Kahraman, B. (2010). Incremental processing of gap-filler dependencies in Turkish: Focusing on the processing asymmetry between subject and object relative clauses. Bulletin of the Graduate School of Education, Hiroshima University. Part. II, Arts and science education. 59 , 239-248. DOI: http://doi.org/10.15027/31042 Acknowledgments: This paper was originally published in Bulletin of the Graduate School of Education, Hiroshima University. Part. II, Arts and science education Issue 59 in Japanese ( 広島大学大学院教育学研究科紀要 第二部 , 文化教育開発関連領域 59 号 ) . To make this research more accessible to a wider audience, the author translated the paper into English. Japanese version can be retrieved from https://ir.lib.hiroshima-u.ac.jp/journals/BullGradSchEducHU-Part2/i/59/item/31042. This paper is a part of author’s doctoral dissertation, and reviewed by Hiromu Sakai (supervisor), Ruiko Ohama, Kazumitsu Chujo, Yukiko Hatasa. I deeply appreciate their constructive comments. All remaining errors, including translation, are my own. Incremental processing of gap-filler dependencies in Turkish: Focusing on the processing asymmetry between subject and object relative clauses Barış KAHRAMAN Abstract: Previous studies have shown that the filler-gap dependencies are processed incrementally in both English and Japanese. In the case of relative clauses (RC), unlike English, the gap precedes the filler in Japanese (henceforth, the gap-filler dependencies). From the existing RC processing data in Japanese, it is hard to judge whether the gap-filler dependencies are also processed incrementally. In the present study, using Turkish subject and object RCs, we investigated whether the processing of gap-filler dependencies is incremental. The results showed that when the RC structure is determined at the RC-verb position, subject RCs were read faster than object RCs at this position. This result suggests that as soon as the gap-site is determined at the RC-verb, the Turkish parser immediately postulates a filler, and starts to construct a dependency even before the filler appears. Thus, the processing of gap-filler dependencies is also incremental in a language like Turkish. Keywords: G ap-filler dependencies, incremental processing, relative clauses, Turkish, Japanese 1. Introduction One of the most important goals of sentence processing research is to explain the constraints and principles that govern human sentence processing, and build a universal model for sentence processing mechanism. To this end, various hypotheses and models of sentence proces sing, mainly in English, have been proposed and tested. Nakayama et al. (2006) point out that SOV languages such as Japanese, which are typologically di stinct from English, are critical for testing the validity of the hypotheses proposed through English and clarifying both universal and language-specific aspects of sentence processing. So far, in Japanese, most studies have contrasted it with English, which has a different basic word order, and less attention has been paid to contrastive studies within typologically similar languages, such as Korean and Turkish. However, even if the basic word order is the same, the use of case particle s and verb conjugations may differ in such languages (see section 1.3). In order to deepen our understanding on the universal and language-specific aspects of human sentence processing, it may be critical to conduct contrastive studies of sentence processi ng between similar languages. This is because it can show explicitly and concretely how the similarities and differences bet ween languages actually relate to the sentence processing. In addition, contrastive studies on sentence processing of typologically similar languages would also make it possible to examine and explain facts and phenomena that are difficult to observe or verify in one language in another. At present, however, few studies of se ntence processing focus on these structural differences within the SOV languages. 1.1. Incremental processing of filler-gap dependencies Relative clauses (hereafter, RCs) have received significant attention in the sentence processing research. One of the main reasons for this is that the syntactic features of RCs exhibit differences and similarities across languages. This allows researchers to explore and explain how filler and gap dependencies, which are essential for sentence comprehension, are constructed in various languages (Miyamoto, 2006). A filler refers to an argument that moves from its original position and appears in a different position within a sentence. A gap refers to the orig inal position of the filler. For example, in the English example (1), who is the filler, and the underline indicates the gap position. In the Japanese example (2), sensei (teacher) is the filler, and the underline represents the gap. Fillers and gaps are observed in structures such as WH-questions, RCs, cleft sentences, scrambled sentences, etc. In processing such structures, in order to enable semantic understanding, it is considered that the filler needs to be interpreted at the gap position where the semantic role is originally assigned, and then the filler and gap should be appropriately linked. This phenomenon is referred to as filler-gap dependencies (e.g., Fodor, 1978). (1) The teacher who i the student saw i was young. (2) Gakusei-ga i mita sensei i wa waka-katta. Student-nom saw teacher-top young-pst ‘ The teacher who the student saw was young.’ In RCs, there is a reversed order of filler and gap between English and Japanese. In English, the filler precedes the gap, while in Japanese, the gap precedes the filler. In this study, we refer to the case where the gap precedes the filler, as gap-filler dependencies , as observed in Japanese. Previous studies in various languages and structures have examined and confirmed that filler-gap dependencies , where the filler precedes the gap, are constructed incrementally. This means that the processing of filler-gap dependencies occurs without delay (i.e. English: Crain & Fodor, 1985; Stowe, 1986; Dutch: Fraizer, 1987; Japanese; Aoshima et al., 2004 ). When the parser 1 detects the filler, it immediately assumes the presence of a gap and begins to establish a dependency between the filler and the gap, without waiting to determine the exact position of the gap. Consequently, the processing of filler-gap dependencies is incremental, as the construction of dependencies initiates before the position of the gap is determined. The study by Aoshima et al. (2004) investigates the processing of scrambled interrogative sentences and provides a notable example that illustrates the incremental processing of filler-gap dependencies in Japanese. Aoshima et al. (2004) compared the reading times of sentences such as (3a) and (3b). In (3b), all arguments are in their canonical position, resulting in the absence of a filler or gap. In contrast, (3a) involves the movement of the dative-marked phrase dono kodomo-ni (which child-DAT) from its canonical position to the sentence-initial position. In this case, the phrase dono kodomo marked with the dative case particle functions as the filler, while its original position immediately following hahaoya-wa (mother-TOP) serves as the gap. 1 It is assumed that humans possess a device, known as a parser, which enables them to efficiently anal yze and identify the underlying syntactic structure of a sentence (Hirose, 2004). (3a) Dono-kodomo-ni hahaoya-wa otetudaisan-ga which-child-DAT mother-TOP housekeeper-NOM chichioya-ni obento-o watasita-to iimasita-ka? father-DAT lunch-ACC handed-COMP told-Q ‘ To which child did the mother tell that the housekeeper handed lunch to the father? ’ (3b) Dono-kodomo-ga hahaoya-ni otetudaisan-ga which-child-NOM mother-DAT housekeeper-NOM chichioya-ni obento-o watasita-to iimasita-ka? father-DAT lunch-ACC handed-COMP told-Q ‘ Which child told the mother that the housekeeper handed lunch to the father? ’ According to Aoshima et al., Japanese native speakers tend to interpret the initial dative noun in a sentence like (3a) as the argument of the first verb. In (3a), the presence of the embedded clause becomes apparent upon encountering otetudaisan-ga (housekeeper-NOM), leading to the interpretation that the first verb belongs to the embedded clause. In other words, dono kodomo-ni is interpreted as the argument of the embedded verb, and a gap is posited within the embedded clause to establish the dependency. However, it is revealed that chichioya-ni (to the father-DAT) actually exists within the embedded clause, contradicting the assumption that the first dative noun (dono kodomo-ni) functions as an argument of the embedded clause. On the other hand, in (3b), there is no need to establish a filler-gap dependency Therefore, Aoshima et al. predicted that the reading time for chichioya-ni in (3a) would be longer than in (3b) when comparing the two sentences. The results of the self-paced reading experiment, which measured word-by-word reading times, supported this prediction. These findings suggest that when the parser detects the filler in Japanese, it attempts to posit a gap within the embedded clause and establish a dependency based on this assumption, indicating that the processing of filler-gap dependencies occurs incrementally. As of now, it is challenging to draw a clear conclusion regarding the incremental processing of gap-filler dependencies in cases where the gap precedes the filler. In other words, it is difficult to determine whether the parser immediately assumes the presence of a filler upon detecting a gap and begins constructing the dependency before confirming the presence of the filler. This difficulty arises from the limitations of existing studies on Japanese relative clause processing, as explained below. 1.2. Processing of gap-filler dependencies It is well known that there is a processing difference between subject relative clauses (SR) and object relative clauses (OR), where the subject and object of the original sentence are relativized, respectively, as shown in (4). This difference is commonly referred to as the processing asymmetry of SR and OR. Numerous studies have demonstrated that SRs are generally easier to process than ORs in various languages, and this processing asymmetry is attributed to the cognitive load involved in constructing the dependency between the filler and the gap (e.g., Frazier & Clifton, 1989; Gibson, 1998; O'Grady, 1997). 2 For instance, Frazier & Clifton (1989) argue that in English, the parser postulates the presence of a gap as soon as it encounters the filler, allowing for quicker processing of SRs compared to ORs. In contrast, Gibson (1998) emphasizes the linear distance (i.e., the number of words) between the filler and the gap as a determinant of RC processing difficulty, while O'Grady (1997) argues that the structural distance (i.e., the number of syntactic nodes) between the filler and the gap plays a crucial role. Although the explanations for the processing asymmetry may differ among researchers, they all share the common assumption that the parser constructs the dependency between the filler and the gap during the processing of RCs. (4a) SR: i Gakusei-o mita sensei i wa waka-katta. Student-acc saw teacher-top young-pst ‘The teacher who saw the student was young. ’ (4b) OR: Gakusei-ga i mita sensei i wa waka-katta. Student-nom saw teacher-top young-pst ‘The teacher who the student saw was young. ’ Studies investigating the processing of SRs and ORs in Japanese have consistently shown that the reading time of 2 The processing asymmetry between SRs and ORs has been found to be influenced by various factors, including the animacy of the head noun, frequency of RCs, discourse function, predictability, and more. However, this does not imply that the processing of dependencies between the filler and the gap is not involved in RC processing. In this study, we posit that the construction of dependencies between the filler and the gap is crucial during the processing of RCs, and factors such as animacy, frequency, predictability, etc. play a role in the processing of these dependencies. the head noun (filler) sensei-wa is faster in SRs co mpared to ORs (Miyamoto & Nakamura, 2003; Ishizuka, 2005; Ueno & Garnsey, 2008). This suggests that the construction of the dependency between the filler and the gap in SRs is easier compared to ORs in Japanese. However, these results alone do not provide a clear indication of the incrementality of the gap-filler dependencies , specifically whether the processing of the dependency begins before the filler is encountered. The reason for this difficulty is elaborated upon below. In Japanese, RCs do not have relative pronouns or specific inflectional forms (RC-markers), and the omission of subject or object ( pro-drop ) is common. As a result, the Japanese parser may initially assume that it is processing the main clause until the RC-verb is encountered. Even if a gap is detected at the position of the RC-verb, the parser may interpret the gap as a pro , indicating that the corresponding filler may not necessarily appear. Therefore, the Japanese parser cannot determine whether a filler will appear later at the RC-verb position, and it may start constructing the dependency between the gap and filler only after the filler is encountered. This suggests that the construction of the gap-filler dependency may not be incremental in Japanese. However, there are also experimental results that report Japanese native speakers produce RCs at a rate of 80% or more when the sentence is presented up to the verb, such as gakusei-ga/o mita... (Ueno & Garnsey, 2008). This indicates that Japanese native speakers assume the presence of a filler at the position of the verb in incomplete sentences. In other words, the Japanese parser may start constructing the dependency between the gap and filler before the filler appears, suggesting that the processing of gap-filler dependencies could be incremental. However, as mentioned above, it is still possible to interpret the gap at the position of the RC -verb as a pro Furthermore, the processing asymmetry between SRs and ORs is consistently observed at the position of RC -head (filler). Therefore, it remains difficult to determine exactly when the processing of the gap-filler dependency starts based on Japanese data alone. To draw clearer conclusions about the incrementality of gap-filler dependencies , it would be necessary to investigate a language in which it can be confirmed at the verb position that the gap is not a pro and that the corresponding filler will definitely appear later. Turkish, like Japanese, is a pro-drop language, but the parser is capable of recognizing that the gap is not a pro at the RC-verb position. Therefore, Turkish is considered an effective language to examine the incrementality of gap-filler dependency processing. 1.3. Subject and object relative clauses in Turkish In Turkish RCs, similar to Japanese, the gap appears before the filler and precedes the head noun. Additionally, case particles are used in Turkish. In SRs, as in the case of Japanese, the embedded object takes the accusative case particle. On the other hand, in ORs, the embedded subject always takes the genitive case particle. Moreover, the RC-verb is inflected by a specific RC-marker in Turkish. In SRs, the verb is accompanied by the marker An , which is primarily used to relativize the subject. In ORs, the verb is accompanied by the marker DIK , which is used to relativize non-subject elements of the sentence (e.g., Kornfilt, 1997). 3 (7a) SR: [ i Öğrenci -yi gör- en] öğretmen i genç-ti [Student-acc see-srp] teacher young-past 4 ‘The teacher who saw the student was young.’ (7b) OR:[Öğrenci -nin i gör- düğ - ü] öğretmen i genç-ti [Student-gen see-nsr-3sg] teacher young-past ‘The teacher who the student saw was young.’ The fundamental difference between Japanese and Turkish lies in the presence of RC-markers. In Japanese, RC-markers are not used, and the parser can only determine the presence of the RC, or the non - pro status of the gap, at the position of the head noun (filler). In contrast, Turkish employs RC-markers like en and düğü , allowing the parser to identify the presen ce of the RC at the RC-verb position. As the gap within the RC is not a pro , the parser can predict that a filler will appear later and process the sentence accordingly. Hence, Turkish serves as a suitable language for investigating the incrementality of gap-filler dependency processing. 3 Due to vowel harmony, the pronunciation and spelling of RC-markers may change (i.e., En and DIG ). 4 This study uses the following glos ses: acc: accusative case; gen: genitive case; loc: locative case ; nsr: non-subject relative marker; past: past tense; prog: present progressive tense; srp: subject relative marker; 3s: third person singular. If the Turkish parser assumes the presence of a filler at the RC-verb position and initiates the construction of the dependency before the filler appears, it is possible to observe the processing asymmetry between SRs and ORs at the RC-verb position. If this can be proved, then we would provide a piece of evidence that the processing of gap-filler dependencies is also carried out incrementally. Building on this premise, the res earch tasks were formulated, and two sentence completion tasks (Experiment 1 and 2) as well as self-paced reading experiments (Experiment 3 and 4) were conducted. 【 Task 1 】 : Ishizuka (2005) argued that in Japanese, when an accusative noun appears at the sen tence-initial position, it is easier to predict the presence of an RC compared to when a nominative noun appears at the beginning of the sentence. This difference in predictability may influence the processing of SRs and ORs. Therefore, in Experiment 1, we aim to verify whether there is a difference in the predictability of RCs between the input of accusative and genitive nouns in Turkish. Subsequently, in Experiment 2, we investigate whether Turkish native speakers predict the presence of RCs before the filler is encountered. 【 Task 2 】 : In Experiments 3 and 4, we will examine whether there is a processing difference between SRs and ORs in Turkish, and identify the specific stage of RC processing where the processing asymmetry emerges. Through this study, we aim to provide empirical evidence regarding the incrementality of processing gap-filler dependencies Additionally, from a cross-linguistic perspective, we can gain insights into how typological characteristics, such as the presence or absence of RC-markers, influence the processing of RCs in SOV languages. 2. Sentence completion tasks 2.1. Experiment 1 2.1.1. Aim The purpose of Experiment 1 was to verify if there is a difference in the predictability of RCs when an accusative noun or a genitive noun is presented. Additi onally, we examined the effect of adverbs following these nouns. 2.1.2. Material In Experiment 1, four conditions were established, as shown in (8a) to (8d), and incomplete sentences were presented up to the point just before the verb izle (to watch). In (8a) an accusative noun, and in (8b) a genitive noun appears after a locative noun. In (8c), an adverb appears soon after the sentence initial accusative noun, while in (8d) the adverb appears after the genitive noun. A total of 128 sentences were used, comprisin g 32 pairs of stimuli based on the four conditions, as presented in Table 1. A Latin square design was employed for presenting the test sentences, and the 32 experimental sentences were divided into four groups for each condition to prevent the duplication of sentences for each participant. Additionally, 40 filler sentences, disti nct from (8), were prepared as filler sentences, and a questionnaire was created by randomly arranging the test and filler sentences. Thus, each participant was presented with a total of 72 incomplete sentences. Table 1 Test sentences in Experiment 1 & 2 (8a) Locative ＋ accusative Fabrika - da ustabaşı - nı (izle - yen)_________ Factory - loc foreman - acc watch - srp (8b) Locative ＋ genitive Fabrika - da ustabaşı - nın (izle - diğ - i)________ Fact ory - loc foreman - gen watch - nsr - 3s (8c) Accusative ＋ adverb Ustabaşı - nı dikkatle (izle - yen)_________ Foreman - acc carefully watch - srp (8d) Genitive ＋ adverb Ustabaşı - nın dikkatle (izle - d iğ - i)________ Foreman - acc carefully watch - nsr - 3s 2.1.3. Participants & procedure Thirty-seven students from a university in Turkey took part in Experiment 1. All participants were native speakers of Turkish with an average age of 21 (SD=1.27). They had not participated in any other experiments apart from Experiment 1. The particip ants were given four sets of questionnaires during a class session and were asked to complete one set of incomplete sentences using natural Turkish. 2.1.4. Results & discussion A total of 122 sentences that were either left incomplete by the participants or deemed unacceptable in Turkish were excluded from the analysis. As a result, a total of 1062 sentences were included in the analysis. The produced sentences were divided into two groups: RCs and Non-RCs (sentences other than relative clauses). In the locative + accusative condition, out of the produced sentences, 63 (24%) were RCs, while the remaining 202 (76%) were Non -RCs, such as simple sentences, subordinate clauses, and adverbial clauses. In the locative + genitive condition, 40 sentences (15%) were RCs, while 235 sentences (85%) were Non -RCs. In the a ccusative + adverb condition, 50 sentences (19%) were RCs, while 208 sentences (81%) were Non -RCs. Lastly, in the genitive + adverb condition, 83 sentences (31%) were RCs, while 181 sentences (69%) were Non-RCs. Table 2 Numbers and rates of RC and Non-RC production RC Non - RC total n (%) n (%) n a. loc ＋ acc 63 24 202 76 265 b. l oc ＋ gen 40 15 235 85 275 c. acc ＋ adv 50 19 208 81 258 d. gen ＋ adv 83 31 181 69 264 total 236 22 826 78 1062 A binomial test was conducted assuming that the production rates of RCs and Non -RCs were equal for each condition. The results of the test were significant in all conditions ( p < .01), indicating that Non -RCs were significantly more frequent than RCs in each condition. Specifically, in the conditions where locative nouns were used, RCs were produced more frequently after a ccusative nouns compared to genitive nouns. Conversely, in the conditions where adverbs were used, RCs were produced more frequently after genitive nouns compared to accusative nouns. Discussion : The results of Experiment 1 show that Turkish native speakers are more likely to produce non-RC structures than RCs in response to incomplete sentences without a verb. It is considered that the proportion of elements produced in the sentence completion tasks reflect their predictability (see Muraoaka (2006) for relevant discussion). Therefore, findings of Experiment 1 suggests that Turkish speakers do not predict the occurrence of RCs unless the RC -verb is present. However, when focusing on the cases where RCs were produced, it is observed that the predictability of RCs varies depending on the presence of adverbs. In the conditions where adverbs were not used, specifically in the locative + accusative and locative + genitive conditions, more RCs were produced after the accusative noun compared to the genitive noun. This may be due to the easier detection of the gap when using accusative nouns compared to genitive nouns. In the locative + accusative condition, the absence of the subject made it easier to notice the presence of the gap, whereas in the locative + genitive condition, there was no missing element, eliminating the need to assume a gap. This suggests that, in situations where a gap exists, the Turkish parser can relatively easily predict the presence of an RC. These findings are consistent with Ishizuka's (2 005) claim for Japanese that "RCs are more predictable when the accusative noun appears at the beginning of the sentence than when the nominative noun appears at the sentence initial position," suggesting that a similar pattern may apply to Turkish. In the genitive + adverb condition, where an adverb was used, more RCs were produced compared to the other conditions. This can be attributed to the fact that in Turkish, it is easier to produce a noun phrase after a genitive noun, such as öğretmenin arabası (the teacher's car ), while it is easier to produce a verb phrase after an adverb following a genitive noun, such as öğretmenin şiddetle eleştirdiği ... (the teacher strongly criticized ...). If a verb phrase is produced after a genitive noun in Turkish, it is not possible to complete it as a simple sentence, and instead, a complex sentence structure must be used. One of the ways to form complex sentences is through the use of RCs. Therefore, it is possible that more RCs were produced in the genitive + adverb condition compared to the other conditions. Based on these findings, Experiment 4 will further investigate how the differences in the predictability of RCs when accusative and gen itive nouns, as well as adverbs are presented, influence the processing of RCs in Turkish. Before conducting Experiment 4, Experiment 2 will examine the structures th at participants predict when RC-verbs are presented after the incomplete sentences used in Experiment 1. 2.2. Experiment 2 2.2.1. Aim In Experiment 2, the aim was to investigate whether Turkish native speakers predict the presence of RCs consisting of gap and filler, when the RC verbs are presented. 2.2.2. Material In Experiment 2, the same methodology as in Experiment 1 (2.1.2) was followed, with the addition of RC-verbs to the incomplete sentences. 2.2.3. Participants & procedure In Experiment 2, a total of 22 participants from a university in Turkey took part. All participants were native speakers of Turkish with an average age of 20 (SD=0.87). They did not participate in any experiments other than Experiment 2. The procedure was identical to Experiment 1. 2.2.4. Results & discussion A total of 14 sentences that were not completed by the participants or were deemed unacceptable in Turkish were excluded from the analysis. Therefore, a total of 690 sentences were included in the analysis. The completed sentences were categorized into two groups: RCs and Non-RCs. Table 3 Numbers and rates of RC and Non-RC production RC Non - RC total n (%) n (%) n a. loc ＋ acc 171 97 5 3 176 b. l oc ＋ gen 171 99 2 1 173 c. acc ＋ adv 167 96 7 4 174 d. gen ＋ adv 167 100 0 0 167 total 676 98 14 2 690 In the locative + accusative condition, 171 (97%) of the completed sentences were RCs, and 5 (3%) were Non-RCs. In the locative + genitive condition, 171 (97%) were RCs, and 2 (1%) were Non-RCs. In the accusative + adverb condition, 167 (96%) sentences were RCs, and 7 (4%) were Non-RCs. All of the structures produced in the genitive + adverb condition were RCs (167 sentences). A binomial test assuming equal production rates of R Cs and Non-RCs was conducted for each condition, and the results were significant in all conditions ( p < .01). This finding suggests that RCs were significantly more frequently produced than Non-RCs in each condition. Discussion : Experiment 1 demonstrated that native speakers of Turkish do not predict the occurrence of RCs in the absence of RC-verbs. However, Experiment 2 confirmed that native speakers of Turkish predict the presence of RCs almost 100% of the time when RC -verbs are presented. This indicates that the Turkish parser promptly identifies the existence of a gap and anticipates the appearance of a filler as soon as the RC -verb is encountered. Based on the obtained results above, it is possible to make following predictions regarding online sentenc e processing. If the Turkish parser strongly predicts the presence of an RC structure at the position of the RC-verb, it is expected to initiate the construction of a gap-filler dependency at that position. In this case, the processing load asymmetry observed in the construction of dependencies would be observed at this position. In other words, differences in processing load between SRs and ORs are expected to manifest at the RC -verb position in Turkish. In Experiments 3 and 4, we will examine whether this prediction can be supported. 3. Self-paced reading experiments 3.1. Experiment 3 3.1.1. Aim The purpose of Experiment 3 is to investigate whether there is a difference in the processing load betwee n subject and object relative clauses in Turkish, and if such a difference exists, to determine where the processing asymmetry is observed. 3.1.2. Material In Experiment 3, two conditions were established as shown in Table 4. In the SR condition, the sentence begins with an accusative noun followed by an RC -verb. In the OR condition, the sentence starts with a genitive noun immediately followed by an RC-verb. The remaining parts of the sentences are the same in both conditions. In self-paced reading experiments, there is a possibility of a spillover effect , where an effect observed in one region may influence the immediately following region. To detect the potential spillover effect , an adjective was inserted between the RC-verb and the head noun in Experiment 3. A total of 24 sets of 48 sentences were prepared as test sentences, as shown in Table 4. The test sentences were presented to participants using a Latin square design, and the 24 pairs of sentences were divided into two for each condition. Additionally, 72 filler sentences that were different from those in Table 4 were prepared, resulting in a total of 96 experimental sentences that were randomly presented to each participant. Furthermore, a norming survey was conducted to ensure that there were no differences in the naturalness of the test sentences. Table 4 Test sentences in Experiment 3 Region 1 2 3 4 5 6 7 (9a) SR: Ustabaşı-nı izle-yen yetenekli işçi söylenenleri harfiyen yap-ıyor-du. Foreman-acc watch-srp skillful worker directives completely do-prog-past The skillful worker who watched the foreman has always obeyed the commands fully.' (9b) OR: Ustabaşı-nın izle-diğ-i yetenekli işçi söylenenleri harfiyen yap-ıyor-du. Foreman-gen watch-nsr-3s skillful worker directives completely do-prog-past The skillful worker who the foreman watched has always obeyed the commands full.' Prediction 1 : If native speakers of Turkish start constructing the gap-filler dependency at the RC-verb position, we would expect to observe processing asymmetry between SRs and ORs at that position. Prediction 2 : If the construction of the gap-filler dependency starts after the filler is encountered, then the native speakers of Turkish would initiate constructing the gap-filler dependency at the RC-head position. In this case, the processing asymmetry between SRs and ORs would be observed at the RC-head. 3.1.3. Participants & procedure Thirty-five students from a university in Turkey participated in Experiment 3. All were native speakers of Turkish with an average age of 23 (SD=4.87). They did not participate in any other experiments besides Experiment 3. The experiment was conducted using Linger 2.94 (developed by Dough Rohde), through a self-paced reading task where participants' reading times were measured for each word in a moving window. Participants were instructed to read the sentences wor d by word and press the space bar on a laptop screen to proceed. After reading each sentence, a yes -no comprehension task was conducted to assess participants' understanding of the sentence content. Prior to the experiment, participants underwent a practic e session to familiarize themselves with the task, and they were instructed to read the sentences at a natural speed as much as possible. 3.1.4. Results & discussion The accuracy rate for SR was 93%, while it was 92% for OR, and this difference was not statistically significant ( F s < 1). The statistical analysis for reading time was conducted exclusively on the trials where participants answered correctly in the comprehension task. Furthermore, to ensure data quality, any reading time data below 250 milliseconds (ms) and data exceeding the mean reading time by more than 3 standard deviation (SD) points were considered outliers and excluded from the analysis. 5 This procedure affected 2.2% of the total data. The reading times for each region (word) are as depicted in Figure 1.The regions of interest are Region 2, where the RC-verb is presented, and Region 4, where the RC-head is presented. 5 In the experiments measuring word-by-word reading times, participants may accidentally press the space bar before fully reading a word. Additionally, it is known that the reaction time for pressing the space bar is around 200 ms . Therefore, to press the space bar aft er recognizing the presented word, it would require at least 250 -300 ms. In the present study, reading times below 250 ms were excluded from the analysis, as they were considered to indicate that adequate comprehension had not taken place. Figure 1. Reading times for SR and OR In Region 1, where the accusative and genitive nouns are presented, and in Region 2, where the RC -verbs are presented, the length of the characters is always shorter in the SR condition. Therefore, to account for difference, residual reading times were calculated for each word based on the time required to read a single letter in these regions, following the approach by Ferreira and Clifton (1986). A within-subject analysis of variance (ANOVA) was conducted to examine the effects. The results showed that in Region 2, there was a significant effect only in the subject analysis [ F 1 (1,34) = 4.57, p < .05; F 2 (1,23) = 1.45, p < .2]. 6 There were no significant differences in other regions ( F s < 2). Overall, the results show that participants read SR verbs faster than OR verbs, and there were no significant differences in the reading times of other regions Discussion : The results of Experiment 3 supported Prediction 1, indicating that if native speakers of Turkish start constructing the dependency between the gap and filler before encountering the filler, then the proc essing asymmetry between SRs and ORs would be observed at the RC-verb position. This finding suggests that the processing of gap-filler dependencies is incremental in Turkish. However, it should be noted that in Experiment 1, more RCs were produced after the accusative noun compared to the genitive noun. This suggests that in Experiment 3, participants may not have predicted the occurrence of RCs after genitive nouns, leading to a potential surprisal effect at the OR verb position, which could have contributed to increased reading times (e.g., Levy, 2008). To confirm this possibility, the results of Experiment 1, particularly the findings related to the use of adverbs, provide valuable insights. Experiment 1 indicated that the 6 F 1 represents the F values obtained from subject analysis and F 2 from item analysis. presence of an adverb following a genitive noun increased the predictability of RCs compared to other conditions. If the results observed in Experiment 3 were primarily driven by a surprisal effect , then introducing adverbs after genitive noun phrases should mitigate or reverse the reading time difference between SR and OR verbs. Conversely, if the processing asymmetry between SRs and ORs observed at the RC-verbs reflects the processing load associated with constructing the gap-filler dependency , then the reading time patterns should remain consistent even when adverbs are used after genitive noun phrases. To examine which of these predictions holds true, Experiment 4 was conducted. 3.2. Experiment 4 3.2.1. Aim Experiment 4 aimed to investigate whether the observed reading time difference at th e RC-verb position in Experiment 3 can be attributed to the surprisal effect or the processing load of constructing the gap-filler dependency 3.2.2. Material In Experiment 4, four conditions were established, following a 2x2 design, as shown in Table 5. Table 5 Test sentences in Experiment 4 Region 1 2 3 4 5 6 7 8 (10a) ADV[-]SR: Yetenekli işçi-yi izle-yen ustabaşı fabrika-da hep gülümser-di. Skillful worker-acc watch-srp foreman factory-loc always smile-past The foreman who watched the skillful worker always smiled at the factory.' (10b) ADV[-]OR: Yetenekli işçi-nin izle-diğ-i ustabaşı fabrika-da hep gülümser-di. Skillful worker-gen watch-nsr-3s foreman factory-loc always smile-past The foreman who the skillful worker watched always smiled at the factory.' (10c) ADV[+]SR: Yetenekli işçi-yi dikkatle izle-yen ustabaşı fabrika-da hep gülümser-di. Skillful worker-acc carefully watch-srp foreman factory-loc always smile-past The foreman who carefully watched the skillful worker always smiled at the factory.' (10d) ADV[+]OR: Yetenekli işçi-nin dikkatle izle-diğ-i ustabaşı fabrika-da hep gülümser-di. Skillful worker-gen carefully watch-nsr-3s foreman factory-loc always smile-past The foreman who the skillful worker carefully watched always smiled at the factory.' In Experiment 4, the adjectives that were previously attached to the head nouns of RCs in Experiment 3 were now placed at the beginning of the sentences. This was done because the anticipated spillover effect was not observed in Experiment 3, and also to eliminate the possibility that the presence of the adjective influenced the lack of differences in reading times at the RC -head position. Additionally, in (10c) and (10d), adverbs were inserted immediately after the accusative and genitive nouns, respectively. As test sentences, a total of 160 sentences, arranged in 40 pairs, were prepared as shown in Table 5. These test sentences were presented using a Latin square design, dividing the 40 sets of sentences into four groups corresponding to each condition. Furthermore, 62 filler sentences were prepared, resulting in a total of 102 experimental sentences that were randomly presented to each participant. A norming study was conducted after the experiment to ensure the naturalness of the test sentences. As a result, significant differences were found among the four sets of test sentences, and thus these sentences were excluded from the final analysis. Prediction 1 : If the results of Experiment 3 were indeed driven by the surprisal effect due to the lack of prediction for RC after the genitive noun, then an interaction effect would be observed, as the reading time difference between SR and OR verb disappears or is reversed in the conditions where adverbs are used. Prediction 2 : If the processing asymmetry observed at the RC-verb position in Experiment 3 is primarily due to the processing load of constructing the gap-filler dependency , then even in the condition where adverbs are used, the reading time of SR verbs should still be shorte r than that of OR verbs, leading to a significant main effect of RC type. 3.2.3. Participants & procedure Fifty-two students from a university in Turkey participated in Experiment 4. All were native speakers of Turkish and their average age was 20.5 (SD=2.57). They did not participate in any experiments other than Experiment 4. The procedure was identical to Experiment 3. 3.2.4. Results & discussion The average accuracy rate for the comprehension task was 89%, and there were no significant differences among the four conditions ( F s < 1). The statistical analysis for reading time was conducted only on the trials where participants answered correctly in the comprehension task. Additionally, two participants and five items with an accuracy rate lower than 75% were excluded from further analyses. This was done because the two participants did not perform the task properly, and the five items were found to be more difficult to understand compared to the other items. Furthermore, as mentioned earlier, the four items that showed differences in naturalne