Skimformer

A collaboration between reciTAL & MLIA (ISIR, Sorbonne Université)

Model description

Skimformer is a two-stage Transformer that replaces self-attention with Skim-Attention, a self-attention module that computes attention solely based on the 2D positions of tokens in the page. The model adopts a two-step approach: first, the skim-attention scores are computed once and only once using layout information alone; then, these attentions are used in every layer of a text-based Transformer encoder. For more details, please refer to our paper:

Skim-Attention: Learning to Focus via Document Layout Laura Nguyen, Thomas Scialom, Jacopo Staiano, Benjamin Piwowarski, EMNLP 2021

Citation

@article{nguyen2021skimattention,
    title={Skim-Attention: Learning to Focus via Document Layout}, 
    author={Laura Nguyen and Thomas Scialom and Jacopo Staiano and Benjamin Piwowarski},
    journal={arXiv preprint arXiv:2109.01078}
    year={2021},
}