Pierre-Antoine Lequeu

The paper proposes explicitly disentangling positional and semantic representations in Transformer encoders, demonstrating that this separation allows for a clearer understanding of how positional information is processed and significantly improves the retention of macroscopic document structure.