On handling textual errors in latent document modeling

As large-scale text data become available on the Web, textual errors in a corpus are often inevitable (e.g., digitizing historic documents). Due to the calculation of frequencies of words, however, such textual errors can significantly impact the accuracy of statistical models such as the popular Latent Dirichlet Allocation (LDA) model. To address such an issue, in this paper, we propose two novel extensions to LDA (i.e., TE-LDA and TDE-LDA): (1) The TE-LDA model incorporates textual errors into term generation process; and (2) The TDE-LDA model extends TE-LDA further by taking into account topic dependency to leverage on semantic connections among consecutive words even if parts are typos. Using both real and synthetic data sets with varying degrees of "errors", our TDE-LDA model outperforms: (1) the traditional LDA model by 16%-39% (real) and 20%-63% (synthetic); and (2) the state-of-the-art N-Grams model by ll%-27% (real) and 16%-54% (synthetic).