Studying the interplay of context and semantic content in the interpretation of adversative
conjunctions with eyetracking

Ghyslain Cantin-Savoie, Denis Foucambert, & Grégoire Winterstein*

Abstract. The French adversative connective mais, much like its English counterpart
but, takes two conjuncts and indicates that they stand in some kind of opposition. The
nature of this opposition is often discussed in the existing literature of adversatives.
Using an eyetracking experiment, we look at where and when this opposition appears
to be manifested in a sentence reading task, using superiority comparatives sentences
that appear degraded without a supportive context, as well as inferiority comparatives
that do not seem to require such specific contextual information. We presented par-
ticipants (n = 28) with four types of sentences, differing in the form of comparative,
and in the information provided by the context. Some contexts contained a pivot prop-
erty to help readers access the opposition of the two conjuncts, while others were
neutral in that regard. Eye movements were recorded with a 250Hz Tobii Pro Fusion
eyetracker, and mixed effect models were used to analyze the following eyetracking
metrics : total fixation time and regression probability for the whole sentences, as well
as first-fixation duration, first-gaze duration, regression-path duration, regressions-out
and regressions-in for each individual words. Even if the helping context is shown to
lower negative acceptability judgment on sentences with mais plus (Winterstein et al.
2014), we found no context effect in online reading processing, finding instead a per-
sisting effect of the less/more dichotomy in all chosen measures, sometimes before
fixation of those words, pointing towards a parafoveal effect of the aforementionned
dichotomy, which merits closer look in future work.
Keywords. Semantics; psycholinguistics; eyetracking; reading; adversatives

1. Introduction. This work is centred on the semantics of the French adversative connective mais,
typically rendered in English as but. The semantics of adversative connectives has been the topic
of much work, in which two main theoretical branches can be distinguished. Those branches can
be characterized by the use of adversatives that they treat as canonical (Winterstein 2017).

On one hand, the contrastive approach considers that so-called contrastive uses of connectives
like but exemplify best the core constraint indicated by adversatives (see a.o. Umbach 2005). What
seems to be required by but is that its conjuncts involve predicates that are comparable, i.e. similar
in some respect and dissimilar in another.

(1) Contrastive John is tall but Richard is small.

On the other hand, inferential approaches take argumentative examples like (2) to be more repre-
sentative of the semantics of adversatives (see a.o. Blakemore 2002, Winterstein 2012). Crucially,

*The authors would like to thank the audience of ELM3 for useful comments. This work was supported by a
FRQSC Soutien à la recherche pour la relève professorale grant (2022-NP-296699). Authors: Ghyslain Cantin-
Savoie, Université du Québec à Montréak (cantin-savoie.ghyslain@courrier.uqam.ca), Denis Foucambert, Université
du Québec à Montréal (foucambert.denis@uqam.ca) & Grégoire Winterstein, Université du Québec à Montréal (win-
terstein.gregoire@uqam.ca).

Proceedings of ELM 3: 75-87, 2025

c©2025 Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein
Published by the LSA with permission of the author(s) under a CC BY license. 75

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


the interpretation of such examples involves some additional inference, called a pivot, that connects
the content of both conjuncts to a third proposition. Those two connections must be dissimilar in
some ways.

(2) Argumentative He’s unattractive but he’s kind so yes, I will marry him.

In section 2, we discuss these approaches in more detail, highlighting a point of contention between
them : the external or internal locus of the dissimilitude effect necessary for felicitous use of
adversatives. We also show in that section that comparatives like the one in (3) allow us to control
whether the required dissimilitude is internal to the conjuncts’ semantics (like it’s the case with
less tall) or if an external proposition is needed to introduce it (like it’s the case for taller).

(3) John is tall but less tall/*taller than Richard.

We then go on to describe a previous study by Winterstein et al. (2014) who found that acceptability
judgment discrepancy between negative (less tall) or positive (taller) comparisons can be alleviated
with the introduction of an external proposition, whilst reading-time differences between positive
and negative comparisons persist.

This reading-time difference is what inspired our present eyetracking study, where we look for
an online reading effect of the positive/negative comparison dichotomy (that we will call valence
from now on) and where we see if this effect persists through the introduction of a preceding
context made to help (or not) attaining required dissimilitude between conjuncts with an external
proposition. Our hypothesis is that there is a valence effect and that there is no effect of context
type on reading time. In section 3, we describe the item-making, participant recruitment, and
material-installing processes we went into to design an eyetracking experiment with the goal of
testing the aforementioned hypothesis. Filters, data transformation, eyetracking measures and
statistical models are described in section 4. The results of those analysis are shown and discussed
in 5.

2. Empirical background: adversatives and comparatives.

2.1. CONTRASTIVE AND ARGUMENTATIVE CASES. Let us start with a closer examination of
examples (1) and (2).

In the contrastive case of (1), John and Richard refer to distinct individuals, who possess mu-
tually exclusive properties (i.e. being small and being tall). In (2), the argumentative case oppose
two properties that are not mutually exclusive (i.e. being unattractive and being kind) predicated of
a single individual (the referent of he). Though the examples differ in many ways, the one differ-
ence relevant to us is the locus of opposition, i.e. the question of determining from where stems the
opposition that is required to license the use of an adversative like but as connective.

For the contrastive use, the locus of opposition is internal to the conjuncts, i.e. supposed to be
supported by the lexical semantics of the material in the conjuncts. This means that for a contrastive
sentence with but to be felicitous, the second conjunct needs to deny a stated or implied proposition
that is brought by the first conjunct itself. In the sentence in (1), John is tall brings an answer to
the question WHO is tall, therefore implying that individuals, like John, can be in the set of tall
people. It also implies the alternative possibility that other people, like Richard, are also part of
the tall people set. This said eventuality is denied by the second conjunct, Richard is small, which

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 76

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


means that the use of but in this context is felicitous (see Umbach 2005 for detailed account along
those lines, crucially exploiting QUD information to account for such cases).

For the argumentative use of but, the locus of opposition is necessarily external, which means
it will be found in the way both conjuncts are linked to a third, contextually given, proposition. As
an example, in (4), He’s unattractive seems to go against I will marry him (indicated with a red
arrow), while he’s kind does not (indicated with a green arrow).

(4) butHe’s unattractive he’s kind

I will marry him

Even if he’s unattractive and he’s kind are not necessarily opposed, it is through the opposing kind
of connections to an external proposition that felicitous use of but is licensed.

2.2. COMPARATIVES CASES. In this work we do not directly engage with the question of deter-
mining which of the two uses of adversatives just described is the most canonical (see Winterstein
2012 for a discussion). We will however assume that some pivot inference is involved in the inter-
pretation of adversatives, even in contrastive cases. To see why, we now introduce the core cases
of this work, which we refer to as comparative cases, are exemplified in (5).

(5) a. John is tall, but less tall than Richard.
b. #John is tall, but taller than Richard.

Comparative cases appear similar to contrastive ones, given how both involve the comparison
of two individuals along some property, like in (5). However, the property needs to be gradable,
which is not a necessity for contrastive uses.

Tall being a gradable adjective, truth-evaluation needs a comparison of some kind. So, when
it is said that John is tall, it inherently implies that John is taller than some threshold (Kennedy &
McNally 2005). Following the method described in 2.1, John is tall answer the question ”WHO is
less tall than John?” by telling us that someone is in the set of things that are less tall than John ;
If there is no one who is less tall than John, it cannot be said that John is tall. Many people could
possibly be in that set, including Richard. The second conjunct of (5-a), [John is] less tall than
Richard denies the inclusion of Richard in the aforementioned set, predicting a correct use of but.
Since [John is] taller than Richard does not deny this implication, an internal-opposition-locus
view of comparative uses of but seems to correctly predict the degraded nature of (5-b).

When showed sentences like those in (5) in French, people judged sentences like (5-b) to be
infelicitous (Winterstein et al. 2014). Reading times for sentences like those in (5-b) were also
shown to be longer compared to (5-a). In the same experiment, another group of participants saw
the same sentences presented after a context that introduces an external opposition pivot in the
form of an equality condition, like in (6-a).

(6) a. Context: We are looking for a stunt double for Richard, an actor, in action scenes for a
movie. The stunt double needs to be exactly the same height as Richard so that it goes
unnoticed. The search is hard, because Richard is tall.

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 77

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


b. Sentence: John is tall, but less tall/taller than Richard.

c. Opposition: butJohn is tall less tall/taller than Richard

John is the stunt double we are looking for

When preceded by contexts like (6-a), sentences like (5-b) were not judged to be infelicitous any-
more. However, reading time differences persisted, suggesting that the presentation of a context
was not enough to facilitate the interpretation of the adversative connective. This suggests that
those examples are somehow on a par with argumentative cases in that their interpretation rests on
the identification of a contextual pivot. The reading times data also suggest that this identification
comes at a later stage of interpretation.

Based on results obtained with the experiment of Winterstein et al. (2014), we hypothesize that
if we show sentences like (5) to participants, reading times will be greater on sentences with plus
(more) like in (5-b) than the reading times on sentences with moins (less) like in (5-a). Duration-
based eye-tracking measures, described below in 4, should be sufficient to deny or confirm this
hypothesis. Secondly, since we think that the locus of opposition is internal, lack of opposition
internal to the conjuncts cause by positive comparison with more should create a problem in online
processes, regardless of context manipulations.1 Therefore, our second hypothesis is that if we
preceded sentences like those in (5) with helping contexts like in (6-a), the reading time effect
of the less/more dichotomy will persist. To test this, we created another type of context, called
”neutral”, that are similar to the one in (6-a) but does not introduce an equality condition that
allows an external opposition of conjuncts. If we are right, we should not see an effect of the
nature of context on the eyetracking measures we observe.

3. Method. To test these hypotheses, we ran an experiment in which participants had to read
comparative sentences with mais preceded by contexts introducing (or not) an equality condition.
The eye movements of the participants were recorded using an eyetracking device. This section
describe the sentences that were shown, the hardware used, participants of the experiments, and
the experimental protocol.

3.1. EXPERIMENTAL ITEMS. Each item had 3 parts : (1) a context, (2) a target sentence, and (3)
a question. There were 60 items in total, of which 40 were distractors, meaning that the sentences
were not comparatives with mais, so that subjects would not easily detect the objective of the
experiment. 10 items and 30 distractors were extracted from Winterstein et al. (2014) and slightly
modified to be adapted to Quebec French, the rest of the items were created for this experiment.

The experiment used a 2x2 design, with the two binary conditions as follows.
First, contexts from experimental items came in two versions (condition Context). One ver-

sion introduced an equality condition like the one shown in (6-a), and was called helping. The other
version presented information irrelevant to the interpretation of the adversative, and is referred to
as the neutral version. We illustrate these version in (7).

1This predict that for uses where we think opposition is external like argumentative cases in (2), context manipula-
tion should have reading time effects. Another experiment will be needed to look into this.

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 78

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


(7) Jeanne Doe est amnésique et ne se rappelle plus qui elle est. Les enquêteurs pensent qu’elle
est la soeur jumelle/ petite soeur de Marc Leclerc, un vieil avocat porté disparu il y a 10 ans
de cela. Pour s’en assurer, ils font passer à Jeanne une batterie de tests.
Translation : Jane Doe is amnesiac and doesn’t remember who she is. Investigators think
that she is the twin / little sister of Marc Leclerc, an old lawyer that disappeared 10 years ago.
To be certain, they make Jane go through a battery of tests.

Each context was modified or created so that the total number of characters in the context would
be inside a 20 character range, between 222 and 242 characters.

Second, target sentences in the experimental items were comparative uses of mais. Each
sentence existed in two versions (condition Valence), depending on the kind of comparative
being used: either a superiority comparative (plus X que / ’more X than’) or inferiority one (moins
X que / ’more X than’). Crucially, each version only differs from the other in the choice of the
comparative adverb. These two versions are shown in (8).

(8) Selon les tests, Jeanne est vieille mais moins/plus vieille que Marc même s’ils ont tous les
deux les cheveux blancs.
Translation : According to the tests, Jane is old, but less old/older than Marc even if they
both have white hair.

The target sentences differ from previously shown examples in (5) in that they are surrounded by
complement phrases, so that critical zones of interest (i.e. around the adversative connective) are
not shown either at the extreme left or the extreme right of the screen when read. Character number
for the target sentences varied between 108 and 128.

Questions were introduced as a way to distract participants into thinking the answer was the
criterion we evaluate, and to determine whether participants were really reading sentences or only
pressing space while moving their eyes. Questions were between 48 and 68 characters, each one
asking for a yes/no answer that indicated comprehension of the previously read context-sentence
pair, like shown in (9) about the sentence in (8) and the context in (7)

(9) Est-il possible que Jeanne Doe soit la soeur jumelle de Marc Leclerc?
Translation : Is it possible that Jane Doe is the twin sister of Marc Leclerc?

Every part of an item, context, sentence or question, had a mean lexical frequency maximum of
737.84 , evaluated as the maximum threshold of the interquartile range. Lexical frequency for
each word was found using Lexique3 database trained on the Frantext corpus, which contains 218
French novels from 1950 to 2000 (New et al. 2004).

3.2. EQUIPMENT. Subjects sat on an adjustable desk-chair so that their heads, regardless of their
individual height, would be at the same spot. Their chin rested on a table model chin-rest, as they
were looking at an Asus TUF Gaming VG259Q computer screen of size 24.5” with a colour depth
of 8-bit, resolution of 1920 per 1080 and a refresh rate of 59.94Hz, placed 60 centimetres away
from their face. Their eye movements were observed with a Tobii Pro Fusion 250 Hz eyetracker.

3.3. PARTICIPANTS. This article analyses the data of 28 participants, 13 male and 15 female, aged
between 19 and 68, with a mean age of 31.5 years old. We invited them by going in their university
class and sending Facebook calls to student groups and friends. Participants were compensated 15

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 79

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


CAD for their participation. Participants were selected to have only Quebec French as their first
language, and to have a normal or corrected vision. Their highest completed levels of education
ranged from high school to PhD.

3.4. PROCEDURE. After arriving and signing a consent form, participants were placed in front of
a screen, rested their chins on the chin-rest, underwent calibration, For which they had to follow a
moving dot displayed on the screen. They then completed a first test item after which they were
given the chance to ask questions before proceeding with the rest of the experimental items.

They were warned that before each item part, a fixation cross would appear and they would
be asked to look directly at it before pressing the space bar to go to the next stimulus. They were
asked to read each item part before pressing space, except after the question part, where they were
asked to answer yes by pressing A or no by pressing L.

To instantiate the experiment, we used the Tobii Pro Lab software, which unfortunately does
not offer latin-square style pseudo-randomization. We thus created four different trial groups, to
ensure that contexts and sentences where shown so that every participants would see the same
number of each condition, and each group would not see the same items in the same conditions
than other groups, as shown in table 1.

Experimental Items Group 1 Group 2 Group 3 Group 4
1-5 helping|less neutral|less helping|more neutral|more

6-10 neutral|more helping|less neutral|less helping|more
11-15 helping|more neutral|more helping|less neutral|less
16-20 neutral|less helping|more neutral|more helping|less

Table 1: Condition repartition amongst participant groups

After showing group 1 items to a participant, the next participant would see group 2 items,
and so on to keep an even spread. After the experiment but before briefing, they would go through
a reading span test in order to evaluate their respective working memory2.

4. Analysis. Since 28 participants read 20 sentences each, we would expect 560 sentence-level
data points. However, item number 19 being one character too long, it showed on 2 lines on the
screen and was therefore removed. Our eyetracking device recorded valid gaze location data for
466 of the 532 remaining sentences. Fixation data was then retrieved using the Tobii-Pro I-VT
fixation filter (TobiiAB 2023). 5 sentences were removed for containing less than 4 fixations each,
resulting in a total of 461 sentence tokens. After removing fixation data on 10 words that were
fixated less than 60 ms each, we had fixation data for a total of 5703 word tokens. Data analysis
was carried out on two separate levels : sentences and words.

4.1. THE SENTENCE LEVEL. At the sentence level, we took total fixation duration and regression
probability as indicators of global sentence-reading difficulty (Goldberg & Kotval 1999). Total
fixation duration is defined as the sum, in ms, of the durations of fixations that happened in the
reading of each sentence-part of the item triads. Regression probability is the number of regressive
saccades made during the recording of a sentence, divided by its total saccade count. Each of those

2We don’t go in details in this article about this test’s procedure, it is described in Marcotte (2014).

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 80

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


measures were took as dependent variables (DV) in their own mixed effect model generated with
lmer() (Bates et al. 2015) in R, sharing the formula shown below in (10).

(10) DV ∼ Valence + Context + ScolCal + Span + (1|Participant) + (1|Media)

Valence is a categorical variable with two possible values : plus (’more’) or moins (’less’), that
represents whether the word following mais (’but’) in item sentences was plus (i.e. a superiority
comparative) or moins (i.e. an inferiority comparative). Our hypothesis is directly linked to this
variable.

Context is a categorical variable with two possible values, helping or neutral, that indicates
whether the context preceding the sentence contains a helping opposition pivot in the form of an
equality condition or not. Our hypothesis is also directly linked to that variable.

ScolCal is an ordinal variable whose values range from 1 to 4. It is used to take into account
the highest school level of participants, 1 for High School, 2 for Cegep 3, 3 for a bachelor’s degree
from university and 4 for masters, PhDs, and higher.

Span is a continuous variable representing the result of the Reading Span Test, to control for
working memory, which can affect several reading behaviours (Traxler et al. 2012), as well as other
potential individual cognitive abilities related to reading (Kim et al. 2021).

Participant and Media are factors with levels representing participant Id and sentence number
4, respectively. The way they are instantiated in the model (1|Variable) means that the intercepts,
but not the slopes, are random, to allow a middle ground between model complexity and statistical
power. They are interpreted by the model as cross-random effects giving us only 1 intraclass
correlation coefficient (ICC) per result table, which is acceptable since all of our participants saw
all types of sentences.

Note that in the model presented in (10) we did not account for interactions between indepen-
dent variables. Models testing for interactions between our two hypothesis-making independent
variables were compared to models without interactions. Since there were no significant differ-
ences between the two in terms of fitting, we chose the simpler version to avoid problems linked
to model complexity.

4.2. THE WORD LEVEL. At the word level, we considered five different measures. For each
word that was fixated at least once, we capture the duration of the first fixation (in ms). First
fixation duration is linked to very fast first order cognitive processes like word and letter recognition
(Rayner 1998). We also included first gaze duration, which is the sum of all fixation durations in
milliseconds on a given word from the first fixation until the subject fixates on something else.
87% of the time, according to our own data, first fixation duration and first gaze duration have the
same value, since many words are only fixated once before a saccade brings the gaze on another
one (Rayner 1998). First-gaze duration are said to be influenced by cognitive processes that are not
as fast as those influencing first fixation duration (Rayner 1998), and also by plausibility of words
in sentence context, which is not the case for first fixation duration (Inhoff 1984).

3Cegep (Collège d’Enseignement Général Et Professionnel, roughly translated to College of General and Profes-
sional Teaching) is a school level unique to Québec with students entering it at around 17 or 18 years old, created to
make higher education accessible and affordable to a wider population.

4As an example, the sentence about Jane Doe previously shown is always number 13, regardless of whether it is in
the plus or moins version.

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 81

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


Fixation # Jane is old but less old than Marc
1 X
2 X
3 X
4 X
5 X
6 X
7 X

Table 2: Fixation order of a hypothetical sentence reading

Regression-path duration for a word is the sum of all fixation durations, from the first one on
the word until a fixation is made on another word to the right of it. In Table 2, showing the ordering
of fixation on a hypothetical sentence reading, the regression-path duration of the word but would
be the sum duration of fixations 4, 5 and 6.

Regression-path duration is used to indicate processing difficulties oftentimes caused by pars-
ing effects, such as syntactic ambiguities (Hyönä et al. 2003).

For each word, we also calculated its probability of being the starting or ending point of an
ocular regression (regression-out, regressions-in).

Each of these measures were taken as the dependent variable of a lmer() model shown in (11),
similar to the one in (10).

(11) DV ∼ Valence + Context + ScolCal + Span + sFreqLex + nbCarac + (1|Participant) +
(1|Media)

The model in (11) and the choices leading to its form are identical to those described in section
4.1, with the exception of 2 additional fixed effects.

nbCarac contains integer numbers representing word length by character number, which is
said to influence if a word is going to be fixated on in the first place (Rayner 1998). sFreqLex
represents the lexical frequency of the word, taken from the Lexique3 database made with the
Frantext corpus, transformed with the scale() R function, to avoid scaling issues brought by the
large value range inherent to the frequency values.

We analyzed the words in two steps. First, each word-level measure took turns being inserted
as dependent variable in the model described above, allowing us to look for global effects of
valence and context on words. Secondly, we indexed words by position in the sentence, taking but
as 0, the word before it as -1, the word after it as 1, and so on. We then selected a segment of the
sentences we call critical zone going from indexes -3 to 4, where there is no variation in word type
per index. As an example, the word before mais (but) at position -1 is always the first iteration of
the gradable adjective. However, words at the position 7 can be articles, nouns or verbs without
any given link between them, so taking position as a category would be a mistake. 5

5. Results.

5Some part-of-speech or semantic related tagging could alleviate this limit, and would be interesting in subsequent
research.

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 82

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


5.1. SENTENCE LEVEL: RESULTS. Results at the sentence level in Table 3 and Table 4 show
that sentences are fixated longer and that participants are more likely to look back when reading
sentences with plus (’more’), which indicates that superiority comparatives increase sentence read-
ing difficulty. However, context type does not affect those metrics, as found by Winterstein et al.
(2014).

Table 3: Model results for the log transform of
total fixation duration of sentences

Table 4: Model results for regression proba-
bility of sentences

Sentence-level results, visualized in Fig. 1 and Fig. 2, confirm that Valence has an effect on
reading times. Since our question is about where and when it happens, we move to results at the
word level.

Figure 1: Regression probability on sentences,
split by experimental conditions

Figure 2: Total fixation duration in millisec-
onds on sentences, split by experimental con-
ditions

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 83

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


5.2. WORD LEVEL: RESULTS. At the word level, when we take all words into account, there is
no significant effect of our variables on either first fixation or first gaze duration. However, if we
test on individual word positions in the critical zone described in section 4.2, we find a significant
effect of Valence on first gaze duration on two word positions, shown on a white background on
Fig. 3.

Figure 3: First gaze duration per word position, white background = significant effect of Valence

The fact that first gaze duration is longer on those two words cannot be explained by saying
that characters in the words or their lengths are harder to decode, since if it were the case there
would also be a first fixation effect, which is not the case. Also, and most importantly, the effect
of the more/less dichotomy is seen on first gaze duration on the first occurrence of the gradable
adjective (’tall’), which is before the occurrence of more/less. Firstly, this indicates a limitation in
metrics definition : we don’t know if the word more/less was fixated before tall in the cases where
first gaze duration is higher. If it is not the case, it would indicate a parafoveal effect of Valence
on the reading of the word tall, which would need subsequent work to confirm.

If we were to hypothesize an effect of Context on a given metric, it would be on regression
path duration, since it is said to represent not only the time spent decoding words, but also the
time spent integrating visual level information with previously known words and facts stored in
memory, such as personal knowledge or information given in a preceding context (Cook & Wei
2019). However, there was no effect of Context on regression-path duration, not on all words
taken as a whole nor on each individual word of the critical zone. There was, however, a signif-
icant Valence effect on the individual word position for the second occurrence of the gradable
adjective, shown on Fig. 4 where the background is white.

Those results mean that when participants look at the second occurrence of the gradable ad-
jective, they take more time and have to go back more in the superiority comparative cases.

Since we showed previously a significant Valence effect on regression probability at the
sentence level, it is no surprise that regressions-out and regressions-in probability on individual
words also have a significant effect of Valence when taken on all words. When we look at the

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 84

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


Figure 4: Regression-path duration on word position, white background = significant effect of
Valence

critical zone, we can see that there are more regressions coming from and out of the first occurrence
of the gradable adjective, as shown in the white background of both Fig. 5 and Fig. 6.

Figure 5: Regressions-out probability for each
word position, white background indicating a
significant effect of Valence.

Figure 6: Regressions-in probability for each
word position, white background indicating a
significant effect of Valence.

These results are not indicative of the fact that participants are regressing from the first occur-
rence of tall to itself, and is therefore indicative of a limitation of our study: there is no chronolog-
ical information about saccade for now that allows us to know where the regressions from tall go,
and where they come from. This would requires subsequent analysis.

There is no Context effect on regressions-in and regressions-out probabilities when taking
all words together. However, looking at the critical zone allows us to see (Fig. 7 and 8) that more
regressions get out of the second named entity (Pierre in the example) when the context is neutral
to target the first named entity (Paul examples). Then again, subsequent analysis is needed to

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 85

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


discover if a significant number of saccades go from Pierre to Paul.

Figure 7: Regressions-out probability for each
word position, white background indicating a
significant effect of Context.

Figure 8: Regressions-in probability for each
word position, white background indicating a
significant effect of Context.

6. Conclusion and outlooks. This study tested hypotheses regarding the reading time increase of
comparative sentences with mais (but) comparing superiority and inferiority comparatives. The-
oretical work about the semantics of adversatives, as well as previous research using self-paced
reading data (Winterstein et al. 2014), led us to think that (1) sentences with plus (more) will have
higher reading times, and that (2) contexts introducing (or not) an external proposition that ex-
plicitly supports the opposition between conjuncts should not affect reading times. Results of an
eyetracking study confirmed these hypotheses, additionally allowing us to point out some limits
and future research avenues.

Sentences with a superiority comparative (plus ’more’) showed an increase in total fixation
time for whole sentences as well as a localized increase of first gaze duration and regression-path
duration, which seems to confirm our first hypothesis. Although context type had no effect on
duration-based metrics and therefore on reading times per say, confirming our second hypothesis,
results from regressions patterns showed the influence of context on where participants decide to
look back from and to. Context type seems to influence whether participants will regress from
and to named entities, whilst most of the localized duration effects were found on the gradable
adjectives.

Limitations to our analysis were made apparent during this study, including lack of tagging on
individual words that would have allowed us to categorize them by semantic function or syntactic
category. Further analysis of our data will allow us to inspect our finding that duration-based
metrics show valence effects on the gradable adjectives while location-based metrics show valence
effects on the named entities. A boundary paradigm study would be needed to confirm parafoveal
effects influenced by the semantics of plus/moins. Finally, another study with different items would
be key to investigate whether different kind of adversative sentences (not only comparatives) with
an external locus of opposition would have duration-based effects of context manipulation.

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 86

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/


References

Bates, Douglas, Martin Mächler, Ben Bolker & Steve Walker. 2015. Fitting linear mixed-effects
models using lme4. Journal of Statistical Software 67(1). 1–48. 10.18637/jss.v067.i01.

Blakemore, Diane. 2002. Relevance and Linguistic Meaning. The semantics and pragmatics of
discourse markers. Cambridge: Cambridge University Press.

Cook, Anne E & Wei Wei. 2019. What can eye movements tell us about higher level comprehen-
sion? Vision 3(3). 45.

Goldberg, Joseph H & Xerxes P Kotval. 1999. Computer interface evaluation using eye move-
ments: methods and constructs. International journal of industrial ergonomics 24(6). 631–
645.

Hyönä, Jukka, Robert F Lorch Jr & Mike Rinck. 2003. Eye movement measures to study global
text processing. In The mind’s eye, 313–334. Elsevier.

Inhoff, Albrecht Werner. 1984. Two stages of word processing during eye fixations in the reading
of prose. Journal of verbal learning and verbal behavior 23(5). 612–624.

Kennedy, Christopher & Louise McNally. 2005. Scale structure, degree modification, and the
semantics of gradable predicates. Language 345–381.

Kim, Young-Suk Grace, Yaacov Petscher & Christian Vorstius. 2021. The relations of online read-
ing processes (eye movements) with working memory, emergent literacy skills, and reading
proficiency. Scientific Studies of Reading 25(4). 351–369.

Marcotte, Sylvie. 2014. Étude en temps réel de la révision de la morphographie du nombre du
verbe chez les étudiants universitaires .

New, Boris, Christophe Pallier, Marc Brysbaert & Ludovic Ferrand. 2004. Lexique 2: A new
french lexical database. Behavior Research Methods, Instruments, & Computers 36(3). 516–
524.

Rayner, Keith. 1998. Eye-movements in reading and Information-Processing: 20 years of Re-
search. Psychological Bulletin 124(3). 372–422.

TobiiAB. 2023. Tobii pro lab. Computer software. http://www.tobii.com/.
Traxler, Matthew J, Debra L Long, Kristen M Tooley, Clinton L Johns, Megan Zirnstein & Eunike

Jonathan. 2012. Individual differences in eye-movements during reading: Working memory
and speed-of-processing effects. Journal of eye movement research 5(1).

Umbach, Carla. 2005. Contrast and information structure: A focus-based analysis of but. Linguis-
tics 43(1). 207–232.

Winterstein, Grégoire. 2012. What but-sentences argue for: a modern argumentative analysis of
but. Lingua 122(15). 1864–1885. 10.1016/j.lingua.2012.09.014.

Winterstein, Grégoire. 2017. Perspectives on Argumentation within Language. Theoretical, Pro-
cessing, Computational and Social aspects.: Université Paris Diderot–Paris 7 dissertation.
(Habilitation à diriger les recherches).

Winterstein, Grégoire, Emilia Ellsiepen, Jaques Jayes & Barbara Hemforth. 2014. Effects of con-
text on the processing of adversative and comparative constructions. In Cuny, .

Proceedings of ELM 3: 75-87, 2025

Ghyslain Cantin-Savoie, Denis Foucambert, and Grégoire Winterstein:
The interplay of context and semantic content in the interpretation of adversative conjunctions. 87

https://journals.linguisticsociety.org/proceedings/index.php/ELM/issue/archive
https://www.elm-conference.net/