Machine Reading, Fast and Slow - When Do Models Understand Language?

Explore a thought-provoking lecture on the complexities of machine reading comprehension and the challenges in developing AI models that truly "understand" language. Delve into behavioral benchmarks, reading comprehension assessments, and the limitations of current transformer models. Examine why scale alone doesn't solve comprehension issues and how pre-training knowledge is underutilized. Investigate the lack of shortcuts in language processing, data-related problems, and testing methodologies. Conclude by considering open problems in the field of machine reading and language understanding.

Intro
Behavioral Benchmarks
Reading Comprehension
How is it going
Scale alone doesnt solve it
Transformers dont use pretraining knowledge
Lack of shortcuts
Data working problem
Testing
Open Problems