Using neural networks in statistical parsing

Abstract:

Statistical parsing is the best tool we have for processing natural language, with applications in fields from machine translation to data mining. Unfortunately, performance to date has been hampered by the scarcity of training data. In this talk, I will describe a new method of encoding the training data. A neural network is trained to predict the training data, and the resulting neural network generalises better than previous approaches. This new method removes restrictions on the complexity of the probability model, and so will allow linguists to explore more complex models. Additionally, the approach is not specific to statistical parsing, and can be usefully applied in other situations where there is a complex probability model and limited training data.

Parking: Visitors requiring a parking pass are asked to contact us at least one working day before the seminar.

Enquiries: sals@ics.mq.edu.au