Panelist: Christopher Manning

A (maybe not yet) unified theory of inference for text understanding

In the 1980s, people in NLU had lofty goals but modest realities;
today we have much better realities, but unfortunately are mainly
failing to enunciate lofty goals. If we contrast the textual inference
problems focused on 30 years ago with where we are today, on some of
the problems we have made real progress, whereas others, such as
elaboration (connecting predicates via reasons or causes) have tended
to be neglected. We are also at an interesting scientifc juncture: To
what extent do we need explicit language and knowledge representations
versus finding that the distributed representations of a bidirectional
LSTM with attention seem to be able to do everything better? Whichever
way that plays out, I think there are still clear things we need to
work more on: less primitive memories and knowledge, developing goals
and plans, understanding inter-sentential relationships, and doing
elaborations from a situation using common sense knowledge.