Phase-specific wordtopics arediscoveredthathelpexplain theconceptual patterns conditioned on the same phase.

Extensive experiments onopen-ended textgeneration (Wikitext-103) andtextsummarization (CNN/DailyMail) demonstrate the generality and effectiveness of our method.

Deep learning models achievetremendous successes across amultitude oftasks, yettheir training and inference often yield high computational costs and long processing times [11,22].