A Reduction for Efficient LDA Topic Reconstruction