Why not be Versatile? Applications of the SGNMT Decoder for Machine Translation

Why not be Versatile? Applications of the SGNMT Decoder for Machine Translation Felix Stahlberg† fs439@cam.ac.uk Danielle Saunders† ds636@cam.ac.uk Gonzalo Iglesias‡ giglesias@sdl.com Bill Byrne‡† bill.byrne@eng.cam.ac.uk, bbyrne@sdl.com †Department of Engineering, University of Cambridge, UK ‡SDL Research, Cambridge, UK Abstract SGNMT is a decoding platform for machine translation which allows paring various modern neural models of translation with different kinds of constraints and symbolic models. In this paper, we describe three use cases in which SGNMT is currently playing an active role: (1) teaching as SGNMT is being used for course work and student theses in the MPhil in Machine Learning, Speech and Language Technology at the University of Cambridge, (2) research as most of the research work of the Cambridge MT group is based on SGNMT, and (3) technology transfer as we show how SGNMT is helping to transfer research findings from the laboratory to the industry, eg. into a product of SDL plc. 1 Introduction The rate of innovation in machine translation (MT) has gathered impressive momentum over the recent years. The discovery and maturation of the neural machine translation (NMT) paradigm (Sutskever et al., 2014; Bahdanau et al., 2015) has led to steady and substantial im- provements of translation performance (Williams et al., 2014; Jean et al., 2015; Luong et al., 2015; Chung et al., 2016; Wu et al., 2016; Gehring et al., 2017; Vaswani et al., 2017). Fig. 1 shows that this progress is often driven by significant changes in the network architecture. This volatility poses major challenges in MT-related research, teaching, and industry. Researchers potentially spend a lot of time implementing to keep their setups up-to-date with the latest models, teaching needs to identify suitable material in a changing environment, and the in- dustry faces demanding speed requirements on its deployment processes. Another practical challenge many researchers are struggling with is the large number of available NMT tools (van Merri¨enboer et al., 2015; Junczys-Dowmunt et al., 2016; Klein et al., 2017; Sennrich et al., 2017; Helcl and Libovick´y, 2017; Bertoldi et al., 2017; Hieber et al., 2017).1 Committing to one particular NMT tool bears the risk of being outdated soon, as keeping up with the pace of research is especially costly for NMT software developers. The open-source SGNMT (Syntactically Guided Neural Machine Translation) de- coder2 (Stahlberg et al., 2017b) is our attempt to mediate the effects of the rapid progress in 1See https://github.com/jonsafari/nmt-list for a complete list of NMT software. 2Full documentation available at http://ucam-smt.github.io/sgnmt/html/. Proceedings of AMTA 2018, vol. 1: MT Research Track Boston, March 17 - 21, 2018 | Page 208 Figure 1: Best systems on the English-German WMT news-test2014 test set over the years (BLEU script: Moses’ multi-bleu.pl). MT and the diversity of available NMT software. SGNMT introduces the concept of predictors as abstract scoring modules with left-to-right semantics. We can think of a predictor as an inter- face to a particular neural model or NMT tool. However, the interface also allows to implement constraints like in lattice or n-best list rescoring, and symbolic models …

Content truncated to 500 words