Bar ış Kahraman and Hiromu Sakai 14 Relative clause processing in Japanese: Psycholinguistic investigation into typological di ff erences 1 Introduction Relative clauses have been one of the most widely studied constructions since the pioneering study by Keenan and Comrie (1977), in the fi eld of language typology. In the fi eld of psycholinguistics, too, the processing of relative clauses has been one of the most extensively studied topics in the past decades. It is therefore natural for psycholinguists to bring typological perspectives into research on relative clause processing. That is, those who are investigating the processing of typologically dis- tinct languages such as English and Japanese are interested in the in fl uence of typological factors on language processing. Researchers initially focused on surface syntactic factors such as constituent order or structural complexities and then gradually shifted their attentions to various other factors, including semantic factors such as animacy of noun phrases or pragmatic factors such as the discourse functions of relative clauses. Importantly, this shift of attention was in line with development of theories on relative clause processing. Some theories attribute sources of processing di ffi culties to the process of fi ller-gap dependency formation . Other theories assume that learning through experience , namely statistical learning , is an essential aspect of human sentence processing that determines processing di ffi culties. In this chapter, we brie fl y review the history of Japanese relative clause processing research and dis- cuss what we have found and what is left for future research. We discuss why it is important to consider various types of typological factors. In quite a few numbers of languages in the world, it has been well-documented that there exist an asymmetry in the comprehension of subject relative clauses (SRCs) like (1a) and (2a) and object relative clauses (ORCs) like (1b) and (2b). (1) English: a. SRC: The student [ who saw the teacher ] was waiting for the bus. b. ORC: The student [ who the teacher saw ] was waiting for the bus (2) Japanese: a. SRC: [ Sensei o mita ] gakusei wa basu o matteita. teacher ACC saw student TOP bus ACC was.waiting ‘ The student who saw the teacher was waiting for the bus. ’ b. ORC: [ Sensei ga mita ] gakusei wa basu o matteita. teacher NOM saw student TOP bus ACC was.waiting ‘ The student who the teacher saw was waiting for the bus. ’ 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 423) SRCs are generally easier to process than ORCs (e.g., Chinese: Lin 2006; Dutch: Mak, Vonk and Schfriers 2006; English: Staub 2010; French: Cohen and Mehler 1996; German: Schriefers, Friederici and Kühn 1995; Hungarian: MacWhinney and Pleh 1988; Japanese: Ueno and Garnsey 2008; Korean: Kwon et al. 2013; Spanish: Betancort, Carreiras and Sturt 2009; Turkish: Kahraman et al. 2010). 1 It should be emphasized that relative clauses of languages in the world show both typological similarities and di ff erences. For example, even though relative clauses are used to ‘ modify ’ the content of head-noun in a broad sense in all of the languages mentioned above, they ful fi ll such function in quite di ff erent ways. In a head-initial language like English, the head-noun ( the teacher ) comes before the relative clause as in (1). In a head- fi nal language like Japanese, on the other hand, the relative clause comes before the head-noun ( gakusei ) as in (2). Such cross- linguistic di ff erences are important not only from typological perspective, but also from the psycholinguistic perspective. Moreover, a relative pronoun (e.g., who ) unambiguously signals the existence of a relative clause in English. Japanese, however, does not have a relative pronoun or a relative clause marker. Therefore, Japanese relative clauses are temporarily ambiguous between relative clauses and other types of subordinate/matrix clauses until the head-noun is encountered. English speaking readers/listeners therefore understand that they are reading/hearing a relative clause when they encounter a relative pronoun, whereas Japanese speak- ing readers/listeners cannot realize the existence of a relative clause until they encounter a head-noun. The exploration of the source of processing asymmetry between SRCs and ORCs in typologically di ff erent languages, thus provides impor- tant insights into characterization of human cognitive mechanisms for language processing by revealing their universal and language-speci fi c aspects. In order to explain the processing asymmetry between SRCs and ORCs, researchers have proposed various accounts (see also Sakamoto, this volume). The recent accounts can roughly be classi fi ed into two groups according to the theories of sentence process- ing they are based on. One is based on the theories of fi ller-gap dependency formation. They mainly examine the in fl uence on relative clause processing by surface syntactic factors, such as constituent order or structural complexity. 2 The other is based on 1 Actually, a few studies reported ORC advantage over SRC in languages like Chinese (Hsiao and Gibson 2003; Lin 2014) and Basque (Carreiras et al. 2010), but in general, SRCs are easier than ORCs. 2 There is also another complexity based account which is called Similarity Interference Hypothesis This account basically assumes that the processing di ffi culty of ORCs stems from the repetition of same types of noun phrases in English (Gordon, Hendrick and Johnson 2001). In SRCs, the thematic roles can be easily assigned because no extra noun phrase intervenes between the relative clause verb and the head-noun. In ORCs, on the other hand, two noun phrases should be held in the memory until the thematic roles are assigned. In Japanese, thematic roles are clearly marked by case markers and semantic ambiguities do not arise except in limited cases. Therefore, we will not discuss the Similarity Interference Hypothesis in this chapter. See also Nakayama, Vasishth and Lewis (2006) on similarity interference in Japanese. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 424) 424 Bar ış Kahraman and Hiromu Sakai various kinds of factors learned through experience, such as; predictability of up- coming elements, frequency of structures, animacy of noun phrases, and discourse functions of relative clauses. We will discuss these accounts in detail. However, since the great majority of the accounts have been proposed based on the data from rela- tive clause processing in English and other European languages, it is quite di ffi cult to evaluate the validity of these accounts by just looking at data from European lan- guages. In this context, cross-linguistic investigation of relative clause processing, especially in languages with typological di ff erences such as Japanese, gains more importance because it allows us to distinguish the validity and the universality of these accounts, and construct a more accurate model of human sentence processing mechanisms. In what follows, we will fi rst discuss accounts based on the theories on fi ller-gap dependency formation, and speci fi cally test the validity of Dependency Locality Theory (Gibson 1998, 2000) and Structural Distance Hypothesis (O ’ Grady 1997) in Japanese. Then, in section 3, we will review accounts based on other factors and the theories on statistical learning through experience in Japanese. Finally, in section 4, we will discuss the limitations and possibilities of future directions of relative clause processing studies in Japanese. 2 Relative clause processing and fi ller-gap dependency First we will examine accounts for relative clause processing based on two represen- tative theories on fi ller-gap dependency formation: the Dependency Locality Theory and the Structural Distance Hypothesis. We will show that both of these theories make the same prediction for the processing di ffi culty of relative clauses in English, whereas these theories make a di ff erent prediction in Japanese. We will then brie fl y introduce studies that attempt to verify these predictions and that examine the in fl u- ence of syntactic factors on relative clause processing in Japanese. Our focus is on Japanese native speakers ’ relative clause processing in this chapter, but those who are interested in the evaluation of the two accounts above in L2 Japanese processing should be referred to Sawasaki and Kashiwagi-Wood ’ s chapter in this volume. 2.1 Dependency Locality Theory and Structural Distance Hypothesis The concept of fi ller-gap dependency plays an important role in the sentence process- ing literature (see also Sakamoto ’ s chapter in this volume). An argument of a verb can be displaced from its original position, appearing in another position in the 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 425) Relative clause processing in Japanese 425 sentence, e.g., WH-questions in English. In the fi eld of sentence processing, a dis- placed element is called a fi ller , and in its original position is a gap . A fi ller-gap dependency is found in many di ff erent constructions, such as clefts, relative clauses, topic sentences, WH-questions, and so on. In order to understand the meaning of these kinds of sentences, listeners/readers set up an association between the fi ller and the gap. This refers to the fi ller-gap dependency (e.g., Fodor 1989). For instance, student is the fi ller, and the underline shows the gap position in English examples below. 3 ! (3) a. SRC: The student i who ___ i saw the teacher . . . b. ORC: The student i who the teacher saw ___ i . . . ! According to the Dependency Locality Theory, the number of discourse referents (words) between the fi ller and the gap is the main source of the processing di ffi culty of relative clauses (Gibson 1998, 2000). This type of account is also conventionally called Linear Distance Hypothesis (Ueno and Garnsey 2008). Gibson argues that the listeners/readers need to hold the fi ller in working memory and retrieve it at the gap position. SRCs are easier to comprehend than ORCs because the number of discourse referents intervening between the fi ller and the gap are fewer in SRCs than ORCs (see Sawasaki and Kashiwagi-Wood ’ s chapter). Since the working memory load is heavier for ORCs, the fi ller is harder to retrieve in ORCs compared to SRCs at the relative clause verb saw in (3). On the other hand, the Structural Distance Hypothesis (O ’ Grady 1997; Hawkins 1999, 2004) considers that the number of the syntactic nodes between the fi ller and the gap, or the embedding depth of the gap, determines the processing load. Accord- ing to O ’ Grady (1997: 136), structural complexity increases by the number of syntactic nodes between the fi ller and the gap, and this causes a processing disadvantage. As illustrated in Figure 1, there are three nodes between the fi ller and the gap in SRCs, whereas there are four nodes in ORCs. Since the computational complexity is heavier in ORCs compared to SRCs, this hypothesis predicts that SRCs are easier to comprehend than ORCs. Both accounts predict that SRCs should be easier to process than ORCs, and they are indistinguishable in the head-initial languages like English. However, in a head- fi nal language like Japanese, these two theories make di ff erent predictions for the processing di ffi culty of SRCs and ORCs. The Dependency Locality Theory predicts that ORCs should be easier to process than SRCs, because the memory load is heavier for SRCs, i.e., more intervening words between the fi ller and the gap in SRCs than ORCs. 3 Some researchers accept a relative pronoun as the fi ller in English. In many languages, no relative pronoun is used but fi llers do exist. In this chapter, we assume the head-noun as the fi ller. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 426) 426 Bar ış Kahraman and Hiromu Sakai (4) a. SRC: [___ i Sensei o mita ] gakusei i wa. . . ! teacher ACC saw student TOP ‘ The student who saw the teacher. . . ’ b. ORC: [ Sensei ga ___ i mita ] gakusei i wa. . . ! teacher NOM saw student TOP ‘ The student who the teacher saw. . . ’ On the other hand, the number of syntactic nodes between the fi ller and the gap is fewer in SRCs than ORCs as illustrated in Figure 2. Therefore, unlike the Depen- dency Locality Theory, the Structural Distance Hypothesis predicts that SRCs should be easier to process than ORCs because SRCs are less structurally complex. Figure 1: Structural distance of subject and object relative clauses in English Figure 2: Structural distance of subject and object relative clauses in Japanese 4 4 Some researchers add a CP node over the IP node (e.g. Ishizuka 2005). In this chapter, it is not a crucial issue. We, thus, based on Murasugi (2000), assume that SRCs and ORCs are IP, and did not add a CP node. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 427) Relative clause processing in Japanese 427 In sum, these two theories make di ff erent predictions in Japanese, and this allows researchers to distinguish between competing hypotheses, which was impos- sible in English. 2.2 Relative clause processing in Japanese from the fi ller-gap dependency formation perspective The theories of fi ller-gap dependency formation make di ff erent predictions about Japanese relative clause processing, and researchers previously examined the validity of these theories (e.g., Miyamoto and Nakamura 2003; Nakamura 2003; Ishizuka 2005; Ueno and Garnsey 2008; Sakamoto and Yasunaga 2009; Mitsugi, MacWhinney and Shirai 2010; Kahraman 2012). For example, Kahraman (2012) used test sentences as shown in (5), and compared the processing di ffi culty of SRCs and ORCs through a self-paced reading task. He compared the mean reading times of each word in the SRC and ORC conditions. (5) a. SRC: Depaato de ryoosin o sagasiteita kodomo wa department store LOC parents ACC looking.for child TOP kyuuni nakidasita. suddenly cried ‘ At the department store, the child who was looking for his parents suddenly began crying. ’ b. ORC: Depaato de ryoosin ga sagasiteita kodomo wa department store LOC parents NOM looking.for child TOP kyuuni nakidasita. suddenly cried ‘ At the department store, the child who the parents were looking for started crying. ’ In both conditions, the test sentences started with a locative adverb. In the SRC con- dition, an accusative marked noun followed the adverb, and in the ORC condition, a nominative marked noun followed the adverb. The rest of the sentences were identi- cal in the two conditions, in which a relative clause verb, a head-noun, an adverb and a matrix verb appeared in sequence. If the linear distance was a more important factor, the ORCs should be easier to process than SRCs at the head noun position, whereas the SRCs should be easier to process than the ORCs if the structural dis- tance is a more important factor. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 428) 428 Bar ış Kahraman and Hiromu Sakai The results of the self-paced reading experiment showed that the reading times of the SRC ’ s head-noun ( kodomo-wa ) were faster than that of the ORC, as shown in Figure 3. Kahraman concluded that SRCs are easier to process than ORCs in Japanese, and the Structural Distance Hypothesis can account for the processing di ffi culty of ORCs (O ’ Grady 1997, Hawkins 1999, 2004), whereas the Dependency Locality Theory cannot (Gibson 1998, 2000). In Japanese, various studies on relative clauses using di ff erent materials consis- tently reported that SRCs are easier to process than ORCs at the head-noun region (Ishizuka 2005; Kahraman et al. 2011a; Mitsugi, MacWhinney and Shirai 2010; Miya- moto and Nakamura 2003; Nakamura 2003; Sakamoto and Yasunaga 2009; Ueno and Garnsey 2008). For example, Miyamoto and Nakamura (2003) inserted an adver- bial phrase between the relative clause verb and the sentence initial noun, and they manipulated the case marking of the head-noun of relative clause. The result of self- paced reading experiment showed that the head-noun of SRCs was read faster than that of ORCs, irrespective of the case marking of the head-noun. Based on these result, Miyamoto and Nakamura argued that the results cannot be explained by the Dependency Locality Theory. Ueno and Garnsey (2008) compared the processing di ffi culty of SRCs and ORCs through an experiment with a self-paced reading task and one with event related brain potentials (ERP). In both experiments, Ueno and Garnsey reported that the SRC were easier to process than the ORC at the relative Figure 3: Reading times of SRCs and ORCs in Kahraman (2012) (The circle and the arrow show where the signi fi cant processing asymmetry was observed) 5 5 In the gloss, case markers are separately listed as independent morphemes, but in the experi- ments, case markers were presented with content words since they form phrases called bunsetsu Therefore, we presented a content word and a case marker within the same region in the fi gures. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 429) Relative clause processing in Japanese 429 clause head noun position and they pointed out that the Structural Distance Hypo- thesis could successfully capture the processing asymmetry between the SRC and the ORC in Japanese. Unlike the above mentioned studies, Sakamoto and Yasunaga (2009) conducted a self-paced reading experiment using relative clause verbs which took dative case marked nouns as their arguments, such as hanasu ‘ talk ’ , au ‘ met ’ , menkaisuru ‘ met ’ , etc. as in (6). (6) a. SRC: Kisya ni menkaisita giin wa kenryoku ga reporter DAT met senator TOP power NOM subete da to omotteita. everything COP that was.thinking ‘ The senator who met the reporter was thinking that the power is everything . ’ b. ORC: Kisya ga menkaisita giin wa kenryoku ga reporter DAT met senator TOP power NOM subete da to omotteita. everything COP that was.thinking ‘ The senator who the reporter met was thinking that the power is everything. ’ Sakamoto and Yasunaga (2009) also found that SRCs were easier to process than ORCs, and concluded that the Structural Distance Hypothesis accurately captured the processing di ffi culty of relative clauses in Japanese. In addition to these studies, Kahraman (2012) and Mitsugi, MacWhinney and Shirai (2010) also compared the processing of SRCs and ORC by L2 learners of Japanese through self-paced reading experiments. The results of these studies showed that not only the native speakers but also Korean and Turkish speaking L2 learners of Japanese read SRCs faster than ORCs (see Sawasaki and Kashiwagi-Wood ’ s chapter for detailed discussion of L2 relative clause processing in Japanese). Overall, these studies show that SRCs are generally easier to process than ORCs in Japanese. It means that the Structural Dis- tance Hypothesis makes a valid prediction for Japanese relative clause processing but the Dependency Locality Theory cannot. So far, we have shown that the studies on Japanese relative clause processing made a very important contribution to the fi eld of sentence processing. However, these studies have mainly examined surface syntactic factors such as linear distance or structural distance and left many other potentially important factors unexamined. Accounts based on the theories of fi ller-gap dependency formation are valid only if other factors that potentially in fl uence the processing load were controlled. Notice, however, that SRCs and ORCs which were used as experimental material in these 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 430) 430 Bar ış Kahraman and Hiromu Sakai studies, repeated below, contain quite a few di ff erences other than linear or struc- tural distance between the fi ller and the gap. (5) a. SRC: Depaato de ryoosin o sagasiteita kodomo wa department store LOC parents ACC looking.for child TOP kyuuni nakidasita. suddenly cried ‘ At the department store, the child who was looking for his parents suddenly began crying. ’ b. ORC: Depaato de ryoosin ga sagasiteita kodomo wa department store LOC parents NOM looking.for child TOP kyuuni nakidasita. suddenly cried ‘ At the department store, the child who the parents were looking for started crying. ’ Excluding the sentence initial adverbs or adjectives, the SRC and the ORC start from noun phrases with di ff erent case markers, a noun phrase with an accusative case marker in SRC and a noun phrase with a nominative case marker in ORC. It is well know that case markers play a very important role not only in determining grammaticality or semantic interpretation but also in processing Japanese sentences. Previous studies on Japanese sentence processing showed that readers use the case marker information very quickly and e ff ectively to make predictions about the argu- ment structure even before the verb appears (e.g., Kamide, Altmann and Haywood 2003; Miyamoto 2002; Yamashita 1997). It is conceivable that these case markers would elicit di ff erent predictions about the upcoming elements and/or structure. This kind of case marking di ff erence and prediction mechanisms should be taken into consideration in Japanese relative clause processing studies. See Sakamoto ’ s chapter on expectancy driven processing. Second, previous studies have shown that the frequency of occurrence of SRCs and ORCs in corpora correlates with ease of relative clause processing in English (Reali and Christiansen 2007). The frequency of SRCs and ORCs might also be di ff er- ent and might have caused the processing asymmetry in Japanese. This possibility should be taken into consideration and examined in Japanese, as well. Third, from the point of view of animacy of noun phrases, the SRC in (5a) and the ORC in (5b) are di ff erent. It is well-known that the prototypical subject is animate whereas the prototypical object is inanimate (Comrie 1989). As a consequence, as we will show in the next section, there are much more instances of SRCs with an animate head-noun compared to those with an inanimate head-noun in corpora 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 431) Relative clause processing in Japanese 431 (Sato 2011). On the contrary, there are more instances of ORCs with an inanimate head-noun compared to those with an animate head-noun. Since the head-noun is an animate noun phrase in (5), they are more expected in the SRC but less expected in the ORC. This might cause the di ff erence in the processing. Fourth, most of previous studies did not provide any discourse context in their experiments. The result of such experiments can be compared only if discourse con- texts in fl uence processing of SRC and ORC in the same way. There are studies, how- ever, that point out the importance of discourse context in the processing of relative clauses (Mak, Vonk and Schfriers 2008; Roland et al. 2012). As it will be explained in detail in section 3, relative clauses play the important discourse function of introduc- ing the referent into the discourse. For instance the ORC in (5b) sounds more natural if the preceding context introduces two reporters, one of them was criticized by the senator but the other was not. This indicates that not only syntactic or semantic factors, but also pragmatic factors should be taken into consideration. 6 Finally, as we have pointed out in the introduction, di ff erences between SRCs and ORCs are observed in the relative clause region in English but in the head- noun region in Japanese. It is true that, in either English or Japanese, relative clause processing involves some kind of association between the fi ller and the gap. They are, however, quite di ff erent in the direction of association; the fi ller comes fi rst in English but the gap comes fi rst in Japanese. See Kahraman et al. (2010), Kahraman (2011), Kwon et al. (2013) and Lin (2006) for detailed discussion. This might also be responsible for similarities and di ff erences in the processing of English and Japanese relative clauses. In summary, most of the above mentioned relative clause processing studies in Japanese did not pay enough attention to these various factors that might potentially in fl uence relative clause processing. This is because most studies were conducted to test the validity of accounts based on theories on fi ller-gap dependency formation. In contrast, more and more researchers have recently begun to pay more attention to factors such as (un)predictability, frequency, animacy, discourse etc., which are closely related to statistical learning through experience. In the next section we will introduce studies examining such factors. 3 Relative clause processing and statistical learning This section brie fl y introduces theories of human sentence processing based on theories that emphasize the role of statistical learning. They share the view that lan- guage processing mechanisms are shaped by learning through experience and that 6 Even young English speaking children are sensitive to the presupposed contexts for relative clauses as shown in Hamburger and Crain (1982). 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 432) 432 Bar ış Kahraman and Hiromu Sakai native speakers form probabilistic expectation for up-coming elements while they process sentences (Gennari and MacDonald 2009; MacDonald and Christiansen 2002; Wells et al. 2009). We fi rst introduce theories that regard frequency or (un)pre- dictability of certain structures or linguistic elements as responsible for determining processing di ffi culty. We then turn to theories that incorporate other types of factors, semantic factors such as animacy, or pragmatic factors such as discourse function. Then we will turn to recent studies in Japanese, which have attempted to examine the in fl uence of various kinds of factors, such as case-marker driven expectation, frequency of occurrence, animacy, and discourse functions, in relative clause proc- essing in Japanese. 3.1 Frequency of occurrence and (un)predictability for up-coming elements Reali and Christiansen (2007) ’ s Frequency of Occurrence Hypothesis basically assumes that frequency of exposure to certain structures in fl uences the di ffi culty of sentence processing. According to this hypothesis, highly frequent structures are processed more easily because readers/listeners are more familiar with frequent structures. Reali and Christiansen conducted a corpus study, and compared the frequencies of relative clauses in English. They found that SRCs have a higher frequency than ORCs. Then, they compared the noun types within relative clauses, and found that proper nouns are used more frequently in SRCs than ORCs, whereas pronouns are used more frequently in ORCs than SRCs. In other words, when the nouns within relative clauses are proper nouns, the frequency of a that-verb-noun chunk, like that saw the teacher was higher than a that-noun-verb chunk, like that the teacher saw . On the other hand, when pronouns were used within relative clauses, the frequency of a that-pronoun-verb chunk, like that I/you/s(he) saw was higher than a that-verb- pronoun chunk, like that saw me/you/him/her . Based on these fi ndings, Reali and Christiansen manipulated the noun types within subject and object relative clauses, and conducted a series of self-paced reading experiments. The results showed that ORCs were read faster than SRCs when pronouns were used within relative clauses. Reali and Christiansen argued that co-occurrences of words, namely, the word- chunks play a central role in the sentence comprehension. Recently, various theories have been proposed about the role of (un)predictability in human sentence processing, such as the Surprisal Hypothesis (Hale 2001; Levy 2008), and the Entropy Reduction Hypothesis (Hale 2006). Although these theories use slightly di ff erent metrics to estimate the (un)predictability of upcoming words or structures, the main logic is that unpredictability for upcoming elements deter- mines the di ffi culty of sentence processing. According to the Surprisal Hypothesis (Hale 2001; Levy 2008), if the predictability of certain linguistic elements is lower than that of other elements, those elements are more di ffi cult to process compared 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 433) Relative clause processing in Japanese 433 to elements with higher predictability (Hale 2001; Levy 2008). Alternatively, Hale (2006) argues that reduction of uncertainty for up-coming elements increases the ease of sentence processing. Hale used a probabilistic context-free grammar and calculated the tree-bank probabilities of derivations for subject and object relative clauses, and found that probabilities of SRCs were higher than that of ORCs at rela- tive pronoun who . This indicates that the uncertainty for up-coming elements was reduced in SRCs compared to ORCs at the relative pronoun position in English. In other words, English speakers expect more SRCs than ORCs at the relative pronoun positions. These theories, thus, attribute the processing di ffi culty of ORCs to their less predictable nature compared to SRCs. Up to this point, we introduced accounts for relative clause processing based on the theories of frequency of occurrence or (un)predictability for up-coming elements. These accounts mainly focused on probabilistic distribution of linguistic elements classi fi ed with their lexical or syntactic properties. However, other researchers found that di ff erences between SRCs and ORCs are not restricted to their lexical or syntactic properties. They proposed that semantic properties such as animacy of noun phrases or pragmatic properties such as discourse functions play important roles in relative clause processing. We will turn to those accounts in the next sub-section. 3.2 Semantic indeterminacy and expectation from discourse function According to Semantic Indeterminacy Hypothesis , distributional patterns of animacy of noun phrases play a crucial role in forming probabilistic expectations in relative clause processing (Gennari and MacDonald 2008). Previous studies, which showed that SRCs were easier to process than ORCs, generally used animate nouns for the head-noun of relative clauses. However, various studies showed that when the head-noun of ORCs was an inanimate noun, ORCs were easier to process than SRCs (e.g., Mak, Vonk and Schfriers 2002, 2006; Traxler et al. 2005; Gennari and MacDonald 2008). Gennari and MacDonald (2008) pointed out that when the head- noun of relative clauses is an animate noun in an SRC, like the reporter that attacked the senator admitted the error , the thematic role of the head-noun ( the reporter ) is agent both for the relative clause verb ( attacked ) and the matrix verb ( admitted ). On the other hand, in the case of an ORC, like the reporter that the senator attacked admitted the error , the head-noun is patient of the relative clause verb and agent of the matrix verb. Gennari and MacDonald argued that the processing di ffi culty of ORCs may stem from the activation of several possible competing structures derived from the distributional pattern of thematic roles of given head-nouns. Moreover, they argued that ORCs with di ff erent animacy con fi gurations may involve di ff erent com- petition processes between structural and semantic analyses, and this may cause a di ff erent amount of indeterminacy and processing di ffi culty. For example, Gennari 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 434) 434 Bar ış Kahraman and Hiromu Sakai and MacDonald conducted a series of sentence completion experiments and found that when the head-noun of ORCs was an inanimate noun, it was most likely to be interpreted as the theme of the relative clause verb and agent of the matrix verb. On the other hand, when the head-noun of ORCs was an animate noun, it is likely to be interpreted as the agent, experiencer, patient or goal of the relative clause verb and theme of the matrix verb. These results suggest that when the head-noun of ORCs is an animate noun in English, there is a greater semantic indeterminacy for that noun, compared to inanimate noun. Furthermore, Gennari and MacDonald, based on these fi ndings, manipulated the head-noun animacy (animate vs. inanimate) and voice (active vs. passive) of ORCs, and conducted a self-paced reading experiment. The results showed that ORCs with inanimate head-nouns were read faster than ORCs with animate nouns. In other words, when the head-noun of ORCs ’ was inanimate, their processing was not as hard as that of ORCs with animate head-nouns. Based on these results, Gennari and MacDonald argued that processing di ffi culty of ORCs found in previous studies stems from the competition processes between structural and semantic analysis, which is closely related to distributional pattern of the head-noun animacy of relative clauses, and proposed Semantic Indeterminacy Hypothesis Fox and Thompson (1990) conducted a corpus study in English, and found that ORCs are generally used when the head-nouns are integrated (or ‘ grounded ’ ) into the ongoing discourse using old discourse referents. In Dutch, Mak, Vonk and Schfriers (2008) showed that when the noun phrase within relative clauses refers to the discourse topic in the previous context, the processing di ffi culty of ORCs is reduced. Based on the fi ndings of Fox and Thompson (1990) and Mak, Vonk and Schfriers (2008), Roland et al. (2012) argued SRCs and ORCs are used for di ff erent purposes in real language. In order to extend and re-examine the fi ndings of Fox and Thomson (1990), Roland et al. used various corpora and conducted a large scale corpus analysis in English. The results showed that an ORC, like the student that the teacher saw . . . , is more likely to appear in contexts in which the noun phrase within the relative clause has already been introduced as an old referent, like the teacher was walking down the campus . On the other hand, Roland et al. found that an SRC, like the student saw the teacher . . .is not likely to appear after such contexts. From these fi ndings, Roland et al. pointed out that a noun phrase within an ORC is tends to be the topic of the ongoing discourse, whereas a noun phrase within an SRCs is not explicitly mentioned in the previous context, as argued in Fox and Thompson (1990). Roland et al. argued that, in terms of givenness of the noun phrases within relative clauses, a noun phrase within an ORC is an old discourse referent and a noun phrase within an SRC is a new referent. They then concluded that ORCs are generally used for grounding modi fi ed nouns to the ongoing discourse by using discourse old referent, whereas SRCs are used for supplying additional information about the modi fi ed noun by using new discourse referent. These observations suggest that SRCs and ORCs are used in di ff erent contexts for di ff erent purposes, 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 (Unicode 9 8/12/14 16:37) WDG-New (170mm 240mm) DGMetaSerifScience (OpenType) 1605 Nakayama pp. 423–456 1605 Nakayama_14_Kahraman (p. 435) Relative clause processing in Japanese 435 and ORCs are more dependent on the context than SRCs because ORCs are used in more speci fi c situations (Fox and Thompson 1990). Furthermore, Roland et al. argued that ORCs might have been more di ffi cult to comprehend than SRCs because most of the previous studies examined the process- ing di ffi culty of these structures without any context. In other words, the lack of appropriate context might have violated the discourse requirements for ORCs, and this kind of unnaturalness might have caused the processing di ffi culty of ORCs. Moreover they pointed out that if the discourse requirements of ORCs are satis fi ed by an appropriate context and the unnaturalness is eliminated; the processing di ffi - culty of ORCs might be reduced. In order to examine this hypothesis and provide empirical evidence, Roland et al. manipulated the contexts before the relative clauses as shown below, and compared the reading times of SRCs and ORCs through a series of self-paced reading experiments. (7) a. Neutral context: There is always something happening in Elmwood Village. b. Topic context: The sculptor collected paintings. (8) a. SRC: The artist that admired the sculptor exhibited portraits at the gallery on Elmwood Avenue. b. ORC: The artist that the sculptor admired exhibited portraits at the gallery on Elmwood Avenue. In the Neutral context condition (7a), there is no particular topic, and no noun phrase within the relative clauses is explicitly mentioned. On the other hand, in the Topic context condition (7b), the embedded noun phrase within relative clauses, namely the sculpture , is subject and topic of the discourse. In a self-paced reading experiment, Roland et al. presented both SRCs and ORCs after the two types of contexts. The results showed that SRCs were read faster than ORCs in the Neutral context condition, as already shown in many previous studies in English (e.g., Staub 2010; see referents therein). In the Topic context condition, on the other hand, the reading times of SRCs and ORCs did not di ff er signi fi cantly. In other words, the processing di ffi culty of ORCs was reduced after the context like (7b), as in Dutch (Mak, Vonk and Schfriers 2008). Roland et al. argued that the results suggest that when the noun phrase within relative clauses is not mentioned in the previous con- text