Study Goal - PowerPoint PPT Presentation

1 / 28
About This Presentation
Title:

Study Goal

Description:

Better performing algorithms produce more effective labels. If high performance labeling algorithms can be identified, the effectiveness of ... – PowerPoint PPT presentation

Number of Views:18
Avg rating:3.0/5.0
Slides: 29
Provided by: comminfo
Category:
Tags: goal | study

less

Transcript and Presenter's Notes

Title: Study Goal


1
(No Transcript)
2
Study Goal
  • Measure the effect of different labels to
    represent clusters of news documents on users
    browsing collections of news stories.

Michael Cole Measuring label effect 3.22.05
3
Practical Benefit
  • The practical benefit of this study is to
    validate an instrument to rank the performance of
    automated labeling algorithms.
  • Better performing algorithms produce more
    effective labels. If high performance labeling
    algorithms can be identified, the effectiveness
    of browsers for very large document collections
    may be significantly improved.

Michael Cole Learning cluster labels 12.11.04
4
Basic Approach
  • Vary the labels while the other representational
    properties are held constant

Michael Cole Learning cluster labels 12.11.04
5
Interface
Michael Cole Learning cluster labels 12.11.04
6
Michael Cole Learning cluster labels 12.11.04
7
Labeling is a Representation Problem
Michael Cole Learning cluster labels 12.11.04
8
Theory
  • The user recognizes some semantic similarity
    between the label and a description of their task
    (Polson Lewis 1990)
  • Here the task understanding includes the user's
    general knowledge of words that are germane to
    the interest that drives the browsing experience.

Michael Cole Learning cluster labels 12.11.04
9
Tasks
  • Difficult Problem
  • Browsing tasks are all quite similar in that they
    rely on a user interest in a topic area. So
    browsing for information on sports is a similar
    task to browsing for business information.
  • Yet, the familiarity of the user with the topic
    is likely to have an impact on the number of
    words that might be recognized as semantically
    similar

Michael Cole Learning cluster labels 12.11.04
10
Tasks
The goal, however, to find good labeling
algorithms over all, or many classes, of users,
so familiarity is ignored in evaluating the main
results of this study. A pretest questionnaire
concerning general news familiarity is
administered as it may be useful in interpreting
results later.
Michael Cole Learning cluster labels 12.11.04
11
Learning Effects
Since only the labels are varied, there is the
probability that a subject will remember the
labels from the previous treatment. To mitigate
this learning effect a Latin Square Design is
used to assign the treatments (Tague-Sutcliffe,
1997).
Michael Cole Learning cluster labels 12.11.04
12
Measurement
  • The measurement of effective labels is the
    selection of clusters that contain in aggregate,
    the highest percentage of relevant documents that
    could be obtained given the test procedure
    (selecting three clusters)

Michael Cole Learning cluster labels 12.11.04
13
Linear Model
  • A linear model is adopted for this study. The
    rationale is based on the assumption that
    scanning the interface and selecting cluster
    representations are independent of one another.
  • There is no sequential process that takes place.

Michael Cole Learning cluster labels 12.11.04
14
Cluster Labeling
  • Not much previous work
  • All of the work has concentrated on creating
    labels strictly from the cluster contents.
  • Lin, Chen, and Nunamaker (1999)
  • Schweighofer, Raubner, and Dittenbach (2001)
  • Popescul and Ungar (2000)

Michael Cole Learning cluster labels 12.11.04
15
Corpus
  • NIST TDT-3 collection
  • gt 37,000 English documents from six news sources,
    including newswires, transcriptions of radio and
    TV broadcasts.
  • 113 human coded topic clusters with gt 7400
    stories
  • About 8000 documents explicitly coded as negative
    examples for some topic.

Michael Cole Learning cluster labels 12.11.04
16
Corpus Cluster Example
0003. Pinochet Trial  Seminal Event WHAT 
Pinochet, who ruled Chile from 1973-1990, is
arrested on charges of genocide and torture
during his reign. WHO Former Chilean dictator
General Augusto Pinochet Judge Baltasar Garzon
("Superjudge") WHERE  Pinochet is arrested and
held in London, then later extradited to Spain.
WHEN  The arrest occurs on 10/16/98 court
negotiations last the rest of the year. Topic
Explication Pinochet was arrested in a London
hospital on a warrant issued by Spanish Judge
Baltasar Garzon.  Pinochet appealed his arrest
and a London court agreed, but the decision was
overturned by Britain's highest court.  After
much legal wrangling over the site of the trial,
the British Courts ruled that Spain should
proceed with the extradition request  Pinochet
continues to fight it.  On topic  stories
covering any angle of the legal process
surrounding this trial (including Pinochet's
initial arrest in October, his appeals, British
Court rulings, reactions of world leaders and
Chilean citizens to the trial, etc.).  Stories
about Pinochet's reign or legacy are not on topic
unless they explicitly discuss this trial. Rule
of Interpretation    ule 3  Legal/Criminal Cases
Michael Cole Learning cluster labels 12.11.04
17
Popescul Ungar (2000)
  • Chi-square test for common terms
  • 1. generate a tree with bags of words at each
    node,
  • 2. calculate confidence that a term in a root has
    a similar frequency in all subtree nodes it has
    no distinguishing power
  • Frequent Predictive words

Michael Cole Learning cluster labels 12.11.04
18
Popescul Ungar results
FP frequent predictive product least/most FP
blend of highest and lowest FP scores
Michael Cole Learning cluster labels 12.11.04
19
Subjects Procedure
  • Drawn from a convenience sample of undergraduates
    and graduates at a large Northeastern university.
  • Fill out pre-test questionnaire.
  • Perform assigned interface tasks.
  • Fill out intermediate questionnaire asking about
    experience with interface.
  • Repeat with another interface and task set
    (twice).

Michael Cole Learning cluster labels 12.11.04
20
Measuring the Effect of Labels
  • Hypothesis 1 Automatically-generated labels will
    be associated with reduced usability and
    effectiveness compared to 'gold standard' labels.
  • Hypothesis 2 Different labeling algorithms will
    be associated with differences in browsing
    effectiveness.

Michael Cole Learning cluster labels 12.11.04
21
Results
Michael Cole Learning cluster labels 12.11.04
22
Results
Michael Cole Learning cluster labels 12.11.04
23
Results
Michael Cole Learning cluster labels 12.11.04
24
Results
Michael Cole Learning cluster labels 12.11.04
25
ANOVA
Michael Cole Learning cluster labels 12.11.04
26
Conclusions
The human-assigned labels that defined the topic
clusters were clearly the most effective
labels. The other two label sets were clearly
distinguishable from the human-assigned labels
and from each other. The instrument is a
promising approach to provide an objective
ranking of the quality of labels and, so,
automated labeling algortihms.
27
References
Popescul, A., and Ungar, L. (2000). Automatic
Labeling of Document Clusters. Unpublished
retrieved September 23, 2004 from
ttp//www.citeseer.com/ Tague-Sutcliffe, Jean.
"The Pragmatics of Information Retrieval
Experimentation, Revisited." In Readings in
Information Retrieval, ed. Karen Sparck Jones and
Peter Willett, 205-216. San Francisco, CA Morgan
Kaufmann, 1997. Originally published in
Information Processing Management 28 (1992)
467-490. Soto,R.(1999) Learning and performing
by exploration label quality measured by latent
semantic analysis, Proceedings of the SIGCHI
conference on Human factors in computing
systemsthe CHI is the limit, p.418-425, May
15-20, 1999, Pittsburgh, Pennsylvania, United
States
28
Tasks
  • Difficult Problem
  • Browsing tasks are all quite similar in that they
    rely on a user interest in a topic area. So
    browsing for information on sports is a similar
    task to browsing for business information.
  • Yet, the familiarity of the user with the topic
    is likely to have an impact on the number of
    words that might be recognized as semantically
    similiar

Michael Cole Learning cluster labels 12.11.04
Write a Comment
User Comments (0)
About PowerShow.com