Semantically enhanced text stemmer (SETS) for cross-domain document clustering

Abstract

This paper focuses on processing cross-domain document repositories, which is challenged by the word ambiguity and the fact that monosemic words are more domain-oriented than polysemic ones. The paper describes a semantically enhanced text normalization algorithm (SETS) aimed at improving document clustering and investigates the performance of the sk-means clustering algorithm across domains by comparing the cluster coherence produced with semantic-based and traditional (TF-IDF-based) document representations. The evaluation is conducted on 20 generic sub-domains of a thousand documents each randomly selected from the Reuters21578 corpus. The experimental results demonstrate improved coherence of the clusters produced by SETS compared to the text normalization obtained with the Porter stemmer. In addition, semantic-based text normalization is shown to be resistant to noise, which is often introduced in the index aggregation stage. © 2013 Springer-Verlag.