6 Make a good prediction or get
ready for a locality penalty
Maybe it’s coming late
Hajime Ono, Kentaro Nakatani
and Noriaki Yusa
1 Introduction
We all make predictions. We are not fortune-tellers, but we can still hope to
make good predictions. It could be that good predictors are also effective commu-
nicators. In a naïve conception of predictions, one may suggest that having more
information in our hand leads to a good prediction. Now, let us consider
those ideas about predictions in the realm of sentence processing. Suppose
you have the following sentence fragments in Japanese.
(1) a. Taro-ga
Taro-NOM
b. Taro-ga tsukue-o
Taro-NOM desk-ACC
c. Taro-ga tsukue-o zookin-de
Taro-NOM desk-ACC dustcloth-with
As soon as a reader encounters a nominative NP at the beginning of the
sentence, such as (1a), we can predict that there must be a predicate (a verb
or a predicative adjective) coming in, but we do not know yet what type of
predicate (argument structure and/or lexical information). Then, after one or
two more constituents are added to the context, as in (1b) and (1c), we can
narrow down the number of possible continuations. When a nominative NP is
followed by an accusative NP, as in (1b), it becomes clear that an intransitive
verb will not appear in the structure, assuming that the reader predicts that
those two constituents have a dependency on a single verb. Still, there are a
number of verbs that could follow, and it may not be easy to predict the specifi c
lexical content of the verb at this point. Note that, for (1c), after the reader
sees another constituent, we can pick out a relatively specifi c predicate such as
huku ‘wipe’ to be the best continuation with confi dence (see an explanation of
the ‘surprisal’ model in Hale (2001) and Levy (2008)). Then, exactly what kind
of information will lead to a good prediction (in fact, it is called an ‘expecta-
tion’ in the literature) is at issue in the literature; obviously, a lot of different
kinds of information could contribute to the prediction-making (Frazier and