5.1. Think-aloud protocols

A think-aloud protocol is an introspective method in which participants are asked to report their thoughts, either as they unfold, or after reading or listening to a text. Verbalization may refer to spontaneously developed thoughts by participants, or to the justifications or explanations requested by researchers. Verbalization can thus be used for assessing the metalinguistic or non-metalinguistic aspects of comprehension, depending on the instructions given. It is also possible to categorize reported thoughts following previously defined criteria, in order to transform them into quantitatively measurable data. For this classification to be feasible, it is necessary to give precise instructions to the participants about the type of thoughts that should be reported, or the moment when these thoughts need to be expressed.

For example, Blanc et al. (2008) studied the conditions in which readers update the mental representations constructed during reading, as new information is provided in the text, sometimes contradicting what has already been presented. As we have already discussed in Chapter 4, readers build the comprehension of a text by developing a mental representation which contains not only the information transmitted by the text itself, but also inferences. These are deductions based notably on world knowledge. In order to adequately reflect the content of the text, mental representations have to be continuously updated, as new pieces of information are continually brought in by the text. This update may correspond to the addition of information to an existing mental representation, but in other cases, it may require revising already formed mental representations and inferences, as is the case when contradictory information emerges.

Blanc et al. (2008) investigated the updating of mental representations constructed while reading newspaper articles reporting dramatic events. They put together six experimental articles, all following the same structure and presenting two plausible causes for the dramatic event. In each text, a critical sentence provided elements either in favor of the first cause or in favor of the second one, or else, neutral elements in relation to these causes. Six other items were similarly constructed, but presented only one plausible cause for the dramatic event. The participants of the experiment had to read the 12 articles, sentence-by-sentence, and answer a question orally: “At this moment, what comes to your mind about the information you have just read?” This question arose every two sentences of the text, as well as after the presentation of the critical sentence, and then again at the end of the text.

The responses were coded depending on whether or not readers mentioned one of the two causes appearing in the text as a reason for the occurrence of the dramatic event. When both causes were mentioned, the existence of a relationship between the causes (or not) was coded as well. Finally, any mentions of comprehension difficulties were also recorded. The analysis of the responses given after the presentation of the second cause showed that the two causes presented in the text were generally taken into account. In addition, the participants were aware that the causes changed as the text progressed, that is, the event was eventually explained by the latter cause, but had earlier been explained by a different cause. Furthermore, when an alternative cause was presented, participants mentioned that they had probably misunderstood information presented earlier in the text, or expressed criticisms as to the way the report had been written. These reactions showed that the emergence of a second cause prompted them to update their mental representation of the text.

An analysis was then carried out on the answers given directly after the critical sentence, which strengthened one cause or the other, or did not strengthen any cause, as well as on the answers given at the end of the article. When the critical sentence strengthened the first cause or transmitted neutral content, the participants included the two causes in their responses, showing that the two had been activated in parallel in their mental representation. On the other hand, when the critical sentence strengthened the second cause, the participants mentioned this cause more than the first one in their responses. These results confirm that the order in which information is presented, rather than the number of times presented, influences the mental representations of readers. Actually, when the first cause was strengthened, the two causes were activated in the memory of the participants, whereas when the second cause was strengthened, only the latter was activated in the end. This can be explained by the fact that the strengthening of the second cause took place directly after its presentation, encouraging participants to forget about the first cause and to accept the second one. This explanation could be corroborated by observing the difficulties reported by the participants. They found the texts more difficult to follow when the critical sentence strengthened the first cause or none of the causes, than when it strengthened the second one.

This example illustrates the interest of think-aloud protocols for the study of processes taking place during comprehension. By observing the thoughts transmitted by the readers as they progressively read a text, it is possible to observe the steps followed by the participants, as well as the changes in their representations. It would not be possible to access these processes only by observing the final representation, as with offline tasks discussed in the previous chapter, or through the use of other online tasks that we will present in the rest of this chapter.

Think-aloud protocols have, however, significant limitations. First, verbalizing our thoughts whilst reading requires high cognitive skills, in order to be able to become aware of our train of thought and to verbalize it. Moreover, we cannot dismiss the fact that the very nature of this task – having to consciously access our thoughts and report them – influences the natural reading process or interferes with text comprehension. Indeed, expressing our thoughts while reading involves distancing ourselves from the text during the time required for expression, before returning to the text. This distancing can compromise the construction of the representation or can add elements which would not have been present if the participants had only read the text. Finally, this task is only appropriate to study those processes which are accessible to consciousness and can be reported on a voluntary basis. For all these reasons, think-aloud tasks are not the best method for studying online processes, and their use has been marginalized since the development of other time-based methods. The rest of this chapter is devoted to them.

5.2. Using time as an indicator of comprehension

The cognitive processes involved in language processing are extremely fast and most of the time, inaccessible to consciousness. In order to study them, it is necessary to access them in an indirect way, by observing indicators or signals of the processes, as explained in Chapter 2. A very commonly used indicator in experimental linguistics – and in cognitive science in general – is the time required to complete a task. The use of this indicator is based on the idea that the time required for processing a linguistic stimulus reflects its degree of difficulty: the more complex the stimulus and/or its processing, the longer processing time it requires.

The term “processing” refers here to the different stages and sub-processes involved in comprehension or production of a linguistic stimulus. This term is deliberately vague, because it depends on the specific aspects of language processing targeted by the different tasks in which time is the dependent variable. Such tasks can be divided into two broad categories. A first category groups the tasks in which the participants must react to a stimulus. In this case, reaction time is measured. A second category concerns the reading tasks themselves. Here, the focus is placed on the time it takes for text segments to be read. Importantly, the time measured does not reflect the same type of processes, depending on the task involved. For example, the reading time of a simple naming task could reflect the time required for deciphering the word, building its phonological representation and then being able to pronounce it aloud. The reading time of a sentence placed in the middle of a text reflects the time needed to decipher every different word, to connect words and to add incoming information to the mental representation already constructed on the basis of previous sentences. Reading this new sentence also potentially involves the derivation of inferences or the need to revise a prior mental representation.

Differences in response times have been documented by studying multiple variables, such as the length of the linguistic stimulus, for example: short words are processed faster than longer words. Another factor is word frequency: frequent words are processed faster than rarer words. Syntactic complexity also seems to play a role: simple sentences are processed more quickly than complex ones (Just and Carpenter 1980; Rayner 1998; Smith and Levy 2013). The processing time of a stimulus makes it possible to deduce information not only about the complexity of the linguistic stimulus itself, but also about the number and the dynamic organization of stages involved in the processing of the stimulus.

In a typical experiment aimed at measuring reaction or reading times, the participants see linguistic stimuli, namely words, sentence fragments or complete sentences, and must perform an action based on these stimuli. As we will see later, these actions may be varied such as deciding whether a string of letters corresponds to a possible word, saying whether a certain word was present in a sentence or even simply reading a sentence. Stimuli are presented on a computer screen by means of experimental software and reading time is measured by asking the participants to indicate their response by pressing a key on a keyboard. In order to get a more precise measurement, it is possible to use a button box, which allows us to measure the response time with a millisecond accuracy. It is also possible to ask participants to respond orally by using a microphone and a voice key for collecting the audio signal and automatically recording the response and the response time. Response time generally corresponds to the time between the initiation of the stimulus and the participants’ response.

The use of response time as a dependent variable requires following certain methodological principles. We will approach these from a theoretical point of view, before taking them up again in the form of practical advice in Chapter 6.

In general, in a task measuring response times, participants can choose between two possible answers: YES and NO, for deciding, for example, whether a string of characters corresponds to an existing word or not. The processes underlying YES and NO responses are different, and for this reason, it is necessary to associate them with different motor responses (see, for example, Rossi (2008)). Typically, YES responses are associated with the participants’ dominant hand, whereas NO responses are associated with the other hand. In order to accurately measure the response time, it is also necessary for the participants to constantly keep their fingers on the response keys during the experiment, so as not to add extra time for identifying the keys. Participants are generally invited to sit in front of a screen, their fingers placed on the keys, and are asked to remain in this position throughout the experiment.

Another important methodological point to understand is that response time per se provides little information about the processes underlying comprehension. Its contribution to the study of an independent variable requires a comparison of at least two conditions, one where the independent variable is present and one where it is absent. In this case, we refer to a subtractive method, in which the difference in the response time between the two conditions reflects the extra time needed for processing a particular type of stimulus or for carrying out a certain process. For example, measuring the time a person takes to read the pronoun she in a sentence would not contribute to drawing any conclusion. In contrast, comparing the reaction time for this pronoun in two different contexts, for example, with reference to secretary and to astronaut, could shed some light on the processes involved when reading the sentence. Here, the reaction time for she would probably take longer after astronaut than after secretary, showing that readers deduce gender-information about a character based on stereotypes pervading society. Firstly, this highlights the need to clearly define the process examined, so as to construct experimental conditions effectively isolating such process. Secondly, it shows that it is essential to choose the appropriate task for measuring the process as directly as possible.

Let us first focus on the need to isolate the process we want to study. Comparing response times between two conditions implies that such conditions must be similar in all respects, apart from the manipulation of the independent variable. It is essential to compare only that which is comparable. Let us go back to the effect of word frequency on processing time, already presented in Chapter 2. There, we saw that it was necessary for the frequent and the less frequent words used in the experiment not to differ on other criteria, such as their length, in order to reliably assess the effect of frequency itself. Thus, the most favorable situation would involve having the same linguistic stimuli repeated across different conditions. In this case, this would be unattainable, since a word cannot simultaneously be very frequent and infrequent. Actually, the different stimuli should be similar on as many points as possible in order to prevent confused variables from jeopardizing the validity of the experiment. All the variables susceptible of influencing response times, such as word length, word frequency, their concreteness, their grammatical category, their position in the sentence and their contextual predictability, should be kept equivalent across the different conditions.

Let us now turn to the choice of the task used to collect response times. This choice is essential, since the processes involved in the different conditions should allow us to reveal the influence of the independent variable. For example, let us imagine a study seeking to determine whether it is faster to indicate an answer on the keyboard using the dominant hand or the other hand. For this study, it would be necessary to build an experiment in which the participants answer half of the time with their right hand, and the other half of the time, with their left hand. For instance, the task could involve verifying simple operations by indicating whether the result is correct (using the forefinger) or incorrect (using the middle finger). In this case, the task of the participants would comprise several stages, such as deciphering the figures and symbols appearing on the screen, resolving the operation, comparing the result with the one shown on the screen, deciding the answer, choosing the finger for pressing the corresponding key and finally pressing the key. Reaction time would also reflect all of these steps. This is why we speak of choice reaction time (or complex reaction time) in cases like this: the response depends on a choice made by the participant. Using choice reaction time for the question we are analyzing might pose two potential problems.

The first problem relates to the fact that the different stages involved in the response should be equivalent for both conditions (right hand or left hand), especially on how difficult the operations are. In order not to threaten the validity of the experiment, all operations should be kept even in terms of the difficulty level. The second problem stems from the large number of steps involved in the task, which can prevent the detection of the desired effect. Indeed, the contribution of the left hand/right hand activation to the reaction time is moderate (see Figure 5.1 for an illustration), as it only represents one stage among others, and is not the longest one to complete. In addition, the hypothetical time difference between the answers given by the right hand and those given by the left hand is probably small. By using this type of task, there is a risk of drowning the effect in the combination of processes involved in the task. In order to study the reaction speed of the two hands as directly as possible, we should aim for the simplest possible experiment, in which the number of processes involved should be kept to a minimum. For example, we could simply present dots on the screen and ask participants to press a single key as quickly as they can when a dot appears. This simple task would make it possible to obtain a more direct measurement of the influence of the hand employed over the reaction time, without tainting the response with incidental processes. As we can see in Figure 5.1, the process of pressing the key would have a more important weight over the reaction time. In this case, we would speak of simple reaction time, since the task would only aim at giving one single response when a stimulus appears.

Unlike the fictitious tasks presented above, the tasks that we will describe in the remainder of the chapter aim to ensure that the participants cannot guess the purpose of the experiment, so as not to compromise its validity. There is always the risk that the participants’ responses can become unnatural as soon as they guess or believe they guess the hypothesis under examination.

Using time as a dependent variable could apply to most offline tasks described in the previous chapter. A possibility would be to measure the time needed for providing the answers in the different tasks we presented, without changing the instructions given to the participants. Measuring time would offer additional information to that already offered by the responses themselves. For example, in an acceptability judgment task, it would be possible to observe that people give similar acceptability scores to different stimuli, but that evaluating some of these stimuli may require a longer amount of time than others. In this case, it could mean that some of the stimuli were more complex. The reason for this greater complexity has yet to be defined, but the response time could act as an indicator that the stimuli or the processes involved in their comprehension differ, despite the similarity of their acceptability score.

In the rest of this chapter, we present a series of tasks aimed at studying online comprehension, for which the response time is at the center of the paradigm, as much as the content of the answer provided. In these tasks, a certain time constraint is placed on the participants, unlike the above-mentioned tasks. This time constraint aims to ensure that the desired processes are measured, regardless of other processes the participants could engage in. To do this, the participants are instructed to respond as rapidly and as accurately as possible. Indeed, without any time constraints, we could not rule out that the participants maximize the number of correct answers by taking their time to answer, or on the contrary, minimize their response time by not paying attention to the accuracy of their answers (speed-accuracy trade-off). By asking the participants to focus on both speed and accuracy, we try to avoid these phenomena, and encourage the participants to respond correctly, and this, in the shortest possible time. Online tasks often rely on a priming mechanism, which is described in section 5.3.

5.3. Priming

The priming effect occurs when a first stimulus (called the prime) is presented and influences the processing of a subsequent stimulus, the target. Priming is said to be positive when the presentation of the prime decreases the target’s processing time, and is said to be negative when the presentation of the prime increases the target’s processing time.

Priming can be explained by the fact that concepts are stored in networks within the memory, on the basis of shared properties. Words, for example, are concepts that are connected by their semantic and phonological properties within the mental lexicon. The presentation of a prime activates all of the properties associated with it, and this activation spreads to other concepts sharing similar properties. When the target appears, some of its properties are pre-activated, which makes it easier to process. For example, seeing a prime picture of a swan facilitates access to the word duck, compared to the word dog, since there are more similarities between a duck and a swan, than between a dog and a swan.

Priming effects are investigated by means of tasks presenting a prime and then a target for which the participants have to provide an answer. In general, the target is shown until the answer is given, but the maximum duration of its presentation time can also be defined beforehand. The answers can be of different types, depending on the task the participants have to carry out. The most common tasks are evaluation, lexical decision or naming tasks, which are presented later in this chapter. Primes can be shown in such a way that participants can see them and process them consciously, or on the contrary, they can be presented for a very short amount of time, so that they are only perceived subliminally. It is also possible to control the capacity of processing the primes by presenting them either in an isolated manner or preceded or followed by a mask (i.e. visual noise, e.g. #####, that blocks their processing). Finally, it is possible to vary the time between the presentation of the prime and that of the target, in order to study the time frame related to the processing of the primes and its influence on the processing of the targets.

The central manipulation in priming tasks is the relation between the prime and the target. This relation can be semantic, phonological, syntactic or even affective. In all cases, it is necessary for the different types of primes to differ only in terms of the variable examined. For this reason, rather than comparing the presence or the absence of a prime on a process, it is recommended to compare the presence of one type of prime and its absence (but a prime, altogether). Later on, we will provide illustrations for this. If a priming effect is present, we can expect to obtain faster response times for the targets in the priming condition than in the control condition.

The priming effect is central in experimental linguistics, because it can shed light on many processes. For example, this effect has allowed us to better understand the organization of the mental lexicon, as well as the way in which we access words during language comprehension and production. Numerous studies have indeed shown priming effects for semantically, orthographically or phonologically related word pairs (e.g. Ferrand and Grainger (1992, 1993) and Dell’Acqua and Grainger (1999)).

We will now turn to the different tasks for testing the priming effect, or more generally, involving a priming mechanism.

5.4. Lexical decision tasks

In a typical lexical decision task, combinations of letters corresponding to words and non-words (sometimes called pseudo-words) are presented, and participants are asked to indicate whether the stimuli are words or not, as rapidly and as accurately as possible. These answers are given by pressing one of two keys at their disposal, representing YES or NO responses. This type of task implies that half of the items presented are words and the other half are non-words.

Lexical decision tasks can be used to study a wide variety of processes. In the field of word recognition, for example, these tasks allow for the manipulation of many variables, such as their visual, phonological or semantic properties, in order to determine the importance of such properties in the recognition process. Lexical decision tasks have notably revealed one of the most robust effects in the field of reading, namely the frequency effect. The frequency of a word corresponds to an estimate of the number of times a person has encountered this word, and is calculated on the basis of the number of occurrences of this word in a corpus. The higher the word frequency, the easier it is to categorize it as a word (e.g. Baayen et al. (2006)).

Most lexical decision tasks involve a priming process. Through this method, it has been possible to demonstrate the phonological priming effect which we have already mentioned. As an example, we will present the study by Carreiras et al. (2005) that revealed certain characteristics of phonological activation during silent reading and whilst reading aloud. Numerous studies have shown that the phonological information associated with words is activated during silent reading (Ferrand and Grainger 1993; Ziegler et al. 2000; Drieghe and Brysbaert 2002) thanks to tasks that combine monosyllabic primes and targets. Since they contain only one syllable, these primes and targets can be phonologically very similar. Carreiras et al. extended the study of the role of phonology to the reading of bisyllabic words, in order to examine two central questions. First, determining whether phonemes and syllables are decoded sequentially or in parallel. Second, exploring whether it is possible to obtain phonological priming only when the overlap between primes and targets is partial.

To answer these questions, Carreiras et al. (2005) chose 120 French words as well as 120 bisyllabic non-words (five to eight letters long) as targets. Then, they created non-words which would serve as primes for the two types of targets. In relation to the first question, the focus was placed on priming the first syllable for half of the targets, and the second syllable for the other half. In order to answer the second question, primes were divided into three categories. The first category corresponded to a prime containing a phonologically similar but orthographically different syllable from the one examined in the target. For example, fomie appeared before faucon in the case of the first syllable, and retôt appeared before gâteau in the case of the second syllable. The second category of primes contained the same first phoneme (and grapheme) as the target syllable of interest (fémie before faucon in the case of the first syllable, and retin before gâteau in the case of the second syllable). The third category was neither phonologically nor orthographically related to the target (pémie before faucon and redin before gâteau).

Each target, whether a word or a non-word, and whether a first-syllable or a second-syllable target, was associated with a prime in each category. Each participant saw 20 targets of each type in each priming condition, in order to ensure that every participant saw all items and all conditions without being presented with the same item more than once. We will return to this notion of groups of items, also called lists, in Chapter 6.

Each target was presented in a random order, and the same procedure was applied to all the different items. Participants were instructed to indicate as rapidly and as accurately as possible whether or not the letter string was a French word. For each of the 240 tests, a mask (XXXXXXXXXXX) appeared on the screen and remained there for 500 milliseconds. Immediately afterwards, the mask was replaced by the non-word prime which remained on the screen for 59 ms, too short a time-lapse for the participants to process the prime consciously. Then, at the end of the 59 ms, the prime was immediately replaced by the target, which remained on the screen until the participant responded.

Analyses were performed on the correct response times for the target words, depending on whether the first or second syllable had been primed. For first-syllable targets, reaction times were faster in the phonological-priming condition than in the first phoneme and unrelated priming conditions. These last two conditions did not differ from each other. No effect, however, appeared for second-syllable targets.

These results showed that – at least in a lexical decision task – phonological priming occurs when the primes have only a partial phonological overlap and the orthographic overlap with the target is minimal. Moreover, the fact that phonological priming only appeared for first-syllable targets supports the idea that phonological processing during reading is sequential.

Lexical decision tasks can also be used for studying the content of mental representations constructed while reading, especially the inferences generated by readers. In this case too, studies are based on the presence of a priming effect, more specifically on the fact that the activation of a concept in the readers’ memory should be transmitted to the associated concepts and make them more accessible. On this basis, it is possible to present texts requiring the derivation of inferences and to test whether an inference has actually been generated by making it appear at different places in the text, in the form of a lexical decision task. The necessary time for responding is supposed to reflect the activation that the concept received while reading the text.

De Vega et al. (1997) used such a lexical decision task to study the ability to infer a character’s emotion by adopting their perspective. The authors created short stories in which the main character was aware of a piece of information (or not) which should have influenced their emotional state. For example, one of the stories featured a woman waiting for her partner who was very late for an appointment. The story elaborated that her partner was inundated with work and that she would not want to put even more pressure on him. In any case, the woman finally decided to call him at home and she was told (in the informed condition) or wasn’t told (in the uninformed condition) that he was playing poker with his friends. The story continued with a neutral sentence explaining that the woman was thinking about her partner and ended with a concluding sentence.

Such a story can activate different representations about the protagonist’s emotion. On reading the first sentences, readers will likely infer that the woman feels sympathy towards her partner. He works a lot, and she doesn’t want to put any additional pressure on him. The protagonist’s emotion should remain the same in the uninformed condition, as she is unaware of the fact that he is not meeting her because he is spending time with his friends while she believes him to be at work. However, in the informed condition, the protagonist’s emotion should no longer be sympathy but a shift to anger, as soon as she realizes the real reason for her partner’s absence.

If the readers infer the characters’ emotions by taking their perspective, then the concept of sympathy should be activated in their mental representation in the uninformed version, and fury should be activated in the informed version. In order to test this, de Vega et al. (1997) asked participants to read the stories sentence-by-sentence and to complete a lexical decision task. This contained an adjective describing the initial emotion (sympathetic) in one experiment and the other emotion (furious) in another experiment. Target words appeared at the end of the neutral sentence. The results showed faster decision times for the initial emotion than for the alternative emotion in the uninformed condition, and the opposite results for the informed condition. These results support the initial hypothesis that readers adopt the characters’ perspective during reading.

In summary, lexical decision tasks have the advantage of being very easy to implement thanks to the use of experimental software that allows the presentation of stimuli, the recording of responses and of reaction times. However, they present a significant limitation, in that the decision process underlying responses can be assimilated to a categorization process (yes, it’s a word or no, it’s not a word) along a lexical familiarity continuum (Ferrand 2001). In a simplified manner, the participants in a lexical decision task can set up strategies in order to classify the strings of letters presented. These strategies depend on various variables, such as the familiarity of the words and the non-words shown, their phonological or orthographic characteristics, or even the instructions given. As a consequence, they may influence the responses and reaction times obtained in the experiment, which may actually depend on different factors from the ones we wish to investigate. It is therefore important to take this limitation into account when choosing the items for a lexical decision task.

5.5. Naming tasks

Naming tasks offer an alternative to the inherent limitations of the categorization process involved in lexical decision tasks. These tasks are extremely simple: participants have to pronounce a word, either immediately after its appearance or in a delayed manner. The word to be pronounced is presented on a screen, and the naming time (i.e. the time between the presentation of the stimulus and the start of the response) as well as the response itself are recorded. This method makes it possible to calculate the correct/incorrect response rate, which can then be analyzed in parallel with response times.

Naming tasks are based on perceptual and on production processes. This is why they measure various processes related not only to silent reading, but also to the production and articulation of words. For this reason, the risk of this method is that is does not allow one to distinguish the role played by those different processes in the response time. A first possibility to solve this problem would be to set up a delayed naming task, in which the response does not immediately take place after the presentation of the word, but after a certain delay. By comparing the latencies obtained from a delayed task with those drawn from an immediate task, we can determine whether the differences between conditions stem from processes related to reading or to naming.

A second possibility would be to add another task to the naming task, in order to confirm the effects using a different technique that does not present the same advantages and limitations. In the literature, we often find a naming task combined with a lexical decision task. As an example, let us refer back to Carreiras et al. (2005) and their study on phonological priming. In their paper, they presented a second experiment, similar in all respects to the one presented above, but this time using a naming task. The results confirmed those obtained with the lexical decision task and showed an additional effect, namely a shorter response time when the target started with the same grapheme as the primer. According to the authors, this new effect can be explained by the fact that the naming task involves an articulatory process which is not present in the lexical decision task. This example illustrates the complementarity of approaches, which we have already emphasized several times in this book, as well as the need to verify the results by using different methods. The particularities of each method can reveal effects which would not emerge using other techniques.

Another example of the application of naming tasks concerns the field of discourse and inferences. One type of inference, predictive inference, involves the activation of information about predictions that can be made based on the text, such as the consequences of an event described in the text, or about future events. Lassonde and O’Brien (2009) investigated the specificity of predictive inferences depending on the contextual support conveyed by the text. According to their hypothesis, the more information a text contains about a specific inference, the higher the probability of observing such an inference, and of increasing its specificity. The specificity of the inference was operationalized as the number of activated lexical items following the reading of a text. If an inference is specific, few items should match it. On the contrary, if the inference is rather general, then more items should be activated.

In order to test this idea, Lassonde and O’Brien (2009) developed three experiments. The first one aimed to test the assumption that predictive inferences are generally not specific, and can include different lexical items. In order to investigate this assumption, participants had to read short stories and perform a naming task at the end of each story. The stories were presented in such a way as to trigger the development of a predictive inference, or not. For example, one story described a young boy, Jimmy, playing with local children throwing rocks at a target. Then, the story continued with the manipulated sentence, which could either activate a predictive inference, “Jimmy missed the target and he accidentally hit the door of a new car”, or transmit some content not activating this type of inference: “a dog came racing across the street and distracted Jimmy from his throw”. Immediately after this sentence, participants saw one of the two possible target words (dent or damage) appear on the screen, and had to pronounce it. Target words were determined before the experiment so as to be the most likely to reflect predictive inferences related to the text. We can nonetheless appreciate that the second target word is less specific than the first one. The results showed that for the two target words, the naming times were faster in the inference condition than in the control condition.

Another experiment aimed to verify that as contextual support increases, the predictive inference becomes more specific and the number of lexical items activated diminishes (Lassonde and O’Brien 2009, experiment 3). The same texts were used after their introduction had been slightly modified, and this time contained information aiming to encourage one of the two target items, compared to one another.

For example, Jimmy’s story stressed the fact that the brand new car of his friends’ family did not have any scratches or blemishes, and that children should be careful not to damage it. The task was the same as in the first experiment, as were the target words presented. The results showed there was only a difference in response times between the inference and the control conditions for the most specific target word (dent), and that there was no difference for the other word (damage). This result suggests that contextual support influences the specificity of predictive inferences. These inferences are captured by fewer lexical items, as the information transmitted by the text progressively constrains them.

5.6. Stroop task

The Stroop task draws its name from the Stroop effect (Stroop 1935), well known in psychology, which illustrates the phenomenon of automatic access to meaning while reading a word. In a classic Stroop task, participants must name the color of the ink that is used to write down the name of a color, such as, for example, the word blue being written in blue or in red. Naming times are slower when the color of the word is incongruent with the name of the color, thus reflecting the impossibility of preventing reading a word and accessing its meaning. In other applications of this task, the words presented in color may refer to concepts other than a color itself (see the example below), or to meaningless strings of letters or symbols. In all cases, naming times in the incongruent condition (sun written in blue, for example) are slower than in control conditions (xxx written in blue), which correspond to an interference effect associated with the Stroop task. But this effect is not the only one at work in a Stroop task. There is also a facilitation effect, whereby naming times in the congruent condition (sun written in yellow) are also faster than in the control condition (xxx written in yellow) (Augustinova et al. 2016). From a methodological point of view, it is therefore very important to choose the relevant control items.

The Stroop effect is useful for studying different aspects of language. Regarding the lexical access while reading, various studies have questioned the automatic semantic encoding postulated by the results of the classic Stroop task described above. By modifying the types of items presented, Besner et al. (1997) showed that semantic activation is not automatic during a Stroop task. As a matter of fact, the interference effect decreased when only one letter of the color name was colored. This effect also diminished if the control words were non-words (in this case, the reading of the words becomes less relevant for the participants) and even disappeared completely when a single letter of the non-words was colored. These results led to the conclusion that the automatic processing of words and access to meaning depend on specific conditions, and that it is possible to not activate such processes.

Another possible application of a Stroop task is found in the study by Eilola et al. (2007), who investigated the processing of emotional words in the mother tongue (L1) and the second language (L2) of bilingual participants. In order to study this question, an emotional Stroop task was implemented. In this task, positive, negative and neutral emotional words are presented in colors. Participants simply had to indicate the color of the word. This task has generally shown that negative and taboo, emotional words are treated differently from neutral words, as their response times are generally longer (McKenna and Sharma 1995; Williams et al. 1996). The mechanism underlying the interference observed in the emotional Stroop task is not yet clear, and different contending explanations have been proposed in the literature (see, for example, Algom et al. (2004) and MacKay and Ahmetzanov (2005)).

In order to extend these findings to the different languages of bilingual subjects, Eilola et al. (2007) recruited advanced Finnish–English bilinguals and presented them with 80 words (20 positive, 20 negative, 20 taboo and 20 neutral) written in four colors, namely yellow, red, blue and green (once per color), in the two languages spoken by the participants. The items were presented in blocks, grouping the items of a single condition and in a single language. Participants thus saw eight blocks, one for each type of item, in one language (L1) and then in the other (L2). We will return to this notion of a block in Chapter 6. The results showed that the participants named the color of negative and taboo words more slowly than that of neutral words. Contrary to what had previously been shown at the discourse level, the effects were similar in L2 and L1, suggesting that the processing of emotionally negative words in L2 does not differ from that in L1, at least among proficient bilinguals.

5.7. Verification task

We now turn to a task specially designed for studying the comprehension of sentences and texts: the verification task. In this task, participants are presented with sentences or short passages, and then asked to indicate whether an item was present in the text or not. The item in question can be of different types, such as a written word or a picture, with specific properties. The nature of the item depends on the research question, as is illustrated below. The response and the response time are recorded in the same way as in the tasks previously discussed.

This type of paradigm makes it possible to investigate the nature of representations constructed while reading. As we have already discussed in this book, mental representations have perceptual properties. Therefore, they are not simply conceptual, but related to our experience. This connection can be revealed, for example, through the use of tasks combining texts and pictures, as proposed by Madden and Zwaan (2003), for exploring the influence of tense categories on mental representations. Madden and Zwaan focused on verb aspect, comparing the perfective aspect, which conveys the fact that an action has been completed, and the imperfective aspect, which conveys the fact that an action is ongoing. For example, in English, the past tense (Eva wrote a book) is perfective, whereas the past progressive (Eva was writing a book) is imperfective.

Madden and Zwaan (2003) constructed 26 experimental sentences describing a character involved in an action that implied duration and obligatory end points. For example, painting your house or going to work implies a certain result (the house is painted, the person has arrived at work). Every chosen action was described in the past tense and in the past progressive. For each sentence, two pictures, one corresponding to the completed action and the other to the action in progress, were created. The sentences and pictures were combined to create four possible conditions: past tense and completed action, past tense and action in progress, past progressive and completed action, past progressive and action in progress. Four lists of items were constructed, each containing a sentence–picture combination, so that every participant saw all the conditions throughout the experiment, but saw every item only once. The task of the participants in this experiment was simply to read each sentence and then to decide, when the picture appeared on the screen, whether or not it represented the scene described in the sentence. Replies and response times were recorded.

A very interesting element of this experiment lies in the fact that the two pictures were compatible with the situation described in the sentence even though one of the two pictures was more suitable than the other, in terms of the verb aspect. All experimental items were intended to elicit a YES decision. The authors’ hypothesis specifically concerned response times, namely that these times would be shorter when the picture was consistent with the aspectual information conveyed by the verb. In order to offer participants the possibility of responding NO, and thereby accomplishing the task, 26 filler sentences were constructed with the same structure as the experimental sentences, followed by a picture which did not correspond to the scene at all.

An analysis of the YES responses showed that, in general, responses were faster in the congruent condition, but that this effect was mainly driven by the perfective sentences. In other words, when the sentence conveyed the idea of a completed action, participants were faster to respond YES to the pictures containing the completed action compared to the ongoing action. The research hypothesis was therefore confirmed, but only partially. Madden and Zwaan (2003) suggested two explanations for these results. The first one argued that it was possible that when reading the description of an action in progress, the participants represented such an action at different stages, and that the picture representing the action in progress did not really correspond to any of these stages. The second explanation suggested that the imperfective aspect of the verb could have encouraged the participants to represent all the stages of the action, leading them to accept both pictures as representing the situation.

5.8. The self-paced reading paradigm

In this section, we will discuss the paradigms that are specifically related to the reading process. The first of these, the self-paced reading paradigm, invites participants to read sentences or texts, either word-by-word, or in segments, or sentence-by-sentence. The participants indicate that the word, the segment or the sentence has been read by pressing a key, which makes the text disappear and the next element appear. In this task, participants can read at their own pace. The instructions generally encourage them to do so in the most naturally possible way, for them to properly understand the text presented.

In order to ensure that the participants read the texts properly, it is necessary to include comprehension questions about the elements that have just been read. These questions are generally associated with filler items, aimed at concealing the experimental manipulation from the participants. It is important to construct suitable questions, neither too simple nor too complicated, since the characteristics of these questions may influence the strategies implemented by participants while reading (Jegerski 2014). Comprehension questions generally appear for only a number of the items.

The advantages of the self-paced reading paradigm are firstly the possibility of getting access to an online measurement of comprehension and secondly the collection of reading times related to segments of text, or even to every word in a text. In addition, this technique is non-invasive, and relatively simple to set up and use.

The disadvantages of this method are mainly related to the fact that the words or segments of text already processed disappear as the text unfolds This does not allow the normal text processing involved in natural reading, which we discuss in further detail in the next section. For the moment, suffice it to say that readers sometimes go back to the text in order to verify information, or to read certain passages over again. By preventing them from going backwards, the method involved in the self-paced reading paradigm requires larger memory capacities than those involved in natural reading. For this reason, the reading times collected in experiments are generally slower than those one would observe in natural reading.

The characteristics of reading should also be taken into account when creating experimental items used in a self-paced reading experiment. For example, it is known that the last segments of a text are read more slowly than the others, due to the finalization of the mental representation. At this point, the different pieces of information included in the representation are linked to one another in order to create a thorough representation. This is why critical elements should not be placed at the end of an item, so as not to confuse the potentially obtained effects with integration effects. Spill-over effects are also common in this type of paradigm. These effects correspond to the fact that the processing of a word or of a segment is not always finished when the person goes ahead with the text, and can continue while processing the next word or segment. For this reason, it can be useful to analyze not only the times related to the critical word or segment, but also those relating to the words or segments directly following the critical sections.

In order to illustrate the use of this method, we will present the study by Kelter et al. (2004, experiment 1), dealing with the influence of the recency of an event on its activation in the mental representation of the text. Various studies have indeed shown that it is not the event’s recent mention in the text that plays a role in its representation, but its recency regarding the current situation described in the text. For example, in a verification task, Carreiras et al. (1997) showed that readers more quickly recognized a role name such as the baker or the teacher introduced in a story when the role was associated with the protagonist in the present rather than in the past. In their research, Kelter et al. examined the activation of past events depending on the time elapsed between the event and the present.

In order to do this, short stories introduced a situation before presenting a first event. Then, a second event was described, which was either long or short. Finally, a third event was presented before the story mentioned a sentence referring back to the first event. The story ended with one or two neutral sentences. For example, one of the stories described a couple getting ready to celebrate Christmas. The husband informed his wife that he disagreed with her choice of Christmas decorations. She got angry. Then, she went into the kitchen and, in the short-term condition, put some cookies on a plate, whereas in the long-term condition, she baked these same cookies. The story continued with a description of the smell of cookies and the Christmas spirit that emanated from them. The target sentence then appeared, referring to the anger experienced earlier by the woman: “At that time, she regretted her anger”.

Participants simply had to read the stories sentence-by-sentence, and the reading time for each sentence was recorded. In order to examine the potential spill-over effects of the experimental sentences on the following reading times, analyses were carried out not only on the target sentences, but also on the filler sentences that followed the experimental ones. While reading times for the filler sentences did not differ between conditions, reading times for the target sentence were slower when a long-term event was described in comparison with a short-term event, as expected. These results support the idea that the organization (and accessibility) of consecutive events in the mental representation of readers reflects the temporal references transmitted by the text.

5.9. Eye-tracking

The eye-tracking technique is similar to the previous paradigm, in that the participants have to read words, sentences or text excerpts. The difference lies in the type of measurement used: this time, this is the participants’ gazes’ direction while reading that is observed. To do this, the participants have to look at a screen on which the stimuli are displayed. Eye movements are recorded by a camera placed in front of them while they are processing the stimuli. A light source illuminates the eye, causing reflections in the pupil and on the cornea, which are detected by the camera. On the basis of the reflections, it is possible to infer the direction of the gaze with impressive spatial (0.5 degree) and temporal accuracy, a measurement being made every millisecond.

When we read, our eyes move forwards and backwards, and dwell for a longer or a shorter period of time on certain words. Contrary to popular belief, eyes do not scan the text in a regular manner while reading. They dwell on certain words and then quickly move on to other words. We refer to saccades when speaking about eye movements, and to fixations when the eyes remain motionless for a short time. During fixation, which lasts approximately 200–300 milliseconds, information can be retrieved and processed, and information processing goes on during the following saccade. The reason for this succession of fixations and saccades is that the acuity of our visual field is high in the central area of the eye, the fovea, but decreases very quickly in the parafoveal and peripheral regions. In order to process words, it is therefore necessary to bring them to the center of the fovea. Moreover, not all words are fixated during reading. Content words are fixated 85% of the time, whereas function words are fixated only 35% of the time, and short words are often skipped (Rayner 2009).

Eye-tracking can be done in different ways. The most natural way is to present the text and to record the eye movements made over it. It is also possible to adapt the text presentation to the person’s eye movements, using the moving window technique (McConkie and Rayner 1975). In this case, only a portion of the text is presented, whose center corresponds to the fixation point and whose width (a certain number of characters) is manipulated in the experiment. The rest of the text is replaced either by the same character (an X, for example) or by other characters such as letters which may be visually similar to the original letters, or not. This technique made it possible to define the size of the perceptual span, which is three to four letters to the left and 14–15 letters to the right of the fixation point, for languages such as English and French. We can see that the perceptual span is rather narrow and that it is asymmetrical, that is, that it is larger towards the side where the eye naturally goes while reading. Another technique used in eye-tracking corresponds to the foveal mask (Rayner and Bertera 1979). In this case, the portion of the text around the fixation point is hidden, whereas the rest of the text remains visible.

On the basis of eye movements, it is possible to study the number and duration of fixations, the size of saccades and (backtracking) regressions, as well as their starting and finishing points, among other measures eye-tracking can offer. When using this method, the sentences or texts are often divided into areas of interest, grouping together the segments manipulated during the experiment that will be compared between the conditions. The variables frequently used for the aggregation of fixation points into reading measurements are numerous. For example, the first fixation duration corresponds to the duration of the first fixation on a word or area of interest. The first-pass reading time or first-run dwell time is the total time spent on a word or area of interest before the gaze goes to the right or to the left. The regression path duration corresponds to the time spent on a word or area of interest before leaving it to the right and includes the time spent re-reading previous portions of the text. The total reading time or dwell time is the sum of all fixations made on a word or area of interest, including regressions from other portions of the text. These variables give different clues as to the processes involved in reading. Some of them, such as the first fixation durations or the first-pass reading time, are considered as early processing measurements, whereas others, such as the total reading time, correspond to late measurements reflecting more elaborate processes (Staub and Rayner 2007).

Measuring eye movements is based on the idea that the time spent on a word corresponds to the time needed for processing this word (Just and Carpenter 1980). This idea is nonetheless questioned by the fact that eye movements can hardly account for the cognitive processes involved in the interpretation of the information just read. As a matter of fact, they partly depend on uncontrollable processes, such as the size or the speed of saccades that lead the eye to land more or less accurately on a given word. Skipped words are also processed, probably in a peripheral way, even if there is no fixation to objectively attest for this. We will not discuss these limitations in further detail, since they go beyond the scope of an introductory book. What is important to remember is that there is no perfect concordance between fixation time and processing time for a word or group of words.

Analyzing eye movements is useful for studying a wide variety of processes. As we saw above, it can contribute to a better comprehension of the natural reading process, as well as the basic characteristics of eye movements. The influence of many variables on the reading process, such as word frequency, contextual predictability, word length or polysemy, could also be investigated thanks to this method. The analysis of eye movements is also interesting for other levels of language processing, such as sentence or text comprehension, the development of mental representations or even the pragmatic processes involved in discourse comprehension.

Measuring eye movements also has the great advantage of being much closer to natural reading than self-paced reading. It also offers the possibility of obtaining fine measurements of the time spent on words or segments, as well as an indication of the processes underlying comprehension while reading, thanks to the observation of regressions, which are not accessible through other methods.

From a technical point of view, measuring eye movements is complex. It requires mastery of the necessary tools, as well as great accuracy in measurement. Accuracy can be granted thanks to prior instrument calibration, but the presence of the experimenter is required at all times to verify the quality of the measurements as the experiment progresses.

From a methodological point of view, the characteristics of natural eye movements while reading should be taken into account while creating the linguistic material. It is essential to verify that the target interest areas are comparable, in particular in terms of positioning on the screen, and that they do not appear at the start of the line, which is where the gaze position is generally adjusted.

The main disadvantage of this measurement is the cost of the equipment required, as well as the time cost, as it is only possible to test one person at a time. Once the data have been acquired, their processing also requires advanced technical and statistical knowledge. The large amount of measures recorded in an eye-tracking experiment may also become a disadvantage in some cases, since the different measurements sometimes offer different results which are not easy to interpret. During the analysis of numerous indicators, it is also likely that one or the other may look different from condition to condition. It is therefore highly advisable to define which indicators will be observed beforehand, as well as the specific hypotheses related to such different indicators. Without doing this, there is a risk of finding spurious results, stemming from the accumulation of statistical tests which increases the probability of finding a difference that does not reflect a real effect.

As an example of the application of the eye-tracking technique, we will present the study by Gordon et al. (2006, experiment 1) on how complex sentences are processed while reading. Their research focused on the comprehension of complex syntactic structures, requiring the simultaneous activation of two noun phrases (NP), before being able to associate them with the different expressions of the sentence. Gordon et al.’s specific hypothesis concerned the similarity between these NPs, the idea being that NPs of the same type could cause interference in their processing. A previous study measuring reaction times demonstrated that retrieving the object NP as in (2) leads to longer reading times than retrieving the subject NP, as in (1). This effect was also larger when the subject and the object were semantically similar (barber-tailor, John-Bill) (Gordon et al. 2001):

1. (1) It was the barber/John who saw the lawyer/Bill in the parking lot.
2. (2) It was the barber/John who the lawyer/Bill saw in the parking lot.

In order to study the comprehension of such structures in a more natural way than in a self-paced reading task, and to determine at which point in the sentence readers find it difficult to deal with such structures, Gordon et al. (2006) used the eye-tracking methodology.

Their participants had to read sentences containing a relative clause (RC) associated with the subject of the sentence and whose relative pronoun was either in the subject position (3) or in the object position (4) of the clause. The RC also contained either a proper name or a role name. Each sentence could thus appear in one of the four versions, two of which contained the same type of NPs (role names) and two contained different types of NPs (role name and proper name). Each sentence appeared isolated on the screen, and the participants’ eye movements were recorded until they indicated that the sentence had been read, by pressing a button. At that point, a comprehension question could appear (in 15% of cases), which the participants simply had to answer orally with YES or NO:

1. (3) The banker that praised the barber/Sophie climbed the mountain just outside of town.
2. (4) The banker that the barber/Sophie praised climbed the mountain just outside of town.

The eye movements associated with the area of the relative clause (from the relative pronoun until the main verb, without including it), as well as those associated with the verb of the main clause, were analyzed. As it had been previously demonstrated in the literature, the results confirmed that relative clauses with object NPs were read more slowly than those with subject NPs. Likewise, the reading time for the verb of the main clause was longer in the object condition than in the subject condition. The difference between these two conditions was itself larger when the NP of the relative clause was a role name (similar to that of the main clause), than when it was a proper name. These effects emerged in early processing measures (first fixation duration and first-pass reading time), suggesting that the type of NP influences its processing, as well as its integration into the main sentence, as soon as it appears. Later processing measures showed that the rereading of the target areas was also influenced by the variables examined.

In summary, eye-tracking, as well as self-paced reading, enables us to study many aspects of reading, from the simplest level of word processing to the more complex level of discourse comprehension. Given the large variety of measures that eye-tracking allows us to collect, this methodology could be interpreted as being more advantageous and interesting, prima facie. The choice to use eye-tracking rather than self-paced reading should, however, be made on the basis of the processes we want to observe and the possibility of investigating such processes that each methodology offers. In cases where the experimental design makes it possible to investigate a question using the self-paced reading paradigm, resorting to the eye-tracking method – with all the technical difficulties it entails – could end up being superfluous.

5.10. The visual world paradigm

To conclude, we present an experimental technique that makes it possible to study comprehension of spoken language, by means of the visual world paradigm. In this paradigm, participants listen to linguistic stimuli while looking at a scene, objects or words on a screen, while their eye movements are recorded. The participants’ task may simply be to listen to a sentence or text while watching a scene, and then to attach any of the objects or words, according to the instructions received. This paradigm is based on the assumption that when we process speech, at the same time as observing a scene or pictures, we tend to relate what we hear to what we see. The eye movements of the participants involved in a visual world task reflect the attention devoted to the different objects or parts of the scene, according to the linguistic content heard. Somehow, this paradigm makes it possible to observe how people interpret the flow of discourse and to establish what they predict on the basis of what they hear.

In this paradigm, the most frequently used measurements are related to the specific regions of the screen participants look at during the task, specifically after listening to a target word. Common measurements are the fixation proportions on areas of interest or the number of saccades directed towards them. Of course, the time window in which fixations, saccades or regressions are analyzed must be defined depending on the research question and the process investigated.

The visual world paradigm makes it possible to study language comprehension at all levels. At the level of word comprehension, research has been carried out, for example, on the areas of phonological processing, word recognition by bilingual speakers, and the effects of context on word recognition. At the discourse level, this paradigm has made it possible to better understand the role of lexical and structural constraints in sentence comprehension. It has also been useful for examining questions related to pragmatics or to linguistic relativity (for a review, see Huettig et al. (2011)).

The advantage of this method is that it allows us to assess language comprehension without requiring reading skills, the use of written material or even metalinguistic abilities. Participants simply look at a screen while words or sentences are presented to them. For this reason, this method is very useful for studying children or people with written language impairments, such as illiterate people.

As an example, we present the study by Engelen et al. (2014) on the resolution of anaphora in children’s narrative comprehension. This study specifically investigated the ability of children, aged 6–11 years, to determine and follow the character being referred to as they listened to a story. It also had the particularity of adopting a natural approach, presenting a story which lasted almost eight minutes, rather than many unrelated items, as is common in most experimental studies. Furthermore, children were split into groups on the basis of their comprehension of the story, which assessed their memorization of literal information and also inference-based information.

The story, told in Dutch, contained four characters (a hedgehog, a rabbit, a squirrel and a mouse), presented simultaneously on a screen in front of the children. The characters had human-like attributes, such as being able to talk. They were all masculine, so that the grammatical gender of a pronoun could not be used as a cue, and the anaphoric pronoun could refer to any of them. In the story, the characters were introduced and then started to interact with one another. A portion of the story is reproduced below (the words in italics are those for which eye movements were analyzed):

“Meanwhile at the lake, the rabbit and the squirrel were sitting on a log. The squirrel wanted to do a running contest with the rabbit. ’I’m sure that from here I can run to the giant rock and back faster than you’, he said. ’Well, let’s see’, the rabbit said. ’Okay’, the squirrel said. ’I’ll count to three and at three we run’. They both got on their marks. The squirrel started counting: ’One… two… three!’ The rabbit dashed away with great speed. But what did the squirrel do? He just stayed there. The rabbit didn’t notice anything and rushed on. The squirrel lay down on the log in the sun. He thought it was a good joke and knew what he’d say when his friend would come back.” (Retrieved from Engelen et al. (2014))

All of the referential expressions in the text could not be analyzed due to the varying levels of difficulty they presented. For this reason, Engelen et al. (2014) chose 42 expressions (28 names and 14 anaphoric pronouns), for which eye movements were analyzed and compared between the groups of children. The main hypothesis was that comprehension of anaphoric pronouns and the ability to follow the protagonist of a story depend not only on literal, but mostly on inferential skills. It was expected that children with good skills in these two areas would look towards the picture representing the character in question when this was referred to by a name or an anaphoric pronoun. For children in the middle group, the glances towards the corresponding picture should be more numerous when the character was designated by a name, rather than by an anaphoric pronoun. Finally, for children with poor comprehension skills, it was expected that they would generally look less at the pictures related to the characters than the other two groups.

When analyzing the results, only two groups of participants could be constructed on the basis of their answers to the comprehension test: a group with good literal and inferential skills and a group with poor skills. As no child revealed skills for being placed in the intermediate group, eye movements were eventually compared between these two groups.

Eye movement analyses were performed on two-second time windows from the appearance of the name or the anaphora. Compared to children with poor comprehension skills, the group with good skills looked at the picture associated with the reference character more, either after hearing names or anaphoras. Interestingly, this difference did not result from the adjustment of the gaze following the hearing of the name or the anaphora, but from the probability of fixating the target picture at the time of hearing it. In other words, the group with good comprehension skills was more inclined to make fixations on the target picture, in advance, and even more when this picture was referenced by an anaphora. According to the authors, and based on other observations (Barr et al. 2011), this reflects the expectations of good comprehension, in terms of the unfolding of a text. This study therefore suggests that children with good comprehension skills are those who can anticipate the content of the text.

5.11. Conclusion

In this chapter, we reviewed the different methods for studying online comprehension. We first presented think-aloud protocols which allow us to access people’s thoughts and reflections during the comprehension of a text. This method is particularly useful for identifying the stages of comprehension, but it has many limitations, in particular due to unnatural protocols. We also discussed the interest of measuring the response time as an indicator of comprehension, especially when studying processes or representations that are inaccessible to consciousness, or when we wish not to draw the attention of participants to the object of study. Response time can be collected while performing various tasks, specifically assessing one or many of the processes involved in language comprehension, and thereby leading to different conclusions being drawn. Most of the tasks used in experimental linguistics have in common the activation of certain characteristics which are present in linguistic stimuli, and which can be shown thanks to priming and interference effects.

5.12. Revision questions and answer key

5.13. Further reading

For more developments on reading time based methods and visual attention based methods, the reader may refer to Kaiser (2013) and Jegerski (2014). Rayner (1998) and Clifton et al. (2007) are references in relation to eye-tracking, its use, as well as the results obtained in the fields of word recognition and sentence comprehension. Staub and Rayner (2007) offer a more accessible text for beginners. For more developments on the visual world paradigm, see Huettig et al. (2011). Finally, we recommend reading the book by Gonzalez-Marquez et al. (2007b) which offers numerous examples of the application of the techniques discussed in this chapter, in the field of cognitive linguistics. For a contribution comparing the interests of online and offline measurements in the study of comprehension, see Ferreira and Yang (2019).