The sentence above illustrates how artists exploit prosodic modulations (here, the absence of explicit cues) to create a comic effect based on syntactic ambiguity. In daily social interactions, speakers use prosody to facilitate online spoken comprehension by cueing the correct parsing of sentences (Lehiste, 1973; Frazier et al., 2006). Prosody reflects subtle modulations that shape speech envelopes (e.g., silences, pitch accents) and allow the speaker to direct the listener’s attention toward relevant information. In addition, prosodic breaks are hypothesized to help listeners segment the signal into intonational phrases and facilitate decoding its syntactic structure. Some approaches argue for an immediate use of prosodic cues in sentence processing (anti-attachment hypothesis: Watson and Gibson, 2005), whereas others argue for a more global analysis of the prosodic structure (informative boundary hypothesis: Clifton et al., 2002). However, to what extent listeners rely on prosody to parse ambiguous sentences remains unclear. Such structures are well illustrated by the classic case of RC attachment ambiguity such as in (1) (for a review, see Fernández, 2003).

Relative clause ambiguous sentences yield two possible interpretations as the RC may attach with the first noun phrase (NP1), as in interpretation (a): “the servant” (HA) or, the second (NP2), as in (b): “the actress” (LA). On the one hand, the Late Closure Principle stated that across languages, listeners generally prefer the latter (LA) interpretation in case of syntactic ambiguity (Frazier, 1979; Fernández, 2003). On the other, Cuetos and Mitchell (1988) showed that Spanish speakers choose a HA interpretation instead, and that forcing an LA interpretation leads to increased reading times, suggesting higher processing costs (Carreiras, 1992; Carreiras and Clifton, 1999; see also Augurzky, 2005). While these findings led to very prolific cross-linguistic research, Gilboy et al. (1995) suggested that the observed variability is mostly due to construction types, reflecting syntactic and semantic aspects of RC constructions. In a comprehensive review of cross-linguistic differences in attachment preference, Grillo and Costa (2014) raised important issues about syntactic characteristics of experimental stimuli used in prior studies. In fact, they convincingly argue that at least some of the studies contain Pseudo-relative (PR) small clauses (Cinque, 1992). In Spanish, for example, the structure “que + Verb phrase” can be either a RC or a PR small clause. There are a number of differences between the two clauses: RCs modify an NP, while PRs are a VP modifier and appear in a more restricted set of contexts (e.g., after perception verbs: (2), adapted from Grillo and Costa, 2014):

Importantly, there is no PR ambiguity: speakers always adopt a HA interpretation. A large proportion of PR sentences may therefore induce a preference shift from LA to HA (Grillo et al., 2015). Grillo and Costa (2014) recently observed that when controlling for structural characteristics (limiting pseudo-relative availability), speakers adopt an LA preference (Frazier, 1979) across languages, although this preference can be modulated toward HA with long RCs [Implicit prosody hypothesis, Fodor, 1998, 2002; see also Hemforth et al. (2015) for a recent account]. Additional factors affecting RC interpretation, including lexical-semantic information (MacDonald et al., 1994), have been discussed. However, most studies have been conducted with written stimuli, which give us little information about auditory language processing.

Auditory RC comprehension has been studied in a variety of languages such as English (Fernández and Sekerina, 2015), German (Augurzky, 2005), Spanish (Teira and Igoa, 2007), and Bulgarian (Stoyneshka et al., 2010). Most of the studies inserted breaks (e.g., silence) after NP1 or NP2 and presented the resulting isolated (without context) sentences using a two alternative forced-choice task (e.g., “Who was on the balcony?”). Overall, these studies suggest that prosody has asymmetrical effects on disambiguation (Fernández and Sekerina, 2015). A break after NP1 strongly reinforces the pre-existing LA preference, but a break after NP2, elicits a weaker (or altogether absent) modulation toward HA, without overriding the initial LA preference. There is an overall agreement that prosodic breaks act as boundaries (Wagner and Watson, 2010), such as a break after NP1, for example, favors binding between NP2 and the RC. Different accounts hypothesize that breaks act as a “grouping” cue (Informational boundary hypothesis: Clifton et al., 2002), or a disjunction cue (anti-attachment hypothesis: Watson and Gibson, 2005). In addition, as pointed out by Fernández (2007), and also reflected in the available data on ambiguous RCs, some breaks seem to have a stronger effect than others. It is possible that a strong universal LA preference (e.g., Late Closure principle¹: Frazier, 1979; see also recent arguments by Grillo and Costa, 2014) prevails over prosody. Alternatively, prosody may play a somehow more important role than previously thought, and that is only the case that uncontrolled confounding variables such as lexical-semantic cues and pragmatics (MacDonald et al., 1994; Desmet et al., 2006) introduce noise and mask the effects of prosody in RC auditory studies.

Along these lines, Fernández and Sekerina (2015) pointed out that the aforementioned studies all use semantically deep sentences such as (1), but do not control² for lexical-semantic or pragmatic factors that may influence the participants’ responses. For example, in sentence (1), an actress is more likely to be on a balcony than a maid. To compensate for this issue, Fernández and Sekerina used an auditory-picture matching task with semantically ‘shallow’ English RCs. In their study, semantically shallow sentences described simple geometric shapes (e.g., “What color is the tip of the triangle that has an umbrella in the middle?”), to be matched with visual displays, which, contrary to animate NPs, did not require engaging in deep semantic processing as one would with animate NPs. The authors compared the proportion of HA to chance level (HA proportion: 50%), and observed that it decreased significantly (by 24%) when a prosodic break was inserted after NP1, but did not find any significant difference with a break after NP2 (despite a numerical increase of 12%). However, their conclusion was based on the assumption that, without any cues, listeners would be equally likely to pick either interpretation. However, the use of long RCs may favor a HA preference, while short RCs favor a LA preference. In sum, in addition to examining the shift of attachment preference compared to chance level, effect sizes should also be estimated in comparison to a baseline, prosody neutral condition.

Only two studies have implemented a prosody neutral condition: Teira and Igoa (2007) in Spanish, and Fernández (2007) in English. Fernández (2007) recorded three versions for each ambiguous sentence with a break after NP1 (180 ms on average), NP2 (247 ms on average), or no break. In contrast, Teira and Igoa (2007) manually deleted the 500 ms breaks and neutralized f0 to create a “null prosody” control condition³. Despite differences in acoustic manipulation, both Fernández (2007) and Teira and Igoa (2007) describe a similar modulation of attachment preference due to the insertion of prosodic breaks, compared to their baseline (43 and 38.8% HA preference, respectively). That is, a weak effect toward HA when a break follows NP2 (48 and 46.5%) and a strong modulation toward LA when a break follows NP1 (16 and 12.7%), compared to the baseline⁴. However, for the swing toward HA interpretation following NP2 breaks the modulation they report comes closer to chance level, but does not actually reverse the interpretation to a HA preference.

We created three experimental conditions varying the presence and position of a prosodic break in ambiguous RC sentences. If listeners rely on prosody to facilitate online comprehension, prosodic break placement after either NP1 or NP2 should change the interpretation to LA or HA, respectively. To test this hypothesis, we suggest an alternative approach by adequately controlling for lexical-semantic bias. First, we matched NP1 and NP2 in frequency for every sentence. Indeed, word frequency (Bybee and Hopper, 2001; Lau et al., 2008), may influence interpretation: for example, very frequent words may be selected by participants over infrequent ones. Second, we pre-tested our sentences in an additional pool of participants in order to make sure stimuli were indeed perceived as ambiguous. We decided to embed the experimental sentences within a context, to make them part of a story. We reasoned that prosodic effects might reveal themselves more naturally under a realistic context (given that, in real life, sentences are rarely processed in isolation). Finally, we implemented a prosody-neutral baseline, which will allow us to quantify the magnitude and symmetry of the change in attachment preference, so that any trend in either way could be revealed.

Our experiment mainly focuses on the interplay between prosody and syntax, for which interpretation responses are the most informative dependent variable. Following previous accounts of short RC attachment in reading (Fodor, 1998, 2002), we predicted a slight preference for LA in the absence of prosodic cues, thus providing an appropriate baseline to measure the effects of prosody. Crucially, an effect of prosody overriding syntactic preference would be quantified by (a) the relative shift in attachment preference as compared to baseline, and (b) a modulation of this preference compared to chance level (HA 50%). We expected that the position of the prosodic break would introduce a change in attachment preference in either direction. If so, a break after the first noun (NP1) was expected to trigger a LA preference because it would bind the second NP with the RC, while a break after the second noun (NP2) should change this interpretation toward HA because it would first bind the two NPs and then the RC to them. Those predictions are compatible with both anti-attachment and informative boundary hypotheses, as the experimental sentences have a maximum of one 200 ms-pause. Reaction times (RTs) may provide us with some additional information. If listeners truly benefit from a prosodic break to disambiguate the sentences, as the anti-attachment hypothesis predicts, then both break conditions should result with a faster RTs compared to the, putatively more ambiguous, baseline.

On average, the participants responded correctly to 90% (±7) of the filler questions, suggesting that they were paying attention to the experimental sentences.

Auditory studies using ambiguous sentences have shown that prosody may modulate the listener’s interpretation to a certain extent. Specifically, prosodic breaks seem to have asymmetrical effects on sentence disambiguation. Together with the recent developments indicating that there may be a universal parsing preference toward LA (Grillo and Costa, 2014; contra Cuetos and Mitchell, 1988), it could be that this preference would take precedence over prosodic cues, consistent with syntax-first approaches (e.g., Late Closure principle; Frazier, 1987). The present study explored the alternative explanation that many factors such as lexical-semantic features may also intervene and bias the results. We suggested that the modest effects of prosody on auditory RC interpretation reported to date might be due to lack of methodological control over some key variables in addition to (or instead of) universal parsing preference of RC structures. Unlike previous studies, stimulus dimensions potentially responsible for noisy and/or artefactual results, such as word frequency, imageability, pseudo-relative availability, and contextual effects, were all controlled for in this report.

Our results show that prosodic cues reinforced the preferred interpretation or reversed it compared to the baseline, but also, for the first time, that listeners actually prefer to attach high (70%) over low, when a break follows NP2. This shift suggests that prosody interacts with syntactic preferences to modulate the interpretation, which is incompatible with a strong “syntax-first” approach. RTs offer some complementary information to the interpretation results. Indeed, RTs were faster when a break followed NP1, though there was only a marginally significant difference between the break-after-NP2 and baseline conditions (as revealed by the pairwise comparisons). These results would be more in line with a global approach to prosody-syntax integration (such as the Informative boundary hypothesis: Clifton et al., 2002; see also Pierrehumbert and Hirschberg, 1990). This approach suggests that listeners seem to weigh out the prosodic cues with previous prosodic boundaries provided by the signal. The Informative boundary hypothesis contrasts with the anti-attachment hypothesis (Watson and Gibson, 2005), which predicts that listeners immediately translate prosodic breaks as a cue not to attach with a previous constituent. However, it is not necessarily incompatible with on-line evidence for immediate integration of prosodic cues using event-related potentials (Steinhauer and Friederici, 2001).

Our study, like Fernández’ (2007) and Teira and Igoa’s (2007) employ short RCs. One should keep in mind that, as evidenced by Clifton et al. (2006), the effects of prosody are stronger when they apply to shorter phrases. and our study employ short RCs. The implementation of a neutral baseline condition was instrumental in evaluating the effect of prosody on sentence disambiguation. Two prior studies have used this approach (Fernández, 2007; Teira and Igoa, 2007). Both reported that prosody could reverse interpretation relative to a baseline (although the statistical evidence confirmed this interpretation only partially). Yet, these studies did not provide evidence for an absolute reversal in interpretation, given that listeners never seemed to prefer a HA over LA interpretation (above and beyond 50%) even in the most favorable conditions. Our results on RC interpretation contrast in that respect, as we demonstrated that listeners do shift their interpretation to a HA preference (70%), when a prosodic break followed NP2. Our methods differ in three major aspects with Teira and Igoa’s and Fernández’s methods: the lexical-semantic controls used to select the stimuli, the use of the cross-splicing method to create the baseline condition, and the choice of fillers. The cross-splicing method (e.g., Steinhauer et al., 1999; Pauker et al., 2011) was employed to prevent possible baseline contamination from the speaker projecting their own interpretation in the production of the materials. However, the proportion of HA responses in our baseline condition (40%) were quite similar to the one in Fernández’ (2007) and Teira and Igoa’s (2007) (42 and 38.8%, respectively). First, this confirms that our baseline material was valid and comparable to theirs, and second, it suggests that the method used for baseline creation (natural or cross-spliced stimuli) has little effect on the participants’ interpretation. Finally, unlike Fernández (2007), our study did not contain additional distractors other than the filler questions. This could have drawn attention to prosodic cues and overestimated their effect. However, this conclusion is unlikely because our effects were larger than Teira and Igoa (2007) who did not include filler sentences or questions at all. Moreover, this unlikely emphasize of attention on prosody would have affected equally the three conditions in a comparable way. Therefore, we suggest the control of lexical-semantic cues seems to be the best candidate to explain the stronger effect of prosodic cues, especially when a break follows NP2, in our study. Therefore, this study not only helps to corroborate the large importance of prosodic cues, but also to establish appropriate controls that help reduce noise in sentence processing studies. Our materials may therefore be used in further studies on RC attachment ambiguities.

Finally, we believe that the novelty of embedding the sentence in a context has helped inducing a more naturalistic-like processing of the materials in the experiment. We argue that the insertion of context sentences gave some legitimacy to the ambiguous RCs. Instead of being presented in isolation and seemingly irrelevant, the experimental ambiguous sentences could be integrated in a story. At this point, however, this must remain only a speculation as we did not measure the effects of breaks in sentences without a context. We base this speculation on the fact that participants were given the opportunity to familiarize with the speaker’s rhythm and voice’s intonations by short stories preceding the experimental sentences for which prosodic modulations were crucial. We hypothesize that speakers implicitly convey one correct interpretation using prosodic cues such as breaks. In turn, listeners are experts in extracting these cues in order to help build adequate syntactic structures and extract meaning, making effective communication generally possible. This may reflect in part the high flexibility that listeners have regarding the great variability in accents or cadences between speakers. On the other hand, in natural scenarios listeners use context to narrows the possibilities of interpretations of the incoming speech to avoid picking the ‘wrong’ interpretation (Gilboy et al., 1995). Though this may also be an important strategy in natural speech, in our experiment contexts were designed not to bias participants toward either interpretation, in order to isolate any effects of prosody. However, the fact that we have obtained a stronger effect of prosodic cues, even compared to Fernández (2007) and Teira and Igoa (2007) who also implemented a baseline, when controlling for lexical-semantic factors suggests that these factors matter, perhaps more than syntactic preference or prosodic cues, during communication. In addition, written studies indicate that context (semantic and pragmatic) has a dramatic effect on RC attachment preferences (Gilboy et al., 1995). Future research could focus on the interplay between prosodic breaks and semantic features, and address whether prosody can still have an impact on interpretation in the face of (or when pitched against) lexical-semantic and contextual expectations.

In sum, our results replicate numerous studies showing that listeners generally prefer to attach low (e.g., Fernández, 2007; Teira and Igoa, 2007) when listening to RC ambiguous sentence in the absence of conflicting cues (40% HA). Listeners also benefit from cues that reinforce this preference (shorter RTs when a break is after NP1). However, this initial preference is not inalterable: interpretation results show that after controlling for lexical-semantic factors and pseudo-relative availability, participants are more sensitive to prosodic cues than previously reported. In fact, even a break after NP2 does not only modulate their interpretation in toward HA relative to baseline, but also beyond a 50% HA preference threshold.

The protocol of the study was approved by the Clinical Research Ethical Committe (Comité Etico de Investigación Clinica)-Parc de Salut Mar, from the University Pompeu Fabra.

LF was responsible for designing the experiment, creating the stimuli, testing the participants, analyzing the behavioral data, and writing the article. She was supervised by EB and SS-F. EB reviewed the stimuli and was present during the recording of the materials. He also closely supervised testing, and participated in writing the article. SS-F gave the final word for the appropriateness of the stimuli, and closely supervised the analyses of the results. He also gave valuable input regarding the theoretical impact of the study, and participated in writing the article. All authors contributed to all for criteria for authorship as defined by Frontiers.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.