Contextual Effects and Locality Preferences in Relative Clause Attachment in Thai Teeranoot Siriwittayakorn (steeranoot@gmail.com) Department of Linguistics, Faculty of Arts, Chulalongkorn University Phayathai Road, Pathumwan, Bangkok, 10330 Thailand Edson T. Miyamoto (miyamoto@alum.mit.edu) University of Tsukuba, Graduate School of Humanities and Social Sciences Tsukuba, Ibaraki, 305-8571 Japan Theeraporn Ratitamkul (Theeraporn.R@chula.ac.th) Department of Linguistics, Faculty of Arts, Chulalongkorn University Phayathai Road, Pathumwan, Bangkok, 10330 Thailand Abstract attachment. For example, based on world knowledge, the RC is likely to be attached to N1 (friend) in there was a Since the early 1990s, there has been a debate on the universality of locality in sentence processing (i.e., the wake for the friend of the teacher who died yesterday. preference to associate a word or phrase to the closest In this paper, we suggest that locality preferences can be possible word). Studies across various languages have obscured by contextual factors. Therefore, we report results investigated ambiguous relative clauses that can be attached factoring out preferences stemming from the intended to either of two nouns to determine the types of languages in meaning of the context surrounding the RC. A corpus count which locality is violated. We report a corpus count and a and a questionnaire provide evidence for contextual effects questionnaire in Thai indicating that intra-sentential contexts can obscure locality. Two reading experiments controlling for in RC attachment in Thai. Moreover, two reading context are also reported in support of locality in Thai. The experiments support the claim that locality is obeyed in Thai finding that context distorted locality raises the possibility when contextual effects are kept under control. that previous reports of locality violations in various languages may be reduced to contextual effects. Thai and its Previous Results Keywords: relative clause attachment; locality; intra- Thai is an SVO (subject-verb-object) language. The word sentential context; Thai order of the target construction is N1 of N2 RC. There are no plural markers or morphological agreement, thus ambiguity Introduction resolution is often based on plausibility. As sentences are read word by word, there is a preference to Although previous results in Thai provided support for a attach a new word (or phrase) to the closet possible word non-local N1 attachment preference (Siriwittayakorn, (locality, for short; Gibson, 1998, for a summary). For Miyamoto, Ratitamkul, & Cho, 2014), there were some example, in (1) the relative clause (RC) can be attached to potential confounds. The first, which is the main concern of either the non-local noun (N1, friend) or the local noun (the the present paper, was that the complex NP came after the noun closest to the RC, i.e., N2, teacher), and English matrix verb (in object position) in the items of a self-paced readers favor N2. reading experiment. This may have led readers to attach the (1) John met the friend of the teacher who was in RC to N1 to make it coherent with the matrix clause Germany (Cuetos & Mitchell, 1988). (Rohde, Levy, & Kehler, 2011). Another concern was that Prima facie evidence against locality comes from reports sentential complements (e.g., literally: “decision of suggesting that N1 is favored in the equivalent of (1) in committee that will extend deadline”) were incorrectly various languages (e.g., Dutch, French, Greek, Italian, classified as RC instances, often as N1 attachment, in a Japanese, Spanish; see Grillo & Costa, 2014, for a corpus count (see Kullavanijaya, 2010, for ways to summary). differentiate RCs and sentential complements in Thai). One problem with previous studies is that they often discussed RC attachment as a purely syntactic phenomenon. Corpus Count However, the surrounding context can affect the intended Since coherence is important in writing (Trabasso, Suh, & meaning of the RC as readers expect clauses to be coherent Payton, 1995), it would not be surprising if RCs are and thus prefer attaching the RC to N1 if it provides a produced and attached according to the surrounding context. reason or justification for the statement in the matrix clause Moreover, saliency as dictated by the animacy and (Rohde, Levy, & Kehler, 2011). concreteness of N1 and N2, has been claimed to affect Text coherence may be achieved in other ways. Even attachment (Desmet, De Baecke, Drieghe, Brysbaert, & when surrounding context does not require a causal Vonk, 2006) and may interact with coherence (e.g., more explanation or justification, it may still affect RC salient nouns may lead to stronger coherence requirements). 686 Thus, we report a corpus count investigating effects of reported; see Desmet, De Baecke, & Brysbaert, 2002, on context, animacy and concreteness on RC attachment. inter- and intra-sentential contexts in RC-attachment). Two native Thai speakers coded all instances Method independently. Disagreements (5.33%) were settled after A total of 4,800 instances of khɔ̌ŋ “of” followed by thî: discussion with a third native Thai speaker. “that” with up to three intervening words were randomly selected (Siriwittyakorn et al., 2014, for details) from the six Results writing genres of the Thai National Corpus (approximately Overall (i.e., for internally- and externally- disambiguated 32 million words; genres: fiction, newspaper, academic text, instances), there was no reliable difference in attachment non-academic text, law and miscellanea; Aroonmanakun, (N1 attachment: 460, 47.28%; N2 attachment: 513, 52.72%; Tansiri, & Nittayanuparp, 2009). Irrelevant instances were where clear, only numbers for N1 are reported from here eliminated (e.g., 145 were sentential complements following on). However, when restricted to internally-disambiguated Kullavanijaya, 2010). instances (i.e., when attachment did not depend on the There were 2,109 instances of the target construction (N1 surrounding context), the bias towards N2 attachment was of N2 RC). Instances were further eliminated if RC reliable (N1: 45.11%; χ2 (1) = 8.12, p = .004; see Table 1). attachment was ambiguous (353 instances, 16.74%); if one For the externally-disambiguated instances, N1 of the head nouns was a pronoun, a proper name or a biasing attachment was more frequent than N2 attachment (N1: 67 noun (e.g., khōn ‘person’, sìŋ ‘thing’; 769 instances, instances, 67.00%; χ2 (1) = 10.9, p = .001). The interaction 36.46%); or if they were repetitions (14 instances, 0.66%). between attachment (N1 or N2) and point of disambiguation We report the results for the remaining 973 instances (internal or external) was also reliable (χ2 (1) = 16.37, p < coded according to attachment (N1 or N2) and lexical .001). features of N1 and N2 (animacy: animate or inanimate; and For animacy and concreteness, results are reported for the concreteness: concrete or abstract; e.g., animate-concrete: internally-disambiguated instances (trends are the same man, animate-abstract: government, inanimate-concrete: when externally-disambiguated instances are included; there house, inanimate-abstract: goodness). was no interaction between animacy-concreteness and More crucially, because we were interested in the context; all ps >.10). The trends in Table 1 replicate the influence of the surrounding context, instances were also effects of animacy and concreteness in Dutch (Desmet et al., classified according to disambiguating point. If attachment 2006). For example, animate N1 attracted RCs when both was resolved within the complex NP (i.e., N1 of N2 RC), it nouns were animate-abstract (N1: 92.11%, p < .001). was coded as internally-disambiguated (e.g., “voice of man Concrete nouns also attracted RCs (e.g., RCs were more that was uttered”). If context surrounding the complex NP frequently attached to N1 when it was inanimate-concrete was needed to determine attachment, it was coded as and N2 was animate-abstract; N1: 77.78%, p = .006). externally-disambiguated (e.g., in “The writer used only Moreover, like in Dutch, there were few instances when words that have beautiful sound to create rhyme of word both nouns were animate-concrete (for Thai, N1: 9; N2: 12). that arouses listeners’ emotion,” “rhyme” is more likely to However, numerical trends were in the opposite direction in be associated with the “beautiful sound” mentioned in the Dutch (out of 1,065 instances, N1: 19; N2: 10), but a direct matrix clause, making it more likely to arouse listeners’ comparison with our results is difficult, since the Dutch emotion and, hence, more likely to be modified by the RC). counts did not differentiate between internally- and Externally-disambiguated instances only involved intra- externally-disambiguated RCs. sentential contexts (the matrix clause, subordinate clauses; inclusion of the 4 instances where an adjacent sentence Discussion determined attachment, 0.41%, did not change the trends Context was found to be a factor that can obscure the N2- Table 1: Attachment distribution in internally-disambiguated tokens (each cell indicates the number of N1 and the number of N2 attachments; *: p < .05; +: p < .10 according to exact binomial tests) Type of N2 Type of N1 animate inanimate Total concrete abstract concrete abstract animate concrete 9-12 11-11 4-4 0-1 24-28 abstract 1-5 35-3* 0-2 0-0 36-10* inanimate concrete 41-46 21-6* 27-47* 9-3 98-102 abstract 76-100+ 64-46 49-122* 45-69* 234-337* Total 127-163* 131-66* 80-175* 54-73 392-477* 687 bias, which was significant only when surrounding-context Six fillers had one correct paraphrase. Since all effects were excluded. The fact that context often favored participants answered these items correctly, results reported N1 (67% of externally-disambiguated instances) is not include all participants’ data. Analyses were conducted on R surprising. To increase text coherence, writers may prefer version 3.0.2 (R Core Team, 2013) using logit mixed-effect N1 attachment as it is the head of the complex NP and is models (see Jaeger, 2008, and references therein). part of the outer clause (e.g., the matrix clause). In two items, a portion of the RC was incorrectly shown In comprehension, readers have been shown to prefer as part of the matrix clause; and in two items, there were texts to be coherent using clausal relations such as causality more than two possible host sites. Therefore, only the and justification (Rohde, Levy, & Kehler, 2011). In the Thai results for the remaining eight items are reported (trends corpus, however, the clausal relations often involved world including all 12 items were similar to those reported). knowledge. Whether such relations are enough to affect All means reported are averages over participants. comprehension even though they do not involve causal relations was tested in the following questionnaire. Results and Discussion As predicted there was a main effect of context as the Experiment 1: Questionnaire preferred attachment site in isolation (77.08%) was less A questionnaire study is reported providing evidence that preferred when context was provided (16.67%; p < .001). intra-sentential contexts not involving causal relations can The effect of context was qualified by an interaction with affect RC attachment preferences during comprehension. type of item (N1-N2, or N2-N1: p = .013) but this is not of interest as it only indicates that the effect was stronger in the Method N1-N2 items (p < .001, according to Bonferroni-adjusted pairwise comparisons with least-square means, function Participants Sixteen Native Thai speakers aged between lsmeans; Lenth, 2013) than in the N2-N1 items (p = .005). 20-30 volunteered to participate. One of them was a The change in preferences indicates that context can graduate student in linguistics who was unaware of the affect RC-attachment during comprehension. To achieve purpose of the study. All participants signed consent forms. coherence, readers do not expect only causality and justification. They also used their world knowledge to relate Stimuli To investigate the influence of context, 12 corpus the meaning of the RC to that of the matrix clause. fragments (N1 of N2 RC), not involving causality or In this paper, we are not directly addressing the relation justification relations, were shown in isolation or with intra- between frequency in corpora and preferences in sentential context (i.e., the entire corpus sentence). An comprehension, but some of the data used to support such a example translated into English is shown in (2). frequentist explanation should be reexamined. For example, (2) a. In isolation: “Rhyme of word that arouses although in a completion questionnaire using corpus listeners’ emotion” fragments (Desmet, Brysbaert, & De Baecke, 2002), the b. With context: “The writer used only words that distribution in corpora matched native speakers’ have beautiful sound to create rhyme of word that continuations; the similar distributions may have been a arouses listeners’ emotion.” result of the surrounding contexts constraining RC The native Thai speakers who coded the corpus sentences attachment in corpora and in the completion task. (see previous section) agreed that N2 (“word”) would be In short, studies investigating RC attachment should take favored in (2a), but N1 (“rhyme”) should be favored in (2b) the influence of context into consideration. This is true for as “rhyme” is more closely related with “beautiful sound.” work using corpora, given the rich contexts that often The twelve corpus segments for which reversals were precede the target construction. But it is also true for most consistent according to a pre-test were included in the experiments showing individual sentences in isolation given questionnaire (e.g., (2)). According to the pre-test, four that intra-sentential context can be a crucial factor affecting items had a bias for N1-attachment when read in isolation, attachment (see also Rohde, Levy, & Kehler, 2011). and a bias for N2-attachment when context was included Although the N2 preference observed in the internally- (i.e., N1-N2 items). In the remaining eight items, the trend disambiguated instances of the corpus is compatible with was in the opposite direction (i.e., N2-N1 items). locality, lexical factors other than animacy and concreteness may have contributed to the N2 preference. To provide Procedure and Analyses Participants were asked to choose clearer evidence for locality in Thai, we conducted reading between two paraphrases for the two possible attachments experiments controlling for contextual and lexical factors, (e.g., “rhyme arouses listeners’ emotion” and “word arouses and report them as Experiments 2 and 3. listeners’ emotion”). Order of the paraphrases was included as a factor together with context (with context or in Experiment 2 isolation). The 12 test sentences were distributed into four An off-line task (i.e., overall preferences after the sentence lists following a Latin Square design. Each participant saw was read) was used to determine whether readers prefer N1 one list with 66 fillers. as the attachment site of the RC when the RC and the matrix clause are unrelated. 688 Method Results and Discussion Overall, the rate of N1 attachment (33.1%; i.e., a 66.9% Participants Eighteen native Thai speakers volunteered to preference for N2) was different from chance (Wilcoxon by participate in the experiment. Three of them had taken an subjects: V1 = 23.5, p = .013; by items: V2 = 6, p < .001) introductory course in linguistics. All participants in this suggesting that participants favored N2 attachment. and the subsequent experiment and in the norming However, it is conceivable that this is not evidence for a questionnaires were undergraduate students at locality preference but an unintended effect of the matrix Chulalongkorn University. They all signed a consent form. clauses used. Readers may avoid attaching the RC to N1 when it is unrelated to the matrix clause so as to avoid two Stimuli Twenty-four ambiguous sentences in which an RC unrelated clauses referring to the same entity. To address can be attached to either of the two nouns were created. To this possibility an on-line experiment was conducted. avoid animacy and concreteness confounds, and to make the results comparable to previous reports for other languages, Experiment 3 the two nouns (N1 and N2) were common human nouns. An An on-line experiment was conducted to show that there is example is given in (3). an N2 preference before readers can determine whether the (3) khó:t khɔ̌:ŋ nák wîŋ thî: wâ:t rû:p sǔaj clauses are coherent (i.e., before the matrix clause is read). coach of runner that draw picture beautifully kāmlāŋ-càʔ ʔɔ̀:k bùat Method FUTURE become-a-monk “The coach of the runner that is good at drawing is Participants A new group of 42 native Thai speakers going to become a monk.” volunteered to participate in the experiment. Thirteen of Matrix clauses unrelated to the RCs were created (e.g., in them had taken an introductory course in linguistics. (3), there is no relation between being good at drawing and becoming a monk). Five native Thai speakers confirmed Stimuli In the 24 pairs of test items used (see (4) for an that they could not find a relation between the topics in the example), the RC modified the subject so that the matrix RC and the matrix clause. None of the five speakers clause would not contaminate the reading times to the RC. participated in any of the experiments reported here. (4) The two interpretations (e.g., “coach is good at drawing,” a. N1 attachment and “runner is good at drawing”) were equally natural lǎ:nchā:j | khɔ̌:ŋ | khūnjǐŋ | thî: | phə̂:ŋ jà: | according to a norming questionnaire in which a new group nephew of duchess that just divorce of 30 native Thai speakers rated the plausibility of the two kàp ʔànōŋ | mɨ̂ a ʔā:thít thî:lɛ́:w | chô:p pāj interpretations on a five-point scale (Wilcoxon: all ps > .10). with Anong(f) when week past like go thîaw | thî: chīaŋmàj Procedure and Analyses Test items were shown in a fixed travel at Chiang Mai random order interspersed with 60 fillers so that at least one “The nephew of the duchess that got divorced from filler intervened between two test items. Anong(f) last week likes traveling to Chiang Mai.” In order to obscure the purpose of the experiment, 28 b. N2 attachment fillers were ambiguous sentences, some of which contained lǎ:nchā:j | khɔ̌:ŋ | khūnjǐŋ | thî: | phə̂:ŋ jà: | the word thî: as an RC marker of an unambiguous RC, a nephew of duchess that just divorce complementizer or a preposition. For 46 fillers, the question kàp jōŋjút | mɨ̂ a ʔā:thít thî:lɛ́:w | chô:p pāj had only one possible answer to verify that participants were with Yongyut(m) when week past like go paying attention (all participants scored over 95%). thîaw | thî: chīaŋmàj Sentences were shown individually without line breaks on travel at Chiang Mai a computer monitor. After each sentence, a question was “The nephew of the duchess that got divorced from displayed on a new screen. This procedure was adopted to Yongyut(m) last week likes traveling to Chiang prevent participants from consulting previous items or Mai.” rereading the sentence when answering the question and All crucial nouns (N1 and N2) were common human thus, noticing the ambiguity. Each question was followed by nouns. Disambiguation was based on plausibility (e.g., two alternatives with the order counterbalanced across involving gender stereotypes; “f” and “m” in the glosses in items. For the test items, the question was about attachment (4) indicate the gender of the preceding noun; e.g., in (4a), (e.g., “Who is good at drawing?”). the RC modifies the nephew as only a man and a woman How often each participant chose N1 attachment was can get divorced according to current Thai laws). included as the dependent variable. Both by-subject and by- To confirm the plausibility biases for each RC, a norming item analyses were conducted on R version 3.0.2 (R Core questionnaire was conducted with a new group of 47 native Team, 2013) using Wilcoxon signed rank test (function Thai speakers. The results indicated that the two plausible wilcox.test in the package stats; R Core Team, 2013). interpretations (e.g., for the nephew and Anong (f) to get 689 divorced, and for the duchess and Yongyut (m) to get Reading Times In the critical region, N2 attachment was divorced) were equally natural, and the two implausible read faster than N1 attachment (β = -37.74, p = .049) and interpretations (e.g., for the nephew and Yongyut (m) to get attachment did not interact with any of the other factors. In divorced, and for the duchess and Anong (f) to get divorced) the remaining regions, there was no reliable effect of were equally implausible (all ps > .10). attachment or interaction with list (main effects of list are Word and bigram frequencies for the disambiguating not of theoretical interest and are not reported). words (e.g., “Anong” in (4a), “Yongyut” in (4b)) obtained from the Thai National Corpus (Aroonmanakun, Tansiri, & Discussion Nittayanuparp, 2009) did not differ (Wilcoxon: all ps > .20). The reading-time results confirmed that with animate nouns, N2 was the preferred attachment site. The results are Procedure and Analyses The 24 pairs of test sentences and compatible with those of the off-line task in Experiment 2, 60 fillers were distributed into two lists according to a Latin confirming the locality preference in attachment in Thai. Square Design. Sentences were divided into nine regions as Because the preference was observed before the matrix indicated by the vertical bars in (4). The critical region predicate was read, we can be confident that the present where the attachment ambiguity was resolved, was always result was not affected by readers trying to make attachment the sixth region (i.e., the underlined region). coherent with the matrix clause. Participants read sentences one region at a time by pressing the space bar. After each sentence, a General Discussion comprehension question was shown on a new screen. The question did not query about attachment to avoid drawing The corpus count and Experiment 1 indicate that context participants’ attention to the point of the experiment. To fit affected attachment both in production and in the width of the screen, sentences were broken into two comprehension. With intra-sentential contexts factored out, lines. For the test items, the nouns and the RC were always the N2 preference in the corpus count was compatible with shown together on the first line (to avoid an N1-attachment locality. The effect of locality was further confirmed by off- bias; see the implicit prosody hypothesis; Fodor, 1998). line and on-line tasks (Experiments 2 and 3). Analyses were performed on R 3.0.2 (R Core Team, Previous results were primarily concerned with causality 2013) using mixed-effects models (package lme4.0; Baayen, and justifications between clauses in comprehension Davidson, & Bates, 2008, and references therein). The p- (Rohde, Levy, & Kehler, 2011). Our results extend the types values were calculated by using Wald chi-square (function of clausal relations involved and indicate that context can Anova in the package car; Fox & Weisberg, 2011). also affect corpus frequencies, thus obscuring local- For all regions, attachment, Latin-Square list and their attachment trends in production. interaction were set as fixed factors. Since the We emphasized coherence but there are many ways that disambiguating words differed (e.g., “Anong” in (4a) and context may affect preferences as sentences are read “Yongyut” in (4b)), their lengths, log frequencies (word and (Spivey, Anderson, & Farmer, 2013, for a summary of bigram), and scores from the norming questionnaire were contextual effects in various constructions). included as additional fixed factors for the critical region. Further studies are needed but if previously-reported N1 Random intercepts were included for participants and items. attachment preferences in various languages can be ascribed Because of convergence limitations, only attachment was to context, then a local attachment preference can be held as included as by-participant random slope, and attachment, a universal principle, without the need for cross-linguistic Latin-Square list and their interaction as by-item random parameterizations in the way people process sentences. slopes. Outliers beyond three standard deviations were In a recent proposal, Grillo and Costa (2014) arrived at a removed (Baayen, 2008) affecting less than 1% of the data. similar conclusion suggesting that N1 attachment is only favored when the matrix clause can give rise to an Results alternative interpretation (pseudo relative small clauses, or pseudo RCs) in which the events in the two clauses are simultaneous and only the N1 interpretation is possible. Comprehension Accuracy Overall (including test items However, the availability of pseudo RCs cannot explain the and fillers) question-response accuracy was 99.04%. None present results. For example, in Experiment 1, since none of of the participants scored less than 94%. the RCs could be interpreted as pseudo RCs, the preference Accuracy for both conditions was high, but N1 reversal are unexpected if pseudo RCs are the only (or the attachment (99.80%) was marginally higher than N2 main) factor leading to N1 preferences. attachment (98.81%) (The results for both by-subjects and Similarly, the items of some studies reporting a non-local by-items are the same: Wilcoxon Vs = 24, ps = .073; by- attachment preference could not be interpreted as pseudo subject medians for both N1 and N2 attachment: 12; by-item RCs (e.g., in Japanese: Kamide & Mitchell, 1997; Yamada, medians for both attachments: 21). Attaching to N1 may Arai, & Hirose, 2014). More interestingly, these studies in reactivate the representation of this noun, thus making it Japanese reported an initial preference for the local noun easier for participants to answer the questions, which always and a late reversal favoring the non-local noun as the matrix included N1 and the matrix predicate. clause was read. This is compatible with the assumption that 690 locality is observed initially but is overridden by text Grillo, N., & Costa, J. (2014). A novel argument for the coherence later as the matrix clause is read. universality of parsing principles. Cognition, 133, 156- The results that pseudo-RCs have been claimed to 187. explain, may be reduced to contextual effects where text Jaeger, T. F. (2008). Categorical data analysis: away from coherence favors attaching the RC to N1 to make its time ANOVAs (transformation or not) and towards logit mixed reference overlap with the time of the matrix clause. models. Journal of Memory and Language, 59, 434-446. Kamide, Y., & Mitchell, D. C. (1997). Relative clause Conclusion attachment: Nondeterminism in Japanese parsing. Journal We examined contextual effects in RC attachment in Thai. of Psycholinguistics Research, 26, 247-254. A corpus count and a questionnaire indicated that context Kullavanijaya, P. (2010). อนุประโยคขยายนาม :คุณานุ ประโยคและอนุ affected attachment. In the corpus results and two reading ประโยคเติมเต็มนาม [Clauses modifying nouns: Relative clauses experiments, there was an N2-attachment preference when and complement clauses] .In A. Prasithrathsint (Ed.), contextual effects were excluded. This is compatible with Controversial constructions in Thai grammar: Relative the assumption that locality is a universal parsing principle, clause constructions, complement clause constructions, which is modulated by context and lexical features such as serial verb constructions, and passive constructions. animacy and concreteness. Bangkok: Chulalongkorn University Press. Lenth, R. V. (2013). lsmeans: Least-squares means (1.10-2). Acknowledgments R package. This work is a part of the first author’s doctoral dissertation. R Core Team (2013). R: A language and environment for statistical computing. Vienna, Austria: R foundation for statistical computing. References Rohde, H., Levy, R., & Kehler, A. (2011). Anticipating Aroonmanakun, W., Tansiri, K., & Nittayanuparp, P. explanations in relative clause processing. Cognition, 118, (2009). Thai National Corpus: A progress report. 339-358. Proceedings of the 7th workshop on asian language Siriwittayakorn, T., Miyamoto, E. T., Ratitamkul, T., & resources, ACL-IJCNLP 2009 (pp. 153-160). Suntec, Cho, H. (2014). A non-local attachment preference in the Singapore. production and comprehension of Thai relative clauses. Baayen, R. H. (2008). Analyzing linguistic data: A practical Proceedings of the 28th Pacific Asia conference on introduction to statistics using R. Cambridge University language, information and computation (pp. 575-584). Press. Bangkok, Thailand: Chulalongkorn University. Baayen, R. H., Davidson, D. J., & Bates, D. M. (2008). Spivey, M. J., Anderson, S. E., & Farmer, T. A. (2013). Mixed-effects modeling with crossed random effects for Putting syntax in context. In R. P. G. V. Gompel (Ed.), subjects and items. Journal of Memory and Language, 59, Sentence Processing. New York, NY: Psychology Press. 390–412. Trabasso, T., Suh, S., & Payton, P. (1995). Explanatory Cuetos, F., & Mitchell, D. C. (1988). Cross-linguistic coherence in understanding and talking about events. In differences in parsing: Restriction on the use of the Late M. A. Gernsbacher & T. Givón (Eds.), Coherence in Closure strategy in Spanish. Cognition, 30, 73-105. spontaneous speech. Amsterdam: John Benjamins Desmet, T., Brysbaert, M., & De Baecke, C. (2002). The Publishing Co. correspondence between sentence production and corpus Yamada, T., Arai, M., & Hirose, Y. (2014, March). frequencies in modifier attachment. Quarterly Journal of Unforced revision in the processing of relative clause Experimental Psychology, 55A, 879-896. association ambiguity in Japanese. Paper presented at the Desmet, T., De Baecke, C., & Brysbaert, M. (2002). The 27th annual CUNY conference on human sentence influence of referential discourse context on modifier processing, Columbus, OH. attachment in Dutch. Memory & Cognition, 30, 150-157. Desmet, T., De Baecke, C., Drieghe, D., Brysbaert, M., & Vonk, W. (2006). Relative clause attachment in Dutch: On-line comprehension corresponds to corpus frequencies when lexical variables are taken into account. Language and Cognitve Processes, 21, 453-485. Fodor, J. D. (1998). Learning to parse? Journal of Psycholinguistic Research, 27, 285-317. Fox, J. & Weisberg, S. (2011). An R companion to applied regression (2nd ed.). Thousand Oaks, CA: Sage. Gibson, E. (1998). Linguistic complexity: Locality of syntactic dependencies. Cognition, 68, 1-76. 691