A Stacked BiLSTM Neural Network Based on Coattention Mechanism for Question Answering

Table 3

Experimental results of different baselines and our proposed model on Train-All data.

Idx	Model	MAP	MRR
1	Probabilistic quasi-synchronous grammar [35]	0.6029	0.6852
2	Tree edit models [2]	0.6091	0.6917
3	Linear-chain CRF [17]	0.6307	0.7477
4	LCLR [18]	0.7092	0.7700
5	Bigram + count [38]	0.7113	0.7846
6	Three-layer BiLSTM + BM25 [6]	0.7134	0.7913
7	Convolutional deep neural networks [39]	0.7459	0.8078
8	BiLSTM/CNN with attention [7]	0.7111	0.8322
9	Attentive LSTM [1]	0.7530	0.8300
10	BiLSTM encoder-decoder with step attention [8]	0.7261	0.8018
11	BiLSTM	0.6982	0.7764
12	Stacked BiLSTM	0.7127	0.7893
13	BiLSTM with coattention	0.7325	0.7962
14	Stacked BiLSTM with coattention	0.7451	0.8114
15	Stacked BiLSTM with coattention (cosine + Euclidean)	0.7613	0.8401