A question about the structure smoothing

[speedcell4]

Hi, thanks for sharing your source code.
According to Table 2, your model on the w/o. Potential Normalization and Structure Smoothing setting performs worse than Pyramid (BERT) (Jue et al. 2020), i.e., 85.7, 84.4, 77.3 vs 86.3, 84.7, 79.2 on the three datasets respectively.
Thus, does this mean what contributes to performance improvement is Structure Smoothing? And potential normalization is used to ensure numerical stability. The PO-TreeCRF itself actually contributes little to improvement?