Lexicalist Approaches to Syntax Day 5 Part I: Origins of - - PowerPoint PPT Presentation

Lexicalist Approaches to Syntax

Day 5 Part I: Origins of Head-Driven Phrase Structure Grammar

Stephen Wechsler The University of Texas at Austin

Course outline

Weeks 1-2. Lexical Functional Grammar

• 1. lexicalism; LFG formalism
• 2. f- to c-structure mapping; nonconfigurationality; head

mobility; grammatical functions

• 3. pronouns and agreement
• 4. anaphoric binding (or LMT?)

Weeks 3-4. Head-driven Phrase Structure Grammar

• 5. complementation; lexicality and argument structure
• 6. raising and control
• 7. unbounded dependency constructions
• 8. complex predicates, resultatives

Origins of HPSG

Head-Driven Phrase Structure Grammar (HPSG) is a kind of phrase stucture grammar. HPSG grew out of Generalized Phrase Structure Grammar (GPSG).

• What is GPSG?
• What problems led to its abandonment?
• How does HPSG solve those problems?

Context-Free Phrase Structure Grammar

Lexicon: a list of pairs < word, category>

• Set of rules: C0 → C1 ... Cn
• Lexical and nonlexical categories

(N, V, S, NP, VP, etc)

• Designated ‘initial symbol’

A Simple CFG

RULES S → NP VP NP → (D) A* N PP* VP → V (NP) (PP) PP→ P NP LEXICON D: the, some A: big, brown, old N: birds, fleas, dog, hunter, I V: attack, ate, watched P: for, beside, with

Sample Tree

S NP VP D A A N PP V NP PP the big brown dog P NP watched D N P NP with N the birds beside D N fleas the hunter

The Chomsky Hierarchy

Generative grammars in a hierarchy of restrictiveness: Type 0: Unrestricted (transformational grammar) Type 1. Context Sensitive Phrase Structure Grammar Type 2. Context Free Phrase Structure Grammar (CFG) Type 3. Finite State Grammar (FSG) Chomsky (1956, Syntactic Structures):

• showed that a FSG is inadequate for natural language
• suggested (but did not prove) that CFG is likely to be inadequate
• proposed a transformational grammar for natural language

The Chomsky Hierarchy

A hierarchy of restrictiveness: Type 0: Unrestricted (transformational grammar) Type 1. Context Sensitive Phrase Structure Grammar Type 2. Context Free (Phrase Structure) Grammar Type 3. Finite State Grammar Gazdar, Klein, Pullum, and Sag (1985):

• claimed that CFG is adequate to describe natural langage
• proposed Generalized Phrase Structure Grammar

Unbounded dependencies: a challenge for CFG I wonder [which man Fido bit __ ] I wonder [which man Mary said that Fido bit __ ] I wonder [which man John thinks that Mary said that Fido bit __ ] *I wonder which man Fido bit the mailman. Chomsky 1956: generate a base tree; use a transformation to move the phrase Gazdar et al 1985: It can be done with CFG, without transformations.

GPSG

phrase structure rules: S → NP VP NP → (D) A* N PP* VP → V (NP) (PP) (S) NP/NP → t (to introduce an NP trace) IQ → NP S/NP (an IQ consists of an NP followed by an S containing an NP trace) I wonder [ [ which man ] [ Fido bit t ]S/NP ]IQ

GPSG

A meta-rule: For each basic rule of the form: α → σ1 … σi … σn there is a derived rule: α/β → σ1 … σi/β … σn basic rules derived rules S → NP VP S → NP VP/NP VP → V NP VP/NP → V NP/NP … etc.

IQ NP S/NP D N NP VP/NP which man N V NP/NP Fido bit t GPSG: Unbounded dependencies are reduced to a chain of local dependencies which are thus expressible in a CFG. This approach is adopted in HPSG.

Problems with GPSG

• 1. Although natural languages appear to be almost entirely

describable with a CFG, certain languages (Swiss German, Dutch and Bambara) contain constructions that are provably beyond the descriptive capacity of a CFG.

• 2. The derivation of cognates show that words have valence

structures— specifications of what sorts of complements they must combine with. devour: [ COMPS ⟨NP⟩ ] Kids devour pie. laugh: [ COMPS ⟨ ⟩ ] Mary laughed (*a banana). HPSG solution: An extension of CFG using feature structures instead of atomic symbols for the nodes of the tree.

A CFG licenses allowable local subtrees, structures consisting of a mother node and her immediate daughters. X → Y Z X Y Z

HPSG: A CFG using feature structures:

ATT1 value1 ATT2 value2 ! " # $ % & → ATT3 value3 ATT4 value4 ! " # $ % & ATT5 value5 ATT6 value6 ! " # $ % & ATT1 value1 ATT2 value2 ! " # $ % & ATT3 value3 ATT4 value4 ! " # $ % & ATT5 value5 ATT6 value6 ! " # $ % &

Problem 1. Certain languages (Swiss German, Dutch and Bambara) contain constructions that are provably beyond the descriptive capacity of a CFG. Solution: Recursion allows for infinite sets of feature structures which allows for the description of languages that are not context-free.

ATT1 value1 ATT2 ATT3 value3 ATT4 value4 ! " # $ % & ! " # # # # $ % & & & &

→

ATT6 value7 ATT8 ATT3 value9 ATT4 value10 ! " # $ % & ! " # # # # $ % & & & & ATT11 value12 ATT13 ATT3 value14 ATT4 value15 ! " # $ % & ! " # # # # $ % & & & &

Problem 2. Words vary as to what sorts of complements they must combine with; and complement selection is shared across cognates. Kids devour pie. (cp. *Kids devour.) Mary laughed. (cp. *Mary laughed a banana). Why was this a problem for GPSG? How does HPSG solve the problem?

Complement selection in GPSG Lexical items have a number assigned to them and can be inserted into phrasal rules that have the same number. Lexicon: V[2] → laugh V[3] → devour Rules: VP → V[2] VP → V[3], NP

Complement selection in HPSG lexicon: words with valence structures: devour: SPR NP COMPS NP

! " # # # # $ % & & & &

laugh: SPR NP COMPS

! " # # # # # # $ % & & & & & &

kids, pie: NP rules: phrase SPR

! " # # # $ % & & & → [1]

phrase SPR [1]

! " # # # # $ % & & & &

phrase COMPS

! " # # # $ % & & & →

word COMPS [1],… [n]

! " # # # $ % & & & [1], … [n]

SPR COMPS

! " # # # $ % & & &

[1]NP SPR [1] COMPS

! " # # # # $ % & & & &

SPR [1]NP COMPS [2]NP

! " # # # # $ % & & & &

[2]NP Kids devour pie

Problem for GPSG: Many morphological processes are sensitive to lexical valence structure. Verb to adjective conversion with German -bar (and English -able) is productive with transitive (accusative-taking) verbs only. adjective gloss verb and cases lösbar ‘solveable’ lösen (nominative, accusative) vergleichbar ‘comparable’ vergleichen (nom., accusative, PP[mit]) *schlafbar ‘sleepable’ schlafen (nominative) *helfbar ‘helpable’ helfen (nominative, dative)

Verb to adjective conversion with German -bar (and English -able) is productive with transitive (accusative-taking) verbs only. In HPSG: a lexical rule relates transitive Verb to -bar Adjective: V[COMPS ⟨[1]NP[acc]⟩] ⇒ Adj[SPR ⟨[1]⟩] In GPSG, Can we say that -bar derivation applies only to verbs with certain category numbers? No, e.g. lösen and vergleichen have different valence frames, hence different numbers. One could then specify a set of numbers. But the numbers by themselves do not contain any information about the presence of a direct object. So such a formulation of the -bar derivation rule would amount to stipulating a seemingly arbitrary set of numbers, and thereby miss the generalization.

Problem for GPSG: Partial fronting:

1. [Erzählen] wird er seiner Tochter ein Märchen können. tell will he.NOM his daughter.DAT a fairy.tale.ACC can 2. [Ein Märchen erzählen] wird er seiner Tochter können. a fairy.tale.ACC tell will he his daughter.DAT can

• 3. [Seiner Tochter

ein Märchen erzählen] wird er können. his daughter.DAT a fairy.tale.ACC tell will he.NOM can ‘He will be able to tell his daughter a fairy tale.’

In GPSG-like approaches the arguments are licensed by a certain phrase structure rule. But in partial fronting the generalization is simply that the fronted and unfronted arguments must add up to the total set belonging to the verb. This shows that the verb has a lexical valence structure. erzählen: [COMPS ⟨NP[dat], NP[acc]⟩]

Conclusion Words have lexical valence structures!