Transcript Novel User Interfaces: The Digital Desk as the Interface of the Future

Untangling Text Data Mining
Stanford Digital Libraries Seminar
May 11, 1998
Marti Hearst
UC Berkeley SIMS
www.sims.berkeley.edu/~hearst
5/11/98
1
Caveat Emptor:
I do information access.
I do not do text data mining (yet).
This talk is an attempt
to explore the relationship
between the two.
5/11/98
2
Talk Outline

Definitions
– What is Data Mining?
– What is Information Access?
– What is Text Data Mining?



Empirical Computational Linguistics
Real text data mining tasks
Conclusions and Future Directions
5/11/98
3
The Knowledge Discovery
from Data Process (KDD)
KDD: The non-trivial process of
identifying valid, novel, potentially
useful, and ultimately understandable
patterns in data.
(Fayyad, Shapiro, & Smyth, CACM 96)
Note: data mining is just one step in the process
5/11/98
4
What is Data Mining?
(Fayyad & Uthurusamy 96, Fayyad 97)



Fitting models to or determining
patterns from very large datasets.
A “regime” which enables people to
interact effectively with massive data
stores.
Deriving new information from data.
– finding patterns across large datasets
– discovering heretofore unknown
information
5/11/98
5
What is Data Mining?

Potential point of confusion:
– The extracting ore from rock metaphor does
not really apply to the practice of data
mining
– If it did, then standard database queries
would fit under the rubric of data mining
• Find all employee records in which employee
earns $300/month less than their managers
– In practice, DM refers to:
• finding patterns across large datasets
• discovering heretofore unknown information
5/11/98
6
Why Data Mining?



Because the data is there.
Because current DBMS technology
does not support data analysis.
Because
–
–
–
–
larger disks
faster cpus
high-powered visualization
networked information
are becoming widely available.
5/11/98
8
DM Touchstone Applications
(CACM 39 (11) Special Issue)

Finding patterns across data sets:
– Reports on changes in retail sales
• to improve sales
– Patterns of sizes of TV audiences
• for marketing
– Patterns in NBA play
• to alter, and so improve, performance
– Deviations in standard phone calling behavior
• to detect fraud
• for marketing
5/11/98
9
DM Touchstone Applications
(CACM 39 (11) Special Issue)

Separating signal from noise:
– Classifying faint astronomical objects
– Finding genes within DNA sequences
– Discovering novel tectonic activity
5/11/98
10
What’s new here?


Sounds like statistical modeling or
machine learning.
Main Difference: scale and
availability (Fayyad 97)
– Datasets too large for classical analysis
– Increased opportunity for access
• end user is often not a statistician
– New issues in sampling
5/11/98
12
Statistician’s Viewpoint
(David Hand 97)

What’s new about DM?
– Returns statisticians to their empirical roots
• exploration rather than modeling
– Hypothesis testing may be irrelevant
• given the large data sizes everything is
significant
– Data was collected for some other purpose
than what it is being analyzed for now
5/11/98
13
Talk Outline

Definitions
– What is Data Mining?
– What is Information Access?
– What is Text Data Mining?



Empirical Computational Linguistics
Real text data mining tasks
Conclusions and Future Directions
5/11/98
15
Information Access
(Information Retrieval more broadly construed)

Problem:
– Huge amounts of online textual information

Goal:
– Build systems to help people discover, create use, reuse, and understand information

Approach:
– Leverage off of users’ smarts
– Combine stats, text analysis, user interfaces
5/11/98
17
Information Retrieval
A restricted form of Information Access



The system has available only pre-existing,
“canned” text passages.
Its response is limited to selecting from these
passages and presenting them to the user.
It must select, say, 10 or 20 passages out of
millions!
5/11/98
18
Needles in Haystacks

The emphasis in IR (and standard
DB) is in answering ad hoc queries.
5/11/98
19
IA vs. KDD Process
5/11/98
20
IA vs. KDD Process
Query/Information Need
5/11/98
21
IA vs. KDD Process
Query/Information Need
Match query
against transformed
data
Show results ranked
in relevance order
5/11/98
22
Talk Outline

Definitions
– What is Data Mining?
– What is Information Access?
– What is Text Data Mining?



Empirical Computational Linguistics
Real text data mining tasks
Conclusions and Future Directions
5/11/98
24
What is Text Data Mining?

Peoples’ first thought:
– Make it easier to find things on the Web.
– But this is information retrieval!

The metaphor of extracting ore from
rock:
– Does make sense for extracting documents
of interest from a huge pile.
– But does not reflect notions of DM in
practice:
• finding patterns across large collections
• discovering heretofore unknown information
5/11/98
25
Real Text DM

What would finding a pattern across a
large text collection really look like?
5/11/98
26
Bill Gates + MS-DOS
in the Bible!
From: “The Internet Diary of the man who cracked the Bible Code”
Brendan McKay, Yahoo Internet Life, www.zdnet.com/yil
(William Gates, agitator, leader)
5/11/98
27
From: “The Internet Diary of the man who cracked the Bible Code”
Brendan McKay, Yahoo Internet Life, www.zdnet.com/yil
5/11/98
28
Real Text DM

The point:
– Discovering heretofore unknown
information is not what we usually do with
text.
– (If it weren’t known, it could not have been
written by someone!)

However:
– There is a field whose goal is to learn about
patterns in text for its own sake ...
5/11/98
29
Observation
Research that exploits patterns in text does
so mainly in the service of computational
linguistics, rather than for learning about
and exploring text collections.
5/11/98
30
Talk Outline


Definitions
Empirical Computational Linguistics
– Special and important properties of text
– Relationship to TDM
– Examples of TDM as CL

Real text data mining tasks

Conclusions and Future Directions
5/11/98
31
Recent Trends in NLP (CL)


Previously: AI, full understanding
Current: Corpus-based, Statistical
• ACL proceedings: from 3 corpus-based papers in
1991 to at least half in 1996
• Stat NLP was tried long ago (Z. Harris)

Simple Often Wins
• Echoes results in IR

Interesting direction:
• Statistics + Linguistics (Klavans & Resnik 96)
5/11/98
32
Text Analysis (CL) Tasks




Word Sense Disambiguation
Automatic Lexicon Augmentation
Discourse Analysis
Parsing
• Phrase Identification
• Phrase Attachments
• Predicate/Argument Structure
• Scope of Conjunctions
• ...
5/11/98
33
Why Text is Tough
– Abstract concepts difficult to represent
(AI-Complete)
– “Countless” combinations of subtle, abstract
relationships among concepts
– Many ways to represent similar concepts
space ship, flying saucer, UFO, figment of imagination
– Concepts are difficult to visualize
– High dimensionality
Tens or hundreds of thousands of features
5/11/98
34
Why Text is Tough

Language is:
– ambiguous (many different meanings for
the same words and phrases)
– different combinations imply different
meanings
5/11/98
35
Why Text is Tough

I saw Pathfinder on Mars with a telescope.

Pathfinder photographed Mars.



The Pathfinder photograph mars our perception
of a lifeless planet.
The Pathfinder photograph from Ford has arrived.
The Pathfinder forded the river without marring
its paint job.
5/11/98
36
Why Text is Easy


Highly redundant in bulk
Just about any simple algorithm
can get “good” results for coarse
tasks
–
–
–
–
5/11/98
Pull out “important” phrases
Find “meaningfully” related words
Create summary from document
Major problem: Evaluation
37
Stupid Text Tricks
– Coarse IR, Clustering
• Don’t need dimension reduction (except
stopwords)
• Don’t need morphological analysis
• Don’t need word sense disambiguation
– Partial parsing:
• Simple, greedy transformation rules
• Cascading finite state machines
– Categorization
• Assume independence
5/11/98
38
Text “Data Cleaning”
Pre-process text as follows:
 Tokenization
 Morphological Analysis
(Stemming)
inflectional, derivational, or crude IR methods

Part-of-Speech Tagging
I/Pro see/VP Pathfinder/PN on/P Mars/PN ...

Phrase Boundary Identification
[Subj I] [VP saw] [DO Pathfinder] [PP on Mars]
[PP with a telescope].
5/11/98
39
CCL Methodology

Describe here the standard
methodology for corpus-based
computational linguistics
algorithms
5/11/98
40
CCL Examples

Place here examples of the kinds of
output generated for
computational linguistics
applications
5/11/98
41
Inducing MetaData for
Documents

Assigning bibliographic metadata
– author, genre, time, region

Subject/Topic assignments
– category labels: MeSH, LoC, ACM keywords

Information Extraction (MUC)
– MUC: terrorist incidents
•
•
•
•
5/11/98
who did the bombing
where did the bombing take place
what weapon(s) were used
when did it happen
42
Inducing MetaData for
Collections



Indexes
Hierarchical Categorization
Overviews of Connectivity
• hyperlinks
• co-citation links

Overviews of Subject Matter
• 2D
• 3D
• dynamic
5/11/98
43
A Main Point:


Empirical CL is usually not helpful
for improving Information Access.
However, it can produce
– metadata
– overviews
– associations
that are indirectly useful for IA.
5/11/98
44
Talk Outline



Definitions
Empirical Computational Linguistics
Real text data mining tasks
– TDM not using text
– TDM using text

Conclusions and Future Directions
5/11/98
45
TDM using Metadata
(instead of Text)
(Dagan, Feldman, and Hirsh, SDAIR ‘96)
– Data:
• Reuter’s newswire (22,000 articles, late 1980s)
• Categories: commodities, time, countries,
people, and topic
– Goals:
• distributions of categories across time (trends)
• distributions of categories between collections
• category co-occurrence (e.g., topic|country)
– Interactive Interface:
• lists, pie charts, 2D line plots
5/11/98
46
Combining Text with Metadata
(images, hyperlinks)

Examples
– Text + Links to find “authority pages”
(Kleinberg at Cornell, Page at Stanford)
– Usage + Time + Links to study evolution of
web and information use (Pitkow et al. at PARC)
– Images + Text to improve image search
5/11/98
47
Talk Outline



Definitions
The New Empirical Computational
Linguistics
Real text data mining tasks
– TDM not using text
– TDM using text

Conclusions and Future Directions
5/11/98
48
Ore-Filled Text Collections


Newspaper/Newswire
Medical Articles
– Patterns associated with symptoms, drugs

Patent Law
– Recent Study Justifying Scientific Funding
– Hypotheses for New Inventions

“Corporate Memory”
5/11/98
49
True Text Data Mining:
Don Swanson’s Medical Work

Given
– medical titles and abstracts
– a problem (incurable rare disease)
– some medical expertise

find causal links among titles
– symptoms
– drugs
– results
5/11/98
50
Swanson Example (1991)

Problem: Migraine headaches (M)
–
–
–
–
–
–
–
–

stress associated with M
stress leads to loss of magnesium
calcium channel blockers prevent some M
magnesium is a natural calcium channel blocker
spreading cortical depression (SCD)implicated in M
high levels of magnesium inhibit SCD
M patients have high platelet aggregability
magnesium can suppress platelet aggregability
All extracted from medical journal
titles
5/11/98
51
Swanson’s TDM


Two of his hypotheses have
received some experimental
verification.
His technique
– Only partially automated
– Required medical expertise

Few people are working on this.
5/11/98
52
Text Collection Overviews

Clusters/Unsupervised Overviews
–
–
–
–
–
5/11/98
Chalmers: BEAD, Networks of Words
Lin,Chen: Kohonen Feature Maps
Xerox PARC: Local Clusters
Pacific Northwest: ThemeScapes
Rennison: Galaxy of News
53
Text Overviews
– Huge 2D maps may be inappropriate focus
for information retrieval
• can’t see what documents are about
• documents forced into one position in semantic
space
• space difficult to browse for IR purposes
– Perhaps more suited for pattern discovery
• problem: often only one view on the space
5/11/98
54
Talk Outline



Definitions
The New Empirical Computational
Linguistics
Real text data mining tasks
– TDM not using text
– TDM using text

Conclusions and Future Directions
5/11/98
55
Conclusions

Currently, what might be construed as Text
Data Mining is really Computational
Linguistics
– Text is tricky to process, but rich and abundant (now)
– There are many CL tools available

Data Mining directly from text
– tells us about language
– produces meta-information that may be useful for
information access
5/11/98
56
Conclusions, continued

Information Access != Text Data Mining
– IA
= finding needle in haystack
– TDM = finding patterns or discovering new information

However, Information Access may
potentially be served by Text Data Mining
techniques:
– automated metadata assignment
– collection overviews

The synthesis of ideas from TDM and IA:
– Perhaps a new field of exploratory data analysis over
text!
5/11/98
57
Promsing Research Directions

Text Data Mining Problems:
– Patterns within sets of documents:
• What is the latest in this field?
• How is this field related to that field?
– Chains of evidence embedded in text:
• What drugs have been tested for this
symptom?
• What effects did this funding have on that
field?
– Human use of information over time,
• How does information diffuse across the web?
5/11/98
58