Coding of the responses

Responses were coded as follows (the target is in brackets when it is not dear from the context what it should be):-

• Correct e.g. silly -* sillier, silly-* silliest

• Bare stem e.g. s illy -* silly (silliei)

• Stem truncation e.g. tidy —> tider, muddy-* muddest • More/most + bare stem e.g. funny —»more funny

• Others e.g. curly-* fluffiest, silly-* silliness (silliest), heavy ->

more heaviest (heaviei)

A variety of identifiable error types come under the ‘others’ category, including semantic substitutions, selection of the wrong suffix and double marking through the use of

more/most with a suffixed form. As the aim of this analysis is to focus on maximal word

errors, I consider just three types of errors that are plausibly caused by maximal word constraints - bare stem, stem truncation, and more/most + bare stem - and lump the other errors in an ‘others’ category. Note that the use of more/most with a bare stem is not strictly speaking an error, given that the comparative or superlative is marked, but it is the inappropriate choice of construction for an adjective of this particular phonological shape.

10.3. Results

The proportion of correct responses for each group is presented in Table 10.3.

Table 10.3. % correct responses

Condition G-SLI LA1 LA2

Comp. 1 syll. Mean (SD) 81.67(33.26) 92.50(10.55) 99.17 (2.89) Comp. 2 syll. Mean (SD) 61.67 (44.28) 77.50 (31.37) 94.17(17.30) Super. 1 syll. Mean (SD) 89.17(19.75) 84.17(23.14) 97.50 (4.52) Super. 2 syll. Mean (SD) 66.67 (38.92) 80.83 (30.59) 100.00 (0.00)

To investigate whether the three groups show the same level of correct performance on the two different suffixes and on stimuli of different syllable number, a 3

(Group: G-SLI, LA1, LA2) x 2 (Suffix: -er, -esf) x 2 (Syllable number: 1, 2) ANOVA was carried out. This revealed significant main effects of group, F (2, 33) = 3.462, p = 0.043 and syllable number, F (2, 33) = 8.593, p = 0.006. The main effect of suffix was not significant, and nor were any of the two-way interactions nor the three-way interaction. Post hoc comparisons (Bonferroni-corrected) were carried out to investigate the differences in performance between the groups. The G-SLI group performed significantly worse than the LA2 group, p = 0.040, but no differently to the LA1 group, p = 0.945. There was no significant difference in performance between the two control groups, p = 0.364. Because -er and -est behave the same way, and they are combined in the analysis that follows.

The finding that there are significantly fewer correct responses for two-syllable stimuli suggests that maximal word effects are present. In order to test this hypothesis, an error analysis was performed. Table 10.4 presents the error scores.

Table 10.4. Error types displaying maximal word effects, expressed as a % of total responses, comparatives and superlatives combined

Error type G-SLI

1 syllable 2 syllable

LA1

1 syllable 2 syllable

LA2

1 syllable 2 syllable

Bare stem Mean 2.08 0.83 0.83 8.33 0.00 0.42

(SD) (5.82) (1.95) (195) (15.28) (0.00) (1.44)

Stem Mean n/a 22.08 n/a 2.50 n/a 0.00

truncation (SD) (33.74) (3.371) (0.0)

More/most* Mean 5.00 3.33 0.00 0.83 0.00 1.25

bare stem (SD) (9.29) (6.15) (0.00) (1.95) (0.00) (4.33)

Because one of the error types, stem truncation, is not possible with one-syllable stimuli, it is not possible to carry out a group x error type x syllable number ANOVA. A 3 (Group: G-SLI, LA1, LA2) x 3 (Error type: bare stem, stem truncation, more/most + bare stem) ANOVA within just the two-syllable stimuli reveals a main effect of group, F (2, 33) = 4.655, p = 0.034, but no significant effect of error type. A significant group x error type interaction, F (2, 33) = 3.822, p = 0.007, indicates that the groups produce different patterns of errors on two-syllable stimuli. I now investigate each of these error types in turn.

For the bare stem errors on two-syllable stimuli, a one way ANOVA by group reveals only a marginally significant effect of group, F (2, 33) = 2.987, p = 0.064, indicating no real group differences in bare stem error production within two-syllable stimuli.

However, the question remains as to whether any of the groups produce bare stem forms as a way of avoiding a three-syllable output. A 3 (Group: G-SU, LA1, LA2) x 2 (Syllable number: 1, 2) ANOVA reveals no significant main effects of group or syllable number, but a significant group x syllable number interaction, F (2, 33) = 3.413, p = 0.045, indicating that the groups produce different numbers of bare stem errors as a function of syllable number. To investigate this interaction further, t-tests within each subject group revealed that only the LA1 controls produced significantly more bare stem errors for two-syllable compared to one-syllable stimuli, t (12) = -1.827, p = 0.050 (1-tailed). This indicates that the LA1 controls, but no other groups, make bare stem errors in response to maximal word effects.

For stem truncation errors, a one way ANOVA reveals a significant group effect, F (2, 33) = 4.579, p = 0.018. Post hoc comparisons (Bonferroni-corrected) reveal that the G- SLI group make marginally more of these errors than the LA1 control group, p = 0.059, and significantly more than the LA2 group, p = 0.028. This indicates that only the G-SLI group respond to maximal word effects by producing stem truncation errors.

For more/most + bare stem errors, a one way ANOVA within the two-syllable stimuli showed no main effect of group. To check whether any of the groups are using this construction to avoid producing a three-syllable output, a 3 (Group: G-SLI, LA1, LA2) x 2 (Syllable number: 1, 2) ANOVA was carried out. This revealed no significant main effects of group or syllable number, and no significant group x syllable number interaction. The error analysis therefore reveals that G-SLI and LA1 children both show maximal word effects, but respond to these pressures in different ways: G-SLI children preferentially truncate the stem whereas LA1 children preferentially omit the suffix.

One obvious question at this point is whether children of any group are more likely to truncate a stem when that first syllable could stand as a semantically-related word on its own. We might predict that *mudder would be produced more often than * heaver because

mud is a word in its own right whereas *heav isn’t. This factor was taken into account

when the stimuli were chosen - five are ‘decomposable’ in that their truncated stem is a semantically-related word in its own right {hairy, funny, curly, muddy, dirty) and five are non-decomposable in that their truncated stem is not a word {happy, silly, tidy, narrow,

hea\ty). Table 10.5 shows the number of stem-truncation errors apportioned to the

Table 10.5. % stem truncation errors according to the decomposability of the stem

Stem type G-SLI LA1 LA2

decomposable Mean (SD) 26.67 (38.46) 4.17 (6.69) 0.00 (0.00)

non-decomposable Mean (SD) 16.67 (30.25) 0.83 (2.89) 0.00 (0.00)

The results indicate that Q-SLI children truncate stems whether or not what remains of the stem is a word in its own right, but a t-test indicates that decomposable stems are more likely to be truncated than non-decomposable ones, t (11) = 2.708, p = 0.020. The LA children do not make enough stem truncations for statistical analysis to be reliable, but they too show a preference for truncating the stem when a real word results from that truncation. While the finding that G-SLI children are more likely to truncate on decomposable stems might suggest a morphological problem, the finding of such large numbers of truncations on non-decomposable stems indicates that stem truncations have a phonological cause.

10.4. Interactions between phonology and derivation: An Optimality