Current efforts in computational historical linguistics are predominantly concerned
with phylogenetic inference. Methods for ancestral state reconstruction have only
been applied sporadically. In contrast to phylogenetic algorithms, automatic reconstruction methods presuppose phylogenetic information in order to explain what has
evolved when and where. Here we report a pilot study exploring how well automatic
methods for ancestral state reconstruction perform in the task of onomasiological
reconstruction in multilingual word lists, where algorithms are used to infer how the
words evolved along a given phylogeny, and reconstruct which cognate classes were
used to express a given meaning in the ancestral languages. Comparing three different methods, Maximum Parsimony, Minimal Lateral Networks, and Maximum Likeli-
hood on three different test sets (Indo-European, Austronesian, Chinese) using binary
and multi-state coding of the data as well as single and sampled phylogenies, we find
that Maximum Likelihood largely outperforms the other methods. At the same time,
however, the general performance was disappointingly low, ranging between 0.66 (Chinese) and 0.79 (Austronesian) for the F-Scores. A closer linguistic evaluation of the
reconstructions proposed by the best method and the reconstructions given in the gold
standards revealed that the majority of the cases where the algorithms failed can be
attributed to problems of independent semantic shift (homoplasy), to morphological
processes in lexical change, and to wrong reconstructions in the independently created
test sets that we employed.