Relative clauses have long offered a window through which to understand language processing, as they entail a long-distance dependency between a noun phrase (NP) “filler” and an unpronounced “gap”. This filler-gap dependency induces a well-attested processing cost [1
], which has been found to be greater for object-extracted relative clauses (ORCs), such as (1), than subject-extracted relative clauses (SRCs), such as (2), across a range of languages (including Spanish [2
] and English [3
]) and experimental methodologies such as pupillometry [4
], eye-tracking while reading [2
], visual world eye-tracking [6
], event-related potentials (ERP: [9
]), positron emission tomography (PET: [10
]), and functional magnetic resonance imaging (fMRI: [12
The monkeyi [that the rabbit bites ___i] grabs the cat.
The monkeyj [that ___j bites the rabbit] grabs the cat.
A number of hypotheses have been proposed to explain this processing asymmetry, or SRC preference (see [15
] for a review). One group of hypotheses attributes the asymmetry to the difference in syntactic structure between SRCs and ORCs. For example, ORCs have been argued to be more structurally complex than SRCs, increasing processing demands [16
]. In addition, constraints on working memory have been proposed to drive the parser to resolve dependencies as soon as possible, favoring SRC interpretations (with an earlier gap) over ORC interpretations (with a later gap) [20
]. Note, however, that this account makes different predictions for languages in which the distance between the filler and the gap is shorter in ORCs than SRCs, e.g., Mandarin [24
]. On the other hand, semantic influences on the SRC preference have also been discovered, casting doubt on the exclusively syntactic nature of the pattern. For example, the SRC preference has been shown to be modulated, even to the point of disappearance, by varying the lexical animacy of the NPs in the sentence [25
]. In general, the difficulty of ORC processing is attenuated when the filler NP is lexically inanimate, compared to when it is lexically animate. In addition, the results of one study [8
] suggest a role for the perceptual
(i.e., context-specific) animacy of lexically inanimate NPs. To the authors’ knowledge, no studies have yet investigated whether a similar context-specific manipulation affects relative clause processing with lexically animate
NPs. However, research on the cognition of event construction has attested to the ability of perceivers to distinguish animate agents
from animate patients
in a very short time and to a preference for assigning agency over patiency to characters in an event [28
]. This “agent advantage” has been shown to manifest in a wide variety of human behavior, from online sentence processing in adults to comprehension in child first-language (L1) acquisition to gestural production in emerging sign languages (for a review, see [30
]). It might be expected, therefore, that the context-specific agency of animate NPs would affect relative clause processing analogously to the effects of lexical and perceptual animacy.
So far, the majority of research on the SRC processing preference has focused on monolingual native speakers [15
]. However, a growing body of work suggests that the preference tends to extend to second-language (L2) processing in learners (e.g., L1 Spanish-L2 English [22
]; L1 German-L2 Dutch [32
]). Offline comprehension and production studies also converge on an SRC preference in both the L1 and L2 of bilinguals (e.g., Spanish-Basque bilinguals [33
]; Russian-English bilinguals [34
]). While some studies have demonstrated an effect of L2 exposure on parsing preferences during comprehension of syntactically ambiguous relative clauses in the L1 (e.g., in L1 Spanish-L2 English bilinguals [35
]), relatively few studies have investigated the effects of bilingualism on the SRC preference in L1 processing.
In a previous study from our lab, Stern and colleagues [6
] used eye-tracking in the visual world paradigm to examine the SRC processing preference in Spanish-English bilinguals’ L1, Spanish. All participants were fluent in both Spanish and English, but participants were divided into two groups based on their history of language exposure. Those who were born in the anglophone U.S. or moved there before age eight were termed “heritage bilinguals”. Heritage bilinguals are usually defined as bilinguals who grew up using a language in the home that differed from the dominant community language (see, for example, [36
]). Participants who moved to the anglophone U.S. at age 17 or older were termed “late bilinguals”. This study [6
] found that the late bilingual group demonstrated the expected SRC preference, evidenced by increased fixations to the target image during SRCs compared to ORCs, while the heritage bilingual group demonstrated basically equivalent processing speed during SRCs and ORCs, i.e., no SRC preference. The late bilingual group demonstrated faster processing speeds than the heritage bilingual group during SRCs, and interestingly, the late bilingual group also demonstrated slower
processing of ORCs than the heritage bilingual group. Recall that the SRC processing preference has previously been observed in monolingual speakers of both Spanish [2
] and English [3
]. Moreover, the lexical items used in this study were highly frequent. Therefore, this group-level processing difference cannot be explained as the outcome of cross-linguistic influence or differences in speed of lexical access. Rather, in order to explain this pattern of results, it was proposed that the heritage bilinguals’ reduced quantity of Spanish input and use caused a general reduction in their predictive processing in Spanish, which slowed down processing during SRCs (the canonical or expected structure) but also reduced the processing cost incurred by encountering ORCs (the less expected structure). Thus, according to this proposal, the heritage bilingual group’s reduced predictive processing compared to the late bilingual group corresponded with both disadvantages and advantages with regard to processing time. However, as a result of the broad criteria used to group participants in that study, the precise causes of this group-level processing difference remained largely unknown.
In a follow-up study using the same experimental paradigm, Stover and colleagues [7
] recruited an additional group of Spanish-English bilinguals who had moved to the anglophone U.S. between the ages of 10 and 16. These participants did not fit clearly into either of the traditional categories of “heritage bilingual” or “late bilingual”; rather, they bridged the gap of language history between these two groups. In addition, rather than dividing participants into groups based on the age of arrival, the follow-up study [7
] conducted an individual-level analysis using a gradient, multidimensional measure of language dominance
(see Section 2.1
]). The results of that study largely supported the findings of [6
] but extended them to the level of the individual: greater Spanish dominance was found to gradiently increase SRC processing speed and gradiently decrease ORC processing speed in a basically linear fashion. Interestingly, the negative effect of dominance on ORC processing speed was more pronounced than the positive effect of dominance on SRC processing speed, suggesting that the processing cost incurred during ORCs actually outweighs the advantage afforded during SRCs. In this way, [7
] demonstrated the utility of multidimensional, individual-level measures in shedding light on bilingual language processing.
However, in [6
], fixation proportions were binned across relatively large time windows, so those studies did not allow fine-grained examinations of the time course
of processing or the way that the effects of language dominance might have varied over the course of processing. This constitutes a particularly important gap with regard to the role of prediction: in order to isolate the effects of prediction during processing, it is crucial to obtain measures before
the onset of relevant linguistic information [38
]. Since fixation proportions in those two previous studies were binned across the entire duration of the relative clause, the precise temporal onset of the SRC preference remains undetermined, as does, consequently, the role of predictive processing. Additionally, [6
] only examined fixations to the target image, ignoring fixations to the two competitor images. However, as will be described in Section 2.2
, fixations to each competitor image indicate particular processing preferences at particular moments during the linguistic stimulus. Therefore, examining fixations to the competitor images, in addition to the target image, has the potential to shed unique light on participants’ time-dependent processing strategies throughout the linguistic stimulus.
In the present study, we used the same experimental paradigm as the two studies described above but treated both language dominance and time as continuous variables in order to gain a finer-grained understanding of the relationship between individual-level language dominance and the time course of processing. An important benefit of continuous monitoring (vs. comprehension questions, or aggregated time bins) is that the gaze record provides insights into how comprehension processes unfold over time on a scale of milliseconds and how the immediate visual context contributes to spoken sentence comprehension [39
]. In particular, we examined fixations before
the onset of the relative clause in order to examine processing preferences in the absence of disambiguating linguistic information, i.e., predictive processing. In addition to manipulating the syntactic structure of the linguistic stimuli (SRC vs. ORC), we also investigated the effects of differences in the depicted agency
of the referents of the NPs in the visual stimuli. While all of the NPs in our stimuli were lexically animate, they varied in how many actions they were performing in each image (zero, one, or two). Our primary research questions can be stated as follows:
Do bilingual listeners demonstrate a syntactic preference for SRCs over ORCs during L1 processing?
Do bilingual listeners demonstrate a semantic preference to interpret animate NPs as agents rather than patients during L1 processing?
Does bilingual language dominance, operationalized as a continuous, individual-level variable, affect the time course of either of these processing preferences?
Based on the studies reviewed above, we expect to observe both a syntactic preference for SRCs and a semantic preference for agency. Moreover, we expect that greater L1 dominance will gradiently exaggerate the syntactic preference for SRCs. While at least one study has found that animacy effects are diminished in L2 RC processing [40
], to our knowledge, no studies have yet examined the influence of bilingualism on the agency preference in event cognition. Therefore, we have no specific predictions regarding the effects of language dominance on the agency preference, so this aspect of our analysis is largely exploratory at this point.
The eye-tracking results indicated two distinct influences on participants’ processing of the linguistic stimuli: a semantically driven preference to assign agency to perceived NPs, and a syntactically driven preference to interpret relative clauses as subject-extracted. Moreover, greater language dominance was found to correspond with a decrease in the semantic preference for agency and an increase in the syntactic preference for SRCs.
In the anticipation region, upon the sole mention of an NP, images depicting the referent of this NP as a thematic agent were preferred over images depicting this referent as a patient. Moreover, this agency preference manifested in a scalar way, such that the two-action image (e.g., the mentioned NP biting and grabbing) attracted more looks than the one-action image. Lastly, the by-condition difference observed in this processing pattern points to a possible distinction between the agency preference and a patiency dispreference in processing.
Since, in the anticipation region, participants received no syntactic information (beyond a single NP), the observed pattern was likely generated by semantic proclivities to assign or expect agency in event construction (and to not
expect patiency). That is, upon hearing an (animate) NP, the expectation is for the referent of that NP to act rather than to be acted upon. This result is consistent with findings on the cognition of event structure where psycholinguistic experiments, visual recognition studies, infant perception studies, and home sign language studies have shown the primacy of assigning agent roles over patient roles [28
In the relative clause region, beginning with the relative pronoun que
, images depicting an SRC interpretation were preferred over those depicting an ORC interpretation. It is possible that this apparently syntactic preference was, in fact, driven by the semantic preference observed in the anticipation region. A long-standing position in linguistics holds that thematic roles project to functional positions in the sentence, such that agents instantiate as subjects while patients instantiate as objects [63
]. While this has mainly been postulated for main clauses, it is possible, indeed likely, that this projection would extend to subordinate clauses as well. In our case, upon mention of the relative pronoun que
, syntactic structure building can begin for a relative clause. However, this process is likely to be influenced by the thematic role expectations originating in the anticipation region, thus maintaining agency for the NP and assigning a subject position to that NP in the subordinate clause. If so, semantic event structure expectations can be said to drive syntactic structure building, favoring SRCs and disfavoring ORCs. The convergence of semantic and syntactic forces might then explain the subject/object asymmetry seen in the relative clause region.
However, an explanation in which the semantic agency preference and the syntactic SRC preference are entirely interdependent is complicated by the fact that language dominance exerted opposite effects on each of these preferences. That is, greater dominance decreased the agency preference but increased the SRC preference. If these two preferences were, in fact, reducible to a single source, it would be puzzling (if not contradictory) that an individual-level factor such as language dominance could decrease one while increasing the other. Therefore, although it is clear that these two preferences often cause converging processing patterns (as described above), it is likely that they emerge from (at least somewhat) distinct sources.
Although this characterization is admittedly vague, a more detailed proposal regarding the mechanisms underlying the observed relationships between dominance and processing would, at this point, be speculative. Nonetheless, some reasoned speculation might be useful to the extent that it can drive future research. On the one hand, it is possible that the observed negative relationship between language dominance and the agency preference is a consequence of the particular languages spoken by the participants in this study: Spanish, participants’ L1 and the language of the experiment, and English, participants’ L2. It has been argued, for example, that Spanish has seen an increase in certain constructions, such as the se
-construction, left-dislocation, and plural impersonal, whose function is to highlight an affected entity (i.e., a prototypical patient, as in the middle construction “the vase broke”) while strongly de-emphasizing the agent, often to the point of omission [65
]. It is possible, then, that the negative effect of Spanish dominance on the agency preference could be particular to Spanish, perhaps from the frequency of such constructions in this language. Of course, this is merely a conjecture, but our point is that there is likely to be variability across languages in the degree to which they instantiate the agency preference.
On the other hand, the negative effect of dominance on the agency preference also suggests a more general interpretation: perhaps there is a trading relation between semantic and syntactic processing strategies during sentence processing. In the case of relative clause processing, perhaps greater reliance on the syntactic preference to interpret the structure of the clause as subject-extracted corresponds with a reduced reliance on the semantic preference to assign agency to NPs. According to this view, to the extent that increased dominance tends to increase the syntactic SRC preference [6
], it would be expected for dominance to correspondingly decrease the semantic agency preference. If we further assume that syntactic structure building strategies are more language-specific than event cognition strategies, then this explanation is intuitive in the sense that decreased language dominance would be expected to decrease syntactic structure building strategies while having little to no attenuating effect on general event cognition strategies.
As stated above, this account is mostly speculative at this point. However, it makes testable predictions. For instance, this account predicts that a similar dominance-mediated trading relation between semantic and syntactic processing strategies should be observed in later-learned languages in addition to the L1. In fact, some evidence in support of this prediction has been reported in previous studies. Although the studies of which we are aware have relied on group comparisons, rather than individual-level measures, a body of work has demonstrated that bilinguals and L2 learners use morphosyntactic cues during processing to a reduced degree compared to monolingual counterparts [66
], while at least one previous study has demonstrated evidence that bilinguals exhibit increased
semantic prediction relative to monolingual counterparts [71
]. Of course, more research is needed to understand the relationship between semantic and syntactic processing in later-learned languages, particularly regarding how this relationship is modulated by continuous factors at the level of the individual. Moreover, even in monolinguals, there might be individual variation in the extent of reliance on syntactic and semantic processing strategies. If our speculation is on the right track, then this variation should be structured such that there should be a negative correlation between indices of semantic and syntactic processing strategies.
Of course, other dimensions of individual variation are also likely to play a role in shaping the relationship between semantic and syntactic processing strategies. Individual differences in cognitive control, for instance, have been found to modulate the resolution of competing semantic and syntactic cues during thematic role assignment, such that comprehenders with greater cognitive control converge on the correct interpretation more quickly [70
]. Regarding relative clause processing, it is thus possible that individuals with greater cognitive control would be able to more quickly resolve competition between the (semantic) agency preference and the (syntactic) SRC preference via more rapid selective inhibition of misleading cues. Investigating this possibility would be a fruitful area for future research. Similarly, individual differences in working memory have been found to modulate parsing preferences during the comprehension of syntactically ambiguous relative clauses [32
]. It would thus be useful to investigate the possible influence of working memory on the relationship between the agency preference and the SRC preference during relative clause processing (see Section 1
). Finally, it is important to point out that the conceptualization of language dominance used in the present study is essentially symmetrical
, in that a score of, say, 50 (relatively Spanish-dominant) has the same meaning regardless of whether Spanish is the L1 or the L2 of the bilingual participant. However, it is likely that order of acquisition plays a role independently of relative dominance at the time of testing. While some work has compared the role of multidimensional language dominance in L1 and L2 processing [74
], more work is needed to understand the potential asymmetricality of language dominance effects.
The results of the present study demonstrate the utility of gradient, multidimensional, individual-level measures in understanding bilingual language processing. As evidenced by self-ratings of proficiency (Table 1
) and offline accuracy scores on the experimental task (Table 4
), all participants in the present study were highly proficient in their L1 (Spanish), the language of the experimental task. However, through the more granular measure of language dominance, in combination with the moment-by-moment tracking of processing via visual world eye-tracking, we were able to detect subtle patterns in the relationship between bilingual experience and L1 processing. Crucially, some of the patterns we observed, particularly among the more English-dominant participants, differed from the patterns observed in monolingual speakers of both Spanish and English
, so they cannot be explained solely via normative comparison to monolinguals. Moreover, it is not clear that any of the processing patterns we observed (e.g., greater reliance on syntactic versus semantic preferences) was “better” than any other, since all patterns entailed both benefits and detriments with regard to processing time, depending on the type of linguistic stimulus being comprehended. Broadly, the present study highlights the complexity of the relationship between the representation of multiple languages in the mind and the processing of those languages in real time, and the importance of avoiding reductionist explanations that rely solely on normative comparison to monolinguals.