Modelling in the Language Sciences
Bart de Boer, Willem Zuidema

Abstract:
Computers can be used for many different purposes in linguistic
research. They can be used for data storage and search. They can be
used as devices for speech analysis or synthesis. They can be used to
present linguistic stimuli to subjects and record their responses. In
all these applications, computers are used as sophisticated tools, and
they are programmed according to purely practical criteria: as long as
they get the job done, the researchers who use the applications do not
care about the internal workings of the software.

However, computing can also become the focus of linguistic
research. Computers can be used to operationalize linguistic theories
by implementing them as computer programs. This is done because
linguistic theories may be so complex that their predictions can no
longer be derived using verbal reasoning or pen-and-paper
analysis. Moreover, turning a linguistic theory into a computer
program forces the researcher to make her assumptions explicit. By
running the program, and studying its behavior under a variety of
circumstances, the researcher can test the theory against empirical
findings and often discover unexpected consequences.

In this chapter, we discuss the use of computational models in the
language sciences. Although formalization has had a central place
since the 1950s in syntax and phonetics in particular, the last two
decades have seen an explosion of interest in mathematical and
computational models in all linguistic subfields: from typology to
language acquisition, from discourse to phonology, linguists are
increasingly viewing formal modelling as an approach that ensures the
internal consistency of theories. However, although many proponents of
modelling believe it makes their field more scientific and objective,
it seems fair to say that the introduction of formal models has so-far
not led to a broad consensus among language researchers. On the
contrary, models have often been at the heart of longstanding
controversies (e.g., those about formalisms vs.  functionalism,
nativism vs. empiricism, single- vs. dual-mechanism).

One reason, we believe, that modelling has played more of a divisive
than a unifying role is that there has been little attention to
questions about modelling methodology: what kind of lessons can we
expect to learn from a model? What makes a good or a bad model? How
may different models of the same linguistic phenomenon relate to
eachother? How could models of different phenomena fit together? 
Thinking about such questions leads one to systematically consider the
role of specific models in a given subfield: Are they consistent with
and complementary to each other? Are the assumptions that go into a
particular model, if not (yet) supported by empirical findings, made
plausible by results from other models?

The situation is not uniform across all linguistic subfields, of
course, but we observe that in fields where 1 or 2 of these questions
have received a lot of attention, the others tend to be ignored even
more. For instance, in syntactic theory there has been an enormous
amount of work (of impressive mathematical sophistication) on
comparing different syntactic frameworks and their ability to model
native speaker intuitions about the grammaticality of carefully
selected (but often highly contrived) sentences. However, in our view,
this field has paid much too little attention to questions about
whether that is really the most important criterion for evaluating
models of language and about relations with cognitive and neural
models. As we will emphasize in this chapter, the ability to reproduce
a selected set of empirical phenomena is certainly not the only
criterion for a good model.

Because it is impossible to cover all linguistic subfields, we will
make our general points about methodology concrete using examples from
two particular domains: the evolution of speech and the learnability
of syntax. In both fields computational modeling has played an
important role, but in both we also believe progress has been hampered
by lack of attention to modeling methodology and the questions one
immediately asks about the relation between existing models when
taking the view on modeling that we develop in this chapter.

For sustaining the success of modelling approaches in linguistic
research, it is crucial that models start living up to their promise:
modellers must make explicit how their models fit in with other
modelling and empirical work, and how their modelling results affect
judgments of plausibility of existing hypotheses that exist in the
field to which they wish to make a contribution. Moreover, they must
do so based on careful consideration of other work, without
overstating their results and misusing the prestige that comes with
mathematical and computational approaches.

In section 2 we will start with some considerations about the
methodology of modelling in linguistics, and introduce the concepts of
model sequencing and model parallelization. In sections 3 and 4 we
will illustrate these concepts with two case studies on modeling in
the evolution of speech and the learnability of syntax
respectively. In section 5 we will then draw some general lessons from
these case studies, and sketch an agenda for future research in
computational modelling of language.