Source-linked AI summary

OpenNMT: Open-source Toolkit for Neural Machine Translation

Guillaume Klein, Yoon Kim, Yuntian Deng, Josep Crego, Jean Senellart, Alexander M. Rush

arXiv:1709.03815v1cs.CL

TL;DR

Neural machine translation needs an open toolkit that supports research while remaining competitive, modular, and practical to train and deploy. OpenNMT addresses this need through a production-oriented sequence-to-sequence toolkit and broader ecosystem, with support for multiple model options, tasks, deployment, training, and community use.

  • Problem

    Neural machine translation research and production require an open toolkit supporting varied architectures and uses while maintaining competitive performance, modularity, and reasonable training requirements.

  • Method

    OpenNMT combines a complete sequence-to-sequence NMT implementation, extensive model and training options, deployment tools, recipes, demos, benchmarking, and community support.

  • Results

    OpenNMT provides a production-grade system and ecosystem used across machine translation, image-to-text, speech-to-text, and summarization.

  • Takeaways & Limitations

    OpenNMT offers a stable, modular framework intended for research, production use, and extension by an active NMT community.

Abstract

from arXiv · show

We introduce an open-source toolkit for neural machine translation (NMT) to support research into model architectures, feature representations, and source modalities, while maintaining competitive performance, modularity and reasonable training requirements.

1 Introduction

Neural machine translation has achieved major improvements and entered production use. OpenNMT is introduced as an open, MIT-licensed toolkit for researchers and engineers.

  • NMT has achieved remarkable improvements, particularly in human evaluation.
  • OpenNMT is an open, MIT-licensed initiative by SYSTRAN and Harvard NLP.
  • The toolkit is designed for researchers and engineers to benchmark, learn from, extend, and build upon.

2 Description

OpenNMT implements a complete sequence-to-sequence NMT approach with numerous model extensions and configurable training options. It targets limitations in efficiency, tooling, features, and documentation found in similar toolkits.

  • OpenNMT implements a complete sequence-to-sequence approach that achieved state-of-the-art results in many tasks, including machine translation.
  • The model supports multi-layer RNNs, attention, bidirectional encoders, word features, input feeding, residual connections, and beam search.
  • Training can be customized with multi-GPU support, retraining, data sampling, and learning-rate decay strategies.
  • OpenNMT addresses efficiency, tooling, feature, and documentation limitations associated with similar toolkits such as Nematus and Google’s seq2seq.

3 Ecosystem

OpenNMT extends beyond its core toolkit into an ecosystem for NMT and sequence modelling, supporting deployment, multiple tasks, training automation, demonstrations, and benchmarking.

  • The ecosystem includes an optimized C++ inference engine based on Eigen for easy and efficient model deployment and integration.
  • OpenNMT has been used for image-to-text, speech-to-text, and summarization in addition to machine translation.
  • Recipes, demo servers, and a benchmark platform automate training, showcase results, and compare approaches.

4 Community

OpenNMT includes a community that supports project use, specific training processes, and discussion of current and future NMT research and development.

  • The OpenNMT community provides support for using the project and addressing specific training processes.
  • It also discusses the current and future state of neural machine translation research and development.
  • The online forum counts more than 100 users, and the project has over 1,000 GitHub stars.

5 Conclusion

OpenNMT is a research toolkit for neural machine translation that prioritises efficiency and modularity while aiming to maintain strong results and support production use.

  • OpenNMT prioritises efficiency and modularity as core toolkit design goals.
  • The toolkit aims to maintain strong machine translation results at the research frontier.
  • OpenNMT is intended to provide a stable framework for production use while enlarging an active community.
Loading 1709.03815v1…