Marasović 2018 – NLP’s generalization problem, and how researchers are tackling it
Notes for MarasovicGradient2018NLP
1. questions
- How should we measure how well our models perform on unseen inputs. It's not enough to test on the same distribution as train.
- How should we modify our models
2. direction 1: inductive bias
- what priors on structure should we build into our models?
- models fail, because they:
- learn language passively
- do not learn anything about the underlying world that language is used to describe
2.1. potential solutions
- use RL to directly optimize METEOR, CIDEr, etc.
- use human in the loop training
3. direction 2: common sense
- models lack social and phyiscal common sense
4. direction 3: generalizing to unseen distributions and tasks
- we are concerned with how well a model is able to extrapolate
- so, trained on one task, how well is the model able to perform another unrelated task
- this is important, because we will never have enough annotated data for all the tasks in the world