Intonation in the processing of contrast meaning in French: An eye-tracking study

Abstract

Listeners rapidly process tonal composition and pitch accent placement within an utterance to create expectations about its pragmatic meaning and information structure. It is still unknown whether the nuclear pitch accent alone or a combination of pitch accent and the following edge tone are needed in order to process intonational meaning in French. This study investigates the online comprehension of the French (L)H*L% rise-fall " implication " contour, which evokes a contrast meaning. Twenty-nine speakers participated in an eye-tracking experiment. The critical stimuli were sentences whose interpretation could be anticipated by successfully processing the implied meaning evoked by the (L)H*L% rise-fall contour on the critical word (hereafter CW). The results showed that participants are able to associate the implication contour with a contrast meaning, and that they start doing this only after the H* peak of the rise-fall intonation movement has been processed, hence when part of the L% falling movement has been perceived.

Full text

Intonation in the processing of contrast meaning in French: an eye-tracking
study
Núria Esteve-Gibert1, Cristel Portes1,2, Amy Schafer3, Barbara Hemforth4, Mariapaola
D’Imperio1,5
1 Aix Marseille Université, CNRS, LPL UMR 7309, 13100, Aix-en-Provence, France
2 IMS, University of Stuttgart, Germany
3 University of Hawai‘i at Mānoa, Hawai‘i
4 Laboratoire de Linguistique Formelle, CNRS, Université Paris Diderot
5 Institut Universitaire de France (IUF)
nuria.esteve-gibert@blri.fr, cristel.portes@lpl-aix.fr, aschafer@hawaii.edu,
barbara.hemforth@linguist.jussieu.fr, mariapaola.dimperio@lpl-aix.fr
Abstract
Listeners rapidly process tonal composition and pitch
accent placement within an utterance to create expectations
about its pragmatic meaning and information structure. It is
still unknown whether the nuclear pitch accent alone or a
combination of pitch accent and the following edge tone are
needed in order to process intonational meaning in French.
This study investigates the online comprehension of the
French (L)H*L% rise-fall “implication” contour, which
evokes a contrast meaning. Twenty-nine speakers participated
in an eye-tracking experiment. The critical stimuli were
sentences whose interpretation could be anticipated by
successfully processing the implied meaning evoked by the
(L)H*L% rise-fall contour on the critical word (hereafter CW).
The results showed that participants are able to associate the
implication contour with a contrast meaning, and that they
start doing this only after the H* peak of the rise-fall
intonation movement has been processed, hence when part of
the L% falling movement has been perceived.
Index Terms: intonation processing, eye-tracking,
implication contour, French
1. Introduction
Intonation is a tool to express pragmatic meaning and
information structure. In English, for instance, the L+H* pitch
accent can signal a contrastively focused element, for which a
speaker selects an element within a set of possible alternatives
[1]. The intonational structure of French differs widely from
that of English: the domain of stress is larger than the word,
being the Accentual Phrase (AP) [2], with prosodic
prominence marking the right edge of the AP. Hence, the
presence of prosodic prominence on a specific element does
not necessarily indicate that that element is contrastively
focused. Instead, other intonation cues like an initial rise (LHi)
on the left edge of the focused constituent or post-focal pitch
range compression is probabilistically employed for
contrastive focus use (e.g. [2]–[4]). Moreover, in French there
is no specific pitch accent indicating that an element is
contrastively focused. Instead, several intonation contours
with various degrees of speaker commitment and attitude
attribution are found to occur in contrastively focused
contexts: a rising H*H%, found in confirmation questions, a
rise-fall-rise H+!H*H%, indicating disbelief on the part of the
speaker relative to her interlocutor’s proposition, or a rise-fall
(L)H*L%
1
, called the “implication contour”, found when there
is a contrast between the interlocutors’ beliefs [5], [6].
Online processing of intonation in contrastively focused
elements has been studied mainly in stress-accent languages
such as English or German, revealing that speakers use pitch
accent location and pitch accent type to create online
expectations about an upcoming referent [7]–[12]. In [9],
participants first heard a sentence like “Put the candy/candle
below the triangle” and then a sentence like “Now put the
candle above the square”, where ‘candle’ could either be
accented or deaccented. They analyzed participants’ looks to
the competitor ‘candy’ when hearing the second sentence and
found that listeners interpreted the noun as anaphoric when it
was deaccented and as non-anaphoric when accented.
[13] found that speakers rely on the placement and type of
pitch accent to infer contrast in American English. Participants
heard a sentence like “Hang the green drum” that was
followed by a sentence like “Now hang the BLUE drum” or
“Now hang the blue BALL”. Their results showed that in the
first case, fixations were speeded to the target image if a L+H*
accent was placed on the adjective while this was not the case
when the same pitch accent was placed on the noun.
Interestingly, in sentences like “Now hang the BLUE ball”,
participants expected the target noun to be “drum” and not
“ball” when they heard a contrastive accent on the adjective,
and produced eye fixations on a set of drums even when
segmental cues had already disambiguated the target.
As for French, intonational meaning has not yet been
explored with online techniques. Moreover, since the contour
under investigation is composed of a rise, which is quite
frequent in AP-final position in French, and a subsequent fall,
we also ask whether listeners can infer the contrast meaning
once the pitch accent rise has been perceived or if they have
need to process the following L% [see 11 for English]. The
present study employed the Visual World eye-tracking
1
Note that we employ a transcription for the implication contour that
is more transparent, since it includes the preceding L(ow) target,
though we include it in parenthesis to indicate that its contrastive
status is still not known

paradigm [14] to investigate if French speakers associate the
LH*L% rise-fall “implication” contour to a specific meaning,
i.e. a contrast between the interlocutors’ beliefs, and if so, at
which point during the contour listeners would be able to
extract the contrastive meaning.
2. Methods
2.1. Participants
Twenty-nine French-speakers were recruited in the Paris
area (7 males). Four additional participants took part in the
study but were excluded due to experimental errors (N=2) or
to exclusive fixations to the center of the screen (N=2).
2.2. Materials
Eighteen test suggestion-response sentence pairs were
created to evoke a dialogue in a card game in which players
guess which cards the other player holds. The critical stimuli
pair had the form of Je pense que tu as un/e X (‘I think you
have a/an X’) – J’ai un/e X (‘I have a/an X’). All suggestions
were produced with a falling LHiL*L% intonation and
included a homophone in phrase-final position, presented with
a visual display that depicted the subordinate alternative (Fig.
1, left panel).
Figure 1: Visual displays in suggestions (left) and
responses (right).
Condition
Target
Competitor (1)
Competitor (2)
Unrel. Dis.
Confirmation-
homophone
Cane
‘Duck’
Canne
‘Stick’
-
Poupée
‘Doll’
Cacahouète
‘Peanut’
Contrast-
homophone
Canne
‘Stick’
Cane
‘Duck’
-
Poupée
‘Doll’
Cacahouète
‘Peanut’
Contrast-
control
Poupée
‘Doll’
Cane
‘Duck’
Canne
‘Stick’
Cacahouète
‘Peanut’
Table 1: Summary with an example of the images
corresponding to each position across conditions.
Critical responses were of three types (Table 1): (a)
confirmation-homophone sentences, produced with a
HiL*L%, including the same homophone as in the suggestion
(e.g., cane ‘duck’) and followed by the segmental
disambiguation bien sûr plus a clarifying phrase, e.g., l’animal
(‘indeed, the animal’); (b) contrast-homophone sentences,
produced with a LH*L%, followed by the segmental
disambiguation plutôt, and clarification to the dominant
alternative of the homophone pair (‘… a cane, instead, for
walking’); (c) contrast-control sentences, produced with a
LH*L% and followed by the segmental disambiguation plutôt,
[…] (‘instead, […]’), but including a non-homophone CW.
Fig. 2 shows the spectrograms of the three possible responses.
In scenes accompanying test responses (Fig. 1, right
panel), the image at the bottom left always coincided with the
suggested word, while counterbalanced in the other positions
were the alternative interpretation of the homophone, the
target non-homophone of the contrast-control sentences, and a
non-homophone unrelated distractor.
Thirty-six filler sentence pairs were also created, with the
same form as test sentence pairs but including non-
homophones in the critical position. There were two types of
filler pairs: (a) filler-confirmation pairs (N=21), with a non-
homophone in the suggestion and the same non-homophone in
the response, produced with a HiL*L% confirmation
intonation; (b) filler-contrast pairs (N=15), with a non-
homophone in the suggestion and a different non-homophone
in the response, produced with a LH*L% contrast intonation.
All intonation contours were felicitous: 27 trials in the
experiment used confirmation intonation for correct guesses,
and 27 trials used contrast intonation for incorrect guesses.
Figure 2: Spectrograms of the responses in the
confirmation-homophone condition (top), the contrast-
homophone condition (middle), and the contrast-
control condition (bottom). In the top panel, the F0
rise at the end of bien sûr is a Praat error.
2.3. Procedure
Participants’ eye fixations were tracked by an Eyelink II
Eye-tracker. First, participants were told a story about a girl
who had to make a guess about the other girl’s cards, and then
the other girl either confirmed or contradicted that guess.
Three practice trials preceded the test phase. The test phase
consisted of 18 test trials (6 per test condition) and 36 filler
trials (21 filler-confirmation and 15 filler-contrast) presented
in randomized order. Test conditions were counterbalanced
across three presentation lists using a Latin-square design.
Participants were randomly assigned to one of the three lists.
2.4. Predictions
We predicted that before the CW, participants would look
equally to the four images because no prosodic or segmental
cues would reveal the intended target (although we anticipated
that there could be some bias for or against the repeated
image, i.e. the duck in Fig. 1). As the beginning of the CW
unfolded, we expected that in the confirmation- and contrast-

homophone conditions participants would look equally at the
two homophone images since no segmental cues to
disambiguate the target nor nuclear tonal ones would have
been perceived yet. However, in the contrast-control condition
participants were predicted to look less at the suggested image
because they would have already perceived segmental
disambiguation. We further predicted that once the entire tonal
configuration had been perceived by the end of the CW, the
two homophone conditions would diverge: participants would
look more at the suggested image if a confirmation intonation
is presented, but less at it if they heard a contrast intonation.
Finally, we expected that during the segmental clarifying
phrase, the confirmation condition (but not the other ones)
would elicit more looks to the suggested image.
3. Results
Fig. 3 plots fixations to the image depicting the suggested
word across the 3 conditions, averaged across participants.
The x-axis shows time (ms) from the onset of the CW, and the
grey region indicates a 200ms offset to account for saccade
planning. Note that looks at the suggested image were similar
across conditions before the CW (before the grey region). At
the beginning of the CW, and so as soon as participants
perceived that in the contrast-control condition there was a
small portion of the segmental disambiguation, their looks at
the suggested image decreased sharply. If we compare the two
homophone conditions, we observe that during the CW
participants increased their looks at the suggested image even
if the intonation signaled contrast. However, shortly before the
end of the CW and coinciding with the presence of the H*
peak in the conditions with contrast intonation (dotted vertical
line in Fig. 3), looks at the suggested image decreased (this
effect being stronger after the CW, coinciding with the region
in which the L% fall is realized).
For the statistical analyses we calculated the proportion of
fixations (out of trials with valid fixations) on each of the four
picture elements, by time steps of 20 ms. We then aggregated
the time segments into larger time regions of interest by
counting the number of 20ms time slots in which a fixation to
the suggested image, the image depicting the homophone
competitor, or to the two non-homophones (one of which was
the target in the contrast-control condition) occurred. Five
regions of interest were analyzed: (1) the region prior to the
CW, (2) the region within the CW preceding the H* peak, (3)
the 100 ms window where the H* peak was perceived (or not
perceived in the HiL*L% condition), (4) the region following
the presence or absence of the H* peak within the CW, and (5)
the first 100 ms of the segmental disambiguating region. Time
is aligned with the beginning of the CW, and since saccade
planning takes approximately 200ms, the effects based on the
CW should start at 200ms. Also, CWs had different durations
and therefore each regions of interest was calculated as a
function of that.
We then calculated logodds of looks to the image
depicting the suggested word vs. looks to the other three areas
of interest in the visual scene. Using this dependent measure
we fit linear mixed-effects models using the lmer function of
the R package lme4. Participants and items were treated as
random effects to accommodate by-subject and by-item
variation in one model (Baayen, 2008; Barr, 2008; Jaeger,
2008). Condition (confirmation-homophone vs. contrast-
homophone vs. contrast-control was included as a Helmert
coded predictor to allow comparison of the contrast-control
level with the other two levels, and the two homophone
conditions to each other. First we present the results for the
comparison between the two homophone conditions
(confirmation-homophone and contrast-homophone) vs. the
contrast-control condition in each region of interest. Second,
we report the results of the comparison between the two
homophone conditions in each region of interest. Table 2
shows ß, t and p values for all comparisons. We determined p
values by chi-square tests from nested model comparisons.
For the first comparison, we observed that there was no
effect of condition in the region preceding the CW (t = .034, p
= .973), but that once the CW started and in all of the
following regions of interest there was an effect of condition
(beginning of the CW: t = -2.006, p < .05; F0 peak: t = -6.148,
p < .001; post-peak within the CW: t = -7.888, p < .001; after
the CW: t = -7.786, p < .001). This shows that participants
looked less at the suggested image in the control condition
than in the other conditions, due to the presence of
disambiguating segmental material, and that this effect began
when the CW started.
For the comparison between conditions with a homophone
in the CW, we observed no effect of condition before the CW
(t = .172, p = 0.863), with equal looks to all images at this
point. Once the CW started and during the region that
preceded the H* peak, we observed a marginal effect of
condition (t = 1.852, p = .06), with an unexpected bias to look
at the suggested image in the contrast LH*L% homophone
condition. This early bias became fully significant in the
following region (when the H* peak was presented) (t = 1.945
p < .05). Notably, after the H* peak had been perceived and
during the following L% fall (i.e. during the rest of the CW),
Figure 3: Proportion of fixations to the suggested image across conditions. The grey region depicts looks when
hearing the critical word, and the dotted vertical line indicates looks when perceiving the H target location.
-0,1
6E-16
0,1
0,2
0,3
0,4
0,5
0,6
0,7
-100
-40
20
80
140
200
260
320
380
440
500
560
620
680
740
800
860
920
980
1040
1100
1160
1220
1280
1340
1400
1460
1520
1580
1640
1700
1760
1820
1880
1940
2000
% Fixations
Time (ms)
Looks to image depicting the previously suggested homophone
L*L
%

this bias disappeared (t = 0.613, p = .506), revealing that in the
contrast LH*L% homophone condition participants started
looking less at the suggested image in response to the falling
contour composed of the H* peak and the following fall, while
within the preceding LH* rising portion this was not the case.
Estimate
SE
t-value
p-value
Pre-CW
Intercept
-0.4373
0.0543
-8.046
Homoph. vs Control.
0.0024
0.0721
0.034
.973
Confirmation vs Contrast
0.0142
0.0829
0.172
.863
Beginning of CW
Intercept
-0.5973
0.0733
-8.145
Homoph. vs Control.
-0.2282
0.1138
-2.006
.045*
Confirmation vs Contrast
0.2432
0.131
1.856
.063•
F0 peak
Intercept
-0.5659
0.1113
-5.086
Homoph. vs Control.
-0.7322
0.1191
-6.148
.000***
Confirmation vs Contrast
0.2758
0.1418
1.945
.043*
Post-peak within CW
Intercept
-0.6377
0.1025
-6.217
Homoph. vs Control.
-0.6378
0.1024
-6.228
.000***
Confirmation vs Contrast
0.0878
0.1432
0.613
.506
After CW
Intercept
-0.6613
0.0970
-6.814
Homoph. vs Control.
-0.9078
0.1166
-7.786
.000***
Confirmation vs Contrast
0.0661
0.1341
0.493
.622
Table 2: Estimates, Standard Errors, t values and p values for
all comparisons in all regions of interest.
4. Discussion
The purpose of this study was to investigate if French
speakers associate the “implication contour” with the specific
meaning of a contrast between the interlocutors’ beliefs, and at
which point within the nuclear contour they have fully
processed this meaning. In other words, we ask whether
listeners can anticipate the contrast by perceiving the LH*
rising portion of the contour or if they have to wait until the
L% falling tone is processed. The results of an eye-tracking
task revealed three main findings. First, and as expected, as
soon as participants perceived segmental cues identifying the
CW they stopped looking at the image depicting the
interlocutor’s suggested word if this was not the target.
Second, there was an early bias for looks to the suggested
image in the contrast-homophone condition. All three critical
conditions presented a homophone as the suggested image,
and the contrast-homophone condition employed the same
contour as the contrast-control condition, yet the difference in
looks begins prior to the segmental divergence between the
two contrast conditions (see Fig. 3). This finding was
unexpected, especially for the region prior to the onset of the
CW. Third, and importantly for our study, participants’ bias to
look at the suggested image decreased in the contrast
homophone condition but only after the H* peak had been
realized, revealing a rapid effect of intonation in reducing the
bias to the suggested image and in supporting a contrastive
interpretation.
In the contrast-homophone condition we saw that
participants increased looks to the suggested image when only
pre-nuclear material was available, and that this effect
continued as the CW unfolded (reaching full significance in
the comparison of the two homophone conditions at the F0
peak). Note that the LH*L% implication contour might be
more acoustically (hence perceptually) salient in terms of its
stronger F0 rise before the L% fall. The increase in perceptual
salience might be the basis of increased participants’ attention
(hence increased looks). Another possible explanation might
be related to the pre-nuclear segment in the rise-fall LH*L%
contour. While the fall HiL*L% contour begins in an H tone,
the rise-fall LH*L% contour begins with an L tone. Until the
beginning of the rise, the LH*L% contour is similar to another
intonation contour of the French inventory, the rising LH*H%
contour (although the rise starts earlier in the LH*L% case)
[15]. This rising LH*H% contour is used for continuation
declaratives, a pragmatic meaning that is very frequently
found in French speech. It could be that participants processed
the first part of the rise-fall LH*L% ‘implication’ contour as
indicating a continuation declarative and not as a contrastive
meaning, since the first meaning is more frequent than the
second one in French. Future analyses will also try to correlate
these findings with participants’ individual differences.
Finally, participants used the prosodic information following
the H* peak in the implication contour to extract the
contrastive meaning. The results show that the change in looks
began only after the H* peak had been processed and in the
tonal region where the L% falling tonal movement begins,
suggesting that contrast meaning processing in French may
require both the H* peak and its subsequent L% fall, or at least
part of it, to have unfolded.
Similarly, [11] found that participants needed to perceive a
L-H% boundary tone to derive the meaning of implied contrast
in English. On the other hand, [7], using a design that made
the implied meaning more easily predictable from the presence
of an L+H* accent, found effects prior to the occurrence of the
boundary tone. Since in French both the H* peak and the fall
occur during the last prominent syllable for the implication
contour, it can be difficult to tease apart whether participants’
decisions were driven by the pitch accent alone or by a
combination of the pitch accent and the boundary tone. This is
especially true since there are open questions about how much
processing time is required to build an implied meaning
subsequent to perceiving the acoustic evidence that evokes it.
An additional complication relates to how quickly an
intonational category can be identified. In a corpus-based
analysis of French rise-fall and rise intonation movements,
[15] found that some rise-falls can be confused with rises due
to delayed H* peak alignment. Our implication contour rise-
fall might contain some instances of delayed H* peak, causing
then some confusion in meaning processing.
In sum, results show that listeners associated the (L)H*L%
French implication contour with a contrast between speakers’
beliefs, and that they started doing this form-meaning mapping
immediately following the H* peak of the rise-fall movement.
Many questions still remain to be answered, especially since
most of the research on the online processing of intonation has
been carried out with stress-accent languages like English,
German or Dutch which are typologically very different from
French. However, we believe the present findings will
contribute to a better understanding of some of the cognitive
processes and individual differences involved in intonation
processing.
5. Acknowledgements
We thank Céline Pozniak for her help running the
experiment. This work, carried out within the Labex BLRI
(ANR-11-LABX-0036), has benefited from support from the
French government, managed by the French National Agency
for Research (ANR), under the project title Investments of the
Future A*MIDEX (ANR-11-IDEX-0001-02).

6. References
[1] J. Pierrehumbert and J. Hirschberg, “The meaning of
intonational contours i n interpretation of discourse,” in
Intentions in Communication, P. Cohen, J. Morgan, and M.
Pollack, Eds. Cambridge, USA: MIT Press, 1990.
[2] S. Jun and C. Fougeron, “A Phonological Model of French
Intonation,” .
[3] J. S. German and M. D’Imperio, “The Status of the Initial
Rise as a Marker of Focus in French,” Lang. Speech, 2015.
[4] A. Di Cristo, “Vers une modélisation de l’accentuation du
français (seconde partie),” J. French Lang. Stud., vol. 10,
pp. 27–44, 2000.
[5] C. Portes, C. Beyssade, A. Michelas, J.-M. Marandin, and
M. Champagne-Lavau, “The dialogical dimension of
intonational meaning : Evidence from French,” J. Pragmat.,
vol. 74, pp. 15–29, 2014.
[6] C. Portes and U. Reyle, “The meaning of French
‘implication’ contour in conversation,” in Speech Prosody,
2014.
[7] C. Kurumada, M. Brown, S. Bibyk, D. F. Pontillo, and M.
K. Tanenhaus, “Is it or isn’t it: Listeners make rapid use of
prosody to infer speaker meanings.,” Cognition, vol. 133,
no. 2, pp. 335–342, 2014.
[8] A. Weber, B . Braun, and M. W. Crocker, “Finding referents
in time: eye-tracking evidence for the role of contrastive
accents.,” Lang. Speech, vol. 49, no. 681, pp. 367–392,
2006.
[9] D. Dahan, M. K. Tanenhaus, and C. G . Chambersc, “Accent
and reference resolution in spoken-language
comprehension,” J. Mem. Lang., vol. 47, pp. 292–314, 2002.
[10] D. G. Watson, M. K. Tanenhaus, and C. a Gunlogson,
“Interpreting Pitch Accents in Online Comprehension: H*
vs. L+H*.,” Cogn. Sci., vol. 32, no. 7, pp. 1232–44, Oct.
2008.
[11] H. Y. Dennison and A. J. Schafer, “Online construction of
implicature through contrastive prosody,” in Speech
Prosody, 2010, pp. 1–4.
[12] W. F. Heeren, S. A. Bibyk, C. Gunlogson, and M. K.
Tanenhaus, “Asking or Telling – Real-time Processing of
Prosodically Distinguished Questions and Statements,”
Lang. Speech, pp. 1–28, 2015.
[13] K. Ito and S. R. Speer, “Anticipatory effects of intonation:
Eye movements during instructed visual search,” J. Mem.
Lang., vol. 58, no. 2, pp. 541–573, 2008.
[14] M. K. Tanenhaus and M. J. Spivey-Knowlton, “Integration
of visual and linguistic information in spoken language
comprehension,” Science (80-. )., vol. 268, no. 5217, p.
1632, 1995.
[15] C. Portes and L. Lancia, “A corpus based investigation of
the contrast between French rise-fall and rise via wavelet
based functional mixed models,” in Laboratory Phonology,
2012.