
least for these short utterances) and at words are presented 200 ms after the
least the first phonological word form. target, we observe semantic interference,
Individual word difficulty affects only but if they are presented 400 ms before
word form retrieval times. the target, we observe semantic
Early semantic activation: facilitation (Bloem, van den Boogaard, &
Evidence for a phase of early La Heij, 2004). Bloem and La Heij
semantic activation in lexical selection proposed a model of lexical access in
and a later phase of phonological which semantic facilitation is localized at
activation in phonological encoding the conceptual level, semantic
comes from picture–word interference interference is localized at the lexical level,
studies (Levelt et al., 1991a; Schriefers, and only one concept is selected for
Meyer, & Levelt, 1990). These lexicalization. They called this the
experiments, discussed in more detail Conceptual Selection Model (CSM). They
later in the section on the time course of account for the effects of SOA with the
lexicalization, used a picture–word assumption that lexical representations
interference paradigm in which decay faster than conceptual
participants see pictures that they have representations.
to name as quickly as possible. At about Whether or not we observe facilitation
the same time they are given an auditorily or inhibition in the picture–word
presented word for which they have to interference paradigm depends on the
make a lexical decision. Words prime details of the experimental set-up. In the
semantic neighbors early on, whereas late most famous example of picture– word
on they prime phonological neighbors. interference, the Stroop task (naming the
This suggests that there is an early stage color in which a word is printed when the
when semantic candidates are active (this word spells out a color name), there is
is the lemma stage), and a late stage when striking inhibition. Usually we find
phonological forms are active. interference with semantically related
The semantic-interference paradigm pairs from the same category, and
provides evidence for two stages, and facilitation with phonologically related
furthermore, that the lexical items pairs. Schriefers et al. (1990) found that
activated by the first stage compete inhibition disappears if participants have to
against each other (Starreveld & La Heij, press buttons instead of naming pictures,
1995, 1996). In semantic-interference suggesting that the interference reflects
studies, participants have to name competition among lexical items at the
13. LANGUAGE PRODUCTION
stage of lemma selection. The details of
the task and the timings involved are also
critical (Bloem & La Heij, 2003; Bloem et
al., 2004).
Evidence from neuroscience
(important; students task)
Different regions of the brain become
activated in sequence as we produce
words (Indefrey & Levelt, 2000, 2004).
Conceptual selection of a word in picture
naming is associated with activation of the
mid-part of the left middle temporal gyrus;
accessing a word’s phonological code is
associated with activation of Wernicke’s
area; and phonological encoding, in terms
of the preparation of syllables, sounds,
and the prosody of the word, is associated
with activation around Broca’s area. As we
shall see, lesions to these areas lead to
different types of impairment to word
naming, with damage to more posterior
regions of the brain resulting in difficulty in
accessing the meanings of words, and
damage to more frontal regions resulting
in difficulty in accessing the sounds of
words. A survey of the imaging literature
also reveals the timings of word retrieval in
naming an object (Indefrey & Levelt,
2004): Visual and conceptual processing
take on average 175 ms; the best-fitting
lexical item, or lemma, is retrieved
between 150 and 225 ms; the
phonological representations are retrieved
between 250 and 330 ms; and the details
of the sounds of the word at around 450
ms (see Figure 13.5).
Electrophysiological evidence also
supports the two-stage model (van
Turenout, Hagoort, & Brown, 1998). Dutch-
speaking participants were shown colored
pictures and had to name them with a
simple noun phrase (e.g., “red table”). At
the same time the participants had to push
buttons depending on the grammatical
gender of the noun, and
Picture 0 ms
↓ Conceptual preparation
Lexical concept
175 ms
↓ Lemma retrieval
Multiple lemmas
↓ Lemma selection
Self-monitoring
400–600
Target lemma 250 ms
275–400
200–400 ↓ Phonological code retrieval
Lexical phonological
150–225 output code
↓ Segmental spell-out
L Segments 350 ms
↓ Syllabification
Phonological word 455 ms
↓ Phonetic encoding
Articulatory scores 600 ms
↓
Articulation
FIGURE 13.5 Time taken (in ms) for different processes to occur in picture naming. The specific
processes are shown on the right and the relevant brain regions are shown on the left. Reprinted
from Indefrey and Levelt (2004).
PET scans of human
brain areas which are
active while speaking
and listening. Top
left—monitoring
imagined speech lights
up the auditory cortex.
Top right—working out
the meaning of heard
words activates other
areas of the temporal
lobe. Bottom
left—repeating words
13. LANGUAGE PRODUCTION 5
activates Wernicke’s area for language comprehension (right), Broca’s area for speech
generation (left), and a motor region producing speech. Bottom right— monitoring speech
activates the auditory cortex.
on whether or not it began with a Stop and try to name the item defined by
particular sound. The electrophysiological (33). You may experience a TOT.
data for the preparation of the motor Example (33) defines the word
movements suggested that the syntactic “sextant.” Brown and McNeill found that a
properties were accessed before the proportion of the participants will be
phonological information. However, the placed in a TOT state by this task.
time delay between the two was very Furthermore, they found that lexical
short—in the order of 40 ms. retrieval is not an all-or-none affair.
Partial information, such as the number of
Evidence from the tip-of-the-tongue syllables, the initial letter or sound, and
the stress pattern, can be retrieved.
phenomenon Participants also often output near
The tip-of-the-tongue (TOT) state is a phonological neighbors like “secant,”
noticeable temporary difficulty in lexical “sextet,” and “sexton.” These other words
access. It is an extreme form of a pause, that come to mind are called interlopers.
where the word takes a noticeable time to TOTs show us that we can be aware of
come out (sometimes several weeks!). the meaning of a word without being
You are almost certainly familiar with this aware of its component sounds; and
phenomenon: You know that you know furthermore, that phonological
what the word is, yet you are unable to get representations are not unitary entities.
the sounds out. TOTs are accompanied There are two theories of the origin of
by strong “feelings of knowing” what the TOTs. These are called the partial
word is. They appear to be universal; they activation and blocking (or interference)
have even been observed in children as hypotheses. Brown (1970) first proposed
young as 2 (Elbers, 1985). The incidence the partial activation hypothesis. This says
of TOTs increases with old age (Burke, that the target items are inaccessible
MacKay, Worthley, & Wade, 1991), and because they are only weakly represented
TOTS are more common in bilingual in the system. Burke et al. (1991) provided
speakers (Gollan & Acenas, 2004; Gollan evidence in favor of this model from both
& Brown, 2006). They appear to be an experimental and a diary study
universal; deaf speakers experience “tip- involving a group of young and old
of-the-finger” states (Thompson, participants. They
Emmorey, & Gollan, 2005).
Brown and McNeill (1966) were the
first to examine the TOT state
experimentally. They induced TOTs in
participants by reading them definitions of
low-frequency words, such as (33):
(33) “A navigational instrument used in
measuring angular distances,
especially the altitude of the sun,
moon, and stars at sea.”
13. LANGUAGE PRODUCTION 6
The tip-of-the-tongue (TOT) state is an extreme form of a pause, where the word takes a noticeable time to
come out.
argued that the retrieval deficit involves weak links between the semantic and the phonological
systems: there is a transmission deficit in getting between the two. A broadly similar approach by
Harley and MacAndrew (1992) localized the deficit within a two-stage model of lexical access,
between the abstract lexical units and the phonological forms. At first sight Kohn et al. (1987)
provided evidence contrary to the partial activation hypothesis in the form of a free association task.
They showed that the partial information provided by participants does not in time narrow or
converge on the target. However, A. S. Brown (1991) pointed out that participants might not say out
loud the interlopers in the order in which they came to mind. Furthermore, in a noisy system there is
no reason why each attempt at retrieval should give the same incorrect answer.
Further evidence that TOTs are associated with a difficulty in retrieving the phonological forms
of words comes from brain imaging. Shafto, Burke, Stamatakis, Tam, and Tyler (2007) had people
aged 19–88 name pictures of famous people. The number of TOTs increased with age and with
atrophy of the left insula, a region of the brain known to be involved (among other things) in
phonological production.
Problems with the lemma model
Although most researchers favor the two-stage model of lexicalization, there is less agreement on
the need for lemmas as a level of amodal, syntactically specified representations mediating between
concepts and phonological forms (Caramazza, 1997; Caramazza & Miozzo, 1997, 1998; Miozzo &
Caramazza, 1997).
One point is that it is not clear that the need for lemmas is strongly motivated by the data. Most
of the evidence really only demands a distinction between the semantic and the phonological levels.
The strongest evidence for lemmas comes from the finding that gender can be retrieved when in the
tip-of-the-tongue state, although this interpretation has been disputed. It should not be possible to
retrieve phonological information for a word without retrieving the syntactic information for that
word such as gender, as the phonological stage can only be reached through the lemma stage. Tip-
of-thetongue data suggest, however, that syntactic
13. LANGUAGE PRODUCTION 7
and phonological information are independent (Caramazza & Miozzo, 1997, 1998; Miozzo &
Caramazza, 1997): Italian speakers can sometimes retrieve partial phonological information when
they cannot retrieve the gender of the word, and vice versa. Importantly, there was no correlation
between the retrieval of gender and phonological information; people are no better at recalling
gender when they correctly recall the initial phoneme of the target in a TOT state than when they
fail to do so. Hence, phonological retrieval does not necessarily depend on syntactic retrieval, and
therefore these results do not support the idea of syntactic mediation. Arguing that lemmas are
unnecessary complications, Caramazza (1997) dispenses with them. He proposes that lexical
access in production involves the interaction of a semantic network, a syntactic network, and
phonological forms (see Figure 13.6). Semantic representations activate both appropriate nodes in
the syntactic network and the phonological network.
Is lexicalization interactive?
Given that there are two stages involved in lexicalization, how do they relate to each other?
Interaction involves the influence of one level of processing on the operation of another. It
comprises two ideas. First, there is the notion of temporal discreteness.
Dell’s (1986) interactive model of speech production
Dell (1986) proposed an interactive model of lexicalization based on the mechanism of spreading
activation. Items are slotted into frames at each level of processing. Processing units specify the
syntactic, morphological, and phonological properties of words. Activation spreads down from the
sentence level, where items are coded for syntactic properties, through a morphological level, to a
phonological level. At each level, the most highly activated item is inserted into the currently active
slot in the frame. For example, the sentence frame might be quantifier–noun–verb. The
morphological frame might be stem plus affix. The phonological frame might be onset–nucleus–coda.
The final output is a series of phonemes coded for position (e.g., /s/ in word-onset position). The
flow of activation throughout the network is timedependent, so that the first noun in a sentence is
activated before the second noun.
The model (see Figure 13.9) gives a good account of speech errors. Several units may be active
at each level of representation at any one time. If there is sufficient random noise an item might be
substituted for another one. As items are coded for syntactic category and position in a word, the
other units that are active at any one time tend to be similar to the target in these respects. There is
feedback between levels. The feedback between the phonological and lexical levels gives rise to
lexical bias and similarity constraints.
A related issue that has recently arisen is the degree to which there is competition within a level
between similar units. Recall that in the IAC model of letter recognition there are withinlevel
inhibitory links leading to competition between similar units. The key issue therefore is whether the
time to produce a word is affected by the activation of similar words. This issue is currently
unresolved, with some researchers arguing for competition, others against it, while yet others claim
that the data can be accounted for by an internal monitor checking planned productions against
internal goals (Dhooge & Hartsuiker, 2012; Melinger & Rahman, 2013).
