Pros

• some initial progress
• semantically interesting topic (people like movies)

Cons

• Graph sparsity

Pros

• the data is there

Cons

• too small: 20k questions
• IR-heavy task where simple algos can do very well
• small splash

Pros

• 200k tables to generate questions from

Cons

• IR-heavy
• many tables may not be interesting
• small splash

Pros

• dense graph
• open domain
• would make a big splash

Cons

• encoding a subset of the KB as a graph for traversal/generation would take significant effort
• IR-heavy: scaling to massive kb's requires significant algorithmic innovation, might detract from reasoning

Pros

• encoding is easy (not as much of an IR focus)
• would make a big splash ("ai solves gre!")
• encoding text is more fun than encoding graphs/KBs (nlp ftw!!)

Cons

• where to get stories? reddit? generate them? have turkers write them? use other short story datasets?
• how to get questions? extract graph from story? turkers (how to encourage compositionality)?

1. you start from squad paragraph, or roc stories, or a photo (we can try out several options).
2. you extract a small structured representation out of that. (graph) - using corenlp tools.
3. two options: a. somehow (not sure how yet) you prime the turkers to ask compositional questions, b. you use the small graph to create auto-generated compositional questions, as we planned to do with wikimovies

Pros

• lots of small graphs

Cons

• lots of unknowns: what works best?
• predicted graphs
• how to prime turkers for compositionality