!"#$ %&'()*+, *+-( THE INSTITUTE OF ELECTRONICS,. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . IEICE Technical Report. INFORMATION AND COMMUNICATION ENGINEERS

Processing object relative clauses in the context. Another support for the discourse function account for the processing load asymmetry. Atsushi Sato/

Barı! Kahraman0

Hiromu Sakai/

/0Graduate School of Education, Hiroshima University E-mail:

/{asato1983, hsakai}@hiroshima-u.ac.jp

[email protected]

Abstract Roland et al. (2007, 2008) argued that the processing load asymmetry between object and subject relative clauses (ORCs/SRCs) arose from discourse factors rather than structural factors. According to Fox and Thompson (1991), the discourse function of ORCs is to anchor a newly introduced noun to the context. The lack of appropriate context thus makes the processing of ORCs more difficult than that of SRCs. We conducted a self-paced reading experiment using Japanese relative clauses and observed that grounding to the context resolved the processing asymmetry between ORCs and SRCs. The result thus provides another support for the discourse function account. Keyword Anchoring, Context, Discourse, Grounding, Relative clause, Sentence processing, Topic ORCs are typically used for grounding modified NP in the

1. Introduction ‘… The types of music that I don’t like are opera, and

discourse context, and subject relative clauses (SRCs) for

screaming heavy metal …’

other purposes, such as supplying additional information (Roland et al. 2007)

about the modified N P. Then, when the embedded NP is

People often use relative clauses (RCs) for introducing

not grounded, ORCs will incur more processing difficulty

a new discourse referent (‘the types of music’) by using discourse topic (‘I’) [1]. In general, personal pronouns

than that of SRCs. According to Roland et al., previous researches that

refer to entities that are topical. Many researchers have

found

found an extinguished difficulty in processing of object

appropriate contexts and this functional dissatisfaction

relative clause (ORC) when RC-internal NP is pronominal

might have caused delay in ORCs. In order to test the

[2, 3, 4]. Roland et al. [5, 6] hypothesize that the

effect of appropriate context, Roland et al. conducted an

asymmetry

the

experiment using two contexts (appropriate vs. neutral,

functional dissatisfaction of ORC. We conducted an

see Table 1). They found no difference at RC combined

experiment inspired by Roland et al.’s insight and found

region (3 and 4) in appropriate condition. In neutral

results in support for their functional account.

condition, reading time of ORCs was significantly longer

1.1. Hypotheses for relative clause processing

than that of SRCs.

in

RC

processing

is

derived

from

The inquiry into RC processing was initiated by examining the source of difficulty in forming filler-gap dependencies [7]. Dominant hypotheses were based on structural

factors,

such

as

linear

distance

[8]

and

structural depth [9]. As the research had progressed, however, other factors were found

that

showed

significant

impact on

the

processing difficulty [Animacy: 10, 11, 12; Frequency:

relative

difficulty

in

ORCs

did

not

supply

Table 1 Eample of Roland et al. (2008) Appropriate condition Neutral condition There was a dinner party The banker was friendly. Saturday night. RC SRC The lady that visited the banker enjoyed the party. ORC The lady that the banker visited enjoyed the party.

13]. In addition to these studies, Roland et al., proposed a new view, which takes the discourse function of RCs as the factor that determines the processing difficulty of

2. Importance of research on Japanese Although Roland et al's [5, 6] results are in favor of

ORCs.

their discourse-based account, their support is not strong

1.2. More on Roland et al. (2007)

enough due to the head-initial nature of English relative

The hypothesis proposed by Roland et al. is based

clause. Since the head noun precedes the relative clause,

primarily on Fox & Thompson’ idea [1]. They argued that

it is not clear at which point the head noun is 'grounded'

.

Table 2 Example Stimuli for the Experiment Appropriate condition Neutral condition Pre-RC Pre-RC Nihonzin-no daigakuinsei-ga hazimete gakkai-de happyou-siNihon-de hazimete kokusaiteki-na gakkai-ga kaisai-sare-ta. ta. "Japanese graduate student gave his presentation at the "The international conference was held in Japan for the first conference for the first time." time." RC SRC {Sono/Nihonjin-no} daigakuinsei-o hihansi-ta rikigaku-no kenkyuusya-wa okotte kaizyou-o ato-ni-si-ta {That/Japanese} graduate student-ACC criticize-PAST dynamics-GEN researcher-TOP angrily hall-ACC leave-PAST "The researcher of dynamics that criticized {the/Japanese} graduate student left the hall angrily." ORC {Sono/Nihonjin-no} daigakuinsei-ga hihansi-ta rikigaku-no kenkyuusya-wa okotte kaizyou-o ato-ni-si-ta "The researcher of dynamics that {the/Japanese} graduate student criticized left the hall angrily. Wrap-up Daigakuinsei-wa batu-ga warusou-ni danzyou-o orita. "The graduate student left the platform uncomfortably." in the context. As a result, we are not sure whether the

region must be directly related to the grounding process.

processing load observed in Roland et al's experiment reflects the lack of appropriate context for grounding

(3) gakusei-ga

head noun into the context. In

student-NOM

addition, the reading time increase is mainly

hihansi-ta

kenkyuusya

criticize-PAST

researcher

“the researcher that the student criticized”

observed at the RC-internal NP region in the neutral condition (squared region in the example below).

(4) gakusei-o student-ACC

(1) The lady that the banker visited ...

hihansi-ta

kenkyuusya

criticize-PAST

researcher

“the researcher that criticized the student”

(2) The lady that visited the banker …

In addition, since the head noun itself does not appear in the previous context, the effects of NP repetition can

The RC-internal subject is already mentioned in the

be strictly controlled. We can thus present more direct

appropriate condition, whereas it is newly introduced in

support for the discourse-based account using Japanese

the neutral condition. Although Roland et al. [6] reported

materials.

that the repetition of non-topic NP rather leads to ‘repeated

noun

penalty’,

it

does

not

exclude

Finall y, the results of experiments using Japanese

the

materials provide insight into the universality of the

possibility that the repetition of topic NP is mainly

discourse-based account. Since Japanese is often argued

responsible for the reduction of processing difficult y. It

to be a discourse oriented language [14], it might be that

would be desirable if the same effect is observed in a

context effects are more robust in Japanese.

region which does not contain the repeated N P.

Based

In short, the discourse-based account can be supported

on

these

considerations,

we

conducted

an

experiment using Japanese relative clauses.

more effectively if we can measure the processing load of the relative clause at the region that is more directly

3. Experiment

related to the discourse ‘grounding’ process and does not

3.1. Method

contain

the

repeated

N P.

experimental

material

for

We these

can

provide

respects

by

ideal using

Japanese relative clauses. Since Japanese relative clauses are strictly head-final, it is reasonable to assume that the grounding of RC head must take place in the head noun region (squared region in the example below). The reading time increase at this

3.1.1. Participants 42 graduate and undergraduate students from the Hiroshima University participated in the experiment.

3.1.2. Materials 20 experimental items were used with 4 conditions. Each

item

contains 3

sentences

(Pre-RC, RC,

and

.

Wrap-up sentences, see Table 1). These items were

3.2. Results

divided into 4 lists by Latin square design. Each list was

3.2.1. Data Trimming

combined with 48 filler sets. Filler sets contained 2 or 3 sentences.

Analysis for reading times was conducted only on correct responses. 3 subjects and 1 item were excluded

In the appropriate condition, first sentence introduced a

because of low comprehension accuracy (subjects: < 70%;

person as the topic in discourse (nihonzin-no daigakuinsei

item: < 60%) and the data exceeding 2 SD was also

“Japanese graduate student”), and the introduced noun

excluded. In total, the excluded data comprises 3.8% of

re-appeared at second sentence as a subject/object in a RC.

all data.

In the neutral condition, humans were not used in the first sentence, but were used in a relative clause in the second sentence. In the appropriate context condition, for the naturalness of RC sentences, and in order to avoid the repeated noun penalty, we used

“sono

(that)”. RC

sentences were divided into regions as follows; Region1 ({sono/nihonzin-no}),

Region2

(daigakuinsei-{o/ga}),

Region3 (hihansi-ta), Region4 (rikigaku-no), Region5 (kenkyuusya-wa), Region6 (okotte), Region7 (kaizyou-o), Region8 (ato-ni-si-ta). Before the experiment, we conducted a norming study in order to test the plausibility of the used materials. 26 sets of relative clauses were compared in pairs by t-test. As a result 4 sets were significantly different, and 2 sets were excluded due to low acceptabilit y. The total number of excluded items is 6. We used a word-by-word self-paced moving window paradigm. Whole sentences were presented to the subjects We

used

Overall accuracy was 89%. ANO VA reveals no reliable difference in accuracies [RC type: F 1 (1, 38) = 0.05, MSe = 0.07,

p > .10, F 2 (1, 18) = 0.41, MSe = 0.79, p > .10;

Context type: F 1 (1, 38) = 1.39, MSe = 0.07,

p > .10,

F 2 (1, 18) = 2.54, MSe = 0.79, p > .10], and No interaction was found [F 1 (1, 38) = 0.16, MSe = 0.07,

p > .10, F 2 (1,

18) = 1.22, MSe = 0.79, p > .10].

3.2.3. Reading time Reading time analyses were conducted only on the RC sentences. The critical region is 5 (relative head region). At region 5(kenkyuusya-wa “researcher-TOP”), the main effect of RC type was significant [F 1 (1, 38) = 18.244, MSe = 150,135,

p < .01, F 2 (1, 18) = 4.202, MSe =

205,876, p = .05]. The main effect of context type was significant only in subject analysis [F 1 (1, 38) = 9.252,

3.1.3. Procedure

word-by-word.

3.2.2. Comprehension Accuracy

a

DELL

laptop,

and

the

experiment was conducted on Linger Ver.2.84 software (developed by Dough Rohde). Yes/No questions followed the end of each set of sentences.

MSe = 150,135,

p < .05, F 2 (1, 18) = 2.899, MSe =

205,876, p = .11]. The relatively small number of items might have some influence on the results of the item analysis. The simple main effect analyses showed the following pattern. The RC type had no effect in the appropriate condition [F 1 (1, 38) = 1.59, MSe = 141,001, p > .10, F 2 (1, 18) = 1.06, MSe = 184,881, p > .10]. In the

.

4. Discussion 4.1. Region 3 The appropriate context eliminated the processing asymmetry between SRCs and ORCs. Ueno & Garnsey [15] argued that the parser is surprised by the initial O-marked NP of the sentence and this surprise effect would be spilt over the verb region. If their argument is on the right track, a grounded referent reduces the markedness of the initial O-marked NP of the sentence.

4.2. Region 5 These results are naturally accounted for by the discourse function hypothesis by Roland et al. [5]. The appropriate context greatly reduced the processing load of ORCs as shown in Fig. Since ‘grounding process’ seems to occur at the RC head region in Japanese, it is reasonable neutral condition, reading times for the ORCs were slower than those for the SRCs [F 1 (1, 38) = 11.27, MSe = 160,003,

p < .01, F 2 (1, 18) = 4.39, MSe = 228,371, p

< .05]. Fig. 3 indicates the difference in reading time between reading times for the SRCs minus those for the ORCs

in

each

condition.

t-test

reveals

significant

difference between ORCs in topic condition and neutral condition [t 1 (38) = 2.63, p < .01, t 2 (18) = 1.73, p < .10]. There was no reliable difference between SRCs in the topic and neutral conditions [t 1 (38) = 0.88, p > .10, t 2 (18) = 0.49, p > .10]. At the embedded verb region (Region 3: hihansi-ta “criticized”), the main effect of the RC type was significant [F 1 (1, 38) = 5.98, MSe = 44,563.3,

p < .05,

F 2 (1, 18) = 4.58, MSe = 66,683.5, p < .05]. In the neutral condition, reading times of SRCs were significantly slower than that of ORCs [F 1 (1, 38) = 5.32, MSe = 47,649.5,

p < .05, F 2 (1, 18) = 6.22, MSe = 67,123, p

< .05]. This tendency was also observed in Ueno & Garnsey [15]. In the appropriate condition, the simple main effect of RC type was not significant [F 1 (1, 38) = 0.25, MSe = 41,677.8,

p > .10, F 2 (1, 18) = 0.004, MSe =

66,268.6, p > .10]. In whole regions, the main effect of the context type was significant [F 1 (1, 38) = 42.123, MSe = 70,557.9,

p

< .01, F 2 (1, 18) = 11.052, MSe = 94,968, p < .01]. There was no reliable main effect of RC type [F 1 (1, 38) = 0.002, MSe = 70,557.9, p > .10].

p > .10, F 2 (1, 18) = 0.10, MSe = 94,968,

to

think

the

processing

load

reduction

observed in our experiment is initiated by the appropriate context. As an alternative to Roland et al.’s discourse function hypothesis, Mak et al. [11, 18] proposed a topichood hypothesis. Their hypothesis assumes that an entity that has higher topicworthiness in an RC is regarded as a syntactic subject. According to the topichood hypothesis, in the appropriate condition, RC-internal NPs are higher in topicworthiness and the parser analyzes them as the subject of the RC. This prediction

accelerates

the

processing of ORCs in the appropriate context because RC-internal NPs are actually in the syntactic subject position. The topichood hypothesis share some intuition with discourse function hypothesis and equally effective in explaining English data. The results of our experiment, however, hypothesis

cannot

be

because

easily

explained

Japanese

relative

by

the

clauses

topic are

head-final and reading time differentce is observed not at the RC internal NP position but in the head noun position. Since topichood hypothesis is concerned about processing of the subject NP in RCs, it cannot explain the processing difficulty of the head noun. Finall y, although the effect turns out to be not significant, we observed 57 ms difference between SRCs and ORCs in neutral condition. This might be caused by structural factor [9]. Dependency Locality Theory (D LT) [8] predicts ORCs are easier to process than SRCs in Japanese. This prediction is inconsistent with experiments in Japanese. But Warren & Gibson [4] revised D LT; discourse

status

also

affects

the

integration

cost.

Although it predicts that ORCs are basically easier in

.

Japanese, D LT might be right under the condition that the

Combined with the previous research using head-initial

discourse status of NP is equal like Ishizuka et al.

relative clauses in English or Dutch [5, 6, 18], our result

observed [19].

indicated some universal characteristics of the discourse

It might be argued that reduction of processing load of

function of object relative clauses. Further research is

ORC is not the effect of context but the effect of

needed to clarify the exact nature of discourse function of

anaphoric expression “sono (that)”. In some researches [2,

relative clauses and their relationship to topichood [11,

13], the processing load of ORCs is reduced or eliminated

18], inference [20, 21], or other discourse related

even under the isolated sentences when RC-internal NPs

concepts..

are pronoun. We leave this possibility open for future

References

research. Van den Broek [20] and his colleagues [21] argued that during discourse processing, the inference has a special role on bridging sentences. Because the sentences in the neutral condition might be more distant than those of the appropriate condition, the backward inference, which bridges

between

discourse

and

a

newly

introduced

sentence, incur heavier processing load in the neutral condition than in the appropriate condition, and this cognitive load might be responsible for the results in our experiment.

If

the

second

sentence

is

headed

by

“Kaizyou-de (at the hall)”, the backward inference might be easier and, as a result, the cognitive load can be reduced. (5) gakkai-ga

kaisai-sare-ta.

conference-NOM

hold-PAS-PAST

“The conference was held.” Kaizyou-de

gakusei-{o/ga}

hihansi-ta

hall- LOC

student-{ACC/NOM}

criticize-PAST

kenkyuusya-wa … researcher-TOP “The reseacher that {criticized

the student/the

student criticized}at the hall …” If the difference of the cognitive load for bridging between the first and second sentence in each condition is the main

source of the difficulty observed in our

experiment, the use of “Kaizyou-de” in the second sentence should eliminate the processing difficulty in neutral condition. We leave this unanswered question for a further stud y.

5. Conclusion In this stud y, we presented a new set of supporting data for the discourse function account for the processing difficulty using head-final Japanese relative clauses.

[1] Fox, B. & Thompson, S., A discourse explanation of the grammar of relative clauses in English conversation, Language, Vol.66, no.2, pp297-316, 1990. [2] Gordon, P.C., Hnedrick, R., & Johnson, M., Memory interference during language processing, Journal of Experimental Psychology: Learning, Memory, & Cognition, 27, pp.1411-1423, 2001. [3] Kaan, E., Effects of NP type on the resolution of word-order ambiguities, Journal of Psycholinguistic Research, 30, pp.529-547, 2001. [4] Warren, T., & Gibson, E., The influence of referential processing on sentence complexity, Cognition, 85, pp.79-112. [5] Roland, D, O’Meara, C., Yun, H., & Mauner, G., Processing object relative clauses: Discourse or frequency?, Poster presented at the 20 th annual CUNY conference on Human sentence processing, University of California at Santa Barbara, 2007, California, US. [6] Roland, D, O’Meara, C., Yun, H., & Mauner, G., Discourse and Object Relative Clauses: The Effect of Topic versus Mention, Paper presented at the 21 s t annual CUNY conference on Human sentence processing, University of North Carolina at Chapel Hill, March 12-15 th , 2008, North Carolina, US. [7] Frazier, L., Syntactic Processing: Evidence from Dutch, Natural Language and Linguistic Theory, 5, pp.519-559, 1987. [8] Gibson, E., The Dependency Locality Theory: A Distance-Based Theory of Linguistic Complexity, In Marantz, A., Miyashita, Y., & O’Neil, W. (Eds.), Image, Language, Brain: Papers from the First Mind Articulation Project Symposium, pp.95-126, 2002. [9] Hawkins, J. A., Processing Complexity and Filler-Gap Dependencies Across Grammars, Language, 75, pp.244-285, 1999 [10] Mak, W.M., Vonk, W., & Schriefers, H., The influence of animacy on relative clause processing, Journal of Memory and Language, 47, 50-68, 2002. [11] Mak, W.M., Vonk, W., & Schriefers, H., Animacy in processing relative clauses: The hikers that rocks crush, Journal of Memory and Language, 54, 466-490, 2006. [12] Traxler, M. J., Morris, R. K., & Seely, R. E., Processing Subject and Object Relative Clauses: Evidence from Eye Movements, Journal of Memory and Language, 47, pp.69-90., 2002. [13] Reali, F., & Christiansen, M.H., Processing of relative clauses is made easier by frequency of occurrence, Journal of Memory and Language, 57,

.

1-23, 2007. [14] Kuno, S., Danwa no Bunpoo, Taisyuukan, 1978. [15] Ueno, M., & Garnsey, S.M., An ERP study of the processing of subject and object relative clauses in Japanese, Language and Cognitive Processes, Volume 23, 646 - 688, 2008. [16] Ishizuka, T., Processing Relative Clauses in Japanese, In Okabe and Nielsen (Eds.), UCLA Working Papers in Linguistics, 13, pp.135-157, 2005. [17] Miyamoto, E. T., & Nakamura, M., Subject/Object Asymmetries in the Processing of Relative Clauses in Japanese, In G. Garding & M. Tsujimura (Eds.), WCCFL 22 Proceedings, pp. 342-355, 2003. [18] Mak, W.M., Vonk, W., & Schriefers, H., Discourse structure and relative clause processing, Memory and Cognition, 36, 170-181, 2008. [19] Ishizuka, T., Nakatani, K., & Gibson. E., Processing Japanese Relative Clauses in context, 19th Annual CUNY Conference on Human Sentence Processing, New York, March 2006. [20] van den Broek, P., Comprehension and memory of narrative texts, In M. A. Gernsbacher (Ed.), Handbook of Psycholinguistics, 1s t edition, Academic press, 1994. [21] van den Broek, P., Fletcher, C.R., & Risden, K., Investigations of Inferential Processes in Reading: A theoretical and methodological integration, Discourse P rocesses, 16, pp.169-180, 1993