Adaptive Feature Abstraction for Translating Video to Text