We present state-of-the-art results for semantic proto-role labeling (SPRL) using an adapted bidirectional LSTM encoder. Predicate-argument structure is represented as pairs of hidden states corresponding to predicate and argument head tokens. The network naturally shares parameters between SPR properties which we show to be helpful in learning additional properties from limited data. An investigation into pretraining or concurrent training of related tasks yields a further absolute gain of over 1 F1 point. We contrast the qualitative behavior of our neural model to the CRF of prior work.
View on arXiv