Classification Tree Generation Constrained with Variable Weights - Foundations on Natural and Artificial Computation - page 280

non-common feature of languages. Only a very small set of non-indoeuropean
languages are ergative. Thus, assigning a high weight to this feature is very
convenient to be able to distinguish Basque from the other languages and, con-
sequently, correctly generate the first leaf of the tree.
Regarding latin languages, they have been classified in western and eastern
languages with a criterion based on phonetic features, more than 'phylogenesy'.
The probably best way to group these languages by filiation would be to relate
all of them to Latin in the same way. Moreover, we know that the Romania
is a linguistic continuum, and it is dicult to find the key features to define a
consistent gathering. In spite of this, many authors would consider a classification
like the one in Fig. 4 as the most closer to the correct one.
Fig. 4. 'Classical' derivation tree of the languages in the example
Such taxonomy has an additional problem: the grouping mixes the genetic,
typogical and areal criteria explained above. While the distinction between Ger-
manic and Latin languages is genetic, the distinction between the different ro-
mance languages is typological. Some different classifications could be obtained
depending on the methodology and the features taken. On the other hand, the
results that Basque obtains, very close to Spanish, are mainly because of areal
reasons, due to the fact that history has made these two very different languages
to coevolve in many features.
The system is challenged to obtain a tree similar to Fig. 4 without the inference
of the intermediate languages. Firstly, we focus in trying to obtain the tree of
Romance Languages with no restriction. The program generates two different
trees.
The best tree is shown in panel A of Fig. 5. This tree leaves Romanian aside
and groups Spanish with Portuguese and Catalan with Italian. The weights it
requires are in a range of 3.66-5.88. To build this tree, the system needs to
decrease the value of the following features: 1(c), 1(d), 1(e), 1(f), 3(e), 3(f), 3(g).
These features are phonetic and syntactical, and they are around 3.66.
The second proposed tree is shown in panel B of Fig. 5. This tree moves Ro-
manian closer to Spanish and Portuguese, and leaves Catalan and Italian in the
other group, with the French aside. However, this option requires a larger range
in the weight values: 3.5-6.98. Here, the traits that had their weight increased
were: 1(c), 1(d), 1(f), 3(a), 3(d), 3(f), 3(g), 4(a).