Experiments with artificially generated noise for cleansing noisy text

Recent works show that the problem of noisy text normalization can be treated as a machine translation (MT) problem with convincing results. There have been supervised MT approaches which use noisy-regular parallel data for training an MT model, as well as unsupervised models which learn the translation probabilities in alternative ways and try to mimic the MT-based approach. While the supervised approaches suffer from data annotation and domain adaptation difficulties, the unsupervised models lack a holistic approach catering to all types of noise. In this paper, we propose an algorithm to artificially generate noisy text in a controlled way, from any regular English text. We see this approach as an alternative to the unsupervised approaches while getting the advantages of a parallel corpus based MT approach. We generate parallel noisy text from two widely used regular English datasets and test the MT-based approach for text normalization. Semi-supervised approaches were also tried to explore different ways of improving the parallel corpus (manually annotated) based MT approach by using the generated noisy text. An extensive analysis based on comparison of our approaches with both the supervised as well as unsupervised approaches is presented.