In this paper, we present results from a deep neural network event classifier that uses lexical semantic features derived from parameters that are underspecified in the event typing. These results demonstrate that the presence or absence of an underspecified feature is a strong predictor of event class, and we propose a model for extending this approach to action recognition (i.e., the recognition of processes enacted by an agent) by using reinforcement learning to learn complex actions from object motions, and then 'factoring out" the specifics of the object to recognize an action denoted by an agent motion, such as a gesture, alone.
@InProceedings{KRISHNASWAMY18.6, author = {Nikhil Krishnaswamy ,Tuan Do and James Pustejovsky}, title = {Learning Actions from Events Using Agent Motions}, booktitle = {Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)}, year = {2018}, month = {may}, date = {7-12}, location = {Miyazaki, Japan}, editor = {James Pustejovsky and Ielka van der Sluis}, publisher = {European Language Resources Association (ELRA)}, address = {Paris, France}, isbn = {979-10-95546-06-1 }, language = {english} }
Summary of the paper