fast er exact decoding and global training for transition
play

Fast(er) Exact Decoding and Global Training for Transition-Based - PowerPoint PPT Presentation

Fast(er) Exact Decoding and Global Training for Transition-Based Dependency Parsing via a Minimal Feature Set Tianze Shi* Liang Huang Lillian Lee* Oregon State University * Cornell University 3 Minimal 3 6


  1. Fast(er) Exact Decoding and Global Training for Transition-Based Dependency Parsing via a Minimal Feature Set Tianze Shi* Liang Huang† Lillian Lee* † Oregon State University * Cornell University 𝑃 π‘œ 3 Minimal 𝑃 π‘œ 3 𝑃 π‘œ 6 Feature Set Theoretical Practical

  2. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Short Version β€’ Transition-based dependency parsing has an exponentially-large search space β€’ 𝑃 π‘œ 3 exact solutions exist πŸ˜„ β€’ In practice, however, we needed rich features ⟹ 𝑃 π‘œ 6 😟 β€’ (This work) with bi-LSTMs, now we can do 𝑃(π‘œ 3 ) ! πŸ˜„ β€’ And we get state-of-the-art results 2

  3. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Short Version β€’ Transition-based dependency parsing has an exponentially-large search space β€’ 𝑃 π‘œ 3 exact solutions exist πŸ˜„ β€’ In practice, however, we needed rich features ⟹ 𝑃 π‘œ 6 😟 β€’ (This work) with bi-LSTMs, now we can do 𝑃(π‘œ 3 ) ! πŸ˜„ β€’ And we get state-of-the-art results 3

  4. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Short Version β€’ Transition-based dependency parsing has an exponentially-large search space β€’ 𝑃 π‘œ 3 exact solutions exist πŸ˜„ β€’ In practice, however, we needed rich features ⟹ 𝑃 π‘œ 6 😟 β€’ (This work) with bi-LSTMs, now we can do 𝑃(π‘œ 3 ) ! πŸ˜„ β€’ And we get state-of-the-art results 4

  5. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Short Version β€’ Transition-based dependency parsing has an exponentially-large search space β€’ 𝑃 π‘œ 3 exact solutions exist πŸ˜„ β€’ In practice, however, we needed rich features ⟹ 𝑃 π‘œ 6 😟 β€’ (This work) with bi-LSTMs, now we can do 𝑃(π‘œ 3 ) ! πŸ˜„ β€’ And we get state-of-the-art results 5

  6. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Short Version β€’ Transition-based dependency parsing has an exponentially-large search space β€’ 𝑃 π‘œ 3 exact solutions exist πŸ˜„ β€’ In practice, however, we needed rich features ⟹ 𝑃 π‘œ 6 😟 β€’ (This work) with bi-LSTMs, now we can do 𝑃(π‘œ 3 ) ! πŸ˜„ β€’ And we get state-of-the-art results 6

  7. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Dependency Parsing root obj xcomp nsubj mark det OUTPUT She wanted to eat an apple INPUT 7

  8. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Transition-based Dependency Parsing … Terminal … states Initial state … … … 8

  9. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Transition-based Dependency Parsing Goal: max score( ) … … = max βˆ‘ score( ) Terminal … states Initial state … … … 9

  10. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Exact Decoding with Dynamic Programming Goal: max score( ) … Exponential to polynomial = max βˆ‘ score( ) … Terminal … states Initial state … … … (Huang and Sagae, 2010; Kuhlmann, 10 GΓ³mez-RodrΓ­guez and Satta, 2011)

  11. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Transition Systems DP Complexity # Action Types Arc-standard 3 𝑃 π‘œ 4 In our Arc-eager 4 𝑷 𝒐 πŸ’ paper Arc-hybrid 3 𝑷 𝒐 πŸ’ Presentational convenience 11

  12. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System State … … 𝑑 2 𝑑 1 𝑑 0 𝑐 0 𝑐 1 Stack Buffer Initial State ROOT She wanted … Terminal State (Yamada and Matsumoto, 2003) ROOT (GΓ³mez-RodrΓ­guez et al., 2008) (Kuhlmann et al., 2011) 12

  13. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Transitions … … … … … … 𝑑 1 𝑑 0 𝑑 0 𝑐 0 𝑐 0 shift reduce β†· reduce β†Ά … … … … 𝑑 1 … … 𝑐 0 𝑐 0 𝑑 0 𝑑 0 Same as arc-standard 13

  14. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Transitions … … … … … … 𝑑 1 𝑑 0 𝑑 0 𝑐 0 𝑐 0 shift reduce β†· reduce β†Ά … … … … 𝑑 1 … … 𝑐 0 𝑐 0 𝑑 0 𝑑 0 Same as arc-standard 14

  15. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Transitions … … … … … … 𝑑 1 𝑑 0 𝑑 0 𝑐 0 𝑐 0 shift reduce β†· reduce β†Ά … … … … 𝑑 1 … … 𝑐 0 𝑐 0 𝑑 0 𝑑 0 Same as arc-standard 15

  16. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Stack Buffer initial ROOT She wanted to eat an apple shift She wanted to eat an apple ROOT shift wanted to eat an apple ROOT She reduce β†Ά wanted to eat an apple ROOT shift She to eat an apple ROOT wanted shift eat an apple ROOT wanted to 16

  17. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Stack Buffer initial ROOT She wanted to eat an apple shift She wanted to eat an apple ROOT shift wanted to eat an apple ROOT She reduce β†Ά wanted to eat an apple ROOT shift She to eat an apple ROOT wanted shift eat an apple ROOT wanted to 17

  18. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Stack Buffer initial ROOT She wanted to eat an apple shift She wanted to eat an apple ROOT shift wanted to eat an apple ROOT She reduce β†Ά wanted to eat an apple ROOT shift She to eat an apple ROOT wanted shift eat an apple ROOT wanted to 18

  19. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Stack Buffer initial ROOT She wanted to eat an apple shift She wanted to eat an apple ROOT shift wanted to eat an apple ROOT She reduce β†Ά wanted to eat an apple ROOT shift She to eat an apple ROOT wanted shift eat an apple ROOT wanted to 19

  20. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Stack Buffer initial ROOT She wanted to eat an apple shift She wanted to eat an apple ROOT shift wanted to eat an apple ROOT She reduce β†Ά wanted to eat an apple ROOT shift She to eat an apple ROOT wanted shift eat an apple ROOT wanted to 20

  21. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Stack Buffer initial ROOT She wanted to eat an apple shift She wanted to eat an apple ROOT shift wanted to eat an apple ROOT She reduce β†Ά wanted to eat an apple ROOT shift She to eat an apple ROOT wanted shift eat an apple ROOT wanted to 21

  22. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Stack Buffer eat an apple ROOT wanted to reduce β†Ά eat an apple ROOT wanted shift to an apple ROOT wanted eat shift apple ROOT wanted eat an reduce β†Ά apple ROOT wanted eat shift an ROOT wanted eat apple 22

  23. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Stack Buffer eat an apple ROOT wanted to reduce β†Ά eat an apple ROOT wanted shift to an apple ROOT wanted eat shift apple ROOT wanted eat an reduce β†Ά apple ROOT wanted eat shift an ROOT wanted eat apple 23

  24. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Stack Buffer eat an apple ROOT wanted to reduce β†Ά eat an apple ROOT wanted shift to an apple ROOT wanted eat shift apple ROOT wanted eat an reduce β†Ά apple ROOT wanted eat shift an ROOT wanted eat apple 24

  25. Background 𝑃(π‘œ 3 ) in theory 𝑃(π‘œ 6 ) in practice Back to 𝑃(π‘œ 3 ) Results Arc-hybrid Transition System Stack Buffer eat an apple ROOT wanted to reduce β†Ά eat an apple ROOT wanted shift to an apple ROOT wanted eat shift apple ROOT wanted eat an reduce β†Ά apple ROOT wanted eat shift an ROOT wanted eat apple 25

Recommend


More recommend