mlsanityData to 7 Oct 2026

Attention Weights in Transformer NMT Fail Aligning Words Between Sequences but Largely Explain Model Predictions

arXiv 2109.05853 ↗First posted 13 Sep 2021cs.CL cs.AIcited by 11

Not on the map. Why?

Abstract (arXiv)

This work proposes an extensive analysis of the Transformer architecture in the Neural Machine Translation (NMT) setting. Focusing on the encoder-decoder attention mechanism, we prove that attention weights systematically make alignment errors by relying mainly on uninformative tokens from the source sequence. However, we observe that NMT models assign attention to these tokens to regulate the contribution in the prediction of the two contexts, the source and the prefix of the target sequence. We provide evidence about the influence of wrong alignments on the model behavior, demonstrating that the encoder-decoder attention mechanism is well suited as an interpretability method for NMT. Finally, based on our analysis, we propose methods that largely reduce the word alignment error rate compared to standard induced alignments from attention weights.

Key facts. Cited by 11 papers in the map (data through 7 Oct 2026). Builds on 22 of its 39 references that are in the map. First posted on arXiv 13 Sep 2021.

Builds on

22 of 39 references.

References · 22

  1. 2017Attention Is All You Needcited by 59,932
  2. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  3. 2015Neural Machine Translation of Rare Words with Subword Unitscited by 2,898
  4. 2019fairseq: A Fast, Extensible Toolkit for Sequence Modelingcited by 1,486
  5. 2017SmoothGrad: removing noise by adding noisecited by 1,097
  6. 2019What Does BERT Look At? An Analysis of BERT's Attentioncited by 994
  7. 2019Attention is not Explanationcited by 916
  8. 2019Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Prunedcited by 844
  9. 2019Is Attention Interpretable?cited by 387
  10. 2015Visualizing and Understanding Neural Models in NLPcited by 199
  11. 2017Visualizing Deep Neural Network Decisions: Prediction Difference Analysiscited by 166
  12. 2021The Curious Case of Hallucinations in Neural Machine Translationcited by 161
10 more
  1. 2019On Identifiability in Transformerscited by 115
  2. 2019Learning to Deceive with Attention-Based Explanationscited by 111
  3. 2019Jointly Learning to Align and Translate with Transformer Modelscited by 75
  4. 2020Analyzing the Source and Target Contributions to Predictions in Neural Machine Translationcited by 71
  5. 2020Fixed Encoder Self-Attention Patterns in Transformer-Based Machine Translationcited by 42
  6. 2020Accurate Word Alignment Induction from Neural Machine Translationcited by 32
  7. 2019Adding Interpretable Attention to Neural Translation Models Improves Word Alignmentcited by 30
  8. 2019Saliency-driven Word Alignment Interpretation for Neural Machine Translationcited by 27
  9. 2019Naver Labs Europe's Systems for the WMT19 Machine Translation Robustness Taskcited by 22
  10. 2019Towards Understanding Neural Machine Translation with Word Importancecited by 19

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2022: 5; 2023: 1; 2024: 2; 2025: 2; 2026: 1.

11 papers.

  1. 2022Towards Faithful Model Explanation in NLP: A Surveycited by 129
  2. 2022Towards Opening the Black Box of Neural Machine Translation: Source and Target Interpretations of the Transformercited by 45
  3. 2022SBERT studies Meaning Representations: Decomposing Sentence Embeddings into Explainable Semantic Featurescited by 33
  4. 2022Optimal Transport for Unsupervised Hallucination Detection in Neural Machine Translationcited by 13
  5. 2024Attention Mechanisms Don't Learn Additive Models: Rethinking Feature Importance for Transformerscited by 8
  6. 2022Word Alignment in the Era of Deep Learning: A Tutorialcited by 3
  7. 2023Predicting Human Translation Difficulty with Neural Machine Translationcited by 1
  8. 2024Non-Fluent Synthetic Target-Language Data Improve Neural Machine Translationcited by 0
  9. 2025Decoding the Multimodal Maze: A Systematic Review on the Adoption of Explainability in Multimodal Attention-based Modelscited by 0
  10. 2025TransAlign: Machine Translation Encoders are Strong Word Aligners, Toocited by 0
  11. 2026Evaluating Explainable AI Attribution Methods in Neural Machine Translation via Attention-Guided Knowledge Distillationcited by 0