Reranking Parse Trees with a SRL system Charles Sutton and Andrew McCallum University of Massachusetts June 30, 2005
Motivation for Joint Processing VERB: barked ARG0: the dog ARG1: the man AM-LOC: TV Miller (2000): single probabilistic model for parsing, relations Why not for SRL? Rather than augmenting grammar, we use reranking approach Uncertainty Long-range features
Reranking by weighted combination Basic MaxEnt SRL model Rerank n-best parse trees from Bikel’s implementation of Collins parser Gildea & Jurafsky tried this with =0.5
Weighted combination Choosing tree by SRL score has lower recall
Training a global reranker MaxEnt reranking of parse tree list, using features from base SRL frame Inspired by Toutanova et al. 2005, but different. Features include: Standard local SRL features Does arg Ax occur in frame? Linear sequence of frame arugments (e.g. A0_V_A1) Conjunctions of heads from argument pairs Parse tree score Parse TreesSRL F1 Gold best 63.9 Trained combination 63.6 Simple combination ( =0.5) 56.9
Ceiling performance Parse TreesSRL F1 Gold best 63.9 Reranked by gold parse F Reranked by gold frame F Best SRL performance of parse-tree reranking system. Parse F1: 95.0
Discussion / Future Work Reranking parse trees has strong limitations Marginalizing over parse trees Should SRL / parsing be joint at all? –If not, how different than MUC relation extraction? (Miller et al., 2000)