Topic Detection Using MFSs

When analyzing a document collection, a key piece of information is the number of distinct topics it contains. Document clustering has been used as a tool to facilitate the extraction of such information. However, existing clustering methods do not take into account the sequences of the words in the documents, and usually do not have the means to describe the contents within each topic cluster. In this paper, we record our investigation and results using Maximal Frequent word Sequences (MFSs) as building blocks in identifying distinct topics. The supporting documents of MFSs are grouped into an equivalence class and then linked to a topic cluster, and the MFSs serve as the document cluster identifier. We describe the original method in extracting the set of MFSs, and how it can be adapted to identify topics in a textual dataset. We also demonstrate how the MFSs themselves can act as topic descriptors for the clusters. Finally, the benchmarking study with other existing clustering methods, i.e. k-Means and EM algorithm, shows the effectiveness of our approach for topic detection.