In recent years a variety of approaches in computing semantic relatedness have been proposed. However, the algorithms and resources employed differ strongly, as well as the results obtained under different experimental conditions. This article investigates the quality of various semantic relatedness measures in a comparative study. We conducted an extensive experiment using a broad variety of measures operating on social networks, lexical-semantic nets and co-occurrence in text corpora. For two sample data sets we obtained human relatedness judgements which were compared to the estimates of the automated measures. We also analyzed the algorithms implemented and resources employed from a theoretical point of view, and we examined several practical issues, such as run time and coverage. While the performance of all measures is still mediocre, we could observe that in terms of of coverage and correlation distributional measures operating on controlled corpora perform best.