Why Island Constraint Is Weaker in Japanese than in English: A Processing Perspective

The constraints on discontinuous dependency in a sentence have been discussed as “island constraint”. Island phenomenon can be observed in many languages; however, while the island effect is quite noticeable in English, it seems obscure in Japanese. This study thus experimentally evaluates the strength of island effect in Japanese complex sentences quantitatively and demonstrates that the island effect in Japanese is weaker than that in English. Furthermore, we attempt to attribute the difference in the strength of the island effect in these two languages to their processing characteristics; that is, the syntactic relationship between two discontinuous elements in English must be computed at the head of an island, with one of the two elements unreceived, while the syntactic computation at the head of an island in Japanese is performed with both elements received. We assume that the processing of discontinuous dependency is thus more costly in English than in Japanese, and that the strong island effect in English reflects the relative difficulty of the syntactic computation. We obtained evidence for our hypothesis from an experiment asking participants to make grammatical judgments on Japanese sentences in which we manipulated their word-order in three ways. We suggest that part of the apparent syntactic phenomenon can be attributed to processing factors.


Introduction: Constraints on Discontinuous Dependency
In a sentence of natural language, two morphemes that are discontinuous in time-series can establish a semantically closer relationship than their adjacent morphemes. The words underlined in (1) are some examples. 1) a) If you don't feel well now, then you must tell me so. b) I don't want anybody to disturb me. c) A review came out yesterday of this article.
d) The woman who you were talking about is my aunt.
Discontinuous dependency is cross-linguistic. In the Japanese examples in (2), the underlined words establish discontinuous dependency (abbreviations, -nom, -top, and -acc mean nominal, topic, and accusative respectively).
If condition-nom ill then go home to rest "If you are ill, then you should go home to rest." b) Negative polarity item Sofu-wa joobu-de ichido-shika kaze-o hiita-koto-ga nai. grandfather-top strong only once cold-acc had fact-nom not "My grandfather is so strong that he has had a cold only once." c) Quantifier floating Gakuee-ga san-satsu hoNya-de hon-o katta. student-nom three at the bookstore book-acc bought "A student bought three books at the bookstore." Discontinuous dependency indicates that the set of sentences of natural language exceeds the generative capacity of finite-state grammar. The phenomenon is thus important for the study of computational aspects of natural language, and it has been intensively discussed in theories of syntax and sentence processing. Discontinuous dependency can cross the clause boundaries of verbal complement clauses. In (3), for example, who can be interpreted as the object of met, and the number of the embeddings is assumed to be unlimited in principle.
3) a) Who (5) can be interpreted as the object of katta (bought). The dependency here crosses the clause boundary of a verbal complement clause in the same manner as in (3a). 5) Japanese discontinuous dependency for verbal complement clause. Sono hon-o Hanako-ga [ S Taroo-ga katta]-to omotteiru. that book-acc name f -nom name m -nom bought-Comp think "That book, Hanako thinks that Taroo bought." (Saito, 1992) Island phenomenon is assumed to be cross-linguistic, though syntactic categories that constitute islands can vary among languages (Goodluck & Rochemont, 1992). We should note here that while the island effect in English is strong, the effect of possible islands in Japanese sentences is obscure. Kuno (1973), for example, judges (6) to be marginal, with the subordinate object preposed from the inside of a complex NP. 6) Japanese discontinuous dependency for complex NP (with nominal complement clause).
?  (7) are judged to be grammatical, with a discontinuous dependency between the inside and outside of a complex NP in (7a) and of an adverbial adjunct clause in (7b).
(7) a) Japanese discontinuous dependency for complex NP (with nominal complement clause).
√Sono hon-o Jon-ga[ NP [ S Mary-ga katta]-toiu uwasa]-o that book-acc John-nom Mary-nom bought-Comp rumor-acc kiita (Nakamura, 2001) heard "(As for) that book, John heard the rumor that Mary had bought (it)." 1 b) Japanese discontinuous dependency for adverbial adjunct clause √Bungakubu-ni Taro-wa [ PP [ S Jiroo-ga nyuugakushita]-node] faculty of letters-dat name m -top name m -nom entered-because odoroita (Mihara, 2000) got surprised "(As for) a faculty of letters, Taro got surprised because Jiroo had entered (it)." The possible difference in the strength of island effects can be theoretically significant because island phenomenon is assumed to be cross-linguistic. However, the comparison of the effects in two languages is difficult for a single researcher. We thus experimentally described the relative strength of island effects 1 The complementizer in (7a), toiu, is optional, and Japanese has no phonetically overt relative pronoun. In our experimental sentences we will discuss below, therefore, the complementizer toiu was not placed at the end of a subordinate clause so that a nominal complement and a relative clause may be distinguished at the head noun. Note here that complementizer to cannot head a nominal complement clause, as in (i). As for (8a) we will demonstrate that the island effect is weaker in Japanese than in English on the basis of our experimental results. As for (8b), we will propose a hypothesis for the degree of island effect from the viewpoint of time-series of sentence processing and discuss our experiment to examine our hypothesis.

Experiment 1: Quantitative Evaluation of Island Effect in
English and Japanese  (2012) successfully indicated the relative strength of the island effects between the multiple constructions. We can now evaluate the relative strength of island effect cross-linguistically by the method of Sprouse, Wagers & Phillips (2012).
In this study, we made Japanese sentences directly corresponding to the experimental sentences in Sprouse, Wagers & Phillips (2012) as in Table 2, and quantitatively described the island effects in Japanese to compare them with those in English.
We made forty-eight Japanese complex sentences including potential syntactic islands, namely, an adverbial adjunct clause, a complex NP, and an indirect question with sixteen sentences for each, in which we manipulated the presence  (2012). Here a discontinuous dependency was present when the sentence-initial object of the subordinate verb was in an embedded clause. The forty-eight experimental sentences and twenty-four fillers were counterbalanced to make the questionnaire asking the acceptability judgments by choosing one of seven point scale ("very good" to "very bad").

Results: Comparison of Island Effects in English and Japanese
The Z-scores for three kinds of Japanese complex sentence and those for corresponding English sentences by Sprouse, Wagers & Phillips (2012)      Language has been attracting researchers from many academic fields because language is specific to human beings. This is one of the main reasons why many theoretical linguists have focused on the cross-linguistic and universal aspects of language. However, our experimental results suggest that a cross-linguistic phenomenon can vary among languages. In the next section, I will propose a hypothesis for the difference of the strength of the island effects between English and Japanese from the view point of sentence processing and discuss the experiment to examine the hypothesis.

Experiment 2: Head Position and Processing Time-Series
In Japanese sentence processing, the syntactic relationship of two discontinuous elements is computed after the correspondence between the two is established.
In (9), which has a complex noun phase, for example, after the dependency between the preposed Sono hon-o (that book-acc) and its gap is established (Aoshima et al., 2004), the gap is recognized to be in a complex NP at the head, uwasa (rumor).

9) Sono hon-o [ S John-ga [ NP [ S Mary-ga katta] uwasa]-o kiita]. ⇓
syntactic computation In (10), corresponding directly to (9), on the other hand, when the syntactic relationship between the discontinuous elements is computed at the head noun, the gap is not yet recognized. 10) a) (As for) that book, John heard the rumor that Mary had bought (it). b) That book, John heard [ NP the rumor that [ S ...(?)... ⇓ syntactic computation A head functions here as an operator, syntactically relating two discontinuous elements to each other. We hypothesize (11) as the processing effect on the grammatical judgment of discontinuous dependency. 11) Processing effect on grammatical judgment of discontinuous dependency The syntactic computation of a discontinuous dependency with one of the two elements unreceived is more costly than that with both of the two received. Sentences with costly processing are likely to be judged as ungrammatical.
We will describe the method of our experiment in detail in the next section.

Participants
Thirty-eight native speakers of Japanese participated in the study. They were paid.

Materials
Four types of complex sentences with six phrases were examined as experimental sentences, namely, sentences with their subordinate clauses as verbal complement (Comp), nominal complement (Ncomp), adverbial adjunct (Adjunct), and relative (Relative) clauses.
Word-order was manipulated in three ways, namely, canonical, scrambled and relativized orders. In (12) of Comp, for example, the sentence in scrambled order was constructed from (12a) in canonical order by proposing the subordinate object shakkin-o (debt-acc), as in (12b). The sentence in relativized order was constructed from (12a) by postponing the object as the head of the relative clause, as in (12c). In (12b), after the filler-gap correspondence is established, the syntactic relationship between them is computed at the complementizer to (indicated in bold face). In (12c), on the other hand, the syntactic relationship be- The clause structures and the sequence of case particles in canonical order are presented in Table 4, and examples of four types of subordinate clauses in three word-orders are shown in Table 5. The phrase frequency was controlled by Asashi News Paper Digital Archives for the four types of subordinate clauses × the six phrase positions. The sequence of accents at the beginning of a sentence may affect the recognition of a clause boundary (Hirose, 2003). In our experimental sentences, therefore, all the human names in P1 and P2 were unaccented. Furthermore, half of the subordinate objects were accented and the other half were unaccented for each type of subordinate clause. Pragmatic plausibility was controlled for the four types of sentences in canonical order so that the availability of context would not affect grammatical judgments. Eighteen sentences were constructed for each type of subordinate clause. Thirty control sentences and twelve experimental sentences with a different purpose from this study were included in the main session. The controls were simple sentences including only one verb, and half of them were ungrammatical, in violation of argument realization. The main session thus included one hundred and fourteen sentences. Three experimental scripts were written in counter-balanced design for the four types of subordinate clauses and the three word-orders. All stimulus sentences were synthesized by computer into vocal sound.  'The ring i that Suzuki had worked at the jewelry store where Kishida bought it i "

Procedure
The stimulus sentences were auditorily presented to the participants by the self-paced sentence-by-sentence listening method. The participants were asked to make grammatical judgments on the sentences by pressing one of the two buttons ("grammatical" or "ungrammatical"). The practice session included nine trials and the presentation order of the stimulus sentences was randomized for each participant. Participants took about twenty minutes to complete the experiment.
The predictions for our experiment are summarized in (13).

13) a) Description of Japanese island effect
If a complex NP and an adverbial adjunct clause behave as island in Japanese, Ncomp, Adjunct, and Relative in scrambled (and relativized) order will be judged grammatical less often than when they are in canonical order.

b) Processing effect on discontinuous dependency
If the grammatical judgments are affected by the difference in processing cost due to the different time-series of a head and the discontinuous elements between English and Japanese, the sentences in relativized order will be judged grammatical less often than those in scrambled order. The island effect in relativized order will be noticeable, as in English.
Four participants made mistakes when judging the control sentences over 70% of the time; their responses were excluded from the subsequent analysis because we judged their linguistic performance to be inadequate for the purposes of the experiment. The mean proportions of the grammatical judgments for the four types of subordinate clauses and the three word-orders are presented in Figure 4. A two-factor ANOVA was performed with the type of subordinate clause and word-order as within factors.  were significantly lower than the mean proportions of those in canonical order.
Furthermore, the mean proportion for Relative was significantly lower than that for Ncomp in scrambled (and relativized) order. For each type of subordinate clause, the mean proportion for relativized order was significantly lower than that for scrambled order. The island effect in relativized order was quite noticeable, as predicted.

Discussion
The lower proportions of grammatical judgments for Ncomp, Adjunct, and Relative in scrambled order than that in canonical order indicate that a complex NP and an adverbial adjunct clause impose constraints on discontinuous dependency in Japanese. The island effect is thus present in Japanese, though it is weak. The more severe island effect in Relative than that in Ncomp parallels the effect in English. This response pattern suggests that the island effect is cross-linguistic. Furthermore, the greater island effect in relativized order than that in scrambled order verifies our hypothesis. We found no significant difference in the grammatical proportion for Comp between canonical and scrambled orders. Here we should focus on the relatively low proportion of grammatical judgments for Comp in relativized order (58.8%). We should note that a similar response is observed in Kluender and Kutas (1993). Kluender and Kutas (1993) performed a rapid serial visual presentation experiment, asking participants to make grammatical judgments of English sentences, some of which are given in (14). Example (14a) is a yes-no question without a discontinuous dependency. Examples (14b) and (14c) include a discontinuous dependency, crossing the boundary of a verbal complement declarative in (14b) and that of an indirect question in (14c).
An indirect question behaves as island in English. The mean proportion of grammatical judgments for (14a) was 93%, and that for (14) was 18%. We should note that the mean proportion for (14b) was only 55%, though the sen- The results of Kluender and Kutas (1993) suggest that the biclausal discontinuous dependency for a verbal complement declarative in English can be fairly costly. Because the processing time-series of a head and the two discontinuous elements for our Comp in relativized order was manipulated to correspond to English, the relatively low proportion for Comp in the word-order is not surprising 2 . We assume that biclausal discontinuous dependency is syntactically constrained cross-linguistically. Our results, however, suggest that some part of the apparent syntactic phenomenon can be attributed to processing factors.

Conflicts of Interest
The author declares no conflicts of interest regarding the publication of this paper.