Transformer + Noisy Channel
Transformer model directly estimates the posterior probability of a target sequence $y$ given a source sequence $x$. The Noisy Channel model operates in the reverse direction. It estimates the likelihood probability $p\left( x \middle| y \right)$ with the help of a language model probability $p\left( y \right)$. To do so, the Noisy channel model applies the Naive Bayes Rule:
\[p\left( y \middle| x \right) = \frac{p\left( x \middle| y \right)\text{.p}\left( y \right)}{p\left( x \right)}\]The model responsible for calculating $p\left( y \middle| x \right)$ is called “Direct Model”, the model responsible for calculating $p\left( x \middle| y \right)$ is called “Channel Model”. Modeling $p(x)$ is ignored since it is constant for all $y$.
The noisy channel approach was widely used in statistical machine translation. However, in this paper: Simple and Effective Noisy Channel Modeling for Neural Machine Translation published by FAIR in 2019 they decided to use the standard Transformer architecture as a channel model. The official code for this paper can be found in the fairseq official GitHub repository: fairseq/noisychannel.
In this paper, a standard Transformer model is used as the channel model, they basically trained the model to translate the target sentence to the source sentence. And they trained another Transformer model as the direct model. And they also trained a transformer language model.
Experiments
Using the English-German WMT’17 dataset for training, news2016 for validation and news2017 for testing, they used:
-
Language Model:They trained two big Transformer language models with adaptive input representations with 12 blocks. one on the German newscrawl data distributed by WMT’18 comprising 260M sentences and another one on the English newscrawl data comprising 193M sentences. Both use a BPE vocabulary of 32K tokens.
-
Direct Model: They use big Transformers where the encoder and decoder embeddings are not shared between them since the source and target vocabularies were learned separately.
-
Channel Model: They trained Transformer models to translate from the target to the source (En-De).
For comparison, they tried the following configurations:
-
DIR: just the direct model.
-
DIR ENS: an ensemble of two direct models.
-
DIR+LM: a direct model + a language Model.
-
DIR+RL: a direct model + a right-to-left seq2seq model.
-
DIR+RL+LM: a direct model + a right-to-left seq2seq model + a language model.
-
CH+DIR: a channel model + a direct model.
-
CH+DIR+LM: a channel model + a direct model + a language model.
Online Decoding
To perform decoding in the noisy channel model, we will need to perform a two-step beam search. For beam size $k_{1}$ in the channel model, we will collect $k_{2}$ possible next words extensions from the direct model for each beam. The, we will score the resulting $k_{1} \times k_{2}$ according to the following equation:
\[\frac{1}{t}\log p\left( y \middle| x \right) + \frac{\lambda}{s}\left( \log p\left( x \middle| y \right) + \log p\left( y \right) \right)\]Where $t$ is the length of the target prefix $y$, $s$ is the source sentence length and $\lambda$ is a tunable weight.
In online decoding, you don’t have the whole target sentence. In this case, you can’t use the right-to-left seq2seq model since it doesn’t know how the target sentence will be like. The following table summarizes the BLEU score over news2016 and news2017 en-de datasets:
The previous results were produces using $k_{1} = 5,\ k_{2} = 10$.
Re-ranking
In re-ranking, you have access to the full target sentence. The purpose is just re-rank them. The following table shows the re-ranking BLEU with different n-best list sizes on news2016 of WMT De-En. As we can see, the noisy channel model configuration obtained the highest scores:
