Random Indexing Distributional Semantic Models for Croatian Language

Distributional semantic models (DSMs) model semantic relations between expressions by comparing the contexts in which these expressions occur. This paper presents an extensive evaluation of distributional semantic models for Croatian language. We focus on random indexing models, an efficient and scalable approach to building DSMs. We build a number of models with different parameters (dimension, context type, and similarity measure) and compare them against human semantic similarity judgments. Our results indicate that even low-dimensional random indexing models may outperform the raw frequency models, and that the choice of the similarity measure is most important.