AN ALIENCE E EASURED IN UPIL IZE C S B M P S ?

(1)

C

AN

S

ALIENCE

B

E

M

EASURED IN

P

UPIL

S

IZE

?

Vincent Boswijk

S2217325

MA in Multilingualism,

Faculty of Arts, University of Groningen

Date: 28-07-2017 Word count: 14100

(2)

A

CKNOWLEDGEMENTS

An intensive half year lies at the basis of this thesis. It has been a period during which I have got to learn many new things and got to experience a new side of academia I would not have thought suited me in the way it turned out to. It has certainly helped to develop my academic skills to new heights. This would however have been impossible with all the help dr. Hanneke Loerts and dr. Nanna Hilton provided during this research project. I want to thank them for their guidance during this project. I also want to thank prof. dr. Petra Hendriks for being willing to serve as a second reader for this thesis.

The project would not have been what it turned out to be without the voice of Marlou Berends, and for that I would like to thank her. It has brought me a lot of fun to listen to all of the (mostly positive) things people had to say about her voice and intonation.

(3)

A

BSTRACT

Can Salience Be Measured in Pupil Size?

Certain aspects of language are more noticeable to the listener than others. Scholars have suggested that this prominence, termed linguistic salience, plays a role for processes of language change (Kerswill & Williams, 2002) as well as for second language acquisition (Smith, 2012) in that salient forms are e.g., more likely to undergo change, or are often acquired earlier than other features. Yet, what makes a linguistic feature salient is a contested issue: some argue that salience can be defined by linguistic traits such as loudness, high word-frequency, or a large articulatory effort, whereas others argue that salience is created by association with social factors (cf. Kerswill & Williams, 2002).

While an exact and univocal definition of the concept of saliency is currently lacking, recent technological advances allow for exploring the potential relationship between salience and cognition. While these advances cannot answer univocally the question of what salience is, the possibility of having a numeric measure of salience can bring us closer to measuring its relationship with linguistic and social factors.

(4)

T

ABLE OF

C

ONTENTS

Acknowledgments 2 Abstract 3 Table of Contents 4 1. Introduction 5 2. Literary review 7

2.1 Grasping the concept of language change 7

2.2 The relationship between salience and language change 12

2.3 The relationship between salience and language learning 13

2.4 Eye tracking and pupillary responses 13

2.5 Research questions 16

3. Materials and Methods 17

3.1 Participants 17

3.2 Materials 17

3.3 The design of the experiment 23

3.4 Procedure 25

3.5 Data processing 26

4. Results 29

4.1 Dilation data 29

4.2 Qualitative interviews about salience 35

4.3 Taking the interviews into account 36

5. Discussion 40

5.1 Research question 1: Can salience be measured in pupil size? 40

5.2 Research question 2: Can salience be seen as extra cognitive load? 41

5.3 Reviewing the definition of salience 42

5.4 Further discussions 43

6. Conclusion 45

References 47

Appendix 1: Questionnaire (in Dutch) 50

Appendix 3: The stimuli and lists 52

Appendix 4: Interview questions (in Dutch) 54

(5)

1. I

NTRODUCTION

In their pioneering work Weinreich, Labov and Herzog (1968) formulate five larger questions that sociolinguistics needs to address to reach a theory that accounts for language change. Of these five, the most fundamental one concerns the still unsolved ‘actuation problem’ (Weinreich, Labov, & Herzog, 1968, p. 102), which asks why and how language change starts out. In the half century that has followed after the pioneering paper, one of the issues that arises when answering this question, is what it is, exactly, that makes one linguistic feature more prone to language change than another. A number of scholars have reached the conclusion that a feature’s degree of salience is what determines its ability to change (cf. Kerswill & Williams, 2002). However, what the concept of salience is, exactly, is something scholars still disagree on.

Nevertheless, recent technological advances allow for exploring the potential relationship between salience and cognition, which is why I have chosen this subject. Although these advances cannot answer univocally the question of what salience is, the possibility of having a numeric measure of salience can bring us closer to measuring its relationship with linguistics and cognition. There is still a lot that is unclear when it comes to this topic and I hope that with this thesis I can contribute to discussions in the field.

In this thesis, the focus will not be on finding an all-solving definition for the concept of linguistic salience and, in doing so, end the ongoing debates in the field. Likewise, the focus will not be on the process of language change itself. Instead, it will focus on the cognitive aspects of linguistic salience. The goal of this, is to create a better understanding of what salience means for the way in which we process language and in what way these processes play a role in the onset of language change. In an attempt to do this, the main research question for this thesis was: Can salience be measured in pupil size? It was hypothesized that the processing of auditory salience can be seen as extra cognitive load, and as such can be measured by an increase in pupil size.

(6)

understand what the cognitive correlates of salience are to, in turn, try to understand its role for language change.

The first part of this thesis is a discussion of the existing literature in the field. In this section, the concept of salience as it is currently understood is discussed, as well as the current usage of eye-tracking techniques. The next chapter discusses the methodology and materials used in this research project. After that, our results are presented. Next, the results found during this project are discussed in the section that is called discussion. Finally, the conclusion to this thesis project will be presented.

Appendices to this thesis are not only attached at the end of this thesis, but are also collected in a Dropbox folder. The folder can be found by copying or filling in the following link in your web browser:

(7)

2. L

ITERARY REVIEW

Before looking into the possible relationship between pupillary responses and linguistic salience, it is of importance to discuss the concept of salience. We will first establish a working definition of salience. In order to do so, this chapter will first look into the historical development of the concept of salience. Furthermore, the concept of pupillary responses will be explored to justify the methodological techniques used in this thesis.

2.1 G

RASPING THE CONCEPT OF

S

ALIENCE

The concept of salience goes back to Schirmunski’s (1930) notion of Auffälligkeit (Kerswill & Williams, 2002), and although the concept is used by many scholars, defining it remains difficult. Many have tried to establish the criteria for salience (e.g., Auer, Barden & Grosskopf (1998) and Kerswill and Williams (2002)), but it is known to be “notoriously difficult to quantify” (Hickey, 2000).

In his work, Schirmunski (1930) tried to make a distinction between two kinds of linguistic features, those that are subject to change or loss (which he calls “primary” or “salient” variables) and those that are more constant (“secondary” or “non-salient” variables) (Auer et al., 1998, p. 164; Kerswill & Williams, p. 83, 2002; Rácz, 2013, p. 28). This distinction can be seen as strongly resembling Labov’s distinction between indicators and markers (Rácz, 2013, p. 28), in that indicators have ‘no social interpretation’ and as such do not raise any notions of value that attach to the way someone speaks (i.e., secondary or non-salient variables), and that markers do convey notions of identity and belonging (i.e. primary or salient variables) (Ibid., p. 25). To establish what it is that makes a feature ‘primary’ or salient, Schirmunski (1930) presented six criteria that need to be fulfilled:

1. Articulatory difference (i.e., there is a distance between two variables)

2. Lexicalisation (i.e., strict rules determine what phonological components can fit together (Siegel, 2010))

3. Categoriality (i.e., whether there is a categorical phonetic difference (Rácz, 2013, p. 29)) 4. Awareness (i.e., speakers are aware of the variability)

5. Writing (i.e., the variability is used in writing)

(8)

Schirmunski’s criteria seem to be largely based on a relationship between dialects and a standardised language, in that he for example focusses on lexicalisation, use in writing and comprehension. I find it hard to believe that these are fixed requirements. Even if one would not understand a certain aspect of some other groups’ language, I believe it can still be perceived as salient, in that it is something unexpected or new. Indeed, Auer et al. (1998) found no strong correlation when it comes to lexicalisation and adaptation of linguistic variables (Rácz, 2013, p. 28). As we will discuss later on in this section, there are others who don’t agree with a fixed set of criteria that all have to be fulfilled, but believe in a more loose set of criteria that might all, in turn trigger salience.

We will see that many others came after Schirmunski. There seems to be a consensus that the definition provided by Trudgill’s Dialects in Contact (1986) is the most careful way of describing and applying the concept of salience (Kerswill & Williams, 2002; MacLeod, 2015). In his work, Trudgill describes salience as something that "attaches" to a linguistic variable and turns it into a marker, or in other words, “a linguistic variable that is subject to social class and stylistic variation at the same time” (Trudgill, 1986, p. 11). Trudgill suspects that whether a feature is prone to processes of imitation (eg., accent imitation) or not, might be a good indicator of that feature’s salience (Ibid., p. 17), and as such, he states that whether a feature is salient or not can be determined by examining the process of imitating (Ibid., p. 37). Furthermore, he states that in language contact, salient features of the target variety are more accommodated to than non-salient features (Ibid., p. 37). In order to establish whether a variable is salient or not, Trudgill proposes the following list of criteria that need to be fulfilled, and it is his list of criteria that has often been used as a starting point in naming criteria for salience: (Trudgill, 1986, p. 11; Kerswill & Williams, 2002, p. 89; Rácz, 2013, p. 27):

1. have a stigmatized variant 2. have a high-prestige variant 3. are undergoing change 4. are phonetically different

5. are maintaining contrasts in phonology

(9)

leading to the situation where a linguistic variable may be called salient if it has certain properties, but, it is however the same set of properties that follows from being salient (Ibid., p. 28).

Instead, Rácz (2013) has proposed to divide the concept of salience into two. On the one hand, we have what he calls cognitive salience, following from a feature’s rate of surprisal. On the other hand, he identifies social salience, which follows from a feature’s social variation (p. 23). This is in contrast with the theories of Trudgill (1986), who has stated that salient features always are markers, and as such subject to both social class and stylistic variation. By dividing the concept of linguistic salience into two, Rácz (2013) creates two levels of salience. As such, the range of what might be called a salient feature is expanded. Rácz's definition of salience is based on the same division. He states that "a segment is cognitively salient if it has a large surprisal value” (Rácz, 2013, p. 37). Furthermore, he states that “a variable that has cognitively salient realisations can be a marker of social indexation, becoming socially salient for the members of the language community" (Ibid.).

Another way of determining a feature’s level of salience is presented by Hickey (2000). According to him, salience is a speakers’ level of awareness towards a linguistic variable (Hickey, 2000, p. 57). In his article, he provides what he calls ‘a preliminary checklist’ for determining salience (Hickey, 2000, p. 62). Note that instead of claiming that a feature has to fulfil all of the criteria on the list, as we have seen with Schirmunski, Trudgill and Auer et al., Hickey merely states that his list is a checklist of criteria, that all may result in salience, even separately. Hickey presents the reader with the following criteria (Hickey, 2000, p. 62-63):

1. Acoustic prominence 2. Homophonic merger 3. System conformity 4. Deletion and insertion 5. Grammatical restructioning 6. Opennes of word-class

7. The loss of vernacular features 8. Retention of conditional realisations

(10)

adds, are more important than the others, and it is usually a combination of several of the triggers that leads to a variable becoming salient (Hickey, 2000, p. 62). It is because of his different attitude towards the different criteria, that this approach has earned a separate status in this overview. Although it is similar to the criteria-list approach, both approaches deal with their criteria differently.

Drager & Kirtley (2016) state that it is more likely for salient variables to gain attention. They add however, that this is not the case if a feature that may be theoretically salient, is perceived as a mistake by the listener. It is then considered to be classified to other pre-existing categories in the brain (Drager & Kirtley, 2016, p. 14). Although this is an interesting theory, that might prove to have implications for this research project, it remains unclear how Drager & Kirtley arrive at this conclusion and whether or not this statement is reliable, is therefore hard to tell at this point. If their theory holds however, we would have to expect no response when people express they perceive the variable in question as a mistake.

Years of academic research have resulted into different approaches to salience. MacLeod (2015) points to what she calls the criteria-list approach and the experimental approach. While the first approach evaluates salience based on the fulfilling of certain criteria, the experimental approach entails more qualitative views on the matter of salience. When trying to quantify the concept of salience, I believe a combination of these two is in order.

As examples of the criteria-list approach, MacLeod (2015) mentions Trudgill (1986), Auer et al. (1998), Hinskens (1996) and Schirmunski (1930). Downsides of this approach are that it is not always clear in what way the criteria are to be applied and that extra-linguistic factors such as gender and level of education are not taken into account (MacLeod, 2015, p. 84).

In the experimental approach, on the other hand, a linguistic variable is salient when it influences speakers’ attitude towards different kinds of speech, either consciously or subconsciously (MacLeod, 2015, p. 85) Salience can then be measured by qualitatively investigating if the linguistic variable is associated with any social indexation. In this experimental approach, linguistic salience should not be seen as a binary property (i.e., a variable is either salient or non-salient) but rather as a gradient one where different variables can have varying levels of salience (MacLeod, 2015, p. 85).

(11)

at the accuracy with which participants were able to correctly categorize trials, which they did in 30% of the trials. Another example is the study of Torbert (2004), who looks into the perception of ethnicity and regional dialect in Southern US English. More specifically, he looks at different realization of /ai/ and /o/. Participants were then asked to judge speaker’s ethnicity on a Likert scale. What is important with the examples mentioned by MacLeod, is that the experimental approach helps generating numerical values to measure salience.

The experimental approach has several advantages. First of all, it is more sensitive to salience that is subjective to context (MacLeod, 2015, p. 85). Also, it is easier to determine statistical significance, because this approach gathers numbers, such as the studies in the examples above (Ibid.). There are however also downsides to this approach. The preparation of one’s experiment and planning the procedure is for example harder than it is when using the criteria-list approach (Ibid.).

Although no univocal definition of salience is currently available, most of those working on the concept, be they propagators of the experimentalist- or the criteria-list approach, seem to agree on what linguistic salience entails. And so, when trying to find a short definition of salience, we can state that it has something to do with ‘standing out’ (Rácz, 2013). It is deemed to be a ‘cognitive or perceptual prominence’ or related to ‘high frequency’ (Kerswill & Williams, 2002). A salient form is ‘noticeable or conspicuous’ (Siegel, 2010), and there is ‘greater awareness’ involved as compared to other variants (Trudgill, 1986). Others point to relationships between salience and ‘loudness’ (Liao, Kidani, Yoneya, Kashino, & Furukawa, 2016), ‘noticing’ (Smith, 2012) and ‘rate of surprisal‘ (Rácz, 2013).

(12)

Different ways of finding out whether or not a feature is salient were presented in this section, and although there are a lot of similarities between them, their point of view and outcome differed. It is my opinion, that a combination of these points of view is in order. The criteria-list approach and the criteria-list of triggers presented by Hickey (2000) can help us to establish different categories of salience and since it is a numerical representation of salience we are interested in, the study fits well to the experimental approach described by MacLeod (2015).

The literature about salience is diverse and one can point out many reasons for a variable to become salient. As I wished to represent this diversity in the study, I decided to include different categories in which we can find different forms of salience. The main goal with this study was to see whether or not we could measure salience. However, since salience has no general definition that all agree on, it seemed logical to include different points of view in this study.

2.2 T

HE RELATIONSHIP BETWEEN

S

ALIENCE AND LANGUAGE CHANGE

As mentioned in the introduction to this thesis, Weinreich, Herzog and Labov (1968) laid out a number of big questions concerning the field of sociolinguistics. Several scholars in the field suggest that one of these questions, seeking why language changes, might be better understood by identifying salience. Amongst these are Kerswill and Williams (2002), who discuss salience in relationship to language change, particularly in a setting of dialect contact. Building on theories of amongst others Trudgill and Siegel, they state that by understanding salience, we might “shed light on the interplay of internal factors and the large range of other factors which impinge on language change” (Kerswill & Williams, 2002, p. 83).

Also one of the scholars theorizing about salience and language change, is Rácz. In his work

Salience in Sociolinguistics (2013), an entire chapter is devoted to this relationship, focussing

on sound change in particular. While building upon models of language change conveyed by Baxter, Blythe, Croft & McKane (2009), Rácz states that a features social salience plays a role in its “propagation” and that by understanding the concept of social salience, we may better understand the dynamics of language change (Rácz, 2013, p. 148).

(13)

underlines however, that speakers not commenting on a feature does not necessarily mean that these features will not result in change. It can in fact be the case that it is exactly with these features where we will be able to find variation (Hickey, 2000, p. 58).

2.3 T

HE

R

ELATIONSHIP BETWEEN

S

ALIENCE AND

L

ANGUAGE

L

EARNING

As stated before, it is not only in language change that salience plays a role. It is suggested that it is also important in language learning. Bardovi-Harlig (1987), for example, states that, in second language learning, salient features are learned faster as opposed to variants that are not salient. In her study, she investigated the rate with which participants acquired both marked (salient) and unmarked (non-salient) constructions, showing a significant difference she ascribed to salience.

Siegel (2010) mentions that salience affects the degree of dialect acquisition (p. 120). He states that salient features occurring in a D1 are more likely to be given up in situations of language shift. This is illustrated by for example Hiramoto (2010), who points to how Japanese immigrants in Hawai’i abandon stigmatized (and as such salient) features from their dialect and replace them by features from the standardized language.

Like Siegel, Auer et al. (1998) also point to an effect of salience in the opposite direction of Bardovi-Harlig’s findings. For example, when considering dialect shift, items that are subject to a high degree of lexicalisation (and thus salient), are most likely to be given up in a language shift (Siegel, 2010). When looking at language or dialect acquisition on the other hand, these features have shown to be harder to acquire (Auer et al., 1998; Siegel, 2010), and not, as Bardovi-Harlig found, acquired faster. Siegel adds that the same holds for salience based on stigmatization and stereotyping, how he comes to this conclusion is however unclear (Siegel, 2010, p. 125). What is obvious, is that salience in the field of language acquisition is just as ambiguous and debated as it is in other regards.

2.4 E

YE

-

TRACKING AND PUPILLARY RESPONSES

(14)

used as an online marker of cognitive effort and I will explain why I decided to use them in this experiment.

Smith (2012) provides us with a brief history of eye-tracking, stating that the use of eye-tracking techniques arose some 100 years ago (p.57). The interest in pupil dilation, however, goes back a lot further, as pointed out by Goldinger & Papesh (2012), who point to the works of Fontana (1765) who discussed pupillary responses to light, and Charles Darwin (1872) who already recognized changes in pupil size to reflect emotions such as fear in animals. Although Smiths history mostly discusses the study of fixations and saccades, with which researchers are able to trace exactly where and when a person is watching (Smith, 2012, p 57), one can find several examples of studies that use modern techniques to examine pupil dilation and cognition, such as Liao, Kidani, Yoneya, Kashino & Furukawa (2016), who have found a correlation between loudness and pupil dilation, Koelewijn, de Kluiver, Shinn-Cunningham, Zekveld & Kramer (2015), who have pointed out that a higher level of listening effort can be seen as an increase in pupil size, and Vogelzang, Hendriks & van Rijn (2016), who have found that in pronoun processing, pupil size reflects ambiguity resolution.

Previous studies have shown that pupillary responses are able to signal many different processes in the brain. Bradley, Miccoli, Escrig & Lang (2008) have for example used picture viewing tasks to show pupillary responses with emotional arousal. Goldinger & Papesh (2012), have shown that when retrieving words, pupil size was bigger for older words than for newly learned words, suggesting that pupil size reflects memory. Finally, Einhäuser, Koch & Carter (2010) found that pupil dilation signalled decision making. In their study, participants were asked to choose between sequentially presented numbers, only presenting their choice at the end. Using pupil dilation, Einhäuser et al. where able to predict the number chosen.

(15)

Normally, pupil size varies somewhere between 2 and 8 mm, and under normal circumstances the pupil is about 3 mm in diameter (Johansson & Balkenius, 2017, p. 1). In her study investigating pupillary responses to light, Ellis (2006) reported that the light reflex was found with a latency of around 220 ms. This latency is ascribed to the fact that the muscular responses that trigger these movements are rather slow ones (Johansson & Balkenius, 2017, p. 1). Interestingly, Ellis (2006) reports that, in the light reflex, the constriction of the pupil is about three times faster than the following dilation. No accounts of the latency of pupillary response to cognitive load were found, however, in the study by Liao et al. (2016) for example, reactions can be observed almost immediately after stimulus onset, lasting for the remaining 4 seconds that were measured. In this study, that can be seen as somewhat similar to that of Liao et al., in that participants were asked to sit in front of a screen and listen to the stimuli while their pupil size was measured, it is assumed that responses will be found in a similar pattern. Pupillary responses due to cognitive effort are usually recognized by changes in diameter of around 0.5 mm, whereas responses to light reflexes might result in contractions of 0.2-2.5 mm’s (Johansson & Balkenius, 2017, p.2). Because the process of pupil dilation is consensual in the eyes, we only have to focus on one of the eyes (Liao et al., 2016, p. 415)

In order to find out whether or not we can observe an increase in pupil diameter or not, we first need to know the initial size of the pupil. Therefore, pupil diameter is typically measured against a baseline, which is established for each participant individually during a timeframe in which no tasks are presented to the participant (Koelewijn et al., 2015; Vogelzang et al., 2016). This is repeated before each trial. Any possible reactions can then be observed as opposed to this baseline.

(16)

blinks from the data set (p. 879). There are however also examples where this is not taken into account, such as Liao et al. (2016).

Van Rijn et al. (2012) add, that movement can influence the process of pupil dilation (p. 2). It is therefore important that participants move their eyes as little as possible during the fixations.

2.5 R

ESEARCH QUESTIONS

After taking the above into consideration, the following research questions and hypotheses were formed.

1. Can salience be measured in pupil size?

Based on previous studies, such as for example the one by Liao et al. (2016), that proved that loudness-invoked salience can be measured in pupil size, I believe that salience should be measureable in pupil size as well. If this is the case, the eye-tracker should help us in providing proof for this hypothesis. We would then see an increase in pupil diameter when participants are presented with salient features as opposed to non-salient features. To find out whether salience can be measured in pupil size, we will test for six different categories of salience, which will be discussed in the next chapter.

2. Can salience be seen as extra cognitive load?

(17)

3. M

ATERIALS AND METHODS

3.1 P

ARTICIPANTS

In total, 41 participants (25 females) were tested. The mean age of the participants was 23.05 years old, ranging from 18 to 29 years old.

For the eye-tracking part of the experiment, the participants were randomly divided into two different groups who would either be exposed to list 1 or list 2 of the experiment. Eight males and thirteen females listened to list 1, their mean age was 23 years old. Eight males and twelve females listened to list 2, their mean age was 23.1 years old.

Six of the participants had an educational background at Dutch HBO level, the remaining 35 were all currently following or had completed education at university level. We did not differentiate between Bachelors and Masters.

Apart from the fact that all participants had Dutch as their L1, linguistic background was not controlled for. However, since regional linguistic variance might influence the extent to which certain traits might be perceived as being salient on an individual level, information about participants’ linguistic background was collected via the questionnaire.

3.2 M

ATERIALS

DETERMINING THE DIFFERENT CATEGORIES

Participants were presented with spoken samples, including six categories of salience based on the literature: Gender, acoustic prominence, unconscious sound change, conscious sound change, loudness and frequency. For each category, eight words were chosen, which were then fitted into different carrier sentences. In this section, we will discuss how the different categories were defined.

(18)

As a means to finding out the frequency of different words, several corpora were used. Instead of using only one, it was decided to use more. The reasoning behind this was that the use of multiple corpora was more reliable than only one. First of these was A frequency dictionary of

Dutch (Tiberius & Schoonheim, 2014). In their dictionary, four existing corpora of both written

and spoken Dutch have been combined in order to create a corpus of the 5000 most frequent words in the Dutch language (Ibid.). The words used in this experiment are chosen from the 100 most common nouns in the Dutch language according to the dictionary by Tiberius and Schoonheim. In order to calculate a word’s frequency, the sample texts in the four corpora are divided into sections of 2000 words, frequency is then established by the number of sections a word appears in and is given in percentages (Ibid.).

Secondly, The Corpus Gesproken Nederlands (Oostdijk & Goedertier, 2003) was used. It contains transcribed records of 900 hours of spoken Dutch. Since the presented stimuli are also spoken variables, it was deemed relevant to look at the frequency of spoken words as well. Frequency is defined as number of occurrences in the corpus.

Finally, Twitter Ngrams (Bouma, 2014) was used to find the frequency with which the words were to be found on Twitter. In this regard, Twitter was treated as a melting pot, where spoken and written speech come together. Apart from that, the language used on Twitter can also be seen as a more recent example of Dutch.

The mean frequency was calculated for each category and the words where selected in such a way that the mean frequencies did not differ significantly from one another.

(19)

For the same reason that it was decided that different stimuli should not be repeated, it was also decided that the different carrier sentences were not allowed to be repeated for one participant, in order to control for possible priming effects. As such, all carrier sentences differ from each other.

GENDER

The Dutch language has a covert gender system, meaning that a noun’s gender does not become clear from its shape or meaning, and has two possible genders: Common (receiving the definite article ‘de’) and neuter (receiving the definite article ‘het’) (Loerts, Wieling, & Schmid, 2013, p. 553). Around 75% of the nouns in Dutch have common gender (Ibid.). The gender category used in the current study presents the listeners with correct and incorrect use of grammatical gender, as incorrect gender use by a native speaker evokes online repair processes. It is because of these results, the incorrect form is deemed to be salient, and the correct form is deemed to be non-salient. Even though 75% of Dutch nouns have common gender, we decided not to distribute the words in this category based on this division. Instead, half of the words have common gender and the others have neuter gender. What is interesting, is that Loerts et al. found that common ‘de’ is used to predict the upcoming noun, while neuter ‘het’ is not. What is more, is that ‘hetNEU tuinCOM’ (correct form: ‘deCOM tuinCOM’, the garden) evokes a larger

response than ‘deCOM meisjeNEU’ (correct form: ‘hetNEU meisjeNEU’, the girl). Based on the

results found by Loerts et al., we might expect the responses for the common words to be stronger than the responses for the neuter words. The fact that neuter ‘het’ is not used as a predictor, is also an interesting finding to keep in mind when analysing the results for this category.

ACOUSTIC PROMINENCE

(20)

UNCONSCIOUS SOUND CHANGES

According to amongst others Auer et al. (1998), the criteria based on which a variable can become salient can be divided into objective and subjective factors. They state that objective factors can be described as things that linguists know, subjective ones as things the speaker knows. In this example, the sound changes involved are examples based on objective factors. The speakers within the speech community are (not yet) aware of the change.

Pinget (2015) states that during the last couple of decades, it has become more frequent in Dutch to pronounce originally voiced word-initial fricatives as voiceless word-initial fricatives (p. 28). Fricative devoicing has been shown to be an example of language change in progress across the Netherlands (Pinget, 2015, p. 28). As an example of this type of phonetic change, Pinget uses the pronunciation of /v/ as either [v] (=voiced) or [f] (=voiceless) (Pinget, 2015, p. 29). Although the change can be seen as advanced throughout the Netherlands, Pinget states that it is not completed at the moment (Pinget, 2015, p. 29). We have deemed this sound change an unconscious one, because there are no stereotypes that set a value to these different realisations of /v/. As such, speakers of Dutch are not aware of this change. Even though Pinget speaks of word initial fricatives only, we have also included examples where the fricative /v/ occurred somewhere in the middle of the word. During the analysis, this division will have to be looked at more thoroughly to see whether or not the results differ.

CONSCIOUS SOUND CHANGES

This type of stimuli also entails a language change. However, in contrast to the previous type of stimuli, speakers are aware of the language change in progress. Part of Auer et al.’s (1998) list of subjective criteria is ‘stereotyping’. If linguistic variants are highly stereotyped within the speech community, they can be deemed salient.

An example of a highly stereotyped variety in Dutch is what is known as the ‘Gooise r’, an approximant realisation of postvocalic /r/ (Bezooijen & van den Berg, 2004), and it is this variety we will use in this experiment. According to Sebregts (2015) the Gooise r has “relatively high sociolinguistic salience” in Dutch (p.66), and most speakers of Dutch will probably share the stereotyped image of the “hockey-meisje” (hockey girl) who uses a Gooise r. As such, it is a fine example of a stereotyped, and thus salient, linguistic variant. As its non-salient counterpart, we will use a rolling r, also known as an alveolar trill.

(21)

stimuli in this category are based on the article by Liao et al., and can serve as a control group, since it has already been shown that there is a pupillary response based on loudness. Even though Liao et al. did not use speech samples but other sounds such as laughter and beeping noises in their article, one might assume that if the set-up of the experiment is correct, we would find similar results. The carrier sentences will be of the same intensities for both the salient and non-salient versions, but the final word will be louder in the salient variant.

FREQUENCY

As discussed above, the frequency of words is important when selecting the right stimuli for this experiment. Since frequency might influence salience, this category of stimuli compares frequent words, which are supposed to be salient, and lesser frequent words, which are supposed to be non-salient (or less salient). As discussed above, I used a frequency dictionary of Dutch (Tiberius & Schoonheim, 2014) to determine the frequency of words in the Dutch language. It was however impossible to also select the non-frequent words from this book, since it is a dictionary containing the most frequent words. Because of that, I did not use the dictionary in the selection of the words in this category. The words in this category are based on Rommers’ (2007) list of high frequent and low frequent word pairs, that except for their frequency were almost identical. The words were selected in such a way, that the mean frequencies of the high frequency group and the low-frequency group were statistically different.

For an overview of the different categories, and their salient and non-salient forms, see Table 1.

(22)

TABLE 1: THE DIFFERENT CATEGORIES

SAMPLES

The samples were recorded in a recording studio using Adobe Audition CS6. A female, aged 26 who was a mother tongue speaker of Dutch helped with the recordings.

After recording all of the samples, they were analysed and edited using Praat (Boersma & Weenink, 2017). In order to make sure that a difference in the carrier sentences (other than the feature of interest) could not interfere with the outcome, the salient and non-salient variants were edited to fit within the same carrier sentence by splicing. As such, the only difference in the sentences presented to the participants was the target salient or non-salient variants. For the words in the unconscious sound change category, I used Praat (Boersma & Weenink, 2017) to check whether the difference between voiced and voiceless realisation of /v/ was actually present, an example of this can be found in Figure 1. For the loudness category the same program was used to alter the intensity of salient words. in such a way, that the only difference between the salient and non-salient variants of this category was the intensity of the target words. The words that had to be salient in this category were altered to have a mean intensity of 80 dB, whereas the rest of the stimuli all had a mean intensity of 60 dB.

In order to be able to make the comparison of the different samples easier later on, the total length, starting time of the soundwave, starting time of the target feature and end time for each sample were established. This information was added to E-Prime (Psychology Software Tools, Inc, 2012) as extra variables.

Category Salient Non-Salient

Gender Violation in grammatical

gender

Correct grammatical gender

Acoustic Prominence /t/ pronounced as [p] /t/ pronounced as [t]

Conscious Sound Change /r/ pronounced as [ɹ] /r/ pronounced as [r]

Unconscious Sound Change /v/ pronounced as [v] (=voiced) /v/ pronounced as [f] (=voiceless)

Loudness Word intensity of 80 dB Word intensity of 60 dB

(23)

QUESTIONNAIRE

In order to collect some basic information from the participants, a questionnaire was made and given to each of the participants before participating. The questionnaire is included (in Dutch) in the Dropbox file containing the appendices, and at the end of this thesis as Appendix 1. We asked people to specify their place of birth and places of residence during their life, in order to be able to generalize about their linguistic background. Although this information is not taken into account in the initial analysis, this might be done later on. An example of why this can be relevant, is the extent to which an approximant /r/ (gooise r) is used in different parts of the Netherlands. While this variant might be very common in the region known as the Randstad, this is less so in the province of Frylân.

3.3 T

HE DESIGN OF THE EXPERIMENT

The experiment comprised two separate parts, a quantitative and a qualitative one. In the quantitative part of the experiment we examined participant’s pupillary reactions to our stimuli. In the qualitative part, we conducted interviews to see to what extent our participants regarded certain features used in the experiment to be salient or not. Ideally, the order of these two different parts of the experiment would have to be altered for different participants to control potential memory effects. This was deemed impossible, however, as van Rijn et al. (2012) mention, memory is known to have an influence on pupil size. In order not to risk any interference on the pupil size that might occur because the participants recognized sentences from the interview, all participants were presented with the eye-tracking part first and the qualitative part second.

(24)

The eye-tracking experiment was created using E-prime version 2.0 and the E-Prime extensions for Tobii eye-tracker, TET (Psychology Software Tools, Inc, 2009). The experiment can be found in the Dropbox folder as Appendix 2. It was designed in such a way that there were two possible lists for the participants. The lists, and the different stimuli can be found in the Dropbox folder and attached to this thesis as Appendix 3. The sound files are also included in the folder. Each list contained 48 stimuli. If a stimulus was salient in list 1, it was non-salient in list 2 and vice versa.

In order to keep participants focussed during the experiment, 14 of the stimuli were followed by a question. Participants were asked if they had heard a word previously in the experiment and could answer with 1 for ‘yes’ and 2 for ‘no’. Question data for the first 12 participants could not be analysed due to some problems with the question slides, which was fixed as soon as it was discovered. These problems did however not influence the rest of the experiment.

The participants were asked to focus on a fixation cross in the middle of the screen. During the 1000 ms before the presentation of each stimulus, the baseline for their pupil size was measured. After a 1000ms fixation, they were presented with the stimuli in a randomized order, and the pupillary response was measured for the following 5000 ms. Finally, each stimulus was followed by a slide in which the participants saw “***” during which they were instructed to blink.

The eye-tracker was instructed to start recording before the baseline was measured and to stop recording when the response was measured for 5000 ms (see Figure 2 below for a flow chart overview, for a schematic overview of the procedure see Figure 3). The participants were asked to blink and move as little as possible during the experiment and preferably only when the Blink slide (“***”) was on screen.

In the qualitative part of the experiment, participants were presented with a salient – non-salient sample pair from each category to determine whether or not the example was in fact salient for the individual participants. These samples were presented to them in a PowerPoint presentation where each slide contained .wav-files for one pair, which can be found in the Dropbox folder

(25)

different samples. In order to control for fatigue during the interview, five different orders of presenting were created. All of these subgroups were presented with a different order of the items. This amount of groups was chosen randomly. The main goal with these subgroups was to make sure that the different categories that were tested in the experiment were discussed randomly. Since there was no way to do this automatically available, I did this manually by making the subgroups. The mean age for these groups were between 21.5 and 24.1 years old. Apart for group one, which contained 9 people, all the groups contained 8 people. It was tried to balance for males and females within these groups as well, but since there were more female participants overall, this was not possible in all groups.

For each category, the participants were then asked to elaborate on the sound files. A document containing the questions (in Dutch) for the interview are found as Appendix 4. Apart from the questions that can be found both in this thesis and online, the Dropbox folder also contains the recordings of the interviews, as well as transcriptions of a number of the interviews.

FIGURE 3: SCHEMATIC FLOW PER STIMULUS

3.4 P

ROCEDURE

(26)

all of them heard the sounds in the same way and that no sound from outside could interfere with the processing of the samples.

After listening to the samples, participants were interviewed about the different categories and recorded using an Olympus Digital Voice recorder WS-200S.

The qualitative part of this experiment used a PowerPoint presentation with six slides, each containing a pair of stimuli from one of the categories. This presentation was used to discuss the different types of variables with the participants. An example of the slides can be found in the Dropbox folder in Appendix 4.

Upon entering, participants were first asked to sit down behind the eye-tracker and to fill in the questionnaire. After that, they were given a short introduction. They were told that they were going to listen to some sounds first and that there would be an interview after that. Subsequently, they were told to put on the headphones and the eye-tracker’s calibration procedure was started. During this procedure, the participant had to follow a red dot on the screen. The calibration procedure is done to makes sure that the eye-movements is recorded properly and that the participants are looking where we want them to look and sitting at the right distance from the screen. Calibration was repeated for every participant. After calibration, the eye-tracking part of the experiment was started.

Upon completion of the eye-tracking part of the experiment, participants were asked to get seated behind a laptop with the PowerPoint presentation made for this part of the experiment. They were allowed to listen to the sounds a couple of times and were asked some questions about the fragments to determine whether or not they perceived one of the two as being salient and why.

A protocol with a step by step guide to the conducting of this experiment (in Dutch) is enclosed here and in the Dropbox folder as Appendix 5.

3.5 D

ATA PROCESSING

(27)

data was also included in this file. These where then uploaded separately into R (R Core Team, 2017) and combined into one data file. The original data set collected from E-Prime and the interviews contained over 1.6 million observations over around 70 variables.

The data file was then cleaned in R, which included steps such as extracting new variables (e.g. Time steps and time in ms.) and artefact removal (e.g. blinks). In line with other articles in the field, such as Vogelzang et al. (2016), trials that contained more than 25% blinks were removed from the data set. After that, the remaining blinks were removed from the data, including 8 time points before and after the blink, corresponding to roughly 65 ms before and after. This was done in order to account for eye movement that occurs around the blink, resulting in flawed data (J. van Rij, personal communication, may 2017). Finally, we checked whether the new data sets for each event was at least 75% of the size of the original dataset. If this was not the case, the event was also removed.

We recorded a baseline for 1000 ms, which is actually longer than needed (J. van Rij, personal communication, may 2017). Therefore, we decided to only keep the last 200 ms before each trial.

(28)

For all the different trials the mean pupil size was established, resulting in a final dataset of 1768 observations, one for each trial-subject combination that was left in the dataset after cleaning. Since one general model, including both the condition (saliency) and category as fixed effects would result in a rather complex model, containing several multiple interaction components, it was decided to create a separate model for each category. In our final data set, the mean pupil size per trial was our dependent variable. Condition (salient/non-salient), trial id (order of presentation), the word length and the total duration of the stimulus were treated as a fixed effect. Random-effects that were included in the models were subject and stimulus. Although we also looked at the effect of age and sex, these variables did not alter the results and were therefore left out of the models. Variables were added to the models step-wise and kept if significant. If not, they were left out of the model.

(29)

4. R

ESULTS

4.1 D

ILATION DATA

Of all 1886 events (41 subjects x 46 items), 118 contained over 25% blinks or were less than 75% of the original size after blink removal, and were as such removed from further analysis. 1768 events were kept for analysis. In this study, the lme4 package (Bates et al., 2015) and its function lmer was used in the R environment (R Core Team, 2017) to fit linear mixed-effect models to the data set.

When comparing the distribution of the pupil size for the (pre-determined, subsequently only referred to as ‘salient’) salient trials vs the (pre-determined) non-salient trials, seen in Figure 4, we can see that overall, the pupil sizes for the salient trials (in red) are larger than the sizes for the non-salient stimuli (in green).

FIGURE 4: PUPIL SIZE DISTRIBUTION SALIENT (RED) VS NON-SALIENT (GREEN)

In Table 1, a summary of the pupil size per category is given. These numbers have been calculated using the time period between 1000 ms to 5000 ms. 1000 ms is chosen because that is approximately the mean stimulus end and 5000 ms is the end of our measurement. The exclusion of the first 1200 ms (200 s baseline and 1000 ms during which the stimulus was heard), was done to make sure that only the reaction was looked at in the analysis and that the time where no reaction was found could not even out the results.

1.5 2.0 2.5 3.0 3.5 4.0 4.5

0.0

0.5

1.0

1.5

Pupil Size Distribution

Pupil size

(30)

TABLE 2: PUPIL SIZE (IN MILIMETERS) PER CATEGORY AND CONDITION

Category Salient Non-salient

mean median min max mean median min max

Gender 2.779 2.750 2.026 3.665 2.761 2.722 1.975 3.517 Acoustic Prominence 2.817 2.809 2.115 3.514 2.782 2.769 1.974 3.737 Unconscious sound change 2.782 2.779 2.003 3.522 2.760 2.705 2.059 3.789 Conscious sound change 2.765 2.735 1.938 3.602 2.742 2.707 1.950 3.635 Loudness 2.834 2.774 2.116 3.738 2.762 2.743 2.118 3.538 Frequency 2.771 2.745 2.088 3.445 2.801 2.771 2.000 3.634

When plotting the pupil size against the time for the different categories, we get the results shown in Figure 5. As can be seen in the plots, pupil size appears to be larger for the salient stimuli in the categories Loudness, Acoustic prominence and Gender. The same, although very slightly, seems to be the case for unconscious sound change and conscious sound change. For frequency, the pupil size for the salient category appears to be smaller than for the non-salient category. These results do however not yet take the answers form the interview into account. For each category, we made a separate mixed effects model. In all models, mean pupil diameter was the dependent variable. Variables were then added to the model step-wise and kept if they were significant. As a result, all models had condition (i.e. salient / non-salient) and trial ID as fixed effects. The model for Conscious Sound Change included the total duration of the sound (TotDuration) as a fixed effect and the model for Loudness included the frequency of the words according to the frequency dictionary of Dutch (Freq1). For all separate models, Stimuli and Subject were included as random effects.

In Tables 3 – 8, the outcomes of our models can be found. First of all, the coefficients and the corresponding t– and p–values of the fixed effects in the different models are listed. It shows that apart from Frequency, all categories revealed an increase in pupil diameter (with

Frequency showing a decrease). However, this difference was only significant for the

(31)

to. Apart from that, we find the standard deviations for the adjustments to the intercept of the random effects. Finally, the tables show the amount of variance that can be explained through our models. Using the MuMIn package (Barton, 2015) in R, we were able to calculate the R2 for our models. As can be seen, the variance explained by our models ranges between 80.6% and 87.3% of the variance in the data set.

(32)

TABLE 3: RANDOM AND FIXED EFFECTS GENDER

Fixed effects

Estimate Std. Error t-value p-value

(Intercept) 2.972 0.045 66.15 0.001

ConditionSalient 0.007 0.017 0.41 0.682

TrialId -0.009 0.001 -16.80 0.001

Correlation of fixed effects

(Intercept) ConditionSalient ConditionSalient -0.197

TrialId -0.281 0.045

Random effects

Groups Name Std. Dev

Subject (Intercept) 0.262

Stimuli (Intercept) 0.022

Residual 0.109

Amount of variance explained

R2 0.873 (87.3%)

TABLE 4: RANDOM AND FIXED EFFECTS ACOUSTIC PROMINENCE

Fixed effects

(Intercept) 2.974 0.043 68.85 0.001

TrialId -0.009 0.001 -15.50 0.001

TrialId -0.288 0.047

Random effects

Residual 0.116

(33)

TABLE 5: RANDOM AND FIXED EFFECTS UNCONSCIOUS SOUND CHANGE

Fixed effects

(Intercept) 2.973 0.043 69.35 0.001

TrialId -0.009 0.001 -15.64 0.001

TrialId -0.316 0.024

Random effects

Residual 0.129

R2 0.825 (82.5%)

TABLE 6: RANDOM AND FIXED EFFECTS CONSCIOUS SOUND CHANGE

Fixed effects

(Intercept) 3.070 0.068 44.87 0.001

TrialId -0.009 0.001 -16.24 0.001

TotDuration -0.124 0.054 -2.31 0.021

(Intercept) ConditionSalient TrialId ConditionSalient -0.023

TrialId -0.239 -0.003

TotDuration -0.761 -0.097 0.069

Random effects

Residual 0.115

(34)

TABLE 7: RANDOM AND FIXED EFFECTS LOUDNESS

Fixed effects

(Intercept) 3.014 0.053 56.91 0.001

TrialId -0.009 0.001 -11.80 0.001

Freq1 -0.002 0.001 -2.48 0.013

TrialId -0.346 -0.019

Freq1 -0.476 -0.036 0.065

Random effects

Subject (Intercept) 2.676e-01

Stimuli (Intercept) 2.494e-08

Residual 1.384e-01

R2 _{0.817 (81.7%)}

TABLE 8: RANDOM AND FIXED EFFECTS FREQUENCY

Fixed effects

(Intercept) 3.032 0.042 73.12 0.001

ConditionSalient -0.033 0.016 -2.10 0.036

TrialId -0.010 0.001 -16.90 0.001

TrialId -0.329 0.014

Random effects

Subject (Intercept) 2.402e-01

Stimuli (Intercept) 08.498e-10

Residual 1.362e-01

(35)

4.2 Q

UALITATIVE INTERVIEWS ABOUT SALIENCE

To understand whether the features that were pre-determined as salient were in fact “noticeable” to the listeners we conducted qualitative interviews. The answers to these give another, equally valuable, answer to whether linguistic features chosen are salient or not.

For the categories Gender, Acoustic prominence and Loudness, participants all answered in line with our expectations and definitions of salience. They all heard a difference, agreed on which variable was salient, and on why it was salient.

For the Frequency category, all participants did hear a difference, however not all of them perceived the salient stimulus as salient. 24 of the 41 participants deemed the salient stimulus to be salient. 21 of them did so for our reasons, the remaining three for other reasons. 2 participants found the non-salient stimulus to be salient. 15 participants did hear a difference between the words, but reported no difference in the level of salience between the two variants (i.e., one wasn’t more noticeable than the other).

For the Conscious sound change category, 7 out of 41 participants did not hear a difference. 5 of the remaining 34 participants deemed the salient stimulus salient, 4 did so for our reasons, 1 for a different reason, namely that ‘there was something odd’ about the pronunciation of sounds other than the target variable. 13 participants found the non-salient stimulus salient. 11 did so for our reasons, 2 for other reasons. The reason given was that they thought of themselves as users of the variant that we deemed salient. As such, what was deemed salient and what was deemed non-salient was the opposite from our expectations. 23 of the participants reported no difference in salience between the two variants.

(36)

4.3 T

AKING THE INTERVIEWS INTO ACCOUNT

The interviews conducted provide us with interesting information about patterns in the way people perceive salience. The models presented so far did however, not include these results. In this section we will include this data for the categories Unconscious Sound Change and

Conscious Sound Change, in order to see if we are able to find significant results for these

categories as well. Gender is not included here, since there was no variance in what the participants perceived as salient and non-salient. Although it would be interesting to include the interview results for the other categories as well, this was not done as of yet for the sake of time.

We followed the same procedure as we did earlier, although this time, we selected only those parts of the data sets that had the same view on salience of the variables. While interviewing the participants, we tried to find out what they perceived as salience. Based on their answers, participants could score a 1 (no difference in level of salience), 2 (our definition of salient was perceived as non-salient and vice versa), or 3 (corresponding to our definitions).

In the Unconscious Sound Change category, when only including the participants who perceived the variants that we hypothesized to be salient, as salient, we found the results presented in Table 9. Although the t- and p-value are somewhat higher than for the initial model, the increase in pupil diameter is still not significant. When only including the participants that did not perceive a difference in the level of salience between the two variants, we found the results in Table 10. Since only one participant scored a 2 in this category, this alternative was

(37)

TABLE 9: EFFECTS FOR PARTICIPANTS WHO CORRESPOND TO OUR DEFINITIONS IN THE UNCONSCIOUS SOUND CHANGE CATEGORY (SCORE 3)

Fixed effects

(Intercept) 2.930 0.081 36.19 0.001

TrialId -0.009 0.001 -7.78 0.001

TrialId -0.358 0.126

Random effects

Residual 0.130

R2 0.791 (79.1%)

TABLE 10: EFFECTS FOR PARTICIPANTS WHO DID NOT PERCEIVE DIFFERENT LEVELS OF SALIENCE IN THE UNCONSCIOUS SOUND CHANGE CATEGORY (SCORE 1)

Fixed effects

(Intercept) 2.972 0.051 58.179 0.001

TrialId -0.009 0.001 -12.851 0.001

TrialId -0.309 0.036

Random effects

Residual 0.130

(38)

The results in Table 11 are found when looking at the participants who in the Conscious Sound

Change category scored a 3. Although not significant, the pupil size for the salient variants is

in fact smaller than the pupil size for the non-salient variants. In Table 12, the results for the participants who found the hypothesised salient variable non-salient and vice versa is presented. As can be seen, the slight increase in pupil diameter is, again, not significant. Finally, in Table 13, the results for participants who did not report a different level of salience between the conditions can be found. As for the other groups, the increase in pupil diameter is not significant. Including the results from the interviews in our analysis, did not result in the finding of new significant results.

TABLE 11: EFFECTS FOR PARTICIPANTS WHO CORRESPONDED TO OUR DEFINITIONS IN THE CONSCIOUS SOUND CHANGE CATEGORY (SCORE3)

Fixed effects

(Intercept) 2.799 0.069 40.84 0.001

ConditionSalient -0.043 0.026 -1.63 0.103

TrialId -0.005 0.001 -4.32 0.001

TrialId -0.376 -0.162

Random effects

Residual 0.065

(39)

TABLE 12: EFFECTS FOR PARTICIPANTS WHO FOUND THE HYPOTHESISED SALIENT VARIANT NON-SALIENT AND VICE VERSA IN THE CONSCIOUS SOUND CHANGE CATEGORY (SCORE 2)

Fixed effects

(Intercept) 2.943 0.087 33.88 0.001

TrialId -0.010 0.001 -10.91 0.001

TrialId -0.259 -0.026

Random effects

Residual 0.107

R2 0.903 (90.3%)

TABLE 13: EFFECTS FOR PARTICIPANTS WHO DID NOT PERCEIVE A DIFFERENCE IN THE LEVEL OF SALIENCE IN THE CONSCIOUS SOUND CHANGE CATEGORY (SCORE 1)

Fixed effects

(Intercept) 3.134 0.097 32.42 0.001

TrialId -0.009 0.001 -11.41 0.001

TotDuration -0.157 0.078 -2.02 0.043

TrialId -0.251 -0.040

TotDuration -0.781 -0.099 0.088

Random effects

Residual 0.126

(40)

5. D

ISCUSSION

The purpose of this study has been to find out whether or not salience could be measured in pupil size. The reason to do so, was to gain a better understanding of the cognitive aspects of salience and to better understand the role it plays in language change. Pupil size was measured while participants listened to sentences, half of which contained a trait that was hypothesised to be salient. Afterwards, mean pupil size for the different categories and conditions was compared in a between-subjects design.

5.1 R

ESEARCH QUESTION

1:

C

AN SALIENCE BE MEASURED IN PUPIL SIZE

?

Three out of our six categories showed a significant change in pupil diameter: Acoustic

prominence, Loudness and Frequency. A link between loudness and auditory salience has

already been pointed towards by Liao et al. (2016), and the fact that we were able to find similar results in our study shows that the design of the experiment was, in fact, working. It is therefore stimulating to see that the categories we called Acoustic prominence and Frequency, also showed a significant result.

What is interesting, is the fact that Drager & Kirtley (2016) pointed out that whenever something is perceived as a mistake, it would not be salient to the listener. The fact that all participants were directly able to pick out the gender violations, but did not show a significant response in pupil dilation for this category, might be a good example of this statement. Many of the participants mentioned that they perceived the gender violations as a mistake that was frequently called ‘annoying’ or something similar. When compared with for example Acoustic

prominence, a category that did show a significant increase in pupil diameter, the point by

(41)

Remember Hickey, who wrote that if people state that they are not aware of a trait’s salience, it does not necessarily mean that there is no subconscious level of noticing (Hickey, 2000, p. 58). When looking at the results for the category that we called Unconscious Sound Change, this is an interesting statement. The interviews proved that over half of the participants were unaware of the difference between the two variants. Interestingly enough, there was a trend that showed an increase in pupil diameter for the salient forms, suggesting that it is indeed possible to pick up on a feature’s level of salience without consciously noticing that there is a difference at all. When adding the interview data, however, we were able to show that for the participants who did not perceive a difference in the level of salience, there was no difference in pupil diameter between our hypothesised salient and non-salient variants. As such, Hickey’s statement that there might be a subconscious level of noticing that a variable is salient does not seem to hold. Further testing is however need in order for our results to be conclusive.

The reaction found in the Loudness category, was by far the largest. Initially, one might think that it is the intensity in itself that triggers this response, however when keeping the interviews in mind, an important aspect of this reaction might be the fact that participants were startled (or surprised) by the intensity of the sound. The difference between the salient and non-salient variant was no less than 20 dB’s. In the future, we might decide that a difference of for example 10dB would suffice. Since we used the mean intensity of the sentences however, I wanted to make sure that there really was a perceivable difference and for that reason I chose to use an intensity of 60 dB and 80 dB respectively. This is, however, something to keep in mind for the future. Nevertheless, the category did show a significant result, suggesting that a sound change that brings a high rate of surprisal is perceived as salient.

5.2 R

ESEARCH QUESTION

2:

C

AN SALIENCE BE SEEN AS EXTRA COGNITIVE LOAD

?

(42)

the brain with information as well. The notion that we had to find a 0.5mms increase in order to find that salience means extra cognitive load does therefore not hold.

However, even when keeping this in mind, the only category where pupil size varies 0.5 mms as opposed to the baseline, is the Loudness category. This can mean two things. Either, the response triggered by salience is not (extra) cognitive load, but a reaction of its own, or the amount of tested subjects and trials was not big enough. Further research on this topic is needed in order to rule out one or the other.

5.3 R

EVIEWING THE DEFINITION OF SALIENCE