Document grouping is a necessity among a large number of articles published on internet. Several attempts have been done to improve this grouping process, while majority of the efforts are based on word appearance. In order to improve its quality, the grouping of documents need to be based on topic similarity between documents, instead of the frequency of word appearance. The topic similarity could be known from its latency, since the similarity of the word interpretation are often used in the same context. In the unsupervised learning process, SOM is often used, in which this approach simplifies the mapping of multi-dimension data. This research result shows that implementation of the latent structure decreases characteristic dimension by 32% of the word appearance, hence makes this approach more time efficient than others. The latent structure, however, when implemented on SOM Algorithm, is capable to obtain good quality result compared to word appearance frequency approach. It is then proven by 5% precision improvement, recall improvement of 3%, and another 4% from F-measure. While the achievement is not quite significant, the quality improvement is able to put the dominance of grouping process, compared to the original classification defined by the content provider.