Discriminative Reranking for Neural Machine Translation

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 7250–7264 August 1–6, 2021. ©2021 Association for Computational Linguistics 7250 Discriminative Reranking for Neural Machine Translation Ann Lee Michael Auli Marc’Aurelio Ranzato Facebook AI Research {annl,michaelauli,ranzato}@fb.com Abstract Reranking models enable the integration of rich features to select a better output hypoth- esis within an n-best list or lattice. These mod- els have a long history in NLP, and we re- visit discriminative reranking for modern neu- ral machine translation models by training a large transformer architecture. This takes as input both the source sentence as well as a list of hypotheses to output a ranked list. The reranker is trained to predict the observed dis- tribution of a desired metric, e.g. BLEU, over the n-best list. Since such a discriminator con- tains hundreds of millions of parameters, we improve its generalization using pre-training and data augmentation techniques. Exper- iments on four WMT directions show that our discriminative reranking approach is effec- tive and complementary to existing generative reranking approaches, yielding improvements of up to 4 BLEU over the beam search output. 1 Introduction Reranking models take a number of different out- put hypotheses generated by a baseline model and select one hypothesis based on more powerful fea- tures. Before the recent re-emergence of neural networks, these models have been well studied for several NLP tasks including parsing (Charniak and Johnson, 2005; Collins and Koo, 2005) and statis- tical machine translation (Och et al., 2004; Shen et al., 2004). Traditional statistical models (SMT) based on n-gram counts made very strong independence as- sumptions where features would only capture very local context information to avoid sparsity and poor generalization. A large n-best list produced by these models would then be passed to a discrimi- natively trained reranker which leverages features engineered to capture more global context (Och et al., 2004) yielding significant improvements to the quality of the translations. On the other hand, modern neural models (NMT) make much weaker independence assump- tions because predictions of standard sequence- to-sequence models depend on the entire source sentence as well as the target prefix generated. However, reranking may still be beneficial for two reasons: First, NMT systems are subject to expo- sure bias (Ranzato et al., 2016), i.e., models are never exposed to their own generations at training time, while a reranking model has been trained on model outputs. Second, beam search with auto- regressive models uses the chain rule to sum indi- vidual token-level probabilities to obtain a target sequence probability. However, individual prob- abilities are based on a limited amount of target context, while a reranking model can condition on the entire target context. Indeed, recent genera- tive reranking approaches applied to NMT, such as Noisy-Channel Decoding (NCD, Yee et al. 2019) which leverages a pre-trained language model and a backward model, show strong improvements over beam search outputs, as demonstrated in recent WMT evaluations (Ng et al., 2019). …

Content truncated to 500 words