Dual Encoder Classifier Models as Constraints in Neural Text Normalization

Abstract

Neural text normalization systems can achieve low error rates; however, the errors they make include not only ones from which the hearer can recover (such as reading $3 as three dollar) but also unrecoverable errors, such as reading $3 as three euros. FST decoding constraints have proven effective at reducing unrecoverable errors. In this paper we explore an alternative approach to error mitigation: using dual encoder classifiers trained with both positive and negative examples to implement soft constraints on acceptability. Since the error rates are very low, it is difficult to determine when improvement is significant, but qualitative analysis suggests that soft dual encoder constraints can help reduce the number of unrecoverable errors.