Check if there's a way to threshold the LayoutLM output

[ljvmiranda921]

Right now the LayoutLM produces a lot of "others" predictions. There might be a better way to reduce that number and parameterize it later.