Inrementl Dependeny Prsing Mihel Fell 9 June 2011 1
Overview - Inrementl Dependeny Prsing - two lgorithms - evlution - enerl ritiism on present pprohes - possile improvements - ummry 2
Dependeny Prsing The mn loves ke. (entene) Det uj Oj The mn loves ke. (Dependeny rph) 3
Dependeny rph Leled, direted grph (W, A) - W: words in the sentene - A: dependeny reltion etween words 4
Dependeny rph Leled, direted grph (W, A) - W: words in the sentene - A: dependeny reltion etween words Well-formedness riteri: - onneted - yli - unique lel - single hed - projetive 5
Inrementl Dependeny Prsing Dependeny prsing - is roust nd performs well - omits phrsl nodes Wht out doing it inrementlly? 6
Inrementl Dependeny Prsing Dependeny prsing - is roust nd performs well - omits phrsl nodes Wht out doing it inrementlly? One possiility: Left-to-right ottom-up dependeny prsing 7
Bottom-up Dependeny Prsing stk input dep. grph 8
Bottom-up Dependeny Prsing hift: () stk input dep. grph 9
Bottom-up Dependeny Prsing hift: () stk input dep. grph Left-Redue (LR): 10
Bottom-up Dependeny Prsing hift: () stk input dep. grph Right-Redue (RR): Left-Redue (LR): 11
Bottom-up Dependeny Prsing Exmple derivtion of 12
Bottom-up Dependeny Prsing Exmple derivtion of (1) 13
Bottom-up Dependeny Prsing Exmple derivtion of (1) (2) 14
Bottom-up Dependeny Prsing Exmple derivtion of (1) (2) (3) LR 15
Bottom-up Dependeny Prsing Exmple derivtion of (1) (2) (3) LR (4) - 16
Bottom-up dependeny prsing Exmple derivtion of (1) (2) (3) LR (4) (5) - RR - 17
Bottom-up dependeny prsing Dependeny grphs with 3 nodes: We hve derived (4). (2), (3) nd (5) n lso e derived. 18
Bottom-up dependeny prsing Dependeny grphs with 3 nodes: We hve derived (4). (2), (3) nd (5) n lso e derived. (1) nd (6), (7) n t e derived 19
Bottom-up Dependeny Prsing Dependeny grphs with 3 nodes: We hve derived (4). (2), (3) nd (5) n lso e derived. (1) nd (6), (7) n t e derived (1): is omined vi Right-Redution hs hed ersed from stk 20
Bottom-up Dependeny Prsing Dependeny grphs with 3 nodes: We hve derived (4). (2), (3) nd (5) n lso e derived. (1) nd (6), (7) n t e derived (1): is omined with vi Right-Redution hs hed ersed from stk (6), (7): no onneting r etween nd To onnet them, we needed to put onto the stk, too. (hene lose inrementlity) 21
Bottom-up Dependeny Prsing Is there wy to prse (1) nd (6), (7) inrementlly? (6), (7): no!
Bottom-up Dependeny Prsing Is there wy to prse (1) nd (6), (7) inrementlly? (6), (7): no! (1): yes, red input from right to left inrementlity?
(1) n e proessed inrementlly Bottom-up Dependeny Prsing Is there wy to prse (1) nd (6), (7) inrementlly? (6), (7): no! (1): yes, red input from right to left inrementlity?
Inrementl Dependeny Prsing Bottom-up nd Top-down in Dependeny Prsing BU: D H x * D H x Dependent D is tthed to its hed H efore H is tthed to its hed 25
Inrementl Dependeny Prsing Bottom-up nd Top-down in Dependeny Prsing BU: D H x * D H x Dependent D is tthed to its hed H efore H is tthed to its hed TD: * H D x H D x Hed H is tthed to dependent D efore D is tthed to its dependent(s) 26
Inrementl Dependeny Prsing Bottom-up nd Top-down in Dependeny Prsing BU: D H x * D H x Dependent D is tthed to its hed H efore H is tthed to its hed TD: * H D x H D x Hed H is tthed to dependent D efore D is tthed to its dependent(s) Insight: We n proess left-dependents inrementlly vi BU proess right-dependents inrementlly vi TD prsing 27
Inrementl Dependeny Prsing Bottom-up nd Top-down in Dependeny Prsing BU: D H x * D H x Dependent D is tthed to its hed H efore H is tthed to its hed TD: * H D x H D x Hed H is tthed to dependent D efore D is tthed to its dependent(s) Insight: We n proess left-dependents inrementlly vi BU proess right-dependents inrementlly vi TD prsing Ar-Eger Dependeny Prsing 28
Ar-Eger Dependeny Prsing hift: () stk input dep. grph 29
Ar-Eger Dependeny Prsing hift: () stk input dep. grph Left-Ar (LA): 30
Ar-Eger Dependeny Prsing hift: () stk input dep. grph Right-Ar (RA): Left-Ar (LA): 31
Ar-Eger Dependeny Prsing hift: () stk input dep. grph Right-Ar (RA): Left-Ar (LA): Redue (R): 32
Bottom-up vs. Ar-Eger stk input dep. grph hift: ( BU ) hift: ( AE ) 33
Bottom-up vs. Ar-Eger stk input dep. grph hift: ( BU ) hift: ( AE ) eft-redue (LR): Left-Ar (LA): LR LA 34
Bottom-up vs. Ar-Eger hift: ( BU ) stk input dep. grph hift: ( AE ) eft-redue (LR): Left-Ar (LA): LR LA 35
Bottom-up vs. Ar-Eger hift: ( BU ) stk input dep. grph hift: ( AE ) eft-redue (LR): Left-Ar (LA): LR LA 36
Bottom-up vs. Ar-Eger Right-Redue (RR): Right-Ar (RA): RR RA 37
Bottom-up vs. Ar-Eger Right-Redue (RR): Right-Ar (RA): RR RA 38
Bottom-up vs. Ar-Eger Right-Redue (RR): Right-Ar (RA): RR RA R 39
Bottom-up vs. Ar-Eger Right-Redue (RR): Right-Ar (RA): RR RA R Ar-Eger Dependeny Prsing n fully simulte Bottom-up Dependeny Prsing! 40
Bottom-up vs. Ar-Eger Right-Redue (RR): Right-Ar (RA): RR RA R Ar-Eger Dependeny Prsing n fully simulte Bottom-up Dependeny Prsing We n lso derive new grphs with AE! (see next slide) 41
Ar-Eger Dependeny Prsing (1) is not derivle with BU prsing, ut it is with AE: 42
Ar-Eger Dependeny Prsing (1) is not derivle with BU prsing, ut it is with AE: 43
Ar-Eger Dependeny Prsing (1) is not derivle with BU prsing, ut it is with AE: RA 44
Ar-Eger Dependeny Prsing (1) is not derivle with BU prsing, ut it is with AE: RA RA - 45
Ar-Eger Prsing: Evlution - smll wedish treenk (5685 sentenes) - evluting inrementlity: numer of onneted omponents on stk during prse ( 1 mens stritly inrementl) 46
Ar-Eger Prsing: Evlution - smll wedish treenk (5685 sentenes) - evluting inrementlity: numer of onneted omponents on stk during prse ( 1 mens stritly inrementl) stritly inrementl 47
Ar-Eger Prsing: Evlution - smll wedish treenk (5685 sentenes) - evluting inrementlity: numer of onneted omponents on stk during prse ( 1 mens stritly inrementl) stritly inrementl mildly inrementl 48
Intermedite ummry - Dependeny prsing works well in prtie - Inrementl dependeny prsing possile in mny ses - Improving the prsing tehnique is essentil - Ar-Eger performs etter thn Bottom-up dep. prsing - Well-formed prsing results show high inrementlity 49
Intermedite ummry - Dependeny prsing works well in prtie - Inrementl dependeny prsing possile in mny ses - Improving the prsing tehnique is essentil - Ar-Eger performs etter thn Bottom-up dep. prsing - Well-formed prsing results show high inrementlity - ut, wht out those strutures (6) nd (7) we ouldn t prse inrementlly? 50
Roust Inrementlity 51
Roust Inrementlity Drwks of storing omponents on stk - psyholinguisti plusiility: why not integrte diretly? 52
Roust Inrementlity Drwks of storing omponents on stk - psyholinguisti plusiility: why not integrte diretly? - prtility: dely of output s stored omponents re not prt of it 53
Roust Inrementlity - Argument Dependeny Model - dependenies etween ver s rguments - proto roles (proto-gent, proto-ptient) - e.g.: noun(nimte & nomintive) noun(proto-gent) dependeny rel. UBJ governs the noun (phrse) unless ontrditory onstrints override this 54
Roust Inrementlity NONPEC node - onnet strutures to NONPEC node while ver hs not een found - NONPEC n hnge into ny other node nd even divide into severl nodes - My even e in the resulting grph
Roust Inrementlity: Evlution - orpus with - uniform sentene pttern - ver-finl suluses 97.3% urte dependeny grphs, ut 56
ummry - Inrementl Dependeny Prsing is possile nd effiient - Ver-end strutures pose prolems to strit inrementlity - Pseudo-strit inrementlity with strt NONPEC node suggested - Integrtes dep. reltions on-the-fly - still seems lot like renmed stk to me (whih n e output) too vgue 57
Thnk you! 58
Referenes Jokim Nivre (2004). Inrementlity in Deterministi Dependeny Prsing Wolfgng Menzel (2009). Towrds rdilly inrementl prsing of nturl lnguge 59