Deconvolutional Latent-Variable Model for Text Sequence Matching