Panelist: Hall Daume III
Title: Reading between the lines

Seven years ago, Ellen Riloff, Amit Goyal and I decided to see how far at-that-time modern NLP techniques could get us on automatically identifying the structure of narrative stories, revisiting Lehnert's Plot Units framework from the early 1980s. While we could do okay-ish on affect states (happy/sad) that were explicitly stated in the narrative, ones that required inference were largely beyond our grasp. Worse, we were virtually unable to identify the formation of goals, much less identify their resolutions. This all points to a larger challenge for NLP: information extraction of stuff that's not explicitly stated in the text, but rather left to inference. A full solution to this problem is likely still well beyond our abilities, but I'll discuss some smaller subtasks we've made progress on and discuss technology (mostly neural network-like frameworks that can combine multiple sources of information) that may help us move forward.