Edited by: Marcela Pena, Catholic University of Chile, Chile
Reviewed by: Barbara C. Malt, Lehigh University, USA; Katharina Spalek, Humboldt-Universität zu Berlin, Germany; Alan Langus, Scuola Internazionale Superiore di Studi Avanzati, Italy
*Correspondence: Federica Cavicchio, Center for Mind/Brain Sciences, Università degli Studi di Trento, Corso Bettini 31, 38068 Rovereto, Trento, Italy e-mail:
This article was submitted to Language Sciences, a section of the journal Frontiers in Psychology.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Research using the visual world paradigm has demonstrated that visual input has a rapid effect on language interpretation tasks such as reference resolution and, conversely, that linguistic material—including verbs, prepositions and adjectives—can influence fixations to potential referents. More recent research has started to explore how this effect of linguistic input on fixations is mediated by properties of the visual stimulus, in particular by visual salience. In the present study we further explored the role of salience in the visual world paradigm manipulating language-driven salience and visual salience. Specifically, we tested how linguistic salience (i.e., the greater accessibility of linguistically introduced entities) and visual salience (bottom-up attention grabbing visual aspects) interact. We recorded participants' eye-movements during a MapTask, asking them to look from landmark to landmark displayed upon a map while hearing direction-giving instructions. The landmarks were of comparable size and color, except in the Visual Salience condition, in which one landmark had been made more visually salient. In the Linguistic Salience conditions, the instructions included references to an object not on the map. Response times and fixations were recorded. Visual Salience influenced the time course of fixations at both the beginning and the end of the trial but did not show a significant effect on response times. Linguistic Salience reduced response times and increased fixations to landmarks when they were associated to a Linguistic Salient entity not present itself on the map. When the target landmark was both visually and linguistically salient, it was fixated longer, but fixations were quicker when the target item was linguistically salient only. Our results suggest that the two types of salience work in parallel and that linguistic salience affects fixations even when the entity is not visually present.
In the basic set-up of a visual world experiment participants hear a word or an utterance while looking at an experimental display on the screen. Their eye movements are recorded as they listen to the sentences or words. Studies using the visual world paradigm (Cooper,
However, the guidance of saccadic eye movements is a complex process that depends on a multitude of factors, including the (bottom-up) visual salience of the item presented on the screen (see, e.g., Itti and Koch,
Although theories developed on the basis of the visual world paradigm have tended to be underspecified with respect to how the bottom-up visual salience of an object affects the integration of linguistic and visual information, more recently the study of such factors has become the focus of research (e.g., see Huettig et al.,
It has been repeatedly demonstrated that linguistically introduced entities have different degrees of accessibility (Grosz,
(1) a. Today we will visit some locations strongly associated with the life of the painter and graphical designer Fortunato Depero.
b. Let us start from the vineyards of the Fedrigotti family that commissioned a lot of work from him.
c. Next we will visit the house where the artist was born.
d. We will end the tour at the Depero Museum.
d′. We will end the tour at the station.
In this example, the abovementioned studies would suggest that in (1), the individual Fortunato Depero, mentioned using a proper name and in salient position in (1a), and then repeatedly mentioned in the subsequent sentences, will be linguistically salient after (1c). Therefore a continuation sentence mentioning Depero, like (1d), will be processed more easily than a continuation sentence that does not contain a reference to that entity, like (1d'). This greater salience will translate in a reduced response time.
The evidence just discussed led to the development of models of linguistic salience stipulating attentional structures that play a role in linguistic salience—similar to the role played by saliency maps with regard to visual salience. The best known among these models, and the most widely used in psycholinguistic research on reference, is the Centering Theory (Gordon et al.,
Studies such as Arnold et al. (
In particular, we were interested in whether there would be interference between visual and linguistic salience in a context in which one landmark is associated to a linguistically salient entity, whereas another landmark is visually salient (bigger or more colorful). The study of eye movements is particularly relevant for this question, since the serial nature of gaze shifts means that the eye can only move to one location at a time. The motor decision of where to move the eye next has been shown to be directly affected by (bottom-up) visual salience, not just task constraints (Gottlieb et al.,
The hypothesis tested here is that in a direction-giving task the landmarks on a map that are associated with linguistically salient entities (although not present on the map) will be the target of more fixations than landmarks not associated with such entities. The interference between linguistic and visual salience—if any—is most likely to be observed in a setting in which the linguistically salient objects are not visually present. In the Arnold et al. (
One possibility is that linguistic salience would modulate the activity in this sensorimotor map—i.e., that the CF List proposed in Centering theory coincides with, or at least interacts directly with, the saliency/priority maps proposed in the visual attention literature. Alternatively, linguistic salience might work concurrently with visual salience, influencing the processing of linguistic and visual information through a separate informational structure rather than directly altering the salience map. More specifically, if only one attentional structure is maintained (i.e., if saliency maps and CF List coincide) we would not expect to observe separate effects of linguistic and visual salience. If, on the other hand, the CF List and the saliency map work in parallel, linguistic salience would be expected to affect response (button press) and dwell (fixation durations) times but not saccadic target selection (time to first fixation).
Our research question on the interaction between Linguistic and Visual salience was investigated using four measures. First, to test the effect of CF List we measured response times, as typically done in studies on the effect of linguistic salience on the interpretation of referring expressions (Gernsbacher,
Finally, we used a more realistic scenario to study the interaction of linguistic and visual salience than the simplified settings normally seen in visual world studies—a direction-giving task in a touristic context, in which subjects are looking at landmarks while the guide tells them about their history and the famous people who lived there, such as the artist Fortunato Depero.
Forty students (18 males, mean age 26.1 years;
Eye movements were recorded using a Tobii x50 eye tracker, with a frame rate of 50 Hz (50 frames per second). Stimuli were presented on a 17-inch TFT monitor at a resolution of 800 × 600. Stimulus presentation was controlled by a PC running E-Prime 1.5 experimentation software. Spoken stimuli were presented binaurally through Sennheiser HD 570 headphones. Response time was collected through mouse button. The mouse was connected to the PC running E-Prime. In this setting the eye tracker sampled the eye position every 24.72 ms.
Visual stimuli consisted of 8 maps with five pictorial landmarks on each of them. There were more landmarks on the maps than there were visual targets as otherwise the last visual target would have been obvious by exclusion. The pictorial landmarks were taken from photographs of natural scenes. The position of the pictorial landmarks was randomized. For example, the first landmark—the train station—could be in any position on the map.
In half of the maps we manipulated visual salience by modifying the size and colors of the last sentence target landmark. The visual salience of each landmark on each map was calculated with Itti-Koch algorithm (Itti et al.,
Eighty sentences were recorded by a male native speaker of Italian (see Appendix). For each map we recorded two different versions: linguistic salience condition (+LS) and no linguistic salience (−LS condition).
In +LS condition, four out of five sentences referred to an item or a person (which we call entity) that was not on the map. For example, the sentence “Today we will visit some interesting location of Depero's artwork” was delivered while the participants attended to a fixation cross. In the following sentences, the word “Depero” was the entity element. To avoid a cognitive load effect, the second sentence, following the fixation cross slide, did not mention the entity element and was the same for all the trials (i.e., “Let's start from the train station”). In both +LS and −LS conditions the last sentence was the same. By the time the last sentence was presented two possible visual targets were left. This was to ensure that participants could not guess what would be the target of the last instruction.
In half of the trials the correct target was visually salient as well. In −LS condition the same maps used in +LS condition were presented on the screen. However, in −LS condition the sentences did not have any entity repetition. In order to control for complexity, in −LS condition we substituted the entities with words of very similar length and frequency (see Results for more information). In the +LS versions the materials would be as in example (2). In this version, one entity which was not visually present on the map (e.g., Fortunato Depero) was introduced using a proper name in the first sentence while the participants attended a fixation cross. Once the sentence ended, the map appeared. The second phrase (the first in which the participants actually saw the map) prompted the participant to localize the train station. This phrase and the corresponding visual target were identical for all of the conditions (See Figure
(2) a. Today we will visit some locations strongly connected with Depero's artwork (fixation cross)
b. Let's start from the train station (visual target = station)
c. next we visit the vineyard of the Fedrigotti family where Depero worked many years (visual target = vineyard)
d. Next we will visit the house where Depero was born (visual target = house)
e. We will end the tour at the castle holding Depero's exhibition (visual target = castle).
In the −LS condition, each phrase referred to different entities instead of always mentioning Depero, as shown in (2′). The final sentence was identical to the condition +LS (see Figure
(2′) a. Today we will visit some sights in Val Lagarina of great cultural interest. (fixation cross)
b. Let's start from the train station (visual target = station)
c. Next we visit the vineyard of the Fedrigotti family where a famous wine is produced. (visual target = vineyard)
d. Then we will visit the house where Rosmini was born (visual target = house)
e. We will end the tour at the castle holding Depero's exhibition (visual target = castle)
Participants listened to a first sentence (mean duration 3891 ms,
The general design of our experiments follows previous studies using the visual world paradigm. Participants sat in front of a screen while listening to a sentence. Our task was a “look and listen” (Huettig et al.,
In the linguistic salience condition, we referred to the target landmark using a description that would associate the object with the linguistically salient entity (e.g., “the vineyard of the Fedrigotti family where Depero worked many years”). In the −LS condition no such associations were provided (e.g., “the vineyards of the Fedrigotti family that produces a famous wine”). Participants saw each of the 8 maps only in one of the two LS conditions. Therefore, they saw each of the 8 maps only once during the experiment. The order of presentation of the 8 maps was counter-balanced across participants.
Participants were seated at approximately 60 cm from the computer screen. The lab was dimly lit. The only two sources of light were the monitor used for stimulus presentation and the monitor of the PC running the eye-tracker. The latter was located behind the participant.
After attending a fixation cross and hearing a first sentence, participants pressed the mouse button and the map appeared on the screen. After that, four sentences followed while a map with five pictorial landmarks was presented on the screen. Participants were asked to press the mouse button when they identified the target landmark mentioned in each of the sentences. They were explicitly told to not move the mouse from landmark to landmark or to click on the landmarks. The next sentence in the sequence was presented only after a button press was made.
Response times were recorded from the sentence target onset (e.g., vineyard) by mouse click.
Fixations were recorded throughout the whole experiment. We measured the log odds of fixations on each target over the total number of fixations. We calculated the log odds of fixations because the dependent variable was the region of the screen (target) to which participants directed the gaze at a given moment in time, a variable that is categorical.
As mentioned above, we measured fixations to each sentence target. For example in the sentences “the vineyard where Depero worked” (+LS) vs. “the vineyard where a famous wine is produced”
(−LS) we measured the fixations to the vineyard pictorial image. In visual world studies it is customary to transform the categorical dependent variable “fixation” into a continuous variable by calculating proportions collapsed over time and over trials in the experiment. However, such analysis has recently been criticized because it violates the assumptions that the dependent variable has an unbounded range and that errors are distributed normally and independently of the mean. In addition to that, in the visual world paradigm the dependent variable fixation has multiple observations within item (the target word) and subjects. Therefore, following Barr (
Each time window was set from the onset of the target word to the end of the longest sentence. Therefore, the time window was 1000 ms for sentence b and 1750 ms for the other sentences.
Since each of the 40 participants was presented with the 8 maps, we had 320 data points for each of the sentences that were read out.
All analyses on fixations were run in R 2.13.2. The mixed effect models for each sentence were run using function lmer in package lme4_0.999375-39. Our models used the “maximal” random effects justified by the experimental design (Barr et al.,
In the last sentence response times varied across the different salience conditions. A linguistic salience × visual salience within subjects ANOVA indicated a significant effect of linguistic salience [
The time course of log odds of fixations (see Methods) for sentence b. is shown in Figure
We did observe an effect of +VS [Est = 2.27, χ2(1) = 32,
Interaction between LS and VS was investigated but not found [Est = 0.2, χ2(1) = 1.1,
To better understand the influence of visual salience on fixations we compared the time to first fixation to the sentence target (the pictorial landmark of the train station) with respect to all the other landmarks not yet mentioned. Time to first fixation was calculated from the beginning of sentence b. A MANOVA (Null hypothesis tested with Pillai's Trace) found a significant effect for +VS [
To further explore these effects, we also considered gaze duration (i.e., the total time in ms in which the gaze stayed in the boundaries of the visual target during sentence b.) as a function of visual and linguistic salience with a MANOVA (Null hypothesis tested with Pillai's Trace). Again, we found a significant effect for visual salience [
Target fixations were longer in the +VS condition (325 ms) and shorter in −VS condition (250 ms). Interestingly, participants tended to fixate the visually salient “distractor” relatively briefly (325 ms) compared to when the last sentence landmark was not visually salient (500 ms).
In order to examine the influence of linguistic salience, we examined the time course of fixations in sentence c. and d. in all the visual and linguistic conditions (see Figures
In +LS condition, the first time the entity was heard by the participants was in sentence a., while they attended a fixation cross. In sentence c., participants heard the entity repeated for the second time from the beginning of the trial. For the first time they heard the entity with the actual map in front of them.
The time course of log odds of fixations for sentence c. in (2) and (2′) is shown in Figure
A MANOVA on the time to first fixation at sentence c. target did not find a significant effect for linguistic salience [
As regards the total duration of fixations at the instruction target, a MANOVA found a significant effect for +LS [
In +LS conditions the entity was repeated in sentence d. The time course of log odds of fixations for sentence d. in (2) and (2′) is shown in Figure
A MANOVA on the time to first fixation did not find a significant effect for linguistic salience [
As regards the total duration of fixations at the instruction target, a MANOVA found a significant effect for +LS [
Sentence e. is the same in 2 and 2′. A linear mixed-effect regression on the log odds of fixations indicates a significant effect of +LS [Est = 1.15, χ2(1) = 26,
A MANOVA on the time to first fixation on the last sentence target found a significant effect for +LS [
We checked for frequency and length of target and entity words in the Italian corpus Repubblica (Baroni et al.,
In this study we investigated the gaze pattern during a task in which bottom up visual salience and linguistic salience have been manipulated. The linguistic salient entity was not visually present on the map but only associated with the visual landmarks presented on the map. The hypothesis tested is whether the CF list containing linguistically introduced information and the visual salience map shared the same attentional structure or work concurrently. If only one attentional structure is maintained (i.e., if saliency maps and CF List coincide) we would not expect to observe separate effects of linguistic and visual salience in terms of reading times, fixation dwells and saccadic target selection (time to first fixation). On the other hand, in case the CF List and the saliency map work concurrently, linguistic salience would be expected to affect reading times and fixation dwells but not saccadic target selection.
As regards visual salience, our findings suggest two main effects. First, participants tended to look at the visually salient item when the map appeared in sentence b. Secondly, they were faster, overall, in looking for the correct target of sentence b., having already sampled and excluded the visually salient item in their first fixation(s).
By the time sentence c. was presented to participants, there was no bottom-up effect due to the visually salient landmark, and linguistic salience was the only significant effect. Linguistic salience has an effect on attention after the second repetition of the entity. As for sentence d., there was no bottom up effect of visual salience and time to first fixation was not significant in any of the conditions. With respect to the total duration of fixation, we observed an effect for +LS. Finally, in sentence e. we found a significant effect of both LS and VS on fixations, whereas for the time to first fixation there was a significant effect of +LS but no significant effect of +VS.
Our results show that linguistic salience affects interpretation in terms of reduced response times in a visual world setting as previously reported by, e.g., Arnold et al. (
Our second key finding is that although both linguistic and visual salience influenced eye movements and response time in a visual world context, the two salience manipulations affected performance in a different way and appeared to operate independently. Visual salience acted quickly, within the first appearance of the map on the screen, while linguistic salience became significant from sentence c., after the entity repetition was repeated for the second time. The effect of visual salience on fixations during sentence b. could be due to the fact that the visually salient item immediately attracts participants' attention. In the +VS condition, the data from time to first fixation and fixation duration showed that participants quickly fixated, and then discarded, the visually salient item when the map appeared. After that, participants found the task salient item, i.e., the train station, quicker. In maps where all the pictorial landmarks had the same visual salience, this “facilitation” effect was not present.
By the last sentence, linguistic salience influenced both the time course of fixations and response time. In contrast, visual salience influenced the time course of fixations at both the beginning and the end of the trial but did not show a significant effect on response times. With regard to the time to first fixation, when the target landmark was both visually and linguistically salient, it was fixated upon slower and at a similar speed compared to the conditions in which the target was visually salient only or not salient at all. For both +LS and −LS conditions, visual salience slowed down the time to first fixation. These results suggested that the two types of salience work in parallel, slowing down the time to first fixations and lengthening the fixation dwells. That is, when the target landmark was both visually and linguistically salient, it was fixated on for longer, but fixations were quicker when the target item was linguistically salient only. This means that VS no longer had bottom up attention grabbing functions. As +VS slowed the time to first fixation and lengthened fixation dwells, we conclude that the two types of salience are relied on two independent mechanisms. This is corroborated by the results of the response times that are slower in both +VS conditions.
It has been argued that participants already knew the last sentence landmark because of the experimental design. Participants fixated longer the last sentence landmark because they would have learned from previous trials that bigger pictures will be the last sentence target. This explanation, though, does not account for the lack of significance for +VS time to first fixations. If participants had learned that the bigger, colorful landmark was the last sentence target they would have fixated on it before every other target and quicker, but the results for time to first fixation in sentence e. showed that this is not the case. Moreover, if the experimental design would have facilitated the individuation of the last landmark, in +VS conditions participants would have been faster in responding to the last sentence. Again, this is not the case, as the response to visual salient landmarks in the last sentence was significantly slower in +VS than in −VS. +LS conditions were significantly faster than −LS conditions, whereas +VS conditions were not necessarily faster than −VS ones. In +LS and +VS condition, we observed both an increase in fixations on the target related to a linguistically salient entity, and reduced response times. The slowest response times, however, were observed in the −LS+VS condition. This evidence suggests that visual and linguistic salience may involve distinct attentional mechanisms. One possibility is that fixations are driven primarily by a visual-spatial salience map which codes the presence of high priority items in a spatial map. This spatial map integrates both exogenous salience (visual salience, in this case) and endogenous salience based on task relevance; the evidence shown here suggests that the salience map may also be affected by linguistic salience. Salience maps guiding eye movements have been reported mainly in visual and oculomotor areas, suggesting that gaze is closely tied to motor and spatial representations. In contrast, response times would seem to depend on a distinct structure such as the CF list postulated by the Centering Theory. Such a structure affects semantic interpretation and does not integrate visual salience. Under cases of uncertainty or ambiguity, visual information might help to disambiguate the interpretation of the scene, leading to the gaze shifts found in visual world experiments. Likewise, mentioning a word might increase the salience of that item, leading to increased priority for it in the spatial map. Nonetheless, the underlying mechanisms guiding sentence processing (semantic) and gaze control (spatial) may be at least partially dissociable.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Federica Cavicchio was supported by the Marie Curie Action – COFUND-TRENTINO, project CoByGest, David Melcher was supported by the European Research Council Starting Grant StG-20111124 (ERC grant agreement no. 313658), Massimo Poesio was supported by GALATEAS ICT Policy Support Programme, Deep Relations Grant (PAT and Expert Systems) and PRESTO (Grant PAT to Delta Informatica).
[Oggi visiteremo alcuni luoghi interessanti per studiare la vita del pittore Depero].
Partiamo dalla stazione di Rovereto.
Poi passiamo dal vigneto della famiglia Fedrigotti, per cui Depero ha lavorato molto.
Dopodiché visitiamo la casa dove Depero e' nato.
[Oggi visiteremo alcune località d'interesse culturale della Val Lagarina].
Partiamo dalla stazione di Rovereto.
Poi passiamo dal vigneto della famiglia Fedrigotti, che possiede una famosa cantina.
Dopodiché visitiamo la casa dove Rosmini e' nato.
[Oggi andiamo a fare una gita in Val di Fiemme].
Partiamo dalla stazione dei treni.
C'è' una cascata molto bella all'inizio della valle.
C'è' una montagna molto bella continuando lungo la valle.
[Oggi andiamo a fare una gita].
Partiamo dalla stazione dei treni.
C'è una cascata molto bella all'inizio del percorso.
C'è una montagna molto bella continuando la gita.
[Oggi andiamo a spasso per Trento].
Partiamo dalla stazione dei treni.
Poi andiamo nella piazza dove gli abitanti di Trento vanno a spasso la sera.
Possiamo andare al bar a bere uno degli spritz migliori di Trento.
[Oggi andiamo a spasso per Trento].
Partiamo dalla stazione dei treni.
Poi andiamo nella piazza dove c'è' una antica fontana del Seicento.
Nella piazza possiamo andare al bar a prenderci uno degli spritz migliori della città.
[Oggi facciamo una passeggiata per Rovereto].
Partiamo dalla stazione dei treni.
Poi visitiamo il museo di arte contemporanea di Rovereto.
Dopodiché passiamo dalla Piazza principale di Rovereto.
[Oggi facciamo una passeggiata per Rovereto].
Partiamo dalla stazione dei treni.
Poi visitiamo il museo di arte contemporanea del MART.
Dopodiché passiamo dalla piazza principale del Nettuno.
[oggi andiamo a vedere il castello di Rovereto].
Partiamo dalla stazione dei treni.
Sulla collina troveremo il castello.
Attraversando la città vecchia arriviamo al castello.
[oggi visitiamo la città di Rovereto].
Partiamo dalla stazione dei treni.
Visitiamo la collina e il suo bosco.
Attraversando la città vecchia arriviamo alla piazza.
[oggi visitiamo la val di Non].
Partiamo dalla stazione dei treni.
Visitiamo i suoi famosi meleti della valle.
Ora vediamo il lago che taglia la valle.
[oggi visitiamo alcune zone di montagna].
Partiamo dalla stazione dei treni.
Visitiamo i famosi meleti del trentino.
Ora vediamo il lago che è tagliato da una diga.
[oggi visiteremo le terme].
Partiamo dalla stazione dei treni.
Attraversando il lago ci avviciniamo alle terme.
Arrivando vicino al monte possiamo vedere le terme.
[oggi vedremo alcune località della Vallagarina].
Partiamo dalla stazione dei treni.
Attraversiamo il lago più famoso della zona.
Arriviamo vicino al monte, il più alto della valle.
[oggi visiteremo il museo di arte contemporanea di Trento].
Partiamo dalla stazione dei treni.
Dalla collina di Povo il museo è stato trasferito a Trento.
Troviamo un aeroplano nei giardini davanti al museo.
[oggi visiteremo alcuni musei a Trento].
Partiamo dalla stazione di Trento.
Sulla collina di Povo troviamo il museo della scienza.
Troviamo un aeroplano nei giardini del museo dell'aviazione.