SGNMT is an open-source framework for neural machine translation (NMT) and other sequence prediction tasks. The tool provides a flexible platform which allows pairing NMT with various other models such as language models, length models, or bag2seq models. It supports rescoring both n-best lists and lattices. A wide variety of search strategies is available for complex decoding problems.

SGNMT is compatible with multiple NMT implementations based on Theano (Blocks) and TensorFlow (the extended seq2seq tutorial and tensor2tensor).


For example, NMT decoding can be started with this command:

$ python --predictors nmt --src_test sentences.txt

where sentences.txt is a plain (indexed) text file with sentences. Rescoring OpenFST lattices with NMT is also straight-forward:

$ python --predictors nmt,fst --fst_path lattices/%d.fst --src_test sentences.txt

See the Tutorial: Basics for more examples.


  • Syntactically guided neural machine translation (NMT lattice rescoring)
  • NMT support in Theano (Blocks) and TensorFlow (Tensor2Tensor)
  • n-best list rescoring with NMT
  • Integrating external n-gram posterior probabilities used in MBR
  • Ensemble NMT decoding
  • Forced NMT decoding
  • Integrating language models (Kneser-Ney, NPLM, RNNLM)
  • Different search algorithms (beam, A*, depth first search, greedy...)
  • Target sentence length modelling
  • Bag2Sequence models and decoding algorithms
  • Joint decoding with word- and subword/character-level models
  • Hypothesis recombination
  • Heuristic search
  • Extensions to NMT training in Blocks (reshuffling, fixing and customizing word embeddings, ...)
  • Neural word alignment (Blocks/Theano)
  • ...


  • Felix Stahlberg, University of Cambridge
  • Eva Hasler, SDL Research


The project is licensed under the Apache 2 license.

Indices and tables