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
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 · showhide
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.