Towards the Evaluation of Feature Embedding Models of the Fusional Languages

Alina Wróblewska, Katarzyna Krasnowska-Kieras, Piotr Rybak

Abstract: An important component of a NLP system with the neural network architecture is an encoder that represents word features as dense vector representations, i.e. feature embeddings. According to the concept of feature embeddings, features sharing common linguistic information should have similar vectors and thus feature similarities can be captured. In this paper we investigate which features should be used in estimating NLP models of the fusional languages – tokens or lemmata. Furthermore, we research the methodological question whether the results of the intrinsic evaluation of feature embeddings are informative for downstream applications, or feature embedding models should be evaluated extrinsically. The presented evaluation experiments are conducted on Polish – a fusional Slavic language with a relatively free word order. However, the evaluation results can be approximately generalised to other Slavic languages, because the studied problems are common to them.