1 do summaries help? a task-based evaluation of multi-document summarization kathleen mckeown,...
Post on 19-Dec-2015
215 views
TRANSCRIPT
1
Do Summaries Help?Do Summaries Help?A Task-Based Evaluation of A Task-Based Evaluation of Multi-Document SummarizationMulti-Document Summarization
Kathleen McKeown, Rebecca Passonneau, David Elson, Ani Nenkova,
Julia Hirschberg
Department of Computer ScienceColumbia University
3
Status of Multi-Document Status of Multi-Document SummarizationSummarization Robust
Many existing systems (e.g. DUC 2004) http://newsblaster.cs.columbia.edu http://www.newsinessence.com
Extensive quantitative evaluation (intrinsic) DUC 2001 – DUC 2005 Comparison of system summary content against human
models
Do system generated summaries help end-users to make better use of the news?
4
Extrinsic EvaluationExtrinsic Evaluation
Task-based evaluation of single document summarization using IR
TIPSTER-II, Brandow et al, Mani et al, Mochizuki&Okumura
Other factors can determine result (Jing et al)
Evaluation of evaluation metrics using similar task as ours
Amigo et al
5
Task EvaluationTask Evaluation
Hypothesis: multi-document summaries enable users to find information efficiently
Task: fact-gathering given topic and questions Resembles intelligence analyst task
Compared 4 parallel news browsing systems Level 1: Source documents only Level 2: One sentence multi-document summaries (e.g., Google
News) linked to documents Level 3: Newsblaster multi-document summaries linked to
documents Level 4: Human written multi-document summaries linked to
documents
6
Results PreviewResults Preview
Quality of facts gathered significantly better Newsblaster vs. documents alone
User satisfaction higher Newsblaster and human summaries vs. documents and 1
sentence summaries
Summaries contributed important facts Newsblaster and human summaries vs. 1 sentence summaries
Full multi-document summarization more powerful than no documents or single sentence summarization
8
Evaluation GoalsEvaluation Goals
Do summaries help users find information needed to perform a fact gathering task?
Do users use information from the summary in gathering their facts?
Do summaries increase user satisfaction with the online news system?
Do users create better fact sets with an online news system that includes summaries than one without?
How does type of summary (i.e., 1-sentence, system generated, human generated) affect quality of task output and user satisfaction?
9
Experimental DesignExperimental Design
Subjects performed four 30-minute fact-gathering scenarios
Prompt: topic description plus three questions
Given a web page as sole resource Space in which to compose response Instructed to cut and paste from summary or article Four event clusters per page
Two centrally relevant, two less relevant 10 documents per cluster on average
Complete survey after each scenario
10
PromptPrompt The conflict between Israel and the Palestinians
has been difficult for government negotiators to settle. Most recently, implementation of the "road map for peace," a diplomatic effort sponsored by the United States, Russia, the E.U. and the U.N., has suffered setbacks. However unofficial negotiators have developed a plan known as the Geneva Accord for finding a permanent solution to the conflict.
Who participated in the negotiations that produced the Geneva Accord?
Apart from direct participants, who supported the Geneva Accord preparations and how?
What has the response been to the Geneva Accord by the Palestinians and Israelis?
11
Experimental DesignExperimental Design
Subjects performed four 30-minute fact-gathering scenarios
Prompt: topic description plus three questions Produced a report containing a list of facts
Given a web page as sole resource Space in which to compose response Instructed to cut and paste from summary or article and
make citation Four event clusters per page
Two centrally relevant, two less relevant 10 documents per cluster on average
Complete survey after each scenario
13
Level 2: 1-sentence summary for each Level 2: 1-sentence summary for each event cluster, 1-sentence summary for event cluster, 1-sentence summary for each articleeach article
14
Full multi-document summariesFull multi-document summaries
Neither humans nor systems had access to the prompt
Level 3: Generated by Newsblaster for each event cluster
Level 4 Human written summary for each event cluster Summary writers hired to write summaries
English or Journalism students with high verbal SAT
15
Levels 3 and 4: full summary for each event Levels 3 and 4: full summary for each event clustercluster
16
Experimental DesignExperimental Design
Subjects performed four 30-minute fact-gathering scenarios
Prompt: topic description plus three questions Produced a report containing a list of facts
Given a web page as sole resource Space in which to compose response Instructed to cut and paste from summary or article and
make citation Four event clusters per page
Two centrally relevant, two less relevant 10 documents per cluster on average
Complete survey after each scenario
17
Study ExecutionStudy Execution
45 Subjects with varied background 73% students (BS, BA, journalism, law) Native speakers of English Paid, with promise of monetary prize for best report
3 studies, controlling for scenario and level order, ~11 subjects/scenario/level
18
Results – What was MeasuredResults – What was Measured
Report content across summary conditions: levels 1-4
User satisfaction per summary condition based on user surveys
Source of report content (summary or article) by counting fact citations
19
Scoring Report ContentScoring Report Content
Compare subject reports against a gold standard Used the Pyramid method [HLT2004]
Avoids postulating an ideal exhaustive report Predicts multiple equally good reports Provides a metric for comparison
Gold standard for report x = pyramid of facts constructed from all reports except x
Relative importance of facts determined by report writers 34 reports per pyramid on average -> very stable
20
Pyramid representationPyramid representation
Tiers of differentially weighted facts
Top: few facts, high weight Bottom: many facts, low
weight Report facts that don’t
appear in pyramid have weight 0
Duplicate report facts get weight 0W=1
W=33
W=34
…
21
Ideally informative reportIdeally informative report
Does not include a fact from a lower tier unless all facts from higher tiers are included as well
22
Ideally informative reportIdeally informative report
Does not include a fact from a lower tier unless all facts from higher tiers are included as well
23
Ideally informative reportIdeally informative report
Does not include a fact from a lower tier unless all facts from higher tiers are included as well
24
Ideally informative reportIdeally informative report
Does not include a fact from a lower tier unless all facts from higher tiers are included as well
25
Ideally informative reportIdeally informative report
Does not include a fact from a lower tier unless all facts from higher tiers are included as well
26
Ideally informative reportIdeally informative report
Does not include a fact from a lower tier unless all facts from higher tiers are included as well
Report LengthReport Length
Wide variation in length impacts scores
We restricted report length < 1 standard deviation above the mean by truncating question answers
28
Results - ContentResults - Content
Summary Level Pyramid Score
Level 1 (docs only) .3354
Level 2 (1 sentence) .3757
Level 3 (Newsblaster) .4269
Level 4 (Human) .4027
Report quality improves from level 1 to level 3. (One scenario was dropped from results as it was problematic for subjects)
29
Statistical AnalysisStatistical Analysis
ANOVA shows summary is marginally significant factor
Bonferonni method applied to determine differences in summary levels
Difference between Newsblaster and documents-only significant (P=.05)
Differences between Newsblaster and 1-sentence or human not significant
ANOVA shows that scenario, question and subject also significant factors
30
Results - User SatisfactionResults - User Satisfaction
6 questions in exit survey required response from a 1-5 scale
Average increases by summary type
Level 1 Level 2 Level 3 Level 4
Average 2.75 3.39 3.47 3.56
31
With full summaries, users read lessWith full summaries, users read less
Question Level 1 Level 2 Level 3 Level 4
A. What best describes your experience reading source articles?
2.83 2.70 3.10 3.10
1. I read a LOT more than I needed to
5. I only read those articles I needed to read
32
With Summaries, easier to write With Summaries, easier to write report and tended to have more timereport and tended to have more timeQuestion Level 1 Level 2 Level 3 Level 4
B. How difficult do you think it was to write the report?
2.27 3.07 2.95 3.0
1. Very difficult
5. Very easy
C. Do you feel you had enough time to write the report?
2.43 3.91 3.38 3.57
1. I needed more time
5. I had more than enough time
33
Usefulness improves with summary qualityUsefulness improves with summary qualityHuman summaries help best with timeHuman summaries help best with timeQuestion Level 1 Level 2 Level 3 Level 4
D. What best describes your experience using article summaries?
n/a 3.16 3.29 4.14
1. They had nothing useful to say
5. Everything I needed to know
E. Did you feel that the automatic summaries saved you time, wasted time, or had no impact on your time budget?
n/a 4.09 3.95 4.14
1. Summaries wasted time
5. Summaries saved me time
34
Multiple Choice Survey Multiple Choice Survey QuestionsQuestionsQuestion Level 2 Level 3 Level 41. Which was most helpful?
Source articles helped most 64% 48% 29%Equally helpful 32% 29% 29%Summaries helped most 5% 24% 43%2. How did you budget your time?
Most searching, some writing 55% 48% 67%Half searching, half writing 39% 29% 19%Most writing, some searching 7% 24% 14%
35
Citation PatternsCitation Patterns
Report writers were significantly more likely to extract facts from summaries with Newsblaster and human summaries
Level 2 Level 3 Level 4
Citations from summaries
8% 17% 27%
36
What we LearnedWhat we Learned
With summaries, a significant increase in quality of report
We hypothesized summaries would reduce reading time As summary quality increases, users significantly more often
draw facts from summary without decrease in report quality Users claim they read fewer full documents with level 3 and
4 summaries Full multi-document summarization better than 1
sentence summaries Almost 5 times the proportion of subjects using Newsblaster
summaries say summaries are helpful than subjects using 1 sentence summaries
37
Need for Follow-on StudiesNeed for Follow-on Studies
Why no significant increase in report quality from level 2 to level 3?
Interface differences Level 2 had summary for each article, level 3 did
not Level 3 required extra clicks to see list of articles
Studies to investigate controlling report length
Studies to investigate impact of scenario and question
40
Need for Follow-on StudiesNeed for Follow-on Studies
Why no significant increase in report quality from level 2 to level 3?
Interface differences Level 2 had summary for each article, level 3 did
not Level 3 required extra clicks to see list of articles
Studies to investigate controlling report length
Studies to investigate impact of scenario and question
41
ConclusionsConclusions
Do summaries help? Yes Our task-based, extrinsic evaluation yielded significant
conclusions
Full multi-document summarization (Newsblaster, human summaries) helps users perform better at fact-gathering than documents only
Users are more satisfied with full multi-document summarization than Google News style 1-sentence summaries