1 / 100

Outline

Localizing Dependencies by Factoring out Recursion: Computational and Linguistic Perspectives Seth Kulick University of Pennsylvania skulick@linc.cis.upenn.edu UCLA, 2/15/02. Outline. The Research Goal - Factoring out Recursion The Problem – Reduced Constructions.

Download Presentation

Outline

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Localizing Dependencies by Factoring out Recursion: Computational and Linguistic PerspectivesSeth KulickUniversity of Pennsylvaniaskulick@linc.cis.upenn.edu UCLA, 2/15/02

  2. Outline • The Research Goal - Factoring out Recursion • The Problem – Reduced Constructions. • An approach based on what a “defective complement” means in the context of TAG • A different approach based on dominance relations • Further consequences of the first approach • Some Future Directions

  3. Tree Adjoining Grammar X X Y Y Y Y Y AUXILIARYTREE INITIALTREE DERIVEDTREE ELEMENTARY TREES

  4. Subject-to-Subject Raising in TAG - 1 IP I’ DP I’ VP I VP I V I’ Johni ti V’ to seems V DP like pizza John seems [ to like pizza] Assumption: Elementary tree is the extended projection of a verb (Frank, 2002)

  5. Subject-to-Subject Raising in TAG - 2 I’ VP I V seems IP DP Johni I’ VP I to ti like pizza

  6. Subject-to-Subject Raising in TAG - 3 I’ I’ VP AP I I V I’ V to I’ seems be certain John seems [ to be certain[ to like pizza] ]

  7. Subject-to-Subject Raising in TAG - 4 I’ IP AP VP I I DP I’ V I’ VP I Johni seems ti V’ to V to I’ V DP be certain like pizza John seems [ to be certain[ to like pizza] ]

  8. Some Properties of TAG (Kroch & Joshi, 1985, Frank 2002) • Building blocks of the grammar - small pieces of phrase structure, called “elementary trees”, combined together in a context-free derivation. • The substantive theory of syntax must be stated over the domain of the elementary trees. • Fundamental TAG Hypothesis: Every syntactic dependency is expressed locally within a single elementary tree. • Cannot create dependencies between elementary trees by transformations. Movement is limited to within the elementary trees. Inter-clausal dependencies follow from adjoining operation.

  9. Wh-Movement in TAG C’ IP C I’ DP do VP I you V C’ think CP DPj C’ IP C What DP I’ that VP I Johni ti V’ tj V likes What do you think that John likes

  10. Some Linguistic Consequences • The substantive theory of syntax must be stated over the domain of the elementary trees. (Minimalist mechanisms, categorial grammar, optimality theory, HPSG, etc.) • Example of what can’t be expressed: “X0 movement may not skip the next landing site” while “this constraint is not relevant” for NPs undergoing long scrambling” (Sabel, 1995) • TAG explanations of locality violations – the formal system does part of the work • TAG enforcement of locality used for psycholinguistics and language acquisition work.

  11. Some Formal/Computational Consequences • “Mildly context-sensitive” grammar formalism (Joshi 1985) • Greater than context-free, less than context-sensitive • Constant Growth property - strings of the language can’t grow “too fast” • Polynomial parsing (the recovery of the syntactic structure underlying a sentence) • Can generate cross-serial dependencies – a1a2a3b1b2b3 – needed for languages such as Dutch • Can “count” only to four - {anbncndn | n >=0} but not five {anbncndnen| n >=0} or higher

  12. Some Consequences – Practical • Implemented Systems • XTAG at Penn – A wide-coverage English parser • Also a parser for French, with work as well on Korean and Chinese • Integration of statistical and structural techniques for parsing – (Schabes 92, Srinivas 97, etc.) • Machine Translation – ( Egedi et al. 94, Han et al. 96) • Natural Language Generation • Descriptions mapped to a TAG elementary tree (Stone and Doran 97)

  13. The Question - 1 • How far can this approach be pushed? • TAG has very nice computational properties, but it can be shown that it’s not adequate for some constructions. • A family of TAG formalisms, with differing formal and linguistic properties – e.g., elementary structures are tree sets instead of trees, constraints on how such structures can combine are loosened, etc.

  14. The Question - 2 • As we look at the various problems for TAG, is there some pattern that emerges that allows us to devise a system that retains the pleasant linguistic and computational properties of TAG while still handling the problematic cases? • One response: Who cares about the computational properties? Just worry about the linguistic properties of the resulting system.

  15. Linguistic and Computational Properties The Connection - 1 • Answer 1: Nicer computational properties make them more suitable for “real” applications • True, but that’s not the focus of this talk • Answer 2: If human languages can be expressed within a system with limited formal power, that is an interesting fact that needs to be explained. • True (subjective), but that’s not the focus of this talk

  16. Linguistic and Computational Properties The Connection - 2 • Answer 3: Computational properties like generative power gives us a way to characterize the type of recursion that a formal system has. • What does “type of recursion” mean? The way in which clauses can compose during a derivation • In particular, generative power can express how well a system implements certain key linguistic intuitions. This helps lead to systems with desirable linguistic properties, particularly with regard to locality.

  17. Outline • The Research Goal - Factoring out Recursion • The Problem – Reduced Constructions. • An approach based on what a “defective complement” means in the context of TAG • A different approach based on dominance relations • Further consequences of the first approach • Some Future Directions

  18. Reduced Constructions in RomanceClitic Climbing (Spanish) Object clitic placement is clause-bound… (1a) Luis insistio en [ comerlas ] Luis insisted on eating them(1b) *Luis las insistio en [ comer ] … except for some (“trigger”) verbs (2a) Luis quiere [ comerlas ] Luis wants to eat them(2b) Luis las quiere [ comer ] Interleaving - A problem for TAG

  19. Reduced Constructions in GermanLong Scrambling Scrambling clause bound except for the trigger verbs (1a)...daß niemand [ das Fahrrad zu reparieren] versprochen hat ...that no one [ the bike to repair] promised has ...that no one has promised to repair the bike (1b)...daß das Fahrrad niemand [ zu reparieren] versprochen hat ...that the bike no one [ to repair] promised has ...that no one has promised to repair the bike

  20. Reduced Constructions in GermanA Center-Embedded Clause (1a)...daß niemand [ das Fahrrad zu reparieren] versprochen ...that no one [ the bike to repair] promised has hat X X daß X niemand X versprochen hat X das Fahrrad zu reparieren

  21. Reduced Constructions in GermanLong Scrambling – Why it’s a problem X X das Fahrrad (1b)...daß das Fahrrad niemand [ zu reparieren] versprochen...that the bike no one [ to repair] promised has hat X More than just the insertion of a recursive structure X daß X niemand X versprochen hat zu reparieren

  22. Outline • The Research Goal - Factoring out Recursion • The Problem – Reduced Constructions. • An approach based on what a “defective complement” means in the context of TAG • A different approach based on dominance relations • Further consequences of the first approach • Some Future Directions

  23. Is A Defective Complement Analysis Possible in TAG? • Defective Complement Analysis – trigger verbs take a complement “smaller” than those taken by other verbs (Strozer 77, Rosen 90, Moore 91, Bleam 94, Wurmbrand 98…) • This acounts for the locality constraints of Reduced Constructions since only the trigger verbs take defective complements • This cannot be done in TAG, and “defective” cannot even have the same meaning in TAG

  24. A Defective Complement Analysis? lo Luis quiere(Luis it wants) Luis quiere(Luis wants) ver(to see) verlo(to see it) This is exactly what cannot be done in TAG – Movement can only be specified on an elementary tree And can’t just make larger elementary trees

  25. What does “defective” mean in TAG? – 1Defective complements in TAG cannot be “defective” • soler – a trigger verb that is also a raising verb: (1a) Luis suele comerlas (Aissen & Perlmutter) Luis tends to buy them(2b) Luislas suele comer Luis tends to buy them • Under the TAG analysis of raising, Luis, las, and comer are all part of the same elementary tree, with Luis already in [Spec, IP] • So if Luis is already above comer in the comer tree, in what way can suele take a “defective” complement?

  26. What does “defective” mean in TAG? - 2 • Querer – a trigger verb that is also a bridge verb: (1a) Juan quiere mostrartelos Juan wants to show them to you(2b) Quete quiere mostrar Juan What wants to show-to-you Juan What does Juan want to show to you? • Under the TAG analysis of wh-movement, Que, mostrar, and te are all part of the same elementary tree, with Que already in [Spec, CP] • So if Que is already above mostrar in the mostrar tree, in what way can quiere take a “defective” complement

  27. What does “defective” mean in TAG? - 3 • versuchen – a trigger verb that is also a bridge verb: (1a) Dieses Buch hat den Kindern niemand [ zu geben] versucht this book has the children no one [ to give ] tried no one has tried to give this book to the children • Under the TAG analysis of wh-movement, Dieses Buch, den Kindern, and zu geben are all part of the same elementary tree, with Dieses Buch already in [Spec, CP] • So if Dieses Buch is already in [Spec, CP] in the geben tree, in what way can versucht take a “defective” complement

  28. Expanding the Scope of the Problem - 1 • Subject Raising and I-C movement:(1a) John seems to like pizza(1b) Whatj does John seem to like tj ? • Subject Raising and experiencer extraction: (1c) To whom does John seem to like pizza ? When the seem and like trees are combined, [Spec, CP] can come from either tree. Adjoining cannot handle this sort of interleaving

  29. Expanding the Scope of the Problem - 2 • Claim: These subject raising cases are another manifestation of the same structural problem as with the clitic climbing and long scrambling • This leads to an intuition of how tree composition should work that will also handle the reduced constructions

  30. Subject-to-Subject Raising in TAG I’ VP I V I’ seems John seems [ to like pizza] IP DP I’ VP I Johni ti V’ to V DP like pizza

  31. A Revised View of Raising - 1 CP VP whatj C’ I IP DP I’ VP I Johni V DP to like tj Whatjdoes John seem [ to like tj?] CP C' C IP I’ does V I’ seem

  32. A Revised View of Raising - 2 CP VP whatj C’ I IP IP I’ DP I’ VP I Johni V I’ V DP to seem like tj Whati does John seems [ to like ti?] CP C' C does

  33. A Revised View of Raising - 3 VP I IP DP Johni to CP C' whatj C I’ does I’ V VP I seem V DP tj like

  34. A Revised View of Raising - 4 CP raising verb clause I’ CP CP I’ I’ I’ I’ result insertion

  35. What a “Defective Complement” means in TAG • In the TAG context, a defective complement has nothing to do with how high a complement projects. The analog of a defective complement analysis is that the recursive structure from the higher clause is inserted lower than in other cases • All inter-clausal dependencies are the same, aside from details: • Subject raising – recursion at I’ • Wh-movement – recursion at C’ • Reduced Constructions – recursion between I’ and C’

  36. Long Scrambling Revisited - 1 CP C' C IP niemand I’ hat IP versucht Dieses Buch hat den Kindern niemand [ zu geben] versucht this book has the children no one [ to give ] tried no one has tried to give this book to the children CP DiesesBuch C’ IP Den Kindern IP PRO zu geben Structurally the same as What does John seem to like, except for extra IP segment

  37. Long Scrambling Revisited – 2 CP C' DiesesBuch C IP Den Kindern IP hat niemand I’ IP versucht PRO zu geben Dieses Buch hat den Kindern niemand [ zu geben] versucht this book has the children no one [ to give ] tried no one has tried to give this book to the children

  38. Long Scrambling Revisited - 3 CP IP Restructuring verb clause CP CP IP IP insertion IP IP IP Higher projections unified, skipping extra IP segment result

  39. Revised Tree Adjoining Grammar X X X Y Y Y Y Y AUXILIARYTREE INITIALTREE RESULTTREE BOTH ELEMENTARY TREES

  40. Hierarchy of Defectiveness CP CP CP C’ I’ CP I’ C’ English non-bridge English bridge English raising CP CP IP I’ I’ IP Romance trigger (can be distinguished from Raising with more detailed phrase structure) German trigger (can also have extra IP segments, not shown here) White space is the recursive subtree, shaded gets merged with corresponding projections in complement clause

  41. What about Locality Constraints? • What about the locality constraints on the different types of non-local movement? • They fall out from the context-free structure of a TAG derivation: • Each derivational step is insensitive to the previous derivational history • Although one elementary can adjoin into another, forming a “derived auxiliary tree”… • …Each step in the derivation can refer only to two elementary trees, not to a “derived tree”.

  42. Outline • The Research Goal - Factoring out Recursion • The Problem – Reduced Constructions. • An approach based on what a “defective complement” means in the context of TAG • A different approach based on dominance relations • Further consequences of the first approach • Some Future Directions

  43. Trees as Monotonic Dominance Links - 1 • A major effort toward resolving some of these issues is the system of Description Tree Grammars (Rambow,Vijay-Shanker,Weir, 2001). Closely related is work by (Frank, Kulick, Vijay-Shanker, 2001) • The idea is that instead of using trees as the elementary objects of the grammar, use “descriptions of trees” (Marcus et. al 83). • Trees consist of immediate and non-immediate dominance links. These have to be maintained throughout the course of the derivation. • For other reasons, the “direction” of the derivation is reversed from TAG.

  44. Trees as Monotonic Dominance links - 2 C’ C’ does IP does IP IP CP IP CP I’ What C’ What C’ John I’ John I’ to like I’ seem I’ I’ seem I’ to like

  45. Trees as Monotonic Dominance links - 3 IP CP IP I’ John I’ What C’ John I’ to be I’certain I’ to like CP C’ C’ What C’ does IP does IP I’ I’ seem I’ seem I’ to be I’certain to like

  46. Trees as Monotonic Dominance Links - 4 • This “solves” the problem of interleaving… • …but it has locality problems. Nothing prevents something from “floating” too far • This is true for all different types of inter-clausal movement – for example, super-raising…

  47. Trees as Monotonic Dominance links - 5 I’ I’ IP IP IP seems IP seems IP John I’ John I’ it is I’ certain it is I’certain Super-raising – no sense of locality *John seems it is certainto like pizza I’ to like pizza to like pizza

  48. Trees as Monotonic Dominance Links - 6 A A C A B C a b b c c B a B C C A A a B B B b C C c

  49. Trees as Monotonic Dominance links - 7 A A A a B C A a b b c c B B B b C C c A B a B C

  50. Trees as Monotonic Dominance links - 8 A A A a a a c b b B B b B C C C c e

More Related