Tree2Seq: Tree-to-Sequence Attentional Neural Machine Translation

We have proposed a novel syntactic ANMT model, "Tree-to-Sequence Attentional Neural Machine Translation" [1]. We extend an original sequence-to-sequence model [2] with the source-side phrase structure. Our model has an attention mechanism that enables the decoder to generate a translated word while softly aligning it with source phrases and words. Here is an online demo of Tree2Seq.

Description

C++ codes of the syntactic Attention-based Neural Machine Translation (ANMT) model.

1. AttentionTreeEncDec.xpp: our ANMT model, "Tree-to-Sequence Attentional Neural Machine Translation"
2. AttentionEncDec.xpp: Baseline ANMT model [3]
3. /data/: Tanaka Corpus (EN-JP) [4]

Requirement

• Eigen, a template libary for linear algebra (http://eigen.tuxfamily.org/index.php?title=Main_Page)
• N3LP, C++ libaray for neural network-based NLP (https://github.com/hassyGo/N3LP)
• Boost, C++ library for tree structure (http://www.boost.org/)
• Option: Enju, a syntactic parser for English (http://kmcs.nii.ac.jp/enju/?lang=en)

Usage

1. Modify the paths of EIGEN_LOCATION, SHARE_LOCATION and BOOST_LOCATION. See Makefile.
2. $ bash setup.sh
3. $./tree2seq (Then, training the AttentionTreeEncDec model starts.)
4. Modify main.cpp if you want to change the model.

(!) Attention: I prepare a small corpus of Tanaka corpus. You need over 100,000 parallel corpus.

Citation

• [1] Akiko Eriguchi, Kazuma Hashimoto, and Yoshimasa Tsuruoka. 2015. "Tree-to-Sequence Attentional Neural Machine Translation". In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (ACL 2016).
• [2] I. Sutskever, O. Vinyals, and Q. V. Le. 2014. "Sequence to Sequence Learning with Neural Networks". In Proceedings of Advances in Neural Information Processing Systems 27 (NIPS2014).
• [3] T. Luong, H. Pham, and C. D. Manning. 2015. "Effective Approaches to Attention-based Neural Machine Translation". In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing (EMNLP2015).
• [4] Tanaka Corpus

Contact

Thank you for your interests. If you have any questions and comments, feel free to contact us.

• eriguchi [.at.] logos.t.u-tokyo.ac.jp
• hassy [.at.] logos.t.u-tokyo.ac.jp