101 uses for information structure (It’s not just for prosody you know!)

Post on 12-Jan-2016

28 views 1 download

description

101 uses for information structure (It’s not just for prosody you know!). Taal- en spraaktechnologie Fall 2005 Lecture 2 Jennifer Spenader. Structure of lecture. Is there really only one way to define new-given? In depth look at one theory of new-given (Prince 1981) - PowerPoint PPT Presentation

transcript

1

101 uses for information structure(It’s not just for prosody you know!)

Taal- en spraaktechnologie

Fall 2005

Lecture 2Jennifer Spenader

2

Structure of lecture

1. Is there really only one way to define new-given?• In depth look at one theory of new-given (Prince 1981)

2. How does new-given relate to choice of lexical and syntactic form?• Definite vs. indefinite forms

– Marking with particles

• What appears as a subject ? (in information structural terms)

3. How do we interpret underspecified forms?• Resolution of anaphoric reference• Interpretation of bridging NPs

4. Why is this useful?

3

New-given revisited

4

New and Given (Prince 1981)

“…the Old/New Information Workshop held at Urbana, Summer 1978, was quickly and quite appropriately dubbed “The Mushy Information workshop” “

Mushy peas

5

Different definitions

are the big names in new-given research but– They are all using same terms but slightly different

definitions that make different categorizations

6

Predictability/Recoverability

• GIVENESS AS PREDICTABILITY/RECOVERABILITY– The speaker assumes that the hearer can predict or could

have predicted that a particular linguistic item will or would occur in a particular position within a sentence

• Most similar to Kuno’s definitionEx. (Prince 81:226)

(1) Mary paid John and he/*0 bought himself a new coat.(2) John paid Mary and he/0 bought himself a new coat.

• Deletability coincides with predictability, yet we don’t want to consider “he” in (2) old, but in (1) new.

7

Saliency

• GIVENESS AS SALIENCY– The speaker assumes that the hearer has or could

appropriately have some particular thing/entity/… in her consciousness at the time of hearing the utterance

• Salient to the hearer• Close to Chafe’s definition of givenness

(3) I saw your father yesterday.

(4) I saw a two-headed man yesterday.• Chafe’s definition would treat both referents as new.

8

Saliency: examples

(5) We got some picnic supplies out of the trunk.

The beer was warm. • Bridging anaphors (associative anaphors) would also

be treated as new.

• But Chafe also claims that only given items can be pronominalized. This seems to be wrong if you don’t allow inferred referents to be given, e.g.:

(6) Harry threw up and Sam stepped in it.

9

Shared knowledge

• GIVENESS AS SHARED KNOWLEDGE

The speaker assumes that the hearer “knows”, assumes, or can infer a particular thing (but not necessarily thinking about it).

• Most similar to Clark & Haviland’s definition– Inferred referents, such as “the beer” would be given

(7) Where were your grandparents born?

10

How are all three ideas related?

• If a speaker thinks the hearer can predict something, then the speaker must also believe the hearer has this element in their conscious mind

• If a hearer has something in their conscious mind, then it isn’t so far fetched to believe that they can draw inferences based on it…

11

12

Basically all these theories claim that speaker’s actively try to take the hearer’s knowledge into account…

…but actually there might be some sort advantage for the Speaker

13

ASIDE: Discourse vs. Word Jumbles

• A discourse is different from a collection of words or sentences because it has coherence

• Coherence is manifested in logical progressions of ideas, continuation of topics, appropriate choice of reference that distinguishes new information from old

• A discourse can be a dialogue, monologue or written text

14

ASIDE: Discourse Model

• Understanding a discourse in part involves building a discourse model of the information contributed

• This includes keeping track of the discourse referents brought up in the discourse, including: – How activated they are– What attributes they have– Links between them– (cf. Discourse Representation Theory)

15

Assumptions used to choose form

“From the point of view of a speaker/writer, what kinds of assumptions about the hearer/reader have a bearing on the form of the text being produced…” (Prince 81:233)

16

Assumed Familiarity

Assumed familiarity

Inferrable EvokedNew

Inferrable Inferrable Inferrable InferrableBrand-newUnused

Brand-new (Unanchored)

Brand-new Anchored

17

Examples of Assumed Familiarity

(8) I bought a beautiful dress. (Brand-new + attribute)(9) A rich guy I know bought a Cadillac. (Brand-new Anchored + attribute)

(10) I went to the post office and the stupid clerk couldn’t’ find a stamp. (Inferrable + attribute)

(11) Have you heard the incredible claim that the devil speaks English backwards? (Containing inferrable + attribute)

(12) Susy went to visit her grandmother and the sweet lady was making Peking Duck (Evoked + attribute)

(13) Hi, I’m home. (Situationally evoked)

18

Points

• This new taxonomy of new-given introduces a large number of distinctions– Are these distinctions necessary?, E.g.– Do they correlate with different linguistic or

prosodic forms?– Are they perceived by hearers? Consistently?

19

ASIDE: Higher level categories

• Linguistic categories like parts-of-speech are generally easy to identify– It’s simple to define what nouns and verbs are for

most languages

• Most syntax, once a theory is agreed upon, can be consistently categorized– i.e. we can all recognize noun phrases with a very

high

20

ASIDE:Verifying higher level categories

• But for things like information structure or speech acts, it’s not as clear what categories actually exist

• Annotation experiments:– A high degree of agreement between annotators is

taken as evidence that the categories identified are cognitively real

21

Definites and anaphors

Correlations with new-given statuses

22

What does a referentially “bad” text

look like?

I live in a house in Gronveldstraat 16.

A house in Gronveldstraat 16 was renovated a few years ago.

The landlord put in central heating and fixed the front of a house in Gronveldstraat 16.

Then a house in Gronveldstraat 16 looked really nice.

23

What makes a text referentially better?

I live in a house in Gronveldstraat 16.

The house was renovated a few years ago.

The landlord put in central heating and fixed the front of the house.

Then it looked really nice.

24

Another BAD text

• A man came into the bar. The bartender began to talk to a man.

25

Referents semantically

• Referents are the actual entities in our discourse model

• Two types of referential forms available

– indefinite reference• introduce new referents with indefinite reference

– definite reference• subsequent mention is done with definite reference

26

indefinite vs. definite

• indefinite noun phrases– “a man”, “some children”

• Indefinites tend to be used to introduce new referents – So can be considered a morphological form for

new information

27

Formal Definiteness

• Formal property of NPs (decidable on form alone)

• Formal definiteness: Marking of the NP– Definite articles (the, de, het)– Demonstrative articles (deze, dit)– Possessive adjectives (jouw, jullie, mijn…)– Personal pronouns (je, jij, ik, hij, zij, hem, haar)– Unmodified proper nouns (Esther, Petra)– Certain quantifiers argued to be definite (all, every)

28

Simple indefiniteness to definiteness

RULE: 1. Introduce items with full, indefinite noun phrases. 2. Later, refer to them with definite noun phrases3. Later, as long as it doesn’t lead to confusion, you

can refer with pronouns

(Ex) I live in a house in Gronveldstraat 16. The house was renovated a few years ago. It looks great.

29

Interpretation

• E.g. DRT– Indefinite noun phrases introduce a new discourse

referent– Definite noun phrases trigger a search for an

already given definite referent that can serve as an antecedent

– Pronouns also trigger a search for a compatible antecedent

30

Definiteness correlates

• Definiteness correlates with other linguistic forms

1. Subjects tend to be definite (Prince 1981, Prince 1993)

2. Only indefinites can occur as the object in There –sentences (Ex. From Prince 1993)

(Ex) A/The man was in the room.

There was a man/the man in the room.

31

There-sentences

(5) a. There were the same people at both conferences.

b. There was the usual crowd at the beach.

c. There was the stupidest article on the reading list.

Revised claim: Some definites are not actually definites

.

32

Conceptual Definiteness

• Definite referents are given in some way– They are recognizable, or familiar?– They are specific, or identifiable?– They are unique in their context, and thus

identifiable (logical definition)

• Formal definiteness and conceptual definiteness don’t always coincide

33

Not definiteness, but givenness

• Prince (1993)’s claim: There-sentences don’t require indefinites. There-sentences require hearer-new referents

(5) a. There were the same people at both conferences.

b. There was the usual crowd at the beach.

c. There was the stupidest article on the reading list.

Definiteness: marks identifiable, specific or unique for the speaker, but the There-sentence for marks the awareness that it is hearer-new

34

Generalizations

• Indefinites tend to be used to introduce hearer-new information (discourse-new)

• Definites are used for discourse-given information

• Definites are sometimes used for speaker known, speaker specific, items, not necessarily known to the hearer.

35

Simplification!!!

• And of course, there are many exceptions to what can be marked as definite

• Given in some way = licenses marking

36

New entities introduced with definite

Greta walked slowly through the woods. She enjoyed the stillness.

37

Anaphoric reference to abstract objects

Greta slowly made her way along the path breathing deeply.

However, it wasn’t enough to quiet her troubled thoughts.

38

Inferrables introduced with definite reference

The sky lit up with lightening and thunder, rumbling in a threatening way. The storm was getting closer.

39

From generation to interpretation

• Until now we’ve been looking at things from the perspective of the speaker

• What about in interpretation? What does the hearer need to do with a definite noun phrase or a pronoun?– RESOLVE IT!

40

Features?

• For NLP = easy to identify features– gender– number– recency

• 90% of pronominal referents have antecedents in the same or preceeding sentence

– parallelism– semantic information– saliency:

• does the referring expression seem to be specifying something in focus?

41

World knowledge

World knowledge supports anaphoric resolution

(Ex) The mother dressed her dotter.

Afterwards she fed her.

Who fed who?

42

World Knowledge 2

(from Sayed, Issues in Anaphor Resolution, http://nlp.stanford.edu/courses/cs224n/2003/fp/iqsayed/project_report.pdf)

(Ex)

There were dresses of several different colors and styles.

They were all pretty, labeled with price tags.

Sally chose a blue one. Mary chose a skimpy one.

43

Ambigous, but we have preferences

(From Beaver, 2003)

(Ex) a. Jane likes Mary .

b. She often goes around for tea with her.

c. She chats with the young woman for ages.

c’. SHE chats with the young woman for

ages.

(Ex) a. John hit Martin.

b. He fell.

c. HE fell.

44

Non-referring NPs…

Ex.Little Johnny threw up and then stepped in it.

Ex.John became a guitarist because he thought

that it was a beautiful instrument.

45

Associative (Bridging) anaphora

We took the picnic things out of the trunk.

The beer was warm.

Gerlof entered the ballroom.

The chandelier sparkled brightly.

46

Resolving Bridging Anaphora

• Requires access to lexical information!

• This is what we will talk about NEXT time!– What type of lexical information is relevant to

language processing?

– What type of computational lexical resources exist• Why are they formed the way they are?• What can they be used for?

47

Applications that benefit from anaphor resolution

• Information extraction systems– Pre-med PULS Medical extraction system

• Question-Answer systems– As a subset of information retrieval systems

• Automatic Summarization systems– As a way to make information search more

efficient– Swe-sum