From local to non-local dependencies Unbounded Dependency - - PowerPoint PPT Presentation

Unbounded Dependency Constructions (UDCs) in HPSG

Introduction to HPSG

• 26. Mai 2009

Kordula De Kuthy

1

From local to non-local dependencies

• A head generally realizes its arguments locally within its head domain.
• Certain kind of constructions resist this generalization, such as, for

example, the wh-questions discussed below.

• How can the non-local relation between a head and such arguments be

licensed? How can the properties be captured?

2

A first example: Wh-elements

Wh-elements can have different functions: (1) a. Who did Hobbs see ?

Object of verb

• b. Who do you think

saw the man?

Subject of verb

• c. Who did Hobbs give the book to

?

Object of prep

• d. Who did Hobbs consider

to be a fool?

Object of obj-control verb

Wh-elements can also occur in subordinate clauses: (2) a. I asked who the man saw .

• b. I asked who the man considered

to be a fool .

• c. I asked who Hobbs gave the book to

.

• d. I asked who you thought

saw Hobbs.

3

Different categories can be extracted: (3) a. Which man did you talk to ?

NP

• b. [To [which man]] did you talk

?

PP

• c. [How ill] has the man been

?

AdjP

• d. [How frequently] did you see the man

?

AdvP

This sometimes provides multiple options for a constituent: (4) a. Who does he rely [on ]?

• b. [On whom] does he rely

? Unboundedness: (5) a. Who do you think Hobbs saw ?

• b. Who do you think Hobbs said he saw

?

• c. Who do you think Hobbs said he imagined that he saw

?

4

Unbounded dependency constructions

An unbounded dependency construction – involves constituents with different functions – involves constituents of different categories – is in principle unbounded Two kind of unbounded dependency constructions (UDCs) – Strong UDCs – Weak UDCs

5

Strong UDCs

An overt constituent occurs in a non-argument position: Topicalization: (6) Kimi, Sandy loves

i .

Wh-questions: (7) I wonder [whoi Sandy loves

i ].

Wh-relative clauses: (8) This is the politician [whoi Sandy loves

i ].

It-clefts: (9) It is Kimi [whoi Sandy loves

i ].

Pseudoclefts: (10) [Whati Sandy loves

i ] is Kimi.

6

Weak UDCs

No overt constituent in a non-argument position: Purpose infinitive (for-to clauses): (11) I bought iti for Sandy to eat

i .

Tough movement: (12) Sandyi is hard to love

i .

Relative clause without overt relative pronoun: (13) This is [the politician]i [Sandy loves

i ].

It-clefts without overt relative pronoun: (14) It is Kimi [Sandy loves

i ].

7

Some properties of UDC constructions

Link between filler and gap with category information needed: (15) a. Kimi, Sandy trusts

i.

b. [On Kim]i, Sandy depends

i.

(16) a. * [On Kim]i, Sandy trusts

i.

• b. * Kimi, Sandy depends

i.

8

And this link has to be established for an unbounded length: (17) a. Kimi, Chris knows Sandy trusts

i.

b. [On Kim]i, Chris knows Sandy depends

i.

(18) a. * [On Kim]i, Chris knows Sandy trusts

i.

• b. * Kimi, Chris knows Sandy depends

i.

(19) a. Kimi, Dana believes Chris knows Sandy trusts

i.

b. [On Kim]i, Dana believes Chris knows Sandy depends

i.

(20) a. * [On Kim]i, Dana believes Chris knows Sandy trusts

i.

• b. * Kimi, Dana believes Chris knows Sandy depends

i.

9

Using the feature slash

To account for UDCs, we will use the feature slash (so-named because it comes from notation like S/NP to mean an S missing an NP)

• This is a non-local feature which originates with a trace,
• gets passed up the tree,
• and is finally bound by a filler

10

An example for a strong UDC

Johni NP we NP know V she

1NP

likes V

• loc|cat|subcat
• 1,2
• i

2NP

h c VP

• loc|cat|subcat
• 1
• c

h S loc|cat|subcat

• h

c VP c h S f h S

The bottom of a UDC: Traces

        word phon

• synsem

   local

1

nonloc

• inherited|slash
• 1
• to-bind|slash

{}

• 

         

• phonologically null, but structure-shares local and slash values
• we’ll talk about to-bind later

12

Traces

Because the local value of a trace is structure-shared with the slash value, constraints on the trace will be constraints on the filler.

• For example, hates specifies that its object be accusative, and this case

information is local

• So, the trace has

synsem|local|cat|head|case acc as part of its entry, and thus the filler will also have to be accusative (21) *Hei/Himi, John likes

i

13

The middle of a UDC: The Nonlocal Feature Principle (NFP)

For each nonlocal feature, the inherited value on the mother is the union

• f the inherited values on the daughter minus the to-bind value on the

head daughter.

• In other words, the slash information (which is part of inherited)

percolates “up” the tree

• This allows the all the local information of a trace to “move up” to the

filler

14

The top of a UDC: Filler-head structures

Filler-head schema phrase dtrs head-filler-struc

• →

           head-dtr|synsem         loc|cat   head verb vform fin

• subcat

   nonloc

• inherited|slash element
• 1
• to-bind|slash
• 1
• 

       filler-dtr|synsem|local 1           

15

The top of a UDC: Filler-head structures

Explanation of the schema

• Filler and trace are identified as the exact same thing (as far as their

local structure is concerned)

• The trace is “bound” by the to-bind feature; this prevents the slash

value from going any higher in the tree

• Only saturated finite verbs (i.e., sentences) license such structures

16

The top of a UDC: Filler-head structures

Example for a structure licensed by the filler-head schema

• local 1
• nloc
• inherited|slash
• . . . ,1,. . .
• to-bind|slash
• 1
• f

h

• nloc|inherited|slash {}
• 17

The analysis of the strong UDC example

Johni NP

ˆ local 3 ˜

we NP know V she

1NP

likes V

" loc|cat|subcat ˙ 1,2 ¸ nonloc|to-bind|slash {} # i

NP

2 " loc 3 nloc|inher|slash ˘ 3 ¯ #

h c

VP

2 6 4 loc|cat|subcat ˙ 1 ¸ nloc " inherited|slash ˘ 3 ¯ to-bind|slash {} # 3 7 5

c h

S

2 6 4 loc|cat|subcat nloc " inherited|slash ˘ 3 ¯ to-bind|slash {} # 3 7 5

h c

VP

" nloc " inherited|slash ˘ 3 ¯ to-bind|slash {} # #

c h

S

" nloc " inherited|slash ˘ 3 ¯ to-bind|slash ˘ 3 ¯ # #

f h

S

ˆ nloc|inherited|slash {} ˜

The analysis of weak UDCs

(22) a. Kimi is easy (for John) to please

i

• b. Kimi is easy to prove that Mary asked Paul to bribe

i.

(23) a. It is easy to please himacc / * henom.

• b. Inom am easy to please

acc.

⇒ No true (non-argument) filler, only coindexed items serving as arguments Subject is role assigned: (24) a. I believe there to be a unicorn in the garden.

• b. * There is easy to believe a unicorn in the garden.

(25) a. [This sonata]i is easy to play

i on that violin.

b. [This violin]i is easy to play this sonata [on

i].

19

Lexical entry of adjective easy

2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 phon <easy> synsem 2 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 4 loc 2 6 6 6 6 6 6 6 6 6 6 6 6 4 cat 2 6 6 6 4 head adj subcat *NP1, “ PP ˆ for˜

3 ,

” VP h inf, inher|slash element “

2NP

ˆ acc˜ :ppro 1 ” i :4 + 3 7 7 7 5 cont 2 6 6 6 4 easy arg1 1ref arg2 3 arg3 4 3 7 7 7 5 3 7 7 7 7 7 7 7 7 7 7 7 7 5 nonloc|to-bind|slash ˘

2

¯ 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5 3 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 7 5

⇒ Lexical entry selects for infinitive complement missing an NP, which is coindexed with the subject

20

A weak UDC analysis

Ii

3NP1

am V easy A

" loc|cat|subcat ˙ 3,4 ¸ nloc|to-bind|slash ˘ 2 ¯ #

to V[inf ] please V[bse]

" loc|cat|subcat ˙ 5,6 ¸ nonloc|to-bind|slash {} # i 6NP1 " loc 2 nloc|inher|slash ˘ 2 ¯ #

h c

VP[bse]

" loc|cat|subcat ˙ 5 ¸ nloc|inherited|slash ˘ 2 ¯ #

h c

4VP[inf ] " loc|cat|subcat ˙ 5 ¸ nloc|inherited|slash ˘ 2 ¯ #

h c

AP

" loc|cat|subcat ˙ 3 ¸ nloc ˆ inherited|slash {} ˜ #

h c

VP

h loc|cat|subcat ˙ 3 ¸ i

c h

VP

ˆ loc|cat|subcat ˜

Multiple traces with strong and weak UDCs

(26) This is a problem which1 John2 is easy to talk to

2 about 1.

This is analyzed by allowing slash to be a set value and binding each trace

• ff in turn
• easy binds trace2
• The head-filler construction at the top of the tree binds trace1

22

Multiple traces example

easy A

h nloc|to-bind|slash ˘ 2 ¯ i

to V[inf ] talk V to PP

h nloc|inher|slash n 2 1

• i

about PP

h nloc|inher|slash ˘ 3 ¯ i

h c c

VP[bse]

h nloc|inherited|slash ˘ 2,3 ¯ i

h c

VP[inf ]

h nloc|inherited|slash ˘ 2,3 ¯ i

h c

AP

2 4loc|cat|subcat D 4 1 E nloc h inherited|slash ˘ 3 ¯ i 3 5 23

Limiting the occurrence of traces

The that-trace effect, one of the island effects: (27) Whoi did he claim that she kissed

i

(28) * Whoi did he claim that

i kissed her.

The trace principle (for English) Every trace must be strictly subcategorized by a substantive head, i.e., its synsem value must be a non-initial member of a substantive head’s subcat list. → That is, traces should be arguments (subcategorized) and they should not be subjects (strictly) ... Other languages may just require subcategorization

24

Subject extraction

But subjects can be extracted out of embedded clauses (29) * Whoi did he claim that

i kissed her.

(30) Whoi did he claim

i kissed her.

⇒ So, we treat subject extraction in a special, traceless manner, allowing it to happen only in embedded clauses

25

Subject extraction lexical rule (SELR):

"word synsem|local|cat|subcat|rest element “ S ˆ unmarked˜ ” #

→

2 6 6 6 6 4 synsem 2 6 6 6 6 4 local|cat|subcat|rest element B @ VP " loc|cat|subcat D ˆ loc

1

˜ E nonloc|inher|slash {} # 1 C A nonloc|inherited|slash ˘

1

¯ 3 7 7 7 7 5 3 7 7 7 7 5

With the modifications in chapter 9, this will of course be reformulated as a rule applying to subj, not subcat

26

A subject extraction analysis

who NP

ˆ local 1 ˜

did V he

3NP

claim V

" loc|cat|subcat ˙ 3,2 ¸ nonloc|inherited|slash ˘ 1 ¯ #

kissed V[fin]

h loc|cat|subcat D ˆ local 1 ˜ , 4 E i

her

4NP

h c

2VP[fin] h loc|cat|subcat D ˆ local 1 ˜ E i

h c

VP

" loc|cat|subcat ˙ 3 ¸ nonloc|inherited|slash ˘ 1 ¯ #

h c c

S

2 6 4 loc|cat|subcat nonloc " inherited|slash ˘ 1 ¯ to-bind|slash ˘ 1 ¯ # 3 7 5

f h

S

Island constraints

With the definition of strict subcategorization in the Trace Principle, certain island constraints are correctly predicted. (31) a. How tall do you think the building is ?

• b. *How do you think the building is

tall? (modifier) (32) a. That is the building [whose design our architects rejected ].

• b. *That is the building [whose our architects rejected [

design]]. (not strict) (33) a. The books that I like, Leslie donated to the library.

• b. *The books, Leslie donated [

that I like] to the library. (head)

28

Multiple unbounded dependencies

(34) a. It will be easy to play even the most difficult sonata on a violin this well crafted.

• b. [A violin this well crafted]1, even [the most difficult sonata]2 will

be easy to play

2 on 1.

(35) a. It is easy to talk to John about this topic.

• b. This is the topic which1 John2 is easy to talk to

2 about 1.

• Traces are bound at different points in the tree (as we saw earlier)
• There are also cases, on the other hand, where traces are bound in unison

29

Multiple unbounded dependencies (cont.)

(36) That was the rebel leader who1 rivals of

1 shot 1.

An example like this is handled rather smoothly because slash is a set- valued feature, and so the two traces can be identified as the same. What distinguishes the following, however? (37) a. *That was the rebel leader who1 rivals of

1 shot the British

consul.

• b. That was the rebel leader who1 agents of foreign powers shot

1.

30

Parasitic gaps

Extraction out of objects is possible in English: (38) Who did John assassinate [rivals of ] ? Extraction out of subjects, however, is only possible in the presence of a second gap: (39) Who did [rivals of ] assassinate ? (40) a. * Who did [rivals of ] assassinate the President? b. Who did [rivals of the president] assassinate ? The first trace is thus parasitic on the second one

31

Capturing parasitic gaps

The subject condition The initial element of a lexical head’s subcat list may be slashed only if that list contains another slashed element.

• All parasitic gaps are contained within subjects.
• Any subject gap must be licensed by a gap from the same subcat list.

32

Summary of HPSG trace theory

• 1. Nonlocal Feature Principle: For each nonlocal feature, the inherited

value on the mother is the union of the inherited values on the daughter minus the to-bind value on the head daughter.

• 2. Trace Principle (English): Every trace must be strictly subcategorized by

a substantive head, i.e., its synsem value must be a non-initial member

• f a substantive head’s subcat list.
• 3. Subject Condition (English): The initial element of a lexical head’s

subcat list may be slashed only if that list contains another slashed element.

33