MIT Press

Table 7

Experimental results for the fine-tuned LongFormer model on the development set depending on the context window size and the label weight.

Window size (nb tokens)	Label weight (MASK, NO_MASK)	R_{di +qi}	ER_di	ER_qi	P_{di +qi}	WP_{di +qi}
32	(1,1)	0.875	0.987	0.850	0.612	0.643
32	(5,1)	0.924	0.992	0.915	0.550	0.583
32	(10,1)	0.954	0.994	0.952	0.470	0.507

512	(1,1)	0.847	0.986	0.817	0.929	0.932
512	(5,1)	0.914	0.993	0.906	0.858	0.866
512	(10,1)	0.937	0.997	0.930	0.767	0.783

4,096	(1,1)	0.860	0.988	0.847	0.925	0.929
4,096	(5,1)	0.916	0.988	0.913	0.843	0.856
4,096	(10,1)	0.935	0.993	0.936	0.795	0.811

Window size (nb tokens)	Label weight (MASK, NO_MASK)	R_{di +qi}	ER_di	ER_qi	P_{di +qi}	WP_{di +qi}
32	(1,1)	0.875	0.987	0.850	0.612	0.643
32	(5,1)	0.924	0.992	0.915	0.550	0.583
32	(10,1)	0.954	0.994	0.952	0.470	0.507

512	(1,1)	0.847	0.986	0.817	0.929	0.932
512	(5,1)	0.914	0.993	0.906	0.858	0.866
512	(10,1)	0.937	0.997	0.930	0.767	0.783

4,096	(1,1)	0.860	0.988	0.847	0.925	0.929
4,096	(5,1)	0.916	0.988	0.913	0.843	0.856
4,096	(10,1)	0.935	0.993	0.936	0.795	0.811