Abstract
Reading requires the assembly of cognitive processes across a wide spectrum from low-level visual perception to high-level discourse comprehension. One approach of unravelling the dynamics associated with these processes is to determine how eye movements are influenced by the characteristics of the text, in particular which features of the words within the perceptual span maximise the information intake due to foveal, spillover, parafoveal, and predictive processing. One way to test the generalisability of current proposals of such distributed processing is to examine them across different languages. For Turkish, an agglutinative language with a shallow orthography–phonology mapping, we replicate the well-known canonical main effects of frequency and predictability of the fixated word as well as effects of incoming saccade amplitude and fixation location within the word on single-fixation durations with data from 35 adults reading 120 nine-word sentences. Evidence for previously reported effects of the characteristics of neighbouring words and interactions was mixed. There was no evidence for the expected Turkish-specific morphological effect of the number of inflectional suffixes on single-fixation durations. To control for word-selection bias associated with single-fixation durations, we also tested effects on word skipping, single-fixation, and multiple-fixation cases with a base-line category logit model, assuming an increase of difficulty for an increase in the number of fixations. With this model, significant effects of word characteristics and number of inflectional suffixes of foveal word on probabilities of the number of fixations were observed, while the effects of the characteristics of neighbouring words and interactions were mixed.
Keywords
The extraction of information from written texts has been the subject of much research over the past few decades. Research on reading addresses various aspects involved in the complex cognitive process, ranging from oculomotor control, comprehension, and understanding to developmental perspectives of reading and ageing (e.g., Kliegl et al., 2006; Liversedge et al., 1998; Liversedge & Findlay, 2000; Paterson et al., 2013; Rayner, 1998, 2009; Rayner et al., 2007, 2012). As an indicator of online cognitive processing, reading consists of processes at multiple layers, from letter perception to word recognition, syntactic parsing, sentence comprehension, and discourse comprehension. Usual reading is a visual process; the investigation of eye movement control in reading has therefore been a backbone of reading research. Eye movement characteristics, such as fixation duration and location as well as the probabilities of the number of fixations on words, have been frequently investigated as measures of eye movement control in reading. Effects of cognitive, perceptual, and oculomotor constraints as well as of sentence and word characteristics are related to the integration of information across fixations in reading. Many effects, especially effects of features of the fixated word n (such as its length, frequency, predictability) and spillover frequency effects of the left word n − 1, are well documented, but other effects, in particular effects related to the preview of the word n + 1, (i.e., the word to the right of the currently fixated word during first pass reading) are still controversial (e.g., Angele et al., 2015; Brothers et al., 2017; Kliegl, 2007; Kliegl et al., 2006; Rayner et al., 2007). They are in the focus of research because evidence for or against these effects are usually taken as support for or against distributed (parallel) versus serial processing which is one of the most controversial theoretical issues in eye-movement control in reading and their implementation in computational models (e.g., for an overview, see Engbert et al., 2002; Radach et al., 2007; Reichle et al., 2003).
Some of the differences between studies may be due to differences between languages. Consequently, perceptual parafoveal preview or cognitive predictability effects have been investigated for several languages, including English, German, Chinese, and Arabic (e.g., the parafoveal preview benefit, Cui et al., 2014; Cutter et al., 2015; Jordan et al., 2014; Li et al., 2009, 2014; Rayner, 1975; Lou et al., 2019; Yan et al., 2012). Using Turkish sentences, the present study investigated effects of features of a fixated word as well as the effects of the features to the left and the right of the currently fixated word on single-fixation durations (SFD) and the number of fixations (skipping, single fixation, or multiple fixations).
The study is set up as a conceptual replication of the statistical model reported in Kliegl et al. (2006; Kliegl, 2007). We included the same covariates and interactions. Kliegl et al.’s statistical model represents a distributed processing account based on two main theoretical assumptions: (1) Words in the perceptual span are processed in a gradient fashion, but parallel up to the meaning of the word. (2) The perceptual span is dynamically modulated by foveal processing difficulty (Henderson & Ferreira, 1990; Inhoff et al., 2000; Kennedy & Pynte, 2005).
Controversial are mainly a negative frequency effect of word n + 1, a positive predictability effect of word n + 1, the interaction between frequency effects of word n − 1 and n, and the interaction between frequency effects of word n and word n + 1 on fixation durations. The lexical characteristics of word n + 1 (i.e., successor effects) could influence fixation durations due to (a) parafoveal lexical preprocessing of the upcoming word, (b) dynamical modulation of the perceptual span, and (c) memory retrieval of word n + 1. Therefore, according to proponents of distributed processing of words within the perceptual span, it was expected that the lower the frequency of word n + 1 the longer the fixation duration on word n. However, the effects should be less pronounced for an infrequent word n due to dynamical modulation of perceptual span (i.e., the shrinkage of perceptual span because of a difficult word n). Another aspect related to dynamical modulation of perceptual span is the smaller parafoveal benefit obtained from word n because of a difficult word n − 1 which would be reflected in a stronger frequency effect of word n if word n − 1 was a difficult (i.e., less frequent) word. Finally, the effect of the predictability of word n + 1 was expected to influence fixation durations positively due to retrieval of a predictable word n + 1 from the memory while fixating word n. In other words, the retrieval process of highly predictable words would increase duration of the fixations on word n (Kliegl, 2007; Kliegl et al., 2006). Recently, current computational models cannot be directly falsified based on the existing, controversial findings for weak effects. In the present study, the theoretical background is the assumption that dynamical modulation of perceptual span by foveal processing difficulty should influence the effects of the characteristics of neighbouring words on eye movement measures. Although the findings of the current study may have implications for design and development of computational models at a general level, we conceive the debate about existing computational models beyond the scope of the study.
As far as we know, this is the first study of this kind with Turkish sentences and also the first study that uses number of fixations (0, 1, and 2+) as a dependent variable and estimates effects with a base-line category logit model.
Turkish is a Ural-Altaic language with an agglutinative morphology and a relatively shallow orthography–phonology mapping. Until 1928, Turkish orthography was based on Perso-Arabic characters and the society had a low overall rate of literacy less than 10% (recently, over 95%). The orthography of modern Turkish is composed of 8 vowels and 21 consonants derived from the Latin Alphabet. Turkish is read from left to right. Word forms in Turkish are formed through the concatenation of derivational and inflectional affixes (predominantly suffixes) to a root morpheme or to an already complex word form (Göksel & Kerslake, 2005; Kornfilt, 2013; Lewis, 2000). Stem formation by affixation to already derived stems is extremely productive; an already highly complex stem can serve as the basis for even more complex word forms. This process often results in relatively long words equivalent to a whole sentence in English. It has been estimated that a comprehensive description of the allowable combinations and morphotactics for Turkish may require the definition of over 40 stem and more than 150 suffix categories (Hankamer, 1989). Table 1 displays a suffix derivation process based on the root noun renk (English: colour).
Sample Turkish words and their English translations.
The relationship between morphemes and their functions is usually of a one-to-one nature, with only very few exceptions. The surface realisations of morphemes are constrained and modified by morphophonemic rules. For example, vowels in affixed morphemes need to agree with the preceding vowel in certain aspects to achieve vowel harmony and there are instances in which vowels within roots and morphemes are deleted. In the same vein, consonants in roots or affixed word forms also sometimes undergo modifications or are deleted (Oflazer et al., 1994). While the majority of Turkish derivational suffixes are unproductive in the sense that they are not perceived by native speakers as available for the production of new lexical entries, a few derivational suffixes are still very productive as they bear a regular meaning and can be used with any stems meeting the necessary criteria. For example, the derivational suffix—CI that attaches to nouns to produce nouns such as lokanta “restaurant”—lokantacı “restaurant owner” is a productive suffix that is used to produce new words, while the derivational suffix—(A)mAK that attaches to verbs to produce nouns such as bas “to step”—basamak “step” is not used to produce new words (Göksel & Kerslake, 2005, pp. 52–53, 58). Inflectional suffixes, in contrast, are highly productive (Durrant, 2013; Göksel & Kerslake, 2005).
Previous research on the role of morphological complexity in reading has predominantly addressed compound words for a set of frequently studied languages, such as English, German, and Dutch. Studies investigating polymorphemic words have been scarce in comparison to research on compound words (Bertram & Hyönä, 2013; Falkauskas & Kuperman, 2015; Juhasz, 2007 among many others). To the best of our knowledge, no studies have so far been conducted on eye movement control in Turkish reading, whereas quite a few studies investigating sentential-level syntactic constructions and word recognition through lexical decision and naming tasks have been reported.
The morphological complexity of Turkish allows the construction of long word forms (see Table 1), which would be unusual for non-agglutinating languages. The study of eye movement patterns in Turkish reading is a methodologically appropriate choice as standard behavioural measures, such as lexical decision tasks and naming tasks may not be appropriate for studying long words. In the present study, we employ the corpus-analytical approach; we presented sentences (excerpted from original resources) to participants and recorded their eye movements, as reported in the following sections. The following section reports an experimental study that aimed at identifying oculomotor patterns in Turkish reading. 1
Method
Participants
Thirty-seven students from Middle East Technical University in Ankara, Turkey (mean age = 21.60 years, SD = 2.36 years, 17 females) participated in the present study for a monetary compensation of approximately €7. All participants were native speakers of Turkish. None of the participants used contact lenses; one participant used glasses during the experiment. The data from two participants were eliminated due to calibration problems. The whole experiment lasted about 30 min. Before the experimental procedure, each participant signed an informed consent form. In addition to the main sample, an independent group of 80 students (mean age = 24.20 years, SD = 3.71 years, 40 females), who were also native speakers of Turkish, participated in a sentential predictability test in return for the same amount of monetary compensation.
Procedure
Participants read 120 nine-word sentences that were presented in 18 pt. Courier New font on a 17-inch cathode-ray tube (CRT) monitor (1,024 × 768 resolution, Video Graphics Array [VGA] connection) controlled by a 3.0 GHz computer. Participants were seated approximately 60 cm from the screen, each letter corresponding to approximately 0.46° of visual angle and approximately 14 pixels. The heads of the participants were positioned on a forehead rest and chin rest to avoid head movements and, hence, calibration problems. Eye movements of the right eye of the participants were recorded with an SR Research EyeLink 1000, tower mount system with a sampling rate of 1,000 Hz.
A standard nine-point grid calibration was used in the eye tracking sessions and the calibration was validated for each participant. Before the experimental session, practice trials were presented, which consisted of sample stimuli. Before the start of each trial, a fixation mark appeared on a blank screen, with the same coordinates as those of the first letter of the sentence. Participants were informed that after a 500 ms fixation on this fixation mark, a sentence would be displayed on the screen. They were further instructed to read the sentences silently for comprehension at normal reading speed. After reading each sentence, the participants fixated on another fixation mark for 500 ms which was displayed at the onset of the sentence display in the right corner. The experiment continued with (a) a yes-or-no question related to the displayed sentence following 15% of the trials, to keep the attention of the participants alive during the session; (b) a fixation marker on the left side of a blank screen indicating the beginning of the next trial; or (c) one of two breaks, after each 46 trials. Three paragraphs were shown in the beginning, the middle, and the end of the experiment as filler stimuli material. The paragraphs were chosen from Bıçakçı (2012), which is a collection of Turkish short stories. The presentation order of the sentences and the paragraphs was randomised.
Materials
Experimental stimuli were 120 sentences chosen randomly from the representatively assembled METU Turkish Corpus, which is a collection of 2 million words of post-1990 written Turkish samples (Say et al., 2004), respecting the following constraints: (1) Each sentence consisted of nine words and 73 characters at maximum to fit onto one line on the screen. (2) No hyphenated words, numbers, or abbreviations were used in the sentences. (3) No punctuation marks were used in the sentences, with the exception of full stops at the end of each sentence. (4) No questions or interjective sentences were selected.
A total of 1,080 tokens of 777 words (types) were used in the 120 experimental sentences. Excluding the first and the last words of each sentence from the analyses, eye movement data on a total of 840 tokens of 589 words were analysed. The suffixation rules set out by Göksel and Kerslake (2005) were used for the identification of suffixes. The derivational suffix count varied from 0 to 2 (M = 0.11, SD = 0.35), the inflectional suffix count varied from 0 to 4 (M = 0.82, SD = 0.92), and the total suffix count varied from 0 to 5 (M = 0.93, SD = 0.96) suffixes. Word, stem, and root length ranges were 1 to 16 letters (M = 5.92, SD = 2.35), 1 to 10 letters (M = 4.20, SD = 1.61), and 1 to 9 (M = 3.94, SD = 1.39) letters, respectively. The frequency values of words, roots, and stems were taken from the BOUN Web Corpus (Bilgin, 2016; Sak et al., 2008), which includes 383,224,629 tokens of 1,337,898 Turkish words (types). The mean word frequency was 1663.66 per million (SD = 4734.48), the mean stem frequency was 1736.29 per million (SD = 4780.95), and the mean root frequency was 1739.16 per million (SD = 4780.36) with a range of 0 to 23250.09 per million. Laplace smoothing (Brysbaert & Diependaele, 2013) and log transformation were performed on the raw frequency values of words for the analyses.
The sentential predictability of stimuli words was measured through a cloze test. The first words of the sentences were not included in the analyses. Each of the eight words of each sentence was predicted by a different participant. Therefore, there were eight conditions consisting of 120 trials in each for different participants. The participants were instructed to predict the upcoming word according to the sentential context. Consequently, each word was predicted by 10 participants. By comparing the whole-word predictions with the actual words, a correctly predicted word received a score of 1, while other words were scored as 0. The probability of predicting a word was calculated over 10 participants, which is the number of participants who predicted that word. Predictability probabilities were then transformed into logit-based predictability scores for the analyses. In what follows, we describe data selection for SFD analyses and the process of building the linear mixed model (LMM) for SFD as the dependent variable.
Data selection
The eye movement data from 83 sentences read by one participant (female, age = 21 years) were not recorded because of technical problems during the experiment, and the data were not included in the eventual analyses. The eye-movement data from the remaining 36 participants were screened for drifts and loss of measurement. Where applicable, drifts were corrected manually by moving fixations up or down or using the drift-correct function of Data Viewer (version 2.6). Otherwise, the trial was excluded from the analysis. During the drift correction process, only vertical corrections were applied to fixations. After the screening and drift correction processes, the data obtained from the reading of 88 sentences were marked as low-quality and were excluded from the analyses. The data from one further participant (male, age = 27) were excluded from the analyses as more than 30% of data from this participant was marked as low quality. The data loss due to low-quality data for the remaining 35 participants was 0.8% (a total of 34 sentences). The participants answered 96% (SD = 5%) of the comprehension questions correctly.
Following the practices reported in the literature (e.g., Rayner, 1998, 2009), first-pass eye movement measures were used in the analyses. Fixations during the first encounter of a word until moving to another word were accepted as first-pass fixations. Fixations after the first fixation on the last word (i.e., re-readings) and words first fixated after fixating any word on the right of it (i.e., skipped words in the first-pass) were not included in the analyses. There were 29,704 words that received at least one fixation in the first pass. Out of 29,704 words, 16,598 of them received single fixations, which constitute 56% of the fixated words in the first pass. After the elimination process, 14,721 words that received single fixations were left, which makes up 50% of all fixated words. Upon inclusion of multiple fixation and skipped cases, there were 26,060 words (69%) that were included in fixation count analysis. The elimination criteria and percent of the loss are summarised in Table 2. Throughout the document, fixation (n) indicates the current fixation, fixation (n − 1) indicates the previous fixation, and fixation (n + 1) indicates the next fixation. Similarly, word (n) indicates the word on which the fixation (n) falls, word (n − 1) indicates the word on the left of the word (n), and word (n + 1) indicates the word on the right of the word (n).
Elimination criteria and percent of fixations.
SFD: single-fixation duration; FC: fixation count.
Some words were eliminated according to more than one criterion (e.g., most of the last words of sentences have a next fixation that was not on a word, but on fixation marker and hence, they were eliminated according to both (4) and (5)). Therefore, valid words included in the analyses are not simply counts subtracted from 37,800 words read by 35 participants.
Gaze duration (GD) elimination criterion (11) was used among fixation count (FC) data for comparable results between FC analysis and GD analysis provided in the supplement.
Word characteristics and eye movement measures were transformed before the statistical analyses to remove skewness in the variables and to reduce model complexity (Hohenstein et al., 2017). All the variables were centred on their means when they were covariates in the models. The characteristics of words (n), words (n − 1), and words (n + 1) were included and tested in the models as covariates. In the following, we present the word characteristics and transformations employed in the models.
Word characteristics
Word length (WL)
Reciprocal values of absolute word length values (i.e., 1/WL) were used in the analyses as covariates.
Word frequency (WF)
Log-transformed values of frequencies per million, estimated according to the Laplace smoothing method, were used in the analyses. Zero frequencies were handled by the Laplace smoothing method, rather than adding one to the frequency per million, as adding one to the frequency per million inflates the estimation of a missing observation in the corpus. However, adding one to frequencies of words increases the theoretical corpus size from the number of word tokens to number of word tokens plus number of word types (Brysbaert & Diependaele, 2013). In the Laplace smoothing method, after adding 1 to each word count, theoretical corpus size was estimated as the number of tokens plus the number of types of words in the corpus, and frequency per million was calculated accordingly. Thus, frequency per million values for words in this study was calculated by following formula (1).
Sentential predictability (WP)
Following the practices in the literature (e.g., Hohenstein et al., 2017; Kliegl et al., 2004), before inclusion in statistical analyses, predictability probabilities (p) were logit transformed and rescaled with 0.5, as shown in equation (2).
As 10 participants predicted each word, p values of 0 were recalculated as 1/(2*10) and those of 1 were recalculated as (2*10-1)/(2*10) (cf. Kliegl et al., 2004).
Inflectional suffix count (IS) of words
Suffixes were identified following Göksel and Kerslake (2005). Suffix counts were included in the analyses as continuous variables, and they were not transformed, except that they were centred on their mean as all covariates.
Eye movement measures
Incoming and outgoing saccade amplitude
Incoming saccade amplitude (ISA) is the distance between fixation (n − 1) and the first fixation on the word (n). Outgoing saccade amplitude (OSA) is the distance between the last fixation on the word (n) in the first pass and the first fixation on the word fixated immediately after leaving word (n). Both incoming and OSAs had to be saccades in the default reading direction (i.e., rightward oriented). Amplitudes were log-transformed (Base 2) before inclusion in the LMM.
Relative fixation location
Relative fixation location (RFL) is the distance between the centre of the word (n) and the letter of the first fixation on that word divided by word length.
Among the transformed values, negative values indicate the first half of the word, positive values indicate the second half of the word, 0 indicates the middle of the word, and −0.5 indicates the blank character space before the word (Hohenstein et al., 2017).
SFD
Natural logarithmic transformation was applied to SFD values to yield a close to normal distribution (Hohenstein et al., 2017).
Fixation count (FC)
The dependent categorical variable of base-line category logit model. Among its categories, zero indicates skipping cases, one indicates single fixation cases, and two indicates multiple fixation cases. In the model, skipping cases were selected as base category.
LMM and baseline-category logit model
To reveal the eye movement characteristics of Turkish reading, we modelled SFD using an LMM 2 with the lme4 package (version 1.1-15; Bates et al., 2015) in the R environment (version 3.4.2, 64-bit build; R Core Team, 2017). We started with the base model presented in Kliegl (2007, p. 532, Table 1). In the model, covariates were Word Frequency (WF); Predictability (WP); Word Length (WL) for words (n), (n − 1), and (n + 1); ISA; OSA; and Relative Fixation Location (RFL). There were five interaction components such that the interactions between WF and WL, WF and WF of word (n − 1) (WF1, hereafter), WF and WF of word (n + 1) (WF2, hereafter), WL and WF2, and WL and p of word (n + 1) (WP2, hereafter). Random factors of the model were participant and sentence. As the inspection of raw data with the lattice package indicated accordingly, word as a random factor was added to the model (version 0.20-38; Sarkar, 2008). The model was tested against the base model using the anova() function with a model selection α-level of likelihood ratio test of 0.2 (i.e., the p level for model comparison was set to .2) to balance Type I error and model power (Matuschek et al., 2017). The resulting model was better significantly (χ2(1) = 301.6, p < 2.2e−16). Random slopes for the participants in the base model were WF, RFL, and p. Raw data were inspected further by visualising the covariates for each participant to add random slopes when graphical inspections indicated different effects for each participant. At each step, only if the model converged and a random slope contributed to the model significantly according to model selection α-level, the slope remained in the model.
To take nonlinear relationships into account, the relationship between covariates and SFD was inspected by visualising it with the languageR package (version 1.5.0; R. H. Baayen, 2013). Each covariate was added to the model as a nonlinear component whenever graphical inspections on raw data indicated a nonlinear relationship. If a nonlinear component contributed to the model significantly according to model selection α-level and the model converged, the component remained in the model; otherwise it was removed from the model.
Inflectional suffix counts (IS) of word (n), word (n − 1) (IS1), and word (n + 1) (IS2) were added to the model to understand the relationship between SFD and morphological complexity of words. As none of the ISs contributed to the model, they were not added to the LMM of SFD. After the construction of the model, random structure was re-evaluated with Principle Component Analysis using the rePCA() function of the RePsychLing package (H. Baayen et al., 2015), following Bates et al. (2015). The model complexity related to the random structure did not exceed the principle component count that cumulatively accounts for 100% of the variance.
The variance inflation factors (VIFs) were calculated to test for multicollinearity (Lefcheck, 2012). There were no fixed factors that had a VIF value greater than 4. Moreover, the average tolerance value of fixed factors was 0.74 (greater than 0.2). Accordingly, no more simplifications were made to this model, after this stage (Field, 2009). Below, the final model was reported with the p values obtained using the lmerTest package (version 3.1.-0; Kuznetsova et al., 2017). Partial effects for graphs were obtained using the remef package (Hohenstein & Kliegl, 2015) and graphs were constructed using the ggplot2 package (Wickham, 2009).
The probabilities of words being skipped, receiving single fixation, and receiving multiple fixations were modelled by a baseline-category logit model (BCLM) using vglm() function of VGAM package (version 1.1-1; Yee, 2015; Yee & Wild, 1996). Base category for the model was skipping instances. In other words, coefficients indicated log-odds of single-fixation cases and multiple fixation cases relative to skipping cases. In model construction process of BCLM, an approach similar to one for LMMs was used. However, in BCLM, skipped instances were included for which oculomotor covariates were not available. In addition, vglm() function of VGAM package does not support a mixed model design. Therefore, in the base BCLM, the oculomotor covariates and random structure of LMMs were not included. The other covariates in the base BCLM were the same as that of LMM. As graphical inspection of raw data did not indicate any nonlinear relationship between fixation counts and covariates, only inflectional suffix counts (ISs) were added to the model. The base model and the model that includes ISs were compared by a likelihood ratio test, using lrtest_vglm() function. The result showed that ISs contributed to the model (χ2(6) = 37.43, p < .001).
Results
Descriptive statistics for fixation patterns in the first-pass reading
Among the valid number of fixations in the first-pass reading without the elimination of fixations due to their durations (i.e., elimination according to criteria (8) and (11) in Table 2), single-fixation cases comprised 65%; right-direction refixation cases comprised 6%, left-direction refixation cases comprised 23% (two-fixation cases comprised 29%); and multiple fixation cases with higher than two fixations comprised 6% of valid fixations. Table 3 lists means and standard deviations of oculomotor measures and word characteristics included in the current study.
Means and standard deviations for oculomotor measures and word characteristics.
SFD model
LMM results with SFD as the dependent variable are presented in Table 4.
LMM parameters estimates for single-fixation durations.
LMM: linear mixed model; WP: predictability; WL: word length; RFL: relative fixation location; OSA: outgoing saccade amplitude; WF: word frequency.
Number of obs: 14,721, groups: Word, 589; Sentence, 120; Participant, 35.
Annotation: * indicates that the effect of the variable is significant; a colon between variables indicates interaction.
Oculomotor measures
Our findings are compatible with a frequently reported effect of fixation location on SFD in the previous literature (i.e., inverted optimal viewing position, the IOVP effect; Nuthmann et al., 2005; Vitu et al., 2001; Yan et al., 2014 for Uighur; and Hyönä et al., 2018 for Finnish). In other words, there was a significant effect of quadratic RFL on SFD; however, the linear component of RFL was not significant. Accordingly, fixation durations were longer when the fixation on the word was closer to the middle of the word compared to the fixation durations when RFL was closer to the edges of the word (linear: b = −0.02, SE = 0.02, t = −0.94, p = .35; quadratic: b = −0.74, SE = 0.05, t = −14.55, p < .001). The effects of ISA (b = 0.15, SE = 0.01, t = 27.68, p < .001) and quadratic OSA (b = 0.03, SE = 0.01, t = 5.24, p < .001) on SFD were significant. SFD increased together with an increase in ISA. Slightly decreasing effect of OSA became an increasing effect on SFD when OSA was higher than approximately six characters. Figure 1 illustrates the effects of fixation locations on SFD.
Oculomotor measures and SFD. Zero value of RFL indicates the centre of the word (the grey horizontal line), negative values indicate the left of the word centre, and the positive values indicate the right of the word centre. Line and 95% confidence band are partial effects, retrieved from LMM estimates by using the remef package (Hohenstein & Kliegl, 2015).
Word characteristics
Word length of word (n − 1) (WL1) (b = 0.18, SE = 0.04, t = 4.57, p < .001) and Word length of word (n + 1) (WL2) (b = 011, SE = 0.04, t = 2.82, p < .001) significantly influenced SFD. WL1 and WL2 had small but significant decreasing effects on SFD. 3 Linear and quadratic components of word length of word (n) did not have a significant effect on SFD (linear: b = −0.001, SE = 0.09, t = −0.01, p = .99; quadratic: b = −0.43, SE = 0.28, t = −1.52, p = .13).
The decreasing effects of WF (b = −0.04, SE = 0.01, t = −5.6, p < .001), WF1 (b = −0.01, SE = 0.004, t = −3.62, p < .001), and WF2 (b = −0.01, SE = 0.004, t = −2.67, p < .05) on SFD were significant.
The effect of predictability of word (n) (WP) (b = −0.02, SE = 0.01, t = −3.37, p < .01) on SFD was significant. SFD increased with a decrease in WP. The positive effects of the predictability of word (n − 1) (WP1) (b = 0.01, SE = 0.01, t = 1.42, p = .16) and the predictability of word (n + 1) (WP2) (b = 0.01, SE = 0.01, t = 1.6, p = .11) on SFD was not significant (see Figure 2 for the effects of word characteristics.).
Effects of word characteristics on SFD. Units of word frequency on x-axis are log-10 power of words per million. Line and 95% confidence band are partial effects, retrieved from LMM estimates by using the remef package (Hohenstein & Kliegl, 2015).
Interaction of frequencies of word n and word n − 1
The interaction of WF and WF1 (b = 0.01, SE = 0.002, t = 1.97, p < .05) significantly influenced SFD. The negative effect of WF was the strongest if the previous word was a low-frequency word (see Figure 3).
Interaction effect of WF and WF1 on SFD. Units of word frequency on x-axis are log-10 power of words per million. WF1 values are grouped as such: high WF1 values are greater than three on logarithmic scale (>1,049 pm, 4,241 obs.), low WF1 values are less than two on logarithmic scale (<99 pm, 5,374 obs.), and medium WF1 values are between two and three on logarithmic scale (between 99 pm and 1,049 pm, 3,391 obs.). Line and 95% confidence band are partial effects, retrieved from LMM estimates by using the remef package (Hohenstein & Kliegl, 2015).
The other interaction terms were not significant.
Fixation count probability model
We modelled the probability of fixation counts that a word could receive using a BCLM with skipping instances as the base category. 4 The model predicted equation (4) log-odds of fixating once on a word vs. skipping it (i.e., single vs. skip) and equation (5) log-odds of fixating more than once on a word vs. skipping it (i.e., multiple vs. skip). Therefore, the coefficients of the model indicate log-odds as follows
where
BCLM parameter estimates for fixation count probabilities.
BCLM: baseline-category logit model; WP: predictability; WL: word length; WF: word frequency.
Residual deviance: 35,941.26 on 52,084 degrees of freedom.
Log-likelihood: −17,970.63 on 52,084 degrees of freedom.
Number of Fisher scoring iterations: 8.
Annotation: * indicates that the coefficient is significant; a colon between variables indicates interaction.
Word lengths
The estimated coefficients 7 for WL single fixation versus skipping were −11.23 (SE = 0.31, χ2(1) = 1280.21, p < .001) and multiple fixations versus skipping was −26.64 (SE = 0.69, χ2(1) = 1494.6, p < .001). The odds ratio of the probability of a single fixation versus the probability of skipping for WL was close to zero, which indicates that a decrease in reciprocal word length (i.e., an increase in the length of word (n)) increased the probability of single fixation. Similarly, an increase in WL increased the probability of multiple fixations. Both effects were significant. The odds ratio of the probability of a single fixation versus the probability of skipping for WL1 was 5.21 (b = 1.65, SE = 0.24, χ2(1) = 46.92, p < .001) and that of the probability of multiple fixations versus the probability of skipping was 15.64 (b = 2.75, SE = 0.31, χ2(1) = 76.21, p < .001). In other words, fixation counts on word (n) tended to increase together with a decrease in WL1. Although the estimated coefficients of WL2 single fixation versus skipping was not significant (b = 0.31, SE = 0.25, χ2(1) = 1.51, p = .22), the probability of multiple fixations versus skipping was influenced by WL2 significantly (b = 0.71, SE = 0.33, χ2(1) = 4.49, p < .05). The odds ratio of the probability of a single fixation versus the probability of skipping for WL2 was 1.36 and that of the probability of multiple fixations versus the probability of skipping was 2.03. The probability of multiple fixations on word (n) relative to the probability of skipping word (n) increased together with an increase in WL2 (see Figure 4).
Effects of word lengths on fixation count probabilities.
Word frequencies
Relative to the probability of skipping, an increase in WF decreased the probability of single fixation with an odds ratio of 0.66 (b = −0.42, SE = 0.03, χ2(1) = 284.6, p < .001) and the probability of multiple fixations with an odds ratio of 0.73 (b = −0.32, SE = 0.05, χ2(1) = 44.36, p < .001). WF2 increased the probability of single fixation versus the probability of skipping significantly with an odds ratio of 1.09 (b = 0.09, SE = 0.02, χ2(1) = 16.81, p < .001). However the effect of WF2 on the probability of multiple fixations and the effect of WF1 on FCP were not significant (see Figure 5).
Effects of word frequencies on fixation count probabilities. Units of word frequency on x-axis are log-10 power of words per million.
Predictabilities
Among predictabilities of words, only the influence of the foveal word (WP) influenced fixation counts significantly (single vs. skipping: b = −0.06, SE = 0.03, χ2(1) = 4.37, p < .05; multiple vs. skipping: b = −0.29, SE = 0.05, χ2(1) = 33.29, p < .001). An increase in WP decreased the probability of single fixation with an odds ratio of 0.94 and the probability of multiple fixations with an odds ratio of 0.75 relative to the probability of skipping. The positive influence of WP1 and the negative influence of WP2 on the probability of fixating a word were not significant (see Figure 6).
Effect of predictability of word (n) on fixation count probabilities.
Inflectional suffix counts
An increase in inflectional suffix count of word (n) (IS) increased the probability of fixating a word with an odds ratio of 1.19 for single fixation versus skipping (b = 0.17, SE = 0.04, χ2(1) = 20.7, p < .001), and with an odds ratio of 1.26 for multiple fixation versus skipping (b = 0.23, SE = 0.04, χ2(1) = 28.09, p < .001). An increase in IS1, however, decreased the probability of multiple fixations on word (n) relative to skipping it with an odds ratio of 0.91 (b = −0.09, SE = 0.04, χ2(1) = 5.24, p < .05). However, the negative effect of IS1 on the probability of single fixation versus skipping and that of IS2 on fixating a word were not significant (see Figure 7).
Effects of inflectional suffix counts on fixation count probabilities.
Interaction of frequency and length of word n
The interaction of WF and WL influenced the probability of fixating a word significantly (single vs. skipping: b = 2.44, SE = 0.2, χ2(1) = 148.6, p < .001; multiple vs. skipping: b = 5.77, SE = 0.41, χ2(1) = 200.51, p < .001). Among frequent words, a decrease in WL accompanied with an increase in the probability of skipping that word and a decrease in the probability of fixating that word once. However, among frequent words, an increase in WL increased the probability of multiple fixations slightly. A similar but stronger pattern was observed among medium frequency words. The positive effect of WL on the probability of multiple fixations and the negative effect on the probability of skipping were stronger among infrequent words than that of frequent and medium frequency words, together with a small negative effect on the probability of single fixations (see Figure 8).
The interaction effect of WL and WF on fixation count probabilities. WF values are grouped as such: high WF values are greater than three on logarithmic scale (>1,049 pm, 726 obs.), low WF values are less than two on logarithmic scale (<99 pm, 1,125 obs.), and medium WF values are between two and three on logarithmic scale (between 99 and 1,049 pm, 669 obs.).
Interaction of frequencies of word n and word n − 1
Although the interaction of WF and WF1 did not influence the probability of multiple fixations versus skipping significantly, there was a significant effect of the interaction on the probability of single fixation versus skipping (b = 0.08, SE = 0.02, χ2(1) = 26.94, p < .001). The increase in the probability of fixating word (n) once relative to skipping it together with a decrease in WF was stronger if word (n − 1) was a medium frequency or an infrequent word (see Figure 9).
The interaction effect of WF and WF1 on fixation count probabilities. Units of word frequency on x-axis are log-10 power of words per million. WF1 values are grouped as such: high WF1 values are greater than three on logarithmic scale (>1,049 pm, 711 obs.), low WF1 values are less than two on logarithmic scale (<99 pm, 1,113 obs.), and medium WF1 values are between two and three on logarithmic scale (between 99 and 1,049 pm, 696 obs.).
Interaction of frequencies of word n and word n + 1
The interaction of WF and WF2 influenced the probability of multiple fixations versus the probability of skipping significantly (b = −0.11, SE = 0.02, χ2(1) = 19.45, p < .001). An increase in WF2 increased the probability of multiple fixations on word (n) and decreased the probability of skipping word (n), if word (n) was an infrequent word. If word (n) was a frequent or medium frequency word, an increase in WF2 accompanied with a decrease in the probability of skipping word (n) but with a stable probability of multiple fixations on word (n). The effect of the interaction was not significant on the probability of single fixation versus that of skipping (see Figure 10).
The interaction effect of WF and WF2 on fixation count probabilities. Units of word frequency on x-axis are log-10 power of words per million. WF values are grouped as such: high WF values are greater than three on logarithmic scale (>1,049 pm, 726 obs.), low WF values are less than two on logarithmic scale (<99 pm, 1,125 obs.), and medium WF values are between two and three on logarithmic scale (between 99 and 1,049 pm, 669 obs.).
Interaction of length of word n and frequency of word n + 1
Similarly, the interaction of WL and WF2 influenced the probability of multiple fixations versus the probability of skipping significantly (b = 0.94, SE = 0.43, χ2(1) = 4.8, p < .05). However, the effect of the interaction was not significant on the probability of single fixation versus that of skipping. The probability of skipping a short word (n) was decreased if WF2 was increasing, while the probability of multiple fixations on a short word (n) remained almost the same. The probability of fixating versus skipping medium length words and that of long words was not influenced by WF2 (see Figure 11).
The interaction effect of WL and WF2 on fixation count probabilities. Units of word frequency on x-axis are log-10 power of words per million. WL values are grouped as such: long WL values are greater than 10 letters (75 obs.), short WL values are less than 6 letters (1,428 obs.), and medium WL values are in between (1,017 obs.).
Interaction of length of word n and predictability of word n + 1
There was a significant effect of the interaction of WL and WP2 on the probability of single fixation versus skipping (b = 0.7, SE = 0.28, χ2(1) = 6.05, p < .05). The increasing effect of WP2 on fixating a short word (n) once was lost for a medium length and long word (n). The effect of the interaction on the probability of multiple fixations versus the probability of skipping was not significant (see Figure 12).
The interaction effect of WL and WP2 on fixation count probabilities. WL values are grouped as such: long WL values are greater than 10 letters (75 obs.), short WL values are less than 6 letters (1,428 obs.), and medium WL values are in between (1,017 obs.).
The pattern of results reported here is in agreement with canonical effects of word difficulty as they relate to WL, WF, and WP: Skipping decreases and multiple fixations increase with increase in word difficulty. The results regarding the lag and successor effects and that of interaction effects were mixed. We will return to them in the “Discussion” section.
Discussion
In the present study, we reported a corpus-based analysis of eye movement control in Turkish sentence reading. The findings are largely compatible with earlier findings reported in the literature for other languages. The immediate effects of word characteristics on SFD were mostly in agreement with the previous findings, except the effect of word length of foveal word which was in the expected direction but not significant (Kliegl, 2007; Kliegl et al., 2006; Radach & Kennedy, 2013; Rayner, 1998, 2009; Rayner et al., 2012). The Fixation Count Probabilities (FCP) were influenced by the characteristics of words. The hypothesised increase in the probability of fixating a word due to increased difficulty of that word was observed in the present study. We have tested the assumption of parallel processing of words within the perceptual span and that of dynamical modulation of perceptual span by foveal processing difficulty presented in Kliegl (2007) by inclusion of the characteristics of neighbouring words and several interactions in the models, respectively. The results regarding these assumptions were mixed. The spillover effect of previous word frequency and parafoveal-on-foveal frequency effect of next word frequency were observed on SFD values—both frequencies showed a negative relationship with SFD. There was a reversed effect of next word frequency on FCP—that is, an increase in the difficulty of the next word due to low frequency decreased the probability of fixating the foveal word once relative to the probability of skipping it. These results imply that if the next word was a low-frequency word, foveal word tended to be either skipped or fixated longer. The pattern of the relationship between next word frequency and FCP was not observed among long or infrequent foveal words. Among long foveal words, the probability of multiple fixations relative to skipping probability was higher independent of the frequency of the next word. This implies a dominance of the effect of the length of foveal word (see Figure 11). Among infrequent foveal words, however, there was an effect of the frequency of the next word on the probability of multiple fixations relative to skipping probability. The probability of multiple fixations was increased relative to skipping probability as the frequency of next word increased if the foveal word was an infrequent word. However, the probability of single fixation was independent of the frequency of next word and it was higher than the other probabilities (see Figure 10). The interaction effect of the frequency of the foveal and the frequency of the next word on FCP of the foveal word provides support for a dynamical modulation of the perceptual span. Another interaction included in the models to test dynamical modulation of perceptual span assumption was the interaction effect of foveal word frequency and previous word frequency on SFD and FCP. Supporting evidence for the assumption that a difficult (i.e., low-frequency) previous word would result in a stronger effect of the frequency of the foveal word was observed. The effect of foveal word frequency among infrequent previous words was more pronounced, on both dependent variables. However, there was no evidence for the expected effects of inflectional suffix counts on SFD, the expected effects of predictabilities of neighbouring words, and the expected interaction effect of length of foveal word and predictability of next word. So, the controversy is not resolved with this study. In the following, we present a more detailed discussion of the results together with controversial effects.
Immediate effects of word characteristics
The frequently reported effects of the characteristics of word (n) on SFD were significant, except for the length of word (n) (WL; Kliegl, 2007; Kliegl et al., 2006; Radach & Kennedy, 2013; Rayner, 1998, 2009; Rayner et al., 2012). An increase in Word Frequency (WF) and Predictability (WP) of word (n) were accompanied by a decrease in SFD. Although the effect of WL on SFD was not significant, the relationship was positive, 8 as expected. The results of FCP revealed that both the probability of single fixation and the probability of multiple fixations versus skipping was significantly influenced by the characteristics of word (n). These effects were compatible with the effects observed in the results of SFD model. In other words, an increase in WP and WF and a decrease in WL (i.e., an increase in word difficulty) increased the probability of fixating word (n). Although there was no effect of inflectional suffix count of word (n) (IS) on SFD, a larger IS count increased the probability of fixating word (n). One reason for not observing an effect of IS on SFD may have been a lack of sufficient statistical power; the number of inflectional suffixes correlates very highly with WL (r = .75). 9 Future research is needed to understand the effect of inflectional suffix count in an experiment in which the confounding effect of word length is eliminated.
Another frequently reported effect, IOVP, was also observed in this study (Kliegl, 2007; Kliegl et al., 2006; Vitu et al., 2001). Given that fixations at the edges of a word are frequently the result of a saccadic error (i.e., an overshoot or undershoot of the intended target), one explanation is that in case of such a misplacement a new saccade is programmed immediately and, as a consequence, fixation durations that do not fall on the optimal viewing position (i.e., approximately the middle of the word) are shorter (Nuthmann et al., 2005). In our study, SFD values were higher if the fixation location was around the middle of the word than SFD values of fixations at the edges of the word. This effect was observed through the significant quadratic effect of RFL on SFD values. Although the oculomotor parameters were not included in the BCLM of FCP because skipping instances did not have those covariates, the effect of RFL on the number of fixations was observed indirectly in the results of the LMM of gaze duration (GD), provided in the online Supplementary Material. Gaze durations of instances that have relative fixation position towards the centre of words were shorter than those of instances that have relative fixation position towards the beginning of words. As SFD data are a subset of gaze duration data, the increase in SFD together with the decrease in GD when fixations are around the centre of words were considered as an indication of decreasing number of fixations.
Lag effects
The effects of the previous word (word n − 1) on the SFD on word (n) reported in Kliegl (2007) were also examined. As the distance between two fixations increases, the parafoveal benefit obtained during previous fixation decreases. Therefore, for fixations that have longer ISAs, SFD values were expected to be longer. The positive relationship between ISA and SFD was significant in the present study. The negative relationship between the length of word n − 1 (WL1) and SFD was investigated further. An inspection of the raw data revealed that short WL1 values were accompanied by long ISA values, because short previous words were skipped frequently (55% of the word n − 1 that have less than four characters). Fixations after skipped words are often longer than after non-skipped words. Similarly, the probability of fixating word (n) was increased together with a decrease in WL1 because of an increase in skipped word (n − 1). For a long word (n − 1), the probability of skipping word (n) was increased which imply multiple fixations on word (n − 1), shorter incoming saccades, and ease in parafoveal processing of word (n).
Although the hypothesised relationship between inflectional suffix counts and SFD was not observed, an increase in IS1 decreased the probability of multiple fixations versus the probability of skipping. As the effect of IS on the probabilities of fixation counts suggests, a word (n − 1) that has a smaller IS1 tends to be skipped. Similar to the effect of WL1, a skipped word (n − 1) increased the probability of multiple fixations on word (n). However, the relationship between IS1 and FCP requires further investigation.
According to the spillover effect, lower values of the frequency of word (n − 1) (WF1; i.e., difficulty of the previous word) have an increasing effect on SFD because of the incomplete processing of word (n − 1; Kliegl, 2007; Radach & Kennedy, 2013; Rayner, 1998, 2009; Rayner et al., 2012). In the current study, we observed a significant increase in SFD for low WF1; the corresponding effect for WP1 was neither in the expected direction nor significant. In Kliegl (2007, Table 1), the WF1 effect was significant overall and in all nine subexperiments; the WP1 effect was significant overall but only in six of nine subexperiments. The effects of WF1 and WP1 on the relative probabilities of fixation counts were not significant but they showed expected influences. In other words, increase in WF1 and WP1 increased the probability of fixating word (n) due to increased skipping probability of an easy word (n − 1) but not significantly.
Another likely source of the relationship between WF1 and SFD is the dynamical modulation of perceptual span depending on the difficulty of a word. Low-frequency values restrict the perceptual span and hence decrease parafoveal benefit. Therefore, the effect of WF on SFD should be stronger for a low-frequency previous word. The interaction of WF and WF1 was included in the model to test dynamical modulation of perceptual span. The interaction was significant. In other words, the stronger WF effect on SFD when previous word was a low-frequency word was observed in the current study. The corresponding interaction was also significant overall and in eight of nine subexperiments in Kliegl (2007, Table 2). The interaction of WF1 and WF influenced the probability of fixating word (n) once versus skipping it significantly. The decreasing effect of WF on the probability of single fixation on word (n) relative to skipping it got stronger as WF1 decreased.
Successor effects
The effect of the frequency and predictability of the next word (WF2 and WP2) on SFD has a somewhat controversial status, for example, Kliegl, (2007, Table 1) reports a significantly negative WF2 effect and a significantly positive WP2 effect; both effects were significant in nine subexperiments. In this study, the effect of WF2 on SFD was also significant in the expected negative direction, but the positive effect of the predictability of word n + 1 (WP2) was not significant, although the estimate was in the expected direction. The effect of WP2 on the probability of fixation counts was not significant. The probability of fixating word (n) once versus skipping it was decreased significantly together with a decrease in WF2. A similar effect for the probability of multiple fixations was found; however, the effect was not significant. In other words, if the next word was an infrequent one, skipping the foveal word increased. However, if the word (n) fixated once and the word (n + 1) was an infrequent word, the duration of single fixation on the foveal word increased.
Dynamical processing in the perceptual span could also be reflected in interaction terms involving the properties of the current and the next word. Kliegl (2007, Table 1) tested interactions between (a) WF and WF2, (b) WF2 and WL, and (c) WP2 and WL. There was no evidence for (a), but (b) and (c) were significant overall, but only in two of nine and six of nine subexperiments for (b) and (c), respectively. Neither of the effects of these interactions on SFD was significant in the present study. The same interactions were tested for their effect on the probability of fixation counts. The probability of multiple fixations versus skipping was influenced by (a) and (b), and the probability of single fixation versus skipping was influenced by (c) significantly. Among frequent and medium frequency words, an increase in WF2 accompanied with a decrease in the probability of skipping the foveal word with a stable probability of multiple fixations. In other words, if the foveal word was a relatively easy word, the effect of WF2 observed on FCP was the same as its main effect. However, among infrequent words, the probability of multiple fixations increased together with an increase in WF2, implying a dominance of the effect of WF on FCP.
The quadratic OSA influenced SFD significantly. According to this effect, for short OSA values, the relationship between OSA and SFD was negative, while the direction of the relationship was positive for long OSA values. An inspection on raw data revealed that OSA values shorter than five characters accompanied with positive Relative Fixation Location (RFL) values (54% of 2,180 words that have OSA values shorter than five characters). In other words, when the fixations were on the right half of the word, OSA values tended to be short. The rest of short OSA values which were observed together with negative RFL values belonged mostly to words that have less than six characters (80% of 1,011 observations). The relationship between OSA, RFL, and WL implies saccadic errors which preceded a corrective saccade that were programmed immediately, resulting in shorter SFD values. For longer OSA values, the expected positive relationship between OSA and SFD was observed, which reflect next saccade programming time (Kliegl et al., 2006). Length of word n + 1 (WL2) had a significant negative effect on SFD. The effect of WL2 on SFD may be conceived as an indicator of the increase in the activation of the next word and hence attraction of attention to it due to its length. This interpretation was supported by the increased probability of skipping word (n) and the decreased probability of multiple fixations on word (n) significantly together with an increase in WL2.
Conclusion
Eye movements during reading of Turkish sentences replicate well-known main effects known for other scripts such as those relating to word frequency, length, and predictability of the fixated word. An additional analysis of the probability of fixation counts revealed compatible effects of frequency, word length, and predictability of word on the probability of fixating a word once, more than once, or skipping it. As expected, longer, less frequent, and less predictable words (i.e., more difficult words) had an increased probability of fixating that word. In addition, the local context of fixation positions relating to the relative fixation location (i.e., the IOVP effect) as well as effects relating to incoming and OSAs are by and large in agreement with previous research.
We also tested effects relating to word properties of the neighbouring words and to interactions between properties of neighbouring words. Usually, these effects are much smaller and the evidence was mixed. Some of the effects replicated both in significance and in direction of effect; for some effects, the estimates were in the expected direction, but not significant; and one of them (i.e., the effect of the previous word predictability on SFDs) was neither significant nor in the expected direction. From a single study, it is difficult to unravel whether discrepancies are due to false positives in earlier research, due to differences in statistical power, or due to genuine differences relating to Turkish script. In perspective, the differences need to be resolved.
The special features of Turkish script are its agglutinative morphology and its relatively shallow orthography–phonology mapping. In the current study, we tested the effect of morphological complexity of words on SFDs and on the probability of fixation counts. Although there was not any evidence for the effect of morphological complexity on fixation duration, the number of suffixes influenced the probability of fixating a word. The effects of the number and type of suffixes of the foveal and neighbouring words on the skipping probability of foveal word require further investigation. The features of Turkish hold much promise for disentangling the distributed processing of words in the perceptual span.
Supplemental Material
200420_Turkish_Reading_Supplement – Supplemental material for Eye movement control in Turkish sentence reading
Supplemental material, 200420_Turkish_Reading_Supplement for Eye movement control in Turkish sentence reading by Ayşegül Özkan, Figen Beken Fikri, Bilal Kırkıcı, Reinhold Kliegl and Cengiz Acartürk in Quarterly Journal of Experimental Psychology
Footnotes
Acknowledgements
We thank Simon P. Liversedge, Tim Slattery, Denis Drieghe, and an anonymous reviewer whose comments and suggestions helped improve and clarify the manuscript.
Declaration of conflicting interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study has been funded by TUBITAK (The Scientific and Technological Research Council of Turkey) 113K723 “Okumada Bilişsel Süreçlerin İncelenmesi: Göz Hareketleri Kontrol Modellemesi için Türkçe Okuma Örüntüleri Derlemi; The Investigation of Cognitive Processes in Reading: Development of a Corpus of Turkish Reading Patterns for Eye Movement Control Modeling.”
Notes
References
Supplementary Material
Please find the following supplemental material available below.
For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.
For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.
