Papers › Alleviating the Inequality of Attention Heads for Neural Machine Translation

Alleviating the Inequality of Attention Heads for Neural Machine Translation

21 Sep 2020COLING 2022 10arXiv:2009.09672archive 2025-07-28

Zewei Sun, Shu-Jian Huang, Xin-yu Dai, Jia-Jun Chen

Recent studies show that the attention heads in Transformer are not equal. We relate this phenomenon to the imbalance training of multi-head attention and the model dependence on specific heads. To tackle this problem, we propose a simple masking method: HeadMask, in two specific ways. Experiments show that translation improvements are achieved on multiple language pairs. Subsequent empirical analyses also support our assumption and confirm the effectiveness of the method.

PaperPDFConference PDF

Code

No code repository is listed for this paper in the archive or in Syntology's graph.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Machine TranslationTranslation

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Machine Translation IWSLT2015 Vietnamese-English HeadMask (Random-18) BLEU 26.85 #1 of 2 Archive leaderboard report
Machine Translation IWSLT2015 Vietnamese-English HeadMask (Impt-18) BLEU 26.36 #2 of 2 Archive leaderboard report
Machine Translation WMT2016 Romanian-English HeadMask (Impt-18) BLEU score 32.95 #10 of 21 Archive leaderboard report
Machine Translation WMT2016 Romanian-English HeadMask (Random-18) BLEU score 32.85 #13 of 21 Archive leaderboard report
Machine Translation WMT2017 Turkish-English HeadMask (Random-18) BLEU score 17.56 #1 of 2 Archive leaderboard report
Machine Translation WMT2017 Turkish-English HeadMask (Impt-18) BLEU score 17.48 #2 of 2 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

Absolute Position EncodingsAdamAttentionBPEDense ConnectionsDropoutLabel SmoothingLayer NormalizationLinear LayerMulti-Head AttentionPosition-Wise Feed-Forward LayerResidual ConnectionSoftmaxTransformer

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections