Relative clause attachment preferences of late bilinguals

Introduction

Parsing or syntactic analysis is a process by which incoming words are structured by the mind into a hierarchical representation according to the grammar of the language (Boland & Blodgett, 2001). It is known that language speakers use their knowledge of language structure (phonology, morphology, syntax, and semantics) to constrain their interpretations of sentences, thus accessing the grammar of a language, and assigning different syntactic structures to a sentence. There are some models for language comprehension such as the garden-path model (Frazier, 1987a, 1987b), and interactive models of ambiguity resolution. In the garden-path model, syntactic analysis is made first, and if later information turns out to be incompatible, it is followed by an analysis for semantic and pragmatic information. If an autonomous syntactic perspective is taken, these preferences may universally reflect the parsing strategies of minimal attachment, which means interpreting sentences through the syntactic structure consistent with the known input and late closure (LC) (Frazier, 1979; Frazier & Rayner, 1982), which means attaching, when possible, the incoming lexical items into the clause or phrase being processed (Frazier, 1979). The rationale behind this model is the need for cognitive systems to minimize cognitive resources used so as to incur the least load on short-term memory. In “interactive” models (Marslen-Wilson & Tyler, 1980), the subjects arrive at the most reasonable analysis of a sentence by continuous interaction of semantic/pragmatic and syntactic information sources. Relevant to these models are the investigations of the processing of relative clause attachment ambiguities which aimed to understand whether the human language processor makes use of a universal parsing strategy in sentence processing.

The idea that a sentence containing a relative clause may sometimes be ambiguous when attached to either the first or the second noun phrase in the main clause, as in the following example, is traceable to the seminal work by Cuetos and Mitchell (1988):

Someone shot the _NP1servant of the _NP2actress who was on the balcony.

As Rodriguez (2004) stated, the ambiguous nature of this particular structure is attributable to the fact that the RC could be attached to the NP1 (early closure or high attachment), the servant, or to NP2 (LC or low attachment), the actress. Parsing preferences in ambiguous RCs have been extensively investigated, and it has been shown that attachment preferences differ across languages.

Cross-linguistic differences

In an early study, Carreiras and Clifton (1993) conducted five experiments which measured reading times for Spanish and English sentences containing a complex NP followed by a relative clause. The results showed that NP1’s outnumbered NP2’s in Spanish, whereas no difference was found in English. This led the authors to suggest that RC attachment preferences do not apply universally. Studies in different languages, including English (Carreiras & Clifton, 1999; Cuetos & Mitchell, 1988), Portuguese (Maia et al., 2007), Norwegian, Swedish, Romanian (Ehrlich et al., 1999), and Arabic (Bidaoui et al., 2016) found that the predominant trend is to attach the RC to NP2 (low attachment). Italian readers seem to only initially prefer NP2 as the agent (De Vincenzi & Job, 1993, 1995). On the contrary, Spanish (Carreiras and Clifton, 1993; Cuetos & Mitchell, 1988; Fernández, 1999), German (Hemforth et al., 1998), Dutch (Brysbaert & Mitchell, 1996), French (Mitchell et al., 1990), Russian (Sekerina & Federova, 2004), Persian (Marefat & Meraji, 2005), and Japanese (Kamide & Mitchell, 1997; Miyao & Omaki, 2006) are claimed to show a tendency to attach the modifying clause to the first NP, which is called high or early attachment. In these languages, the head of the complex noun phrase (NP1) is preferred as the subject of the RC. As evidenced in the literature, attachment preferences may vary across languages. In the same vein, the language pair in this study differs according to RC attachment preferences. The relative clause is more likely to be attached to the second noun phrase (NP2) in English, and to the NP1 in Turkish.

A number of views have been suggested to explain the possible reasons for the differences in attachment types across languages. One of the early theories is Frazier and Clifton’s (1996) Construal Theory which suggests that ambiguity in modifier attachment is resolved primarily by lexical and semantic, rather than purely syntax-based or locality-based information. In the Implicit Prosody Hypothesis (Fodor, 1998, 2002), it is hypothesized that speakers of different languages show varying attachment preferences due to differences in their internalized phonologies. Recency Principle is a structure-based view in which the new material is attached to the most recently processed constituent (Gibson et al., 1996). It is considered a variant of LC, in that it predicts that the parser attaches the RC to the nearest phrase that is currently being processed, which is the case in English (Carreiras & Clifton, 1999; Fernández, 2003). Predicate Proximity Principle (Carreiras & Clifton, 1996), on the contrary, suggests that ambiguous modifiers are attached to constituents as structurally close as possible to the main predicate of the sentence, as in languages with a relatively freer word order, such as German, Greek, and Turkish. Studies on Japanese (Ito et al., 2021) may give some insight about RC attachment expectation in Turkish since it has a similar word order, and in both languages, the RC precedes both NPs that can be modified. Despite belonging to the same language family, however, there is counterevidence from Korean (Lee & Kweon, 2004), in which high attachment is preferred irrespective of the task type, thus differing from previous studies in Japanese. In languages with a more rigid word order, however, Recency Principle seems to be a more plausible explanation for attachment preferences among the models and principles mentioned above.

Bilingualism

Bilinguals’ approach to attachment preferences has been the subject of psycholinguistic studies in bilingualism. Considering the variety in the definition of bilinguals (Felser & Clahsen, 2009; McDonald, 2006), and the effects of age of acquisition (AoA) in syntactic parsing (Pavlenko, 2012), late bilinguals’ differences from native speakers (NSs) in RC attachment preferences warrant special attention.

To explore this issue, Papadopoulou and Clahsen (2002) examined how L2 learners parse temporarily ambiguous sentences containing relative clauses and found that their parsing tendencies were different from those of NSs. The authors attributed this result to a difference in the material to be processed, which had optional transitive verbs and intransitive verbs, and thus argued against unequivocal support for the transfer claim. This view was supported in a subsequent study by Papadopoulou and Clahsen (2003), where parsing strategies of L2 learners were further examined. In that study, L2 learners were found to be influenced by their L1 parsing strategies, which led the authors to conclude that L2 learners “integrate information relying more on lexical cues than the NSs and less on purely structurally based parsing strategies.” This view is known as Shallow Structure Hypothesis (SSH) (Clahsen & Felser, 2006) which claims that even highly proficient L2 speakers differ from NSs in that they have problems processing abstract syntactic representations in real time and are guided by semantic, pragmatic, probabilistic, or surface-level information to a greater extent than NSs. Later, Clahsen and Felser (2018) revised that SSH was not meant to imply a black-and-white picture and that they expected L1 influence on L2 grammatical processing to be limited. In a follow-up study on AoA, it was found that it does not affect language but is responsible for selectively modulating grammar processing (Kırkıcı & Clahsen, 2013). This view complies with SSH (Clahsen and Felser, 2018).

On the contrary, there is also some evidence against L2-specific parsing strategies. For example, Witzel et al. (2012) examined the reading patterns of NSs and high-level (Chinese) non-native speakers (NNSs) on three English sentence types involving locally ambiguous structural configurations. The results indicated that both first-language and second-language (L2) sentence comprehension is at least partially shaped by structure-based parsing strategies. In other studies, in offline tasks, both native-like and non-native-like preferences have been found. Non-native-like preferences may be attributable to language exposure, proficiency, or cross-linguistic transfer. Although Frenck-Mestre (2002) found that increased exposure to the L2 resulted in native-like preferences for L2 speakers of French, fewer instances of native-like behavior are seen in online tasks. Reading times in these studies suggest that L2 users make use of a default, low-cost recency strategy (Dekydtspotter et al., 2008; Dussias, 2003). However, L2 speakers may display no preferences at all, perhaps due to a deficiency in structural processing (Papadopoulou & Clahsen, 2003). These studies show that bilinguals’ strategies in L2 processing may differ from those in the native language, as previously found by Clahsen and Felser (2006).

Belonging to the Altaic language family, Turkish differs from English in some respects. To start with, it is an agglutinative language with a flexible word order. It is also a head-final language, and this feature shows itself in RC clauses. Unlike English, relative clauses in Turkish precede the predicates. The verb of the RC is nonfinite, and uninflected for tense and aspect. The RC participles –(y)An and –DIK are used as relativizing suffixes in subject and object relative clauses, respectively. Both encode non-future tense and follow vowel harmony. Also, –(y)AcAK is used in RCs for the future tense (Göksel & Kerslake, 2005). In this study, we focus on subject RCs comprised of complex noun phrases marked with genitive case (NP1 _genitive + NP2). Below is a typical subject RC in Turkish.

[_RCAdamı seven] [_NP1kadının] [_NP2çocuğu] küçüktü.

The son of the woman who loved the man was young.

RC parsing in Turkish has been understudied. Kırkıcı (2004, p. 9), for example, investigated attachment preferences in sentences which contain a relative clause followed by a complex noun phrase with two potential attachment sites such as RC-NP2-NP1 as in the following active sentence:

Yazar, [_RCparklarıyla ünlenen] [ _NP1ülkenin] [_NP2başkentini] ayrıntısıyla anlattı.

The author explained in detail [_NP2the capital] of the [_NP1country] which was famous for its parks.

This ambiguous sentence can be interpreted in two ways. The author explained in detail that:

1. The country is famous for its parks 2. The capital is famous for its parks.

The results of that study showed that attachment preferences may differ depending on whether the NPs are human or nonhuman, suggesting that RC ambiguity resolution was highly influenced by the semantic features of the constituents involved. The author maintained that processing high attachment sentences take less time as compared with low attachment. Refuting the parsing account for Turkish on grounds of syntactic or locality-based constraints alone, these results are considered to be in compliance with Predicate Proximity Principle and Construal Theory.

In addition, it is maintained that languages with more flexible sentential orders, such as VOS, VSO, SOV, or OSV, may have relatively strong Predicate Proximity factors, which might result in a high preference for two-word-NP-site RC attachment ambiguities. As Turkish is such a language, it is conceivable that high attachment can be an option. Supporting this view, Turan (2020) used the eye-tracking method and comprehension questions to investigate attachment preferences in Turkish relative clauses. The analysis of reading times revealed faster processing of sentences with high attachment than those with low attachment, a finding supported by the results of the comprehension task. Turan associated this with less cognitive workload and maintained that the Turkish parser gives priority to syntactic information while also using lexical-semantic information of the main verb in ambiguous sentences. Uludağ (2020b) investigated NSs of Turkish on their attachment preferences in globally ambiguous sentences to understand the influence of whether animacy would influence their preferences. The results showed a high attachment preference for RCs irrespective of animacy factor in the sentences, supporting Predicate Proximity Hypothesis. Akal (2021), on the contrary, tested the attachment preferences of Turkish NSs to observe the effect of Recency and Predicate Proximity Principles in two experiments with NPs directly followed by the main verb or adjuncts, which intervened between NP2 and the main verb. He argued that low attachment was a stronger option for the participants, but with the reservation that the rate of NP2 increased with the growing distance between the verb and its arguments. In both experiments, NP1 attachment was prevalent, which lends support for the former position possibly due to the difference in task type. Dinçtopal-Deniz (2010) tested the processing of ambiguity in temporarily and globally ambiguous sentences with three groups in online self-paced reading tasks and offline pen-and-paper questionnaires. Monolingual groups preferred low attachment with both animate and inanimate antecedents, in line with the predictions of the Construal Hypothesis. Not all studies came out with clear attachment preferences. For example, Başer (2019) investigated RC attachment preferences in two offline (pen-and-paper) tasks and an online (self-paced reading) task. She reported that monolingual Turkish speakers had no clear attachment preferences, as long as several confounding factors were controlled, including semantic associations, task requirements, and techniques. Similarly, Başer and Hohenberger (2020) argued that semantic relations and associations play an important role in ambiguity resolution in Turkish.

Carreiras and Clifton (1999) argued that in English as L1, the data are mixed; both an NP2 preference in questionnaires (Cuetos & Mitchell, 1988) and with eye-tracking measures (Henstra, 1996) and a lack of preference (Carreiras & Clifton, 1993) have been reported. Carreiras and Clifton (1999) stated that English may differ from other languages in that an RC is commonly introduced by the complementizer that, rather than a relative pronoun. This may cause English relative clauses to be treated less like pronouns and more like complement phrases, obeying LC. This makes English NP2’s relatively heavier, and the relative clause correspondingly lighter, which may have led to specific attachment sites. RCs’ length proved to have influence on the attachment preferences. As seen in Hemforth et al. (2015), in English, German, and Spanish, in sentences with longer RCs, more NP1 (high attachment) was seen.

L1 transfer has also been investigated in L2 speakers of English with different L1s. For example, Marefat and Meraji (2005) reported that monolingual speakers of Persian and English had language-specific attachment preferences in their own native language, while bilinguals transferred L1 preferences in the parsing of L2 sentences with RC ambiguity.

The effect of the task type in determining preferences in RC attachment was a subject of interest in the literature. Miyao and Omaki (2006) argued that, in online tasks, L2 learners were more likely to have low attachment preference. The study by Ito et al. (2021) combined off- and online tasks to demonstrate how L1 Japanese late intermediate learners of L2 English specify an antecedent of an RC in L1 Japanese and L2 English. The results suggest that native Japanese speakers learning English may unconsciously transfer their L1 attachment principle to L2 English. This result is consistent with Felser et al.’s (2003) report of offline tasks, revealing that L1 German and Greek advanced learners of L2 English tend to follow Predicate Proximity Principle in the target language as well as in their L1s, even though their L1 English-speaking participants showed an RC preference. The present finding suggests that, like German and Greek speakers, native Japanese speakers transfer the Predicate Proximity Principle in their L1 to comprehend the ambiguous RC in the L2. In Dinçtopal-Deniz (2010), attachment preferences of highly proficient learners of English were investigated in online and offline tasks. The participants performed differently in the two tasks, that is, in the online task, the preference was for NP1 for animate antecedents, and NP2 for inanimate ones, and in the offline task, there was a consistent NP1 preference. This result was associated with SSH, which predicts that L2 learners depend more on lexical-semantic information as compared with syntactic information. Uludağ (2020a) compared Turkish learners of English and NSs of English in an eye-tracking study with temporarily ambiguous RCs as test stimuli. The results showed that L2 learners, like NSs, depended on syntactic cues, suggesting a challenge for the assumptions of SSH.

To the best knowledge of the authors, no research has been conducted to date with Turkish–English bilinguals’ attachment preferences using a translation task. As RC attachment has predominantly been investigated through question-answer tasks in the literature, we adopted a different task, that is, translation task, to reveal whether semantic as well as syntactic analysis will shape the interpretation of the sentences and see if preferences would remain the same. Focusing on late bilinguals who learned their L2 (English) through formal education, we aimed to explore the effect of ambiguity on parsing relative clauses in L2 on the role of transfer and to identify whether the processing could be performed in a native-like way in the advanced stages. Based on the previous literature regarding the correlation between reading habits and the success of second language processing, we also investigated the reading habits of the participants. We wanted to explore if attachment preferences would differ depending on reading habits as suggested by Esen-Aygün and Kızılaslan-Tuncer (2021) who maintained that individuals with developed reading habits are “successful readers and successful readers use various strategies to increase the quality of reading action.”

The research questions are as follows:

Method

Results

As the homogeneity tests yielded no significant results, we performed a 2 (Task type: Translation vs. Response) × 2 (Ambiguity: Global vs. Local) repeated-measures ANOVA on the data.

The first main effect was found significant, F(1, 42) = 58.929, p < .001. This result shows that our participants’ NP preferences differed when answering questions about the sentences as compared with when translating them. More specifically, they used more NP1s (high attachment) for answering the questions about the sentences than when translating them. Table 1 shows that the participants preferred more NP2s when translating the sentences.

OPEN IN VIEWER

Table 1.

Mean number of NP1 and NP2 preferences in the translation of locally and globally ambiguous sentences.

The second main effect which compared local versus global ambiguity, however, was not significant, meaning that the participants had no particular NP preference regardless of sentences being either globally or locally ambiguous, F(1, 42) = 0.592, p = .45. Table 2 shows that NP1s outnumbered NP2s when answering the questions about the sentences.

OPEN IN VIEWER

Table 2.

Mean number of NP1 and NP2 preferences in answering questions about locally and globally ambiguous sentences.

The interaction between task type and ambiguity was found significant, F(1, 42) = 9.60, p < .01, that is, the total number of NP1 and NP2 preferences differ across globally and locally ambiguous sentences. NP2s outnumbered NP1s in the translation task for globally and locally ambiguous sentences, while the reverse was true for response task for globally ambiguous sentences only. There were no clear preferences in giving responses to locally ambiguous sentences.

Finally, we investigated the effects of reading habits on the participants’ NP preferences when processing globally and locally ambiguous relative clauses. The results showed no significant interaction between reading habits and task type, F(3,39) = 2.21, p > .05, and ambiguity, F(3,39) =  1.19, p > .05, respectively.

Discussion

The first research question was whether bilinguals’ attachment preferences differ across the translation and question-answer tasks. The results of the analysis showed a significant difference in our participants’ attachment preferences when performing two different tasks. More specifically, in the translation task, our participants preferred more NP2s (low attachment) in both locally and globally ambiguous sentences. However, in the question-answer task, there were more NP1s (high attachment). This shows that our participants’ greater use of NP2s when translating ambiguous RCs seems to be a typical behavior displayed by NSs of English. This preference is in line with Recency Principle, which favors NP2 as the attachment site. It can also be interpreted as a cognitive strategy, which enables the participants to apply the least effortful option to resolve ambiguity. Therefore, it can be said that late bilinguals in our study performed similarly to NSs when processing RC ambiguity.

Low attachment preference in the translation task may also be traceable to the learning process of English as a foreign language, which encourages students to attach the relative pronoun to the closest preceding noun phrase. Also, the use of translation in the ELT world is associated with the Grammar-Translation Method (GTM) developed many centuries ago. This method continues to be extensively used in many countries, including Turkey, despite the 1997 curriculum innovation which encouraged the integration of the communicative approach into language teaching. However, many Turkish teachers still use translation as part of their teaching methodology (Kırkgöz, 2008), particularly in English preparatory schools of Turkish universities (Coşkun, 2011). Therefore, this attachment habit might have led them to translate the sentences using low attachment, as an automatic response to English relative clauses comprised of complex noun phrases in genitive form marked with ‘s. This result also replicates that found in an offline study comparing NSs and proficient Chinese speakers of English. In that study, Witzel et al. (2012) argued that NSs and NNSs employ similar strategies in RC ambiguity resolution. Marefat and Farzizadeh (2018) reported similar results with Persian-speaking proficient L2 speakers in a self-paced reading task and likewise argued that proficient L2 users adopted the strategies of English NSs. Our results also support Uludağ’s (2020a) eye-tracking study revealing RC processing in L2 learners. The participants, similar to ours in profile, were found to employ syntactic cues comparable to NSs.

In the question-answer task, however, we obtained a reverse pattern, which may be associated with L1 influence, as Turkish favors more NP1 attachments. This was surprising, because our participants were all late proficient bilinguals, and we expected to obtain homogeneous responses as regards local and global ambiguity in both tasks, hypothesizing a parallelism between their translational decisions and responses regarding RC attachments. However, it turned out that they differed in their attachment preferences in RC ambiguity. Therefore, we explain these results in light of the findings of the second research question, exploring the differences between the processing of locally and globally ambiguous RCs. We attribute this difference to the gap between their competence and performance in L2. We can speculate that they employed automatic processing in resolving the ambiguity in the RCs in the translation process, under the influence of their previous language training experience in which they were taught to focus on grammar use and choose the closest antecedent as the attachment site, as mentioned before. During the question-answer task, however, they seem to have considered semantic cues as well as syntax as they can evaluate the sentence as a whole rather than attaching the relative pronoun to the antecedent closest to it as in the example below:

[_NP1The employee] of the [_NP2publisher] who was badly punched called the police.

Due to their previous foreign language learning habits prioritizing NP2 as the attachment site, the participants may have focused on NP2 (publisher), which immediately precedes the pronoun when translating the sentence. However, in the question-answer task, the participants might have noticed that the two NP’s are both potential attachment sites, thus revisiting the sentence and selecting out of two options rather than the immediate one. This result led us to suggest that the integration of internalizing syntactic and semantic processing in L2 may take longer than simply applying the grammar rules, echoing the basic principles of SSH, which hypothesizes that “L2 processing may prioritize semantic, pragmatic, or other types of non-grammatical information, with L2 speakers potentially being more sensitive to these types of information compared to L1 speakers” (Clahsen & Felser, 2006). This view aligns with Papadopoulou and Clahsen (2003) who investigated advanced learners of Greek with different L1s in performing online and offline tasks. Participants displayed a fluctuating pattern between native-like and L2-learner performances across tasks, providing evidence for incomplete acquisition of target language structures and underuse of syntactic cues in processing RC attachments.

The difference in our participants’ attachment preferences in the two tasks is also compatible with Gilboy et al.’s (1995) observation that RC attachment tendencies (high vs. low) vary considerably across different RCs in a single language, reflecting differences in focus on structure between in Tasks 1 and 2. Our participants’ NP1 attachment preference in globally ambiguous RCs may be attributed to task type. The question-answer task may have guided them into analyzing the sentences in their native language, causing L1 preference transfer. This result complies with Marefat and Meraji (2005). As argued above, as a head-final language, Turkish prioritizes high attachment. Therefore, our results are in line with a recent study by Uludağ (2020b) which supports the Predicate Proximity Principle. Similarly, Turan’s (2020) eye-tracking study reported high attachment for Turkish speakers in processing RCs in Turkish.

Finally, we aimed to explore if late bilinguals’ reading habits had any influence on their attachment preferences in both tasks and found no apparent effect. We hypothesized that the participants who read more extensively would be able to deal more successfully with ambiguity resolution, particularly in dealing with globally ambiguous RCs. However, we found no evidence that reading extensively facilitates the resolution of such ambiguity.

Conclusion

This study investigated late bilinguals’ parsing ambiguity in RCs from a psycholinguistic perspective by employing a translation task, in addition to the question-answer task used in previous studies in the parsing literature. Students of translation and interpreting, considered advanced L2 English learners with L1 Turkish (N = 43) were asked to parse locally and globally ambiguous sentences containing relative clauses in two offline tasks: translation and question answer. Also, the correlation between their preferences and reading habits was investigated.

First, our results showed that the type of the task determined their attachment preferences, such that they used more NP2s (low attachment) in the translation task, which we attributed to being accustomed to former language learning habits and also, their lack of syntactic parsing strategies in L2. This preference did not seem to be affected by the type of ambiguity, that is, local versus global. This result could also be supportive of Recency Principle in L2 English. Second, in the question-answer task, the reverse pattern was observed. This can be interpreted either as L1 influence, or alternatively as underuse of syntactic information in L2, indicative of SSH. Finally, we found no correlation between attachment preferences and reading habits.

This study was the first to examine parsing ambiguity in RCs through a translation task, allowing a comparison of late bilinguals’ competence and performance in L2 parsing, as well as exploration of the role of transfer, and the identification of whether the processing could be performed in a native-like way in the advanced stages. Further studies may investigate ambiguity parsing by using online tasks backed by the eye-tracking method.