Dual-Channel Reasoning Model for Complex Question Answering

Table 1

The performance of our model and competing approaches on the HotpotQA dataset.

Model	Answer		Sup fact		Joint
	EM	F1	EM	F1	EM	F1
Baseline	45.60	59.02	20.32	64.49	10.83	40.16
NMN	49.58	62.71	—	—	—	—
KGNN	50.81	65.75	38.74	76.69	22.40	52.82
BERT Plus	55.84	69.76	42.88	80.74	27.13	58.23
DFGN	56.31	69.69	51.50	81.62	33.62	59.82
DFGN	55.19	68.68	49.72	80.67	31.53	58.26
DFGN/BERT	55.17	68.49	49.85	81.06	31.87	58.23
Our model
Baseline-Dual	49.56	64.15	47.61	83.45	26.44	55.41
CGDe-Baseline	51.23	65.42	46.71	83.03	27.69	55.77
FGIn-Baseline	50.42	64.82	48.93	84.10	27.95	56.47
DFGN_Dual	55.42	68.90	50.70	81.70	31.76	58.83