Improving Text Normalization via Unsupervised Model and Discriminative Reranking

Various models have been developed for normalizing informal text. In this paper, we propose two methods to improve normalization performance. First is an unsupervised approach that automatically identifies pairs of a non-standard token and proper word from a large unlabeled corpus. We use semantic similarity based on continuous word vector representation, together with other surface similarity measurement. Second we propose a reranking strategy to combine the results from different systems. This allows us to incorporate information that is hard to model in individual systems as well as consider multiple systems to generate a final rank for a test case. Both word- and sentence-level optimization schemes are explored in this study. We evaluate our approach on data sets used in prior studies, and demonstrate that our proposed methods perform better than the state-of-the-art systems.