Finnish resources for evaluating language model semantics
Viljami Venekoski, Jouko Vankka · DSpace repository (University of Tartu) · 2017
Distributional language models have consistently been demonstrated to capture semantic properties of words.However, research into the methods for evaluating the accuracy of the modeled semantics has been limited, particularly for less-resourced languages.This research presents three resources for evaluating the semantic quality of Finnish language distributional models: (1) semantic similarity judgment resource, as well as (2) a word analogy and (3) a word intrusion test set.The use of evaluation resources is demonstrated in practice by presenting them with different language models built from varied corpora.