MIT Press

Table 2:

Question answering results. Accuracy averaged over the 20 bAbI tasks. Using tanh is worse than ReLU (line 13 vs. 15). RUM 150 λ = 0 without an update gate drops by 1.7% compared with line 13.

	Model	Acc.(%)
	Word Level
1	LSTM (Weston et al., 2016)	49.2
2	uRNN (ours)	51.6
3	EURNN (ours)	52.9
4	LSTM (ours)	56.0
5	GRU (ours)	58.2
6	GORU (Jing et al., 2017a)	60.4
7	RUM λ = 0 (ours)	73.2
8	DNC (Graves et al., 2016)	96.2
	Sentence Level
9	EUNN/attnEUNN (ours)	66.7/69.5
10	LSTM/attnLSTM (ours)	67.2/80.1
11	GRU/attnGRU (ours)	70.4/77.3
12	GORU/attnGORU (ours)	71.3/76.4
13	RUM/attnRUM λ = 0 (ours)	75.1/74.3
14	RUM/attnRUM λ = 1 (ours)	79.0/80.1
15	RUM/attnRUM λ = 0 w/ tanh (ours)	70.5/72.9
16	MemN2N (Sukhbaatar et al., 2015)	95.8
17	GMemN2N (Perez and Liu, 2017)	96.3
18	DMN+ (Xiong et al., 2016)	97.2
19	EntNet (Henaff et al., 2017)	99.5
20	QRN (Seo et al., 2017)	99.7

	Model	Acc.(%)
	Word Level
1	LSTM (Weston et al., 2016)	49.2
2	uRNN (ours)	51.6
3	EURNN (ours)	52.9
4	LSTM (ours)	56.0
5	GRU (ours)	58.2
6	GORU (Jing et al., 2017a)	60.4
7	RUM λ = 0 (ours)	73.2
8	DNC (Graves et al., 2016)	96.2
	Sentence Level
9	EUNN/attnEUNN (ours)	66.7/69.5
10	LSTM/attnLSTM (ours)	67.2/80.1
11	GRU/attnGRU (ours)	70.4/77.3
12	GORU/attnGORU (ours)	71.3/76.4
13	RUM/attnRUM λ = 0 (ours)	75.1/74.3
14	RUM/attnRUM λ = 1 (ours)	79.0/80.1
15	RUM/attnRUM λ = 0 w/ tanh (ours)	70.5/72.9
16	MemN2N (Sukhbaatar et al., 2015)	95.8
17	GMemN2N (Perez and Liu, 2017)	96.3
18	DMN+ (Xiong et al., 2016)	97.2
19	EntNet (Henaff et al., 2017)	99.5
20	QRN (Seo et al., 2017)	99.7

View Large

This Feature Is Available To Subscribers Only