Learning Document Semantic Representation with Hybrid Deep Belief Network

Table 1

Document classification accuracy and document retrieval precision.

Dataset	Model	L_r	Document classification					Document retrieval
			Output units number					Output units number
			50	100	128	512	1000	50	100	128	512	1000
20 Newsgroups	RSM	0.01	63.11	66.82	67.67	66.52	69.31	62.82	67.80	68.64	65.10	67.48
		0.001	60.94	62.91	71.55	73.68	73.05	56.10	58.32	69.17	69.65	69.34
		0.0001	36.59	62.75	63.33	70.40	71.38	28.46	58.56	59.39	66.37	65.14
	DBN	0.01	72.77	72.35	72.49	72.56	72.39	67.52	67.78	68.90	69.27	69.81
	DocNADE	0.01	73.83	73.40	73.54	74.75	74.11	68.50	68.78	70.17	70.66	70.82
	HDBN	0.01	76.57	76.68	76.35	76.49	76.89	76.91	71.89	72.99	73.50	73.91
BBC data	RSM	0.01	94.04	94.65	95.41	95.11	95.87	93.44	93.29	92.98	93.90	94.81
		0.001	95.11	96.48	96.64	94.65	96.64	94.66	95.34	96.18	95.79	96.03
		0.0001	60.40	94.65	94.80	96.79	94.95	70.88	91.76	94.05	95.88	95.42
	DBN	0.01	95.31	94.81	96.03	95.06	95.05	93.18	94.10	94.30	94.13	94.69
	DocNADE	0.01	95.60	96.77	96.93	96.89	96.93	94.94	95.44	96.28	95.98	96.13
	HDBN	0.01	97.40	97.40	98.01	97.09	97.25	95.42	96.02	96.94	96.48	96.94