Using Deconvolutional Variational Autoencoder for Answer Selection in Community Question Answering

Answer selection in community question answering is a challenging task in natural language processing. The main problem is that there is no evaluation for the answers given by the users and one should go through all possible answers for assessing them, which is exhausting and time consuming. In this paper we propose a latent-variable model for learning the representations of the question and answer, by jointly optimizing generative and discriminative objectives. This model uses variational autoencoders (VAE) in a multi-task learning process with a classifier to produces a representation for each answer by which the classifier could classify it's relation with correspond question with a high performance. The experimental results on two public datasets, SemEval 2015 and SemEval 2017, recognize the significance of the proposed framework, especially for the semi-supervised setting. The results showed that the proposed model outperformed F1 of state-of-the-art method up to about 8% for SemEval 2015 and about 5% for SemEva1 2017.