2
Mean proportion "f"
1
Lexical knowledge is available, but not always used, very early
Amanda Rysling, John Kingston, Adrian Staub, Andrew Cohen, Jeffrey Starns, and Anthony Yacovone
(Depts. of Linguistics and Brain and Psychological Sciences, University of Massachusetts, Amherst.)
Understanding when lexical information is available and when it is used are key to the
construction of models of perceptual processes sensitive to lexical effects. Previous work
(Fox 1984) has shown that early responses to an ambiguous word-initial segment do not
recruit lexical information to identify that segment. Such findings have been argued to
support models with incremental increases in lexical activation, in which lexical information
would not inform responses until late because words are activated gradually (McClelland &
Elman 1986). Here we present work that demonstrates that lexical knowledge is available to
inform responses earlier than has been reported, but is not recruited unless listeners
develop an appropriate response strategy.
In three studies, participants categorized syllable-initial fricatives from a 20-step /s/ to /f/
continuum in the same three contexts: lexical /s/-biasing /aɪd/ (side, *fide), lexical /f/-biasing
/aɪl/ (file, *sile), and no lexical biasing /aɪm/ (*sime, *fime). All fricatives were recorded in the
context of the /aɪ/ vowel alone, and all lexical contexts were recorded with /h/ at the
beginning of the syllable (i.e. hide, *hile, *hime), to avoid coarticulatory bias. The /s/ and /f/
fricatives were then spliced together with these contexts to make the stimuli.
In Exp. 1 (n=27), a free-response study, participants categorized fricatives without time
pressure. Responses faster than the 0.005 quantile and slower than the 0.995 quantile were
excluded from analysis. A mixed effects logistic regression with treatment coding of Context
and by-participant random slopes and intercepts was fitted to the proportion of "f" responses.
The expected lexical bias effects were obtained: participants responded "f" more often in the
/f/-biasing /aɪl/ context (f-bias vs. no bias; β=0.99, z=4.67, p<0.001), and "s" more often in
the /s/-biasing /aɪd/ context (s-bias vs. no bias; β=-0.26, z=-2.22, p<0.05). An interaction of
RT quantile and f-bias vs. no bias was found (β=0.002, z=2.60, p<0.01), such that there was
no lexical bias in the earliest quantile, but larger lexical biases in later ones.
In Exp. 2 (n=29), a response-signal study, response time quantiles from Exp. 1 were
used to determine four delays at which participants were prompted to respond after the
stimulus onset; these were at 375, 675, 975, and 1350ms after the beginning of a stimulus
(375ms corresponded to when lexically biasing final consonant became audible, and a time
when there was no lexical bias in the free response study; 675ms corresponded to the offset
of all stimuli). Participants were shown a visual cue prompt, and were given 300ms to
respond with a button press.
375
675
975
1350
1.00
Response prompts were randomly
0.75
presented to participants so that
0.50
they could not predict when to
0.25
respond. Only responses that fell
0.00
1.00
within the 300ms target window for
0.75
each delay were analyzed. A
0.50
mixed effects logistic regression
0.25
with the fixed effects plotted in
0.00
Figure 1, all of their two-way
1.00
interactions, and random slopes
0.75
and intercepts by subjects was
0.50
fitted to the proportion of "f"
0.25
responses. As expected, Step (10.00
1 5 9 12 16 20 1 5 9 12 16 20 1 5 9 12 16 20 1 5 9 12 16 20
20; β=4.11, z=13.64, p<0.001)
[s-f]
was significant. Lexical f-bias was
Context ide ile ime
significant (f-bias vs. no bias;
Figure 1: Mean proportion "f" responses by prompt Delay,
β=1.61, z=7.29, p<0.001), and it
Session, Context, and Step
interacted with response prompt
delay (f-bias vs. no bias x Delay;
β=0.31, z=6.20, p<0.001), such that the proportion of "f" responses was greater in the /f/biasing context at all delays, and more so for the later delays than the earlier, as is visible
3
ile
density
ide
across the rows of Figure 1. No main effect of s-bias vs. no bias, Delay, Session, or their
interactions with other fixed effects reached significance. Responses within the 300ms target
window of the 375ms delay fell within the early interval in which Fox did not find lexical
effects (this, along with the RTs of Exp. 3, is shown in Figure 2). Exp. 2 thus showed that
lexical biases are available to, and indeed do, inform responses quite early in speech
processing, and appear when the listener is
2
3
0.0125
under time pressure to respond.
0.0100
0.0075
Exp. 3 (n=25), also a response-signal
0.0050
study, used the same stimuli and procedures
0.0025
0.0000
as Exp. 2, with different response delays of
0.0125
0.0100
175 and 375ms after stimulus onset. Only
0.0075
on-time
responses
were
analyzed.
0.0050
Inspection of the distributions of on-time
0.0025
0.0000
responses at the 375ms delay across Exps.
0.0125
0.0100
2 and 3 revealed that participants in Exp. 3
0.0075
developed a qualitatively different response
0.0050
0.0025
strategy from those in Exp. 2: responses at
0.0000
375ms in Exp. 2 were all clustered in the
0
100
200
300 0
100
200
300
On-Time RT (ms)
second half of the 300ms response interval,
Response F S
while responses in Exp. 3 were distributed
throughout that window, and actually were more numerous in the first half, as shown in
Figure 2. Assuming an average of 200ms to
plan and execute a button press, participants
Figure 2: Response time distributions at the
in Exp. 3 launched the majority of their
375ms prompt in Exps. 2 and 3
responses without completely integrating coda
consonants or recognizing the wordhood of the stimuli. In contrast, paricipants in Exp. 2
waited as long as they could while still responding within the interval, and thereby gained
greater lexical support for their responses. A mixed effects logistic regression of a similar
structure to that fitted for Exp. 2, but without Delay and with the addition of Window Half to
the fixed effects, was fit to the responses at 375ms. It found a significant effect of Step (1-20;
β=2.32, z=10.59, p<0.001), no significant effect of Context, but an interaction between
lexical bias and time within response window (Context x Half; β=0.11, z=2.88, p<0.01), such
that responses in the second 150ms of the on-time window for the 375ms delay, though less
frequent than in the first 150ms, did show a lexical bias effect.
The present studies join Cutler et al. (1987) in demonstrating that changes to the
experimental setting can largely eliminate lexical bias effects. But unlike Cutler et al.'s study,
the present studies found that even when the demands of the task did not require or
encourage the use of lexical information, such information was recruited, as long as
participants did not respond earlier than it became available. We hypothesize that
participants in Exp. 2 were encouraged to develop a lexically-supported response strategy
by the presence of later response intervals, which forced them to wait long enough to have
integrated stimulus codas before giving judgments. A fourth study is planned with 175, 375,
and 675ms delays to test whether requiring responses late enough to have heard the entire
stimulus, and so encouraging the use of lexical information, will indeed cause a shift in the
375ms delay such that lexical information is again reliably used.
These studies demonstrate that lexical information is available to listeners earlier than
previously thought, and that whether listeners choose to leverage this information is
inextricably tied to when they choose to respond. Participants in previous free-response
categorization studies, which did not find lexical bias effects as early as we have found them,
likely enacted a strategy of waiting to ascertain the wordhood of a stimulus before
responding, and then responded later relative to the time windows we probed here. When
participants are forced to respond earlier than is comfortable, but still given the option of
waiting for lexical information, whether they do so seems to be a function of whether they
perceive the need to do so in order to complete the experimental task.
ime