Íslenskt mál og almenn málfræði - 01.01.2023, Side 195

matically annotated by hand, building on previously proposed generalizations on
the associations between case, structural role and meaning. I refer to the resulting
frequency list as the IceCASE (Icelandic Child Argument Structure Evidence)
corpus. I then implemented a maximal approach to hypothesis formation, both
in terms of the size of the set a rule applies to, rule specificity and the conceptual
categories assumed (e.g. Rissman and Majid 2019). When a productive rule is not
discovered, the set is further divided (Belth et al. 2021). Rule productivity was
incrementally evaluated using the Tolerance Principle (Yang 2016), a parameter-
free learning model which provides a threshold for the number of exceptions a
productive rule can tolerate:
(5) Let a rule R be defined over a set of N items. R is productive if and only if
e, the number of items not supporting R, does not exceed θn:
e ≤ θn =  n
 ln n
The Tolerance Principle draws on the idea that productive rules must be applic-
able to a large enough number of candidates, without having too many excep-
tions. Additionally, the nature of the computation yields a higher proportion of
tolerable exceptions in smaller sets. The corpus study therefore provides insights
into the linking problem and the distributional information children could use to
discover associations between argument form and role (e.g. Pearl and Sprouse
2021). It also revives the broad generalizations between case, structure and mean-
ing which have been discussed by linguists (e.g. agents are always nominative,
patients tend to be accusative and goals are usually dative) but have remained theo -
retically elusive, and their existence even contested (Sigurðsson 2012), on the
basis of correlations not being exact (Maling 2002, Thráinsson 2007, Wood 2015,
McFadden 2020).
In addition to testing the potential of the Tolerance Principle for the deriva-
tion of Icelandic case marking rules, a major goal of the chapter is to provide an
overview of the case frames present in Icelandic child language. This overview
reveals various important facts, notably the early prominence of dative goals as
opposed to motion themes (Figure 3), in part due to optional reflexives (Figure 4).
The results furthermore indicate that many of the broad patterns between
case, meaning and syntactic role emerge in the experimental data of Chapters 2
and 3 in the dissertation as well as in previous literature, can in fact be derived
from child language data assuming the Tolerance Principle. This refines the role
of statistical majority in productivity and redefines the value of exceptions in case
theory. In certain contexts, the lack of productivity is also informative for the
more unusual results, such as the broad scope of dative productivity with direct
objects. In general, I show how non-default productivity can be derived in the
form of nested subrules (dative objects) and structured exceptions (dative sub-
jects), and challenge the view that case, both structural and inherent, does not
have meaning. The use of the Tolerance Principle in the analysis of the child lan-
guage corpus furthermore sheds light on various more specific puzzles relating
Project rationale and core ideas 195