>
> |
| META TOPICPARENT | Fall2007ReadingGroup |
|---|
-- ElifAktolga - 07 Dec 2007
Do Summaries Help? A Task-Based Evaluation of Multi-Document Summarization
| Date | Place | Author | Keyword |
| 2005 | SIGIR | K. McKeown?, R. J. Passonneau, and D. K. Elson | summarization, evaluation |
Summary
In this paper, the usefulness of Columbia's multi-document summarization system Newsblaster is evaluated.
Design of Evaluation
Report writing Task:
- Three questions on each of four topics
- 30 minutes to write the report
- Use only the provided information (4 clusters: 2 have related news, 2 have relevant news)
- One of these conditions: no summaries; one sentence summaries; Newsblaster summaries; human summaries
- Report content quality is measured
Method of Evaluation
Pyramid method for scoring the reports:
- look over all available summaries
- look for clause-length "summarization content units" (SCUs)
- weight them by the number of summaries they occur in
- arrange SCUs in a pyramid; fewest SCUs at the top (contained in all the summaries), most of them at the bottom (contained only in 1 summary)
- good summaries will contain all the SCUs at the topmost tier and further below
- pyramid score is the ratio of the summary's score (sum of SCU weights in the summary) to the optimal report score
Results
- reports are significantly better if the Newsblaster summaries were provided than when no summary was provided
- user satisfaction: higher with multi-document summaries (human or Newsblaster) than without any summaries
- the better the summary was, the more it was used by the users
Reference
|