The objective of this experiment is to determine what pluralization rules are acquired by our learner given a sample of common nouns and their plurals. The formation of English plurals is unusually regular. There are very few irregular plural nouns. This property of English might lead one to propose learning mechanisms that exploit the statistics of regular plurals by training on a large number of examples so that any new test noun is sufficiently similar to a known one to produce the closest matched plural ending.
But there is evidence that the statistical property may not be essential to the acquisition of regular rules. For example, Marcus et. al. [9] and Clahsen [4] showed that the German -s plural behaves like a regular rule despite the fact that the rule applies to fewer than 30 common nouns. This observation raises the question of how a child can acquire regular rules from very few examples. The experiment will show that our learner can acquire generalizations that closely resemble those described in linguistics books after seeing on the order of 10 examples.
The input of this experiment consists of 22 noun-plural pairs. The particular number and choices of words are not very important as long as there are some examples of singular nouns ending in different phonemes. We pick a few examples for each of five types of plural formation:
[s] [z] [I.z] semi-regular irregular
------------------------------------------------------------
cake(s) bottle(s) box(es) house(s) man/men
cat(s) boy(s) bush(es) leaf/leaves foot/feet
chief(s) dog(s) church(es)
cup(s) girl(s) dish(es)
fruit(s) gun(s) glass(es)
month(s) horse(s)
nose(s)
The first three columns are the regular plural forms. The semi-regular column contains words whose singular forms end in voiceless fricatives ([s] and [f]), which become voiced in their plural forms ([z] and [v] respectively). The irregular nouns involve internal vowel changes instead of adding affixes.
The 22 pairs are fed to the learner sequentially in a random order
once. The results presented here are typical because the final rules
acquired are observed to be not sensitive to the order of
presentation.
After the presentation of all 22 pairs, the learner has acquired five rule-classifiers and four exceptions. The phoneme bits of the classifiers are as follows:
1. [dc.dc.[+voice,-strident].z]
2. [dc.dc.{y,e,I,v}.z]
3. [dc.dc.[-voice,-strident].s]
4. [dc.dc.[-voice,-coronal].s]
5. [dc.[+coronal,+strident].I.z]
Rule-classifier 3 is acquired after the presentation of 7 pairs. Rule-classifier 1 is acquired after 9 pairs. Although the list of 22 pairs is sufficient for the acquisition of the English plural rules, the list is far from minimal. For example, rule-classifier 5 is acquired after the learner encounters only 4 examples of nouns ending in [I.z]. The remaining three [I.z] examples are redundant. The irregulars also do not affect the acquisition of the regular rules. They are represented as specific rote-classifiers. The correlations among the irregulars are grouped into three exceptional classes: (1) foot/feet and man/men, (2) leaf/leaves, and (3) house/houses.
Notice that we can almost read off the standard English pluralization rules from these classifiers. There are, however, two differences. First, the standard English pluralization rules are typically ordered (see, for example, [1]):
a. If the noun ends in a phoneme containing the features [+strident,
+coronal] (i.e., one of the sounds [s], [z], [sh], [zh], [ch], [j]),
the plural affix is [I z].
Otherwise,
b. If the noun ends in a [+voice] phoneme, the affix is [z].
c. If the noun ends in a [-voice] phoneme, the affix is [s].
In our system, the classifiers are activated in parallel, with the most excited ones gaining control over the data registers.
The second difference is that the unvoiced-plural rule c is
represented by a disjunction of two rule-classifiers, 3 and 4, in our
system. Rule-classifier 3 covers nouns ending in consonants [t], [k],
[p], or [th]. Rule-classifier 4 covers nouns ending in the
strident [f] or the non-coronal stops [k] and [p]. Similarly, the
voiced-plural rule b is split into rule-classifiers 1 and 2.
The learner also exhibits intermediate behaviors similar to those of young children [2]. After rule-classifier 1 and rule-classifier 3 are acquired, the performance program produces plurals like *foot[s] and *man[z]. Upon presentation of the nonce word ``wug,'' it gives wug[z]. For nonce words ending in a strident like ``tass'' or ``gutch,'' it gives the unaltered singular forms as plurals.
There is however a remaining mystery regarding the ``add-[I.z]'' rule. Berko in her study of children's learning of English morphology made the following observation. While the first-graders can apply the ``add-[z]'' and ``add-[s]'' pluralization rules productively to new words, they fail to apply the ``add-[I.z]'' rule to nonce words like ``tass'' or ``gutch.'' When asked to produce the plural of a nonce word ending in [s] or [ch], they either repeat the word in its singular form or fail to respond. In no cases do they give wrong answers like tass[z], tass[s], or gutch[z], and only in few cases do they respond with gutch[s]. The children fail to use the [I.z] rule productively despite the fact that they can recognize and use real words like ``glass'' and ``glasses'' correctly.
Although the intermediate result of experiment 1 is consistent with
Berko's interpretation of the developmental data, the result depends
on the higher density of English plurals ending in non-stridents.
Contrary to Berko's interpretation, our theory predicts that the
learner would have no difficulty in acquiring the add-[I.z] rule
before the add-[s] or add-[z] rules if it were given the plurals
ending in stridents first.