Bio-inspired Grammatical Inference - Foundations on Natural and Artificial Computation - page 318

to the child. We have also pointed out the relevance of semantics in the first
stages of children's language acquisition. And finally, we have seen that the
Chomsky hierarchy has some limitations, specially when we try to locate natural
languages in this hierarchy; since regular and context-free grammars seems not
to be very adequate to model natural language syntax, linguists have tried to find
other grammatical formalisms that have interesting linguistic and computational
properties, such as Mildly Context-Sensitive.
Taking into account all these ideas, we propose to use linguistic studies to im-
prove models and techniques used in GI. Thanks to ideas coming from linguistic
studies on natural language acquisition, models in GI could be more realistic;
these models could take into account more aspects about children's language
acquisition. Moreover, such ideas could also improve the results obtained in the
field of GI. In that way, we would use a bio-inspired model for language learning.
First of all, we propose that GI algorithms take into account not only positive
data , but also corrections during the learning process. We consider that the
most convincing proposal to the question of what kind of data is available to
children, is the one proposed by Chouinard and Clark [12]. To consider that only
explicit disapproval counts as negative data is not realistic. As Chouinard and
Clark showed, adults correct children in a very different way, taking into account
the meaning that the child intends to express. Moreover, a very large number
of examples of such kind of meaning-preserving corrections can be found in real
conversations between children and adults (for example, in CHILDES database).
The second proposal is also unconvincing, since corrective and non-corrective
replies are mixed in their analysis, and hence, learning from “reply-types” would
require that children do complex statistical comparisons in order to learn which
sentences are correct. Therefore, as Chouinard and Clark proposed in [12], we
consider that meaning-preserving corrections are available to children, and they
can help them to learn some aspects of natural language syntax.
In order to see the effect of corrections in language leaning, we propose to
incorporate the idea of corrections to the studies of GI. We have already done
some work in this direction. Our first approach has consisted on considering
only syntactic corrections based on proximity between strings. Since this idea
was totally new in GI, we started by learning deterministic finite automata [8],
in the framework of query learning (i.e., the learner is able to ask queries to
the teacher, and the teacher has to answer correctly to these questions). Later,
these results were extended to learn other classes with interesting properties,
such as bal ls of strings (which are defined by using the edit distance) [7]. In
both cases, when the learner asks for a string that does not belong to the target
language, the teacher returns a correction (in the first case, such correction is
based on the shortest extension of the queried string, and in the second case,
such correction is based on the edit distance). In both cases, we could show that
results can be improved thanks to corrections. Therefore, such works show that
new challenging results can be obtained in the field of GI if corrections are taken
into account in the learning process.