text summarization dragomir r. radev school of information, department of electrical engineering and...

Text summarization

Dragomir R. RadevSchool of Information, Department of Electrical Engineering and

Computer Science, and Department of Linguistics

University of Michigan

http://www.si.umich.edu/~radev

TutorialACM SIGIR

New Orleans, LouisianaSeptember 9, 2001

Part IIntroduction

The BIG problem

• Information overload: 1.39 Billion URLs catalogued by Google

• Possible approaches:– information retrieval– document clustering– information extraction– visualization– question answering– text summarization

Some concepts

• Abstracts: “a concise summary of the central subject matter of a document” [Paice90].

• Indicative, informative, and critical summaries

• Extracts (representative sentences)

Informative summaries

. .

.

. . .

Lines sometimes blurred

Net Tax Moratorium Clears House

The House passed a bill to extend the current moratorium on new Internet taxes until 2006. The moratorium forbids states from trying to find new ways of taxing Internet use, like imposing taxes on monthly access charges for Internet service providers.

House Votes to Ban Internet Taxes for 5 More Years

By LIZETTE ALVAREZ

WASHINGTON, May 10 -- In a Republican bid to woo the high-technology industry and please taxpayers, the House today rushed to the floor and then handily passed a bill to extend the current moratorium on new Internet taxes until 2006.The moratorium, which is due to expire in October 2001, forbids states to try to find new ways of taxing Internet use, like imposing taxes on monthly access charges for Internet service providers.The legislation passed today, which faces an uncertain future in the Senate, does not directly address the question of sales taxes; it would not stop states from trying to collect taxes for goods sold on the Internet.By failing to address sales taxes, however, the measure alarmed some traditional retailers, as well as state governments that say they have found it nearly impossible to collect taxes for goods sold online."The single largest contributor to our economic prosperity has been the growth of information technology -- the Internet," saidRepresentative John R. Kasich, an Ohio Republican. "Why would we try to tax something, why would we try to abuse something, why would we try to limit something that generates unprecedented growth, wealth, opportunity and unprecedented individual power?"Critics of the bill say the moratorium, while seemingly benign, ignores the thorny question of how state and local governments can best collect taxes on the billions of dollars of merchandise sold over the Internet each year. These taxes are expected to provide a crucial future source of revenue for states, especially as more consumers buy goods online.The bill's opponents -- a consortium of retailers, small-business groups and governors -- say that consumers who buy merchandise over the Internet can easily circumvent the sales and "use" taxes that would be collected automatically if the same merchandise is bought at a bricks-and-mortar retail store.The National Governors' Association is working on the best way to collect electronic sales tax. Estimates have put the loss in sales tax revenue to the states at $8 billion a year by 2004.

http://www.nytimes.com/library/tech/00/05/biztech/articles/11tax.html

Retailers and small businesses have complained that the current system unfairly places at a disadvantage the traditional retailers that do not sell their wares online and must charge sales tax."It's easy to imagine how these kinds of losses can affect state and local governments' ability to provide essential services," said Representative William D. Delahunt, a Massachusetts Democrat, citing the concerns of many governors. "They will be compelled to cut back local services or raise income taxes or property taxes."The bill even drew criticism from a few Republicans. Representative Ernest J. Istook Jr. of Oklahoma circulated a letter stating, "The Internet should not be singled out to be taxed, nor to be freed from tax."Still, the House voted overwhelmingly, 352 to 75, to pass the bill. A number of Democrats approved the measure after they received assurance that Congress would hold hearings concerning sales taxes and would try to come up with a solution.The moratorium "has absolutely nothing to do with the sales tax -- we will have the opportunity to have that debate," said Representative Robert Goodlatte, a Virginia Republican.The House bill faces a murkier future in the Senate. Senator John McCain, chairman of the Commerce Committee, who advocates a permanent tax moratorium, canceled a hearing on the bill last month after Republican senators, some of them former governors, expressed reservations about extending the moratorium.The legislation also faces opposition from the Clinton administration, which signaled support today for a two-year moratorium. The full House today rejected a two-year extension in a separate vote.Gov. George W. Bush, the likely Republican presidential nominee, has said he will support an extension of the moratorium. But the governor must tread carefully around the issue because Texas, which does not have a state income tax, would stand to lose substantial revenue ifsales taxes are not made workable on the Internet.A spokesman for Al Gore said the vice president supported a two-year extension of the moratorium "at a minimum." If a five-year moratorium is put into place, "it should include flexibility" to adjust federal policies on Internet taxation "to take into account the fast-paced change in the Internet world.”

Types of summaries

• dimensions

• genres

• context

Dimensions

• Single-document vs. multi-document

Genres

• headlines• outlines• minutes• biographies• abridgments• sound bites• movie summaries• chronologies, etc.

[Mani and Maybury 1999]

Context

• Query-specific

• Query-independent

What does summarization involve?

• Three stages (typically)– content identification– conceptual organization– realization

Spärck Jones’s three sets of factors

• Input factors (source form, subject type, unit)

• Purpose factors (situation, audience, use)

• Output factors (material, format, style)

[Spärck Jones 99]

http://transend.labs.bt.com/prosum/word/index.html

ProSum

• Profile-based summarization

• Control of summarization length

• Retention of user-defined text

• Customizable heading treatment

• Customizable table treatment

• Customizable text differentiation

Example (New York Times)

Net Tax Moratorium Clears House

The House passed a bill to extend the current moratorium on new Internet taxes until 2006.The moratorium forbids states from trying to find new ways of taxing Internet use, like imposing taxes on monthly access charges for Internet service providers.


By LIZETTE ALVAREZ



Microsoft Autosummarize outputHouse Votes to Ban Internet Taxes for 5 More Years

The moratorium, which is due to expire in October 2001, forbids states to try to find new ways of taxing Internet use, like imposing taxes on monthly access charges for Internet service providers.By failing to address sales taxes, however, the measure alarmed some traditional retailers, as well as state governments that say they have found it nearly impossible to collect taxes for goods sold online.

10% summary


By LIZETTE ALVAREZ



Microsoft Autosummarize outputHouse Votes to Ban Internet Taxes for 5 More Years

The moratorium, which is due to expire in October 2001, forbids states to try to find new ways of taxing Internet use, like imposing taxes on monthly access charges for Internet service providers.The legislation passed today, which faces an uncertain future in the Senate, does not directly address the question of sales taxes; it would not stop states from trying to collect taxes for goods sold on the Internet.By failing to address sales taxes, however, the measure alarmed some traditional retailers, as well as state governments that say they have found it nearly impossible to collect taxes for goods sold online.The National Governors' Association is working on the best way to collect electronic sales tax. Representative Ernest J. Istook Jr. of Oklahoma circulated a letter stating, "The Internet should not be singled out to be taxed, nor to be freed from tax."Senator John McCain, chairman of the Commerce Committee, who advocates a permanent tax moratorium, canceled a hearing on the bill last month after Republican senators, some of them former governors, expressed reservations about extending the moratorium.

25% summary


By LIZETTE ALVAREZ



OutlineIntroduction

Traditional approaches

Multi-document summarization

Knowledge-rich techniques

Evaluation methods

The MEAD project

Language modeling

I

II

III

IV

V

VI

VII

Part II Traditional approaches

Human summarization and abstracting

• What professional abstractors do

• Ashworth:• “To take an original article, understand it

and pack it neatly into a nutshell without loss of substance or clarity presents a challenge which many have felt worth taking up for the joys of achievement alone. These are the characteristics of an art form”.

Borko and Bernier 75

• The abstract and its use:– Abstracts promote current awareness– Abstracts save reading time– Abstracts facilitate selection– Abstracts facilitate literature searches– Abstracts improve indexing efficiency– Abstracts aid in the preparation of

reviews

Cremmins 82, 96

• American National Standard for Writing Abstracts:– State the purpose, methods, results, and conclusions

presented in the original document, either in that order or with an initial emphasis on results and conclusions.

– Make the abstract as informative as the nature of the document will permit, so that readers may decide, quickly and accurately, whether they need to read the entire document.

– Avoid including background information or citing the work of others in the abstract, unless the study is a replication or evaluation of their work.

Cremmins 82, 96

– Do not include information in the abstract that is not contained in the textual material being abstracted.

– Verify that all quantitative and qualitative information used in the abstract agrees with the information contained in the full text of the document.

– Use standard English and precise technical terms, and follow conventional grammar and punctuation rules.

– Give expanded versions of lesser known abbreviations and acronyms, and verbalize symbols that may be unfamiliar to readers of the abstract.

– Omit needless words, phrases, and sentences.

Cremmins 82, 96• Original version:

There were significant positive associations between the concentrations of the substance administered and mortality in rats and mice of both sexes.

There was no convincing evidence to indicate that endrin ingestion induced and of the different types of tumors which were found in the treated animals.

• Edited version:

Mortality in rats and mice of both sexes was dose related.

No treatment-related tumors were found in any of the animals.

Redundancy of English

• 75% redundancy of English [Shannon 51]

• [Burton & Licklider 55] show that humans are as good at guessing the next letter after seeing 32 letters as after 10,000 letters.

Morris et al. 92

• Reading comprehension of summaries• Compare manual abstracts, Edmundson-

style extracts, and full documents• Extracts containing 20% or 30% of original

document are effective surrogates of original document

• Performance on 20% and 30% extracts is no different than informative abstracts

Extraction models

• Extracts vs. abstracts

• Linear model• Text structure

based• New techniques

Compression Ratio =|S|

|D|

Retention Ratio =i (S)

i (D)

Information content

Text compaction techniquesMissam ad amicum pro onsolatione epistolam, dilectissime, vestram ad me forte quidam nuper attulit.

Quam ex ipsa statim tituli fronte vestram esse considerans, tanto ardentius eam cepi legere quanto scriptorem ipsum karius amplector, ut cuius rem perdidi verbis saltem tanquam eius quadam imagine recreer.

Erant, memini, huius epistole fere omnia felle et absintio plena, que scilicet nostre conversionis miserabilem hystoriam et tuas, unice, cruces assiduas referebant.

Complesti revera in epistola illa quod in exordio eius amico promisisti, ut videlicet in omparatione tuarum suas molestias nullas vel parvas reputaret; ubi quidem expositis prius magistrorum tuorum in te persequutionibus, deinde in corpus tuum summe proditionis iniuria, ad condiscipulorum quoque tuorum Alberici videlicet Remensis et Lotulfi Lumbardi execrabilem invidiam et infestationem nimiam stilum contulisti.

Missam ad amicum pro onsolatione epistolam, dilectissime, vestram ad me forte quidam nuper attulit.


Text compaction techniques

Missam ad amicum pro onsolatione epistolam, dilectissime, vestram ad me forte quidam nuper attulit.


Missam vestram nuper attulit.

Erant, scilicet nostre conversionis miserabilem hystoriam referebant.

Luhn 58

• Very first work in automated summarization

• Computes measures of significance

• Words:– stemming– bag of words

WORDSF

RE

QU

EN

CY

E

Resolving power of significant words

Luhn 58

• Sentences:– concentration of

high-score words

• Cutoff values established in experiments with 100 human subjects

SIGNIFICANT WORDS

ALL WORDS

* * * * 1 2 3 4 5 6 7

SENTENCE

SCORE = 42/7 2.3

Edmundson 69

• Cue method:– stigma words

(“hardly”, “impossible”)

– bonus words (“significant”)

• Key method:– similar to Luhn

• Title method:– title + headings

• Location method:– sentences under

headings– sentences near

beginning or end of document and/or paragraphs (also [Baxendale 58])

Edmundson 69

• Linear combination of four features:

1C + 2K + 3T + 4L

• Manually labelled training corpus

• Key not important!0 10 20 30 40 50 60 70 80 90 100 %

RANDOM

KEY

TITLE

CUE

LOCATION

C + K + T + L

C + T + L

1

Paice 90

• Survey up to 1990• Techniques that

(mostly) failed:– syntactic criteria

[Earl 70]– indicator phrases

(“The purpose of this article is to review…)

• Problems with extracts:– lack of balance– lack of cohesion

• anaphoric reference• lexical or definite

reference• rhetorical

connectives

Paice 90

• Lack of balance– later approaches

based on text rhetorical structure

• Lack of cohesion– recognition of

anaphors [Liddy et al. 87]

• Example: “that” is– nonanaphoric if

preceded by a research-verb (e.g., “demonstrat-”),

– nonanaphoric if followed by a pronoun, article, quantifier,…,

– external if no later than 10th word,else

– internal

Brandow et al. 95

• ANES: commercial news from 41 publications

• “Lead” achieves acceptability of 90% vs. 74.4% for “intelligent” summaries

• 20,997 documents• words selected

based on tf*idf• sentence-based

features:– signature words– location– anaphora words– length of abstract

Brandow et al. 95

• Sentences with no signature words are included if between two selected sentences

• Evaluation done at 60, 150, and 250 word length

• Non-task-driven evaluation:

“Most summaries judged less-than-perfect would not be detectable as such to a user”

Lin & Hovy 97

• Optimum position policy

• Measuring yield of each sentence position against keywords (signature words) from Ziff-Davis corpus

• Preferred order

[(T) (P2,S1) (P3,S1) (P2,S2) {(P4,S1) (P5,S1) (P3,S2)} {(P1,S1) (P6,S1) (P7,S1) (P1,S3)(P2,S3) …]

Kupiec et al. 95

• Extracts of roughly 20% of original text

• Feature set:– sentence length

• |S| > 5

– fixed phrases• 26 manually chosen

– paragraph• sentence position in

paragraph

– thematic words• binary: whether

sentence is included in manual extract

– uppercase words• not common

acronyms

• Corpus:• 188 document +

summary pairs from scientific journals

Kupiec et al. 95

• Uses Bayesian classifier:

• Assuming statistical independence:

k

j j

k

j j

kFP

SsPSsFPFFFSsP

1

121

)(

)()|(),...,|(

),(

)()|,...,(),...,|(

,...21

2121

k

kk FFFP

SsPSsFFFPFFFSsP

Kupiec et al. 95

• Performance:– For 25% summaries, 84% precision– For smaller summaries, 74%

improvement over Lead

Salton et al. 97

• document analysis based on semantic hyperlinks (among pairs of paragraphs related by a lexical similarity significantly higher than random)

• Bushy paths (or paths connecting highly connected paragraphs) are more likely to contain information central to the topic of the article

Salton et al. 97

…

…

Salton et al. 97

Overlap between manual extracts: 46%Algorithm Optimistic Pessimistic Intersection Union

Globalbushy

45.60% 30.74% 47.33% 55.16%

Globaldepth-first

43.98% 27.76% 42.33% 52.48%

Segmentedbushy

45.48% 26.37% 38.17% 52.95%

Random 39.16% 22.07% 38.47% 44.24%

Marcu 97-99

• Based on RST (nucleus+satellite relations)

• text coherence• 70% precision and

recall in matching the most important units in a text

• Example: evidence

[The truth is that the pressure to smoke in junior high is greater than it will be any other time of one’s life:][we know that 3,000 teens start smoking each day.]

• N+S combination increases R’s belief in N [Mann and Thompson 88]

2Elaboration

2Elaboration

8Example

2BackgroundJustification

3Elaboration

8Concession

10Antithesis

Mars experiences

frigid weather

conditions(2)

Surface temperatures typically average

about -60 degrees

Celsius (-76 degrees

Fahrenheit) at the

equator and can dip to -

123 degrees C near the

poles(3)

4 5Contrast

Although the atmosphere

holds a small

amount of water, and water-ice

clouds sometimes develop,

(7)

Most Martian weather involves

blowing dust and carbon monoxide.

(8)

Each winter, for example, a blizzard of

frozen carbon dioxide

rages over one pole, and a few meters of

this dry-ice snow

accumulate as

previously frozen carbon dioxide

evaporates from the opposite

polar cap.(9)

Yet even on the summer pole, where

the sun remains in the sky all day long,

temperatures never warm

enough to melt frozen

water.(10)

With its distant orbit (50 percent farther from the sun than Earth) and

slim atmospheric

blanket,(1)

Only the midday sun at tropical latitudes is

warm enough to

thaw ice on occasion,

(4)

5Evidence

Cause

but any liquid water formed in this way would

evaporate almost

instantly(5)

because of the low

atmospheric pressure

(6)

Barzilay and Elhadad 97

• Lexical chains [Stairmand 96]

Mr. Kenny is the person that invented the anesthetic machine which uses micro-computers to control the rate at which an anesthetic is pumped into the blood. Such machines are nothing new. But his device uses two micro-computers to achineve much closer monitoring of the pump feeding the anesthetic into the patient.


• WordNet-based

• three types of relations:– extra-strong (repetitions)– strong (WordNet relations)– medium-strong (link between synsets is

longer than one + some additional constraints)


• Scoring chains:– Length– Homogeneity index:

= 1 - # distinct words in chain

Score = Length * Homogeneity

Score > Average + 2 * st.dev.

Other approaches

• Salience-based [Boguraev and Kennedy 97]

• Computational linguistics papers [Teufel and Moens 97]

Part III Multi-document summarization

Mani & Bloedorn 97,99

• Summarizing differences and similarities across documents

• Single event or a sequence of events

• Text segments are aligned

• Evaluation: TREC relevance judgments

• Significant reduction in time with no significant loss of accuracy

Carbonell & Goldstein 98

• Maximal Marginal Relevance (MMR)

• Query-based summaries

• Law of diminishing returns

C = doc collectionQ = user queryR = IR(C,Q,)S = already retrieved

documentsSim = similarity

metric used

MMR = argmax [ (Sim1(Di,Q) - (1-) max Sim2(Di,Dj)]DiR\S DiS

Radev et al. 00

• MEAD• Centroid-based• Based on sentence

utility

• Topic detection and tracking initiative [Allen et al. 98, Wayne 98]

TIME

1. Algerian newspapers have reported that 18 decapitated bodies have been found by authorities in the south of the country.

2. Police found the ``decapitated bodies of women, children and old men,with their heads thrown on a road'' near the town of Jelfa, 275 kilometers (170 miles) south of the capital Algiers.

3. In another incident on Wednesday, seven people -- including six children -- were killed by terrorists, Algerian security forces said.

4. Extremist Muslim militants were responsible for the slaughter of the seven people in the province of Medea, 120 kilometers (74 miles) south of Algiers.

5. The killers also kidnapped three girls during the same attack, authorities said, and one of the girls was found wounded on a nearby road.

6. Meanwhile, the Algerian daily Le Matin today quoted Interior Minister Abdul Malik Silal as saying that ``terrorism has not been eradicated, but the movement of the terrorists has significantly declined.''

7. Algerian violence has claimed the lives of more than 70,000 people since the army cancelled the 1992 general elections that Islamic parties were likely to win.

8. Mainstream Islamic groups, most of which are banned in the country, insist their members are not responsible for the violence against civilians.

9. Some Muslim groups have blamed the army, while others accuse ``foreign elements conspiring against Algeria.’’

1. Eighteen decapitated bodies have been found in a mass grave in northern Algeria, press reports said Thursday, adding that two shepherds were murdered earlier this week.

2. Security forces found the mass grave on Wednesday at Chbika, near Djelfa, 275 kilometers (170 miles) south of the capital.

3. It contained the bodies of people killed last year during a wedding ceremony, according to Le Quotidien Liberte.

4. The victims included women, children and old men.

5. Most of them had been decapitated and their heads thrown on a road, reported the Es Sahafa.

6. Another mass grave containing the bodies of around 10 people was discovered recently near Algiers, in the Eucalyptus district.

7. The two shepherds were killed Monday evening by a group of nine armed Islamists near the Moulay Slissen forest.

8. After being injured in a hail of automatic weapons fire, the pair were finished off with machete blows before being decapitated, Le Quotidien d'Oran reported.

9. Seven people, six of them children, were killed and two injured Wednesday by armed Islamists near Medea, 120 kilometers (75 miles) south of Algiers, security forces said.

10. The same day a parcel bomb explosion injured 17 people in Algiers itself.

11. Since early March, violence linked to armed Islamists has claimed more than 500 lives, according to press tallies.

ARTICLE 18854: ALGIERS, May 20 (UPI) ARTICLE 18853: ALGIERS, May 20 (AFP)

Vector-based representation

Term 1

Term 2

Term 3

Document

Centroid

Vector-based matching

• The cosine measure

kk

kk

k kk

cd

kidfcdCDsim

22 .

)(..),(

CIDR

sim T

sim < T

CentroidsC 00022 (N=44)

(10000)diana 1.93princess 1.52

C 00025 (N=19)(10000)albanians 3.00

C 00026 (N=10)(10000)universe 1.50

expansion 1.00bang 0.90

C 10007 (N=11)(10000)crashes 1.00

safety 0.55transportat

ion0.55

drivers 0.45board 0.36flight 0.27buckle 0.27

pittsburgh 0.18graduating 0.18automobile 0.18

C 00035 (N=22)(10000)airlines 1.45

finnair 0.45

C 00031 (N=34)(10000)el 1.85

nino 1.56

C 00008 (N=113)(10000)space 1.98

shuttle 1.17station 0.75nasa 0.51

columbia 0.37mission 0.33mir 0.30

astronauts

0.14steering 0.11safely 0.07

C 10062 (N=161)microsoft 3.24justice 0.93

department

0.88windows 0.98corp 0.61

software 0.57ellison 0.07hatch 0.06

netscape 0.04metcalfe 0.02

MEAD

...

...

MEAD

• INPUT: Cluster of d documents with n sentences (compression rate = r)

• OUTPUT: (n * r) sentences from the cluster with the highest values of SCORE

SCORE (s) = i (wcCi + wpPi + wfFi)

[Barzilay et al. 99]

• Theme intersection (paraphrases)

• Identifying common phrases across multiple sentences:– evaluated on 39 sentence-level

predicate-argument structures– 74% of p-a structures automatically

identified

Other multi-document approaches

• Reformulation [McKeown et al. 99]

• Generation by Selection and Repair [DiMarco et al. 97]

• Topic and event distinctions [Fukumoto & Suzuki 00]

Part IV Knowledge-rich

approaches

Overview

• Schank and Abelson 77– scripts

• DeJong 79– FRUMP (slot-filling from UPI news)

• Graesser 81– Ratio of inferred propositions to these

explicitly stated is 8:1

• Young & Hayes 85– banking telexes

Radev and McKeown 98

MESSAGE: ID TST3-MUC4-0010 MESSAGE: TEMPLATE 2 INCIDENT: DATE 30 OCT 89 INCIDENT: LOCATION EL SALVADOR INCIDENT: TYPE ATTACK INCIDENT: STAGE OF EXECUTION ACCOMPLISHED INCIDENT: INSTRUMENT ID INCIDENT: INSTRUMENT TYPEPERP: INCIDENT CATEGORY TERRORIST ACT PERP: INDIVIDUAL ID "TERRORIST" PERP: ORGANIZATION ID "THE FMLN" PERP: ORG. CONFIDENCE REPORTED: "THE FMLN" PHYS TGT: ID PHYS TGT: TYPEPHYS TGT: NUMBERPHYS TGT: FOREIGN NATIONPHYS TGT: EFFECT OF INCIDENTPHYS TGT: TOTAL NUMBERHUM TGT: NAMEHUM TGT: DESCRIPTION "1 CIVILIAN"HUM TGT: TYPE CIVILIAN: "1 CIVILIAN"HUM TGT: NUMBER 1: "1 CIVILIAN"HUM TGT: FOREIGN NATIONHUM TGT: EFFECT OF INCIDENT DEATH: "1 CIVILIAN"HUM TGT: TOTAL NUMBER

Generating text from templates

On October 30, 1989, one civilian was killed in a reported FMLN attack in El Salvador.

Input: Cluster of templates

T1 Tm

Conceptual combiner

T2 …..

Combiner

Paragraph planner

Planningoperators

Linguistic realizer

Sentence planner

Sentence generator

Lexical chooserLexicon

OUTPUT: Base summary

SURGE

Domainontology

Excerpts from four articles

JERUSALEM - A Muslim suicide bomber blew apart 18 people on a Jerusalem bus and wounded 10 in a mirror-image of an attack one week ago. The carnage could rob Israel's Prime Minister Shimon Peres of the May 29 election victory he needs to pursue Middle East peacemaking. Peres declared all-out war on Hamas but his tough talk did little to impress stunned residents of Jerusalem who said the election would turn on the issue of personal security.

JERUSALEM - A bomb at a busy Tel Aviv shopping mall killed at least 10 people and wounded 30, Israel radio said quoting police. Army radio said the blast was apparently caused by a suicide bomber. Police said there were many wounded.

A bomb blast ripped through the commercial heart of Tel Aviv Monday, killing at least 13 people and wounding more than 100. Israeli police say an Islamic suicide bomber blew himself up outside a crowded shopping mall. It was the fourth deadly bombing in Israel in nine days. The Islamic fundamentalist group Hamas claimed responsibility for the attacks, which have killed at least 54 people. Hamas is intent on stopping the Middle East peace process. President Clinton joined the voices of international condemnation after the latest attack. He said the ``forces of terror shall not triumph'' over peacemaking efforts.

TEL AVIV (Reuter) - A Muslim suicide bomber killed at least 12 people and wounded 105, including children, outside a crowded Tel Aviv shopping mall Monday, police said. Sunday, a Hamas suicide bomber killed 18 people on a Jerusalem bus. Hamas has now killed at least 56 people in four attacks in nine days. The windows of stores lining both sides of Dizengoff Street were shattered, the charred skeletons of cars lay in the street, the sidewalks were strewn with blood. The last attack on Dizengoff was in October 1994 when a Hamas suicide bomber killed 22 people on a bus.

1

2

3

4

Four templates

MESSAGE: ID TST-REU-0001 SECSOURCE: SOURCE Reuters SECSOURCE: DATE March 3, 1996 11:30 PRIMSOURCE: SOURCE INCIDENT: DATE March 3, 1996 INCIDENT: LOCATION Jerusalem INCIDENT: TYPE Bombing HUM TGT: NUMBER “killed: 18'' “wounded: 10” PERP: ORGANIZATION ID

MESSAGE: ID TST-REU-0002 SECSOURCE: SOURCE Reuters SECSOURCE: DATE March 4, 1996 07:20 PRIMSOURCE: SOURCE Israel Radio INCIDENT: DATE March 4, 1996 INCIDENT: LOCATION Tel Aviv INCIDENT: TYPE Bombing HUM TGT: NUMBER “killed: at least 10'' “wounded: more than 100” PERP: ORGANIZATION ID

MESSAGE: ID TST-REU-0003 SECSOURCE: SOURCE Reuters SECSOURCE: DATE March 4, 1996 14:20 PRIMSOURCE: SOURCE INCIDENT: DATE March 4, 1996 INCIDENT: LOCATION Tel Aviv INCIDENT: TYPE Bombing HUM TGT: NUMBER “killed: at least 13'' “wounded: more than 100” PERP: ORGANIZATION ID “Hamas”

MESSAGE: ID TST-REU-0004 SECSOURCE: SOURCE Reuters SECSOURCE: DATE March 4, 1996 14:30 PRIMSOURCE: SOURCE INCIDENT: DATE March 4, 1996 INCIDENT: LOCATION Tel Aviv INCIDENT: TYPE Bombing HUM TGT: NUMBER “killed: at least 12'' “wounded: 105” PERP: ORGANIZATION ID

43

21

Fluent summary with comparisons

Reuters reported that 18 people were killed on Sunday in a bombing in Jerusalem. The next day, a bomb in Tel Aviv killed at least 10 people and wounded 30 according to Israel radio. Reuters reported that at least 12 people were killed and 105 wounded in the second incident. Later the same day, Reuters reported that Hamas has claimed responsibility for the act.

(OUTPUT OF SUMMONS)

Operators

• If there are two templatesAND

the location is the sameAND

the time of the second template is after the time of the first template

ANDthe source of the first template is different from the source of the second template

ANDat least one slot differs

THENcombine the templates using the contradiction operator...

Operators: Change of Perspective

Change of perspective

March 4th, Reuters reported that a bomb in Tel Aviv killed at least 10 people and wounded 30. Later the same day, Reuters reported that exactly 12 people were actually killed and 105 wounded.

Precondition:The same source reports a change in a small number of slots

Operators: Contradiction

Contradiction

The afternoon of February 26, 1993, Reuters reported that a suspected bomb killed at least six people in the World Trade Center. However, Associated Press announced that exactly five people were killed in the blast.

Precondition:Different sources report contradictory values for a small number of slots

Operators: Refinement and Agreement

Refinement

On Monday morning, Reuters announced that a suicide bomber killed at least 10 people in Tel Aviv. In the afternoon, Reuters reported that Hamas claimed responsibility for the act.

Agreement

The morning of March 1st 1994, both UPI and Reuters reported that a man was kidnapped in the Bronx.

Operators: Generalization

Generalization

According to UPI, three terrorists were arrested in Medellín last Tuesday. Reuters announced that the police arrested two drug traffickers in Bogotá the next day.

A total of five criminals were arrested in Colombia last week.

Other conceptual methods

• Operator-based transformations using terminological knowledge representation [Reimer and Hahn 97]

• Topic interpretation [Hovy and Lin 98]

Part V Evaluation techniques

Overview of techniques

• Extrinsic techniques (task-based)

• Intrinsic techniques

• Can you recreate what’s in the original? – the Shannon Game [Shannon 1947–50].

– but often only some of it is really important. • Measure info retention (number of keystrokes):

– 3 groups of subjects, each must recreate text:• group 1 sees original text before starting. • group 2 sees summary of original text before

starting. • group 3 sees nothing before starting.

• Results (# of keystrokes; two different paragraphs):

Group 1 Group 2 Group 3approx. 10 approx. 150 approx. 1100

Hovy 98

• Burning questions:1. How do different evaluation methods compare for each type of summary? 2. How do different summary types fare under different methods? 3. How much does the evaluator affect things?

4. Is there a preferred evaluation method? Shannon Q&A

Original 1 1 1 1 1

Abstract Background 1 3 1 1 1Just-the-News 3 1 1 1

Regular 1 2 1 1 1Extract Keywords 2 4 1 1 1

Random 3 1 1 1

No Text 3 5

1-2: 50% 1-2: 30%2-3: 50% 2-3: 20%

3-4: 20%4-5:100%

Classification

Hovy 98

• Small Experiment– 2 texts, 7 groups.

• Results:– No difference!– As other

experiment…– ? Extract is best?

Precision and Recall

Relevant Non-relevant

System:relevant

A BSystem:

non-relevantC D

Precision and Recall

CA

A R

:Recall

BA

A P

:Precision

)(

2

RP

PRF

Jing et al. 98

• Small experiment with 40 articles

• When summary length is given, humans are pretty consistent in selecting the same sentences

• Percent agreement

• Different systems achieved maximum performance at different summary lengths

• Human agreement higher for longer summaries

SUMMAC [Mani et al. 98]

• 16 participants• 3 tasks:

– ad hoc: indicative, user-focused summaries

– categorization: generic summaries, five categories

– question-answering

• 20 TREC topics• 50 documents per

topic (short ones are omitted)

SUMMAC [Mani et al. 98]

• Participants submit a fixed-length summary limited to 10% and a “best” summary, not limited in length.

• variable-length summaries are as accurate as full text

• over 80% of summaries are intelligible

• technologies perform similarly

Goldstein et al. 99

• Reuters, LA Times• Manual summaries• Summary length

rather than summarization ratio is typically fixed

• Normalized version of R & F.

C)B,A(A

A R'

min

)R(P

PR F

'

''

2

Goldstein et al. 99

b)(

bp p'

1

)(

b)(g

gs

g

gs

)()(

'

''

• How to measure relative performance?

p = performanceb = baselineg = “good” systems = “superior” system

Radev et al. 00

---S10

---S9

---S8

---S7

---S6

---S5

+--S4

---S3

+++S2

-++S1

System 2System 1Ideal

Cluster-Based Sentence Utility

Cluster-Based Sentence Utility

---S10

---S9

---S8

---S7

---S6

---S5

+--S4

---S3

+++S2

-++S1


9(+)67S4

432S3

8(+)9(+)8(+)S2

510(+)10(+)S1


Summary sentence extraction method

CBSU method

CBSU(system, ideal)= % of ideal utility covered by system summary

Interjudge agreement

Judge1 Judge2 Judge3

Sentence 1 10 10 5

Sentence 2 8 9 8

Sentence 3 2 3 4

Sentence 4 5 6 9

Relative utility


Sentence 1 10 10 5

Sentence 2 8 9 8

Sentence 3 2 3 4

Sentence 4 5 6 9

RU =

Relative utility


Sentence 1 10 10 5

Sentence 2 8 9 8

Sentence 3 2 3 4

Sentence 4 5 6 9

17

RU =

Relative utility


Sentence 1 10 10 5

Sentence 2 8 9 8

Sentence 3 2 3 4

Sentence 4 5 6 9

1317

RU =

= 0.765

Normalized System Performance

1.000

0.765

0.765

Judge 3

0.7560.7890.722Judge 3

0.8831.0001.000Judge 2

0.8831.0001.000Judge 1

AverageJudge 2Judge 1

D = (S-R)

(J-R)

System performance

Interjudge agreement

Normalized system performance Random performance

Random Performance

D = (S-R)

(J-R)

Random Performance

D = (S-R)

(J-R)

n !

( n(1-r))! (r*n)!systemsaverage of all

Random Performance

D = (S-R)

(J-R)

n !

( n(1-r))! (r*n)!systemsaverage of all

{12}{13}{14}{23}{24}{34}

Examples

0.833 - 0.732

0.841 - 0.732= 0.927D {14} =

(S-R)

(J-R)=

Examples

0.833 - 0.732

0.841 - 0.732= 0.927D {14} =

(S-R)

(J-R)=

0.963D {24} =

1.0

J = 0.841

0.5

0.0

J’ = 1.0

0.5

R’= 0.0

R = 0.732

S = 0.833

S’ = 0.927 = D

Normalized evaluation of {14}

Cross-sentence Informational Subsumption and Equivalence

• Subsumption: If the information content of sentence a (denoted as I(a)) is contained within sentence b, then a becomes informationally redundant and the content of b is said to subsume that of a:

I(a) I(b)

• Equivalence: If I(a) I(b) I(b) I(a)

Example

(1) John Doe was found guilty of the murder.

(2) The court found John Doe guilty of the murder of Jane Doe last August and sentenced him to life.

Cross-sentence Informational Subsumption

967S4

432S3

898S2

51010S1

Article 3Article 2Article 1

Toxic spill in SpainAP, NYTTDT-3 corpus, topic 67833F

General strike in Denmark

AP, PRI, VOATDT-3 corpus, topic 7815110E

Explosion in a Moscow apartment building (Sept. 13, 1999)

AP, AFP, UPIclari.world.europe.russia1897D

Explosion in a Moscow apartment building (Sept. 9, 1999)AP, AFPclari.world.europe.russia652C

The FBI puts Osama bin Laden on the most wanted listAFP, UPIclari.world.terrorism453B

Algerian terrorists threaten BelgiumAFP, UPI

clari.world.africa.northwestern252A

topicnews sourcessource

# sents

# docsCluster

Evaluation

0.75

0.8

0.85

0.9

0.95

1

10 20 30 40 50 60 70 80 90 100

Compression rate (r)

Ag

ree

me

nt

(J) Cluster A

Cluster B

Cluster C

Cluster D

Cluster E

Cluster F

Inter-judge agreementversus compression

4A2-8----A1-7

4A2-7----A1-6

2A2-4A2-2-A2-1-A1-5

4A2-10-A2-10A2-10A2-10A1-4

4A2-10----A1-3

3A2-5--A2-5A2-5A1-2

3A2-1-A2-1A2-1-A1-1

- score+ scoreJudge5

Judge4

Judge3

Judge2

Judge1

Sent

Evaluating Sentence Subsumption

Subsumption (Cont’d)

SCORE (s) = i (wcCi + wpPi + wfFi) - wRRs

Rs = cross-sentence word overlap

Rs = 2 * (# overlapping words) / (# words in sentence 1 + # words in sentence 2)

wR = Maxs (SCORE(s))

Subsumption analysis

0107070112112

7323520284454633

11035837910163614

610731880450240705

-+-+-+-+-+-+#judges agreeing

Cluster F

Cluster E

Cluster D

Cluster C

Cluster B

Cluster A

Total: 558 sentences, full agreement on 292 (1+291), partial on 406 (23+383)Of 80 sentences with some indication of subsumption, only 24 had agreement of 4 or more judges.

Results

10% 20% 30% 40% 50% 60% 70% 80% 90%

Cluster A 0.855 0.572 0.427 0.759 0.862 0.910 0.554 1.001 0.584

Cluster B 0.365 0.402 0.690 0.714 0.867 0.640 0.845 0.713 1.317

Cluster C 0.753 0.938 0.841 1.029 0.751 0.819 0.595 0.611 0.683

Cluster D 0.739 0.764 0.683 0.723 0.614 0.568 0.668 0.719 1.100

Cluster E 1.083 0.937 0.581 0.373 0.438 0.369 0.429 0.487 0.261

Cluster F 1.064 0.893 0.928 1.000 0.732 0.805 0.910 0.689 0.199

MEAD performed better than Lead in 29 (in bold) out of 54 cases.

MEAD+Lead performed better than the Lead baseline in 41 cases

Donaway et al. 00

• Sentence-rank based measures– IDEAL={2,3,5}:

compare {2,3,4} and {2,3,9}

• Content-based measures– vector comparisons of summary and

document

Proposed TIDES evaluation

• Creation of corpora

• Development of evaluation software

• TREC-style evaluation

• Intrinsic and extrinsic evaluations

• Multilingual summaries (over time)

• Question-answering evaluation

Part VIIThe MEAD project

Background

• Summer 2001

• Eight weeks

• Johns Hopkins University• Participants: Dragomir Radev, Simone

Teufel, Horacio Saggion, Wai Lam, Elliott Drabek, Hong Qi, Danyu Liu, John Blitzer, and Arda Çelebi

Technical objectives

• Develop a summarization toolkit including a modular state-of-the art summarizer: single-document, multi-document, generic, query-based

• Develop a summarization evaluation toolkit allowing comparisons between extractive and non-extractive summaries

• Produce an annotated corpus for further research in text summarization

Sample scenarios

• Evaluate an existing summarizer

• Build a summarizer from scratch

• Test a summarization feature

• Test a new evaluation metric

• Test a machine translation system

Resources

• manual summaries (extracts and abstracts)• baseline summaries• automatic summaries• manual and automatic relevance judgements• XREF, lemmatized, tagged versions of the corpus• manual and automatic query translations• sentence segmentation• sentence alignments• XML DTDs, converters• subsumption judgements• guidelines for judges• guidelines for building summarizers• evaluation software• modular, trainable summarizer

<?xml version='1.0'?><!DOCTYPE QUERY SYSTEM "../../../dtd/query.dtd" ><QUERY QID="Q-241-E" QNO="241" TRANSLATED="NO"><TITLE>Fire safety, building management concerns</TITLE></QUERY>

<?xml version='1.0'?><!DOCTYPE QUERY SYSTEM “../../../dtd/query.dtd" ><QUERY QID="Q-241-C" QNO="241" TRANSLATED="NO"><TITLE>¨¾¤õ·NÃÑ,¤j·HºÞ²z</TITLE></QUERY>

Sample Chinese Query

Sample English Query

Sample Retrieval Result for Full-length Documents

<?xml version='1.0'?><!DOCTYPE DOC-JUDGE SYSTEM "/export/ws01summ/dtd/docjudge.dtd" ><DOC-JUDGE QID="Q-241-E" SYSTEM="SMART" LANG="ENG"> <D DID="D-20000126_008.e" RANK="1" SCORE="135.0000" CORR-DOC="D-20000126_012.c"/> <D DID="D-19980625_007.e" RANK="2" SCORE="99.0000" CORR-DOC="D-19980625_006.c"/> <D DID="D-19990126_017.e" RANK="3" SCORE="98.0000" CORR-DOC="D-19990126_018.c"/> <D DID="D-19981007_018.e" RANK="4" SCORE="91.0000" CORR-DOC="D-19981007_023.c"/> <D DID="D-19980121_004.e" RANK="5" SCORE="78.0000" CORR-DOC="D-19980121_009.c"/> <D DID="D-19971016_004.e" RANK="6" SCORE="72.0000" CORR-DOC="D-19971016_005.c"/>

Sample Retrieval Result for Lead-Based Summary (5%)

<?xml version='1.0'?><!DOCTYPE DOC-JUDGE SYSTEM"/export/ws01summ/dtd/docjudge.dtd" ><DOC-JUDGE QID="Q-241-E" SYSTEM="SMART" LANG="ENG"> <D DID="D-20000126_008.e" RANK="1" SCORE="14.0000" CORR-DOC="D-20000126_012.c"/> <D DID="D-19991214_002.e" RANK="2" SCORE="11.0000" CORR-DOC="D-19991214_001.c"/> <D DID="D-19980810_006.e" RANK="3" SCORE="10.0000" CORR-DOC="D-19980810_003.c"/> <D DID="D-19990505_028.e" RANK="4" SCORE="9.0000" CORR-DOC="D-19990505_034.c"/> <D DID="D-19980115_009.e" RANK="4" SCORE="9.0000" CORR-DOC="D-19980115_013.c"/>:

querySMART

LDC Judges

Rankeddocumentlist

Rankeddocumentlist

IR results

document

Summarycomparison

Correlation

Summarizer

Baselines

Single-document situation

Extract

1. Co-selection2. Similarity

LDC Judges

Summarycomparison

Manual sum.

Summarizer

Baselines

documentcluster

Multi-document situation

1. Co-selection2. Similarity

Extracts

Summaries produced

• Single-document extracts– automatic (135 runs on 18,146 documents

each): 10 compression rates, Word/Sentence, English/Chinese/Xlingual, 10 summarization methods

– manual (80 runs on 200 documents each): 10 compression rates, Word/Sentence, (3 judges + average)

Summaries produced

• Multi-document summaries– 3 lengths, 3 judges, 14 queries (out of 40)

• Multi-document extracts– automatic (160 extracts) = 8 compression rates

(5-40%,50-200AW) x 20 clusters– manual (320 extracts) = 8 compression rates x

10 clusters x (3 judges + average)

List of summarizers

• MEAD, Websumm, Summarist, LexChains, Align

• English, Chinese

• Single-document, Multi-document

MEAD architecture

Feature scorer Relation scorer

……………

… … … … …

……………

……………

SVM

Extractor…

……

Subsumption

WEBSUMM:

Some of them are taking temporary shelter at Lung Hang Estate Community Centre in Sha Tin, and Shek Lei Estate Community Centre and Princess Alexandra Community Centre in Tsuen Wan.

Emergency relief by SWD

The Social Welfare Department has provided relief articles and hot meals to 114 people who were affected by the rainstorm or mudslip throughout the territory. The people, comprising adults and children, come from 30 families. Some of them are taking temporary shelter at Lung Hang Estate Community Centre in Sha Tin, and Shek Lei Estate Community Centre and Princess Alexandra Community Centre in Tsuen Wan. The Regional Social Welfare Officer (New Territories East), Mrs Lily Wong, visited victims at Lung Hang State Community Centre this (Thursday) afternoon to offer any necessary assistance. Six victims have so far requested for Comprehensive Social Security Allowance and the applications are being processed. Social workers also escorted an 88-year old man who was feeling unwell to the Prince of Wales hospital for medical checkup.

MEAD:

The Social Welfare Department has provided relief articles and hot meals to 114 people who were affected by the rainstorm or mudslip throughout the territory. The Regional Social Welfare Officer (New Territories East), Mrs Lily Wong, visited victims at Lung Hang State Community Centre this (Thursday) afternoon to offer any necessary assistance.

RANDOM:

The Social Welfare Department has provided relief articles and hot meals to 114 people who were affected by the rainstorm or mudslip throughout the territory. Some of them are taking temporary shelter at Lung Hang Estate Community Centre in Sha Tin, and Shek Lei Estate Community Centre and Princess Alexandra Community Centre in Tsuen Wan.

LEAD:

The Social Welfare Department has provided relief articles and hot meals to 114 people who were affected by the rainstorm or mudslip throughout the territory. The people, comprising adults and children, come from 30 families.

510

2030

4050

6070

8090

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

% agreement

compression

Humans: Percent Agreement (20-cluster average) and compression

510

2030

4050

6070

8090

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

p/r

compression

Random

Humans

Humans: precision/recall (cluster average) and compression

Kappa

• N: number of items (index i)

• n: number of categories (index j)

• k: number of annotators

)(1

)()(

EP

EPAP

N

i

n

jij k

mkNk

AP1 1

2

1

1

)1(

1)(

2

1

1

)(

Nk

mEP

N

iijn

j

510

2030

4050

6070

8090

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

K

compression

Humans: Kappa and compression

2 46 54 60 61 62 112 125 199 323 398 447 551 827 883 885 1014 1197 241 1018

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

K

cluster no

Kappa, human agreement, 40%

112125

199241

323398

551883

10141197

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

K

cluster no

MEAD

Humans

Multi-document summaries of length 50 words, kappa on 10

clusters

A B C D E FG H

I J

R

J

0.45

0.5

0.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

Relative utility (upper and lower bounds), Q125, 5%

R

J

R 0.648 0.65 0.652 0.465 0.626 0.727 0.509 0.497 0.644 0.566

J 0.715 0.666 0.859 0.726 0.876 0.944 0.909 0.776 0.71 0.869

A B C D E F G H I J

A B C D E FG H

I J

R

J

0.45

0.5

0.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95


R

J

R 0.69 0.685 0.679 0.523 0.642 0.741 0.541 0.553 0.699 0.595

J 0.827 0.73 0.866 0.828 0.838 0.913 0.861 0.876 0.736 0.874

A B C D E F G H I J

A B C D E FG H

I J

R

J

0.45

0.55

0.65

0.75

0.85

0.95


R

J

R 0.74 0.738 0.724 0.653 0.695 0.77 0.647 0.679 0.764 0.664

J 0.836 0.754 0.878 0.954 0.91 0.952 0.919 0.954 0.811 0.904

A B C D E F G H I J

Relative Utility (RU) per summarizer and compression rate (Single-document)

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

Compression rate

Su

mm

ariz

er J

R

WEBS

MEAD

LEAD

J 0.785 0.79 0.81 0.833 0.853 0.875 0.913 0.94 0.962 0.982

R 0.636 0.65 0.68 0.711 0.738 0.765 0.804 0.84 0.896 0.961

WEBS 0.761 0.765 0.776 0.801 0.828

MEAD 0.748 0.756 0.764 0.782 0.808 0.834 0.863 0.895 0.921 0.968

LEAD 0.733 0.738 0.772 0.797 0.829 0.85 0.877 0.906 0.936 0.973

5 10 20 30 40 50 60 70 80 90

Relative Utility (RU) per compression rate (Multi-document)

0.61

0.63

0.65

0.67

0.69

0.71

0.73

0.75

0.77

0.79

0.81

Compression rate

RUR

S

J

R 0.6116 0.6302 0.6614 0.6894

S 0.6928 0.7246 0.7476 0.766

J 0.6886 0.7296 0.7582 0.7904

5 10 20 30

Relevance correlation (RC)

22)()(

))((

ii

ii

iii

yyxx

yyxxr

Relevance Preservation Value (RPV) as a function of compression rate (RANDOM)

0.44

0.54

0.64

0.74

0.84

0.94

Summary length (%)

RPV

Query 112

Query 125

Query 241

Query 323

Query 551

AVERAGE (10 queries)

Query 112 0.5 0.64 0.8 0.86 0.91 0.93 0.95 0.97 0.98 0.99

Query 125 0.44 0.66 0.78 0.87 0.91 0.91 0.96 0.97 0.98 0.99

Query 241 0.68 0.77 0.87 0.91 0.94 0.96 0.97 0.98 0.99 1

Query 323 0.63 0.78 0.85 0.9 0.93 0.95 0.97 0.98 0.99 1

Query 551 0.52 0.69 0.79 0.88 0.92 0.94 0.95 0.97 0.98 0.99

AVERAGE (10 queries) 0.553 0.687 0.8 0.874 0.912 0.932 0.956 0.973 0.984 0.992

5 10 20 30 40 50 60 70 80 90

FD

ME

AD

WE

BS

LE

AD

RA

ND

SU

MM

Q125

Q551

AV

G(1

0Q

)

Q112

Q241

Q323

0.77

0.82

0.87

0.92

0.97

RPV

Summarizer

Query

Relevance Preservation Value (RPV) for different summarizers (English, 20%)

Q125

Q551

AVG(10Q)

Q112

Q241

Q323

Q125 1 0.92 0.82 0.8 0.78 0.79

Q551 1 0.9 0.88 0.81 0.79 0.81

AVG(10Q) 1 0.903 0.843 0.802 0.8 0.775

Q112 1 0.91 0.88 0.8 0.8 0.77

Q241 1 0.93 0.89 0.84 0.87 0.85

Q323 1 0.92 0.91 0.85 0.85 0.88

FD MEAD WEBS LEAD RAND SUMM

FD

ME

AD

SU

MM

ALG

N

LE

AD

RA

ND

Q112

Q323

Q551

AV

G(1

0Q

)

Q125

Q241

0.58

0.63

0.68

0.73

0.78

0.83

0.88

0.93

0.98

RPV

Summarizer

Query

Relevance Preservation Value (RPV) for different summarizers (Chinese, 20%)

Q112

Q323

Q551

AVG(10Q)

Q125

Q241

Q112 1 0.87 0.76 0.74 0.72 0.71

Q323 1 0.66 0.84 0.59 0.58 0.6

Q551 1 0.91 0.75 0.72 0.75 0.74

AVG(10Q) 1 0.85 0.755 0.738 0.733 0.744

Q125 1 0.87 0.75 0.72 0.71 0.75

Q241 1 0.93 0.85 0.83 0.83 0.85

FD MEAD SUMM ALGN LEAD RAND

FDMEAD

WEBSLEAD

SUMMRAND 5%

10%20%

30%40%

0.55

0.6

0.65

0.7

0.75

0.8

0.85

0.9

0.95

1

RPV

Summarizer

Compression rate

Relevance Preservation Value (RPV) per compression rate and summarizer (English, 5 queries)

5%

10%

20%

30%

40%

5% 1 0.724 0.73 0.66 0.622 0.554

10% 1 0.834 0.804 0.73 0.71 0.708

20% 1 0.916 0.876 0.82 0.82 0.818

30% 1 0.946 0.912 0.88 0.848 0.884

40% 1 0.962 0.936 0.906 0.862 0.922

FD MEAD WEBS LEAD SUMM RAND

SUMMLEAD

MEADRAND

WEBS

with cutoff

no cutoff0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

RPV

Summarizer

Correlation method

Relevance Preservation Value (RPV) with and without cutoff (English, 5%)

with cutoff

no cutoff

with cutoff 0.48 0.55 0.61 0.29 0.6

no cutoff 0.61 0.59 0.74 0.44 0.63

SUMM LEAD MEAD RAND WEBS

SUMMLEAD

MEADRAND

WEBS

with cutoff

no cutoff0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

RPV

Summarizer

Correlation method


with cutoff

no cutoff

with cutoff 0.65 0.65 0.76 0.56 0.7

no cutoff 0.73 0.71 0.84 0.66 0.72


SUMMLEAD

MEADRAND

WEBS

with cutoff

no cutoff0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

RPV

Summarizer

Correlation method


with cutoff

no cutoff

with cutoff 0.71 0.74 0.88 0.72 0.8

no cutoff 0.79 0.8 0.92 0.78 0.82


AS

GE

ME

AD

ME

AD

OR

IG

ME

AD

002

ME

AD

003

ME

AD

S00

2

Q55

1

Q11

2

Q-A

VG

Q12

5

Q32

3

Q24

1

0.85

0.86

0.87

0.88

0.89

0.9

0.91

0.92

0.93

RPV

MEAD policy

Query

Relevance Preservation Value (RPV) per MEAD policy (5 queries)

Q551

Q112

Q-AVG

Q125

Q323

Q241

Q551 0.88 0.9 0.89 0.89

Q112 0.86 0.91 0.9 0.9 0.9

Q-AVG 0.886 0.916 0.908 0.908 0.9125

Q125 0.87 0.92 0.91 0.91 0.91

Q323 0.89 0.92 0.91 0.91 0.91

Q241 0.93 0.93 0.93 0.93 0.93

ASGEMEAD MEADORIG MEAD002 MEAD003 MEADS002

Properties of evaluation metricsKappa,P/R,accuracy

RU Wordoverlap,cosine, lcs

Relevancepreserv.

Agreement Humanextracts

X X X

Agreement humanextracts – automaticextracts

X X X

Agreement humansummaries/extracts

X

Non-binary decisions X X X

Full documents vs.extracts

X X

Systems with differentsentence segm.

X X

Multidocument extracts X X X

Full corpus coverage X X

Part VII Language modeling

Language modeling

• Source/target language• Coding process

Noisy channel Recovery

e f e*

Summarization using LM

• Source language: full document• Target language: summary

Berger & Mittal 00

• Gisting (OCELOT)

• content selection (preserve frequencies)• word ordering (single words, consecutive

positions)• search: readability & fidelity

g* = argmax p(g|d) = argmax p(g) . p(d|g)g g

Berger & Mittal 00

• Limit on top 65K words• word relatedness = alignment• Training on 100K summary+document

pairs• Testing on 1046 pairs• Use Viterbi-type search• Evaluation: word overlap (0.2-0.4)• transilingual gisting is possible• No word ordering

Berger & Mittal 00

Sample output:

Audubon society atlanta area savannah georgia chatham and local birding savannah keepers chapter of the audubon georgia and leasing

Banko et al. 00

• Summaries shorter than 1 sentence• headline generation• zero-level model: unigram probabilities• other models: Part-of-speech and position• Sample output:

Clinton to meet Netanyahu Arafat Israel

Knight and Marcu 00

• Use structured (syntactic) information

• Two approaches:– noisy channel– decision based

• Longer summaries

• Higher accuracy

Conclusion

• Summarization is coming of age

• For general domains: sentence extraction

• IR techniques not always appropriate: NLP needed

• New challenges: language modeling, multilingual summaries

APPENDIX

Conferences

• Dagstuhl Meeting, 1993 (Karen Spärck Jones, Brigitte Endres-Niggemeyer)

• ACL/EACL Workshop, Madrid, 1997 (Inderjeet Mani, Mark Maybury)

• AAAI Spring Symposium, Stanford, 1998 (Dragomir Radev, Eduard Hovy)

• ANLP/NAACL, Seattle, 2000 (Udo Hahn, Chin-Yew Lin, Inderjeet Mani, Dragomir Radev)

• NAACL, Pittsburgh, 2001 (Jade Goldstein and Chin-Yew Lin

• DUC, 2001 (Donna Harman and Daniel Marcu)

Readings

http://mitpress.mit.edu/book-table-of-contents.tcl?isbn=0262133598

(A detailed bibliography is available at the end of this handout)

Advances in Automatic Text Summarization by Inderjeet Mani and Mark T. Maybury (eds.)

1 Automatic Summarizing : Factors and Directions (K. Spärck-Jones )

2 The Automatic Creation of Literature Abstracts (H. P. Luhn)

3 New Methods in Automatic Extracting (H. P. Edmundson)

4 Automatic Abstracting Research at Chemical Abstracts Service (J. J. Pollock and A. Zamora)

5 A Trainable Document Summarizer (J. Kupiec, J. Pedersen, and F. Chen)

6 Development and Evaluation of a Statistically Based Document Summarization System (S. H. Myaeng and D. Jang)

7 A Trainable Summarizer with Knowledge Acquired from Robust NLP Techniques (C. Aone, M. E. Okurowski, J. Gorlinsky, and B. Larsen)

8 Automated Text Summarization in SUMMARIST (E. Hovy and C. Lin)

9 Salience-based Content Characterization of Text Documents (B. Boguraev and C. Kennedy)

10 Using Lexical Chains for Text Summarization (R. Barzilay and M. Elhadad)

11 Discourse Trees Are Good Indicators of Importance in Text (D. Marcu)

12 A Robust Practical Text Summarizer (T. Strzalkowski, G. Stein, J. Wang, and B. Wise)

13 Argumentative Classification of Extracted Sentenses as a First Step Towards Flexible Abstracting (S. Teufel and M. Moens)

14 Plot Units: A Narrative Summarization Strategy (W. G. Lehnert)

15 Knowledge-based text Summarization: Salience and Generalization Operators for Knowledge Base Abstraction (U. Hahn and U. Reimer)

16 Generating Concise Natural Language Summaries (K. McKeown, J. Robin, and K. Kukich)

17 Generating Summaries from Event Data (M. Maybury)

18 The Formation of Abstracts by the Selection of Sentences (G. J. Rath, A. Resnick, and T. R. Savage)

19 Automatic Condensation of Electronic Publications by Sentence Selection (R. Brandow, K. Mitze, and L. F. Rau)

20 The Effects and Limitations of Automated Text Condensing on Reading Comprehension Performance (A. H. Morris, G. M. Kasper, and D. A. Adams)

21 An Evaluation of Automatic Text Summarization Systems (T. Firmin and M J. Chrzanowski)

22 Automatic Text Structuring and Summarization (G. Salton, A. Singhal, M. Mitra, and C. Buckley)

23 Summarizing Similarities and Differences among Related Documents (I. Mani and E. Bloedorn)

24 Generating Summaries of Multiple News Articles (K. McKeown and D. R. Radev)

25 An Empirical Study of the Optimal Presentation of Multimedia Summaries of Broadcast News (A Merlino and M. Maybury)

26 Summarization of Diagrams in Documents (R. P. Futrelle)

Collections of papers

• Information Processing and Management, 1995

• Computational Linguistics (in progress), 2002

Web resources

http://www.summarization.comhttp://www.cs.columbia.edu/~jing/summarization.htmlhttp://www.dcs.shef.ac.uk/~gael/alphalist.htmlhttp://www.csi.uottawa.ca/tanka/ts.htmlhttp://www.ics.mq.edu.au/~swan/summarization/

http://www.summarization.com/

http://www.cs.columbia.edu/~jing/summarization.html









http://www.dcs.shef.ac.uk/~gael/alphalist.html











http://www.csi.uottawa.ca/tanka/ts.html

http://www.ics.mq.edu.au/~swan/summarization/

Ongoing projects

• Columbia

• ISI

• JHU, Michigan

• CMU, JPRC, etc.

• Sheffield

• elsewhere ...

Existing companies/systems

• Microsoft

• British Telecom

• http://extractor.iit.nrc.ca/

• inXight

• http://www.islandsoft.com/products.html (IslandInTEXT )

• www.pertinence.net

Available corpora

– SUMMAC corpus• send mail to [email protected]

– <Text+Abstract+Extract> corpus• send mail to [email protected]

– Open directory project• http://dmoz.org

– MEAD corpus• send mail to [email protected]

Possible research topics

• Corpus creation and annotation

• MMM: Multidocument, Multimedia, Multilingual

• Evolving summaries

• Personalized summarization

• Web-based summarization

Number Relationship type Level Description1 Identity Any The same text appears in more than one

location2 Equivalence (paraphrasing) S, D Two text spans have the same

information content3 Translation P, S Same information content in different

languages4 Subsumption S, D One sentence contains more

information than another5 Contradiction S, D Conflicting information6 Historical background S Information that puts current

information in context7 Cross-reference P The same entity is mentioned8 Citation S, D One sentence cites another document9 Modality S Qualified version of a sentence10 Attribution S One sentence repeats the information of

another while adding an attribution11 Summary S, D Similar to Summary in RST: one

sentence summarizes another

Cross-document structure theory

DOC 1

Word levelPhrase level Paragraph/sentence levelDocument level

DOC 2 DOC 3

phrasal link

word link

cross-sentential link

cross-document link

1. Clustering 2. DocumentAnalysis

3. LinkAnalysis

4. Summarization

Principles of Summarization

• Put a disclaimer indicating that (automated) summaries may not preserve the emphasis and meaning of the document.

• Preserve attribution.• Always give users a pointer to the original

document.• Indicate that the summary has been generated

automatically.• In case of conflicting sources, give all points of

view.

Bibliography

THE END

text summarization dragomir r. radev school of information, department of electrical engineering and...

Documents

new internet taxes

question of sales taxes

imposing taxes

income taxes

property taxes

current moratorium

house bill

permanent tax moratorium