What is a natural syntactic model for frame-semantic composition? - PowerPoint PPT Presentation

What is a natural syntactic model for frame-semantic composition? Timm Lichte, Laura Kallmeyer & Rainer Osswald University of Düsseldorf, Germany CTF14, August 26, 2014 SFB 991 1 / 26

Overview “natural” syntax counterpart for frames? properties of frames properties of grammars EDL vs. LDL (extended domain of locality) (limited domain of locality) EDL: case studies in LTAG (directed motion construction, secondary predicates) 2 / 26

What does natural mean? Sparse and transparent in terms of the syntax-semantics interface, and similar with respect to compositional aspects: syntax and semantics are homomorphic classical example: Montegovian semantics + Categorial Grammar λ y λ x . love ′ ( x , y ) V \ NP / NP Currying, functional application “ordered argument systems” (Dowty, 1989) frame semantics + ??? 3 / 26

Formal properties of frame semantics Frames are formalized as extended typed feature structures (Petersen, 2007; Kallmeyer & Osswald, 2013) no inherent ordering on the attributes of the same node no overt/explicit distinction between arguments and modifiers actor   1 locomotion mover  actor 1  locomotion 0     mover 0 1   manner   path path path path     manner walking walking 4 / 26

Formal properties of frame semantics Frames are formalized as extended typed feature structures (Petersen, 2007; Kallmeyer & Osswald, 2013) no inherent ordering on the attributes of the same node no overt/explicit distinction between arguments and modifiers Frames are composed by unification, not by functional application.   locomotion   locomotion � �   person    actor  actor 1 1 � �     person name John     mover ∪ = 0 1 1 0     name John     mover 1   path path       path path   manner walking   manner walking 5 / 26

Formal properties of grammars Fundamental distinction between two classes of grammar frameworks: limited domain of locality (LDL) extended domain of locality (EDL) Another recently discussed distinction that is othogonal: lexical vs. phrasal (Müller & Wechsler, 2014) 6 / 26

Formal properties of grammars: LDL LDL (limited domain of locality) predetermined derivational order (specified in the lexicon) indicator: valency lists, which are stepwise processed CG, (binarized) HPSG, SBCG, MG � ��� V subcat c h 2 NP � 2 �� � V subcat m h John AP � 2 �� � V subcat c h sometimes 1 PP � 2 , 1 �� � V subcat walks into the house 7 / 26

What are ordered valency lists good for? Implement the obliqueness hierarchy (Keenan & Comrie, 1977) subject ⇒ direct object ⇒ indirect objects of ⇒ obliques ⇒ genitives ⇒ object comparison List of applications (Müller, 2007, §3.1) binding theory passive ellipsis free relative clauses secondary predicates 8 / 26

Formal properties of grammars: EDL EDL (extended domain of locality) no predetermined derivational order capability to immediately access arbitrarily distant parts of a sentence within one lexical entry or syntactic rule LTAG, RRG, some versions of CxG, Dependency Grammar RRG: CLAUSE LTAG: S CORE NP VP RP NUC PP VP PP PRED V V walks walks 9 / 26

Formal properties of grammars: EDL EDL (extended domain of locality) no predetermined derivational order capability to immediately access arbitrarily distant parts of a sentence within one lexical entry or syntactic rule LTAG, RRG, some versions of CxG, Dependency Grammar CxG (Goldberg, 2013, 2014):   intransitive motion construction   Form: V { Subj, Oblique path }     | | |   Function: move agent path 10 / 26

LTAG: Introduction Ingredients: a set of elementary trees two combinatorial operations: substitution (replace a leaf node) adjunction (replace an inner node) S NP VP NP VP PP VP PP ADV John VP* into the house V sometimes walks EDL ⇒ the attachment order of the NP and the PP is independent! 11 / 26

LTAG and frames Kallmeyer & Osswald (2013): lexicon: pairs of elementary trees and base-labelled typed fea- ture structures Elementary trees are enriched with interface features , which contain base labels from the frame representation. unification of interface features � identification of base labels parallel composition of derived trees and larger frames S [ e = 0 ]   bounded-locomotion NP [ i = 1 ] VP [ e = 0 ] actor 1       mover 1   0 VP [ e = 0 ] PP [ i = 2 , e = 0 ]   goal 2     path path     V [ e = 0 ] manner walking walked 12 / 26

LTAG and frames: example (1) John walked into the house. S [ e = 0 ]   bounded-locomotion NP [ i = 1 ] VP [ e = 0 ] actor  1      mover 1   PP [ i = 2 , e = 0 ] VP [ e = 0 ] 0   goal 2     path path   V [ e = 0 ]   manner walking NP [ i = 3 ] walked N John � � person 3 name John 13 / 26

LTAG and frames: example (1) John walked into the house. S [ e = 0 ]   bounded-locomotion � �   NP [ i = 1 ] person VP [ e = 0 ]   actor 1   name John     PP [ i = 2 , e = 0 ] N VP [ e = 0 ] 0   mover 1     goal 2     V [ e = 0 ] John  path path    manner walking walked 14 / 26

LTAG and frames: example (1) John walked into the house. S [ e = 0 ]   bounded-locomotion � �   NP [ i = 1 ] person VP [ e = 0 ]   actor 1   name John     PP [ i = 2 , e = 0 ] N VP [ e = 0 ] 0   mover 1     goal 2     V [ e = 0 ] John  path path    manner walking walked   event PP [ i = 5 , e = 4 ] � �   4 path   path   endp P NP [ i = 5 ] v � � 5 in-region w into part-of ( v , w ) 15 / 26

LTAG and frames: example (1) John walked into the house.   bounded-locomotion S [ e = 0 ] � �   person   actor 1   name John   NP [ i = 1 ] VP [ e = 0 ]     mover 1     0 � �   N VP [ e = 0 ] PP [ i = 2 , e = 0 ] goal 2 in-region w       � � path   V [ e = 0 ] NP [ i = 2 ] John P  path    endp v     manner walking walked into part-of ( v , w ) NP [ i = 6 ] � � house Det N 6 in-region region the house 16 / 26

LTAG and frames: example (1) John walked into the house.   S [ e = 0 ] bounded-locomotion � �   person   actor 1   NP [ i = 1 ] VP [ e = 0 ] name John       mover 1   N VP [ e = 0 ] PP [ i = 2 , e = 0 ]   � �   0 house   goal 2   in-region  w  John V [ e = 0 ] P NP [ i = 2 ]     � � path    path  walked into Det N  endp  v     manner walking the house part-of ( v , w ) 17 / 26

LTAG and frames: factorization with metagrammars Lexical entries can be further decomposed/factorized using metagrammars (e. g. XMG, see the other talk!). class n0Vpp(dir) S [ e = 0 ] class n0V class DirPrepObj NP [ i = 1 ] VP [ e = 0 ] S [ E = 0 ] VP VP [ e = 0 ] PP [ i = 2 , e = 0 ] VP [ E = 0 ] NP [ I = 1 ] ≺ VP PP [ I = 1 , E = 0 ] V ⋄ [ e = 0 ] VP [ E = 0 ] � V ⋄ walked V ⋄ [ E = 0 ]     bounded-translocation bounded-locomotion goal 0  1   actor 1        path path mover � � 0 1 event     0 goal  2  actor 1   path path 18 / 26

Comming back to EDL vs. LDL They are different: representation of valency; order of derivation ⇒ EDL with set-like valency, LDL with list-like valency transparency of the syntax-semantics interface ⇒ EDL more transparent than LDL But are there fundamentally different ramifications? depictive secondary predicates ⇒ probably yes: see next slides. passive (probably no) binding theory ellipsis free relative clauses idioms, multi-word expressions 19 / 26

Depictive secondary predicates A case of cross modification : the modifier is disconnected from the modified phrase: (2) He i walked into the house naked i . What are the scope possibilities of depictives? EDL-analysis (LTAG, on the next slides): The depictive can ‘see’ the whole frame of the matrix sentence. But the valency status of frame components is not accessible! LDL-analysis (HPSG, Müller 2002; Müller 2008): The depictive only ‘sees’ the members of the valency list (in subcat ). non-cancellation approach: arguments are not removed during the derivation, but they remain there as “ghosts” 20 / 26