Parsing and reinforcement learning
Supervisor
Suitable for
Abstract
Incremental dependency parsers such those described inhttp://www.mitpressjournals.org/doi/pdf/10.1162/coli.07-056-R1-07-027
typically try to predict what parsing action to take next by training a classifier which will look ahead in the input, and at the current parse state, and make a choice between actions. In many current non-linguistic applications, however, this kind of "what action do I perform next" question is answered by using reinforcement learning, where the system learns a reward function from training data that should bias towards that action most likely to lead to a successful conclusion.
This project aims to experiment to see whether a reinforcement learning decision component could lead to better parsing performance than the more usual classifier-based decisions.