2020/12/29 by Liang Ding, Ding, Liang, Longyue Wang +9 · 44 citations
Computer Science · Mathematics · #Artificial intelligence #Autoregressive model #Computation and Language (cs.CL) #Computer science #Divergence (linguistics) #FOS: Computer and information sciences #Lexical choice #Lexical item #Linguistics #Mathematics #Multimodal Machine Learning Applications #Natural Language Processing Techniques #Natural language processing #Raw data #Speech recognition #Statistics #Topic Modeling #cs.CL
paper · pdf · doi:10.48550/arxiv.2012.14583
published in arXiv (Cornell University) (Cornell University) · ICLR 2021
openalex publication_date 2020/12/29 · openalex created_date 2021/01/05 · arxiv created 2021/01/27 · arxiv updated 2021/01/28 · openalex updated_date 2026/07/28
Knowledge distillation (KD) is essential for training non-autoregressive translation (NAT) models by reducing the complexity of the raw data with an autoregressive teacher model. In this study, we empirically show that as a side effect of this training, the lexical choice errors on low-frequency words are propagated to the NAT model from the teacher model. To alleviate this problem, we propose to expose the raw data to NAT models to restore the useful information of low-frequency words, which are missed in the distilled data. To this end, we introduce an extra Kullback-Leibler divergence term derived by comparing the lexical choice of NAT model and that embedded in the raw data. Experimental results across language pairs and model architectures demonstrate the effectiveness and universality of the proposed approach. Extensive analyses confirm our claim that our approach improves performance by reducing the lexical choice errors on low-frequency words. Encouragingly, our approach pushes the SOTA NAT performance on the WMT14 English-German and WMT16 Romanian-English datasets up to 27.8 and 33.8 BLEU points, respectively. The source code will be released.