Duplicate Question Detection based on Neural Networks and Multi-head Attention
Heng Zhang, Liangyu Chen · 2019
It is well known that using only one neural network can not get a satisfied accuracy for the problem of Duplicate Question Detection. In order to break through this dilemma, different neural networks are ensembled serially to strive for better accuracy. However, many problems, such as vanishing gradient or exploding gradient, will be encountered if the depth of neural network is blindly increased. Worse, the serial integration may be poor in computational performance since it is less parallelizable and needs more time to train. To solve these problems, we use ensemble learning with treating different neural networks as individual learners, calculating in parallel, and proposing a new voting mechanism to get better detection accuracy. In addition to the classical models based on recurrent or convolutional neural network, Multi-Head Attention is also integrated to reduce the correlation and the performance gap between different models. The experimental results in Quora question pairs dataset show that the accuracy of our method can reach 89.3%.