Information Structure Preferences in Focus-Sensitive Ellipsis: How Defaults Persist

1.1 Information structure background

Informally, information structure serves to organize content by its role in the discourse. This organization may be achieved in many ways. Typically, discourse-old/given information precedes discourse-new information (Chafe, 1970; Firbas, 1966; Prince, 1992), a trend that governs how discourse participants expect a sentence to unfold (Clark & Haviland, 1977). In English, and many other languages, special or non-canonical organizations are indicated by special or non-canonical markings (e.g., Horn, 1984; Ward & Birner, 2004). For example, the cleft structure in (1b) focuses Bob in a way that a canonical SVO structure (1a) does not.

Similarly, constituents may be marked as prominent or focused via pitch accents: suprasegmental targets that signal the information structural status of an element in the current discourse. In the tones and break indices (ToBI) framework, for example, pitch accents include high or low pitch targets, H* and L*, and their combination, for example, L+H*, with the tone that associates with the stressed syllable of an accented word marked with an asterisk (e.g., Beckman & Elam, 1997; Pierrehumbert & Hirschberg, 1990). Accented words are also often longer in duration and spoken with more intensity than unaccented ones. There is debate within the prosodic community as to whether H* and L+H* accents are different categories or merely different points on a continuum of more or less high, more or less steep H* accents (e.g., Bartels & Kingston, 1994; Ladd, 2008; Ladd & Morton, 1997; Ladd & Schepman, 2003). For convenience, we assume that the L+H* accent corresponds to a category and describe accents of this type as contrastive, but those who do not believe in the category may assume that the accents referred to are on the higher, steeper end of the H* continuum. Non-contrastive H* accents will be lower, more rounded, and less clearly preceded by a low fundamental frequency target.

Most of the semantic literature on focus assumes that a focused constituent contains a pitch accent, though focus may project beyond the accented constituent (e.g., Büring, 2012; Jackendoff, 1972; Selkirk, 1984). Although such accents do not always affect the truth-conditional meaning of a sentence, especially in the absence of focus-sensitive operators like only (Beaver & Clark, 2008; Horn, 1969), they can strongly affect the appropriateness of an utterance. The most prominent and focused element in (2a) is Bob, which corresponds to the information sought by the wh-question Who ate the cake? Here, Bob is appropriately marked as being new (or non-presupposed) information by accenting. However, the same string of words becomes infelicitous as a response to (2) if any of the discourse old/given information (ate the cake) receives pitch accent (2b).

We observe the opposite pattern in (3), where the object is now queried by the wh-question. Together with (2), the example shows that what should be prosodically marked as new information, and hence what appears in focus, depends on the context and the discourse status of the element (e.g., Jackendoff, 1972).

In case context is absent or weak, prominence in English is usually located in the default position for primary (nuclear) sentence accent, on the most deeply embedded constituent at the right edge of a sentence (Chomsky & Halle, 1968; Cinque, 1993; Halle & Vergnaud, 1987; Selkirk, 1984). For instance, the most natural response to the question in (4) is (4b), as the entire sentence is new information in the discourse. Although one might reply with (4a), it would require some degree of accommodation on the part of the addressee, for example, the speaker believes that eating the cake was or should have been available in common ground.

Of course, the tendency to put prominence late in the sentence is not absolute, as pitch accent placement is governed by many factors, including the phonological requirement to place an accent within each phonological phrase, or the desirability of alternating rhythm (e.g., Ladd, 2008; Selkirk, 1995). Thus, the presence of a pitch accent does not necessarily signal that an element is in focus. ¹ As a consequence, the auditory processor may not be certain whether prominence in the acoustic input transparently marks the location of focus within an utterance, and may thus consider default locations for foci during interpretation.

In this project, we propose that default locations for focus interfere with the identification of focus in non-default positions from overtly marked constituents (5).

In other words, we expect that hearers may be tempted to default to standard locations of focus, even when there are overt cues, like pitch accent, to the contrary. This is not to say that hearers are completely insensitive to the pitch accent, or that they wholly rely on information structural defaults. The enduring focus principle may instead reflect an efficient processing strategy which serves to reduce uncertainty in the message by relying in part on conventional mappings between given and new information. We expect that this principle is particularly robust in cases of weak or impoverished context, where the hearer has little independent reason for assigning given and new status besides the sentence itself. We speculate on the origins and mechanism of enduring focus further in the General Discussion section below.

Some prior research is already compatible with the enduring focus principle. Stolterfoht, Friederici, Alter, and Steube (2007) showed event-related potential evidence of an expectation for focus on the object in German, with markers of revision of structure when the focus particle nur “only” or a later contrastive remnant indicated a different focus position. Carlson et al. (2009) found a lasting bias for focus on the object or the latest argument in studies of the processing of sluicing ellipsis sentences. In a series of acceptability rating studies on corrections in spoken dialog, Clifton and Frazier (2016) manipulated whether a correction to a statement (Speaker A: Mary brought the pie) was syntactically parallel with the corrected statement (Speaker B: No, Susie brought the pie) or not (Speaker B: No, the pie was brought by Susie). In addition to an advantage for parallelism, they found an interaction in which non-parallel corrections became more acceptable when the contrastive term (Susie) was situated in a position bearing default focus, for example, the object position. The effect of default focus position was present even when given information (the pie was brought) carried the most prominent pitch accent, suggesting that default syntactic positions for focus continue to influence which constituent is in focus. All of these results are highly compatible with the enduring focus principle, and suggest that focus marking in English is determined by both syntactic position and pitch accent. We now turn to the use of information structure in resolving ellipsis, and the effect that enduring focus plays in interpreting focus-sensitive constructions.

1.2 Focus-sensitive coordination

Constructions with let alone, much less, and certain uses of never mind, exemplify focus-sensitive coordination structures. At a descriptive level, these coordinators appear to coordinate two elements of nearly any syntactic type, except fully inflected clauses (Hulsey, 2008). They are typically produced with contrastive focus on the second conjunct (the remnant: a contractor in (6a)), as well as the element in the matrix clause with which it contrasts (the correlate: a carpenter in (6a)), as shown in (6).

In mainstream dialects of English, the coordinator is licensed in negative environments of various sorts, as well as pragmatically adverse situations, but not in positive environments. ²

Besides their grammatical properties, focus-sensitive coordination structures require comparison along a contextually salient scale for the correlate and the remnant, such that the negation of the correlate strongly implies the negation of the remnant—for example, that John did not call a carpenter strongly implies that he did not hire one (6b). These scales are typically ad hoc and rely heavily on dimensions that can be implicitly ascertained from the context, even when such dimensions are complex (7).

That the entailment pattern is strongly encoded in the ordering of the correlate and the remnant can be seen more directly in the infelicitousness of (8b), which violates the logical entailment between two cookies and five cookies, compared to the minimally different variant (8a).

While Fillmore et al. (1988) used the seemingly idiosyncratic grammatical properties of let alone and related coordinators to argue for a construction-based grammar, recent studies have converged on an ellipsis account of their syntax (Harris, 2016; Hulsey, 2008; Toosarvandani, 2010). For instance, Harris (2016) proposes that the syntax of let alone coordination introduces stripping ellipsis (Frazier, Potter, & Yoshida, 2012; Sailor & Thoms, 2014), in which the second conjunct is a coordinated vP (a VP with an external argument, in the sense of Chomsky, 1995; Kratzer, 1996; Marantz, 1984) or CP (a complementizer phrase) having undergone mandatory ellipsis of all but the focused elements. We assume that there are several sizes of a clause, and refer to the kind of ellipsis in let alone constructions as clausal ellipsis, as in the case of fragment ellipsis (Merchant, 2005; Weir, 2014). In this analysis, example (7a) is the result of moving the remnant coffee to a focused position (FocP) and eliding the remainder of the clause; a simplified bracketed structure is provided in (9).

Since evidence for the idea that let alone structures are distinct from ordinary coordination structures has been documented elsewhere, we simply illustrate the point with two examples from and Harris (2016) and Hulsey (2008), respectively. First, in gapping ellipsis under ordinary coordination (and, or), an active verb may be omitted or expressed in its uninflected form (10a). The uninflected form (Sue eat veal) takes low scope under negation, consistent with the gapping interpretation (Sue—veal), whereas the inflected form (Sue eats veal) coordinates above the negation (Siegel, 1984). In contrast, let alone cannot coordinate a clause with an active inflected verb (10b), suggesting that it coordinates material below negation or at least up to the vP, assuming the subject-internal hypothesis (e.g., Koopman & Sportiche, 1988; Kuroda, 1988).

The second example comes from sprouting, in which the remnant of the ellipsis, for example, caviar in the replacive ellipsis example (11a) or what in the sluicing ellipsis example (11b) lacks an overt correlate in the host clause (Chung, Ladusaw, & McCloskey, 1995). Let alone structures (11c) pattern with cases of clausal ellipsis (11a–b) in allowing sprouting, and not with ordinary coordination, in which sprouting is not permitted (11d).

The purpose of the present paper is not to defend any particular syntactic analysis of this construction, but we believe the close correspondence in its processing preferences to other kinds of ellipsis structures, particularly stripping and sluicing, is important. In particular, we believe that it sheds light on the character of the relations that must be established by the processor when interpreting ellipsis in general, briefly reviewed below.

1.3 Processing ellipsis

Although much can be said about approaches to processing ellipsis structures (e.g., Phillips & Parker, 2014), we assume that there are three basic tasks that the processor must solve in order to interpret any ellipsis involving a remnant (12). For concreteness, we adopt the view that a basic covert syntactic/logical structure exists at the ellipsis site at some level of representation, in keeping with a great deal of syntactic literature (e.g., Merchant, 2001) and experimental work (e.g., Frazier, 2008; Poirier, Wolfinger, Spellman, & Shapiro, 2010).

We use example (13) to illustrate the three tasks in more depth:

Conceptually, steps 2 and 3 each depend on the step that precedes it. For step 2, assuming that remnants must be of the same syntactic type as the correlate, the remnant needs to be parsed before the processor can retrieve an appropriate correlate. ³ Similarly, for step 3, the basic argument position for the remnant trace must be determined before the elided structure can be formulated, and we assume that the processor prefers analyses with parallel structures between clauses, as has been observed for basic conjunction with and (e.g., Frazier, Taft, Roeper, Clifton, & Ehrlich, 1984) and gapping ellipsis (e.g., Carlson, 2001, 2002).

We are particularly interested in step 2, which perhaps relies the most heavily on default resolution strategies. One such strategy was discussed in Frazier and Clifton’s (1998) study of sluicing ellipsis, in which the processor appears to prefer the closest correlate, which was later generalized as the Locality bias (14).

We have remained agnostic about whether the nearest relation should be explicated in structural position, temporal precedence, or linear distance, as these co-vary in our experimental manipulations below. Given the relationship between focus and structural embedding, we suspect that the relationship is structural, although we leave open whether other factors, such as temporal precedence, may independently influence which correlate is paired with the remnant. The experiments discussed below were not designed to tease apart these factors (see comments in the General Discussion section).

In support of the Locality Bias, Frazier and Clifton (1998) observed a reading time advantage for sluices with a viable correlate for the remnant who in object position (someone), even when there was another possibility in subject position (somebody); see also Carlson et al. (2009) and Harris (2015) for congruent results.

In a follow up paper, Carlson et al. (2009) found that the closest noun phrase (NP), the object the co-pilot, was the preferred correlate for ambiguous sluices (16). This was true whenever the prosodic marking was consistent with the most local correlate, as with contrastive accent on the object in (16a), on the verb (talked) in (16b), or on both the subject (captain) and the object noun in (16c). Only when the subject alone received contrastive accent (16d) was the subject NP consistently understood as the correlate.

However, the bias towards subject correlates in subject accent cases (16d) was markedly lower than the bias toward the object in all other cases (about 60% compared to 80%). Carlson et al. (2009) suggest that the default syntactic position for focus continues to exert an influence on interpretation even when overt marking of focus does not support that interpretation. When the overt cues and structural defaults for focus assignment converge, the interpretation is robust. When they diverge, however, the consistency of judgments is weakened considerably, as comprehenders may have to choose between what their ears perceive and their expectations about the information structure of the sentence. In other words, their results strongly support the idea that pitch accent position guides correlate resolution in sluicing, but that the processes that identify prominence are subject to enduring focus (5).

The bias for more local correlates has been corroborated in studies of other types of ellipsis, including focus-sensitive ellipsis (Harris & Carlson, 2016), as well as the closely related case of stripping/replacive ellipsis (Carlson, 2013). In a corpus study of let alone in British and American English, Harris and Carlson (2016) found that approximately 84% of remnants of let alone ellipsis in their sample were associated with local correlates, with a somewhat weaker bias for NP remnants. In two self-paced reading experiments, they found evidence that the Locality bias is also observed during silent online sentence processing. They manipulated the location of a contrastive adjective like nicest to modify the subject (non-local) noun (17b), or the object (local) noun (17a), thereby varying which NP formed a likely semantic/discourse contrast with the remnant containing an adjective like meanest. As expected, reading times on the remnant (the meanest one) and the spill over region (and no one at the hospital) after the let alone coordinator were increased following a non-local contrast compared to the expected local contrast.

In all, a range of experimental studies provides compelling evidence that sentence processing is facilitated when there is a local correlate for the remnant of clausal ellipsis. In the case of let alone, this preference has been confirmed across a wide range of written material, both in corpora and in silent reading. However, previous studies have not explored how the Locality bias might interact with overt focus marking in focus-sensitive coordination structures. This study aims to distinguish between two possibilities regarding the relationship between focus (as realized through pitch accent) and Locality.

The first possibility identifies focus and information structure as the primary source for Locality preferences in clausal ellipsis, and allows a significant role for enduring focus in the processing of contrastive material. This possibility makes two predictions. First, constituents that are accompanied by contrastive pitch accents should be preferred as correlates to the remnant. Second, assuming that default focus positions can intercede in the assessment of focus structure, accented constituents in non-local positions should be understood as correlates less readily and less easily than those in local positions. We assume that the syntactic structure of the antecedent clause indicates the location of default accent, in that the most deeply embedded constituent will receive nuclear pitch accent (Cinque, 1993).

The second possibility is that focus-sensitive coordination structures require contrastive accents for interpretation, because they are explicitly contrastive. Thus, listeners might well be attuned to explicit pitch accents to the exclusion of default preferences. On this view, the placement of pitch accents within the first clause should fully determine which NP serves as the correlate for the remnant. This possibility would require us to either abandon the enduring focus principle, despite the arguments made above, or limit its domain to cases in which prosodic marking is optional or otherwise less central to the task of interpretation. Assuming a relatively direct pathway from a contrastive pitch accent to a contrastive focus structure, this possibility has a clear appeal for the design of the auditory processing system, as listeners would have little cause to doubt which NP provides the intended correlate.

To investigate these possibilities, we present results from a corpus study of spoken radio transcripts, and two auditory rating studies. The results together support the central predictions of our first information structure account of enduring focus, in which explicit contrastive accent strongly influences, but does not wholly override, default focus position.

2.1 Syntax of remnant

The most common remnant types were NP (48%) and VP (32%) remnants, with prepositional phrases (PPs), sentence complements, bare Verbs, Adverbs, Adjectives, and Determiners making up the remaining 20% of cases, each category comprising less than 8% of the data. The syntactic distribution of the remnant categories closely follows the distribution found in the British National Corpus and the Contemporary Corpus of American English (COCA) for both let alone and much less constructions (Davies, 2008; Harris & Carlson, 2016), which suggests that this NPR sample, though small, is representative of speech and text recorded in other, previously explored, corpora.

2.2 Locality of correlate

As in Harris and Carlson (2016), the majority of correlates appeared in the most Local position with respect to the remnant for all remnant categories (88% of total cases). NP remnants accounted for the greatest deviance from the Locality bias with 16% of cases being Non-Local, followed by PP remnants at 14% Non-Local. Examples of Non-Local correlates for each of the three coordinator types is provided below, where the correlate and remnant are marked in bold, and the more local, non-correlate is presented in italics.

2.3 Prosodic contour

Prosodic analysis of the auditory files revealed that every correlate was produced with an accent, and that the pitch accent was judged to be contrastive 79% of the time (specifically, a steep, high L+H* accent). Similarly, every remnant bore a pitch accent, and most had contrastive accents (73% of cases). For the two major syntactic categories of remnants, NPs and VPs, a majority of examples had contrastive accents on both the correlate and the remnant (56% for NPs, 67% for VPs). Only a small number of examples had non-contrastive accents on both correlate and remnant (9% for NPs, 7% for VPs) and the remaining examples had contrastive accents on either the correlate or the remnant and non-contrastive accents on the other. A prosodic contour with contrastive accents on the correlate and remnant was clearly the most common rendition of the construction. Non-local and local examples had similar percentages of accented correlates and remnants to the whole dataset, though the number of non-local examples was small (just 25 of 211 examples). Sample pitch tracks for cases with local and non-local correlates are provided below in Figure 1.

OPEN IN VIEWER

Figure 1.

Pitch track of examples with a local correlate (a) and a non-local correlate (b).

Interestingly, most examples (85%) had pitch accents on the coordinator phrase itself (let alone, much less or never mind). This had not been expected but emerged as a phenomenon to analyze through observation of the sound files. These were usually non-contrastive accents (27% were contrastive). However, there were occasional examples where the accent on the coordinator was the most prominent element in the phrase.

Another common pattern was a prosodic boundary before the coordinator, appearing in 97% of the examples. Close to half of these prosodic boundaries (42%) were intermediate phrase boundaries, the smaller of the two prosodic phrase types in the ToBI system (94% of them ending with L- phrase tones). Another 40% of examples with boundaries had Intonational Phrase (IPh) boundaries with L-L% boundary tones, 14% had IPh boundaries with continuation rises (L-H%), and 4% had IPhs ending in H-L% plateaus.

2.4 Discussion

Although the relatively small size of the dataset makes it less than conclusive, the auditory corpus data validates the intuitive role of contrastive accent in focus-sensitive coordination structures. Correlate–remnant pairs almost always bore pitch accents of some kind, usually contrastive accents. This finding supports the use of L+H* accent on the correlates and remnants in the experiments that follow as a natural and consistent way to produce FSC sentences, as well as indicating the intended position of focus and marking the contrast between correlate and remnant. Furthermore, we find that the Locality bias is observed in speech recorded for radio interviews, although there are a healthy number of exceptions, indicating that the preference is not absolute.

The fact that correlates and remnants were uniformly marked with clear pitch accents in this dataset provides some naturalistic support for the idea that accents are nearly obligatory in focus-sensitive coordination structures. In addition, it lends some plausibility to the alternative possibility that listeners could in principle rely solely on overt accent placement to determine focus structure, at least in the case of focus-sensitive coordination. In that case, there would be no clear utility to enduring focus, so that comprehenders would resist assigning focus based on its default position.

We explore the relationship between pitch accent marking and the Locality bias in two controlled experiments that follow. As a preview, we find that explicit pitch accent fails to override the preference for the most local correlate, but that this preference is active only when the remnant is compatible with the most local correlate. We interpret these results in terms of the continued, and powerful, role of default focus in sentence comprehension even when prominence is marked overtly.

3.1 Participants

Fifty-six subjects were recruited on Amazon’s Mechanical Turk (AMT). Subjects self-reported as native speakers of English, and were compensated with $2 for completing the experiment, which typically took approximately 25 minutes to complete. Although subjects were distributed evenly across four counterbalanced lists in a within-subjects Latin Square design, the data for one subject were corrupted and not replaced.

3.2 Materials

Twenty quartets like (21) were constructed, crossing Accent location of a contrastive L+H* accent (Subject vs. Object) and Remnant type (Subject vs. Object). Remnant types were disambiguated by the animacy of the second conjunct, in that remnants with Subject correlates were names of people like Kayla in (21b, d), or definite descriptions like the coach, whereas remnants with Object correlates were inanimate nouns, for example, the final (21a, c). Each Remnant type followed a matrix clause (Danielle didn’t pass the quiz) with a contrastive pitch accent located either on the Subject or on the Object NP. Although differences in animacy might create differences between the saliency of the subject and object, the animacy manipulation does allow us to unambiguously assess which correlate the participant took as the remnant. The design minimizes the effect of animacy as a confounding factor, in that pitch accent location was crossed with animacy of the remnant, which allows us to test for an interaction between conditions. In addition, we accented the let alone coordinator with an H* accent, and placed a L-H% continuation rise at the end of the matrix clause and the end of the remnant, as was seen in the majority of examples from the auditory corpus study. All items are provided in Appendix A.

All sentences were recorded by the second author and analyzed for adherence to these prosodic specifications. Any sentences that did not have the intended contour were re-recorded. Acoustic measurements to substantiate the listed features are detailed in Tables 1–2. Specifically, words with contrastive pitch accents were significantly higher and longer on average than those words in the other conditions, and all had the characteristic narrow peak shape of this type of accent, as illustrated in Figure 2. Objects in subject accent conditions were usually deaccented, while subjects in object accent conditions bore mild H* accents, as recorded sentences beginning with deaccented material sound quite odd.

OPEN IN VIEWER

Table 1.

Average fundamental frequency measurements in Hz for critical words and boundaries for Experiments 1–2 (standard deviations under 20Hz for peaks, 10 Hz for boundaries). The third column contains the average values for L- and L% or H% boundary tones that appeared at the end of the host clause before the coordinator.

Condition	Subject or NP1 peak	Object or NP2 peak	Boundary tone values (L-, L/H%)	Remnant peak
Experiment 1
a	249	318	150, 212	249
b	241	306	153, 209	243
c	347	164	154, 212	260
d	337	163	157, 211	245
Experiment 2
a.i	284	148	141, 149	216
a.ii	284	148	141, 153	218
a.iii	282	147	141, 152	224
b.i	206	273	140, 157	214
b.ii	192	267	140, 155	214
b.iii	190	266	139, 158	222

OPEN IN VIEWER

Table 2.

Average duration measurements in milliseconds (ms) for critical words for Experiments 1–2 (standard deviations under 200 ms for words, under 60 ms for pauses).

Condition	Subject or NP1	Object or NP2	Pause	Remnant noun phrase
Experiment 1
a	395	703	141	636
b	381	722	137	714
c	458	593	116	630
d	452	586	123	715
Experiment 2
a.i	528	603	138
a.ii	527	602	152
a.iii	515	599	159
b.i	462	694	126
b.ii	447	692	128
b.iii	453	698	140

OPEN IN VIEWER

Figure 2.

Pitch excursions of (21), labeled with the Tones and Breaks Indices annotation system of Beckman & Elam (1997). The first two panels show contrastive pitch accent on the object, with either a matching object (a) or mismatching subject (b) remnant. The final two panels show contrastive pitch accent on the subject, with either a mismatching object (c) or matching subject (d) remnant.

The materials additionally varied whether the animate remnant could plausibly be understood as the object of the matrix verb. For example, it might be plausible for Ryan to lift either Mary or the barrel in (22a), but it is implausible, not to mention gruesome, for the patient to eat his family in (22b).

In order to be able to examine this factor, we balanced the materials to make one half of the items plausible and the other half implausible, as judged by the authors. If comprehenders attempt to initially pair the remnant with the closest, that is, object, correlate no matter the meaning of the nouns and regardless of pitch accent location, then the interaction between Accent location and Remnant type could be driven by items with implausible object interpretations for animate nouns (22b).

A separate post-hoc norming study tested whether participants perceived the location of prominence in the experimental sentences for Experiments 1 and 2. The audio files for conditions (21a) and (21c) of each item were truncated in the pause before let alone, so that the remnant and contrastive construction would not influence perceptions of the first clause (e.g., Danielle didn’t pass the quiz.). Thirty participants listened to these files (along with several other sentence types in one of two counterbalanced lists) and indicated where they thought the most prominent phrase in the sentence was by answering questions like (23) for the item in (21):

The responses options included sole prominence on the subject (23a) or object (23b), dual prominence on both (23c), and other (23d), with a text box for participants to indicate another phrase. As only one response was provided in the text box (23d), it was removed from the data. One item was coded incorrectly and subsequently removed. Participants identified the subject as the sole location of prominence in 69% of subject accent cases, whereas they selected the object as the sole location of prominence for object accent sentences in 76% of cases; this difference was not significant in a binomial test, p = 0.30. Subjects selected the dual prominence response in 29% of subject accent cases, and in 22% of object accent cases, a difference which was also not significant in a binomial test, p = 0.13. Further inspection revealed that these responses were limited to a handful of listeners: five respondents chose dual prominence for almost every example (regardless of condition), whereas the rest chose it only rarely. The sentences which actually contained multiple accents, the object accent conditions, did not receive significantly more dual prominence responses overall. The norming results suggest that for most listeners, the matrix clauses had clear pitch accents of highest prominence on the intended NPs.

The primary goal of Experiment 1 was to determine whether the Locality bias would persist in sentences bearing clear, contrastive pitch accents. If correlate resolution is guided primarily by overt pitch accent placement, there should be no preference for local correlates. If, however, Locality is at least partially independent of explicit pitch accent, as predicted by a theory of auditory processing subject to enduring focus, then the bias for local correlates should be retained. In addition, the design creates two conditions with a potential mismatch: Subject accent paired with the Object remnant (21c), and the Object accent paired with the Subject remnant (21b). If perceivers’ assessment of focus marking is influenced by default focus in addition to explicit accent, then mismatching sentences should not be penalized equally; mismatching Object remnants (21c) should show higher acceptability ratings than mismatching Subject remnants (21b), since the object position is the default focus position.

3.3 Methods

Experimental items were presented in counterbalanced and individually randomized order, interspersed with 24 items from unrelated experiments and 16 non-experimental filler sentences, for a total of 60 items per participant in each session. Each participant heard only one condition of each item and heard equal numbers of items in each condition over the experiment, following a Latin Square design. Sentences were rated on 7-point Likert scale with labeled endpoints (1 = Totally unnatural; 7 = Totally natural).

3.4 Results

The data were collected and analyzed in R 3.2.3 (R Development Core Team, 2015). Means and standard errors are summarized in Table 3. A bar plot of z-score transformed scores, showing the difference of each condition from the grand mean (M = 5.12) is provided in Figure 3, along with the results from Experiment 2.

OPEN IN VIEWER

Table 3.

Experiment 1. Raw means/z-score transformations. Standard errors are in parentheses.

	Local correlate	Non-local correlate	Mean	Non-local Penalty
Local accent	5.84/0.43 (0.08)	4.15/–0.12 (0.10)	4.99 (0.08)	1.69
Non-local accent	5.58/0.27 (0.09)	4.91/–0.57 (0.10)	5.25 (0.07)	0.67
Mean	5.71 (0.06)	4.53 (0.07)

OPEN IN VIEWER

Figure 3.

Ratings for each condition expressed as z-score transformations for Experiments 1 and 2.

The ratings were analyzed in a linear mixed effect regression model, treating Accent location, Remnant type and their interaction as fixed effect predictors, with maximal random effect structures, that is, by-subject and by-item random slopes and intercepts. Levels of the predictor variables were given deviation coding (testing each group for significant deviations from the grand mean, as in traditional analyses of variance) in which the hypothetically most natural condition (Object correlate–Object accent) was treated as the statistical baseline. As there is some debate regarding how best to compute p-values from fixed-effect parameters in these models (Bates, 2006), we adopted the convention that t-values over |2| are to be considered significant. The model was computed on the raw acceptability rating scores, although qualitatively identical results were obtained in a model fit to centered z-score values.

The mixed effects model yielded several effects; see Table 4. First, there was a main effect of Locality: Non-Local (Subject) correlates were rated as less natural than Local (Object) correlates, t = -6.73. Second, we observed an effect of Accent location, in which Object focus elicited lower naturalness ratings than Subject focus, t = 2.49. Crucially, there was a differential effect of Accent location. While Non-Local accent degraded Local correlates (d = -0.67), it improved Non-Local accent correlates (d = 1.69), t = 4.10. No other effects were observed.

OPEN IN VIEWER

Table 4.

Experiment 1: Fixed effects of the linear mixed effects regression model. Parameters with t-values above |2| were considered significant, and are marked with an asterisk.

Parameter	Estimate	Standard error	t-value
(Intercept)	5.119	0.131	39.04*
Remnant type	−0.589	0.088	−6.73*
Accent location	0.126	0.051	2.48*
Remnant type × accent location	0.252	0.062	4.10*

As mentioned above, half of the items contained plausible object interpretations of the subject remnant and half contained implausible ones. However, when Plausibility was added as a fixed effect to the model above, it did not add or remove any significant effects. ⁴ In all, the lack of a robust effect for plausibility is compatible with the idea that comprehenders’ ratings were guided by pitch accent placement, semantic parallelism, and a general preference for local correlates, instead of a failed attempt to integrate the remnant as an object. Nonetheless, the issue is not fully settled, and may require additional testing using online methods.

As noted, Local and Non-Local contrasts usually differed not only by animacy, but also by definiteness: sentential subjects were proper names (Kayla) in 65% of the items and definite descriptions denoting individual roles (the patient) or professions (the math teacher) in the remaining 35% of the items. As noted by a reviewer, accommodating an appropriate scale for proper names might be pragmatically more demanding or less constrained than for descriptions. For example, in (22), finding a scale for the Non-Local contrast condition, Ryan couldn’t lift the jug, let alone Cindy, requires accommodating some relation between Ryan and Cindy along an appropriate dimension such as strength. Finding a scale for the object contrast condition, Ryan couldn’t lift the jug, let alone the barrel, requires the comparison of the jug and the barrel, which might be facilitated by their conventional real-world properties, for example, weight or size. In other words, the availability of real-world knowledge might have eased the computation of the necessary scalar contrast independently of the location of the correlate.

Although we have attempted to control for this issue in the next experiment, we conducted an additional post-hoc analysis on the items, in which the non-orthogonal contrast of Subject type (Proper name, Definite) was added as a predictor to the linear mixed effects regression model above. The scores and resulting model are displayed in Tables 5 and 6, respectively. Model comparison revealed that adding Subject type resulted in a marginally improved model fit, χ ² (4) = 9.02, p = 0.06. Although the predicted results are qualitatively identical to the previous model, the more complex model revealed a significant three-way interaction between Locality, Accent Placement, and Subject type. While the penalty for violating Locality was the same within the Local accent condition for Proper names (d = 1.68) and Definite description (d = 1.70) subjects, the penalty for violating Locality with a Non-Local accent was twice the size for Proper names (d = 0.81) compared to Definite descriptions (d = 0.41), t = 2.30. No other effects were observed.

OPEN IN VIEWER

Table 5.

Experiment 1: Raw means/z-scores transformations including non-orthogonal contrast of subject type. Standard errors are in parentheses.

Remnant type
Subject	Accent placement	Local	Non-local	Non-local penalty
Proper name	Local accent	5.88/0.45 (0.10)	4.20/–0.54 (0.12)	1.68
	Non-local accent	5.58/0.28 (0.11)	4.77/–0.21 (0.12)	0.81
	Remnant type mean	5.73/0.36	4.49/–0.37
Definite	Local accent	5.75/0.37 (0.13)	4.05/–0.63 (0.19)	1.70
	Non-local accent	5.58/0.27 (0.16)	5.17/–0.21 (0.17)	0.41
	Remnant type mean	5.67/0.32	4.61/–0.30

OPEN IN VIEWER

Table 6.

Experiment 1: Fixed effects of the linear mixed effects regression model with the non-orthogonal contrast of subject type as an additional predictor. Parameters with t-values above |2| were considered significant, and are marked with an asterisk.

	Estimate	Standard error	t-value
(Intercept)	5.128	0.133	38.54*
Remnant type	−0.571	0.089	−6.41*
Accent placement	0.138	0.054	2.57*
Subject type	0.027	0.057	0.47
Remnant type × accent placement	0.285	0.059	4.81*
Remnant type ×subject type	0.059	0.050	1.19
Accent placement × subject type	0.040	0.050	0.80
Remnant type × accent placement × subject type	0.111	0.047	2.30*

It is important to note that the predicted interaction between Remnant type and Accent placement emerges in both types of items, with either Proper names or Definite descriptions as the subject; the Proper name simply produced a larger penalty for violating Locality. This pattern coheres with our intuition that the likely pairing between the correlate and the remnant can be affected by many factors, including the semantic parallelism between the correlate and the remnant, and the availability of a scalar contrast between the two terms (as manipulated in Harris & Carlson, 2016).

To address this concern directly, a post-hoc completion study manipulating the type of sentential subject (Proper name, Definite, Definite with adjective) was conducted over the Internet using Ibex Farm (http://spellout.net/ibexfarm/). Participants were instructed to complete written sentence fragments truncated after the let alone coordinator (e.g., {Patrick / the slacker / the smartest slacker} didn’t read the article, let alone …) with the first natural-sounding completion that came to mind. Sentence fragments were based on the Experiment 1 items, but modified so that the object was always definite and so that subjects varied according to the three conditions. Materials were presented with 56 unrelated experimental items and five highly constrained catch items. Items were counterbalanced in three equally distributed lists and presented in individually randomized order.

The data consist of 432 responses from 24 native speakers of English who provided sensible completions of catch items, and were annotated for the syntactic category of the remnant (either VPs or NPs), and, in the case of NP remnants, for Locality. There was an overall 65% bias towards VP completions, consistent with previous studies (Harris, 2016; Harris & Carlson, 2016). A local contrast was vastly preferred in conditions with both proper name (98%) and definite description (100%) sentential subjects. When the sentential subject contained an adjective (smartest), 65% of completions contrasted with the most local noun, significantly fewer than the other conditions, χ ² (2) = 8.81, p < 0.05, which is compatible with Harris and Carlson’s (2016) study that used similar adjectives to create non-local contrasts. As mentioned, we suspect that correlate selection is modulated by a great many factors, including the accessibility of a contextually salient scale, locality, and pitch accent location. However, there were no clear differences between responses to proper names and definite descriptions as subject correlates in our materials.

3.5 Discussion

Using a contrastive L+H* accent to mark subject or object nouns as focused, the experiment manipulated the compatibility between accent location and the parallelism of the remnant and correlate (by varying the animacy of the NPs). As predicted by a theory subject to enduring focus, structures that violated Locality were rated as less natural than those that did not—that is, remnants that contrasted with the matrix object were preferred to those that contrasted with the non-local matrix subject. The penalty for non-local correlates was reduced, though not eliminated, when the matrix subject was the only contrastively accented constituent in the matrix clause. The results offer initial support for the principles outlined above, in that the default location for focus continues to influence the interpretation of focus-sensitive constructions, even when focus-marking prosody is present. The pitch accents did affect ratings, though, so it is also clear that listeners were aware of their presence. More generally, these results point to the multiple factors that must be resolved when resolving focus-sensitive phenomena.

4.1 Participants

Fifty-six subjects were recruited from AMT to complete an auditory rating and questionnaire study on the Qualtrics platform. Subjects were compensated $4 for completing the experiment, which took less than 30 minutes to complete on average. Eight subjects were removed for counterbalancing purposes, leaving a total set of 48 subjects evenly distributed into six lists, counterbalanced across conditions.

4.2 Materials

Twenty-four sextets like (25) were recorded; items crossed Accent location (a. Non-Local accent vs. b. Local accent) and Remnant type (i. Local, ii. Non-Local, iii. Non-Local RC). As with Experiment 1, sentences with the Local and Non-Local levels of Remnant type were disambiguated by the animacy of the noun in the remnant. In the Non-Local RC condition, the remnant began with the same noun as the Non-Local condition (a book) but followed it with a relative clause ending in VP ellipsis. The ellipsis syntactically disambiguated the remnant towards the Non-Local correlate, as both consist of a subject relative clause headed by an inanimate noun (an article that exposed the governor vs. a book that did). All items are provided in Appendix B.

All sentences were recorded by the second author and analyzed for adherence to certain prosodic specifications, that is, presence of L+H* contrastive accents on the earlier (Non-Local) or later (Local) nouns and the remnants, and a L-L% or L-H% prosodic boundary before let alone. Continuation rises followed the initial adverbial phrases or clauses (Although… in (25)), and L-L% boundaries before the remnant were as common in the corpus study as L-H% boundaries. The longer remnants (a book that did) had contrastive accent on the head noun (book) and all remnants ended in L-H% continuation rises. Any sentences that did not have the intended contour were re-recorded. Acoustic measurements to substantiate the listed features are in Tables 1 and 2. Specifically, words with contrastive (L+H*) pitch accents were significantly higher and longer on average than those words in the other conditions, and all had the characteristic narrow peak shape of this type of accent.

The same post-hoc norming study presented in Experiment 1 also tested whether the intended prominence contour on the experimental sentences from Experiment 2 was audible to participants. The audio files for conditions (25a) and (25b) of each item had the adverbial clause removed and were truncated at the pause before let alone (e.g., John didn’t write an article that exposed the governor.), so that the let alone coordinator and the remnant type could not influence the perception of the prosodic contours on main clause. Participants indicated the location of prominence by answering questions like (26).

The options included sole prominence on the head noun of the relative clause (26a) or on the final noun in the relative clause (26b), dual prominence on both (26c), and other (26d), with a text box for participants to indicate another phrase. Subjects located prominence largely as expected. For the head NP accent sentences (25a), 79% of responses identified the prominence solely on the relative clause head, whereas for the final NP sentences (25b), 77% of responses chose sole prominence on the final noun. Subjects chose dual prominence in about 20% of responses, and the difference between conditions was not significant in a binomial test, p = 0.74. The same five subjects were primarily responsible for the majority of dual prominence responses as in the norming task to Experiment 1. These results suggest that for most listeners, the clauses hosting let alone ellipsis had clear pitch accents with the highest prominence on the intended NPs.

All of the possible correlates here were positioned within the predicate of the matrix clause, that is, the relative clause NP (an article that exposed the governor). If the Locality bias in Experiment 1 simply reflects a preference for object correlates or a dispreference for subject correlates, then the Local and Non-Local conditions in this experiment should not differ. If instead Locality reflects the default focus structure of a sentence, then we expect lower ratings for the Non-Local conditions, in which the remnant is paired with a higher NP (an article).

4.3 Methods

Experimental items were presented in one of 12 counterbalanced lists in a single pseudo-randomized order, with no consecutive items of the same type or condition. The experimental items were interspersed with 84 items from unrelated experiments for a total of 108 stimuli. Each participant heard only one version of each experimental item and heard equal numbers of items in each condition over the experiment. The same Likert scale as in Experiment 1 was used to collect naturalness ratings.

4.4 Results

Responses were collected in Qualtrics and analyzed in R; means and standard errors are presented in Table 7, and z-score transformed values showing the deviation from the mean is presented in Figure 3. The data initially were subjected to a linear mixed effects regression model with random slopes and intercepts, as in the previous experiment. However, as this model failed to converge, we used a simplified model with random intercepts only. With respect to the deviation contrast coding, the Local NP (25.i) level was selected as the statistical baseline for the Remnant type factor, and the Local accent level was treated as the baseline for the Accent location factor (25b), as in the previous experiment. We thus tested whether the effects of Non-Local accent conditions, Non-Local correlate conditions, or their interaction were greater than what would be expected from normal deviation from the grand mean.

OPEN IN VIEWER

Table 7.

Experiment 2. Raw means/z-score transformations for each condition. Standard errors are presented in parentheses.

Remnant type
Accent location	Local noun phrase (NP)	Non-local NP	Non-local relative clause	Mean
Non-local	5.40/–0.08 (0.10)	5.67/0.10 (0.10)	5.77/0.14 (0.10)	5.60 (0.06)
Local	5.79/0.18 (0.10)	5.12/–0.27 (0.10)	5.43/–0.06 (0.10)	5.45 (0.06)
Mean	5.59 (0.08)	5.39 (0.08)	5.58 (0.07)

As predicted, the Non-Local NP remnant (the president) was rated as less acceptable than the Local NP baseline (a book), t = -2.50, which did not differ in acceptability from the Non-Local RC remnant (a book that did), t = 1.21 (see Table 8). Non-Local accent (pitch accent on the relative clause head) was rated less acceptable overall than Local accent, t = -2.17. We attribute this effect to an experimental bias toward the Non-Local conditions: Non-Local nouns provide the correct correlate in two of the three remnant conditions, thereby producing higher ratings overall for the accent pattern. In addition, Accent location affected naturalness ratings differently for the Non-Local remnant types. The ratings penalty for Local accent locations was reversed when paired with a Local remnant (d = -0.55, t = -3.88), but accent location did not differentially affect Non-Local RC remnants, (d = 0.30, t = -1.40), suggesting that the effect of Locality is limited to syntactically ambiguous forms.

OPEN IN VIEWER

Table 8.

Experiment 2: Fixed effects of the linear mixed effects regression model. Parameters with t-values above |2| were considered significant, and marked with an asterisk.

	Estimate	Standard error	t-value
(Intercept)	5.520	0.142	39.00*
Local accent	−0.077	0.036	−2.17*
Non-local noun phrase (NP)	−0.127	0.051	−2.50*
Non-local relative clause (RC)	0.061	0.051	1.21
Local accent × non-local NP	−0.196	0.051	−3.88*
Local accent × non-local RC	−0.071	0.051	−1.40

The overall pattern is perhaps best understood as a penalty for a mismatch between pitch accent and the likely remnant–correlate pairing. Indeed, in a model that transformed the conditions to reflect the relative match or mismatch between the pitch accent and the remnant, there were only two main effects: one showing a naturalness penalty for remnants whose correlates are not marked with matching, that is, contrastive, accent; and another showing a naturalness penalty for syntactically ambiguous remnants with Non-Local correlates. In this model, the Match factor reflected cases in which the location of the pitch accent matched semantically with the remnant (Match: 25a.ii, 25a.iii, 25b.i vs. Mismatch: 25a.i, 25b.ii, 25b.iii). Remnant type was specified with the levels Local NP (25. i), Non-Local NP (25. ii), and Non-Local RC (25.iii). The statistical baseline for model was the Match–Local condition, that is, John didn’t write an article that exposed the GOVERNOR, let alone the PRESIDENT in (25.b.i). Responses were given the same random effects and contrast coding as before, resulting in a decrease in naturalness ratings for Mismatching ( β ^ = −0.20, standard error (SE) = 0.04, t = −5.70), and Non-Local ( β ^ = −0.13, SE = 0.05, t = −2.50) conditions. No other fixed effects besides the Intercept ( β ^ = 5.52, SE = 0.14, t = 39.00) were significant, including the Non-Local RC condition, which confirms the interpretation that Non-Local remnants are not costly when syntactically unambiguous.

In support of this interpretation, post-hoc pairwise t-tests with Bonferroni correction revealed a penalty for the condition with Local accent and Non-Local correlate (25b.ii) compared to conditions with matching accents (25b.i and 25a.ii, 25a.iii) in by-subject analyses, p’s < 0.05. The penalty for mismatching accent also appeared as fully significant in by-items analysis for conditions with Non-Local correlates (25a.ii, 25a.iii), p’s < 0.05, and as a non-significant trend for the condition with a Local correlate (25b.i), p = 0.097. In concert with the auditory corpus study and the previous experiment, listeners appear to expect the correlate and remnant to match in their accent status, and find deviations from that expectation less natural.