Natural language processing systems have made impressive strides in producing useful task-oriented language tools - but even the most sophisticated NLP systems fall dramatically short of the human brain in robustness and nuance of language understanding. A particular area of need in NLP is that of sentence composition: the combinatory capacity that allows humans to generate and understand the meanings of infinite sentences based on their parts. 

 

This talk will discuss two threads of work tapping into insights from brain science to address the gap between humans and NLP systems. The first of these threads draws on analysis techniques from cognitive neuroscience in order to assess the strengths and limitations of otherwise opaque NLP systems in composing sentence representations. The second thread implements computational models based on brain data from human language comprehension experiments, in order to gain insight into the mechanisms underlying sentence understanding by humans.