Conditional Video-Text Reconstruction Network with Cauchy Mask for Weakly Supervised Temporal Sentence Grounding.