Rotation Reveals the Importance of Configural Cues in Handwritten Word Perception

Abstract

A dramatic perceptual asymmetry occurs when handwritten words are rotated 90° in either direction. Those rotated in a direction consistent with their natural tilt (typically clockwise) become much more difficult to recognize, relative to those rotated in the opposite direction. In Experiment 1, we compared computer-printed and handwritten words, all equated for degrees of leftward and rightward tilt, and verified the phenomenon: The effect of rotation was far larger for cursive words, especially when rotated in a tilt-consistent direction. In Experiment 2, we replicated this pattern with all items presented in visual noise. In both experiments, word frequency effects were larger for computer-printed words and did not interact with rotation. The results suggest that handwritten word perception requires greater configural processing, relative to computer print, because handwritten letters are variable and ambiguous. When words are rotated, configural processing suffers, particularly when rotation exaggerates natural tilt. Our account is similar to theories of the “Thatcher Illusion,” wherein face inversion disrupts holistic processing. Together, the findings suggest that configural, word-level processing automatically increases when people read handwriting, as letter-level processing becomes less reliable.

In reading research, a long-standing question is whether word perception is affected by familiar, global visual configurations (Cattell, 1886; Johnson, 1975), in addition to letter-level analysis. Certain findings suggest a role for configural processing. For example, “wordshape” features are sometimes sufficient for identification (Beech & Mayall, 2005; Huey, 1908). Words with distinctive outlines are also more efficiently processed parafoveally (McConkie & Rayner, 1975). However, wordshape effects are inconsistent (Paap, Newsome, & Noel, 1984) and direct manipulation by cAsE MiXinG has relatively little impact on perception (Besner & Johnston, 1989). On the whole, to the degree that configural processes affect word perception, their effect seems quite small.

Although the literature suggests a minimal role for configural processes in reading, it is important to note that virtually all studies have involved computer-generated words, with letters separated and pristine. As noted by Barnhart and Goldinger (2010), little research has examined perception of handwritten words, such as ideal. Relative to synthetic print, letters in handwriting are noisy and ambiguous, and their physical forms are affected by neighboring letters. In perception, as a general rule, when bottom-up cues become less reliable, top-down processing tends to increase (Becker & Killion, 1975). Given handwritten words, configural processes may assume a more prominent role. In the present study, we tested this hypothesis, taking our inspiration from the Thatcher effect in face perception (Thompson, 1980), wherein people cannot adequately appreciate facial feature misorientations when faces are inverted. When the same face is shown in its normal orientation, the inverted eyes and mouth are striking. According to many accounts (e.g., Bartlett & Searcy, 1993), when faces are inverted, configural processes are impaired, which reduces appreciation for component features (although see Talati, Rhodes, & Jeffrey, 2010).

Prior researchers have adapted the Thatcher Illusion to examine configural processes in reading. Wong, Twedt, Sheinberg, and Gauthier (2010) asked participants to make “same-different” judgments for sequentially presented upright and inverted word and nonword pairs that were normal or had been “Thatcherized” by rotating one or two internal letters 180°, relative to the rest. They observed a robust Thatcher Illusion: Discrimination between normal and Thatcherized stimuli was reduced when presented upside-down. Interestingly, the effect was larger for words than nonwords, and it interacted with word frequency. Both findings suggest that top-down processes superseded the complex bottom-up analysis necessary for successful discrimination of ambiguous low-frequency/nonword items (see Parks, 1983; Rock, 1988).

In two experiments, we evaluated the importance of configural processing in word perception by rotating synthetic and handwritten words 90° clockwise or counter-clockwise. Lewis (2001) found that RTs to categorize faces as “Thatcherized” or “not Thatcherized” increased gradually as faces were rotated, suggesting that the configural processing used to recognize upright faces gradually shifts to a component processing strategy with rotation. With words, Koriat and Norman (1985) observed similar increases in recognition time as a consequence of rotation (see also Pashler, Ramachandran & Becker, 2006). From a working hypothesis that handwritten word perception relies more heavily on configural information, we expected that rotating handwritten exemplars would have a larger perceptual effect, relative to typewritten words.

In preparing our experiments, we immediately noticed a powerful perceptual asymmetry: When handwritten words were rotated clockwise, they were far more difficult to read, relative to those rotated counter-clockwise (see Figure 1), suggesting that rotation interacts with the natural tilt of the handwriting. Whether right- or left-handed, most people produce handwriting with a rightward tilt, as in the upper row of Figure 1. When these words are rotated clockwise, their natural tilt exaggerates the degree of rotation, whereas it mitigates the effect of counter-clockwise rotation. Thus, when words are rotated in a tilt-consistent direction, their constituent letters wind up more inverted, relative to being rotated in the opposite direction. As in the Thatcher Illusion, greater inversion reduces appreciation for relations among features. In reading, Jolicouer, Snow, and Murray (1987) studied rotated letter recognition, using letters from two different fonts, one more familiar than the other. Rotating letters clockwise, they observed a linear increase in recognition times up to 120° rotation; this effect was larger for the unfamiliar font. Had they also rotated letters counter-clockwise, they may have discovered the perceptual effect at the heart of this paper: Their unfamiliar font was Profil, which has a slight rightward tilt, similar to handwriting.

Figure 1.

Sample stimuli in right- and left-tilting handwriting and print at −90° and +90° orientations. To fully appreciate the perceptual effect, refrain from tilting your head as you read. To experience a reversal of the effect, turn the figure upside-down.

In a pilot experiment, we showed participants printed and cursive words for 500 ms in standard orientation, rotated +90°, or rotated −90°. All items were produced with common rightward tilt (Figure 1, upper row). Presuming that configural processing is mainly important for handwritten words, we expected printed word recognition to vary little with rotation. Conversely, we expected rotation to impair handwritten word perception, especially for words rotated clockwise. Indeed, printed words were identified with 90% accuracy and minimal effects of rotation. Cursive words were identified with 85% accuracy when displayed horizontally. When rotated counter-clockwise, accuracy decreased to 57%. When rotated clockwise, accuracy fell to 31%.

In a second pilot experiment, we added words from a volunteer whose handwriting has an unusual, leftward tilt (Figure 1, second row). Hypothesizing that our initial finding reflected direction of tilt, we now expected more errors following counter-clockwise rotation. The printed and the right-tilting words replicated the previous results. The left-tilting words were relatively hard to read even when horizontal, with 74% accuracy. When rotated, the results flipped, relative to right-tilting words: Accuracy was only 8% for words rotated counter-clockwise, and 37% for words rotated clockwise. Together, these pilot studies verified the phenomenological observation (which is easily appreciated in Figure 1), and motivated the experiments reported next, which included greater stimulus control.

Experiment 1

Our pilot studies suggested that handwritten words elicit greater reliance on configural properties, relative to synthetic words. By our hypothesis, reading handwriting differs from reading computer text, in part by requiring greater reliance on word-level knowledge (Barnhart & Goldinger, 2010). However, the results may only reflect the inversion of component letters, rather than differences in configural processing. Do configural processes have any explanatory value, beyond differences in letter rotation? In Experiment 1, we addressed this question by standardizing the average angles of tilt across all printed and handwritten exemplars. Using image-editing software, we adjusted the tilt for left- and right-tilting handwritten words such that their deviations from absolute vertical were equal (although in opposite directions), and we added identical tilt to typewritten exemplars. The “letter inversion” hypothesis predicts that, under these conditions, equivalent Orientation × Tilt interactions should arise for printed and cursive items. Alternatively, if handwritten words require greater configural processing, the Orientation × Tilt interaction should be larger for cursive words, relative to printed words. Therefore, support for the configural processing hypothesis requires a three-way Orientation × Tilt × Script interaction.

Method

Participants

Thirty-seven Arizona State University undergraduates participated for course credit. All were native English speakers with normal or corrected vision.

Stimuli

We generated 180 five-letter words for the experiment; all were selected from Coney (2005), with equal numbers of low- and high-frequency words (see Appendix). The mean frequency counts per million (Brysbaert & New, 2009) for the low- and high-frequency words were 3.9 and 87.5, respectively, F(1, 179) = 22.9, p < .0001. Each word was produced in both computer-printed and cursive forms as in Figure 1. Printed words appeared in 24-pt Courier New font. Handwritten words were collected from two volunteers using a Logitech io2 digital pen; this looks and feels like a standard ball-point pen, but contains a small camera which reads a fine dot pattern printed on special paper, converting pen strokes into a digital trace. The volunteers were given multiple opportunities to rewrite all items until they appeared legible and natural. The trace files were converted to images and were enlarged (comparable to 24-pt font) and sharpened using GIMP2 software.

Once the words were prepared, ten words were randomly sampled, per script; these were used to estimate the average angles of letter tilt for the left- and right-tilting handwriting. The measure tool in GIMP2 was used to determine the angular deviation from vertical of each sampled exemplar. When available, angles were calculated using the longest ascender or descender within the word. The average angle for left-tilting stimuli was 43° from vertical and the average angle for right-tilting stimuli was 40°. We chose the mean, 41.5°, as our standardized tilt angle for all stimuli in Experiment 1. For the printed stimuli, the “shearing” feature in Adobe Illustrator® CS5 was used to adjust letter tilt 41.5° from vertical, in both leftward and rightward directions. The images were then sharpened using the unsharp mask filter in GIMP2. Finally, 12 lists were generated, counterbalancing script, tilt direction, and orientation.

Apparatus

Stimuli were presented using E-prime 1.2 software on a CRT monitor with screen resolution 1024×768. Responses were collected via keyboard.

Procedure

Participants were instructed that they would briefly see words in printed and cursive forms, in different orientations. They were to identify each word, typing it as quickly and accurately as possible. They were asked to refrain from tilting their heads, and were monitored by the experimenter. Each trial began with a 750-ms fixation cross, shown at the spatial beginning of the upcoming word. Thus, the cross was located above center for words rotated clockwise, below center for words rotated counter-clockwise, and left of center for words shown horizontally. The cross was replaced by the stimulus word for 500 ms, and participants typed their responses, encouraged to guess if necessary. Each trial was followed by accuracy feedback. Following nine practice trials, the experiment included 180 randomized trials.

Results

One participant was excluded from analysis due to excessive errors. We conducted a repeated-measures ANOVA on error rates (see Figure 2) with Script (print, cursive), Tilt (left, right), and Orientation (0°, −90°, +90°) as within-subject variables. To ensure normality of the data, error rates were arcsine-square-root transformed prior to analysis (Cohen, Cohen, West & Aiken, 2003).¹ We observed a large main effect of Script, F_S(1, 35) = 1494.5, p < .001, η²_p = .97, with more errors for cursive words.² Rotation in either direction produced substantial errors, relative to upright words, creating an Orientation effect, F_S(2, 34) = 159.1, p < .001, η²_p = .90.³ A Tilt effect, F_S(1, 35) = 49.1, p < .001, η²_p = .58, reflected greater errors for words with the atypical, leftward tilt.

Figure 2.

Experiment 1. Error rates (± SEM) as a function of script, tilt, and frequency for words at different orientations. LF and HF denote low- and high-frequency words; LT and RT denote left- and right-tilting text, respectively.

All main effects were qualified by interactions. A Script × Tilt interaction, F_S(1, 35) = 51.7, p < .001, η²_p = .60, reflected a larger tilt effect for cursive words (15.8%) than for printed words (7.9%), F_S(1, 35) = 18.3, p < .001, η²_p = .26. An Orientation × Script interaction, F_S(2, 34) = 101.1, p < .001, η²_p = .86 was observed: The average rotation effect was 6.5% for printed words, but was 42.1% for cursive words, F_S(1, 35) = 132.9, p < .001, η²_p = .79. An Orientation × Tilt interaction, F_S(2, 34) = 93.4, p < .001, η²_p = .85, was observed: Error rates were 17.5% higher following tilt-consistent rotation, relative to tilt-inconsistent rotation, F_S(1, 35) = 25.0, p < .001, η²_p = .44. Finally, the critical three-way interaction was reliable, F_S(2, 34) = 19.4, p < .001, η²_p = .54. The Orientation × Tilt interaction was larger for cursive words, F_S(2, 34) = 84.9, p < .001, η²_p = .83, than for printed words, F_S(2, 34) = 4.48, p < .05, η²_p = .09.

Because the experimental materials were divided into low- and high-frequency words, we assessed whether the foregoing results differed according to frequency. There was a main effect of Frequency, F_I(1, 178) = 10.9, p < .01, η²_p = .06, with 5.1% more errors to low-frequency words, relative to high-frequency words. Among the five possible interactions, only Script × Frequency was reliable, F_I(1, 178) = 36.4, p < .001, η²_p = .34; the frequency effect was 18.8% larger when words were written in print, relative to cursive. We consider this further in the General Discussion.

Discussion

Experiment 1 verified that rotation disproportionately affects handwritten word perception. Printed words elicited few errors across orientations, despite a small effect of orientation. In contrast, cursive words in standard orientation were recognized quite well (80.2% correct), but rotation in either direction reduced perception, with tilt-consistent rotation producing especially dramatic effects. As noted, we hypothesize that this asymmetry reflects the handwriting tilt. It has been suggested that, during mental rotation (Shepard & Cooper, 1982), people “internally sample” all intermediate positions between an abstract, standard position and the actual position. From this perspective, when a word is rotated clockwise, it becomes a rectangular perceptual object with a “global orientation” of +90°. Each letter, however, is tilted farther, requiring mental rotation beyond the salient global orientation. If the observer mentally “corrects” the word by rotating it −90°, she will not reach the actual orientations of the letters. Conversely, when the same word is mentally rotated clockwise from a −90° orientation, the “correct” orientation of the letters will be achieved prior to reaching 0°, allowing an opportunity to appreciate the letters and their transitions.

Moving beyond our pilot studies, Experiment 1 involved equivalent tilt across all items, including the typewritten words. We did observe rotation effects with typewritten words, although the effect was substantially larger for handwritten words. This finding suggests that, if rotation disrupts configural processes (Lewis, 2001), then such processes are especially important for handwritten words. Handwritten words may be less amenable to a component-processing strategy (which rotation necessitates) by virtue of their connectedness and context-conditioned variation (Barnhart & Goldinger, 2010). Thus perception of rotated cursive often fails, especially in cases of tilt-consistent rotation.

Experiment 2

Experiment 1 produced the anticipated three-way interaction: Tilt-consistent rotation had stronger effects for handwritten word perception, relative to typewritten word perception. One potential concern, however, is that performance was too accurate to the printed words, such that potential rotation effects were obscured by ceiling effects. Experiment 2 was conducted to replicate Experiment 1, but with visual noise added (to all items). Although there was some risk that degrading the handwritten words could elicit floor effects, such an outcome would “work against” the hypothesized pattern, as it would allow the printed words more “room” to produce a Rotation × Tilt interaction, and would reduce that interaction for the handwritten words.

Method

Participants

Forty-seven undergraduates participated for course credit. They were native English speakers with normal or corrected vision.

Stimuli and Procedure

The stimuli and procedure were identical to those of Experiment 1, with one exception. For all items, we used Adobe Photoshop to change 80% of pixels (in a uniform rectangle around each word) to randomly-selected levels of gray-scale, giving the appearance of words in visual static.

Results

Because the principal goal of Experiment 2 was to reduce performance, particularly for the printed words, we first compared error rates (see Figure 3) across Experiments 1 and 2. In a repeated-measures ANOVA, there was a main effect of Experiment, F_S(1, 81) = 21.6, p < .001, η²_p = .22, with more errors in Experiment 2. We conducted another ANOVA on the printed words alone, finding a similar main effect, F_S(1, 35) = 11.5, p < .001, η²_p = .12. Mean error rates to printed words in Experiments 1 and 2 were 7.9% and 13.4% respectively. Most important, when printed words were rotated, performance moved away from ceiling, with accuracy ranging from 74-90%.

Figure 3.

Experiment 2. Error rates (± SEM) as a function of script and tilt for degraded words at different orientations. LF and HF denote low- and high-frequency words; LT and RT denote left- and right-tilting text, respectively.

In other regards, the results of Experiment 2 were similar to Experiment 1, as verified by a repeated-measures ANOVA on arcsine-transformed error rates with Script (print, cursive), Tilt (left, right), and Orientation (0°, −90°, +90°) as within-subject variables. We observed a large main effect of Script, F_S(1, 46) = 1717.9, p < .001, η²_p = .98, with more errors for cursive words. A Tilt effect, F_S(1, 46) = 142.4, p < .001, η²_p = .76, again reflected higher errors for words with the atypical, leftward tilt. Rotation in either direction produced high error rates, relative to horizontal words, producing an Orientation effect, F_S(2, 45) = 153.3, p < .001, η²_p = .87.

As in Experiment 1, all main effects were qualified by interactions, in the same directions. Script interacted with Tilt, F_S(1, 46) = 83.7, p < .001, η²_p = .65, with a larger tilt effect for cursive words (18%) than printed words (1%), F_S(1, 46) = 49.2, p < .001, η²_p = .68. An Orientation × Script interaction, F_S(2, 45) = 40.3, p < .001, η²_p = .64, reflected an average rotation effect of 8.3% for printed words, but 27.6% for cursive words, F_S(1, 46) = 39.5, p < .001, η²_p = .51. The Orientation × Tilt interaction, F_S(2, 45) = 55.4, p < .001, η²_p = .71, reflected error rates being 15.2% higher following tilt-consistent rotation, relative to tilt-inconsistent rotation. Finally, the critical three-way interaction was again reliable, F_S(2, 45) = 8.3, p < .01, η²_p = .22, as the Orientation × Tilt interaction was larger for cursive words, F_S(2, 45) = 64.5, p < .001, η²_p = .74, than printed words, F_S(2, 45) = 6.1, p < .01, η²_p = .13.

We again assessed whether the results varied according to word frequency. There was a main effect of Frequency, F_I(1, 178) = 16.7, p < .001, η²_p = .09, with 7.7% more errors to low-frequency words. As in Experiment 1, among the possible interactions, only Script × Frequency was reliable, F_I(1, 178) = 40.9, p < .001, η²_p = .38; the frequency effect was 23.1% larger when words were written in print, relative to cursive.

Discussion

Experiment 2 replicated Experiment 1, with less concern about ceiling effects for printed words: Rotation again interacted with tilt, with maximal errors when directions of word rotation were consistent with inherent tilt. Although the same general pattern was observed for synthetic and handwritten text, the effect was substantially larger for handwriting. As we consider in the General Discussion, the difference between reading these forms is not only a matter of scale; there is a qualitative phenomenological difference between reading rotated print and rotated handwriting.

General Discussion

In contrast to previous studies, the present results suggest that configural processes can affect word perception, but their role is typically minimized by the use of synthetic print. In handwritten words, there are many “coarticulatory” cues that degrade letter-level information, naturally shifting perceptual priority to the overall, word-level Gestalt. When configural processes were impaired by rotation, the effects were dramatic. In Experiment 1, rotating synthetic words had small (but reliable) effects on printed word recognition, with tilt-consistent rotation increasing errors approximately 5% more than tilt-inconsistent rotation. With handwritten words, tilt-consistent rotation increased perceptual errors by approximately 30%, relative to tilt-inconsistent rotation. Experiment 2 replicated this pattern, while added visual noise alleviated concerns regarding ceiling effects for printed words.

In both experiments, the perceptual effects of rotation did not interact with word frequency, although frequency did affect overall accuracy. We suggest that the present effects are largely pre-lexical (although further evidence would be required to make a strong claim). This suggestion is partly motivated from the phenomenological experience of attempting to read rotated, handwritten words. Stated simply, while performing this task, there are numerous trials wherein the stimulus appears and disappears, and the observer experiences no perceptual coherence at all. (Perhaps this experience was simulated, in reduced form, while viewing the sample items in Figure 1.) In such cases, it is difficult even to guess a couple of letters; making a sophisticated, frequency-biased guess is simply impossible. By contrast, with the computer-printed words, this qualitative experience almost never happens: People nearly always perceive at least a subset of the letters, and can issue a guess. Indeed, in both experiments, frequency effects were quite large for printed words, and were absent for handwritten words, which underscores the qualitative difference between conditions.

As noted earlier, we suggest that the “rotated script” effect is conceptually similar to the Thatcher Illusion, using words instead of faces: When handwritten words are rotated, their configural integrity is compromised, reducing appreciation for individual letters and their transitions. Although we doubt that similar neural specializations connect these effects (Carbon, Grüter, Weber, & Lueschow, 2007), they appear psychologically similar, as part-whole relations are disrupted by disorientation. Indeed, we recently created a modified version of the Thatcher illusion, modeled after the current experiments with words (Barnhart & Goldinger, manuscript in preparation). Specifically, we altered facial photographs, such that each person’s eyes and mouth were rotated 27°, either clockwise or counter-clockwise. When viewed upright, both versions appear equally grotesque. Upon rotating the images, a perceptual effect emerges, analogous to the rotated script effect: When the faces are turned in the same direction as their tilted features, they appear more normal.

Certainly, the perceptual experiences differ widely between our modified Thatcher Illusion and the rotated script effect. One is expressed as a failure to feel revulsion, the other as a failure of lexical access. Nevertheless, they have underlying similarity: When faces are rotated in the same direction as their internal features, people fail to fully appreciate that “something is wrong.” When handwritten words are rotated in the same direction as their internal features, people often fail to appreciate anything at all. In both phenomena, the importance of configural processing to perception is emphasized by its selective disruption. The present results suggest that word perception can be affected by holistic, word-level processes, but those effects are strongest when letter-level information becomes noisy. Handwriting provides a natural vehicle for investigating such perceptual compensation processes.

Appendix. Stimulus words, with frequency per million from Brysbaert and New (2009)

Low-Frequency Words
agony (3.75)	flask (1.12)	polio (0.88)
algae (0.43)	fleet (10.59)	pouch (1.71)
alias (4.08)	flood (5.71)	prior (8.27)
ankle (8.02)	frost (4.80)	quart (2.08)
basin (1.88)	gable (1.08)	rabbi (6.71)
berry (3.49)	globe (5.22)	ridge (7.08)
bosom (3.12)	graft (1.65)	ruler (3.18)
brass (12.12)	graph (0.75)	salve (0.57)
broom (4.76)	gravy (5.27)	satin (2.61)
camel (5.02)	grove (3.86)	sheer (4.57)
canon (0.98)	harem (0.88)	siren (6.55)
cargo (9.00)	haste (2.08)	slope (2.94)
cedar (2.49)	havoc (1.35)	snail (1.76)
chaos (9.39)	hedge (1.55)	spark (6.27)
chart (9.47)	ideal (7.33)	spire (0.10)
chasm (0.35)	idiom (0.24)	stain (6.20)
cleat (0.43)	jelly (7.12)	stair (1.35)
clove (0.51)	lemon (12.02)	stout (1.04)
comet (3.08)	linen (2.92)	stump (2.45)
coral (2.37)	links (2.76)	swamp (8.98)
crumb (1.80)	maker (4.92)	taper (0.18)
curve (4.61)	miner (1.45)	tense (10.24)
dairy (2.78)	navel (1.24)	theft (6.75)
decay (2.06)	noose (2.18)	thorn (5.10)
deity (0.55)	opium (2.24)	troop (5.80)
donor (4.08)	organ (7.25)	trout (4.02)
eagle (11.49)	pause (5.39)	tunic (0.55)
error (9.27)	peach (6.35)	vista (2.78)
essay (6.14)	plank (2.04)	wharf (1.27)
ether (2.12)	plumb (1.69)	whiff (2.49)
High-Frequency Words
adult (14.29)	grass (16.78)	round (66.53)
badge (15.25)	green (72.47)	salad (17.02)
beach (56.63)	group (73.76)	scout (12.88)
beast (24.55)	guilt (14.90)	shirt (46.37)
birth (27.55)	horse (92.88)	shore (19.86)
blade (13.00)	hotel (103.22)	short (85.63)
board (64.16)	issue (34.14)	slave (18.43)
booth (20.37)	knife (46.80)	small (124.96)
brain (77.02)	light (165.20)	smoke (65.43)
bunch (58.88)	lobby (12.69)	sound (143.39)
candy (35.78)	loose (41.78)	space (66.06)
chair (49.24)	lunch (104.12)	spike (12.25)
cliff (21.57)	major (104.76)	split (38.31)
coast (26.69)	mercy (25.31)	stake (16.61)
couch (23.47)	metal (19.45)	stick (97.12)
court (100.73)	minor (12.82)	stone (40.63)
crown (13.69)	money (640.76)	store (81.92)
devil (41.33)	moral (13.51)	storm (30.86)
dough (15.88)	motor (13.16)	story (220.78)
dozen (24.14)	night (866.04)	sweet (145.2)
drill (13.75)	nurse (44.98)	teeth (47.84)
earth (99.49)	party (233.14)	thick (13.98)
equal (13.37)	peace (69.61)	thief (24.27)
extra (59.16)	phase (12.33)	thing (1088.67)
fairy (16.69)	phone (269.73)	track (55.75)
fence (16.06)	plant (27.61)	train (95.06)
flash (15.35)	plate (25.65)	trash (22.47)
floor (100.63)	proof (34.39)	trial (49.37)
giant (27.06)	purse (19.76)	water (225.06)
glass (60.71)	ranch (17.55)	wrong (523.10)