Phanxuan Phuc

fakerphan

I’m a AI Engineer in Vin Bigdata. And I summarize research papers in these topics:

2 mins read

Published by: NAVER LABS Europe & Université Grenoble Alpes

Multilinguality in Transformers

The following paper: Do Multilingual Neural Machine Translation Models Contain Language Pair Specific Attention Heads? asks a very good question. To answer it, the publishers tried to measure the importance of the self-attention heads in the encoder and the encoder-decoder attention heads of a many-to-one transformer. The NMT model was able to translate French, German, Italian, Spanish, and Korean sentences to English. It uses a variant of the Transformer-Big architecture with a shallower decoder: 16 attention heads, 6 encoder layers, and 3 decoder layers on TED2020 dataset.

Denoting $\left| I \right|,\ \left| J \right|$ as the number of source tokens and/or target tokens depending on whether we looked at the self-attention of encoder or the encoder-decoder cross attentions, The metrics used for importance are three:

  • Confidence:
    It is the mean of its maximum attention weights.
\[\text{conf}\left( \text{head} \right) = \frac{1}{\left| I \right|}\sum_{i \in I}^{}{\max_{j \in J}\alpha_{i,j}}\]
  • Variance:
    It’s measured by how much each individual position $i$ is away from the expected position $\mu_{i}$:
\[\text{var}\left( \text{head} \right) = - \sum_{i \in I}^{}{\sum_{j \in J}^{}{\alpha_{i,j}\left( \mu_{i} - j \right)^{2}}}\ \ \ \ \ \ \ \ \mu_{i} = \sum_{j \in J}^{}{\text{j.}\alpha_{i,j}}\]
  • Coverage:
    It measures the amount of attention a source token has received.
\[\text{cov}\left( \text{head} \right) = \sum_{j \in J}^{}\left( \sum_{i \in I}^{}\alpha_{i,j} \right)^{2}\]

According to the paper, the most important heads are language-independent as you can see in the following figure:

Note:
Even though most important heads are language-independent, in the paper they showed that it is possible to find the rare heads specific to a language pair via the extensive SBS (sequential backward selection) procedure.