Stress Matters Revisited: A Boundary Change Experiment

Abstract

Breen and Clifton (2011) argued that readers’ eye movements during silent reading are influenced by the stress patterns of words. This claim was supported by the observation that syntactic reanalysis that required concurrent metrical reanalysis (e.g., a change from the noun form of abstract to the verb form) resulted in longer reading times than syntactic reanalysis that did not require metrical reanalysis (e.g., a change from the noun form of report to the verb form). However, the data contained a puzzle: the disruption appeared on the critical word (abstract, report) itself, although the material that forced the part of speech change did not appear until the next region. Breen and Clifton argued that parafoveal preview of the disambiguating material triggered the revision, and that the eyes did not move on until a fully-specified lexical representation of the critical word was achieved. The present experiment used a boundary change paradigm (Rayner, 1975) in which parafoveal preview of the disambiguating region was prevented. Once again, an interaction was observed: syntactic reanalysis resulted in particularly long reading times when it also required metrical reanalysis. However, now the interaction did not appear on the critical word, but only following the disambiguating region. This pattern of results supports Breen and Clifton's claim that readers form an implicit metrical representation of text during silent reading.

In recent years, there has been increasing interest in questions about the nature of phonological representations during silent reading. This trend began with proposals by Fodor (1998) and Bader (1998), who argued that readers generate an implicit prosodic representation during silent reading; that is, even during silent reading, there is a voice in readers’ heads ‘reading aloud’. Support for the Implicit Prosody Hypothesis (Fodor, 1998) includes demonstrations that syntactic parsing decisions can be influenced by implicit prosodic phrasing (Hwang & Steinhauer, 2011; Hwang & Schafer, 2009) and implicit sentence rhythm (Kentner, 2012).

The current study was designed as a follow-up to Breen and Clifton (2011), who investigated whether readers form metrical representations of words during silent reading. To do so, they employed stress-alternating noun-verb homographs, which are words that can serve as nouns or verbs depending on their stress pattern. For example, the word abstract, when produced with stress on the first syllable (ABstract), is a noun, but when produced with stress on the second syllable (abSTRACT), it is more likely to be interpreted as a verb. In two experiments with very different manipulations, Breen and Clifton induced readers to generate expectations about the stress patterns of these ambiguous words, and demonstrated, in both cases, a reading time cost when these expectations were not met.

In their first experiment, Breen and Clifton had participants read limericks in which they had placed stress-alternating homographs. The stress pattern of the alternating homograph was either consistent or inconsistent with the stress pattern of the limerick. For example, in (1a), the lexical stress pattern of the noun form of present is strong-weak (PREsent), which is consistent with the stress pattern of the limerick. Conversely, in (1b), the stress pattern of the verb form of present is weak-strong (preSENT), which is inconsistent with the stress pattern of the limerick. Eye-tracking results demonstrated longer reading times on the critical word present in (1b) than in (1a), suggesting that readers could not move on from the critical word until they had accessed its correct metrical form.

• 1
  1. There once was a penniless peasant Who couldn't afford a nice present
  2. There once was a penniless peasant Who went to his master to present

In a second experiment, Breen and Clifton placed stress-alternating homographs in garden-path contexts like (2). In the critical sentence (2a), in order to resolve the syntactic garden path, participants had to also engage in metrical reanalysis. That is, although both (2a) and (2b) require syntactic reanalysis when the reader encounters the disambiguating material (i.e., the best ideas), only (2a) requires a simultaneous metrical reanalysis, such that the reader needs to change the stress pattern from strong-weak ABstract to weak-strong abSTRACT.

• 2
  1. The brilliant abstract the best ideas from the things they read.
  2. The brilliant report the best ideas from the things they read.
  3. The brilliant abstract was accepted at the prestigious conference.
  4. The brilliant report was accepted at the prestigious conference.

These sentences were compared with their non-garden-path counterparts (2c) and (2d), which require neither syntactic nor metrical reanalysis.

Breen and Clifton observed the expected effect, in the form of longer reading times for simultaneous metrical and syntactic reanalysis (2a) than for syntactic reanalysis (2b) alone, with no corresponding difference between (2c) and (2d). However, the data contained a puzzle: The interaction appeared on the critical word (abstract) itself, although the material that forced the part of speech change (the best ideas) appeared in the next region of the sentence. In order to explain this unexpected data pattern, Breen and Clifton argued that parafoveal preview of the disambiguating material triggered the revision, and that the eyes did not move on from the critical word until a fully-specified lexical representation of the critical word was achieved.

There is ample evidence that the parafoveal preview of upcoming information influences reading behavior. For example, Drieghe and colleagues (Drieghe, 2008; Drieghe, Rayner, & Pollatsek, 2005; Rayner, Slattery, Drieghe, and Liversedge, 2011) have demonstrated that the more parafoveal processing of an upcoming region that a reader has, the more likely the reader is to skip that region. More controversially, Gordon, Plummer, and Choi (2012) have argued that skipping of words in the parafovea happens only when those upcoming letter strings are recognized as words.

In the Breen and Clifton (2011) materials, the disambiguating region immediately past the critical word was generally short, often a function word such as the or was, and presumably could be recognized parafoveally with reasonable frequency. Breen and Clifton assumed that skipping the disambiguating region would indicate that it had been processed, and in fact recognized, while the reader was fixating the critical word. They observed that disruption on the stress-shifting critical word indeed appeared only when the disambiguating region was skipped which accounts for the surprising location of the disruption.

The goal of the current experiment was to explicitly test Breen and Clifton's claim that parafoveal preview of the disambiguating material in their second experiment led to the very early observed effects of metrical reanalysis. To do this, we adopted the boundary change paradigm proposed by Rayner (1975). With this method, readers are denied correct preview of upcoming sentence material. Specifically, as readers are fixating region n, the information in region n+1 is incorrectly displayed. Although readers are often unaware of the manipulation, reading times demonstrate that incorrect preview of region n+1 results in slower reading times on that region when it is subsequently fixated, compared to conditions in which readers have correct preview of region n+1 (for reviews see Rayner, 1998, 2009; Schotter, Angele, & Rayner, 2012).

To implement this manipulation, we conducted an eye-tracking experiment with materials based heavily on those used in Breen and Clifton (2011)'s second experiment. The critical difference was that, in the current experiment, readers were subjected to incorrect preview of the critical disambiguating material, as shown in Table 1. If parafoveal preview of the disambiguating sentence material in Breen and Clifton (2011) allowed readers to engage in metrical reanalysis of the critical word while they were fixating it, then denying readers preview of the disambiguating material should delay metrical reanalysis until readers actually fixate the disambiguating region. In this case, the interaction between the need for syntactic and metrical reanalysis, observed before the disambiguating material by Breen and Clifton (2011), should appear only on or after the disambiguating region. Specifically, we predict that syntactic reanalysis will take the form of longer reading times on the disambiguating region (i.e., the best ideas) in 2a and 2b than on the same region (i.e., was accepted) in 2c and 2d. Moreover, we predict that metrical reanalysis will be reflected in an interaction that will take the form of longer reading times on the disambiguating region in 2a than the same region in 2b, but no corresponding difference between reading times on the same region in conditions 2c and 2d.

Table 1.

Sample Set of sentences used in the experiment.

Condition	Disambiguation	Critical Word	Example	Preview
a	Noun	Alternating	The brilliant^l abstract^2 was accepted^3 at the prestigious conference.^4	men nascgfut
b	Noun	Non-alternating	The brilliant^1 report^2 was accepted^3 at the prestigious conference.^4	men nascgfut
c	Verb	Alternating	The brilliant^1 abstract^2 the best ideas^3 from the things they read.^4	blu droh fbuwc
d	Verb	Non-alternating	The brilliant^1 report^2 the best ideas^3 from the things they read.^4	blu droh fbuwc

Note. The material in the final column was substituted for the underlined region until the reader fixated that region. The incorrect preview was matched to the correct sentence material on ascenders and descenders. Superscripts in the Examples indicate region numbers.

Method

Participants

Twenty-two participants contributed data to the analyses. All were undergraduates at the University of Massachusetts Amherst. Thirty-five more participants were run but excluded from analysis. Fifteen were excluded because they could not be tracked by the eye-tracker; another twenty completed the experiment but failed to contribute a minimum of ten trials per condition after trials containing eyetracking problems were eliminated. All participants spoke English as their native language, and had normal or corrected-to-normal vision. All received extra course credit in exchange for participation.

Materials

Thirty-two sets of items like those in Table 1 were constructed based on the sentences from Experiment 2 in Breen and Clifton (2011), supplemented by a few homographs from their Experiment 1. Twenty-one of the sets had the same initial regions (adjective/noun and noun/verb) as used in the earlier paper, while minor changes were made in the remaining 11 items to make them more natural sounding. The full set of items can be found in the Appendix.

All four versions of each item began with The. The second word was ambiguous between an adjective and noun interpretation, but was biased by frequency of occurrence towards an adjective reading (e.g., brilliant). Celex counts (Baayen, Piepenbrock, & van Rijn, 1993) confirmed the adjective bias, such that these words occurred more often as adjectives (36 per million) than nouns (1.34 per million), t = 3.5, p < .01. The third word was the critical word, which was ambiguous between a noun and verb interpretation. Specifically, it was either a stress-alternating noun-verb homograph (e.g., abstract, which has a strong-weak pattern in its noun form, and a weak-strong stress pattern in its verb form) or a non-alternating noun-verb homograph (e.g., report). Celex counts confirmed this balance, such that the stress-alternating homophones occurred as often as nouns (28 per million) as verbs (23 per million), t = .57, p = .57; in addition, the non-alternating homophones occurred as often as nouns (74 per million) as verbs (47 per million), t = 1.71, p = .10 (see Table 2). Moreover, norming procedures reported in Breen & Clifton, 2011 for the 21 items that had the same initial words as the current items showed that subjects overwhelmingly completed fragments like The brilliant abstract or The brilliant report in ways that indicated they took the fragments to be adjective-noun sequences.

Table 2.

Mean values (and standard deviations) of lengths (in number of letters) and frequencies (in occurrences per million words) of critical words by condition

Word type	Length	Celex Frequency
Alternating homophone as Noun	6.6 (0.98)	28 (42)
Non-alternating homophone as Noun	6.7 (0.86)	74 (79)
Alternating homophone as Verb	6.6 (0.98)	23 (23)
Non-alternating homophone as Verb	6.7 (0.86)	47 (54)

The continuation of conditions (a,b) disambiguated the critical homograph as a noun. The continuation of critical conditions (c,d) disambiguated the critical homograph as a verb. The word immediately following the critical homograph generally disambiguated it, as it was typically a verb, a determiner, or a preposition. This word, plus the following two or three words when needed to create a region of eight or more characters in length, were treated as the disambiguating region in analyses. The letters in the disambiguating region were replaced with quasi-random letters that matched the original letters as ascenders or descenders and also maintained the spacing pattern of the original letters.

Conditions were distributed across lists in a Latin-square design so that each participant saw two conditions from each experimental item. Importantly, these two conditions resolved the noun/verb ambiguity in different ways. Specifically, condition (a) (Alternating Verb) was always paired with condition (d) (Non-alternating Noun). Similarly, condition (b) (Non-Alternating Verb) was always paired with condition (c) (Alternating Noun). In this way, the only overlap between the two conditions was the first two words, meaning that participants were unlikely to notice the repetition of items.

Included with the experimental items were 32 fillers, which had similar structures to the experimental items in that they began with the determiner The and continued with an ambiguous adjective/noun (e.g., poor). However, the fillers differed from the experimental items in that the third word was always unambiguously a noun (e.g., country). In this way, two-thirds of all sentences (critical items and fillers) were disambiguated, either early or late, as Determiner-Adjective-Noun, so that readers were encouraged to interpret the ambiguous noun-verb homographs as nouns. In addition, there were 30 items from another experiment which contained no local ambiguities. While the fillers contained display changes, these 30 items did not.

Procedure

Participants were tested individually. After providing written informed consent, participants were taken into the testing room, shown the eyetracker, and given instructions to read the sentences normally and occasionally answer questions about what they read. Participants were seated 55 cm from a CRT monitor that displayed the sentences. Their eye movements were recorded using an EyeLink 1000 (SR Research, Toronto, Ontario, Canada) eyetracker controlled by a PC running the University of Massachusetts EyeTrack software (http://www.psych.umass.edu/eyelab). Eye position was sampled at 1kHz. Although participants viewed the stimuli with both eyes, only data from the right eye were collected.

Before beginning the experiment, the experimenter calibrated the eyetracker, recalibrating between trials when necessary. Trials were initiated when the participant fixated on a box at the left edge of the screen. After this fixation, the sentence appeared as a whole, on a single line, and participants’ eye movement locations and times were recorded. The eye contingent boundary technique was used, and all changes on trials included in analysis occurred within five ms of the participants’ eyes crossing the invisible boundary. Each character subtended a visual angle of .32 degrees. When they had finished reading a sentence, participants were instructed to look to a colored square on the right edge of the monitor and to pull a trigger on a game controller. Questions, which appeared after 25% of items, appeared as a whole immediately after this trigger-pull. Participants answered “yes” or “no” by pulling one of two triggers on the game controller. The experiment lasted between 30 and 45 minutes.

Results

Overall question-answering accuracy was 87%, (SD = 8.8%) indicating that the participants were engaged in the experimental task.

Before being subjected to analysis, the eye-tracking data were cleaned of blinks, track loss, and extremely long fixations, with fixations in excess of 800 ms being deleted (Rayner, Sereno, Morris, Schmauder & Clifton, 1989). Short fixations were incorporated into the nearest fixation within 3 characters; if there was no close fixation, the short fixation was deleted. In addition, trials in which the change from incorrect to correct preview was either too slow or too fast were deleted (20% of all trials).

In the following, we report the following six standard eye-tracking measures: first fixation, first pass time, go-past time, second-pass time, proportion of regressions in, and proportion of regressions out (Rayner, 1998; Rayner, et al. 1989). First fixation time is the duration of the first fixation anywhere in the region. First-pass time is the sum of all fixation durations made from first entering to first leaving a region, eliminating trials on which no such fixations occurred. Go-past time is the sum of all fixation durations made from first entering a region to first leaving it to the right. Second-pass time is the sum of all fixation durations made on a region after first leaving it to the right (and in order to reflect the actual time spent in re-reading, trials on which re-reading does not occur are scored as zero durations). The proportion of regressions into a region is computed as the probability of refixating a region after leaving it to the right. Finally, the proportion of regressions out of a region is the probability of regressing out of that region given that it was fixated during the first pass.

First fixation, first pass, and go-past times were analyzed on a single-trial basis in a series of linear mixed-effects models using the languageR package (Baayen, 2008) implemented in the R statistical programming language (R Core Development Team, 2012). Models of fixation proportions, including the proportion of trials in which readers regressed out of, or into, a region, were analyzed in a series of mixed-effects logistic regressions. Three fixed effects, as well as their interactions, were included in all models. These fixed effects were Part of Speech (disambiguation as a noun or verb), Stress Alternation (whether or not the stress pattern differed for the noun and verb versions of the target), and Sequence (trial number). The first two effects are the effects of interest as they reflect syntactic and metrical reanalysis, respectively. The Sequence effect was included to determine whether the other two main effects, or their interaction, changed over the course of the experiment. All three fixed effects were centered. In addition to the fixed effects, we also included subject and item intercepts as random effects (note that the disambiguating material following the critical word defined an item, and thus, Part of Speech was a between-items effect). We report models in which subject and/or item slopes for the fixed effects improved model fit, as determined by a chi-square test comparing the more complex model to a simpler, nested version (Baayen, 2008). For models with only subject and item intercepts as random effects, we present significance levels estimated using Markov Chain Monte Carlo sampling (Baayen, Davidson, and Bates, 2008). When the best model included random slopes, MCMC sampling is not available; therefore, we interpret t-values of 2.0 or greater as significant.While Sequence affected most measures (with faster reading later in the experiment), it seldom interacted with the other factors. In the interests of simplicity, effects of Sequence will not be reported, aside from the significant interactions. The parameters of the best-fitting models are presented in Table 4 (excluding Sequence parameters).

Table 4.

Parameter estimates of models

Fixed Effect	Est	SE	t/z	pMCMC	Est	SE	t/z	pMCMC	Est	SE	t/z	pMCMC	Est	SE	t/z	pMCMC
	Region 1				Region 2				Region 3				Region 4
	First Fixation (ms)
Intercept	208	9.61	21.73	*	257	8.43	30.47	*	246	6.59	37.30	< .001	233	5.25	44.36	< .001
Noun vs. Verb	-0.65	6.72	-0.10	n.s.	7.73	6.78	1.14	n.s.	3.77	6.89	0.55	0.81	3.41	6.39	0.53	0.57
NonAlt vs. Alternat	-1.18	5.78	-0.20	n.s.	23.06	7.44	3.10	*	0.91	4.91	0.19	0.60	7.43	5.54	1.34	0.17
NV* Alternating	-8.05	12.83	-0.63	n.s.	0.21	11.34	0.02	n.s.	-15.17	9.82	-1.54	0.15	4.16	11.09	0.38	0.68
	First Pass (ms)
Intercept	263	19.31	13.63	< .001	319	12.37	25.85	*	454	16.96	26.75	*	672	47.29	14.22	*
Noun vs. Verb	-5.21	12.11	-0.43	0.64	15.75	10.89	1.45	n.s.	-8.10	18.46	-0.44	n.s.	97.62	57.61	1.69	n.s.
NonAlt vs. Alternat	3.49	7.96	0.44	0.85	26.45	14.77	1.79	^	5.98	17.89	0.33	n.s.	-17.30	22.76	-0.76	n.s.
NV* Alternating	-6.94	15.93	-0.44	0.69	-2.95	28.58	-0.13	n.s.	9.25	28.58	0.32	n.s.	36.95	45.38	0.81	n.s.
	Go Past (ms)
Intercept					4067	23.71	17.15	< .001	780	56.04	13.67	< .001	1086	101.11	10.74	*
Noun vs. Verb					35.60	19.59	1.82	0.09	273	57.17	4.78	< .001	377	95.65	3.94	*
NonAlt vs. Alternat					25.05	15.15	1.65	0.08	43.93	27.51	1.60	0.09	37.40	56.09	0.67	n.s.
NV* Alternating					-23.0	30.32	-0.76	0.56	18.33	55.06	0.33	0.64	248.97	112.22	2.22	*
	Regressions Out (Prop)
Intercept					-1.97	0.16	-12.10	< .001	-0.73	0.21	-3.39	< .001	-0.80	0.17	-4.79	< .001
Noun vs. Verb					0.19	0.23	0.81	0.42	1.32	0.25	5.26	< .001	0.43	0.16	2.72	0.006
NonAlt vs. Alternat					-0.11	0.18	-0.61	0.54	0.17	0.18	0.98	0.32	0.06	0.14	0.43	0.70
NV* Alternating					-0.29	0.37	-0.80	0.42	-0.37	0.31	-1.19	0.26	0.35	0.27	1.26	0.21
	Regressions In (Prop)
Intercept	-0.93	0.24	-3.91	< .001	-0.46	0.20	-2.29	0.02	-1.62	0.16	-9.94	< .001
Noun vs. Verb	0.80	0.20	3.92	< .001	1.31	0.19	7.09	< .001	-0.02	0.21	-0.09	0.93
NonAlt vs. Alternat	-0.31	0.21	-1.50	0.14	0.26	0.14	1.92	0.06	0.03	0.16	0.21	0.84
NV* Alternating	0.24	0.28	0.88	0.38	-0.14	0.27	-0.60	0.71	-0.03	0.32	-0.10	0.92
	Second Pass (Subjects) (ms)
Intercept	150	22.24	6.73	*	184	24.47	7.51	*	108	17.45	6.21	*
Noun vs. Verb	115	20.40	5.65	*	148	14.58	10.12	*	52.55	17/76	2.96	*
NonAlt vs. Alternat	0.35	18.72	0.02	n.s	40.21	14.59	2.76	*	13.51	19.30	0.70	n.s
NV* Alternating	25.95	41.71	0.62	n.s	54.12	29.17	1.86	^	62.78	34.79	1.81	^
	Second Pass (Items) (ms)
Intercept	151	10.27	14.74	*	185	10.77	17.18	*	109	10.73	10.16	*
Noun vs. Verb	121	20.53	5.88	*	155	21.54	7.21	*	58.08	21.46	2.71	*
NonAlt vs. Alternat	-2.22	19.37	-0.12	n.s	37.93	18.98	2.00	*	10.35	17.12	0.60	n.s.
NV* Alternating	19.71	38.73	0.51	n.s	60.74	37.96	1.60	^	58.25	34.25	1.70	^.

Note. *: or ^{^} or n.s.: random slopes were used in model, sopMCMC cannot be calculated

^*indicates estimated significance beyond the .05 level

^{^}indicated marginal significance

Because individual second pass times have an extremely skewed distribution, they were averaged over items and averaged over subjects, and the resulting averages were analyzed in separate mixed model analyses with subjects and with items as random factors (including random slopes when they improved the model fit). The first analyses to be reported reflect the initial reading of each region together with regressions out of the region. Analyses of re-reading times will be presented later.

Region 1: The first region of analysis consisted of the determiner and the temporarily ambiguous adjective/noun. As this region was identical across all conditions, we predicted no differences across conditions in early reading time measures for the main effects of Part of Speech and Stress Alternation. We observed no main effects or interactions on first fixation times, ts <1 (see Table 3 for reading times, and Table 4 for model parameters). In addition, we observed a significant interaction of Sequence and Stress Alternation on first pass time, t = -2.79, as well as a significant 3-way interaction of Sequence, Stress Alternation, and Part of Speech. t = -2.51. The material was identical across all conditions in this region, suggesting that these two interactions may be spurious, but since each temporarily ambiguous adjective/noun was used twice in the experiment, the interaction could reflect readers’ memory for the previous occurrence of the word.

Table 3.

Mean Values (and standard errors) of Eyetracking Measures

	Region 1		Region 2		Region 3		Region 4
Condition	Noun	Verb	Noun	Verb	Noun	Verb	Noun	Verb
				First Fixation
Alternating	212 (6)	207 (5.0)	266 (6)	274 (7)	248 (5)	247 (5)	235 (5)	240 (7)
Non-alternating	211 (5)	212 (6.0)	244 (5)	249 (6)	243 (5)	249 (6)	230 (5)	230 (6)
				First Pass
Alternating	276 (10)	265 (10)	328 (11)	343 (11)	464 (14)	451 (15)	625 (26)	734 (27)
Non-alternating	274 (9)	272 (10)	299 (8)	316 (9)	461 (11)	443 (13)	659 (27)	718 (27)
				Go-Past
Alternating			410 (17)	432 (18)	669 (27)	941 (41)	867 (37)	1384 (68)
Non-alternating			372 (14)	414 (17)	640 (25)	883 (36)	976 (44)	1183 (51)
				Regressions Out
Alternating			0.14 (0.02)	0.13 (0.020)	0.28 (0.027)	0.50 (0.030)	0.27 (0.030)	0.40 (0.029)
Non-alternating			0.13 (0.02)	0.17 (0.022)	0.21 (0.024)	0.50 (0.030)	0.31 (0.030)	0.35 (0.029)
				Regressions In
Alternating	0.26 (0.026)	0.46 (0.030)	0.31 (0.03)	0.58 (0.029)	0.19 (0.023)	0.19 (0.023)
Non-alternating	0.31 (0.028)	0.47 (0.030)	0.26 (0.03)	0.52 (0.030)	0.20 (0.024)	0.18 (0.023)
				Second Pass
Alternating	86 (11)	213 (18)	113 (12)	293 (21)	72 (10)	155 (19)
Non-alternating	102 (11)	204 (20)	106 (12)	225 (16)	92 (12)	113 (16)

Region 2: The second region consisted of only the target word (abstract, report) which was ambiguous between a noun and verb interpretation. On this region, there were effects of stress alternation, with longer reading times on alternating homographs (abstract) than non-alternating homographs (report). This effect was significant for first fixation time and marginal for first pass and go-past times. This result may be due to the fact that the non-alternating homographs were more frequent than the alternating homographs. There was, in addition, a suggestion of an effect of Part of Speech for go-past times such that the reading time cost was higher when the target was going to be disambiguated as a noun than when it was going to be disambiguated as a verb. However, as the material that the reader could see did not distinguish between noun and verb disambiguation, we conclude that this effect is spurious.

Region 3: This region contained material which served to disambiguate the target word as a noun or verb. There were no main effects or interactions on first fixation times. However, there was a main effect of Part of Speech on both go-past times and on the probability of a regression out of the region. This result indicates that readers first interpreted the homograph in Region 2 as a noun and then had to revise their analysis upon encountering the disambiguating information in Region 3. Moreover, there was a marginal effect of stress alternation on go-past times in this region, as in Region 2, with longer reading times on the stress-alternating homographs than the non-alternating homographs. However, the interaction that appeared prior to the disambiguating region in Breen and Clifton (2011) did not appear in Region 3. Finally, we observed interactions of Part of Speech and Sequence in go-past times, t = -2.45, and in regressions out, t = -2.11, indicating that, while the Part of Speech (i.e., garden-path) effect remained substantial over the course of the experiment, its absolute size decreased in later trials.

Region 4: This region was the final region of the sentence as well as the spillover region for reading time effects from the disambiguating Region 3. We observed a main effect of Part of Speech on go-past times, and on the probability of a regressive eye movement, such that readers took longer to read this region, and were more likely to regress out of it, in sentences where the ambiguous target had been disambiguated as a verb than when it had been disambiguated as a noun. Critically, we now observed the predicted interaction of Part of Speech with Stress Alternation such that go-past times on this region were even longer when the change to a verb required a concomitant change in stress pattern from strong-weak stress (ABstract) to weak-strong stress (abSTRACT), (Figure 1). Finally, we observed two interactions of Sequence with Stress Alternation, though in opposite directions. Specifically, in first pass time, we observed an interaction such that, although reading time following a Non-alternating homographs did not change over the course of the experiment, reading times following Alternating homographs decreased, t = -2.84. Conversely, we observed an interaction in the probability of regressive eye movements due to the fact that, although the probability of a regression decreased over the course of the experiment for both Non-alternating and Alternating homographs, it decreased more for the Non-alternating homographs, t = 2.12. We suspect that these counteracting effects reflect a change over trials in the tradeoff between regressing out of a problematic region and refixating that region (Rayner & Sereno, 1994). We note that the critical and significant go-past time interaction between Part of Speech and Stress Alternation did not significantly change across trials.

Figure 1.

Go-past times on the disambiguating region (Region 3) and the spillover region (Region 4). Error bars represent the standard error of the mean. The target word (in Region 2) was either Stress Alternating (e.g., abstract) or Non-alternating (e.g., report).The material in Region 3 disambiguated the target as either a noun (e.g., was accepted) or as a verb (e.g., the best ideas).

Re-reading times

Re-reading times include second-pass times and the probability of a regression back into a previously read region. Because second-pass times are often zero, their distribution does not meet assumptions of normality. Therefore, to analyze these data, we averaged individual trials over subjects and items, and submitted both sets of averages to separate mixed-effects model analyses.

Region 1: The late reading time measures, including second pass times by subjects and by items and the probability of regression into the region, demonstrated significant costs of verb disambiguation. These effects indicate that the syntactic manipulation was successful, in that readers regressed and reread more often when the target was disambiguated as a verb.

Region 2: The results in this region were similar to that of the previous region. Second pass times (over subjects and items), and probability of regression into the region demonstrated significant main effects of syntactic reanalysis. Also, the critical word was re-read more frequently and for a longer time when it had a stress alternation than when it did not, and the marginal interactions between Part of Speech and Stress Alternation in second pass times suggested that the effect of Stress Alternation was larger when syntactic reanalysis to a verb was required than when it was not.

Region 3: As observed in Region 2, second pass times were longer when disambiguation was toward a verb than toward a noun. Further, we observed marginal interactions of Part of Speech and Stress Alternation on second pass times, reflecting the significant go-past time interaction observed in Region 4.

Discussion

The current experiment was a replication of Breen and Clifton (2011), which demonstrated that readers compute a metrical representation of words during silent reading. In addition, the experiment tested Breen and Clifton's claim that the early observed effects of metrical reanalysis were the result of readers using parafoveal preview of disambiguating information to correctly determine the meaning of the target word while they were fixating it.

The results demonstrate a successful replication of Breen and Clifton, as evidenced by the fact that syntactic reanalysis which required simultaneous metrical reanalysis (e.g., changing from the noun form of abstract to the verb form) led to longer reading times than syntactic reanalysis that did not require metrical reanalysis (e.g., changing from the noun form of report to the verb form). We suggest that eye movements reflect the creation of a full and coherent linguistic representation of material that is read, and that the requirement to change any part of the representation – syntactic, morphological, or phonological, including metrical – disrupts eye movements.

Importantly, in contrast to Breen and Clifton's second experiment, the interactive effect of syntactic and metrical reanalysis did not appear until readers had an accurate display of the disambiguating information. In fact, the critical interaction of Part of Speech and Stress Alternation did not appear until the final region of the sentence. This may be surprising, in that the effect of Part of Speech did appear in go-past times in the disambiguating region, indicating that syntactic reanalysis was initiated immediately. The delayed effect of the requirement to change stress of the critical word presumably reflects the slowed lexical processing observed when preview is blocked, but is also reminiscent of an effect that Folk and Morris (1995) observed in reading heterophonic homophones such as tear. Their data indicated the disruption of eye movements when following material forced the normally-unpreferred interpretation of such an item was delayed, appearing largely as re-reading of the heterophone itself. Thus, although reading of a word that requires a change in stress assignment is immediately disrupted when the information that triggers the change is processed when the word is fixated, the disruption appears as a delayed re-reading of the word when the triggering information is encountered only later.

These results provide support for Breen and Clifton's hypothesis that, in their original study, readers were using preview of upcoming disambiguating material to trigger syntactic and metrical reanalysis of the target word while they were still fixating it. As such, in the current study, where the parafoveal preview was inaccurate, readers could not start the process of reanalysis until they fixated the actual disambiguating material.

These results also provide support for the claim that readers are creating an implicit prosodic representation of words when reading silently. They extends results like those reported by Folk and Morris (1995), that heterophonic homographs such as tear disrupt eye movements more than homophonic homographs such as calf when they are used with their less frequent meaning. This earlier evidence indicated that the segmental content of the “little voice” in our heads has a functional role in reading. Our evidence adds to other recent evidence (Kentner, 2012; Hwang & Steinhauer, 2011) that the prosodic composition of implicit speech plays such a role, and that it can even direct on-line parsing decisions.

Appendix

1a/b. The brilliant abstract/report was accepted at the prestigious conference.

1c/d. The brilliant abstract/report the best ideas from the things they read.

2a/b. The secretive project/design was closely guarded by the military.

2c/d. The secretive project/design an image of mystery and privacy.

3a/b. The vicious combat/defeat resulted in many casualties.

3c/d. The vicious combat/defeat their enemies with strength and malice.

4a/b. The unusual permit/practice was not sanctioned by the local government.

4c/d. The unusual permit/practice the behaviors that most people find unacceptable.

5a/b. The strong contrast/partner was hard to ignore.

5c/d. The strong contrast/partner with their weaker friends.

6a/b. The handy object/answer was helpful to the student.

6c/d. The handy object/answer to their less able superiors.

7a/b. The awkward subject/challenge was discussed by the family.

7c/d. The awkward subject/challenge themselves to difficult tasks.

8a/b. The phony address/return was discovered by the authorities.

8c/d. The phony address/return the messages with fake sincerity.

9a/b. The epic details/reports of Greek heroes are well known.

9c/d. The epic details/reports the story of Odysseus’ journey.

10a/b. The local contests/questions occupied the entire street.

10c/d. The local contests/questions the new zoning laws.

11a/b. The subtle conduct/disguise of the secret agent alerted no one to his presence.

11c/d. The subtle conduct/disguise themselves in a manner that does not look suspicious.

12a/b. The hefty decrease/budget was a surprise to the school board.

12c/d. The hefty decrease/budget their calories in an effort to lose weight.

13a/b. The intelligent convict/sentence exhibited surprising coherence.

13c/d. The intelligent convict/sentence the criminals after a fair trial.

14a/b. The witty addict/puzzle was a hit at all of the parties.

14c/d. The witty addict/puzzle their friends with their jokes and stories.

15a/b. The lovely entrance/delight was appreciated by the all of the guests.

15c/d. The lovely entrance/delight the people who see them arrive.

16a/b. The mysterious invite/notice was the topic of much gossip.

16c/d. The mysterious invite/notice the scrutiny of others.

17a/b. The speedy relay/process was finished very quickly.

17c/d. The speedy relay/process the news as soon as it arrives.

18a/b. The foreign import/visit was a hot topic of conversation around town.

18c/d. The foreign import/visit the things they miss from home.

19a/b. The vulgar insult/surprise was a shock to the audience.

19c/d. The vulgar insult/surprise the people with their inappropriate comments.

20a/b. The angry protest/lament was closely watched by the police.

20c/d. The angry protest/lament their treatment by the police and courts.

21a/b. The exotic extract/bottle was sold at the expensive shop.

21c/d. The exotic extract/bottle the flavors of herbs and flowers for cooking.

22a/b. The righteous rebel/struggle was supported by a foreign government.

22c/d. The righteous rebel/struggle against oppressive regimes and dictators.

23a/b. The sneaky pervert/escape was featured on all the local news channels.

23c/d. The sneaky pervert/escape the truth when it stands in their way.

24a/b. The evil suspect/promise was distressing to the public.

24c/d. The evil suspect/promise their victims will be freed by the police.

25a/b. The unfortunate reject/recruit was run out of town.

25c/d. The unfortunate reject/recruit the offers of help from strangers.

26a/b. The sophisticated survey/design was used by many researchers.

26c/d. The sophisticated survey/design their surroundings with a critical eye.

27a/b. The religious convert/reform was applauded by many citizens.

27c/d. The religious convert/reform the beliefs of their friends and neighbors.

28a/b. The weekly inserts/offers were particularly annoying this month.

28c/d. The weekly inserts/offers useless coupons in every issue.

29a/b. The sturdy compress/bandage was used to stop the bleeding.

29c/d. The sturdy compress/bandage the wounds of their fallen comrades.

30a/b. The suspicious record/shadow was discovered to be fake.

30c/d. The suspicious record/shadow their friends and colleagues.

31a/b. The tacky present/display was covered in sequins and feathers.

31c/d. The tacky present/display themselves tastelessly when in public.

32a/b. The modest discount/value was appreciated by the customers at the store.

32c/d. The modest discount/value the importance of their ideas and opinions.