Classical Structured Prediction Losses for Sequence to Sequence Learning

Proceedings of NAACL-HLT 2018, pages 355–364 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Classical Structured Prediction Losses for Sequence to Sequence Learning Sergey Edunov∗, Myle Ott∗, Michael Auli, David Grangier, Marc’Aurelio Ranzato Facebook AI Research Menlo Park, CA and New York, NY Abstract There has been much recent work on train- ing neural attention models at the sequence- level using either reinforcement learning-style methods or by optimizing the beam. In this paper, we survey a range of classical objec- tive functions that have been widely used to train linear models for structured prediction and apply them to neural sequence to se- quence models. Our experiments show that these losses can perform surprisingly well by slightly outperforming beam search optimiza- tion in a like for like setup. We also report new state of the art results on both IWSLT’14 German-English translation as well as Giga- word abstractive summarization. On the large WMT’14 English-French task, sequence-level training achieves 41.5 BLEU which is on par with the state of the art.1 1 Introduction Sequence to sequence models are usually trained with a simple token-level likelihood loss (Sutskever et al., 2014; Bahdanau et al., 2014). However, at test time, these models do not pro- duce a single token but a whole sequence. In order to resolve this inconsistency and to potentially improve generation, recent work has focused on training these models at the sequence-level, for instance using REINFORCE (Ranzato et al., 2015), actor-critic (Bahdanau et al., 2016), or with beam search optimization (Wiseman and Rush, 2016). Before the recent work on sequence level train- ing for neural networks, there has been a large body of research on training linear models at the ∗ Equal contribution. 1An implementation of the losses is available as part of fairseq at https://github.com/ facebookresearch/fairseq-py/tree/ classic_seqlevel sequence level. For example, direct loss opti- mization has been popularized in machine transla- tion with the Minimum Error Rate Training algo- rithm (MERT; Och 2003) and expected risk min- imization has an extensive history in NLP (Smith and Eisner, 2006; Rosti et al., 2010; Green et al., 2014). This paper revisits several objective func- tions that have been commonly used for structured prediction tasks in NLP (Gimpel and Smith, 2010) and apply them to a neural sequence to sequence model (Gehring et al., 2017b) (§2). Specifically, we consider likelihood training at the sequence- level, a margin loss as well as expected risk train- ing. We also investigate several combinations of global losses with token-level likelihood. This is, to our knowledge, the most comprehensive com- parison of structured losses in the context of neural sequence to sequence models (§3). We experiment on the IWSLT’14 German- English translation task (Cettolo et al., 2014) as well as the Gigaword abstractive summarization task (Rush et al., 2015). We achieve the best re- ported accuracy to date on both tasks. We find that the sequence level losses we survey perform similarly to one another and outperform beam search optimization (Wiseman and Rush, …

Content truncated to 500 words