Information Structure Resolution as an Anaphora Problem

1
Information Structure Resolution as an Anaphora
Problem

• Stefan Bott

2
What Info Structure has to do with anaphora

• There is a general agreement the focus-background
  partitioning depends on the context, i.e. is
  anaphoric (in a general and intuitive sense!).

3
What Info Structure has to do with anaphora

• There is a general agreement the focus-background
  partitioning depends on the context, i.e. is
  anaphoric (in a general and intuitive sense!).

E.g (1) a. What about Paul? b. The company
FIRED bkPaul.
4
What Info Structure has to do with anaphora

• There is a general agreement the focus-background
  partitioning depends on the context, i.e. is
  anaphoric.
• The reading aloud challenge (recoverability)

5
What Info Structure has to do with anaphora

• There is a general agreement the focus-background
  partitioning depends on the context, i.e. is
  anaphoric.
• The reading aloud challenge (recoverability)
• There are, however, important differences between
  background elements and other anaphora.

6
What Info Structure has to do with anaphora
Pronouns and donkey sentences in DRT (2) a)
Every linguist discusses a donkey sentence b)
It is a rather complex sentence.

• There is a general agreement the focus-background
  partitioning depends on the context, i.e. is
  anaphoric.
• The reading aloud challenge (recoverability)
• There are, however, important differences between
  background elements and other anaphora.

z,
z?.
x
y
linguist (x)
donkey_s (y)discuss(x,y)
7
What Info Structure has to do with anaphora
Why does that not extend to backgrounded
material? (3) a) Every linguist discusses a
donkey sentence b) ok Donkey sentences are
beautiful b) ok Linguists LIKE bgdonkey
sentences

• There is a general agreement the focus-background
  partitioning depends on the context, i.e. is
  anaphoric.
• The reading aloud challenge (recoverability)
• There are, however, important differences between
  background elements and other anaphora.

z, ...
Why can the backgrounded donkey sentences be
anaphoric outside the sub-DRS???
donkey_s(z) anaphoric(z,y)
x
y
linguist (x)
donkey_s (y)discuss(x,y)
8
What Info Structure has to do with anaphora
Why does that not extend to backgrounded
material? (3) a) Every linguist discusses a
donkey sentence b) ok Donkey sentences are
beautiful b) ok Linguists LIKE bgdonkey
sentences

• There is a general agreement the focus-background
  partitioning depends on the context, i.e. is
  anaphoric.
• The reading aloud challenge (recoverability)
• There are, however, important differences between
  background elements and other anaphora.

The accessibility problem
z, ...
Why can the backgrounded donkey sentences be
anaphoric outside the sub-DRS???
donkey_s(z) anaphoric(z,y)
x
y
linguist (x)
donkey_s (y)discuss(x,y)
9
What Info Structure has to do with anaphora
Other examples/problemsGeneric (kind refering)
NPs(4) a. Pirates are scum b. Nevertheless
Mary MARRIED bga pirate.Bridging (5) a. The
motor broke. b. So, John bought a NEW car.
antecedent underspefication (cf. Van Deemter
1993)(6) a) Mozart wrote a lotta pieces for the
viola. b) He must have LOVED bgstring
instruments

• There is a general agreement the focus-background
  partitioning depends on the context, i.e. is
  anaphoric.
• The reading aloud challenge (recoverability)
• There are, however, important differences between
  background elements and other anaphora.

10
What Info Structure has to do with anaphora
Other examples/problemsGeneric (kind refering)
NPs(4) a. Pirates are scum b. Nevertheless
Mary MARRIED bga pirate.Bridging (5) a. The
motor broke. b. So, John bought a NEW car.
antecedent underspefication (cf. Van Deemter
1993)(6) a) Mozart wrote a lotta pieces for the
viola. b) He must have LOVED bgstring
instruments

• There is a general agreement the focus-background
  partitioning depends on the context, i.e. is
  anaphoric.
• The reading aloud challenge (recoverability)
• There are, however, important differences between
  background elements and other anaphora.

The non-identity problem
11
What Info Structure has to do with anaphora

• Hypothesis
• Backgrounded material can be treated as anaphora
  if
• The context model is rich enough (bridging,
  underspecification, )
• We have a way to explain why some anaphora do not
  respect accessibility constraints (cf donkey
  sentences)

12
What Info Structure has to do with anaphora
ANOTHER QUESTION If IS resolution turns out
to be a purely anaphoric process, why does NL
bother to code it syntactically, phonologically,
morphologically, typographically .
13
What Info Structure has to do with anaphora
ANOTHER QUESTION If IS resolution turns out
to be a purely anaphoric process, why does NL
bother to code it syntactically, phonologically,
morphologically, typographically .
The theoretical status problem
14
Two sources to inherit from

• Event Semantics and Bachs (1986) algebraic
  treatment of event (and other referents)
• Van der Sandts (1992) theory of presupposition /
  presupposition accommodation as Anaphora
  Resolution

15
Event Semantics Bach (1986)

• Different (sorted) domains of reference
• - individuals (incl. plural individuals)
• - stuff which the individuals are made of
  (properties/kinds)
• - atomic events
• - bounded processes (bits of processes)

16
Event Semantics Bach (1986)

• Each of these domains has the structure of a
  complete semi-joint lattices

17
Event Semantics Bach (1986)

• Each of these domains has the structure of a
  complete semi-joint lattices

En example for Individuals (singular and plural)
John Paul Ringo
John Paul
John Ringo
Paul Ringo
Paul
John
Ringo
18
Event Semantics Bach (1986)

• Additional components
• Temporal relations overlap and precedence
• A homomorphism which maps individuals to the
  stuff they are made of and atomic events to
  bounded events (the bits of process

19
Event Semantics Bach (1986)

• Some modifications we will make to this proposal
  (in part taken from Carlson 2001)
• 1) There are 4 domains
• entities (individuals)classes of entities
  (kinds, properties)
• events
• classes of events

20
Event Semantics Bach (1986)

• Some modifications we will make to this proposal
  (in part taken from Carlson 2001)
• 2) The semi-joint lattice structure of each
  domain can be (partially) represented by a simple
  part-of relation
• run move
• eat cake eat
• cat mammal animal

21
Event Semantics Bach (1986)

• This bit of theory should help to (partially)
  solve the non-identity problem
• (6) a) Mozart wrote a lotta pieces for the
  viola. b) He must have LOVED bgstring
  instruments
• viola string instruments
• ? string instruments is a legal antecedent for
  viola

22
Event Semantics Bach (1986)

• (9) Mantenir viva una planta no és fàcil. Jo
  tenia un amic que els bitxos sels hi menjaven
  totes. Quan no era el pulgó, eren les erugues.
  Quan matava les erugues, li sortia la cochinilla,
  un nom lleig, també... Un nom dinsecte guarro. I
  la mosca blanca.
• Al principi estava preocupat, el tio, però al
  final els hi va acabar agafant carinyoi tot,
  bgals bitxos. (57)
• Buenafuente, cited in Mayol (2002)

23
Van der Sandt (1992)

• Presupposition resolution is the same mechanism
  as anaphora resolution.
• (7) John has a wife. Johns wife is a singer.
• The possesive introduces an existential
  presupposition.

24
Van der Sandt (1992)

• If the (existential) presupposition does not meet
  an antecedent, the existence of the presupposedly
  existing referent will be introduced by
  accommodation.

25
Van der Sandt (1992)
(7) a) John has a wife.

• If the (existential) presupposition does not meet
  an antecedent, the existence of the presupposedly
  existing referent will be introduced by
  accommodation.

x, y
john(x)wife(y)has(y,x)
26
Van der Sandt (1992)
(7) a) John has a wife. b) Johns wife is a
singer

• If the (existential) presupposition does not meet
  an antecedent, the existence of the presupposedly
  existing referent will be introduced by
  accommodation.

x, y, z
john(x) wife(z)wife(y) zyhas(y,x) of(y,x)
singer(z)
27
Van der Sandt (1992)
(8) Johns wife is a singer

• If the (existential) presupposition does not meet
  an antecedent, the existence of the presupposedly
  existing referent will be introduced by
  accommodation.

x, y
john(x)wife(y)of(y,x)singer(y)
y is introduced via presupposition accomodation
28
Van der Sandt (1992)

• Presupposition accommodation can introduce a
  referent in a higher DRS than the one in which
  the presupposition is triggered
• (10) a) Every critic liked Jacksons film.
• b) It is a nice piece of work.

29
Van der Sandt (1992)
(10) a) Every critic liked Jacksons
film.

• Presupposition accommodation can introduce a
  referent in a higher DRS than the one in which
  the presupposition is triggered
• (10) a) Every critic liked Jacksons film.
• b) It is a nice piece of work.

y
Jackson (y)
x
z
critic (x)
film (z)of (z,y)like (x,z)
30
Van der Sandt (1992)
(10) a) Every critic liked Jacksons
film.

• Presupposition accommodation can introduce a
  referent in a higher DRS than the one in which
  the presupposition is triggered
• (10) a) Every critic liked Jacksons film.
• b) It is a nice piece of work.

y
Jackson (y)
Presuppositions can be introduced in a higher DRS!
x
z
critic (x)
film (z)of (z,y)like (x,z)
31
Van der Sandt (1992)
(10) a) Every critic liked Jacksons
film.

• Presupposition accommodation can introduce a
  referent in a higher DRS than the one in which
  the presupposition is triggered
• (10) a) Every critic liked Jacksons film.
• b) It is a nice piece of work.

y, z
Jackson (y)
x
critic (x)
film (z)of (z,y)like (x,z)
32
Van der Sandt (1992)
(10) a) Every critic liked Jacksons film. b) It
is a nice piece of work.

• Presupposition accommodation can introduce a
  referent in a higher DRS than the one in which
  the presupposition is triggered
• (10) a) Every critic liked Jacksons film.
• b) It is a nice piece of work.

y, z, v
Jackson (y) vznice_piece_o_work(z)
x
critic (x)
film (z)of (z,y)like (x,z)
33
Class is a question of presupposition

• Hypothesis 1 The existence of an discourse
  referent presupposes the existence of the class
  it belongs to.

34
Class is a question of presupposition

• Hypothesis 1 The existence of an discourse
  referent presupposes the existence of the class
  it belongs to.

Example The indefinite NP a cat (11) A cat
Introduces an discourse referent x, such that
cat (x). But it also presupposes that there is a
class of entities which share the property of
being cats.
35
Class is a question of presupposition

• Hypothesis 2There are different types of
  discourse referents entities, events and their
  respective classes

36
NOW

• its time to extend DRT

37
Extending DRT Typed variables

• (12) El gat menja peix.

x, y, e
gat(x)peix(y)menjar(e)agent(e,x)patient(e,y)t
ime(e, now)aspect(e, .....)
38
Extending DRT Typed variables

• () El gat menja peix.

Now we associate each referent to domain it
belongs to. Note that we have already introduced
e as a referent for the event eat.
x, y, e
gat(x)peix(y)menjar(e)agent(e,x)patient(e,y)t
ime(e, now)aspect(e, .....)
39
Extending DRT Typed variables

• Types entities/objects individual oi class
  oc
• events individual ei class
  ec

xoi, yoi, eei
gat(x)peix(y)menjar(e)agent(e,x)patient(e,y)t
ime(e, now)aspect(e, .....)
40
Extending DRT Typed variables
Types entities/objects individual oi class
oc events individual ei
class ec
xoi, yoi, eei
gat(x)peix(y)menjar(e)agent(e,x)patient(e,y)t
ime(e, now)aspect(e, .....)
BUT WHAT ABOUT ENTITY AND EVENT CLASSES???
41
Extending DRT Typed variables
According to Hypothesis 1 every individual
presupposes the existence of its corresponding
class. So the problem is parallel to Van der
Sandts presupposition accommodation.

• Types entities/objects individual oi class
  oc
• events individual ei class ec

xoi, yoi, eei
gat(x)peix(y)menjar(e)agent(e,x)patient(e,y)t
ime(e, now)aspect(e, .....)
42
Extending DRT Typed variables

• Types entities/objects individual oi class oc
• events individual ei class ec

xoi, yoi, z1oc, z2oc, eei, fec
gat(x) instance(x,z1)peix(y)
instance(y,z2) menjar(e) instance(e,f)
agent(e,x)patient(e,y)time(e, now)aspect(e,
.....)
43
Accessibility and Non-identity revisited

• (3) a) Every linguist discusses a donkey
  sentence b) Linguists LIKE bgdonkey sentences

44
Accessibility and Non-identity revisited
(3) a) Every linguist discusses a donkey
sentence

• (3) a) Every linguist discusses a donkey
  sentence b) Linguists LIKE bgdonkey sentences

• (3) a) Every linguist discusses a donkey
  sentence b) ok Donkey sentences are beautiful
  b) ok Linguists LIKE bgdonkey sentences

x
y
linguist (x)
donkey_s (y)discuss(x,y)
45
Accessibility and Non-identity revisited
(3) a) Every linguist discusses a donkey
sentence

• (3) a) Every linguist discusses a donkey
  sentence b) Linguists LIKE bgdonkey sentences

• (3) a) Every linguist discusses a donkey
  sentence b) ok Donkey sentences are beautiful
  b) ok Linguists LIKE bgdonkey sentences

Typing of the variables Introducction of classes
via presupposition accomodation
xoi, c1oc
yoi, c2oc
linguist (x) inst(x,c1)
donkey_s (y)discuss(x,y) inst(y,c2)
46
Accessibility and Non-identity revisited
(3) a) Every linguist discusses a donkey
sentence

• (3) a) Every linguist discusses a donkey
  sentence b) Linguists LIKE bgdonkey sentences

• (3) a) Every linguist discusses a donkey
  sentence b) ok Donkey sentences are beautiful
  b) ok Linguists LIKE bgdonkey sentences

c1oc, c2oc

Projection into a higher DRS
Xoi
yoi
linguist (x) inst(x,c1)
donkey_s (y)discuss(x,y) inst(y,c2)
47
Accessibility and Non-identity revisited
(3) a) Every linguist discusses a donkey
sentence b') Linguists LIKE bgdonkey
sentences

• (3) a) Every linguist discusses a donkey
  sentence b) Linguists LIKE bgdonkey sentences

• (3) a) Every linguist discusses a donkey
  sentence b) ok Donkey sentences are beautiful
  b) ok Linguists LIKE bgdonkey sentences

c1oc, c2oc, v, w
linguists(v) wc2donkey_s(w) like(v,c2)
Xoi
yoi
linguist (x) inst(x,c1)
donkey_s (y)discuss(x,y) inst(y,c2)
48
Advantages

• The non-identity problem is solved, as well as
  the accessibility problem
• IS resolution only needs a logical form of the
  discourse. Interpretation does not influence IS.
  IS does not depend on truth.

49
Contrastive foci

• What about contrastive foci?
• (13) a. John insulted Bill
• b. Then BILL insulted JOHN
• The problem is that all elements in () are
  anaphoric

50
Contrastive Focus Layer2?

• In Bott (2004) (my PhD Thesis proposal) I
  proposed a backoff model for IS resolution
• 1) Simple foci can be directly linked to the
  referential status of the referents contained
  within this focus by the fact that these
  referents cannot be discourse old and discourse
  accessible. The same is true for the background
  of those simple foci. I will call such foci layer
  1 foci, in order to avoid implications that terms
  as presentational would bring about.
•  

51
Contrastive Focus Layer2?

• 2) All foci that cannot be directly determined by
  the simple referential status are receive
  contrastive or parallel reading. Such foci
  involve relational information between discourse
  referents and must, hence, be interpreted on a
  compositional semantic level. I will argue that
  such foci will necessarily depend on semantic
  relations, but such relations are not necessarily
  whole propositions. I will call such foci layer 2
  foci.

52
Contrastive Focus Layer2?

• In simple words
• First Try to find an anaphoric antecedent for
  each word and mark every word with an
  antecedent as background.
• If this leads to sentence without focus, back off
  into layer2
• Layer2 uses information about argument
  realisation for each predicate.

53
Contrastive focus and event semantics

• (13) a. John insulted Bill
• b. Then BILL insulted JOHN
• Context representation (event class domain)

arg1(e, john')
insult(e)
arg1(e, john')
insult(e) arg1(e, john')
arg1(e, john') arg2 (e, bill')
insult(e) arg2 (e, bill')
insult(e) arg1(e, john') arg2 (e, bill')
54
Contrastive focus and event semantics

• (13) a. John insulted Bill
• b. Then BILL insulted JOHN
• Backgrounding
• Pruning out impossible antecedents

arg1(e, john')
insult(e)
arg1(e, john')
insult(e) arg1(e, john')
arg1(e, john') arg2 (e, bill')
insult(e) arg2 (e, bill')
insult(e) arg1(e, john') arg2 (e, bill')
55
Contrastive focus and event semantics

• (13) a. John insulted Bill
• b. Then BILL insulted JOHN
• Backgrounding
• Pruning out impossible antecedents

arg1(e, john')
insult(e)
arg1(e, john')
insult(e) arg1(e, john')
arg1(e, john') arg2 (e, bill')
insult(e) arg2 (e, bill')
insult(e) arg1(e, john') arg2 (e, bill')
56
Contrastive focus and event semantics

• (13) a. John insulted Bill
• b. Then BILL insulted JOHN
• Backgrounding
• Pruning out impossible antecedents

insult(e)
57
Contrastive focus and event semantics

• (13) a. John insulted Bill
• b. Then BILL insulted JOHN
• Backgrounding
• Only insult will be backgrounded

insult(e)
58
Backoff (Layer2) or pure event semantics?

• (13) a. John insulted Bill
• b. Then BILL insulted JOHN
• ProblemIn the entity individuals domain John and
  Bill will be activated after a) has been uttered
• Bill and John should be backgrounded in b)

59
Backoff (Layer2) or pure event semantics?

• (13) a. John insulted Bill
• b. Then BILL insulted JOHN
• Solution1 Only event semantics (the event type
  domain) is relevant for IS resolution

60
Backoff (Layer2) or pure event semantics?

• (13) a. John insulted Bill
• b. Then BILL insulted JOHN
• Solution2 A backoff model
• Principle of focus requirement
• ALL SENTENCES MUST HAVE A FOCUS

61
Backoff (Layer2) or pure event semantics?

• (13) a. John insulted Bill
• b. Then BILL insulted JOHN
• Solution2 A backoff model
• If all elements in the a sentences have an
  antecedent, the sentence is left without a focus
  and backgrounding will only take place on the
  basis of the event type domain

62
Evidence against the pure event semantics
solution

• (14) Q Did you see John?
• A John is working in Leeds now.
• Would we like to analyze John as part of the
  focus in the answer?

63
Evidence against the pure event semantics
solution

• (14) Q Did you see John?
• A John is working in Leeds now.
• Would we like to analyze John as part of the
  focus in the answer?

Representation of the context (the event class
domain) see(e) agent(e, you') patient (e,
john') Qu(e) This implies the partial events
see(e) agent(e, you')see(e) patient (e,
john')..patient (e, john')
64
Evidence against the pure event semantics
solution

• (14) Q Did you see John?
• A John is working in Leeds now.
• Would we like to analyze John as part of the
  focus in the answer?

answer . agent(john)
Representation of the context (the event class
domain) see(e) agent(e, you') patient (e,
john') Qu(e) This implies the partial events
see(e) agent(e, you')see(e) patient (e,
john')..patient (e, john')
65
Evidence against the pure event semantics
solution

• BUT, there is an interesting observation we can
  make (but probably I am wrong)
• In such cases the anaphoric element can
  probably not be a tail (r-dislocated in catalan),
  although it is probably a link (l-dislocated)
• (15) Q Que lhas vist a en Joan?
• A ?Ara viu a Leeds, en Joan.
• A En Joan, ara viu a Leeds.

66
Why is does NL mark topics?

• (16) Després hi ha els que sofereixen A mi
  mhauries de fitxar, Buenafuente! Jo sí que ten
  explicaria, dhistòries!
• Buenafuente, cited in Mayol (2002)
• Tentative answer If a topic is not really
  anaphoric to the discourse context, a
  syntactically (phonologically, ) marked topic
  will trigger presupposition accommodation.

67
(some) References

• Bach, Emmon (1986). The algebra of events
  Linguistics and Philosophy 9, 5-16.
• Parsons, Terence (1990) Events in the semantics
  of English A study in subatomic semantics.
  Cambridge, MIT Press.
• Carlson, Greg N. (2001) Weak indefinites. In
  Coene, Martine and Yves Dhulst (eds.) From NP to
  DP, 195210. (http//semanticsarchive.net/Archive/
  WNhNDdiZ/Indefinites.pdf)
• Steedman, M. (2000) The Syntactic Process. MIT
  Press, Cambridge, MA.
• van Deemter, K. (1993) What's New? A Semantic
  Perspective on Sentence Accent. Journal of
  Semantics 11, 1993, pp.1-31.
• Van der Sandt, R.A. (1992) Presupposition
  projection as anaphora resolution. Journal of
  Semantics 9 333- 377.
• Zuo Y. W. Zuo (2001) The Computing of
  Discourse Focus. Lincom Studies in Pragmatics 6.