RNN-encoder decoder for machine translation

Introduction

In phrase based machine translation, phrase pairs are typically extracted using unsupervised alignment methods. These alignment methods which are typically generative in nature are unable to incorporate information about linguistic integrity and other measures of quality of phrase pairs. Hence, the extracted phrase pairs are often fairly noisy. One method of using these phrase pairs without deviating too far from the conventional phrase extraction procedures is to use additional features for each phrase pair and then learn the weights for these features using discriminative training where the goal is to discriminate between good and bad hypotheses. With the recent use of neural networks in machine translation, we have the capability to represent variable length sentences into a fixed size vector representation. This representation can be created based on any measure of quality we deem to be useful. Once we obtain vector representations of sentences/phrases based on some property of the language (syntax, semantics), it is relatively easy to ask ourselves how good a phrase pair is. This project builds on this work and other work in neural machine translation to estimate the phrasal similarity of phrase pairs. Evaluation will be conducted by using this metric as another feature in phrase based translation and for phrase table pruning.

Phrase Similarity

A natural question to ask when using unsupervised alignment for phrase extraction is : How do we trust the quality of the phrase pairs that we have extracted ? Obviously, this questions assumes the presence of a goodness function for phrase quality. One standard metric is to ask if the source and the target components of a phrase pair convey the same information (semantics). There are several linguistically motivated methods to do this. We will describe a method that encodes semantic information into a fixed sized vector using an unsupervised method and then use it to measure phrase pair similarity. Before we proceed further, let us look at some of the kinds of problems we may encounter in an actual phrase table with phrase pairs.

• Rare phrases: Rare phrase pair occurrences provide a sub-optimal estimate for phrase translation probabilities.
  $p(\text{sorona} ;;|;; \text{tristifical}) = 1$
  $p(\text{tristifical} ;;|;; \text{sorona}) = 1$

• Independence assumptions : The choice to use one phrase pair over an another is largely independent of previous decisions.

• Segmentation : Phrase segmentation is generally not linguistically motivated and a large percentage of the phrase pairs are not good translations.
  (, veinte délares, era ||| you! twenty dollars, it was)
  (Exactamente como ||| how they want to)

The RNN encoder-decoder framework

This project draws motivation from @schwenk_continuous_2012 which demonstrates the use of a feed forward neural network in phrase based machine translation. Specifically, it implements the RNN encoder-decoder framework for estimating phrase similarity described in @cho_learning_2014. It is worth noting that a similar exposition appears in @sutskever_sequence_2014, while @bahdanau_neural_2014 extends this framework to perform joint alignment and translation.

The RNN encoder-decoder framework consists of a pair of recurrent neural networks (RNNs). The goal of the encoder is to encode a variable length input sentence to a fixed size vector representation. The decoder on the other hand, takes a fixed size vector representation and produces a variable length sentence. In a probabilistic framework, this is akin to learning the conditional distribution $p(y_1, \cdots , y_{T'} | x_1, \cdots , x_{T})$ where $x$ and $y$ are tokens from some language pair and $T$ is not necessarily equal to $T'$.

Encoder

The encoder is a straightforward RNN which consists of a hidden state $h$ which is updated at every time step $t$ while operating on some sequence $x = (x_i,\cdots,x_T)$. That is $$h_{t} = f(h_{t-1}, x_t) \nonumber$$ where $f$ is some non-linear activation function. We will use a special form of the LSTMs (@hochreiter_long_1997) as our activation function as described in @cho_learning_2014 unless otherwise stated. The hidden state obtained for the entire input sequence will be called the context vector $c$.

Decoder

The decoder is a modified RNN that will produce the next symbol $y_t$ given the current hidden state $h_t$, previous symbol generated $y_{t-1}$ and the context vector from the encoder. $c$ The hidden state in turn is created as a function of the previous symbol generated $y_{t-1}$, the context vector from the encoder $c$ and the previous hidden state $h_{t-1}$. That is, $$\begin{aligned} & h_{t} = f(h_{t-1}, y_{t-1}, c) \ & p(y_t | y_{t-1}, \cdots , y_1, c) = g(h_t, y_{t-1}, c)\end{aligned}$$ where $f$ is an LSTM-like non-linear activation function and $g$ is a logistic sigmoid (softmax) function (which is required to produce valid probabilities).

Evaluation

We will explore the validity of the task by two techniques on the translation quality of a parallel dataset (to be decided):

1. Phrase features : In a phrase based model, we plan to use this model to estimate phrasal similarity and use it as an additional feature in the phrase table. The SMT system will then use this feature for decoding during tuning and at test time. We hope to achieve better translation quality with this feature.

2. Phrase table pruning : As an additional test of the usefulness of this feature, we will use it to prune an existing phrase table. The main idea is that the phrase table generally contains a large number of phrase pairs that are bad (given our goodness function), and any phrase pair below a certain threshold can be eliminated. This evaluation method will remain an optional investigation and will be performed if time permits.

Software

We plan to use two major software components for this project :

1. Theano : A python based library that allows relatively easy implementation of GPU/CPU based operations for implementing neural networks.

2. Moses : The phrase based translation system in Moses will be used and we plan to add an additional feature function for decoding. If found useful, we will contribute this feature to the Moses code base.

Interim Report

This section reports updates with this project since the proposal was submitted. There are two major components that have been finalized

1. Dataset and Baseline : The French-English translation task from WMT-2013 will be used to evaluate this project. The dataset from the corresponding task and partitions corresponding to train, dev and test will be used to build the phrase table, tune and evaluate respectively. The baseline system will be a phrase based SMT system built using Moses.

2. Preliminary RNN implementation : A prototype of the RNN that will be used in this task has been built and exists at https://github.com/noisychannel/mt-rnn. This version uses Theano and is built so that it can be adapted for any task. At the current moment, an example which allows one to predict tags with the ATIS data and evaluate using the CONLL evaluation setup is included.