Glossa Psycholinguistics

1.1 The role of implicit prosody in ambiguity resolution

While listeners can rely on prosodic cues when processing sentences, readers do not have access to prosody. However, Fodor (2002a, 2002b) proposed the hypothesis that the parser internally generates default prosody, which is then mapped onto a syntactic representation to help decide which analysis to accept (i.e. the Implicit Prosody Hypothesis; see also Bader, 1998; Breen, 2014; Frazier & Gibson, 2015; Hirose, 1999; Swets et al., 2007).

Implicit prosody cannot be measured directly, because it is an internal construct of the parser. However, its existence can be attested by examining whether the prosodic factors that influence the explicit prosody of a sentence also modulate syntactic processing in silent reading. If the parser generates a prosodic structure internally, it is expected to guide the resolution of temporal ambiguity.

1.2 Lexical accent and prosodic phrasing in Japanese

Before going into the details of the previous evidence for implicit prosody, we will describe the prosodic system in Japanese, in order to help the reader understand the critical manipulation of the experiment. In Japanese, the presence or absence of pitch accent (a pitch fall) is specified lexically. For example, Hosoka⌝wa and Moriyama are both family names, but only the former is accented.¹ The difference in accentedness is important, because at most one accented word can form a minor phrase. Specifically, accented words such as Hosoka⌝wa in (1a) must form an independent minor phrase, while unaccented words such as Moriyama in (1b) can form a minor phrase together with the following word (Fujisaki & Sudo, 1971; Kubozono, 1993; McCawley, 1968; Poser, 1984). This difference can be seen in the presence of the phenomenon called initial lowering, whereby the first mora of the minor phrase is realised as a low (L) tone followed by a high (H) tone. As indicated in boldface in (1), the second word Morishi⌝ta shows an initial lowering pattern in (1a), because it is the onset of a minor phrase, given that Hosoka⌝wa forms an independent minor phrase. In contrast, it does not show initial lowering in (1b).

Two minor phrases (e.g. Hosoka⌝wa and Morishi⌝ta), as shown in (1a), form one major phrase (Beckman & Pierrehumbert, 1986; Poser, 1984; Selkirk & Tateishi, 1991). This can be diagnosed using downstep, a phenomenon whereby the pitch peak of accented/unaccented words is realised as lower when they are preceded by an accented word within the same major phrase (compared to when they are preceded by an unaccented word). Specifically, the second word (Morishi⌝ta) shows downstep, but the third word (new.medicine-acc) does not, because the latter is at the beginning of another major phrase. The major phrase boundary provides important cues for parsing, because it often aligns with syntactic boundaries (e.g. Selkirk & Tateishi, 1991).

1.3 Experimental evidence for implicit prosody

Hirose (2003) conducted a self-paced reading experiment in which Japanese speakers silently read sentences such as (2) phrase-by-phrase. The lexical accent of the matrix subject was manipulated so that it forms a major phrase in (2a) but not in (2b). Since Japanese speakers have a strong preference for the single-clause analysis up to the first verb trust-pst, the head of the relative clause friends-dat produces a garden-path effect in both types of sentences (Inoue, 2006). When required to reanalyse the single-clause structure to the relative-clause structure, Japanese speakers need to take the matrix subject out of the clause, because the subject who truly trusted the new medicine is friends. Thus, in (2a), the major prosodic boundary aligns with the structural boundary, whereas in (2b), it does not. Hirose (2003) reported that the reading time of the head noun was significantly shorter in (2a) than in (2b). They interpreted this result as evidence that the major phrase boundary facilitated structural revision, because the major phrase boundary was ‘recycled’ as a relative-clause boundary.

One of the concerns with Hirose’s (2003) experiment was that the results could reflect the difference in the number of syntactic phrases in the stimuli (i.e. 8 phrases in (2a) and 7 phrases in (2b)). In self-paced reading tasks, participants tend to read the later part of sentences faster (e.g. Aoshima et al., 2009; Nakatani & Gibson, 2010). Thus, the faster reading times in (2a) may be due to a general speed-up effect. Furthermore, since two phrases were presented together only in the subject region of (2a), the participants may have been more likely to analyse them as a single chunk, separately from the rest of the sentence. In other words, they may have recognised the clause boundary more easily in (2a), due to an artefact of the presentation method.

2.1 Stimuli

The sentences in (3) show an example set of target sentences. The +Accent+Accent condition in (3a) and the –Accent condition in (3c) were prepared in a way similar to Hirose’s (2003) experiment.

In +Accent+Accent condition, the matrix subject forms a major phrase, because both names are accented words. The –Accent+Accent condition was created by replacing the first accented names of the +Accent+Accent condition with unaccented names. The unaccented name Moriyama-cnj forms a minor phrase with its following name Morishi⌝ta-nom, creating a minor phrase boundary. The difference in reading times between these conditions can be attributed to processing differences due to prosodic phrasing, since the number of syntactic phrases is the same. The –Accent condition in (3c) involves the same unaccented names as in (3b). This unaccented name combines with its following phrase new.medicine-acc to form a minor phrase.

For the manipulation of the lexical accent of the matrix subjects, we used common family names. Although Hirose (2003) used several types of nouns, including proper nouns (e.g. Hosokawa), occupational nouns (e.g. fucho ‘chief nurse’) and kinship terms (e.g. kaasan ‘mother’), semantically neutral names are more appropriate, because Japanese speakers have difficulty in recovering from the garden path effect when the subject noun is semantically biased towards the agent of a following verb phrase (Inoue, 2006). The names always had four morae, with the accent nucleus either on the third mora (i.e. +Accent) or absent (i.e. –Accent).

Following a Latin Square design, 18 triplets of the target sentences were distributed into three lists, so that each participant saw only one sentence of each triplet. These target sentences were intermixed with 54 filler sentences (36 of them were simple single-clause sentences, with the remaining 18 having various structures).

2.2 Participants

We recruited 32 native speakers of Tokyo-Japanese for our experiment (mean age = 20.4 years, SD = 1.5). Participants reported no history of language and/or neurological disorders. The experiment was approved by the Ethics Committee of Tokyo Metropolitan University (No. H6-126). Written informed consent was obtained from all participants before the experiment.

2.3 Procedure

Each phrase of the sentences was presented as a separate region, using PCIbex (Zehr & Schwarz, 2018), in contrast to Hirose’s (2003) experiment, in which the entire conjoined subject phrase was presented as a single region. We chose this method to avoid any effect of the presentation mode on the syntactic analysis.

Participants were asked to read each phrase silently and to press the space bar to move to the next phrase.² At the end of each trial, participants were presented with a comprehension question and asked to respond by pressing the ‘Yes’ or ‘No’ button. The response was followed by feedback indicating whether their answer was correct or incorrect.

2.4 Prediction

The region of primary interest is Region 6. Because the parser prefers the syntactically simpler, single-clause structure, the head noun of the relative clause would induce a garden-path effect in Region 6, where the relative-clause structure is found to be correct. If the presence of the major phrase boundary facilitates the computation of the relative-clause analysis, as Hirose (2003) proposed, the reading times should be faster in the +Accent+Accent condition than in the –Accent+Accent condition.

However, it is also possible that this difference reflects a difference in the reanalysed structure. In the +Accent+Accent condition, the parser only takes the matrix subject out of the clause, and, thus, it can preserve an existing structure as much as possible, as shown in (4). In contrast, it is possible that the parser reanalyses the structure by detaching the matrix subject and the object into a different clause from other clause mates, as shown in (5), given that the major clause boundary is placed after the object ‘medicine-acc.’ Since the late-opening structure is not preferred in Japanese (Mazuka & Itoh, 1995; Oishi et al., 2007), the effect of structure should be examined, in order not to confound it with the ‘recycling’ effect.

Region 7 serves this purpose. Here, prescription-acc signals unambiguously that it is the object of the matrix clause. If the parser chose the late-opening structure in the –Accent+Accent condition, it should experience another garden-path effect. This is due to a syntactic constraint in Japanese that restricts multiple accusative-marked arguments within the same clause (Harada, 1973). Consequently, the reading time should be longer in the –Accent+Accent condition than in the +Accent+Accent condition. In contrast, if the parser chose the early-opening structure in the –Accent condition, it should avoid a second garden-path effect in Region 7. This is because new.medicine-acc is analysed as an object inside the relative clause after the first garden-path effect in the early-opening structure, allowing prescription-acc to be easily incorporated as a matrix object. In the –Accent condition, a second garden-path effect should occur, because the parser is expected to choose the late-opening structure, given the absence of a prosodic boundary between Moriyama-nom and new.medicine-acc.

2.5 Data and statistical analysis

We analysed only the trials in which the comprehension question was answered correctly. Reading time data that exceeded 2.5 standard deviations from the mean at each region were discarded.

The reading times were log-transformed and analysed using linear mixed-effects (LME) models, with Accent as a fixed factor and participants and items as random factors. The fixed factor was coded with the Helmert scheme, as shown in Table 1. The effect of Type is a factor of primary interest in testing the role of the major phrase in facilitating structural reanalysis. The effect of Boundary (i.e. the presence vs. absence of the prosodic boundary after the matrix subject) involves the comparison between different numbers of phrases; thus, interpreting this effect requires caution. Item Order (the nth trial) was included as a covariate to assess task adaptation (e.g. Prasad & Linzen, 2021), but its significance was not of interest.

The maximal model was built using the lmer function of the lme4 package in R (Bates et al., 2015), and then compared with more parsimonious models with simpler random effects in the backward stepwise method (Matuschek et al., 2017). The p-values were calculated based on the Satterthwaite method (Kuznetsova et al., 2017).

2.6 Results

The accuracy of the task did not differ significantly among the three conditions (+Accent+Accent: 91.1% (SE: 2.06), –Accent +Accent: 86.5% (SE: 2.48), –Accent: 92.7% (SE: 1.88)). Figure 1 shows the results of the self-paced reading experiment.⁴ The LME model showed a marginally significant effect of Type in Region 6 (p = 0.052), indicating that the +Accent+Accent condition was read faster than the –Accent+Accent condition (Table 2). In Region 7, the effect of Boundary was significant, as reading times were shorter in the +Accent+Accent and –Accent+Accent conditions than in the –Accent condition (Table 3). There was no significant difference between the +Accent+Accent and –Accent+Accent conditions. Note that the apparent numerical trend suggests shorter reading times for the –Accent+Accent condition (i.e. no evidence for the second garden-path effect in this condition).

Figure 1: The mean reading times (ms). Error bars represent the standard error of the mean.

As Figure 1 shows longer reading times for the +Accent+Accent condition in Regions 1–3, an additional analysis was performed to examine this effect. The +Accent+Accent condition took longer to read than the other conditions in Regions 2 and 3 (Region 2: $\widehat{\beta }$ = –0.06, t = –1.89, p = 0.05; Region 3: $\widehat{\beta }$ = 0.08, t = 2.24, p < 0.05). No significant difference was found in Region 1.