Evaluating Centering for Information Ordering Using Corpora M.Sc. - - PowerPoint PPT Presentation

▶
evaluating centering for information ordering using
SMART_READER_LITE
LIVE PREVIEW

Evaluating Centering for Information Ordering Using Corpora M.Sc. - - PowerPoint PPT Presentation

Apr 30, 2023 154 likes •679 views

Evaluating Centering for Information Ordering Using Corpora M.Sc. Seminar: Discourse Coherence Theories and Modeling Jonathan Poitz Department of Computational Linguistics, Saarland University May 13th, 2013 Jonathan Poitz (CoLi Saarland)

slide-1
SLIDE 1

Evaluating Centering for Information Ordering Using Corpora

M.Sc. Seminar: Discourse Coherence Theories and Modeling Jonathan Poitz

Department of Computational Linguistics, Saarland University

May 13th, 2013

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 1 / 35

slide-2
SLIDE 2

Motivation

It shows a beautiful sunset. Most of my vacations I spend on Malta. I took this picture on the west coast of the island last year.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 2 / 35

slide-3
SLIDE 3

Motivation

Most of my vacations I spend on Malta. I took this picture on the west coast of the island last year. It shows a beautiful sunset.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 2 / 35

slide-4
SLIDE 4

Papers to be presented

1 Evaluating Centering for Information Ordering Using Corpora

(KARAMANIS, POESIO, MELLISH, OBERLANDER 2008)

2 Specifying the Parameters of Centering Theory: a Corpus-Based

Evaluation using Text from Application-Oriented Domains (POESIO et al. 2000)

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 3 / 35

slide-5
SLIDE 5

Overview

1

Centering Theory The Theory Principles of Centering Parameters of Centering

2

Application on Information Ordering Motivation The Metrics

3

Experimental Evaluation Test Data Results Conclusion

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 4 / 35

slide-6
SLIDE 6

A possible approach - Centering Theory Questions to be answered

1 How appropriate is the Centering Theory for

Information Ordering?

2 Which aspects of it are most useful?

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 5 / 35

slide-7
SLIDE 7

Centering Theory The Theory

What is Centering?

  • Developed by Barbaba Grosz, Candy Sidner & Aravind Joshi (1986,

1995)

  • Goal was to offer a new approach for anaphora resolution
  • Two different data structures for each utterance:
  • CF(U) – Forward looking center holding all Discourse Entities (DE) of

utterance n with the preferred center CP(U) defined as its first item

  • CB(U) – Backward looking center defined as highest ranked element of

CF(Un − 1) that also occurs in CF(Un)

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 6 / 35

slide-8
SLIDE 8

Centering Theory The Theory

An Example

Example Most of my vacations I spend on Malta. CF: {vacations, I, Malta} CB: {n.a.}

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 7 / 35

slide-9
SLIDE 9

Centering Theory The Theory

An Example

Example Most of my vacations I spend on Malta. CF: {vacations, I, Malta} CB: {n.a.} I took this picture last year. CF: {I, picture, last year} CB: {I}

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 7 / 35

slide-10
SLIDE 10

Centering Theory The Theory

An Example

Example Most of my vacations I spend on Malta. CF: {vacations, I, Malta} CB: {n.a.} I took this picture last year. CF: {I, picture, last year} CB: {I} It shows a beautiful sunset. CF: {picture, sunset} CB: {picture}

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 7 / 35

slide-11
SLIDE 11

Centering Theory The Theory

The transitions

Comparing the forward and backward looking centers of an utterance several transitions were defined by Brennan, Friedman & Pollard in 1987.

  • Continue
  • Retain
  • Smooth-Shift
  • Rough-Shift
  • NOCB (Kibble & Power 2000)

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 8 / 35

slide-12
SLIDE 12

Centering Theory The Theory

An example

1 This cake is amazing and I am very

hungry!

2 A friend prepared it yesterday. 3 He is a good cook. 4 He also sometimes prepares delicious

pasta.

5 Unfortunately I’m allergic to pasta. 6 Green elephants eat breakfast. 1 CB(U) : n.a. 2 CB(U) : cake 3 CB(U) : friend 4 CB(U) : friend 5 CB(U) : pasta 6 CB(U) : n.a.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 9 / 35

slide-13
SLIDE 13

Centering Theory The Theory

An example

1 This cake is amazing and I am very

hungry!

2 A friend prepared it yesterday. 3 He is a good cook. 4 He also sometimes prepares delicious

pasta.

5 Unfortunately I’m allergic to pasta. 6 Green elephants eat breakfast. 1 CB(U) : n.a. 2 CB(U) : cake 3 CB(U) : friend 4 CB(U) : friend 5 CB(U) : pasta 6 CB(U) : n.a.

Continue CB(Un) = CB(Un − 1) and CB(Un) = CP(Un)

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 9 / 35

slide-14
SLIDE 14

Centering Theory The Theory

An example

1 This cake is amazing and I am very

hungry!

2 A friend prepared it yesterday. 3 He is a good cook. 4 He also sometimes prepares delicious

pasta.

5 Unfortunately I’m allergic to pasta. 6 Green elephants eat breakfast. 1 CB(U) : n.a. 2 CB(U) : cake 3 CB(U) : friend 4 CB(U) : friend 5 CB(U) : pasta 6 CB(U) : n.a.

Retain CB(Un) = CB(Un − 1) or CB(Un − 1) = undef. and CB(Un) = CP(Un)

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 9 / 35

slide-15
SLIDE 15

Centering Theory The Theory

An example

1 This cake is amazing and I am very

hungry!

2 A friend prepared it yesterday. 3 He is a good cook. 4 He also sometimes prepares delicious

pasta.

5 Unfortunately I’m allergic to pasta. 6 Green elephants eat breakfast. 1 CB(U) : n.a. 2 CB(U) : cake 3 CB(U) : friend 4 CB(U) : friend 5 CB(U) : pasta 6 CB(U) : n.a.

Smooth Shift CB(Un) = CB(Un − 1) and CB(Un) = CP(Un)

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 9 / 35

slide-16
SLIDE 16

Centering Theory The Theory

An example

1 This cake is amazing and I am very

hungry!

2 A friend prepared it yesterday. 3 He is a good cook. 4 He also sometimes prepares delicious

pasta.

5 Unfortunately I’m allergic to pasta. 6 Green elephants eat breakfast. 1 CB(U) : n.a. 2 CB(U) : cake 3 CB(U) : friend 4 CB(U) : friend 5 CB(U) : pasta 6 CB(U) : n.a.

Rough Shift CB(Un) = CB(Un − 1) and CB(Un) = CP(Un)

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 9 / 35

slide-17
SLIDE 17

Centering Theory The Theory

An example

1 This cake is amazing and I am very

hungry!

2 A friend prepared it yesterday. 3 He is a good cook. 4 He also sometimes prepares delicious

pasta.

5 Unfortunately I’m allergic to pasta. 6 Green elephants eat breakfast. 1 CB(U) : n.a. 2 CB(U) : cake 3 CB(U) : friend 4 CB(U) : friend 5 CB(U) : pasta 6 CB(U) : n.a.

NOCB CB = n.a.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 9 / 35

slide-18
SLIDE 18

Centering Theory The Theory

Rules and Constraints

  • Constraint 1: All utterances of a segment except for the first have

exactly one CB.

  • Rule 1: If any element of Un − 1 is realized by a pronoun in Un, the

CB(Un) is too.

  • Rule 2: Sequence preferences are Continue > Retain > Smooth Shift >

Rough Shift

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 10 / 35

slide-19
SLIDE 19

Centering Theory The Theory

Rules and Constraints

  • Constraint 1: All utterances of a segment except for the first have

exactly one CB.

  • Rule 1: If any element of Un − 1 is realized by a pronoun in Un, the

CB(Un) is too.

  • Rule 2: Sequence preferences are Continue > Retain > Smooth Shift >

Rough Shift Example for rule 1 Peter is in a bad mood. He visited his friend John. He didn’t like Peter’s visit.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 10 / 35

slide-20
SLIDE 20

Centering Theory The Theory

Rules and Constraints

  • Constraint 1: All utterances of a segment except for the first have

exactly one CB.

  • Rule 1: If any element of Un − 1 is realized by a pronoun in Un, the

CB(Un) is too.

  • Rule 2: Sequence preferences are Continue > Retain > Smooth Shift >

Rough Shift Example for rule 1 Peter is in a bad mood. He visited his friend John. He didn’t like Peter’s visit. Example for rule 2 Peter likes to play the drums. He and Paul are friends of Mary. She doesn’t like what he eats.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 10 / 35

slide-21
SLIDE 21

Centering Theory Principles of Centering

Details of Centering I

  • Centering Theory was developed and refined by several researcher

groups

  • Therefore no single version exists, but lots of different parameters and

principles were introduced to suit the individual purposes

  • Cheapness: (Strube & Hahn 1999) defined as CB(Un) = CP(Un − 1)
  • Meant to improve anaphora resolution, but also used by the authors for

calculating quality of ordering

  • Rule 2 was modified to always favor sequences that satisfy Cheapness

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 11 / 35

slide-22
SLIDE 22

Centering Theory Principles of Centering

Details of Centering II

  • Continuity and NOCB: CF(Un) should contain at least one member of

CF(Un − 1) resp. NOCB accounting for the case of Continuity violation

  • Coherence: subsumes Continue and Retain transitions and is satisfied

if backward looking center stays the same

  • Salience: in turn combines Continue and Smooth Shift and therefore

cases where the backward looking center is also the most highly ranked element

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 12 / 35

slide-23
SLIDE 23

Centering Theory Principles of Centering

An example

1 This cake is amazing and I am very

hungry!

2 A friend prepared it yesterday. 3 He is a good cook. 4 He also sometimes prepares delicious

pasta.

5 Unfortunately I’m allergic to pasta. 6 Green elephants eat breakfast. 1 CB(U) : n.a. 2 CB(U) : cake 3 CB(U) : friend 4 CB(U) : friend 5 CB(U) : pasta 6 CB(U) : n.a.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 13 / 35

slide-24
SLIDE 24

Centering Theory Parameters of Centering

Details of Centering III

One can also modify the way centering deals with the texts it’s applied on

  • Utterance: entire sentences, finite clauses or verbed clauses
  • Ranking: linear order, grammatical function or discourse status
  • Realization: direct realization (including anaphoras) or indirect via

bridging relationships. Example Most of my vacations I spend on Malta. I took this picture on the west coast of the island last year. It shows a beautiful sunset.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 14 / 35

slide-25
SLIDE 25

Centering Theory Parameters of Centering

An example

1 This cake is amazing and I am very

hungry!

2 A friend prepared this cheesecake

yesterday.

3 He is a good cook. 4 He also sometimes prepares delicious

fusilli.

5 Unfortunately I’m allergic to pasta. 6 Green elephants eat breakfast. 1 CB(U) : n.a. 2 CB(U) : cake 3 CB(U) : friend 4 CB(U) : friend 5 CB(U) : pasta 6 CB(U) : n.a.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 15 / 35

slide-26
SLIDE 26

Centering Theory Parameters of Centering

Centering Theory Exercise

Let’s practice that!

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 16 / 35

slide-27
SLIDE 27

Application on Information Ordering Motivation

Information Ordering with Centering

  • Information Ordering mainly relevant for text-to-text and concept-to-text

systems

  • Karamanis et al. only used pure centering based approaches
  • Originally for anaphora resolution, but Brennan, Friedman and Pollard

(BFP) also consider it relevant for text generation systems

  • Ordering follows principles of Rule 2 → Avoiding Shifts, prefering

Continues and Cheapness

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 17 / 35

slide-28
SLIDE 28

Application on Information Ordering Motivation

Information Ordering in Detail

Information Ordering workflow:

1 The module gets as input an (unordered) set of items (in this context CF

lists)

2 Produces several candidate orderings 3 Computes a score for each of them with a preselected metric based on

features from CT

4 The highest ranked item is then the optimal ordering and returned as

  • utput

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 18 / 35

slide-29
SLIDE 29

Application on Information Ordering The Metrics

The metrics

  • M.CHEAP: Number of violations against Cheapness. The less

violations the higher ranked the candidate will be.

  • M.BFP: Number of Continue transitions. If two top ranked candidates

exist, Retain etc. will be used.

  • M.NOCB: Sum of NOCBs. Like in M.CHEAP

, the less violations the higher ranked it will be.

  • M.KP: Violations against Cheapness, Coherence and Salience +

NOCBs

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 19 / 35

slide-30
SLIDE 30

Experimental Evaluation Test Data

Experimental Evaluation

Experiments on 3 different corpora with varying parameter settings The workflow:

  • Utterances converted into CF lists
  • CF lists of one segment were defined as OSO - Original Sentence

Ordering ...

  • ... and used as input for the information ordering algorithm
  • If any output got a higher ranked metric score than the OSO → metric

got penalized Classification Error Rate Better(M,OSO)+ Equal(M,OSO)/2

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 20 / 35

slide-31
SLIDE 31

Experimental Evaluation Test Data

MPIRO and MPIRO-CF

  • Dimitromanolaki & Androutsopoulos (D & A) derived facts form MPIRO

concept-to-text generation system

  • MPIRO was created in an effort to provide multilingual descriptions of

pictures of art

  • Organized in sentences and sets and ordered by a domain expert
  • Then the authors constructed a corpus consiting of 122 orderings on

each of which an instantiation of CT was applied.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 21 / 35

slide-32
SLIDE 32

Experimental Evaluation Test Data

MPIRO-CF Corpus

This exhibit is an amphora. This exhibit was decorated by the Painter of Kleofrades. The Painter of Kleofrades used to decorate big vases. ... CF : exhibit, amphora CB : n.a. Tr.: n.a. Cheap: n.a. CF : exhibit, Painter of K. CB : exhibit Tr.: Continue Cheap: yes ...

  • Used parameters:
  • utterance: each database fact
  • rank: linear order
  • realization: arguments of preceding fact

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 22 / 35

slide-33
SLIDE 33

Experimental Evaluation Test Data

Gnome-LAB Corpus

  • Poesio et al. (2004) created the GNOME corpus containing object

descriptions of museum pieces.

  • As opposed to MPIRO GNOME does not consist of artificial data but of

texts written by humans

  • More than one domain contained in original corpus, but only the

museum label domain was used

  • As the data is natural the syntax tends to be more complex than the

MPIRO examples

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 23 / 35

slide-34
SLIDE 34

Experimental Evaluation Test Data

Gnome-LAB Corpus

Item 144 is a torc Its present arrangement, twisted into three rings, may be a ...; it should probably be a single ring, worn around the neck. The terminals are in the form of goats’ heads.

  • Used parameters:
  • utterance: finite units (corpus annotated with DU)
  • rank: linear order & grammatical function
  • realization: direct realization

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 24 / 35

slide-35
SLIDE 35

Experimental Evaluation Test Data

NEWS and ACCS Corpus

  • 400 articles from the North American News Database and National

Transportation Safety Board

  • Corpus constructed by Barzilay & Lapata (2005) for evaluation of their

Entity Grid Model

  • Data was parsed in order to obtain constituent structure of the

sentences

  • Used parameters:
  • utterance: sentences
  • rank: linear order + grammatical function
  • realization: direct realization (allowing coreferential pairs like Microsoft

and the company)

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 25 / 35

slide-36
SLIDE 36

Experimental Evaluation Results

Classification Error Rate

MPIRO.CF GNOME.LAB NEWS ACCS MEAN

Corpora

Error Rate 20 40 60 80

M.NOCB M.BFP M.KP M.CHEAP

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 26 / 35

slide-37
SLIDE 37

Experimental Evaluation Results

Classification Error Rate - Analysis I

MPIRO.CF GNOME.LAB NEWS ACCS MEAN

Corpora

Error Rate 20 40 60 80

M.NOCB M.BFP M.KP M.CHEAP

Implications:

  • Why do M.CHEAP and M.KP perform

so poorly?

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 27 / 35

slide-38
SLIDE 38

Experimental Evaluation Results

Classification Error Rate - Analysis I

MPIRO.CF GNOME.LAB NEWS ACCS MEAN

Corpora

Error Rate 20 40 60 80

M.NOCB M.BFP M.KP M.CHEAP

Implications:

  • Why do M.CHEAP and M.KP perform

so poorly?

→ M.KP contains the sum of violations

against several different principles, namely Cheapness, Coherence, Salience and Continuity

→ M.CHEAP adds up the violations

against Cheapness. Strube and Hahn (1999) introduced it to further improve anaphora resolution.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 27 / 35

slide-39
SLIDE 39

Experimental Evaluation Results

Classification Error Rate - Analysis II

MPIRO.CF GNOME.LAB NEWS ACCS MEAN

Corpora

Error Rate 20 40 60 80

M.NOCB M.BFP M.KP M.CHEAP

Implications:

  • Why does M.BFP perform so much

better than M.KP?

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 28 / 35

slide-40
SLIDE 40

Experimental Evaluation Results

Classification Error Rate - Analysis II

MPIRO.CF GNOME.LAB NEWS ACCS MEAN

Corpora

Error Rate 20 40 60 80

M.NOCB M.BFP M.KP M.CHEAP

Implications:

  • Why does M.BFP perform so much

better than M.KP?

→ M.BFP doesn’t use several principles

for scoring an ordering, but compares the number of Continue transitions → correspond roughly to the concept of Continuity

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 28 / 35

slide-41
SLIDE 41

Experimental Evaluation Results

Classification Error Rate - Analysis III

MPIRO.CF GNOME.LAB NEWS ACCS MEAN

Corpora

Error Rate 20 40 60 80

M.NOCB M.BFP M.KP M.CHEAP

Implications:

  • ... but still has a higher error rate

than M.NOCB. Why?

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 29 / 35

slide-42
SLIDE 42

Experimental Evaluation Results

Classification Error Rate - Analysis III

MPIRO.CF GNOME.LAB NEWS ACCS MEAN

Corpora

Error Rate 20 40 60 80

M.NOCB M.BFP M.KP M.CHEAP

Implications:

  • ... but still has a higher error rate

than M.NOCB. Why?

→ M.NOCB solely uses the violations

against Continuity as measure for computing the metric

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 29 / 35

slide-43
SLIDE 43

Experimental Evaluation Results

Classification Error Rate - Conclusion

MPIRO.CF GNOME.LAB NEWS ACCS MEAN

Corpora

Error Rate 20 40 60 80

M.NOCB M.BFP M.KP M.CHEAP

Implications:

  • Strength of M.NOCB is further

backed up by doing pairwise tests against all other metrics

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 30 / 35

slide-44
SLIDE 44

Experimental Evaluation Results

Classification Error Rate - Conclusion

MPIRO.CF GNOME.LAB NEWS ACCS MEAN

Corpora

Error Rate 20 40 60 80

M.NOCB M.BFP M.KP M.CHEAP

Implications:

  • Strength of M.NOCB is further

backed up by doing pairwise tests against all other metrics

→ M.NOCB is significantly better in 10

  • ut of 12 cases

→ Additional to that M.NOCB is the

simplest of all used metrics

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 30 / 35

slide-45
SLIDE 45

Experimental Evaluation Conclusion

Conclusion

Questions to be answered

1 How appropriate is the Centering Theory for Information Ordering?

  • M.NOCB still has CER ranging from 15 % to 31 %
  • Continuity only partially models entity coherence
  • Other coherence-inducing factors need to be extracted

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 31 / 35

slide-46
SLIDE 46

Experimental Evaluation Conclusion

Conclusion

Questions to be answered

2 Which aspects of it are most useful?

  • Cheapness and the centering transitions didn’t prove to be of use for the

task of Information Ordering

  • Continuity however seems to play a far bigger role in entity coherence

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 31 / 35

slide-47
SLIDE 47

Experimental Evaluation Conclusion

Conclusion

Pure Centering based approach wasn’t able to built a tool that performs robustly and independently. M.NOCB can serve as a good baseline for other researchers to test their more elaborate metrics against.

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 32 / 35

slide-48
SLIDE 48

Experimental Evaluation Conclusion

The end ...

Thank you very much for your attention!

Any questions, feedback etc?

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 33 / 35

slide-49
SLIDE 49

Experimental Evaluation Conclusion

Exercise number 2 - Information Ordering

1 This cake is amazing and I am very

hungry!

2 A friend prepared this cheesecake

yesterday.

3 He is a good cook. 4 He also sometimes prepares delicious

fusilli.

5 Unfortunately I’m allergic to pasta. 6 Green elephants eat breakfast. 1

M.NOCB: 1

2

M.CHEAP: 1 (w/o NOCB)

3

M.KP: 9

4

M.BFP: 1 Continue

1: 3,1,4,6,5,2 2: 1,2,5,4,3,6 3: 6,5,4,3,2,1 4: 1,3,2,5,4,6 5: 1,2,3,4,6,5 6: 4,6,1,2,5,3 7: 6,5,4,3,2,1 8: 1,6,5,2,3,4

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 34 / 35

slide-50
SLIDE 50

Experimental Evaluation Conclusion

References

  • Evaluating Centering for Information Ordering Using Corpora,

Karamanis et al., 2008 (http://www.mitpressjournals.

  • rg/doi/pdf/10.1162/coli.07-036-R2-06-22)
  • Specifying the Parameters of Centering Theory: a Corpus-Based

Evaluation using Text from Application-Oriented Domains, Poesio et al., 2000 (http://acl.ldc.upenn.edu/P/P00/P00-1051.pdf)

  • Functional Centering – Grounding Referential Coherence in Information

Structure, Strube and Hahn, 1999 (http://dl.acm.org/citation.cfm?id=973328)

  • A Reformulation of Rule 2 of Centering Theory, Rodger Kibble, 2001

(http://acl.ldc.upenn.edu/J/J01/J01-4007.pdf)

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 35 / 35

slide-51
SLIDE 51

Experimental Evaluation Conclusion

Discussion

  • What could be other coherence inducing factors besides Continuity?

Jonathan Poitz (CoLi Saarland) Evaluating Centering for Information Ordering May 13th, 2013 36 / 35