Parafoveal processing influences word frequency and predictability effects on eye movements during

reading

Christopher J. Hand

GlasgowLanguageProcessin

g

Background• The ease or difficulty associated with processing a

particular word influences when the eyes move from one fixation to another

• Studies of eye movement (EM) behaviour have demonstrated that both low-level oculomotor factors and high-level linguistic factors influence EMs during reading– See Rayner (1998) for a review

• Word Length– Major influence on reading time– As word length increases, fixation time increases.

• (Just & Carpenter, 1980; Rayner, Sereno, & Raney, 1996)

Background

• However, after controlling for word length, two variables in particular strongly influence fixation time on a word:– Word Frequency– Contextual Predictability

Word frequency effects on eye movements in reading

• Words can be defined as high or low frequency dependent on their frequencies of occurrence being above or below certain thresholds– Published norms, i.e., Francis and Kučera (1982)

– On-line resources, i.e.,100-million word British National Corpus (BNC) (1995; http://www.natcorp.ox.ac.uk)

Word frequency effects on eye movements in reading

• Many studies have demonstrated that readers look longer at low-frequency (LF) words than at high-frequency (HF) words.– Inhoff & Rayner, 1986; Just & Carpenter, 1980; Raney & Rayner, 1995;

Rayner & Raney, 1996; Rayner & Duffy, 1986; Rayner et al., 1996; Rayner, Ashby, Pollatsek & Reichle, 2004; Rayner, Fischer & Pollatsek, 1998; Sereno & Rayner, 2000.

• “Spillover” effects– Fixation on LF word n inflates fixation time on word n + 1 (Rayner &

Duffy, 1986).– Argued that spillover reflects not an effect of frequency, but is

caused by less parafoveal preview benefit when the parafoveal word is LF (Henderson & Ferreira, 1990)

Contextual predictability effects on eye movements in reading

• Research has established that there is a substantial effect of a word being more predictable from prior context on– Reading times– Skipping Behaviour

• Balota, Pollatsek & Rayner, 1985; Ehrlich & Rayner, 1981.

• Contextually constrained (i.e., predictable) words are fixated for less time and skipped more often than unconstrained (i.e., unpredictable) words

• Ehrlich & Rayner, 1981; Binder, Pollatsek & Rayner, 1999; Rayner et al., 2004; Rayner & Well, 1996.

Contextual predictability effects on eye movements in reading

• Unfortunately…– There is no agreed metric of contextual constraint.– Constraint is typically measured via word-rating

experiments and / or Cloze tasks:

Word Rating

To be a dentist, you must not be afraid of touching people’s teeth

Unpredictable 1 2 3 4 5 Predictable

Cloze Probability

To be a dentist, you must not be afraid of touching people’s ______

Examining the effects of word frequency and predictability simultaneously

• Reaction Time Studies– Stanovich & West (1979, 1983) West & Stanovich (1982)– Typically reported an interactive pattern of frequency

and predictability effects

• Event-Related Potential (ERP) Study– Sereno, Brewer, & O’Donnell (2003)– Evidence to suggest an interaction between frequency

and context in early ERP component (~132-192 ms post-stimulus).

Examining the effects of word frequency and predictability simultaneously

• Although the separate effects of word frequency and contextual predictability on EMs during reading have been demonstrated many times, fewer EM studies have examined these variables simultaneously.

Examining the effects of word frequency and contextual predictability simultaneously

• A frequency × predictability interaction on EM behaviour during reading suggests that these variables affect the same stage of processing.

• Word frequency affects early lexical processing (Sereno & Rayner, 2000)

• Debate as to whether context affects early, lexical processing or later, post-lexical processing.

• Important for models of EM control in reading.

Examining the effects of word frequency and contextual predictability simultaneously• Rayner et al. (2004)

– Simultaneously varied the frequency and predictability of target words to determine whether these to variables yielded additive or interactive effects on

• Fixation durations on target words• Probability of skipping the target word.

• Fixation Duration Data– Significant main effects of frequency and predictability across

various measures– No interaction (all Fs < 1).

• Probability of fixating the target word– Significant frequency × predictability interaction

Limitations of Rayner et al. (2004)

• Aspects of the Rayner et al. (2004) study were of concern:– Number of experimental items per condition– Length of each experimental item– Style of contexts; anomalous targets vs. unpredictable targets.

• Number of experimental items– Rayner et al. (2004) used 16 pairs of HF and LF target words.– The results of this is that each participant only saw 8 target words in

each condition.

• Length of each experimental item– All experimental items in Rayner et al. (2004) were short

(maximum length 72 characters, including spaces and punctuation).

Limitations of Rayner et al. (2004) cont.

• Content of experimental materials– A final concern over the materials used by Rayner et al.

(2004) concerns the anecdotal or “conventional” nature of their experimental materials.

HF-P | LF-U“June Cleaver always serves meat and potatoes | carrots for dinner.”

LF-P | HF-U“Bugs Bunny eats lots of carrots | potatoes to stay healthy.”

Present Study

• Number of Items– 88 experimental items

• Half of items contained HF targets, the other half contained LF target.

• In addition, half of the targets were presented in a predictable context, the other half in an unpredictable context.

– Thus, participants see 22 experimental items per experimental condition vs. 8 per condition in Rayner et al. (2004).

Present Study

• Length of materials– Materials were a maximum of 120 characters, including

spaces and punctuation vs. 72 characters in Rayner et al. (2004).

– Materials were presented over two lines of visual display, max. line length 60 characters

– Target words were always in the middle of the second line of display

– It is argued that using longer materials allows for a stronger development of context with which to constrain target words.

Present Study

• Example materials

HF-P | LF-U

Guests were arriving and Jen’s flat was a sty. She picked up

her clothes from the floor | couch and quickly cleaned the bathroom.

LF-P | HF-U

Clare had been on her feet all day. Armed with a pizza and

a video, she laid down on the couch | floor for a relaxing evening.

Present Study - Materials

HF-P HF-U LF-P LF-U

Mean BNC freq 145 145 4 4

Predictability (1-7) 6.19 4.07 6.11 3.69

Cloze probability 0.57 0.02 0.50 0.01

Mean Word Length 5.47 5.47 5.47 5.47

No. of syllables 1.68 1.68 1.80 1.80

• Target words were never line or sentence final• Target words were never immediately preceded or followed

by a punctuation mark.

Present Study

• Subjects– 64 participants– Normal or corrected-to-normal vision– Native English speakers– No serious reading disorders, i.e., dyslexia.

• Method– Dual-Purkinje eye tracker (Generation 5.5)– Materials displayed over two lines of visual display

(maximum line length 60 character spaces)– 4.14 characters subtended 1º of visual angle.

Results

• 3.6% of total experimental trials were rejected due to track losses or excessive blinking

• A 2 (frequency; high, low) × 2 (context; predictable, unpredictable) ANOVA was performed both by participants (F1) and items (F2)

• A range of standard EM measures were examined– First fixation duration (FFD), single fixation duration

(SFD), gaze duration (GD), total time (TT), probability of fixating the target, and spillover fixation duration.

Forward Fixation Data

• For FFD, SFD and GD, highly significant main effects of word frequency and predictability were found by both participants and items

• However, no evidence of an interaction was found on these measures (all Fs < 1)

SFD• Significant 26 ms main effect of frequency

– F1 (1,63) = 104, p < 0.0001; F2 (1,43) = 148, p < 0.0001.

• Significant 10 ms main effect of predictability– F1 (1,63) = 13.8, p < 0.001; F2 (1,43) = 12.1, p < 0.01.

• No evidence of interaction– Both Fs < 1.

Single Fixation Data

Frequency × Predictability - SFD

240

250

260

270

280

290

300

Predictability

SFD

(m

s) HF target

LF target

HF target 259 269

LF target 285 294

Pred Unpred

Total Time

• Similar to forward fixation duration measures– Highly significant main effects of word frequency and predictability

were found by both participants and items– no evidence of an interaction was found on these measures (all Fs <

1)

TT• Significant 45 ms main effect of frequency

– F1 (1, 63) = 71.0, p < .0001; F2 (1, 43) = 51.7, p <.0001

• Significant 39 ms main effect of predictability

– F1 (1, 63) = 56.0, p < .0001; F2 (1, 43) = 37.1 , p < .0001

• Frequency × Interaction was marginally significant by participants [F1 (1, 63) = 2.86, p = 0.096], but non-significant by items [F2 (1, 43) < 1].

Total Time Data

Frequency × Predictability - TT

260

280

300

320

340

360

380

400

Predictability

TT

(ms) HF target

LF target

HF target 297 328

LF target 334 380

Pred Unpred

Probability of Fixation Data

• Significant main effect of word frequency

– F1 (1, 63) = 5.63, p < 0.05 ; F2 (1, 43) = 4.89, p < 0.05

• Non-significant main effect of predictability– Both Fs < 1

• Significant frequency × predictability interaction

– F1 (1, 63) = 11.3, p < 0.01; F2 (1, 43) = 6.45, p < 0.05

• Nature of the interaction was such that HF-P targets were more likely to be skipped than the other three conditions

Spillover Data

• Non-significant main effect of frequency– Both Fs < 1

• Main effect of predictability was marginally significant by participants [F1 (1,63) = 3.71, p = 0.059], but was non-significant by items [F2 (1,43) = 1.98, p > 0.15]

• Frequency × predictability interaction was significant by subjects [F1 (1, 63) = 7.25, p < 0.01], and was marginally significant by items [F2 (1, 43) = 3.81, p = 0.06]

• Nature of interaction was such that spillover fixation duration was shortest after HF-P targets than the other three conditions

Summary of Results…so far

• Additive effects of word frequency and predictability on target word fixation duration measures

• BUT – Interactive effects on the probability of fixating

the target word– Interactive effects on spillover fixation duration

So…?

• An interaction between frequency and predictability effects may be an elusive effect, that does not manifest itself in the EM record.

• However, research has demonstrated that the ability to extract information from words viewed parafoveally is influenced by the frequency and predictability of that parafoveal word– Inhoff & Rayner (1986)– Balota et al. (1985)

Parafoveal Preview

• Readers obtain more parafoveal preview benefit from HF parafoveal words than LF parafoveal words.– Fixation times on parafoveal word n + 1 were

significantly lower for HF words previously viewed parafoveally (Inhoff & Rayner, 1986)

• Readers obtain more parafoveal preview benefit when words viewed parafoveally are contextually predictable

• Balota et al., 1985.

Parafoveal processing

• It may be the case that parafoveal preview operates in conjunction with the effects of frequency and predictability.

• Parafoveal preview typically manipulated by gaze-contingent display change paradigms

• Parafoveal preview benefit can also be indexed dependent on the distance of the fixation prior to fixating the target word and the beginning of the target word (launch site).

Frequency × Predictability × Launch Site

• Data from original EM analyses was conditionalised post-hoc on the launch site to the target.

• Word identification span typically does not exceed 7-9 characters– (McConkie & Zola, 1987)

• Three ‘groups’ of launch site– 1-3 characters– 4-6 characters– 7-9 characters

Frequency × Predictability × Launch Site

• Present Study - 64 Ss × 88 Itms = 5632 data points– Rayner et al. (2004) - 54 Ss × 36 Itms = 1944 data points– After skips / rejected trials, 4489 data points in initial analyses

Data Points Percentage data from

initial analyses

1-3 chars 916 20.4%

4-6 chars 1359 30.3% = 76.5%

7-9 chars 1156 25.8%

10-12 chars 536 11.9%

13+ chars 523 11.7%

Frequency × Predictability × Launch Site - Results

• A 3 (launch site; 1-3 chars, 4-6 chars) × 2 (frequency; high, low) × 2 (context; predictable, unpredictable) ANOVA was performed both by participants (F1) and items (F2)

• A range of standard EM measures were examined– FFD, SFD, GD, TT, probability of fixating the

target, and spillover fixation duration.

Frequency × Predictability × Launch Site – Forward Fixation Data

• For FFD, SFD, and GD, highly significant main effects of frequency, predictability, and launch site were observed by both participants and items (all ps < 0.0001)

• Non-significant frequency × predictability interactions (all Fs < 1)

• Significant three-way interactions between frequency, predictability, and launch site (all ps < 0.05)

Frequency × Predictability × Launch Site – Single Fixation Duration

• Significant main effect of frequency– F1 (1,63) = 77.6, p < 0.0001; F2 (1,43) = 107, p < 0.0001.

• Significant main effect of predictability– F1 (1,63) = 19.1, p < 0.0001; F2 (1,43) = 13.7, p < 0.001.

• Significant main effect of launch site– F1 (2,63) = 50.0, p < 0.0001; F2 (2,43) = 32.2, p < 0.0001

• Non-significant frequency × predictability interaction– Both Fs < 1

• Significant three-way interaction between frequency, predictability and launch site– F1 (2,63) = 7.19, p < 0.01; F2 (2,43) = 7.49, p < 0.01

Launch site Frequency Predictability Freq × Pred1-3 chars p<0.0001 p<0.0001 p<0.054-6 chars p<0.0001 p=0.10 p<0.01

7-9 chars p<0.01 F<1 F<1

SFD Frequency × Predictability by Launch Site

200

220

240

260

280

300

320

1-3 4-6 7-9

Launch Site (chars)

SF

D (

ms

)

HF-P

HF-U

LF-P

LF-U

Frequency × Predictability × Launch Site – Total Time

• Significant main effect of frequency– F1 (1,63) = 36.4, p < 0.0001; F2 (1,43) = 30.0, p < 0.0001.

• Significant main effect of predictability– F1 (1,63) = 55.3, p < 0.0001; F2 (1,43) = 32.2, p < 0.001.

• Significant main effect of launch site– F1 (2,63) = 25.0, p < 0.0001; F2 (2,43) = 29.1, p < 0.0001

• Frequency × predictability interaction was marginally significant by participants; non-significant by items– F1 (1,63) = 3.06, p = 0.09; F2 < 1

• Non-significant three-way interaction between frequency, predictability and launch site by participants; trend by items– F1 (2,63) = 1.82, p > 0.15; F2 (2,43) = 2.07, p = 0.13

Launch site Frequency Predictability Freq × Pred

1-3 chars p<0.0001 p<0.0001 p<0.05

4-6 chars p<0.0001 p<0.001 F<1

7-9 chars p<0.05 p<0.001 F<1

TT Frequency × Predictability by Launch Site

220

260

300

340

380

420

1-3 4-6 7-9

Launch Site (chars)

SF

D (

ms

)

HF-P

HF-U

LF-P

LF-U

Frequency × Predictability × Launch Site – Spillover

• Non-significant main effect of frequency– Both Fs < 1.

• Non-significant main effect of predictability– Both Fs < 1.

• Main effect of launch site significant by participants; marginally significant by items– F1 (2,63) = 7.87, p < 0.001; F2 (2,43) = 2.56, p = 0.08.

• Frequency × predictability interaction significant by participants; marginally significant by items– F1 (1,63) = 5.17, p < 0.05; F2 (1,43) = 3.61, p = 0.06.

• Non-significant three-way interaction between frequency, predictability, and launch site– F1 < 1; F2 (2,63) = 1.81, p > 0.15

Frequency × Predictability × Launch Site – Probability of Fixation

• Non-significant main effect of frequency– F1 (1,63) = 1.71, p > 0.15; F2 (1,43) = 1.32, p > 0.25.

• Non-significant main effect of predictability– F1 (1,63) = 1.46, p > 0.20; F2 (1,43) = 1.30, p > 0.25.

• Significant main effect of launch site– F1 (2,63) = 18.1, p < 0.0001; F2 (2,43) = 14.8, p < 0.0001.

• Significant frequency × predictability interaction– F1 (1,63) = 20.4, p < 0.0001; F2 (1,43) = 7.63, p < 0.01.

• Significant three-way interaction between frequency, predictability, and launch site by participants; trend by items– F1 (2,63) = 7.76, p < 0.001; F2 (2,43) = 2.04, p = 0.14

Discussion

• Initial analyses found no interaction between the effects of frequency and predictability on target word reading time measures (all Fs < 1)

• Interactive effects found on spillover fixation duration and probability of fixating the target word

Discussion

• Conditionalised analyses on the basis of parafoveal preview – as indexed by launch distance to target word – revealed a significant three-way interaction between the effects of preview, frequency and predictability

• Re-analyses revealed that when parafoveal preview is accounted for, word frequency and contextual predictability interact on– Forward fixation times (i.e., FFD, SFD, GD)– Spillover fixation duration– Probability of fixating target words.

Discussion

• For LF words, predictability effects only occur at a close launch site (1-3 chars)

• For HF words, predictability effects occur at further launch sites (1-3 chars & 4-6 chars)

• The finding of an interactive pattern of effects when parafoveal preview is accounted for provides clear evidence of predictability effects at an early, lexical stage of processing.

The nature of contextual constraint

• Recent research has suggested that readers are able to utilise low-level statistical probability information, i.e., the statistical likelihood of word n occurring given word n – 1 (McDonald & Shillcock, 2003)

• Future research into context effects should perhaps eschew subjective tasks such as Cloze probability / word-rating tasks in favour of a more computational approach to context construction– Latent semantic analyses of large text corpora such as

the 100-million word BNC.

Conclusion

• By employing temporally-precise recording techniques, such as EMs and ERPs (Sereno & Rayner, 2003) in conjunction with carefully controlled stimulus materials, it is possible to delineate a precise time-course of contextual predictability effects.

• Accurately resolving this matter has important consequences for models of EM control during reading and will aide resolve contention between equivocal theories of language processing.

Thanks (shameless theft of Dr. Fraser Smith’s (2006) idea)

Sébastien MielletPaddy O’Donnell Sara Sereno