Attention weights in transformer NMT fail aligning words between sequences but largely explain model predictions
Visualitza/Obre
10.18653/v1/2021.findings-emnlp.39
Inclou dades d'ús des de 2022
Cita com:
hdl:2117/369772
Tipus de documentComunicació de congrés
Data publicació2021
EditorAssociation for Computational Linguistics
Condicions d'accésAccés obert
Llevat que s'hi indiqui el contrari, els
continguts d'aquesta obra estan subjectes a la llicència de Creative Commons
:
Reconeixement 4.0 Internacional
Abstract
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.
CitacióFerrando, J.; Costa-jussà, M.R. Attention weights in transformer NMT fail aligning words between sequences but largely explain model predictions. A: Conference on Empirical Methods in Natural Language Processing. "Findings of the Association for Computational Linguistics: EMNLP 2021". Stroudsburg, PA: Association for Computational Linguistics, 2021, p. 434-443. DOI 10.18653/v1/2021.findings-emnlp.39.
Versió de l'editorhttps://aclanthology.org/2021.findings-emnlp.39/
Fitxers | Descripció | Mida | Format | Visualitza |
---|---|---|---|---|
2021.findings-emnlp.39.pdf | 1,184Mb | Visualitza/Obre |