Vine Parsing and Minimum Risk Reranking for Speed and Precision - PowerPoint PPT Presentation

Vine Parsing and Minimum Risk Reranking for Speed and Precision Markus Dreyer David A. Smith Noah A. Smith Johns Hopkins University

Design Goals Speed Precision CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking

E L M d e h t o o m S U -best unlabeled Unlabeled Parser parses E L U × L -best labeled M d e h t o o m S Labeler parses r a e n i l g - o L Reranker Collins (2000) Charniak and Johnson (2005) CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking

Unlabeled Parser Split-Head assumption Projectivity Vine Grammar Eisner and N. Smith (2005) CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking

CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking

CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking

CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking

CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking

CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking

2 + nB right r o f s d 2 n u o b O ( nB left ) t n e r e f f D i right n e r d l h i c t e f l d n a f f o e d a r T y c a r u c c A - d e e p S : s d n u o b e s o o h C d e h c u o t n u s e i c n e d n e p e d a l n i g i r o f o % 0 9 CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking

Reranker Minimum Risk Training Deterministic annealing D. Smith and Eisner (2006) CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking

M e d i a n 6 7 . 6 Labeled Japanese (82.9), Portuguese (75.3), Bulgarian (74.8), Chinese (71.6), German (71.0) M e d i a n 7 7 . 5 Unlabeled Japanese (86.0), Portuguese (82.4), Bulgarian (82.0), Swedish (79.5), Chinese (77.6) CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking

Summary Unlabeled Labeler Reranker Parser  Parsing constraints  Linear-time inference and decoding  Minimum Risk reranking Future Work  High precision ,  Better estimation  Better labeler (label bigrams) mediocre recall  More fine-grained parsing constraints (length bounds given head) CoNLL-2006  M. Dreyer, D. A. Smith, N. A. Smith  Vine Parsing and Minimum Risk Reranking