Connecting What to Say With Where to Look by Modeling Human Attention Traces.