3.3 Lexical Semantic and Usage-based Approaches
3.3.2 Usage-Based Approaches
In this section, I describe usage-based approaches that use indirect negative evidence in or-
der to account for the retreat from overgeneralization (Brooks and Tomasello 1999; Bower-
man and Croft 2008). Unlike models that assume linking rules between the verb semantics
and verb argument structure, these approaches rely on the input data to a higher degree.
Some usage-based models (e.g., Bowerman and Croft 2008) acknowledge that children’s
causative errors are productive. The causative errors are then resolved over time when
a more frequent form of the verb in the input replaces the child’s rules that stem from
overgeneralization. This model relies on one crucial aspect, frequency. For verbs that are
not overridden by a more frequent form, the learner may assume over time that the verb
can only take the structure in which it occurs; i.e., over time, the learner assumes that if
a verb has not occurred as a transitive, the transitive form of the verb is ungrammatical
altogether. I elaborate on the role of frequency in these models below.
Usage-based indirect negative evidence approaches typically follow a statistical pre-
emption (Bates and MacWhinney 1987; Bowerman and Croft 2008; Goldberg 1995; Pinker
1989) or entrenchment (Ambridge et al. 2008; Braine and Brooks 1995) model. I discuss
statistical preemption approaches below, before describing entrenchment models of learn-
ing. Under this approach, there are two types of verbs that are overgeneralized in learning
(e.g.,killanddie), and verbs that do not (e.g.,disappear). For suppletive verbs, the learner makes causative errors because the suppletive causative form of the verb has not yet been
learned. For verbs that do not have a suppletive causative form, the make-causative is said
to preempt the child’s overarching hypothesis; for example,make X disappear preempts
the overgeneralized form ofdisappear X. Statistical preemption falls under an indirect neg-
ative evidence approach because the learner must first assume that the transitive causative
form of a verb is not possible in order to preempt it with another frequent grammatical
form in the input.
Another line of work proposes the entrenchment of verb argument structure (Alishahi
and Stevenson 2008; Ambridge 2013; Ambridge and Lieven 2011; Ambridge, Pine, Row-
land, Jones, and Clark 2009; Ambridge, Pine, Rowland, and Young 2008). For instance,
Ambridge et al. (2008, 2009) argue that the possible argument structures for a verb are
stored each time they occur, and the more times the verb occurs, the higher the probabil-
ity that the verb will only be used in the argument structure frames in which it has been
encountered. Thus, when the child hears a verb like fall occur only as an intransitive,
over time, the probability of the verb occurring as a transitive decreases until the learner
assumes it can only occur as an intransitive. Furthermore, verb semantics also play a
role in addition to entrenchment under this analysis, and the process of entrenchment
is also claimed to include the entrenchment of semantic classes under some proposals
(e.g., Ambridge 2013). The idea behind using entrenchment as a learning mechanism for
the causatives problem is that children are said to learn the semantics of verbs through
the linguistic environments they are found in. In other words, the verbs are entrenched in
their meaning by their semantic environment. This semantic entrenchment constrains the
learner from the beginning, and the learner assumes that the other verb uses in meaning
are impossible. The notion of entrenchment, hence, involves the use of indirect negative
evidence, as the absence of one verb form is taken as evidence for its ungrammaticality.
approach. Ambridge et al. 2008, 2009, 2011 argue that intransitive only verbs that are more
frequent in the input are rated as less acceptable when used as a transitive causative than
less frequent verbs in the same class. For instance, in testing adults, Ambridge et al. 2011
find a difference in the acceptability betweenfallandtumble as causatives. Participants
were found to rate the more frequent verbfall a 1.46 on a 7 point Likert scale, and the
less frequent verbtumblea 2.68 out of 7. I note here that the aforementioned results only
suggest a possible effect of verb frequency, and not a robust distinction in the structural
properties of these verbs. However, in order to confirm that adults at least show effects of
entrenchment, one needs to rule out whether these effects could stem from other semantic
properties of the verb. I address this question in Section 3.3.3.
Furthermore, Ambridge et al. 2008 argue that other situational factors play a role; 5
and 6 year old participants rate the causative form of the novel verb meaninglaughsignif-
icantly lower than its intransitive form (p= 0.001); however, they do not rate the causative
forms of novel verbs of disappearing and falling differently from their intransitive forms
(p= 0.1 andp= 0.9 respectively). 9 and 10 year old children always significantly preferred
the grammatical forms over the ungrammatical forms. Ambridge et al. 2008 argue that
these results indicate that the level of direct external causation determines the level of
acceptability of causatives. In Section 3.4, I show how these facts follow from an account
that generalizes from positive evidence.
Although this approach discusses the mechanisms through which children can retreat
from positing an incorrect causative form to verbs that do not allow them, the analysis
does not discuss the reason for this overgeneralization. Why do children propose a rule
for verbs that do not have a causative form in the first place? A linking rules approach
addresses this problem by claiming that younger children have not fully mastered the spe-
cific properties of each semantic class, but under an entrenchment approach, the reason
for overgeneralization is not entirely clear.
count for the retreat from overgeneralization, we can evaluate whether these approaches
can account for the problem at hand. In the following sections, I provide arguments
against the use of indirect negative evidence to account for this problem by showing that
these approaches do not address the productivity of the causative errors, nor are they able
to adequately account for retreat from overgeneralization. I also illustrate, using child
data, that even if children could learn from indirect negative evidence in theory, their de-
velopmental trajectory suggests otherwise. In Section 3.4, I argue instead that children
overgeneralize when there is sufficient positive evidence to form a productive rule that
can apply to verbs for which they have not heard the causative form.