Contrastive Learning for Robust Neural Machine Translation with ASR Errors

Abstract

AbstractNeural machine translation (NMT) models have recently achieved excellent performance when translating non-noised sentences. Unfortunately, they are also very brittle and easily falter when fed with noisy sentences, i.e., from automatic speech recognition (ASR) output. Due to the lack of Chinese-to-English translation test set with natural Chinese-side ASR output, related studies artificially add noise into Chinese sentences to evaluation translation performance. In this paper we eliminate such strong limitation and manually construct natural ASR output on popular Chinese-to-English NIST translation test dataset, NISTasr which consists of 680 documents with 7,688 sentences. To train an NMT model being robust to ASR output, we take contrastive learning framework to close the gap among representations of original input and its perturbed counterpart. Experimental results on NIST Chinese-to-English translation show that our approach significantly improves the translation performance of ASR output while having negligible effect on the translation performance of non-noised sentences. It also shows that the translation performance gap is quite big (e.g., 44.28 in BLEU v.s. 35.20) when translating non-noised sentences and sentences of ASR output.KeywordsNeural machine translationAutomatic speech recognitionContrastive learning