ATR denotes: Addition-Subtraction Twin-Gated Recurrent Unit. It relies on a twin gate mechanism that utilizes addition and subtraction operation to generate an input gate and a forget gate respectively.

The idea is proposed in our EMNLP18 conference paper: Simplifying Neural Machine Translation with Addition-Subtraction Twin-Gated Recurrent Networks. If you use ATR cell, please consider citing it:

@InProceedings{D18-1459,
  author = 	"Zhang, Biao
		and Xiong, Deyi
		and su, jinsong
		and Lin, Qian
		and Zhang, Huiji",
  title = 	"Simplifying Neural Machine Translation with Addition-Subtraction Twin-Gated Recurrent Networks",
  booktitle = 	"Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
  year = 	"2018",
  publisher = 	"Association for Computational Linguistics",
  pages = 	"4273--4283",
  location = 	"Brussels, Belgium",
  url = 	"http://aclweb.org/anthology/D18-1459"
}

To help others quickly learn the structure, we implement a word-based LM model on PTB dataset using both tensorflow and pytorch framework. We didnot pay too much efforts to optimize the hyperparameters. By LM task, we aim to show how the model works and give an intuitive comparison among LSTM, GRU and ATR.

Our main application of ATR is the machine translation task, where source sentence semantic encoding, target conditional language modeling, and complex semantic reasoning to capture source-target translation correspondence are required. This is a very suitable platform for testing RNN-based models, particularly in terms of model capacity. For more details, please see our Zero system.

Throughout our experiments, we mainly use shallow RNN models. Though recent researches could claim encouraging performance with deep architectures, we still believe a shallow model is the right way to demonstrate its capability.

Given current input x and previous hidden state h_rev, ATR composes them as follows:
\begin{equation}
p = W x
q = U h_prev
i = sigmoid(p + q)
f = sigmoid(p - q)
h = i * p + f * h_prev
\end{equation}
where W and U are the only weight matrices, just as a vanilla Elman structure.

Notice that the computation of forget gate f is very important and sensitive, you should subtract x related p with h_prev related q, i.e. p - q, so that the previous hidden state can help adjust the forget gate value to avoid value explosion. The equations above are slightly different from those in our paper, mainly on the computation of the forget gate f where we made a typo in the paper :(.

In short, these two gates are order-sensitive!

• You can also apply non-linearity to the hidden states, which sometimes can improve the performance.

  For example,

  • you can change h = i * p + f * h_prev to h = tanh(i * p + f * h_prev)); or
  • you can change h = i * p + f * h_prev to h = i * tanh(p) + f * h_prev; or
  • you can just apply tanh to h for output: o = tanh(h) and use o for the following models.

• You can also replace the input gate with 1. - f in a general manner.

Again, these desings are not necessary.

See lm, snli, cws for language model, natural language inference and chinese word segmentation respectively.

The lm part is implemented in tensorflow and pytorch, while others are implemented in theano.

• Implement the ATR unit in cuda-level. Welcome contributions in this direction.

For any question, please feel free to contact Biao Zhang.