How focus particles and accents affect attachment

This project shows that focus and information structure, as indicated by the focus particle “only” and pitch accents, influence syntactic attachment, in contrast to the well-known effects of prosodic boundaries on attachment. One written questionnaire, one completion study, and several auditory questionnaires show that the position of “only” strongly affects attachment preferences in ambiguous sentences, while contrastive pitch accents have smaller effects. The two types of focus marking do not interact but independently impact attachment. These results support a modified version of the Focus Attraction Hypothesis, with ambiguous material drawn to attach to the most important information in a sentence. This research shows that information structure can affect sentence structure as well as discourse coherence.


Introduction
Sentence processing involves the combination and use of many different types of information, including prosodic and semantic information. One central question for the field is which types of information affect which parts of processing. This work shows novel evidence that the focus particle only can influence the attachment of phrases into an evolving syntactic structure, a core part of the sentence processing endeavor.
Numerous works have shown that prosodic boundaries affect the attachment of phrases in a sentence (e.g., Lehiste 1973;Price et al. 1991;Pynte & Prieur 1996;Schafer 1997;Carlson et al. 2001;Snedeker & Trueswell 2003;Kraljic & Brennan 2005;Watson & Gibson 2005; among others). For example, when a phrase could attach higher or lower within a sentence, like the temporal Prepositional Phrase on Monday in (1), a prosodic boundary (at #) between the ambiguously-attached phrase and the low attachment site favors high attachment into the Verb Phrase (VP) headed by claimed: (1) Sandy claimed that Bill had arrived # on Monday.
This result illustrates a natural mapping between prosodic and syntactic structure. The listener needs to decide whether or not to close a syntactic phrase, the VP headed by arrived. A prosodic boundary tells them that a prosodic phrase ends at that position, while the lack of a prosodic boundary means the prosodic phrase continues. And although prosodic and syntactic phrase edges do not always coincide, having their final boundaries at the same place in a string of words is favored (Nespor & Vogel 1986;Speer, Kjelgaard & Dobroth 1996;Selkirk 2000). It is intuitively plausible that the two levels of grouping of words, syntactic and prosodic, should be related.
The question of whether information structure, as cued by the presence and position of focus, can affect attachment has been less well studied, with only a handful of papers addressing it (Schafer et al. 1996;Lee & Watson 2011;Lee & Garnsey 2012;Carlson & Tyler 2018;Carlson & Potter in press). Most of this research has found that accented heads of attachment sites do attract a later modifier. For example, people hearing (1) with a contrastively accented first verb, claimed, are more likely to interpret the modifier as attaching high than those hearing a contrastively accented second verb, arrived (Carlson & Tyler 2018). Carlson & Potter (in press) showed that preceding wh-questions, which also indicate focus position within a sentence, can affect attachment similarly, increasing attachments to the constituent focused by the wh-question. These effects all support the Focus Attraction Hypothesis in (2), first formulated in Schafer et al. (1996: 149). The underlying motivation for the hypothesis is that perceivers prefer for ambiguously attached material to modify more important information in the sentence rather than less important information.
(2) Focus Attraction Hypothesis: 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.
But questions remain about how focus influences attachment. We will explore in this project whether and how focus particles like only can influence attachment. Because the position of only affects a sentence's information structure, we suspect that it should affect the attachment of an ambiguous modifier. But there is some doubt about what predictions the Focus Attraction Hypothesis in (2) makes about the impact of only on attachment. The hypothesis specifically concerns focused attachment sites drawing modifier attachment. However, only does not, itself, cue focus or bear focus. The semantics of the focus particle, its structural position, and the position of a focused element with which it associates all impact the final truth conditions and information-structural interpretation (Rooth 1992;Beaver & Clark 2008). Thus, a sentence containing only has multiple information-structural cues which conspire to yield the final information-structural parse.
The central question we address is how these cues impact attachment. Will only draw attachment to the syntactic domain it c-commands, its scope? Or will the position of a focused element below only be where attachment is drawn-which the current Focus Attraction Hypothesis would seem to predict?
We find that the scope of only draws attachment, as does a focused element below it, independently. We therefore will present a revision of the Focus Attraction Hypothesis, generalized to cover the effects of other information-structural cues beyond foci. Ultimately, this research aids in clarifying how different aspects of information structure affect sentence processing generally, including what levels of representation they influence. Schafer et al. (1996) found that relative clauses which could attach to one of two nouns (propeller or plane in (3)) were more likely to attach to whichever noun was contrastively pitch accented.

Pitch accent effects on attachment
(3) … the propeller of the plane [which the mechanic was so carefully examining] RC They explained these results with the Focus Attraction Hypothesis, on which the focus that accents indicated led to the modifier being attracted to a particular attachment site.  found a similar result in a different ambiguity. In sentences like Jimmy comforted the girl that he had insulted [at the end of the party], the final bracketed modifier could attach to the first verb (comforted) or the second verb (insulted). The sentences were produced with both verbs bearing H* accents, and then either the first or second accent was replaced with an F0 contour linearly interpolated between surrounding points. The modifier was drawn to whichever verb retained an accent, though this effect was smaller than effects of prosodic boundaries after the first or second verb.  also attributed their results to the Focus Attraction Hypothesis.
By contrast, Lee & Watson (2011), studying ambiguously-attached relative clauses, argued that salience is the primary factor explaining effects of L+H* accents on attachment. They found that longer relative clauses showed larger changes in attachment rates due to accent than shorter ones. They reasoned that the salience of an accented attachment site aided perceivers more under conditions of greater processing load such as with longer relative clauses; oddly, though, the complexity of relative clauses did not modulate accent effects. In their final experiment, where some questions probed aspects of the sentence meaning besides relative clause attachment, participants preferred to provide the accented noun as an answer, even when it was an irrelevant and wrong answer to the question. Lee and Watson thus argued for a version of their Salience Hypothesis based on post-sentence selection mechanisms: perceivers prefer to provide accented words as answers to any question, about attachment or otherwise, because of their salience.
Their results are nevertheless broadly consistent with the Focus Attraction Hypothesis. Carlson & Tyler (2018) examined the generality of accent effects on attachment in three different attachment ambiguities (4).
(4) a. Sammy heard that Bill had called after the meeting. b. Jimmy comforted the girl that he had insulted at the end of the party. c. Paula phoned a friend from Alabama.
All three structures showed effects of L+H* accent placement on attachment, and these effects were independent of prosodic boundaries, though boundaries usually affected attachment These results taken together show evidence that focus impacts attachment. In the current project, we examine whether the Focus Attraction Hypothesis predicts the effects of only on attachment, both alone and with pitch accents.

Focus markers: accents and only
Pitch accents in English cue focus, marking the discourse status of an accented element (Pierrehumbert & Hirschberg 1990;Rooth 1992;Schwarzschild 1999). Information structure can also be marked through the interaction of focus and focus-sensitive constructions, such as question-answer pairs, the operator only, negation, and modals, among others (Rooth 1985;1992;Hajičová, Partee, & Sgall 1998). In focus-sensitive constructions, the presence and location of focus materially affects the truth-conditional meaning and/or contextual felicity of that utterance. Here, we concentrate on the focus particle only.
Only is sensitive to focused elements within its scope, which is roughly the syntactic material c-commanded by the particle The meaning of an utterance with only depends on the location of focus within its scope, and on the exhaustivity conveyed by the operator (Rooth 1992;Beaver & Clark 2002;2003;Beck 2016). Consider (5a-b), from Rooth (1985), against the context in (5). Answers (5a-b) only differ in the location of focus (indicated by capitalization), yet vary in their truth value: (5a) is false in this context, as Mary also introduced Tom to Sue, while (5b) is true, since the only person Mary introduced Bill to was Sue. Or it could indicate a wider focus, like the embedded clause (8b) or the higher VP (8c), as per Selkirk (1984)'s focus projection rules. The focus alternatives generated in each of these cases are of different syntactic sizes. With only present, though, its position would delineate the largest syntactic domain within which alternatives could be constructed.
It also affects the resolution of ambiguous ellipsis structures like gapping and bare argument ellipsis (Stolterfoht et al. 2007;Carlson 2001;2013;Sauermann, Filik, & Paterson 2013) by focusing a potentially contrastive NP. But these results do not guarantee that only would also affect the attachment of an ambiguous modifier. One reason for this research, then, is to discover whether focus particles do indeed affect attachment, and if so, how.
In the following experiments, we explore effects of the focus particle only on attachment, alone or along with pitch accents. The studies primarily use ambiguous sentences in which a final time-based adverbial phrase could attach into either the VP in a complement clause (in Susie heard that Bill had called on Monday, the one headed by called) or into the higher VP (headed by heard). Experiment 1a is a written questionnaire varying the position of only before either verb in order to establish whether this affects attachment. Experiments 1b-c then explore the likely focus structures entertained in the presence of only and contrastive accents, first through a written completion study and then an auditory questionnaire. In Experiments 2a-b, we use auditory questionnaires to study effects of only on attachment, with and without contrastively accented verbs; Experiment 2a tests these factors with a prosodic boundary before the adverbial, and Experiment 2b tests them without the prosodic boundary. Finally, in Experiment 3, we explore the effects of contrastive accents on multiple verbs as well as only before the first verb.
The overall pattern of results suggests that only does affect attachment, that it indeed has stronger effects than contrastive accents on attachment, and that it draws attachment into the highest syntactic domain under the particle. We propose revising the Focus Attraction Hypothesis in order to account for these results.

Experiment 1a
This written questionnaire tested whether the presence and position of the focus particle only would affect attachment. The Focus Attraction Hypothesis predicts that only before a VP should draw attachment of the ambiguous modifier to that VP, leading to more high attachments with only before Verb1, but only if people take the particle to specifically focus the nearest verb. If focus occurs elsewhere within the scope of only, then it is not clear what the Focus Attraction Hypothesis predicts. If focus is taken to be on the lowest non-adjunct constituent in the sentence, consistent with the default focus pattern for English (Selkirk 1984;Cinque 1993), then it will be on the lower verb, which should favor low attachment regardless of the position of only.

Materials
The experimental items were 20 sentences as in (9), with matrix verbs taking a complement clause followed by an ambiguously attached final PP. We placed only before either the matrix or the embedded verb.
(9) a. Sammy only heard that Bill had called during Monday's emergency meeting. b. Sammy heard that Bill had only called during Monday's emergency meeting.
These sentences were adapted from Carlson and Tyler (2018), with the final time phrases lengthened to an average of 10.4 syllables to encourage high attachment (see Appendix A).
We expect that the long final phrases will be read with implicit prosody placing them in their own prosodic phrase (Fodor 1998;Hirose 2003;Breen 2014), and their consequent prosodic separation from the nearest attachment site will increase high attachments Clifton et al. 2002;2006).
Each experimental sentence was followed by a two-choice comprehension question as in (10).

Procedure
The 20 experimental sentences, 52 fillers and 44 unrelated experimental items were combined in a latin square design to create 6 lists of 116 items. The lists were pseudo-randomized so that items from the same experiment or condition never appeared consecutively.
Participants from AMT clicked on a link to a Qualtrics survey. In Qualtrics, they read a short introduction explaining the experiment, answered the survey, and answered some demographic questions. Most participants completed the experiment within 30 minutes. They were provided with a code to enter into AMT to qualify for payment.

Results and Discussion
The data was 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 Verb1 only condition as = 0.5, and the Verb2 only condition coded as = -0.5, and was introduced into the model as a fixed effect. The model included the maximal random effects structure justified by the design that would converge (Barr et al. 2013): random intercepts for frame and for participant, and random slopes for participant by position of only. 1 The model results are shown in Table 1. 1 The specific R syntax was (dataModelExp1a = glmer(Response~ OnlyPositionContrastCoding + (1 + OnlyPosition-ContrastCoding|participant) + (1|frame), data = dataSet, family = "binomial")).
There were significantly more high attachment responses in the Verb1 only condition (61.5%) than in the Verb2 only condition (26.6%; β = 2.16 ± 0.31, z = 6.87, p ≤ 0.001). 2 Indeed, in this written questionnaire, the position of only strongly affected the attachment of the final adverbial. Only before the first verb led to an increase of more than 30 percentage points in high attachment answers, vs. the condition with only before Verb2. One possible explanation is that the focus particle was taken to indicate focus on the nearest verb, which drew the attachment of the modifier to the relevant VP. On this explanation, these conditions would be analogous to those of Carlson and Tyler (2018), in which contrastive accents on the first or second verb drew attachment into the VP headed by the accented verb. These results would also follow neatly from the Focus Attraction Hypothesis.
But is it likely that participants took the verb following only to be focused? The particle only obligatorily associates with focused elements somewhere within its scope (Beaver & Clark 2002;2003;Rooth 1992;Beck 2016). However, in these experimental items, focus position is not explicitly cued, so there are many potential elements that only could associate with.
In (11a-b) we show the focus position (in brackets) and an alternative in parentheses, if participants interpret only as associating with the nearest verb. But when only precedes the matrix verb, it could associate with other focused elements, like the complement clause or the entire matrix VP (11c-d). Given the range of potential foci that only could associate with, it is unclear whether participants actually interpreted the nearest verb as being focused. In addition, English default focus is on the rightmost element of a clause (Cinque 1993;Selkirk 1995;Arregi 2016 only is unlikely to have been the first choice of perceivers for sentences like (9). To explore which element participants would actually take to be focused under only, we conducted a completion experiment.

Experiment 1b
This experiment explored what focus structure readers would apply to sentences with complement clauses in the presence of only. We asked participants to complete complex sentences that contained only before either verb, with a negated clause showing a contrast set. We expected that participants would offer completions showing what they understood to be focused in the preceding material.

Materials
The materials were adapted from the 20 sentences in Experiment 1a without the final adverbials.
The focus particle only appeared before the first or second verb, as in (12). 3 (12) a. Kathie only claimed that Alex had lied, she didn't ____________________. b. Kathie claimed that Alex had only lied, he didn't _____________________.
The sentences were followed by a contradicting clause starting with the pronoun he or she. The original sentences were standardized so that first-and second-clause subjects were of different genders within each item (see Appendix B). The pronouns disambiguated the intended level of the continuation as the clause containing only. The contradiction paradigm was chosen because the exhaustivity of only rules out alternatives to the stated event (Rooth 1992;Kiss 1998;Umbach 2004), and was modeled on the procedure in Hurley & Bishop (2016). What part of the original sentence participants contradicted would indicate where within the scope of only they interpreted focus.
In addition to the 20 experimental items, participants also saw 80 filler items. Twenty-eight of the fillers were ambiguous (with ambiguities related to causal relations or ellipsis) and 52 were unambiguous sentences with various structures. Across the experiment, 45 items were completions, with the rest forced-choice or rating items.

Participants
Twenty-eight self-reported American English-speaking people participated in the experiment on Qualtrics through a link from AMT, and were paid $4.00. The data of 3 participants were discarded for providing four or more ungrammatical or incoherent completions. Any additional ungrammatical answers were also removed from the data.

Procedure
The 20 experimental completion items were combined in a latin square design with 80 filler items to create two pseudorandomized lists of 100 items. The lists were presented on Qualtrics after recruitment of participants through AMT as in Experiment 1a. Most participants completed the experiment within 30 minutes.

Results and Discussion
First, we examined the range of verbs provided in the answers. With only before Verb1, the most frequent completions included either say or know, and those verbs appeared in nearly 50% of completions. This makes sense, since the initial verbs were verbs of speech, thought, or discovery taking tensed complement clauses, and say and know are common verbs of that type. An average of 5.9 different first verbs per item appeared in completions.
With only before Verb2, the most frequent verbs varied more, including take, have, leave, and get. The most frequent verbs appeared in only 30% of the completions, and on average, 8.7 different verbs appeared in completions per item. This reflects the higher variability of the original second verbs, which did not all fit within a semantic or syntactic class and included intransitives, passives, be-verbs with adjective complements, and particle verbs like step down.
We then examined contrasts between the contradiction clause and the previous sentence: did the completion just contrast with one of the verbs, with the whole VP, or with the complement (if any) of the verb (Complementizer Phrase, CP)? The results of this analysis are in Table 2.
With only before Verb1, 99% of contrasts were with the first verb, the first VP, or its complement; with only before Verb2, 96% of contrasts were with elements of the lower clause. Completions with other contrasts were so rare as to plausibly be mistakes. Thus the initial pronoun and particle position successfully conveyed at what level to complete the contradiction.
The contrasts were coded by both authors, with disagreements resolved by discussion.
Examples of contrasts in the condition with high only are shown in (13). In (13b), we took claiming and proving to be distinct acts, so this was classified as V1 focus (the relegation of all other material to a pronoun clarified the lack of additional contrasts). In (13c), knowing is different from claiming, and the CPs also contrast in content, so we classified this as having focus on the entire VP1. In (13d), though, claiming and saying are quite similar, both being verbs of speech, with say just less specific than claim. Therefore we categorized the CP following say as what contrasts with the original embedded clause (lying and stealing jewelry being distinct).
As  However, the attraction effect from only in Experiment 1a was robust: the VP preceded by only was much more likely to be modified by the adverbial than the other VP. This is interesting; we might have predicted that high only would not draw a modifier to VP1 because it was consistent with focus at so many positions. Instead, only was apparently even more effective than accents in drawing attachment to a VP (cf. accent effects in Carlson & Tyler (2018) which were in the range of 5%).

Experiment 1c
Experiment 1b showed that the basic sentences from Experiment 1a, without the final adverbial, In this and following auditory experiments, we primarily vary L+H* accents. 4 Schafer et al. (1996) found that H* accents affected relative clause attachment in the same direction as L+H* accents, but were less effective. Related work by Lee & Watson (2011)

Materials
The 20 sentences from Experiment 1a, without final adverbials, were recorded with contrastive L+H* accents on Verb2. The position of only before Verb1 or Verb2 was the sole difference between the conditions (15). (15) a. Sammy only heard that Bill had CALLED. b. Sammy heard that Bill had only CALLED. Table 3 show the average F0 and duration properties of the sentences,

Acoustic measurements in
and Figure 1 shows a sample prosodic contour. In this and the other auditory experiments reported here, audio files were recorded in .wav format and converted to high bitrate .mp3s (320kbit/s) using ffmpeg 5 . The acoustic measures and pitch tracks show that the two conditions had similar contours, with the final verb (Verb2) always reaching an F0 peak higher than the F0 on the first verb and usually higher than the subject. The final verbs also had fairly long durations. The height of the F0 peak on the initial subject varied, with (a) conditions (V1 only) averaging slightly higher subject peaks than (b) conditions (V2 only). We do not believe this variation in initial H* accents affected sentence interpretation.
Each recorded sentence was followed by a two-choice comprehension question as in (16).  The answers explicitly labeled which event the exhaustivity of only applied to: a hearing event, with the first VP as the particle's domain, or a calling event, with the second VP as the domain.
The high answer assumed the widest focus consistent with Verb2 accent: the entire first VP including complements. We expected that condition (a) with only high would get more high answers than low answers, with low answers preferred for condition (b). If instead a narrower focus given Verb2 accent was preferred, we would expect low responses to dominate for both conditions.
Besides the 20 experimental items, participants also heard 94 filler items with a range of structures, including 40 pure fillers, 20 conjoined possessive phrases, 20 ambiguous VP ellipsis sentences, and 24 stripping ellipsis or it-clefts. These fillers involved single-talker and two-talker dialogues, and participants were asked to either rate the naturalness of, or to disambiguate, the utterances.

Participants
Forty-five native American English-speaking participants began the experiment. Forty-one of them completed it and were paid $3.50. The data from three participants who achieved less than

Results and Discussion
The data was 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 interpretation (1 = high interpretation; 0 = low interpretation). The independent variable was contrast coded, with the V1 only condition as = 0.5, and the V2 only condition coded as = -0.5, and introduced into the model as a fixed effect. The model included the maximal random effects structure justified by the design that would converge (Barr et al. 2013): random intercepts and random slopes for frame and for participant by position of only. 6 The model results are shown in Table 4. 7 There was a clear effect of the position of only in these sentences, with 77.5% high answers for the condition with high only but 33.2% high with low only (33.2%; β = 2.14 ± 0.34, z = 6.31, p ≤ 0.001). This is reassuring, showing that only's position affects the interpretation of an utterance by changing the alternatives considered by the listener. 6 The specific R syntax was (dataModelExp1c = glmer(Response~ OnlyPositionContrastCoding + (1 + OnlyPosition-ContrastCoding|participant) + (1+ OnlyPositionContrastCoding|frame), data = dataSet, family = "binomial")). 7 High answer rates by item ranged from 30% to 68%; items 1 and 11 had relatively low rates of high answers, and items 14 and 20 had high rates. High answer rates by subject ranged from 25% to 90%.   (17). (17) Which follow-up sentence gives you new information, instead of presenting information implied by the sentence you will hear? a. She didn't hear that he had done anything else. b. According to her source, he hadn't done anything else.
In that pilot, we found no significant effect of only position. We suspect that the complexity of the question and answers obscured any possible effect. The formulation of the question and answer pairs in Experiment 1c was more straightforward, allowing the observed effect to surface.
Overall, Experiments 1a-c have shown several things. First, the focus particle only does affect attachment preferences, with a relatively strong effect in the initial written study, Experiment 1a.
We do not yet know how it might interact with accents or whether it works in auditory processing as well. Second, placing only before a predicate is compatible with many different focus structures (Experiment 1b), so Experiment 1a's effects were unlikely to be due to readers usually placing narrow focus on the adjacent verb. This makes effects of only different in their mechanism from accents on the heads of attachment sites. Third, placing only before different predicates while holding accent position constant did affect interpretation (Experiment 1c) through differences in the domain over which alternatives are calculated. These findings provide a background for interpreting the upcoming auditory studies, which vary the presence and position of both only and pitch accents.

Experiment 2
To learn more about effects of only on attachment, we turned to auditory studies allowing us to ask how the focus particle might interact with accents.

Experiment 2a
This auditory questionnaire experiment tested conditions with only placed before the first or second verb, plus the same conditions with the corresponding verb contrastively accented.
The first question was whether the position of the focus particle would draw attachment to each verb, and the second was whether accenting the same verb would intensify this effect. On the original Focus Attraction Hypothesis, we would predict that only before a verb should draw attachment of the ambiguous modifier to that verb only if people take the particle to specifically focus the nearest verb. However, only can associate with any focused phrase within its scope (as Experiment 1b showed). With contrastive accent on the same verb preceded by only, then, we might predict that attachment to the indicated phrase would be increased: more high attachments with only and accenting of Verb1, and more low attachments (fewer high attachments) with only and accenting of Verb2. Alternatively, the focus particle position could unambiguously indicate focus within its scope, making accenting of the same position redundant: this would lead to no effect of accent.

Materials
Twenty experimental items (18) Table 5, and sample pitch tracks in Figure 2. The audio files were recorded in .wav format and converted to high bitrate .mp3s (320kbit/s) using ffmpeg.
The acoustic measurements and pitch tracks show that each verb, when contrastively accented, had average F0 peaks over 30 Hz higher and average durations at least 30 ms longer than when unaccented. All conditions had a low F0 and a pause of over 150 ms before each final adverbial, consistent with a L-ip boundary. Conditions without contrastive accents showed general declining F0 from the first subject to the boundary. The main additional variation in the prosodies was on the initial subject, which was higher and longer in conditions without contrastively accented verbs. This resulted from a different metrical structure: the most felicitoussounding contour with the first verb less prominent had the initial subject more prominent. The pitch tracks for conditions (a/c) look as if the initial subjects had contrastive L+H* accents, though we suspect this derives from each sentence starting at a relatively constant F0 range and then moving to a delayed H* peak in subjects with initial stress. (Also, Carlson & Tyler 2018 showed that accenting subjects did not affect attachment preferences for this sentence type.)    Each recorded sentence was accompanied by a two-choice comprehension question as in (19).

Condition
What happened? a. Sammy heard something on Monday.
(high attachment) b. Bill called on Monday.
(low attachment) Participants also heard 79 filler items, including 20 unambiguous filler sentences with a range of structures, followed by forced-choice interpretation questions; 20 ellipsis sentences with two possible completions; 24 sets of conjoined clauses in which participants were asked to pick from two sound files the one which best expressed a certain meaning; and 15 sentences with ambiguous pronouns and two possible completions.

Participants
Fifty-nine self-reported native speakers of American English were recruited and paid $3.75 through AMT. The data from 7 participants was excluded due to less than 90% performance on unambiguous attention checks or more than 2 items skipped. This left 52 participants whose data were analyzed.

Procedure
The 20 experimental sentences were combined in a latin square design with 79 filler items to create 8 pseudo-randomized lists of 99 items. The experiment was presented using AMT and Qualtrics. The questions and answers (or completions) for each item were visible on screen along with an audio player which participants clicked to hear the sound files. There was no time limit on answers. Most participants completed the experiment within 20 minutes.

Results and Discussion
The data was 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 variables were presence of the focus operator only before the matrix verb (contrast coded as = 0.5) or the embedded verb (coded as = -0.5) and the presence of a contrastive pitch accent on the verb adjacent to only (coded as = 0.5) or the absence of a contrastive pitch accent (coded as = -0.5).
These variables were each introduced into the model as fixed effects, and the model included the maximal random effects structure justified by the design that would converge (Barr et al. 2013): random intercepts for frame and for participant, and random slopes for participant by only position. 8 The model results are shown in Table 6. 8 The specific R syntax was (dataModelExp2a = glmer(Response~ OnlyPositionContrastCoding * AccentPresenceCon-trastCoding + (1 + OnlyPositionContrastCoding|participant) + (1|frame), data = dataSet, family = "binomial")).
The results are shown in Figure 3. Conditions with only before the first verb received more high attachment responses than those with the particle before Verb2, as shown in a main effect of only height (β = 1.75 ± 0.29, z = 6.02, p < .001), and the presence of an accent increased high attachments slightly regardless of position (β = 0.39 ± 0.16, z = 2.64, p = 0.02). There was no interaction between these factors (p ≥ 0.78). 9 The major finding is that the position of only strongly affected attachment of the adverbial in an auditory study, with a change of over 20 percentage points in attachment responses. Only drew attachment to the VP whose verb it preceded. In fact, the effect of focus particle placement was larger than expected given previous results with accents (e.g., Carlson & Tyler 2018), which involved shifts of less than 10 percentage points in responses. These results thus replicate the 9 High attachment rates by item varied between 35% and 69%. Items with low rates of high attachment included 9, 11, and 19; those with high rates were 3, 8, and 20. High attachment rates by subject ranged from 5% to 95%.   Verb2 results of the written Experiment 1a in an auditory domain, and suggest that accents have a lesser effect on attachment than focus particles.
Accenting either verb slightly increased high attachments. But we expected an interaction between the two factors, with an accent on the second verb increasing attachment to the lower VP and thus decreasing high attachments (or no effect of accent).
One possible explanation for the lack of interaction is that the prosodic boundary interfered with the perception of accents. All conditions had an intermediate phrase boundary after Verb2.
Perhaps the presence of this boundary influenced the perception of the preceding word. Accents and prosodic boundaries both affect the duration and pitch contours of words, and words at the ends of prosodic phrases can be heard as accented or vice versa (Byrd & Riggs 2008;Lee & Garnsey 2012

Experiment 2b
This experiment was a version of Experiment 2a with the sentences re-recorded without prosodic boundaries before the final adverbial. We expected this to lower the overall rate of high attachments, as a pre-PP boundary favors high attachment Watson & Gibson 2005). We asked whether the absence of this prosodic boundary would change the effects of accent placement compared to Experiment 2a.

Materials
This experiment contained the same 20 items as Experiment 2a (see Appendix C), recorded in the four conditions in (20).  Table 7 and an example set of pitch tracks in Figure 4.    The prosodic contours were very similar to those in Experiment 2a, with the same large F0 and duration differences between contrastively accented and unaccented verbs. The primary difference was that there was no pause between the second verb and the adverbial, and the Verb2 durations averaged over 150 ms shorter than the Experiment 2a versions. These two properties suggest a lack of an ip boundary before the adverbial. As before, the initial subject measurements show higher F0s and longer durations in the conditions without accented verbs.
Each recorded sentence was followed by a comprehension question and two answers as in (21). The audio files were recorded in .wav format and converted to high bitrate .mp3s (320kbit/s) using ffmpeg. Participants also heard 79 filler items, including 24 ambiguous comparative ellipsis sentences, 20 sentences containing conjoined clauses with ambiguous causality, 23 partial sentences with two possible completions, and 12 unambiguous filler sentences.

Participants
Fifty-nine self-reported American native English-speaking participants were recruited through AMT and paid $3.12 for their participation. Data from the 52 participants who had over 90% performance on unambiguous attention check items was analyzed.

Procedure
The 20 experimental sentences and 79 filler items were combined into 12 pseudo-randomized lists of 99 items in a latin square design. The questions and answers (or completions) for each item were visible on the screen along with an audio player which participants clicked to hear the sound files. There was no time limit on answers. The study was presented in Qualtrics and most participants completed the experiment within 20 minutes.

Results
The results are shown in Figure 5.
The data was 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  Table 8. 10 The specific R syntax was (dataModelExp2b = glmer(Response~ OnlyPositionContrastCoding * AccentPresenceCon-trastCoding + (1 + OnlyPositionContrastCoding||participant) + (1|frame), data = dataSet, family = "binomial")).  Overall, there were many fewer high attachment responses than in Experiment 2a, because the lack of a prosodic boundary after the second verb favors low attachment Watson & Gibson 2005). 11 As in Experiment 2a, the position of only significantly affected attachment, as conditions with the particle before Verb1 had more high attachments: significant main effect of only (β = 1.43 ± 0.31, z = 4.55, p < .001). Also as in Experiment 2a, accenting either verb slightly raised the rate of high attachment: significant main effect of accent presence (β = 0.54 ± 0.19, z = 2.85, p = .005). There was no significant interaction between the two factors (p > 0.43), so accenting Verb2 still did not increase low attachments as was expected.

Discussion
This experiment replicated Experiment 2a very closely. Placing only before either verb focused that VP, leading to a change in attachment preferences of 15 or more percentage points. The significant effect of accent position was small and consistent: accenting either verb, with only before that verb, slightly increased attachments to the first VP. The only difference between the experiments was in the overall level of high attachments, as the prosodic boundary before the PP increased high attachments in Experiment 2a as expected.
Together, Experiments 2a-b revealed effects of the presence and location of both only and contrastive pitch accents on attachment preferences. The focus particle only likely does not cue the verb it precedes as focused, since Experiment 1b showed that only is compatible with multiple alternative focus structures. Nonetheless, only still drew attachment to the VP that it immediately preceded. This suggests that the resolution of attachment ambiguities can be guided not just by direct cues for focus, but also by other information-structural cues like only.
However, Experiments 2a-b also showed no interaction between the position of only and accent presence. The location of only drew attachment, but, unexpectedly, accents adjacent to only increased high attachment slightly whether the particle and accent were low or high. Prior results led us to expect that accenting Verb2 should lead to fewer high attachments than when only marks low focus alone (Carlson & Tyler 2018): we expected an interaction between accents and only, with the accents drawing attachment in different directions depending on their position.
It's possible that even without a prosodic boundary in the recordings, Verb2 accents gave the impression of a prosodic boundary (due to the falling F0 after the high peak), which then increased high attachments (see Hirose (2019) on ambiguity of a particular acoustic F0 movement; also Lee & Garnsey (2012) and Jun & Bishop (2015)). This is the reverse of the perceptual illusion suggested after Experiment 2a. There, we hypothesized that a prosodic boundary after Verb2 11 High attachment rates by item varied from 10% to 46%. Items with relatively low high attachment rates were 4, 6, 9, and 19, and ones with relatively high rates were 8, 14, and 20. High attachment rates by subject varied from 0% to 65%. led to perceived accentuation of Verb2, but Experiment 2b ruled that out. Here, we suggest that the accent on Verb2 led to a perceived boundary after it, whether or not it was present. That perceived boundary then raised high attachments in this condition compared to ones without a Verb2 accent.
Another possibility is that a confound in the Verb2 only no-accent conditions (e.g., (18/20c):

Experiment 3
In Experiment 3, we explored a fuller range of conditions, with single and multiple focus markers, in order to better understand how they might interact. In particular, we compared sentences with and without only before Verb1 in three prosodic conditions: with Verb1 contrastively accented, Verb2 accented, or both verbs accented. We predicted that the conditions with only high would generally show more high attachments than those without. The different accent conditions tested whether the highest cue to focus position has an invariant effect, with more high attachments whenever a contrastive accent is on Verb1; or whether multiple cues matter, and conditions with both verbs contrastively accented should differ from conditions with either accent alone.
Finally, if high only and low accent are interpreted together, then this condition might show a low attachment preference similar to the low accent condition without only.

Materials
This experiment tested ambiguous attachment sentences in six different conditions, as in (22).
The first three varied the presence of contrastive accents on Verb1, both verbs, or Verb2. The next three conditions had the same accents as (a-c) but also had only before Verb1. All conditions had ip boundaries before the final phrase. To accommodate six conditions, we increased the set of sentences by four for a total of 24 (see Appendix C). Acoustic measurements to substantiate this description are shown in Table 9 and sample pitch tracks are shown in Figure 6. Measurements for accented verbs are shown in bold. The audio files were recorded in .wav format and converted to high bitrate .mp3s (320kbit/s) using ffmpeg. The acoustic measurements and pitch tracks show the same critical properties as those for Experiments 2a-b: when each verb was contrastively accented, its average F0 height was at least 80 Hz higher and its average duration at least 70 ms longer than when unaccented, whether the condition had both verbs accented or only one. Low F0 levels at the end of Verb2 and pauses of over 150 ms support the presence of an ip boundary before each adverbial. As in prior experiments, the F0 averages on the initial subject were higher in conditions with Verb1 not contrastively accented than in those with it accented, for metrical reasons.
Each auditory sentence was accompanied by a written two-choice comprehension question, as in (23). dialogues with it-clefts or ellipsis for rating of pronoun binding, and 30 pure fillers of different structures). Participants were asked to either rate the acceptability of these items, or indicate which of two interpretations was more likely.

Participants
Sixty-five self-reported American native speakers of English were recruited through AMT and the data of 44 participants who passed attention checks was analysed. The 61 participants who completed the experiment were paid $5.00 for participation.

Procedure
The 24 experimental sentences were combined with 78 additional items into twelve pseudorandomized lists in a latin square design. The questions and answers (or rating scale) for each item were visible on screen while participants heard the sound files. There was no time  Table 9: F0 averages for items in Experiment 3 in Hz and duration averages in ms by condition; accented verb measurements shown in bold.
limit on answers. The study was presented using the Northwestern University survey tool as in Experiment 1c and most participants completed the experiment within 30 minutes.

Results and Discussion
The data was 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 variables were the presence of only before the matrix verb (contrast coded as = 0.5) or its absence (coded as = -0.5) and the location of contrastive pitch accent(s), which was helmert coded.
This involves coding the three levels in this factor as two fixed effects. In one, the low accent condition (helmert coded as -0.5) was contrasted with the high and both accent conditions (each helmert coded as 0.25), and in the second, the high accent condition (helmert coded as -0.5) was contrasted with the both accent condition (helmert coded as 0.5). The model included the maximal random effects structure justified by the design that would converge (Barr et al. 2013): random intercepts for frame and for participant, and random slopes for participant by only position and accent position. 12 The model results are shown in Table 10. 12 The specific R syntax was (dataModelExp3 = glmer(Response~ OnlyPositionContrastCoding * (HighContrastHelmert-Coding + LowContrastHelmertCoding) + (1 + OnlyPositionContrastCoding * (HighContrastHelmertCoding + Low-ContrastHelmertCoding)|participant) + (1|frame), data = dataSet, family = "binomial")).  The results of this experiment (in Figure 7) showed that both focus markers affected attachment preferences. 13 There was a significant effect of the presence of only: only before Verb1 increased high attachments compared to conditions without the particle (β = 2.27 ± 0.30, χ² = 7.54, p < 0.001). As in earlier experiments, this was a robust effect, with a change of more than 30 percentage points in high attachment. There was also a significant effect of contrastive accent position: low accents led to fewer high attachments than conditions with high or both accents (β = 0.65 ± 0.29, χ² = 2.34, p = 0.03). There was a marginal effect of high vs. 13 High attachment rates by item ranged from 31% to 68%. Items with relatively low rates were 9 and 15; items with relatively high rates were 8, 20, and 22  only is what draws attachment. Instead, this condition showed over 50% high attachments, as did all three conditions with only before Verb1. This demonstrates that it is the position of the focus particle and its syntactic scope, not the later focused element with which it associates, that draws attachment.
These results do not show the odd pattern found in Experiments 2a-b, in which accenting Verb2 appeared to increase high attachments, because different conditions are being compared.
In Experiment 3, all conditions had at least one contrastive accent in them. In that context, the conditions with Verb2 accent alone showed more low attachment than any other accented conditions.

Conclusions
These experiments show that the information structure of a sentence can influence its syntactic structure, in addition to the known effects of prosodic phrasing on attachment. Prior work concentrated on how pitch accents, as cues of focus, affected attachment (Schafer et al. 1996;Lee & Watson 2011;Carlson & Tyler 2018), while Carlson & Potter (in press) showed that preceding wh-questions also affected attachment by focusing specific phrases in a target sentence. Here we found that focus particles affect attachment, too, and explored how they interact with accents. 14 The statistical patterns are the same if reversed helmert coding is used, so that high accent conditions are compared to both and low accent conditions: a significant effect of only presence, and of each comparison, but no interaction.
The results of our experiments show that the Salience Hypothesis of Lee & Watson (2011) does not explain the full range of focus effects on attachment, due to its concentration on accents as markers of salience or on choosing accented answers in post-sentence questions. In Experiments 1-3, modifiers were attracted to the highest syntactic domain in the scope of only, even when the nearest verb was not accented. Thus, it seems that the attraction of a modifier to an attachment site is not driven by the perceptual salience of a focused element (within the critical sentence or in forced-choice answers). The Focus Attraction Hypothesis, in spirit at least, is more consistent with these findings, because it identifies focus position as the key factor underlying the attraction effect. Whether focus is cued by a pitch accent, a preceding wh-question, or the focus-sensitive operator only, perceivers preferentially attach a modifier to a potential attachment site that is focused.
However, the Focus Attraction Hypothesis must be modified to actually account for how the focus particle only works to attract a modifier. Since only associates with a focused element within its c-command domain, the most obvious way for the hypothesis to work would have been for a modifier to be drawn to the position of that focused element. Instead, attachment in the presence of only is drawn near the position of the focus particle itself. With pitch accents, what draws attachment is focus on the head of a phrase into which the modifier could attach, but only is neither the head of a phrase nor accented. It is therefore unlikely that the modifier is literally attaching to the particle. Instead, it attaches into the VP preceded by only. The original Focus Attraction Hypothesis from Schafer et al. (1996) stated that "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" (p. 149). This is consistent with how pitch accents (Carlson & Tyler 2018) and wh-questions (Carlson & Potter in press) affect attachment, because each of those mark focused phrases into which the modifier could attach.
For the operation of only, though, we suggest the following revision, italicized below: Interestingly, none of the experiments supported the idea that the lowest focus, the one most local to the ambiguous modifier, would draw attachment more strongly than other focused elements. Several studies on focus expectations in ellipsis sentences have shown a preference for low/late focus in English and German (e.g., Frazier & Clifton 1998;Carlson 2002;Stolterfoht et al. 2007;Carlson, Dickey et al. 2009;Harris & Carlson 2018;among others). This expectation was strong enough to affect interpretation preferences in ambiguous ellipsis sentences despite salient accents (or clefting) placing focus elsewhere. Such focus markers did influence interpretation, but did not overturn a general bias toward a low focus position. Why, then, doesn't the lowest focus marker draw attachment? That would seem most consistent with this other body of results on the surface.
In fact, we think there are important differences between these situations. In ellipsis processing, each remnant of ellipsis needs to contrast with a phrase in the complete antecedent clause (or clauses). So ellipsis experiments all involve people working backwards from the second member of a contrast to find the first member, and using various sources of information to do so, including overt focus markers and expectations about the position of focus. Thus the processor figures out the antecedent material to be copied or reaccessed in place of the ellipsis (Frazier & Clifton 2005;Martin & McElree 2008). But the antecedent material has already been syntactically structured, and its information structure probably computed. In the current experiments, by contrast, no sentence has been completed and the processor is working to finish the process of syntactic attachment. The active phrase needs to be attached into some other phrase that already exists, and focused attachment sites are privileged above non-focused sites.
There are interesting parallels between this research and Gotzner (2019)'s findings. She studied calculation of exhaustive implicatures while manipulating the presence of only and type of pitch accent (H* vs. L+H*). Specifically, given a contrast set of two individuals, and the statement CARL likes herring or Only Carl likes herring, she asked whether people would reject the statement that the other individual liked herring. As expected, the focus particle only led to high rates of exhaustive interpretation (rejecting the statement), as did a L+H* accent or the two together. H* accents also suggested exhaustivity, but less strongly. Crucially, the combination of only plus contrastive L+H* accents did not increase exhaustive interpretations over either marker alone, though it did raise the certainty of the interpretation. Gotzner suggested that once the particle had made its contribution to the interpretation, the intonation had little left to do. This is consistent with the small effects of contrastive pitch accents on attachment when only already indicated a particular VP, in Experiments 2a-b. In two different domains of language processing, then, the focus particle only has a stronger influence on interpretation than pitch accents, despite their similar effects on sentence meaning.
One contribution of this research is that we can see that focus, expressed by accents or other means, overlaps with prosodic boundaries in the type of effects it can have on sentence processing. It would have been reasonable for prosodic boundaries, which mark the grouping of words into prosodic phrases, to affect attachment while focus did not. It is intuitively sensible for processing to favor matching of syntactic and prosodic groupings of words; for a modifier to be attracted to prominent or contrastive information in the sentence is less obvious. But research on focus attraction shows that there is a preference to attach a modifier to a focused attachment site or syntactic domain. This broadens the range of factors that we understand to affect the basic syntactic structuring of a sentence. Conversely, this research also broadens our expectations about what focus can influence in sentence processing. In ellipsis research,  found that prosodic boundaries did not affect the resolution of bare remnant ellipsis while accents on contrasting phrases did, and therefore suggested that different prosodic features affect different types of ambiguities. They proposed a separate roles hypothesis, on which pitch accent effects are restricted to domains like ellipsis processing or pronoun resolution, rather than basic syntactic structure. This hypothesis is therefore shown to be too strong in its differentiation of where accents (and focus) vs. prosodic boundaries have effects.
The Focus Attraction Hypothesis, as modified to account for the effects of the position of only within a sentence, is the most consistent with these results as well as the prior literature (Schafer et al 1996;Lee & Watson 2011;Carlson & Tyler 2018;Carlson & Potter in press). We believe that the reason that focused elements or the highest syntactic domain of the operator only draw attachment is because they are the most important information in the sentence. Given the choice, people prefer attaching a modifier to important information rather than that which is given or backgrounded. A similar idea prompted the Main Assertion Hypothesis of Frazier & Clifton (2005) for ellipsis processing: an ellipsis clause following a sentence boundary is more likely to relate to the main assertion of the antecedent even if that is high in the sentence. Similarly, Traxler & Frazier (2008) used this idea to explain attachment of a PP into a main clause rather than to a subordinate clause.
Finally, there is the question of why the modifier is drawn to the highest predicate in the scope of only, though only can associate with a focused element that is high or low within its scope. In these sentences out of the blue, it appears that the widest syntactic domain within which focus could be located under only is what draws attachment. Perhaps the widest scope is preferred because it satisfies the widest Question Under Discussion (QUD: Roberts 1996). That is, in a discourse context, a sentence with only might be answering a more specific QUD and be consistent with a narrow focus; but these sentences out of context led perceivers to take the highest VP and all of its complements as focused, allowing the broadest QUD to be answered.
Interestingly, though, accent effects on attachment do not seem to involve listeners taking the widest focus possible for an accent position. Accents on the lower verbs in our sentences could have been interpreted as indicating wide focus on the entire sentence. Instead, accents on verbs