Focus attracts attachment

Abstract

There is growing evidence that pitch accents as well as prosodic boundaries can affect syntactic attachment. But is this an effect of their perceptual salience (the Salience Hypothesis), or is it because accents mark the position of focus (the Focus Attraction Hypothesis)? A pair of auditory comprehension experiments shows that focus position, as indicated by preceding wh-questions instead of by pitch accents, affects attachment by drawing the ambiguous phrase to the focus. This supports the Focus Attraction Hypothesis (or a pragmatic version of salience) for both these results and previous results of accents on attachment. These experiments show that information structure, as indicated with prosody or other means, influences sentence interpretation, and suggests a view on which modifiers are drawn to the most important information in a sentence.

1. Introduction

Spoken language contains many different sources of information, from the lexical items to the rhythm and prominence pattern of the prosodic contour. We know that listeners need to fit the lexical items they hear into a syntactic structure. One important issue in psycholinguistics is how and when semantic and pragmatic information, such as the information structure of a sentence, influences sentence processing. In this project, we explore the way that preceding wh-questions can affect syntactic attachment in English, and compare this to previously-shown effects of pitch accents on attachment. It turns out that wh-questions and pitch accents both influence the attachment of an ambiguously-attached modifier. This suggests that these effects occur because of their similarity in indicating to a listener the position of focus within a sentence, supporting the Focus Attraction Hypothesis first proposed in Schafer, Carter, Clifton, & Frazier (1996). We compare this hypothesis to the Salience Hypothesis of Lee & Watson (2011), and find that a pragmatic version of salience would also be consistent with the effects.

1.1. Prosodic boundary effects on attachment

Much research in the processing of spoken sentences has found that prosodic boundaries can influence syntactic ambiguity resolution (e.g., Carlson et al., 2001; Cooper & Paccia-Cooper, 1980; Lehiste, 1973; Price, Ostendorf, Shattuck-Hufnagel, & Fong, 1991; Speer, Kjelgaard, & Dobroth, 1996; Watson & Gibson, 2005; see Cutler, Dahan, & van Donselaar, 1997 and Wagner & Watson, 2010 for reviews). This body of research shows that the location or category of a prosodic boundary can influence the likelihood that a syntactic break would be posited, or where a phrase would be attached.

One set of findings shows that, when a structural ambiguity involves the presence or absence of a syntactic boundary at a given position, the presence of a prosodic boundary facilitates or favors the presence of a syntactic boundary at that position. Examples like (1) have two parses: an early closure parse in which the initial clause ends after checks and a late closure parse in which the initial clause ends after door. While the late closure structure in (1b) is usually preferred in English, Speer et al. (1996) found that a prosodic boundary after checks facilitated the processing of the early closure structure in (1a) and that a prosodic boundary after door facilitated the late closure structure. A prosodic contour in which it is ambiguous whether the boundary position is after checks or door allowed the usual preference for late closure to re-emerge.

(1)

1. Whenever the guard checks, the door is locked. (early closure)

2. Whenever the guard checks the door, it’s locked. (late closure)

In attachment ambiguities, prior work reveals that attachment of a modifier to a potential attachment site is discouraged by an intervening prosodic boundary, and encouraged by a phrasing that prosodically groups the attachment site together with the modifier (e.g., Carlson, Clifton, & Frazier, 2001; Clifton, Carlson, & Frazier, 2002; Pynte & Prieur, 1996; Schafer, 1997). In a sentence like (2), for example, a prosodic boundary between arrived and on Monday makes high attachment (to the VP headed by the higher verb heard) more likely. Conversely, a prosodic boundary earlier, putting arrived and on Monday in the same prosodic phrase, encourages low attachment (to the VP headed by arrived).

(2) Sam heard # that Bill had arrived # on Monday.

1.2. Previous accent effects on attachment

Prosodic phrasing is not the only dimension of prosody which has been shown to impact attachment ambiguities. Schafer et al. (1996) found effects of accents on attachment preferences for relative clauses following a complex NP, as in sentences like (3).

(3) a. The sun sparkled on the PROPELLER of the plane [that the mechanic was so carefully examining].

b. The sun sparkled on the propeller of the PLANE [that the mechanic was so carefully examining].

The bracketed relative clauses could modify either the first noun (N1), propeller, or the second noun (N2), plane. Their first experiment, testing conditions like (3a-b), found that H* accents on N1 or N2 increased attachments to the accented noun.¹ A second experiment tested H* or contrastive L+H* accents on N2 along with contrastively accented or not contrastively accented relative clauses, and found that the contrastive L+H* accents were more effective in drawing attachment to N2. Relative clauses without contrastive accents were also more likely to attach to N2 than accented ones. The authors suggest that this effect is due to the conditions with contrastively accented relative clauses tending to have stronger phonological boundaries after N2, as reflected in lengthening between N2 and the accent in the relative clause and in break indices. The presence of a prosodic boundary (or a stronger prosodic boundary in these conditions) before the relative clause would decrease N2 attachments. Overall, the results of the two experiments support the authors’ Focus Attraction Hypothesis, which is phrased as follows: “A phrase that is neither a complement nor syntactically obligatory is preferentially taken to modify a focused phrase unless this violates linguistic (grammatical or pragmatic) constraints” (Schafer et al. 1996, p. 149). The ultimate idea is that the focus status of an attachment site makes it important to the sentence’s main assertion, and ambiguously attached material is drawn to this important material.

Lee and Watson (2011) propose an alternative explanation, the Salience Hypothesis: “Accented words also tend to be lengthened, their segmental content is better articulated, and they tend to be produced with greater intensity. All of these acoustic properties enhance the signal by making it louder or more prominent. We call this the Salience Hypothesis.” (p. 267). Salience is a general perceptual notion within psychology dealing with why some stimuli are more likely to grab a person’s attention and are more activated than others (e.g., Itti & Koch, 2000; Parkhurst et al., 2002; Treisman & Gelade, 1980; Weichselgartner & Sperling, 1987). It spans various methods of perception, having been studied extensively in vision, but also in hearing (e.g., Kayser et al., 2005), smell, etc. In auditory language, a word should be perceptually salient or prominent if it is longer, louder, and has a higher fundamental frequency (F0) peak than others, all properties which are likely if a word bears an H* or L+H* pitch accent (though a L* accent would lower the F0). If accent attachment effects were driven by perceptual salience, as Lee and Watson suggest, then these effects could be linked to a broad literature within psychology and perception about what makes a stimulus salient and how attention is allocated.

Several variants of the Salience Hypothesis are discussed or alluded to by Lee and Watson (2011). They initially suggest that “focused elements attract attachment because focused words are acoustically and pragmatically more salient than non-focused elements” (267). Thus we can consider a Phonological Salience Hypothesis variant, in which it is the acoustic and phonological salience of an accented word which draws modifier attachment.² In this paper, we concentrate most on this variant of the Salience Hypothesis. But we could alternatively consider a Pragmatic Salience Hypothesis, in which the pragmatic importance of an accented word attracts modifiers. We will postpone discussion of such a variant until the General Discussion because it seems to us that the predictions made by a Pragmatic Salience Hypothesis would be the same as those of the Focus Attraction Hypothesis.

Ultimately, Lee and Watson set aside these variants of the Salience Hypothesis in which there is a direct connection between the salience of an accented element and the resolution of syntactic attachment ambiguities. They instead endorse a task-specific version of the Salience Hypothesis based on question-answering. Under this Question-Answering Salience Hypothesis variant, accent attraction is due to “a post-sentence selection process in which the increased salience of accented attachment sites leads to increased selection of the accented word” (p. 267). In this variant, then, the effects of an accent on modifier attachment are indirect and should be seen as an effect of salience on question-answering strategies rather than an effect of focus or a difference in syntactic processing.

Lee and Watson (2011) provide evidence in support of the Salience Hypothesis from four experiments. Their initial experiment replicated the first experiment of Schafer et al. (1996) while adding conditions with both nouns accented and without accent on either of the possible noun attachment sites. The relative clause was drawn to the accented noun in the N1 or N2 accented conditions, while conditions with no accents or both nouns accented received intermediate responses. Additional experiments found that longer relative clauses were more likely to be attached high (to N1) than shorter relative clauses. Lee and Watson trace this effect to better memory for prominent accented words, especially in conditions of additional processing load. However, it is also plausible that longer relative clauses were more likely to be perceived as having a prosodic boundary before them, which would increase high attachments, as in Schafer et al.’s explanation of the effects of accenting the relative clauses themselves. Additionally, as pointed out by an anonymous reviewer of this manuscript, if the focus attraction effect were due, at least in part, to processing constraints, then we would also expect that more complex structures would increase the attraction effect. However, Lee and Watson did not find this to be the case; there was no effect of complexity, comparing object relative clause extraction to subject relative clause extraction, on the focus attraction effect. This slightly weakens the memory-based explanation of this effect.

Lee and Watson (2011)’s final experiment changed methodology to a completion task, similar to a task used in Schafer et al.’s first experiment. Participants answered questions about experimental items that probed the attachment of the relative clause or simply asked about the object of the matrix clause. Instead of choosing from given written responses, participants were asked to produce answers from memory. Participants still tended to attach the relative clause to the accented noun, but they also often answered with the accented noun even for questions where that was not the correct answer. The overall results certainly suggest that accented nouns are more easily remembered after the sentence than unaccented ones, corroborating the conclusions of Fraundorf, Watson, and Benjamin (2010). These results could be an effect of focus conveyed by accents, or, as suggested by Lee and Watson, the phonological salience of the accented items. However, it is possible that both are true: that accented words are more easily remembered than unaccented ones, and thus salient when choosing answers to questions, and that accents also draw the attachment of modifiers like relative clauses to the focused attachment sites. Thus, these results remain compatible with the Focus Attraction Hypothesis.

Lee & Garnsey (2012) primarily studied the effects of prosodic boundaries on the attachment of relative clauses. But one of their experiments also varied the presence of a contrastive accent on either of the two possible nouns the relative clause could modify. They found an interaction between prosodic boundary position and accent position, with the accents only affecting attachment in the conditions with the largest prosodic boundary late, i.e. between the second noun and the relative clause itself. Because of this, they suggest that effects of pitch accents on attachment may be indirect and mediated by the prosodic phrasing. However, Carlson & Tyler (2018) found effects of accents on attachment that were not dependent on prosodic boundaries, and concluded that accents and boundaries had independent effects. It is possible that Lee & Garnsey’s finding is an accidental null result of the usually small effect of accents. It also could be that the relatively complex interplay of boundaries and accents in their stimuli ended up canceling each other out.

Carlson and Tyler (2018) took up the question of whether accents affect attachment in a range of structures, not just sentences with relative clause ambiguities. They tested sentences in which two different VPs were the possible attachment sites (e.g., Sammy heard that Bill had arrived # on Friday); and sentences in which a final PP followed a verb and its object, so the VP and the object NP were the possible attachments (e.g., Paula phoned a friend # from Alabama). Both structures showed accent attachment effects, with the ambiguous constituent attaching more often into the phrase whose head bore a contrastive accent. One of the experiments also showed that accent attachment effects cannot be explained by people preferring to choose answers containing an accented word over those without (the Question-Answering Salience Hypothesis of Lee & Watson 2011). The answers paraphrasing the sentence both contained all of the potentially accented words, so this task-specific version of salience would predict equal choices of answers, but the responses were still chosen at different rates depending on the attachment favored by accent position. Thus, the results in Carlson and Tyler (2018) are compatible with both Schafer et al.’s Focus Attraction Hypothesis and the Phonological Salience Hypothesis variant. Their results are not, however, consistent with the task-specific Question-Answering Salience Hypothsis variant ultimately endorsed by Lee and Watson (2011). Consequently, we set that variant aside for the most part and concentrate on the Phonological Salience Hypothesis.

1.3. Focus

One difficulty in distinguishing between the Phonological Salience Hypothesis and the Focus Attraction Hypothesis is that accented words in English are likely to be both focused and perceptually salient (usually long, loud, and higher in pitch). This usual co-occurrence of salience and focus can make it unclear which property is responsible for an accent’s effect on other aspects of processing. One of the main predictions that differs between the two hypotheses, though, is that the Focus Attraction Hypothesis predicts that different means of marking focus in a sentence, means that do not obviously increase the salience of the focused word, should have similar effects on attachment to the already-shown effects of accent. The Phonological Salience Hypothesis, on the other hand, predicts that modifier attachment should not be affected by focus-marking strategies that rely on cues that are not salient. The present research attempts to distinguish between these hypotheses by using a means of marking focus without contrastive pitch accents: question-answer congruence in wh-questions.

Focus is a grammatical notion expressing the fact that information within a sentence is not equal in how it relates to other constituents in the sentence and the discourse (Rooth, 1992; Schwarzschild, 1999). Material which is new or contrastive is focused (as in John bought a LIZARD, not a POODLE), and so is distinguished from material which is already given and noncontrastive. Focus, and information structure more generally, we know to be part of the grammar because it can be expressed with a variety of syntactic and/or phonological mechanisms. These include specific focus-marking constructions such as shi ‘be’ in Mandarin Chinese (4a), sentence positions for focused material such as the preverbal position in Hungarian (4b), and varying the placement of prosodic accents (4c).

(4) a. Shi [Zhangsan]f zai Beijing xue yuyanxue.

BE Zhangsan at Beijing study linguistics

‘[Zhangsan]f studies linguistics in Beijing (and not my brother).’

(Mandarin Chinese, Hole 2012, cited in Chen, Lee & Pan 2016, p. 736, ex. (6a))

b. Mari egy kalapot nézett ki magának.

Mary a hat.ACC picked out herself

‘It was a hat that Mary picked for herself.’

(Hungarian, Kiss 1998, p. 249, (8a))

c. Marina SLUŠALA muzyku. vs. MARINA slušala muzyku.

Marina listened-to music

‘Marina listened to the music,’ with focus on the verb or subject, respectively

(Russian, Jasinskaja 2016, p. 713, (4b-c))

In English, pitch accents are an important cue for focus, relatively directly marking the discourse status of a focused element. However, information structure in English can also be marked through the interaction of focus and focus-sensitive constructions, including it-clefts, question-answer pairs, the operator only, negation, and modals, among others (Hajičová, Partee, & Sgall, 1998; Rooth, 1985, 1996). Here, we focus on question-answer pairs.

The felicity of question-answer pairs is constrained by their information-structural properties, such that a question-answer pair must be congruent (Rooth, 1992). For example, while the responses in (5a) and (5b) have the same truth values, only the (a) response is an appropriate and congruent response to the question in (5). Question-answer congruence requires a focused element in an answer to correspond to the wh-word in the question. Thus, (5a) is a congruent response because the focus in the answer, the object Bill, corresponds to the wh-word in the question, the object who. In contrast, response (5b) is incongruent: the focused subject Mary does not correspond to the object wh-phrase of the question.

(5) Who did Mary cut down to size?

a. Mary cut BILL down to size.

b. #MARY cut Bill down to size.

c. Mary cut Bill down to size.

d. Mary cut BILL down to size on MONDAY.

The example responses in (5a-b) both contained focus which is clearly indicated on one element within the response: a pitch accent shown by the use of CAPS. However, the same string of words without a contrastive pitch accent, as in (5c), would also be an acceptable response to the question. (5c) would be pronounced with downstepping H* accents on major content words, but contains no cues to the position of focus that are incongruent with the question in (5). Therefore the information structure of the question in (5) can be taken to indicate the relevant assignment of focus to (5c). Note that question-answer congruence is a minimum bound requirement: it merely requires that the focus in the answer address the question raised. It does not require that all new information in the response directly answer the question. The answer in (5d) is not a violation of congruence, as a focused element, BILL, does answer the question, thereby satisfying question-answer congruence. The prepositional phrase on Monday in the answer in (5d) merely supplies additional, albeit unrequested, information, in a separate focused phrase. Overall, question-answer congruence can be used to indicate the position of focus in an answer without the use of a perceptually salient cue, like a contrastive pitch accent or a nuclear (but non-contrastive) pitch accent.

The variety in the means available to languages for expressing focus beyond a simple increase in salience therefore makes it clear that focus, as a theoretical concept, must be distinguished from salience. Even if we were to concentrate on focus cued by prosodic salience, we see that there is not a one-to-one mapping from pitch accents to focus. Consider the sentences in (6), from Beaver and Clark (2008, pg. 15; originally from Selkirk, 1995). In each of (6a)-(6e), the NP bats is pitch-accented. However, in each, that pitch accent can correspond to a distinct focus; each answers a distinct question. Crucially, in each of (6b)-(6e), the salient element, bats, is not the focus. Thus, there is a double dissociation between salience and focus: a salient element does not always bear focus, and focus can be cued by means other than salience, e.g. clefting, focus particles, or question-answer congruence.

(6)

• a. Mary bought a book about [BATS]_F.

• (What did Mary buy a book about?)

• b. Mary bought a book [about BATS]_F.

• (What kind of book did Mary buy?)

• c. Mary bought [a book about BATS]_F.

• (What did Mary buy?)

• d. Mary [bought a book about BATS]_F.

• (What did Mary do?)

• e. [Mary bought a book about BATS]_F.

• (What’s been happening?)

Recent psycholinguistic work has shown that both the focused item and its likely alternatives are activated during sentence processing (see Braun & Tagliapietra, 2010; Fraundorf et al., 2010, 2013; Husband & Ferreira, 2016; Spalek, Gotzner, & Wartenburger 2014), and that focus influences the processing of pronoun reference (Arnold, 1998; Foraker & McElree, 2007; etc.) and ellipsis resolution (e.g. Carlson, 2001, 2002; Carlson, Dickey et al., 2009; Frazier & Clifton, 1998; Sag, 1980). In a particularly elegant early demonstration of the effect of focus, Cutler (1976) examined sentences in which focus had been produced in a specific position, but with the actual accented word replaced by an unaccented counterpart from another recording. Thus the prosodic context surrounding the replaced word exhibited the properties of prosodic focus on the replaced word. Such sentences still showed faster response times in a phoneme monitoring task for the target word, relative to non-focused counterparts.³ This shows that focus indicated through the prosody of the whole sentence, even without perceptual salience of the focused word, influences later processing. Similarly, Cutler and Fodor (1979) showed that a preceding wh-question facilitated phoneme monitoring within the focused word just as a pitch accent did in other studies.

From the perspective of the Focus Attraction Hypothesis, it would be reasonable for a focused word or phrase to attract the attachment of a modifier, due to its increased importance within the sentence. If focus is what explains effects of accents on attachment, then the results link to a literature within linguistics regarding its semantic and discourse effects (e.g., Beaver & Clark, 2008; Kadmon, 2001; Rooth, 1992; Schwarschild, 1999; among many others), which is almost entirely distinct from the literature on the perceptual salience of sensory stimuli.

Question-answer congruence can indicate the information structure of an utterance without changing the perceptual salience of a particular head of an attachment site. Consequently, the Phonological Salience and Focus Attraction hypotheses make diverging predictions about attachment attraction in these constructions. Because the association of a wh-question and its answer provides cues to a sentence’s information structure, the Focus Attraction Hypothesis predicts that question-answer congruence should consequently affect the attachment of an ambiguous modifier: a VP-focused wh-question should lead to more high attachments in a target sentence than an object noun-focused wh-question. Conversely, because question-answer congruence doesn’t involve the perceptual salience of a particular head of an attachment site, the Phonological Salience Hypothesis would not predict any effect on modifier attachment. If an attachment effect is observed, there would be no obvious way to link these attachment effects as a natural class under the Phonological Salience Hypothesis. Experiments 1–2 explore these predictions.

2. Experiment 1

This experiment explored whether focus conveyed through a preceding wh-question context, rather than prosody, could affect attachment. Wh-questions targeting different parts of the sentence preceded responses with an attachment ambiguity and a constant prosodic contour without contrastive accents. Specifically, the preceding questions asked about the two possible attachment sites, the VP and the object noun in this case. In the responses, neither the object noun nor the VP were contrastively pitch accented. If the Focus Attraction Hypothesis is correct, then focus conveyed by a wh-question should draw the attachment of an ambiguous modifier just as focus conveyed by contrastive pitch accents does. The Phonological Salience Hypothesis, on the other hand, would predict no attraction to the focused word, and an attraction effect would be difficult to explain, as neither head of either potential attachment site is more prosodically salient than the other.

2.1. Method

2.1.1. Materials:

Twenty sentences like (8), with a final PP that could attach to either the nearest noun (friend) or the earlier verb (phoned), were recorded by the first author with a neutral prosodic contour containing no contrastive accents. An intermediate phrase (ip) boundary separated the final adverbial from the rest of the sentence. Another speaker recorded two different wh-questions for each target sentence as in (7).

(7) Wh-questions

1. What did Paula do? (VP focus question)

2. Who did Paula phone? (Noun focus question)

(8) Answer: Paula phoned a friend ip from Alabama.

The VP-focused question always asked about the entire VP, replacing it with do, and using the wh-word what as in (7a). The Noun-focused question used either who or what, depending whether the object NP was headed by an animate or inanimate noun, as in (7b). The full set of target sentences and wh-questions appears in Appendix A.

The recorded wh-questions and answers were examined for adherence to the intended prosodic contours; any disfluent or anomalous items were re-recorded. Both speakers had been trained in prosodic analysis. All answers had the most prominent accents on the subject followed by downstepping H* accents on most other content words, as well as an intermediate phrase (ip) boundary between the object N and the PP. This boundary was instantiated with a low (L-) pitch target and a pause, followed by reset of F0 on the final accent in the adverbial. Acoustic measurements to substantiate this description are shown in Table 1, and pitch tracks illustrating a pair of sample dialogues are shown in Figure 1.

Table 1.

Average acoustic measurements for target sentences from Experiments 1–2

	Subject	Verb	Object	Boundary	Final PP
F0 peak (Hz)	313	230	200	168 (L-)	225
duration (ms)	391	364	594	162 (pause)	866
intensity peak (dB)	82	81	78		79

Figure 1.

Pitch tracks of a sample item from Experiments 1–2: pauses indicated by (ps) in the text rowsa.

a. VP question.

b. NP question

The questions and answers were concatenated into single sound files with a pause between them of 350 to 450 ms, depending on what sounded best given the speech rates of the two utterances. After hearing each short dialogue, participants chose between paraphrases of the meaning of the answer sentence following a new written wh-question, as in (9), with parenthetical labels added here for clarity:

(9) Which is true?

1. Paula phoned from Alabama to reach a friend. (high)

2. The friend who Paula phoned was from Alabama. (low)

The high attachment answer corresponded to modification of the verb (i.e., phoned), and the low attachment answer to modification of the object noun (friend). The answer choices in this experiment were in long form, meaning that both answers contained all of the critical words in the sentence, including the heads of both attachment sites. Participants thus could not choose answers solely by whether they contained the focused words or not, avoiding the possible confound identified by Lee and Watson (2011) in the Question-Answering Salience Hypothesis. The two answer choices for the comprehension questions appeared equally often in first or second position over the experiment.

2.1.2. Participants:

There were 59 subjects, recruited and paid through Amazon Mechanical Turk (AMT). All participants reported being native speakers of English who had grown up in America. Their responses to the comprehension questions for 30 unambiguous filler items were examined for accuracy. All participants had over 90% accuracy on the comprehension items and so there were no exclusions. Participants were paid $2.50 for participation.

2.1.3. Procedure:

The 20 experimental item dialogues were interspersed with 30 unambiguous filler sentences, 20 sentences with another attachment ambiguity, and 40 sentences with ellipsis structures, for a total set of 110 sentences. These were arranged into 4 pseudo-randomized lists, with no two consecutive items from the same sub-experiment and items from any one condition separated by multiple items. The lists were set up with a Latin square design, so that participants heard an equal number of items per condition, but only heard one condition of each item.

Participants from AMT clicked on a link to a Qualtrics survey. Once in Qualtrics, they read a short introductory text explaining the experiment, answered the survey, and finished by providing the answers to some demographic questions. Most participants completed the experiment within 20 minutes. They were then provided with a code to enter into AMT to qualify for payment.

2.2. Results and Discussion:

The data were analyzed using a linear mixed-effects model with a binomial link function (Bates, Maechler, Bolker, & Walker, 2015). The dependent variable was the participant response disambiguating the stimuli as high or low attachment (1 = high attachment; 0 = low attachment). The independent variable was contrast coded, with the object focus question coded as = −0.5, and the VP focus question coded as = 0.5, and was introduced into the model as a fixed effect. The model included the maximal random effects structure justified by design that would converge (Barr, Levy, Scheepers, & Tily, 2013): random intercepts for sentence frame and for participant, and random slopes for participant by question type.⁴

With a preceding VP focus question, participants chose the VP attachment (high) answer 73% of the time, and with a Noun focus question, participants chose the VP attachment (high) answer 61% of the time. This difference in responses was significant (β=0.70 ± 0.16, z=4.51, p≤0.001). The model results are shown in Table 2. The general bias towards high attachment is consistent with prior research on this structure (Britt et al., 1992; Carlson & Tyler, 2018; Ferreira & Clifton, 1986; Pynte & Prieur, 1996; Schafer, 1997; Sedivy & Spivey-Knowlton, 1994). Any high bias would be increased by the presence of the ip break before the PP, as such prosodic breaks have been shown elsewhere to increase high attachment rates (e.g. Clifton, Carlson, & Frazier, 2002).

Table 2.

Statistical analysis of results of Experiment 1

	Estimate	Std. Error	z value	Pr(>|z|)
(intercept)	1.0	0.34	3.07	0.002
Question Focus Type Contrast	0.70	0.16	4.51	<0.001

These results show that a preceding wh-question about the VP increased VP attachment interpretations for the ambiguous sentence, while a preceding wh-question about the object noun led to more NP attachment interpretations. This parallels the effects of contrastive accents on interpretation of these ambiguous sentences (Carlson & Tyler 2018), as accenting the verb led to increased VP attachments, and accenting the object noun increased NP attachments. The interpretation percentages are in roughly the same range as earlier auditory studies as well, since VP attachments ranged between 50-70% in different prosodic conditions.

These results support the Focus Attraction Hypothesis (Schafer et al. 1996; then used by Carlson & Tyler, 2018), since placing a pitch accent on the head of a phrase or having the phrase be the answer to a wh-question both result in the phrase being interpreted as focused. Perceivers prefer to attach an ambiguous modifier into a phrase whose head is focused or which itself is focused, most likely in order to add to what is perceived as the most important information in the sentence. This appears to be true regardless of the specific means of indicating focus, whether through prosody or the context of a preceding question.

While it might be possible for the Phonological Salience Hypothesis to explain this result, it is less obvious how it would do so. The heads of the attachment sites in the target sentence were not more acoustically or phonologically salient in the different conditions, as a single recording with a neutral prosodic contour was used. Instead, they were focused by forming the core of answers to wh-questions, a process which the theory of focus already explains. It seems redundant for a theory of salience to also specify that answers to preceding wh-questions count as phonologically salient, despite not differing in acoustic properties from non-salient portions of a sentence. Alternatively, the Pragmatic Salience Hypothesis, under which any focused portion of a sentence counts as salient, would predict these results, as would an Assumed Salience Hypothesis.

Additionally, the task-specific, Question-Answering Salience Hypothesis variant ultimately advanced by Lee and Watson (2011) fails to predict these results. Every alternative answer to the comprehension questions contained the relevant potentially focused words and phrases. Thus, the Question-Answering Salience Hypothesis would predict that participants should have chosen between the alternative answers at chance, because neither of them uniquely contained or was made up of an accented word. This prediction was not borne out in the results. The design of the experiment ensured that participants could not simply choose the single answer that contained either an accented or a focused word, but had to choose between answers based on their understanding of the entire auditory sentence. This replicates the finding in Carlson and Tyler (2018), Experiment 4, which also included long answers (containing all potentially accented words) to comprehension questions in a different ambiguity.

One might wonder whether there was an issue of incongruence between the questions and answers in this experiment due to the final PP being new information and prosodically accented. Focus in English is by default assigned to the rightmost element of a constituent, and this could lead the ambiguously attached PP to be construed as part of a single focus phrase. For the VP focus conditions, this would result in a focus structure as depicted in (10b), rather than the multi-focus structure we assume (10a). For the Noun focus conditions, this would again result in a focal structure with a single focus phrase, as in (11b), rather than the multi-focus structure in (11a). We believe there is independent evidence against a requirement that the ambiguously attached PP be construed as part of a single focus phrase, e.g. (10b) or (11b).

(10) What did Paula do? (VP Focus question)

a. Paula [_Foc phoned a friend] [_Foc from Alabama].

b. Paula [_Foc phoned a friend from Alabama].

(11) Who did Paula call? (N focus question)

a. Paula phoned [_Foc a friend] [_Foc from Alabama].

b. Paula phoned [_Foc a friend from Alabama].

First, there is no obligation that all focused material appear in a single focused phrase in our stimuli. The construals in (10b) and (11b), if possible, cannot be obligatory. As discussed above with the examples in (5), question-answer congruence does not require that an answer contain only a single focus addressing the question. Thus, it is possible for the PP to be construed as a focused element which is independent of the focus phrase cued by the preceding question, all while still satisfying question-answer congruence, as in (10a) and (11a).

In the same vein, we believe that the intermediate phrase prosodic break after the head Noun friend would argue against the focus structures depicted in (10b) and (11b). Consider (12), a VP focus question that contains a VP-modifying PP, and the responses illustrated in (12a) and (12b), which differ in the presence or absence of a prosodic break preceding the ambiguously attached PP. If the responses both could be equally easily construed as in (10b), we would expect them to be equally natural responses. The PP could attach to the NP and still yield a single focus phrase. Instead, we find the response in (12b) to be markedly less natural than (12a). If anything, it seems that (12b) contrasts calling from Kentucky, as indicated in the question, with calling from Alabama. However, none of the other ordinary cues of such a contrast are present in (12b), such as the conjunction but in (12c), thereby yielding the oddness of (12b). We take this to mean that the prosodic break in (12b) indicates to the listener that the subsequent PP does not form a single focused phrase with the VP, and that two foci are present. Thus, for our stimuli, which all include a prosodic break preceding the PP, the construals in (10a) and (11a) are more likely than those in (10b) and (11b).

(12) What did Paula do from Kentucky?

a. Paula phoned a friend from Alabama.

b. #Paula phoned a friend ip from Alabama.

c. Paula phoned a friend, but from Alabama.

Second, the single focus structures depicted in (10b) and (11b) predict that the attachment sites available for the ambiguously attached PP are constrained in the Noun focus but not the VP focus conditions. In the Noun focus conditions the PP would have to attach low; if the PP were to attach high, to the VP rather than to the NP, the PP would not form a subconstituent of the NP and so could not further constitute a single focus phrase. With this in mind, note that in the Noun focus conditions, the PP was attached high 61% of the time. Under this high attachment interpretation, the PP and the NP could not be construed as a single focused element as depicted in (11b). In the VP focus conditions, either low or high attachment would be consistent with construing the PP as part of the focus as in (10b): both attachment sites are inside the focused constituent. This would leave unexplained the high attachment rate of 73% we found in the VP-focused condition. If both attachment sites satisfied the requirement that only a single focused phrase appear in the answer, then why should there be a higher high attachment bias in the VP-focused condition than in the Noun-focused condition?

Thus, the single focus phrase construals depicted in (10b) and (11b) make the incorrect prediction for attachment in the Noun-focused conditions and fail to explain the overall pattern of results found. On the other hand, if the PP can be construed as a focused element independent of the focus cued by the preceding question, as argued above and as in (10a) and (11a), then the Focus Attraction Hypothesis explains our results: the ambiguously attached PP is attracted to attach to the head of the focused phrase that answers the wh-question, either the focused VP or the focused NP.

3. Experiment 2

This experiment was a replication of Experiment 1 with a minor modification. The setup of Experiment 1 allowed for the possibility that listeners could answer the comprehension question on the basis of the initial auditory wh-question, without making use of the auditory answer sentence. The wh-questions and the eventual written answer choices could fit together to make a coherent discourse (e.g., What did Paula do? Paula phoned from Alabama to reach a friend and Who did Paula phone? The friend who Paula phoned was from Alabama). Although we considered it unlikely that perceivers would ignore the auditory answer and written wh-question which followed each auditory wh-question, Experiment 2 was designed to avoid this potential explanation of the results.

3.1. Method:

3.1.1. Materials:

This experiment used the same 20 items as Experiment 1, and the same recordings of those items. The example in (13) repeats the setup, with wh-questions about the VP or the object NP, and the same target sentence as an answer.

(13) Wh-questions

1. What did Paula do? (VP question)

2. Who did Paula phone? (NP question)

Answer: Paula phoned a friend ip from Alabama.

Experiment 2 differed from 1 in the written answer choices that followed the recorded dialogues. Whereas Experiment 1 had only two answer choices per item, this experiment had four, as in (14).

(14) Which is true?

a. Paula phoned from Alabama.

b. The friend was from Alabama.

c. Paula wrote a letter to Fred.

d. Fred got a letter from Paula.

Answers (14a-b) expressed the same meanings as the two answer choices in Experiment 1, but used a shortened form due to the additional choices participants had to read. The VP attachment (high) interpretation was expressed as in (14a), and the NP attachment in (14b). The shortened VP attachment answer still could function as an answer to the VP focus wh-question from the recordings, but the short noun attachment answer did not appropriately answer the earlier noun focus question, as in the infelicitous sequence of the question in (13b), Who did Paula phone?, followed by the answer in (14b), #The friend was from Alabama. Instead, it was only coherent if the intervening auditory target sentence was processed.

Two additional answer choices were also created. The one illustrated in (14c) was a possible answer to the VP focus wh-question, as long as the material in the target sentence was not also processed: What did Paula do? Paula wrote a letter to Fred is coherent, though not if Paula phoned a friend from Alabama was understood in between. The answer choice shown in (14d) was intended to be an answer to the object focus wh-question, Who did Paula phone? Due to an error in item construction, these choices were not appropriate answers to the noun focus question, with or without the intervening target sentence. Instead, they were pure distractors which could not be accurate if participants processed either or both of the preceding auditory question and answer. The full set of answer choices for each item are shown in Appendix B. The four answer choices for the comprehension questions appeared equally often in first through fourth positions over the experiment.

3.1.2. Participants:

There were 56 participants through AMT. All but one reported being native speakers of American English; the data from one speaker of British English were dropped. No other exclusions were necessary. Subjects were paid $2.50 for participation in this experiment.

3.1.3. Procedure:

The 20 experimental dialogues were combined with 40 two-sentence items to be rated for coherence, 24 items to be rated for the confidence of the speaker, and 24 one-sentence items to be rated for naturalness, for a total of 108 items. These were combined into 12 pseudo-randomized lists with items from the same sub-experiment separated from each other. The lists were set up with a Latin square design, so that participants heard equal number of items per condition, but only heard one condition of each item. The experiment was presented to participants using AMT and Qualtrics, as in Experiment 1.

3.2. Results and Discussion:

First, across all lists and items, no participant ever chose either of the two distractor answers, neither the always-incorrect ones like (14d) nor the ones like (14c) which could answer the VP focus question if the auditory target sentence was ignored. All participants instead chose one of the two answer choices which matched the information in the intervening target sentence. This suggests that in Experiment 1, participants had similarly answered the written comprehension questions after having heard and understood the target sentence along with its preceding wh-question. This interpretation is strengthened by the fact that the NP attachment answer choice (14b) in this experiment also required the target sentence to be processed, as it could not felicitously answer the wh-question on its own due to its shortened nature. The explanation for the Experiment 1 results on which listeners processed only the wh-question and the written paraphrases, ignoring the intervening auditory target answer, found no support in this experiment from either of the two conditions in which it might have.

The data were analyzed using a linear mixed-effects model with a binomial link function (Bates et al., 2015). The dependent variable was the participant response disambiguating the stimuli as high or low attachment (1 = high attachment; 0 = low attachment). The independent variable was contrast coded, with the object focus question coded as = −0.5, and the VP focus question coded as = 0.5, and was introduced into the model as a fixed effect. The model included the maximal random effects structure justified by design that would converge (Barr et al., 2013): random intercepts for frame and for participant, and random slopes for participant by question type.⁵

With the preceding VP focus wh-question, participants chose the VP (high) attachment interpretation 68% of the time, vs. 52% following the noun focus wh-question. This is a significant difference (β=1.01 ± 0.16, z=6.41, p≤0.001). We attribute the general bias for high attachment to the presence of the ip break before the PP as well as the structure’s intrinsic bias. The model results are shown in Table 3. As in Experiment 1, the wh-question focused on the VP increased high attachments. This suggests that prior focus, conveyed by a wh-question about the VP as a whole or the object NP in particular, affected the choice of attachment for the final PP in the target sentences. Specifically, focus drew attachment, with more people interpreting the modifier as affecting each phrase more often when it was focused than when it was not. These results therefore support the Focus Attraction Hypothesis over the Phonological and Question-Answering versions of the Salience Hypothesis.

Table 3.

Statistical analysis of the results of Experiment 2

	Estimate	Std. Error	z value	Pr(>|z|)
(intercept)	0.60	0.32	1.90	0.06
Question Focus Type Contrast	1.00	0.17	5.96	<0.001

An anonymous reviewer raised the possibility that the results of Experiment 1 were the result of another type of question-answering bias. They proposed that participants might have been biased to select answers that began with information which would answer the original wh-question.We think this suggestion is unlikely because it seems to ignore the fact that the actual sequence of events in a trial was: auditory wh-question (e.g., What did Paula do?), auditory answer (Paula phoned a friend from Alabama), with concurrent visual display of a written wh-question (Which is true?) and two written answers (Paula phoned… or A friend…). We suggest that the intervening answer and following wh-question would be automatically processed, and would be likely to decrease the effectiveness of the initial auditory wh-question on answer choices. It is likely that participants read the questions less thoroughly later in the experiment, but the questions did differ between the experimental items and the various types of fillers in the experiment and so were informative for some period of time. Finally, the design and results of Experiment 2 obviates the possibility that such a bias might provide a general explanation for the modifier attraction effect. If participants were simply selecting an answer which began with a phrase that answered the initial wh-question, ignoring the intervening answer and question, then we would expect that answers 14a and 14c would have been selected equally often in the VP question conditions and answers 14b and 14d would have been selected equally often in the NP question conditions. Because the distractor answers were never selected, we claim that participants were not swayed by such a bias to ignore the intervening material.

4. General Discussion

These experiments add to the growing body of evidence that comprehenders utilize not just prosodic phrasing, but also the information-structural properties of an input sentence in the disambiguation process. Earlier work has concentrated on the role that pitch accents, one type of cue for focus, play in this process (Carlson & Tyler, 2018; Schafer et al., 1996). In this work we have shown that another focus indicator, a preceding wh-question, is similarly influential in resolving the constituent structure of a sentence by affecting attachment preferences.

From the experiments presented here, we conclude that the Phonological Salience Hypothesis does not account for the full range of effects that focus cues have on the disambiguation process, because it specifically traces effects to the perceptually salient properties of accents. In Experiments 1–2, the focal structure of the input sentence was cued by a preceding wh-question in conjunction with the constraints on question-answer congruence. Despite no contrastive pitch accenting in the critical sentence and identical recordings across conditions, listeners’ attachment of a modifier was influenced by the information structure signalled by the preceding question: the modifier was attracted to the focused phrase, a VP or an NP. Similarly, related research (Potter & Carlson 2019; Carlson & Potter 2020, under rev.) has found that the position of the focus particle only also draws modifier attachment. Thus, it appears that modifier attachment attraction cannot be fully explained in terms of the perceptual salience of an accented element. The only results predicted by the Phonological Salience Hypothesis are those where a modifier is attracted by a pitch accent, leaving the present results, with a distinct type of focus-marking impacting attachment, unexplained.

The Focus Attraction Hypothesis makes better predictions about modifier attraction across these three types of situations by identifying the position of focus as the key factor underlying the attraction effect, regardless of how the focus is conveyed. Whether focus is cued by a pitch accent, a focus particle, or a preceding wh-question, the comprehender will preferentially attach a modifier to a potential attachment site that is marked for focus. Thus the Focus Attraction Hypothesis captures the full picture of results and reveals that the natural class of information-structural cues all have similar effects on attachment.

Psycholinguistic research has become accustomed to taking account of the prosodic phrasing of sentences, understanding that it can affect syntactic phrasing and interpretation. But the prominence and focus pattern of sentences has been linked less often to syntactic interpretation (being seen as more relevant to semantic and pragmatic issues), and auditory studies are sometimes presented without detailed information about accent positions and types. These studies, along with the work by Schafer et al. (1996), Lee and Watson (2011), and Carlson and Tyler (2018), show that focus and accents can be a factor influencing syntactic structure, not just information structure. These results therefore highlight the methodological point that auditory experimental studies of attachment ambiguities, and perhaps other ambiguous structures, need to consider accent position and type, keep them consistent across conditions, and carefully describe them for the reader.

Further work may show that focus affects attachment in some structures but not others, as we have not exhaustively explored the processing of all structures with attachment ambiguities. But this project and the previous studies by Schafer et al. (1996), Lee and Watson (2011), and Carlson and Tyler (2018) have shown effects of focus on attachment into NPs and VPs, with verbs, nouns, or entire phrases focused. So the focus attraction effect is not restricted to the processing of ambiguously attached relative clauses, as in the earliest work, and seems to be a relatively broad effect. Similarly, there are additional markers of focus whose effect on attachment could be explored in future work.

In principle, a Pragmatic Salience Hypothesis (or an Assumed Salience Hypothesis) could also explain our results. Focused words are generally also salient, either through an assortment of acoustic measures or due to surrounding context like wh-questions, which is why it is so difficult to disentangle the two properties. It would certainly strengthen the case for the Focus Attraction Hypothesis if we could tie the effects of focus or salience on attachment preferences to specific predictions of the theory of focus (Kadmon, 2001; Rooth, 1992; Schwarzschild, 1999) such as the consideration of alternatives, which we are not able to do with our data. Converging work within sentence processing has found evidence for the generation and rejection of focus alternatives (Braun & Tagliapietra, 2010; Fraundorf et al. 2010, 2013; Husband & Ferreira, 2016; Spalek, Gotzner, & Warttenburger 2014), which shows that the theory of focus has observable psycholinguistic effects. It is also true that the perception or expectation of focus can lead to the perception of prominence even in the absence of phonological accent (Vainio & Järvikivi 2006; Bishop, 2012). Carlson & Potter (2020; Potter & Carlson 2019) have recently found that the focus particle only influences attachment preferences also, which the Focus Attraction Hypothesis would predict. Ultimately it may be a matter of which literature, that on focus or that on salience, has the most predictive force as researchers continue to explore attachment and syntactic processing. We happen to believe that focus, a linguistic theory which interacts with syntax and semantics, provides a more likely explanation of attachment effects on syntactic structure.

An important general question is why focus attraction should work at all—why does an ambiguously-attached phrase become drawn to a position in a focused (and salient) phrase (or a phrase whose head is focused)? We believe that people prefer to attach a modifier to a focused phrase because that phrase is important to the sentence. That is, given the choice of attaching a modifier to backgrounded, given, or simply less informative material vs. to a focused phrase in the sentence which is thus marked as important, comprehenders prefer to attach it to the focused material. This is related to the Main Assertion Hypothesis in Frazier and Clifton (2005), in which an ellipsis clause across a sentence boundary from its antecedent prefers to relate to the main assertion in the antecedent, usually material high in the syntactic tree. Traxler and Frazier (2008) also invoke a pragmatic preference for attachment to the main assertion of the sentence to explain the results of studies of PP attachment to verbs within main or subordinate clauses. Of course, these other cases involve a preference for a specific type of syntactic position, rather than one picked out by focus, so there are also differences. But the general idea of attaching material so that it is relevant to the most important information in the sentence has been proposed before, and we suggest that it extends to the cases studied here.

Appendix A

Questions and target sentences from Experiments 1a-b

1a. What did Paula do? Paula phoned a friend from Alabama.

1b. Who did Paula phone? Paula phoned a friend from Alabama

2a. What did Nathan do? Nathan followed a woman in a Ferrari.

2b. Who did Nathan follow? Nathan followed a woman in a Ferrari.

3a. What did the engineer do? The engineer repaired a machine with a metal lever.

3b. What did the engineer repair? The engineer repaired a machine with a metal lever.

4a. What did the detective do? The detective startled a thief with a mallet.

4b. Who did the detective startle? The detective startled a thief with a mallet.

5a. What did Alfie do? Alfie threatened a thug with a weapon.

5b. Who did Alfie threaten? Alfie threatened a thug with a weapon.

6a. What did Jenny do? Jenny sketched a child with crayons.

6b. Who did Jenny sketch? Jenny sketched a child with crayons.

7a. What did the producer do? The producer recorded a band with lousy equipment.

7b. Who did the producer record? The producer recorded a band with lousy equipment.

8a. What did Darrin do? Darrin greeted a visitor with a great smile.

8b. Who did Darrin greet? Darrin greeted a visitor with a great smile.

9a. What did Aaron do? Aaron spotted a girl from the party.

9b. Who did Aaron spot? Aaron spotted a girl from the party.

10a. What did the girl do? The girl frightened a boy with a snake.

10b. Who did the girl frighten? The girl frightened a boy with a snake.

11a. What did Alison do? Alison entertained a toddler with many toys.

11b. Who did Alison entertain? Alison entertained a toddler with many toys.

12a. What did the judge do? The judge ruled against a lawyer with no justification.

12b. Who did the judge rule against? The judge ruled against a lawyer with no justification.

13a. What did the driver do? The driver upset a passenger with an angry frown.

13b. Who did the driver upset? The driver upset a passenger with an angry frown.

14a. What did Heather do? Heather jostled a man with a large bag.

14b. Who did Heather jostle? Heather jostled a man with a large bag.

15a. What did the assistant do? The assistant described a problem in the report.

15b. What did the assistant describe? The assistant described a problem in the report.

16a. What did James do? James read an article on a train.

16b. What did James read? James read an article on a train.

17a. What did Amy do? Amy tripped a tourist with a cane.

17b. Who did Amy trip? Amy tripped a tourist with a cane.

18a. What did Ryan do? Ryan pursued a suspect on a motorcycle.

18b. Who did Ryan pursue? Ryan pursued a suspect on a motorcycle.

19a. What did the cops do? The cops observed an intruder with binoculars.

19b. Who did the cops observe? The cops observed an intruder with binoculars.

20a. What did Bobby do? Bobby read aloud a letter to his girlfriend.

20b. What did Bobby read aloud? Bobby read aloud a letter to his girlfriend.

Appendix B

The expanded set of 4 answers per item for Experiment 2.

1. a. Paula phoned from Alabama.

b. The friend was from Alabama.

c. Paula wrote a letter to Fred.

d. Fred got a letter from Paula.

2. a. Nathan drove a Ferrari.

b. The woman drove a Ferrari.

c. Nathan washed the woman’s windows.

d. The woman got her windows washed.

3. a. The engineer used a metal lever.

b. The machine had a metal lever.

c. The engineer looked in the mailbox.

d. The mailbox was empty.

4. a. The detective had a mallet.

b. The thief had a mallet.

c. The detective called his secretary.

d. The secretary answered the phone.

5. a. Alfie used a weapon.

b. The thug had a weapon.

c. Alfie waved at Bill.

d. Bill waved back.

6. a. Jenny sketched with crayons.

b. The child had crayons.

c. Jenny texted her boyfriend.

d. The boyfriend read the texts.

7. a. The producer used lousy equipment.

b. The band had lousy equipment.

c. The producer underpaid the band.

d. The band was exploited.

8. a. Darrin had a great smile.

b. The visitor had a great smile.

c. Darrin ignored the student.

d. The student was scared of Darrin.

9. a. Aaron was at the party.

b. The girl had been at the party.

c. Aaron visited his aunt.

d. His aunt made pancakes.

10. a. The girl used a snake.

b. The boy had a snake.

c. The girl pushed the boy.

d. The boy fell over.

11. a. Alison used many toys.

b. The toddler had many toys.

c. Alison sang for the audience.

d. The audience clapped.

12. a. The judge had no justification.

b. The lawyer had no justification.

c. The judge stared at the defendant.

d. The defendant smiled creepily.

13. a. The driver had an angry frown.

b. The passenger had an angry frown.

c. The driver passed a policeman.

d. The policeman pulled over the driver.

14. a. Heather had a large bag.

b. The man had a large bag.

c. Heather recognized an old neighbor.

d. The neighbor didn’t recognize Heather.

15. a. The assistant pointed out a problem.

b. The report had a problem.

c. The assistant asked for a raise.

d. The raise was not forthcoming.

16. a. James was on a train.

b. The article was about a train.

c. James wrote a short story.

d. The story was published.

17. a. Amy used a cane.

b. The tourist had a cane.

c. Amy tutored the child.

d. The child was late.

18. a. Ryan was on a motorcycle.

b. The suspect was on a motorcycle.

c. Ryan surprised the waitress.

d. The waitress finished work.

19. a. The cops used binoculars.

b. The intruder had binoculars.

c. The cops arrested the banker.

d. The banker complained.

20. a. Bobby read to his girlfriend.

b. The letter was to his girlfriend.

c. Bobby designed a webpage.

d. The webpage was cluttered.