Glossa Psycholinguistics

2.1 Misinterpretations in garden path sentences

The processing of temporary, local syntactic ambiguity in garden path sentences has been well studied in the psycholinguistics literature as a window into the internal mechanisms of the human parser (e.g., Bever, 1970; Frazier & Fodor, 1978; Frazier, 1979). Consider the sentence in (1), where the complementizer while signals the start of an adjunct clause. Because the presence of an adjunct clause entails the presence of the matrix clause that dominates it, the beginning of the sentence signals the parser to expect a sentence which is minimally composed of two clauses – an adjunct clause followed by the matrix clause (Kazanina et al., 2007). However, up through the second NP, headed by baby (bolded region), it is ambiguous whether the boundary between the two clauses comes early after dressed or late after cute, as both (1a) and (1b) are possible continuations of the sentence fragment. Here, if the parser commits to the late-closure analysis of the adjunct clause VP but the actual structure of the sentence is one where that VP closes early, this presents a serious challenge for comprehension.

Fortunately, the parser is not deterministic such that it simply breaks down when it is led down the garden path (Marcus, 1980; Berwick & Weinberg, 1984). Instead, it has the capacity to recover from an initial misparse and build the correct underlying structure, though often with great effort (Fodor & Frazier, 1980; Marcus et al., 1983; Pritchett, 1992; Schneider & Phillips, 2001; Sturt et al., 2001). In reading studies, comprehenders have been reported to systematically display patterned behavior while processing ambiguous regions of the sentence that reflect this process of recovery from being garden-pathed, also called reanalysis. For example, Frazier and Rayner (1982) examined the on-line behavioral responses to parsing difficulty in sentences like (2) in an eye-tracking-while-reading paradigm.

Frazier and Rayner (1982) found that the parser initially interprets a mile as the direct object of jogs, although the correct analysis requires the adjunct clause VP to be closed early at the verb, with the NP a mile being analyzed as the subject of the matrix clause instead. This violation of the parser’s expectation is reflected in longer looking times at the disambiguating region seems like, followed by regressive eye movements to an earlier region of the sentence. This sequence of behaviors has been taken to be a sign of a successful reanalysis because readers do not appear to experience any further difficulty reading the rest of the sentence. Representationally, this reanalysis process appears to allow the parser to both discard the initially incorrect analysis from further consideration and also build the globally correct analysis. In sum, reanalysis was understood to be a mechanism that allows comprehenders to arrive at the correct interpretation of garden path sentences despite the initial difficulty processing them.

However, this all-or-nothing view of reanalysis has been challenged by studies that more closely examine the post-repair representations of garden path sentences (Gibson, 1991; Gibson & Pearlmutter, 1998, 2000; Ferreira, Christianson, & Hollingworth, 2001; Ferreira, Bailey, & Ferraro, 2002; Ferreira, Lau, & Bailey, 2004). For example, Christianson et al. (2001) reports that after reading sentences like (3), comprehenders often failed to correctly answer simple questions like (4) even after reporting high confidence in their answers.

If reanalysis indeed facilitated the parser’s recovery from its initial mistake of attaching the baby as the direct object of dressed, then the post-repair representation of the sentence should no longer license the interpretation that Anna dressed the baby. Therefore, these comprehension errors provide evidence that the process of reanalysis as traditionally understood could sometimes fail at yielding the correct interpretation of the sentence. At the same time, however, comprehenders consistently answered questions like (5) correctly, suggesting that the failure of reanalysis is very particular and local.

In light of this finding, early works by Ferreira and colleagues have made a broader claim about the relationship between sentence processing, the syntactic structure, and comprehension (e.g., Ferreira et al., 2001). In Ferreira et al., (2002), the syntactic structure is claimed to be fragile and not by itself a reliable source of sentence meaning for comprehenders (see also Traxler, 2011, 2012, 2014). Instead, the communicative context in which the sentence is situated ultimately helps make a sentence’s meaning available to comprehenders. In most natural conversations, the context of an utterance reinforces the interpretation generated from the sentence’s underlying structure, which allows people to comprehend utterances rapidly and accurately in real time. However, when comprehenders are deprived of the context, or if the available contextual information interferes with the interpretation generated from the underlying structure, then the normal routines of comprehension are disrupted, resulting in representations that are merely “good enough.”

Ferreira and Patson (2007) formalize this notion in what they call the Good Enough approach to human language comprehension, founded on the view that the language processing system, by design, builds representations of utterances that are simply “good enough” for the task at hand (e.g., comprehension in conversation). For example, when the processing of a sentence becomes costly, comprehenders may resort to constructing a shallow representation of the sentence that may sometimes conflict in interpretation with the one generated from the faithfully parsed syntactic structure. In this way, the Good Enough approach attempts to unify the phenomena of comprehension errors beyond just those involving sentences with local syntactic ambiguity, such as thematic role reversals in passives and Moses illusions (Ferreira, 2003).

However, there are strong reasons to maintain the belief that the parser is capable of building complex structures under pressure, and that the syntactic structure still plays a significant role in determining the possible interpretations that are made available to comprehenders. Decades of psycholinguistics research on the role of grammatical constraints in parsing have found that comprehenders are sensitive to the structural relations between abstract linguistic units in a sentence that cannot be represented at a shallow level of processing. For example, readers have been shown to respect island constraints when positing gaps (Stowe, 1986; Traxler & Pickering, 1996; Phillips, 2006) and Binding Principles when forming pronoun-antecedent dependencies (Sturt, 2003; Kazanina et al., 2007; Kush & Dillon, 2021) in real time, neither of which are easily reducible to lexical identity, word order, pragmatic inference, and the like. In fact, even in the case of passives, which are often claimed to be especially susceptible to interference from extra-syntactic strategies like the agent-first heuristic (e.g., Ferreira, 2003), recent evidence suggests that the comprehension difficulty is driven instead by task-specific demands giving rise to “active-like” representations in offline processing, and the robustness of these interpretations to memory decay (Grillo et al., 2018; Paolazzi et al., 2019). Given this evidence, the claim from early formulations of the Good Enough approach – that the syntactic structure can become a fragile source of meaning for comprehenders when the sentence is too complex – warrants greater scrutiny.

This issue was investigated in Slattery et al.’s (2013) study of the syntactic representations that underlie “good enough” interpretations. Slattery et al. (2013) lay out two competing hypotheses about the nature of post-repair syntactic representations when misinterpretations are observed. The first hypothesis is that the parser fails to build the correct structure in reanalysis, resulting in a degenerate structure that tolerates misinterpretations. The second hypothesis is that the correct structure is fully represented, but the locally coherent analysis lingers in some form and interferes with comprehension. Early proposals of the Good Enough approach (e.g., Christianson et al., 2001; Ferreira et al., 2001) are most consistent with the first possibility, where an ill-formed and/or underspecified syntactic representation can drive comprehenders to become more reliant on external heuristics that operate quickly at the surface level of the sentence. However, if the correct syntactic structure is fully recovered after undergoing reanalysis, that presents a potential challenge to this account: it would need to make an even stronger claim about the ability of surface-level heuristics to dominate structural information as the primary source of sentence meaning (Fujita, 2021).

Slattery and colleagues tested both possibilities in two separate eye-tracking-while-reading experiments. In the first experiment, they exploited the gender mismatch effect (Sturt, 2003) as a diagnostic for whether the parser successfully builds the correct syntactic structure, using a pair of sentences such as in (6).

In the absence of the post-adjunct comma, the classic garden path effect is predicted at the matrix verb grew for both sentences. But critically, if the parser is able to recover the correct structure after undergoing reanalysis, an additional difficulty at the pronoun himself is predicted when the matrix subject does not agree in gender (i.e., when the matrix subject is David’s mother) because the parser’s search for the antecedent of the reflexive should be constrained by the c-command constraint on reflexive binding (Chomsky, 1981). In other words, if the post-repair syntactic representation minimally specifies the c-command relationship between the matrix subject and the reflexive, and if the matrix subject is David’s mother, then the parser will attempt to interpret himself as David’s mother and experience difficulty doing so. On the other hand, if reanalysis is “good enough” and simply yields a coarse-grained representation of the sentence such that the antecedent search is not syntactically constrained, then the mere presence of the noun David preceding the reflexive pronoun should be a sufficient cue to erroneously license the interpretation of David as the antecedent of himself. Interestingly, a slowdown at the pronoun was observed only when the pronoun did not agree in gender with the matrix subject, suggesting that the parser does indeed build rich and detailed representations of the syntactic structure even when comprehension appears to be degraded.

At this point, if we still maintain the assumption that interpretations are constrained by the syntactic structure, then the parser must have built a structure that can license a misinterpretation at some point during processing, in addition to ultimately building the globally correct structure after reanalysis. This assumption bears out in Slattery et al.’s (2013) second experiment, which investigated whether the post-repair interpretation is consistent with that associated with the locally coherent parse by manipulating the material past disambiguation, such as in (7).

The adjunct verb dried off in the first sentence of (7) is temporarily ambiguous in transitivity until the matrix verb was (peed on) signals that the NP headed by truck is not the embedded object – i.e., that the transitive analysis of dried off is no longer viable. If that incorrect analysis is nevertheless available after disambiguation, comprehenders should experience difficulty processing himself in the following sentence, since the interpretation that Frank has finished drying himself off presupposes that he had been drying himself off before – an interpretation that is only available if the matrix verb dried off in the preceding sentence had incorrectly been interpreted as a reflexive at some point. Indeed, readers slow down at the pronoun well after the garden path effect in the first sentence had subsided, suggesting that some aspect of the initial misanalysis is persisting past the point of disambiguation.

Thus, Slattery et al. (2013) provide strong evidence that it is not the failure to construct the correct structure but rather the failure to discard the incorrect analysis that leads to lingering misinterpretations. This is consistent with previous findings for the semantic persistence effect as described in Sturt (2007), which studied reading behavior in garden path sentences when they followed a context sentence. Sturt reports that given a context sentence like (8), late reading measures indicative of semantic integration were longer at the critical disambiguating region (was actually) when the content of the second sentence was semantically inconsistent (8b) compared to when it was consistent (8a).

The increased difficulty in processing (8b) relative to (8a) indicates that the initial semantic interpretation from the preceding context sentence must have been available at the point of reanalysis to interfere with semantic integration. Critically, this finding points to the possibility that this same mechanism of semantic persistence across two sentences may also characterize the lingering effect observed for garden path sentences before and after reanalysis as well.

To address this question of whether it is the semantic interpretation or the syntactic structure that lingers from the incorrect local parse to interfere with semantic/syntactic reanalysis, previous research has turned to the contribution of lexical factors. For garden path sentences with early/late-closure ambiguity like While Anna dressed the baby played in the crib, two qualities of the erroneous local analysis have been reported to contribute to the failure of reanalysis: the semantic fit of the adjunct NP as the direct object of the adjunct verb (Trueswell & Tanenhaus, 1994; Pickering & Traxler, 1998; Tabor et al., 2004; Hare et al., 2009) and the transitivity bias of the verb (MacDonald, 1994; MacDonald et al., 1994; Garnsey et al., 1997; Gibson & Pearlmutter, 1998, 2000). For example, if the baby is a plausible candidate as the object of dressed and if dressed is frequently used in transitive structures, comprehenders would have greater difficulty accurately answering a question that asks whether Anna dressed the baby.

However, this line of evidence has not been successful in disentangling the role of semantic vs. syntactic reanalysis in the lingering effect. For example, proponents of serial models with incremental semantic processing (e.g., Sturt, 2007) claim that the incorrect local analysis is first pursued before the globally correct analysis is made available to the parser, and so it is merely the semantic interpretation from that initial misanalysis which lingers. Under this account, the semantic processor may make stronger commitments to the initial misanalysis if the structure is highly probable and the interpretation is highly coherent, strengthening its resistance to revision during semantic reanalysis. On the other hand, proponents of parallel models that support the coexistence of multiple local structures (Tabor & Hutchins, 2004; Tabor et al., 2004; Ferreira et al., 2004; Lau & Ferreira, 2005) claim that a locally coherent parse that was temporarily entertained could fail to decay if it receives sufficient activation from plausibility and frequency. Under this account, the lingering of the incorrect structure alongside the correct structure in the syntactic representation is what may drive the misinterpretations, despite the parser having successfully built the correct structure. Given this difficulty of disentangling the role of semantic versus syntactic reanalysis from the contribution of lexical information, we propose in the next section a method of isolating the role of semantic reanalysis in the lingering effect by manipulating post-lexical prosodic information, drawing on the well-known relations between pitch accent, focus, and semantic processing.

2.2 Pitch accent and meaning

In the Autosegmental Metrical (AM) model of English intonational phonology, the pitch contour of an utterance is understood to be derived from interpolation between discrete tonal targets which are anchored to syllables that are phrase-level peaks of prominence and phrase-final edge tones (Pierrehumbert, 1980; Beckman & Pierrehumbert, 1986; Gussenhoven, 2004). Phrasal prominence is determined from a hierarchically organized metrical structure, which defines a strong-weak patterning over words at the phrase level (Ladd, 2008). In English, the default prominence pattern has an obligatory prominence on the rightmost content word in the prosodic phrase. For a simple sentence (with no embedded clause), produced in a neutral speaking style, the verb and its (non-clausal) complements constitute one prosodic phrase, and a non-pronominal subject NP may optionally constitute a separate prosodic phrase, or may be integrated into the phrase containing the verb (Büring, 2016). An obligatory prominence at the right edge of a prosodic phrase marks the location of the main phrasal stress and is termed the nuclear prominence (Chomsky & Halle, 1968). If a phrase is long enough, there may be additional optional prenuclear prominences assigned to words in metrically strong positions earlier in the phrase (Calhoun, 2006).

A word with phrasal prominence licenses a tonally specified pitch accent (Truckenbrodt, 1995), which is obligatorily assigned to the word with nuclear prominence, and is optionally assigned as a “rhythmic” or “ornamental” feature of words with prenuclear prominence (Calhoun, 2006; Ladd, 2008; Büring, 2016). In the AM model, pitch accents may consist of a single tone (L* or H*), or a sequence of two tones (e.g., L+H*), associated with the syllable with word-level primary stress in the prominent word. Additional tones are assigned at the right edge of a prosodic phrase, with a single tone (L- or H-) at the edge of a low-level prosodic phrase which is followed by a second tone (L% or H%) at the end of a high-level prosodic phrase. The following example illustrates the default hierarchical metrical structure, accent assignment, and phrase-edge tones for the sentence Chicago is a city in Illinois.

In addition to metrical structure, pragmatic and semantic factors also play a role in the distribution of phrasal prominence. For instance, accent placement is sensitive to distinctions in the information status of words within a prosodic phrase; specifically, words that are accessible from the prior discourse context (discourse-given) are generally not eligible for nuclear prominence. In conversations, the canonical ordering of sentence constituents is given information before new information, with the result that the new information is located in the default position for nuclear pitch accent (Birner & Ward, 1998; Calhoun, 2012). For example, in (10), the bolded portion of Speaker B’s response is new in the discourse because it is the answer to Speaker A’s question, while the preceding content is old in the discourse because it was presupposed in the question.

Prominence assignment is also conditioned by semantic focus. A word with contrastive, narrow, or corrective focus is preferentially assigned nuclear prominence, even if it occurs in the non-final position in the sentence (Büring, 2016). For example, in the sentence SUZY showed up to class, the prominence on SUZY can be interpreted as a marker of contrastive focus. In the framework of alternative semantics (Rooth, 1992), focus is understood as a semantic notion which references a set of focus alternatives (f-alternatives) and triggers a set of propositions which contrast in the focused element (e.g., JOHN showed up to class).¹

In English, tonally distinct pitch accents are also used to mark these graded distinctions in givenness and semantic focus (Baumann & Riester, 2012). In particular, new and/or contrastive information is often marked with an H* or L+H* (Watson et al., 2008).² For example, in (11), Speaker B’s response has the nuclear prominence on cake, and would typically be assigned a H* or L+H* pitch accent marking narrow focus (as the answer to the prompting question).

Experimental works on prosodic focus over the last several decades provide supporting evidence of a systematic relationship between pitch accents and focus, both in production and comprehension (see reviews in Cutler et al., 1997; Wagner & Watson, 2010; Wagner, 2020; but see also Chodroff & Cole, 2019 and Roettger et al., 2019 for evidence of a weaker relationship). For example, studies using eye-tracking in the visual world paradigm have found that listeners can rapidly access information-structural meanings from pitch accents in complex comprehension tasks (Ito & Speer, 2008; Kurumada et al., 2014). Furthermore, interpretations of focus-related semantic meaning have been found to trigger deeper semantic processing that has consequences for encoding and retrieval in memory. In a series of recognition tasks, Fraundorf et al. (2010) found that listeners were better able to recall a word from a prior discourse when the word was accentually prominent. Critically, listeners performed better on the recognition task when the choice was between contrasting referents, meaning that contrast can be encoded in pitch accents in a way that facilitates future identification and discrimination.

These findings indicate that listeners rapidly and automatically compute the rich semantic and pragmatic meanings of pitch accents in sentence processing. However, as already noted, not all pitch accents mark semantic focus on the accented word. Rather, prosodically licensed semantic focus scopes over syntactic constituents in a derivational process known as focus projection (Gussenhoven, 1992, 1999; Selkirk, 1995). Revisiting the example in (11), because the most prominent pitch accent in Speaker B’s answer (by virtue of being the only one present) occurs on cake, it specifically marks that object noun phrase as the answer focus. The focused constituent (a.k.a. the domain of focus) is marked with []_FOC brackets in (12), borrowing from Selkirk (1995). It is due to this ability for focus to project from the accented word and percolate up the syntactic structure that the same response by Speaker B in (12) is also an appropriate answer to a VP-focus question like (13).

Theories of the intonational grammar of American English generally agree that, to mark focus, the last pitch accent in the focus domain must bear the nuclear pitch accent and all subsequent material in the sentence must be deaccented (e.g., Jackendoff, 1972; Schwarzschild, 1999; Ladd, 2008). This focus-to-accent relationship becomes clearer in sentences where focus and default accenting locate the nuclear pitch accent on different parts of the utterance. For instance, the default accenting of (14) assigns prominence to CLASS, the rightmost word in the intonational phrase eligible to receive a pitch accent. However, having a contrastive focus over Suzy in (14) requires that the nuclear pitch accent occurs within the domain of focus, which is realized in (15) with an accent on Suzy. In turn, the accent on class is removed, as per the rule of post-focal deaccenting.

That the domain of focus must minimally contain the nuclear pitch accent is also reflected in the observation that focus is not realized by simply placing a pitch accent just anywhere in the domain of focus. For example, given Speaker A’s question in (16), the verb phrase visited a friend in Chicago in Speaker B’s answer constitutes the domain of focus. This may be realized prosodically with an H* on Chicago, as in (16a), but it would be inappropriate to have the nuclear pitch accent occur on any other word in the focus domain, such as on visited (16b). Since visited is phrase-medial and a predicate preceding an argument, assigning it nuclear (i.e., rightmost) prominence makes available only the narrow verb-focus reading, as shown in (17).

Given these constraints on the interpretation of prosodic focus and the observations on the effect of pitch accent on semantic processing in real time, we can make predictions about the interpretation of prominence in more complex sentences. Consider again the sentence While Anna dressed the baby played in the crib. Over the course of processing this sentence, the parser is expected to have built both the incorrect and the correct structure – either simultaneously (parallel) or in succession (serial), depending on the theory – due to the preference for late closure. Critically, the incorrect local analysis that was temporarily entertained may persist depending on the strength of local coherence as noted earlier. Here, we posit that the likelihood of persistence may also be affected by prosodic factors as well.

For example, the sentence can be produced with a sharply rising pitch accent on the adjunct verb dressed, as in While Anna DRESSED the baby played in the crib. When the parser is garden pathed and initially interprets the baby as the object of dressed, the accent on the adjunct verb occupies a non-canonical, phrase-medial position in that structure. Assuming that the parser considers a focus interpretation when there is structural evidence to do so in the absence of relevant context, the phrase-medial pitch accent is strong evidence for the presence of narrow focus on the transitive interpretation of the verb.³ This assumption generates an interesting prediction for a serial parser with incremental semantic processing: the prominence on the adjunct clause verb should yield an interpretation of focus on the incorrect late-closure parse, triggering deeper semantic processing for the erroneous analysis and enriching it with focus-related semantic meaning before reanalysis is triggered.

A likely arrangement of ordinary semantic representations (Carlson, 1984) for the early/late-closure parses in each accent condition is illustrated in (18), with the region of temporary local syntactic ambiguity in bold. If the pitch accent is interpreted as narrow focus (in 18a and 18c-d), the focus semantic meaning is also provided, where focus contributes to the semantic value with a set of alternatives. Following Rooth’s (1992) Focus Interpretation Principle, focus is understood to introduce a free variable at the level of the domain of focus, bound through the existential closure as an argument.⁴

Here, we see an asymmetry in the interpretation of focus between the local parses within the verb accent condition (18a-b), compared to within the subject accent condition (18c-d). Whereas the late-closure parse in the verb accent condition is enriched with focus semantic meaning which must later be revised (18a), the focus semantic meaning from the late-closure parse in the subject accent condition (18c) need not be revised in the course of garden path reanalysis, because a focused subject (Anna) is equally compatible with the early- and late-closure interpretations. In other words, the focus may in fact not be a target of reanalysis entirely (since it lies outside the region of ambiguity). Thus, we predict no additional semantic processing for the erroneous analysis in the subject accent condition, assuming that the interpretation of the pitch accent on the subject is unaffected by the process of ambiguity resolution which follows it. In this sense, the subject accent contour serves as the baseline condition that allows us to isolate the effect of deeper semantic processing on the initial misanalysis from the lower-level lexical effects on local coherence. Additionally, although (18c-d) construes the accent on the subject as contrastive, it is also possible for the subject accent contour to be interpreted as a rhythmic or “ornamental” accent, i.e., unrelated to focus (Büring, 2007). But whatever the interpretation of the subject accent may be, our hypothesis makes the same prediction: there should be no direct consequences for semantic reanalysis since the accented subject lies outside the region of ambiguity (in other words, there is no need to revise the initial interpretation of the subject accent). In that sense, the critical manipulation that separates the verb accent condition from the baseline subject accent condition is that the verb accent facilitates an interpretation of focus for the erroneous analysis which must later be revised, in order to be consistent with the syntactic reanalysis of the structure of the adjunct VP.

In sum, our hypothesis is as follows: if the lingering effect is driven by the failure of semantic reanalysis (i.e., the failure to discard the erroneous semantic interpretation), then enriching the erroneous semantic interpretation with focus-related semantic meaning should strengthen this effect. The crucial comparison here is comprehension accuracy between the condition where the erroneous garden path interpretation is marked for contrastive focus (verb accent condition) and the condition where it is not (subject accent condition).⁵ If the pitch accent on the verb facilitates an interpretation of narrow semantic focus in the incorrect local parse, then the erroneous semantic interpretation will become more likely to linger than in the baseline condition with the pitch accent on the subject. We predict this lingering misinterpretation to be reflected in lower accuracy rates on comprehension questions that ask whether the second NP is the embedded object (e.g., Was the baby dressed?), an interpretation that is only licensed by the incorrect late-closure parse. Critically, if this effect of pitch accent on the depth of semantic processing drives the persistence effect independently of the extent of locally determined coherence, then the location of the pitch accent should emerge as a significant predictor of accuracy on comprehension questions after controlling for item-level lexical effects from semantic fit and transitivity bias.

3.1 Design and norming

The critical items for the auditory comprehension experiment consisted of twenty-four garden path sentences with closure ambiguity, such as (19). The adjunct clause consisted of a complementizer (since, while, or when) followed by a two-syllable name with a strong-weak stress pattern and an ambiguously transitive verb in the past tense. The matrix clause consisted of a subject noun phrase that was also a possible direct object candidate for the adjunct clause verb, followed by a matrix verb phrase.

An initial set of twenty-six sentences were normed for biases in semantic fit (how often the interpretation of the second NP as the embedded object is preferred) and transitivity bias (how often the adjunct verb is followed by a direct object) in two independent off-line experiments. The norming experiments are described in more detail as follows:

Semantic fit was normed in a three-alternative forced-choice task, in which participants were asked to choose the more plausible event between the early-closure interpretation such as (20), the late-closure interpretation such as (21), or an option for both being equally plausible. Participants in the semantic norming task did not rate locally ambiguous sentences such as (19).

For each item, the proportion of responses preferring the early-closure interpretation against those preferring the late-closure interpretation was calculated (median = 0.53, IQR = 0.38). Semantic fit scores of each item were then centered at 0.5 (representing equal bias) and downweighted by the number of “both” responses.⁶

Transitivity bias was normed in a sentence completion task, in which participants saw fragments of each sentence up to the adjunct clause verb, such as While Anna dressed the baby __, and were asked to type in a continuation that completes the sentence. Responses that failed to complete the sentence were excluded from analysis and the rest were coded for the presence of a direct object following the adjunct clause verb. The proportion of responses involving direct objects was calculated (median = 0.38, IQR = 0.60) and again centered at 0.5.⁷

The centered measures of semantic fit and transitivity bias were then z-scored to be entered into the statistical model as independent item-based predictors.⁸ Figure 1 shows the distribution of semantic fit and transitivity bias scores for all items used in the experiment. The variability among items along both dimensions allows the model to statistically control for their effects on comprehension accuracy and isolate the effect of pitch accent placement.

Figure 1: Norming scores for transitivity bias and semantic fit of the sentences used as stimuli for the speeded auditory comprehension experiment.

3.2 Audio stimuli

The stimuli for the auditory perception experiment were recorded using a final set of twenty-four normed sentences.⁹ A trained male speaker recorded each sentence with two accent patterns – one with prominence on the adjunct subject, such as (22), and another with prominence on the adjunct verb, such as (23).

In the construction of the auditory stimuli, we considered the possibility that an interpretation of a prosodic phrase boundary after the adjunct verb may immediately resolve the local attachment ambiguity, much like the role of a comma in reading. The effect of an intervening intonational phrase boundary on blocking attachment is well-documented and may be signaled via a combination of a pause, pitch reset, and pre-boundary lengthening (Kjelgaard & Speer, 1999; Schafer et al., 2000). In order to make these cues ambiguous, so as to induce the garden path effect, a set of late-closure variants of the stimuli, such as (24) and (25), were also recorded by the same speaker. From the late-closure recordings, the durations of the words over the region of ambiguity that are potential targets of pre-boundary lengthening – the adjunct verb (dressed) and the post-verbal noun (the baby) – were measured and referenced in the resynthesis of the early-closure stimuli used in the experiment.

In total, four recordings were created for each sentence. Figure 2 shows the distribution of durations for the two words that are potential candidates for pre-boundary lengthening in the four source recordings as well as in the resynthesized stimuli.

Figure 2: Distribution of durations (in milliseconds) for words in the ambiguous region that are candidates of pre-boundary lengthening: the adjunct verb (left) and the post-verbal noun (right). The duration distributions are plotted for recordings in the subject accent condition (top) and the verb accent condition (bottom). The source recordings are plotted in grey and the resynthesized stimuli which average between the source recordings are plotted in black and placed in the middle.

The duration of the adjunct clause verb and the post-verbal noun were resynthesized in Praat (Boersma & Weenik, 2020) to their averages between the early- and late-closure variants, using the early-closure recording as the source recording. After controlling for pre-boundary lengthening cues in this way, the end-of-sentence region (material after the post-verbal noun) from the subject-accent recordings was spliced into the verb-accent recordings to control for any sentence wrap-up effects between the two accent conditions. Lastly, pitch movement over the main clause was flattened to induce an unambiguous perception of accentual prominence in the adjunct clause. The pitch contours of the subject accent stimuli and the verb accent stimuli for a sentence used in the experiment are shown in Figures 3 and 4 respectively, where the dotted line after the ambiguous region marks the location of the splice.

Figure 3: Pitch contour of the subject-accent stimuli for While ANNA dressed the baby stopped crying.

Figure 4: Pitch contour of the verb-accent stimuli for While Anna DRESSED the baby stopped crying.

Because prominence on a word has consequences for the prosodic structure of the entire phrase, acoustic measures between the conditions were not further controlled in order to preserve naturalness. In other words, the overall shape of the pitch contour over the ambiguous region was unaltered, aside from controlling for pre-boundary lengthening cues. Figure 5 reports the measurements for pitch (max F0) on the adjunct verb and the post-verbal noun in both conditions from the resynthesized stimuli that were played to participants.

Figure 5: The mean, standard deviation, and distribution of max F0 values on the adjunct subject and the adjunct verb for the resynthesized stimuli in the two accent conditions.

For each sentence, a corresponding comprehension question similar to that used in Christianson et al. (2001) was constructed. These questions asked whether the post-verbal noun was the direct object of the adjunct clause verb, such as in (26). Because the post-verbal noun in the critical sentences was always the subject of the matrix clause, the correct answer to all the comprehension questions for the critical trials was NO.

In addition to the critical items, forty-eight distractor sentences were recorded by the same speaker to balance the transitivity bias of the adjunct verb (ambiguous, obligatorily transitive, obligatorily intransitive), the location of prosodic prominence (first noun, first verb, second noun), and the sentence’s syntactic structure (early- vs. late-closure of the adjunct clause VP) across all trials presented to participants in the experiment. The comprehension questions were also overall balanced in the correct answer choice and varied in the syntactic position of the entity being asked about. Lastly, two catch trials which simply instructed participants to answer YES or NO to the following question were included to filter out inattentive listeners.

As discussed previously, our hypothesis predicts a significant negative effect of verb pitch accent on comprehension accuracy. If a pitch accent on the verb leads to an interpretation of focus that triggers deeper semantic processing for the incorrect analysis, listeners should be more likely to answer the comprehension question in a way that is consistent with that erroneous interpretation. For example, after listening to the sentence While Anna DRESSED the baby stopped crying, participants are predicted to incorrectly answer YES to the question Did Anna dress the baby? more frequently than in the corresponding utterance with an accent on the subject ANNA. Importantly, this effect of pitch accent location on accuracy should reach significance after controlling for transitivity bias and semantic fit. Given previous findings for these two factors contributing to the persistence of the merely local interpretation, we also expect a transitively biased adjunct verb followed by a plausible direct object candidate to conspire to strengthen the initial misanalysis, making misinterpretation more likely.

3.3 Procedure

Each experiment session began with two familiarization trials in which participants listened to two sentences with the same structure as the critical sentences, presented auditorily with no accompanying text.¹⁰ Following the familiarization trials, the session continued with 74 trials, including 24 critical trials and 48 distractor trials, presented in pseudorandomized order that evenly spread out the critical trials. Each trial began with the audio presentation of a recorded sentence, with a blank screen, after which participants were given 5000ms to answer a comprehension question about the sentence that they just heard. The question was presented in written form and participants responded by clicking “YES” or “NO” on the screen. If an answer was not given within the time window, the next trial automatically proceeded. There were also two randomly sequenced catch trials which simply explicitly instructed participants to answer YES or NO to the following comprehension question. The experiment was run in a within-participant design where participants heard the critical sentences in twelve subject-accent and twelve verb-accent conditions. The experiment lasted around 10 to 15 minutes. The answers and the response times to the comprehension questions were recorded for analysis.

3.4 Results

Prior to analysis, critical trials where participants timed-out or answered the comprehension question in under 500ms were excluded (<2%).¹¹ Additionally, out of 61 participants, one was excluded due to low accuracy (<75%) on the distractor trials (mean accuracy on the distractor trials for the included participants was 95%). Data from the remaining sixty participants were included in the analysis.

The grand mean of within-speaker accuracies was above chance at 50% but well below ceiling in both the verb accent (78%) and subject accent (83%) conditions, replicating the semantic persistence effect. To test whether the difference between the two conditions was significant after controlling for the lexical-level predictors, a logistic mixed-effects regression model was fitted to the accuracy of the responses using the lme4 package in R (Bates et al., 2015), with pitch accent condition, semantic fit, and transitivity bias as fixed effects without interaction terms, and subject and item as random effects.¹² Convergence issues due to overparameterization of the random effects structure were addressed through the iterative model reduction process suggested in Bates et al. (2018), whereby the correlation between random effects was dropped first and individual random effects were removed in the order of lowest variance until the model converged, in order to keep the random effects structure as informative as possible. The final model included a random slope for Pitch by subject, and random intercepts by item and by subject.¹³ To test the significance of each predictor, the full model was compared via a log likelihood-ratio test to a depleted model with the predictor of interest removed.¹⁴ Table 1 reports the log-odds estimates and standard errors from the final model and the chi-squared statistics and p-values from the log likelihood-ratio tests.

A post-hoc analysis of response times (mean = 2080, SD = 759) was also conducted to determine whether listeners were garden-pathed in both pitch accent conditions, as intended by the resynthesis design. Figure 6 shows the distribution of response times aggregated by item between the two conditions, where each line tracks an item’s mean response time across conditions. The response time distribution for a subset of the filler items that were non-garden-path sentences of similar length with an unambiguously intransitive adjunct verb (e.g., When Tyler sneezed, the driver passed a tissue) is also plotted to the right for comparison.

Figure 6: Response time distributions of the critical trials by condition (raincloud plot to the left) and a subset of the filler trials (boxplot to the right). Each dot represents the mean response time of an item. The slopes of the lines connecting the dots represent the effect of pitch accent condition on response time.

In addition, a linear mixed-effects regression model was fitted to log-transformed response time with pitch accent placement, semantic fit, transitivity bias, and accuracy as fixed effects without interaction terms, and subject and item as random effects with varying intercepts.¹⁵ Table 2 reports the estimates, standard errors, and t-values from the model. The model estimates a large effect on response time of semantic fit and transitivity bias such that responses are slower for items with high semantic fit and high transitivity bias scores, whereas the location of the pitch accent has a small effect on response time.¹⁶