Decoding Layer Saliency in Language Transformers

Elizabeth Mary Hou, Gregory David Castanon
Proceedings of the 40th International Conference on Machine Learning, PMLR 202:13285-13308, 2023.

Abstract

In this paper, we introduce a strategy for identifying textual saliency in large-scale language models applied to classification tasks. In visual networks where saliency is more well-studied, saliency is naturally localized through the convolutional layers of the network; however, the same is not true in modern transformer-stack networks used to process natural language. We adapt gradient-based saliency methods for these networks, propose a method for evaluating the degree of semantic coherence of each layer, and demonstrate consistent improvement over numerous other methods for textual saliency on multiple benchmark classification datasets. Our approach requires no additional training or access to labelled data, and is comparatively very computationally efficient.

Cite this Paper

BibTeX
@InProceedings{pmlr-v202-hou23a, title = {Decoding Layer Saliency in Language Transformers}, author = {Hou, Elizabeth Mary and Castanon, Gregory David}, booktitle = {Proceedings of the 40th International Conference on Machine Learning}, pages = {13285--13308}, year = {2023}, editor = {Krause, Andreas and Brunskill, Emma and Cho, Kyunghyun and Engelhardt, Barbara and Sabato, Sivan and Scarlett, Jonathan}, volume = {202}, series = {Proceedings of Machine Learning Research}, month = {23--29 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v202/hou23a/hou23a.pdf}, url = {https://proceedings.mlr.press/v202/hou23a.html}, abstract = {In this paper, we introduce a strategy for identifying textual saliency in large-scale language models applied to classification tasks. In visual networks where saliency is more well-studied, saliency is naturally localized through the convolutional layers of the network; however, the same is not true in modern transformer-stack networks used to process natural language. We adapt gradient-based saliency methods for these networks, propose a method for evaluating the degree of semantic coherence of each layer, and demonstrate consistent improvement over numerous other methods for textual saliency on multiple benchmark classification datasets. Our approach requires no additional training or access to labelled data, and is comparatively very computationally efficient.} }
Endnote
%0 Conference Paper %T Decoding Layer Saliency in Language Transformers %A Elizabeth Mary Hou %A Gregory David Castanon %B Proceedings of the 40th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2023 %E Andreas Krause %E Emma Brunskill %E Kyunghyun Cho %E Barbara Engelhardt %E Sivan Sabato %E Jonathan Scarlett %F pmlr-v202-hou23a %I PMLR %P 13285--13308 %U https://proceedings.mlr.press/v202/hou23a.html %V 202 %X In this paper, we introduce a strategy for identifying textual saliency in large-scale language models applied to classification tasks. In visual networks where saliency is more well-studied, saliency is naturally localized through the convolutional layers of the network; however, the same is not true in modern transformer-stack networks used to process natural language. We adapt gradient-based saliency methods for these networks, propose a method for evaluating the degree of semantic coherence of each layer, and demonstrate consistent improvement over numerous other methods for textual saliency on multiple benchmark classification datasets. Our approach requires no additional training or access to labelled data, and is comparatively very computationally efficient.
APA
Hou, E.M. & Castanon, G.D.. (2023). Decoding Layer Saliency in Language Transformers. Proceedings of the 40th International Conference on Machine Learning, in Proceedings of Machine Learning Research 202:13285-13308 Available from https://proceedings.mlr.press/v202/hou23a.html.

Related Material