Neural word alignment

Neural versions of SMT-style word alignment models like the IBM model hierarchy or HMM-based alignment models.

This project is based on TensorFlow’s high-level machine learning API (tf.estimator).

Getting started

This repository contains implementations of neural alignment models, and the infrastructure for training and basic inference. We do not provide functionality for data generation, preprocessing, tokenization etc. Instead, we maintain compatibility with the Tensor2Tensor project, ie. training and dev sets created by T2T data generators can be used in Nizza.