Comparison of generality based algorithm variants for automatic taxonomy generation

Abstract

We compare a family of algorithms for the automatic generation of taxonomies by adapting the Heymann-algorithm in various ways. The core algorithm determines the generality of terms and iteratively inserts them in a growing taxonomy. Variants of the algorithm are created by altering the way and the frequency, generality of terms is calculated. We analyse the performance and the complexity of the variants combined with a systematic threshold evaluation on a set of seven manually created benchmark sets. As a result, betweenness centrality calculated on unweighted similarity graphs often performs best but requires threshold fine-tuning and is computationally more expensive than closeness centrality. Finally, we show how an entropy-based filter can lead to more precise taxonomies.