Visual Computing
Text Visualization
Based on slides by
Chris North, Virginia Tech
Jeffrey Heer, Stanford University
Text & Document Visualization
• Text not pre-attentive
• Text = Abstract Concepts = Very High Dimensionality
–
–
–
–
Multiple & ambiguous meanings
Combinations of abstract concepts more difficult to visualize
Different combinations imply different meanings
Language only hints at meaning
 based on common understanding “How much is that doggy in the window?”
• Facilitate Information Retrieval
– Collection Overview
– Visualize which parts of query satisfied by document / collection
– Understand why documents retrieved
• Cluster Documents Based on Words in Common
– Finds overall similarities among groups of documents
– Picks out some themes, ignores others
• Map Clusters onto 2D or 3D Representation
– Minimize time/effort to decide which documents to examine
What is text data?
Documents
•
•
•
•
Articles, books and novels
Computer programs
E-mails, web pages, blogs
Tags, comments
Collection of documents
• Messages (e-mail, blogs, tags, comments)
• Social networks (personal profiles)
• Academic collaborations (publications)
Text as Data
Words are (not) nominal?
• High dimensional (10,000+) More than
equality tests
• Words have meanings and relations
– Correlations: Hong Kong, San Francisco, Bay Area
– Order: April, February, January, June, March, May
– Membership: Tennis, Running, Swimming, Hiking, Piano
– Hierarchy, antonyms & synonyms, entities, …
Text Processing Pipeline
• Tokenization: segment text into terms
– Special cases? e.g., “San Francisco”, “L’ensemble”, “U.S.A.”
– Remove stop words? e.g., “a”, “an”, “the”, “to”, “be”?
• Stemming: one means of normalizing terms
–
–
–
–
Reduce terms to their “root”; Porter’s algorithm for English
e.g., automate(s), automatic, automation all map to automat
For visualization, want to reverse stemming for labels
Simple solution: map from stem to the most frequent word
• Result: ordered stream of terms
The Bag of Words Model
• Ignore ordering relationships within the
text
• A document ≈ vector of term weights
– Each dimension corresponds to a term (10,000+)
– Each value represents the relevance
– For example, simple term counts
• Aggregate into a document x term matrix
– Document vector space model
Document x Term matrix
• Each document is a vector of term weights
• Simplest weighting is to just count occurrences
Antony and
Cleopatra
Julius
Caesar
The
Tempest
Hamlet
Othello
Macbeth
Antony
157
73
0
0
0
0
Brutus
4
157
0
1
0
0
Caesar
232
227
0
2
1
1
Calpurnia
0
10
0
0
0
0
Cleopatra
57
0
0
0
0
0
mercy
2
0
3
5
5
1
worser
2
0
1
1
1
0
WordCount (Harris 2004)
WordCount™ is an interactive presentation of the 86,800 most
frequently used English words.
http://wordcount.org
Term Vector Theory for Information Retrieval (IR)
Vector Space Model
IR systems assign weights to terms by considering
1.
2.
local information from individual documents
global information from collection of documents
Systems that assign weights to links use Web graph information to properly
account for the degree of connectivity between documents.
In IR studies, the classic weighting scheme is the Salton Vector Space Model,
commonly known as the "term vector model".
This weighting scheme is given by
Term Weight =
where
tfi = term frequency (term counts) or number of times a term i occurs in
a document.
dfi = document frequency or number of documents containing term i
D = number of documents in the database.
Many models that extract term vectors from documents and queries are derived
this equation.
Computing Weights
Term Frequency
t = term were are searching for
tftd = count(t) in d
dft = # docs containing t
N = # of docs
TF.IDF: Term Freq by Inverse Document Freq
tf.idftd = tftd × log(N/dft)
• This is the relative importance in the document
• Word is more important in the fewer document it appears.
— We are more interested in a words that appear often in a single
document not in the collection as a whole
Term vectors for Group of Docs with tf-idf
weights
Visualizing Document Content
Tag Cloud: Word Counts
Wordle
http://www.wordle.net/create
During the campaign, Palin gave an energetic speech in Dayton, Ohio.
She appealed to women voters by evoking, of all things, the presidential
campaign of Democrat Hillary Clinton, saying, “Hillary left 18 million
cracks in the highest, hardest glass ceiling in America. But it turns out the
women of America aren’t done yet.”
Here’s that speech:
Text from: California Watch
Weaknesses of Tag Clouds
• Sub-optimal visual encoding (size vs. position)
• Inaccurate size encoding (long words are
bigger)
• May not facilitate comparison (unstable layout)
• Term frequency may not be meaningful
• Does not show the structure of the text
Word Tree: Word Sequences
TextArc – Brad Paley
http://textarc.org/
Arc Diagrams – M. Wattenberg
Les Misérables character interaction. Each character is represented by a circle and
the connecting arc represents co-occurrence in a chapter. The character's size
indicates the number of appearances they have over the entire work.
Literature Fingerprinting
Problem: Authorship Attribution
• Determine, if a text was written by an
author or not.
• A common problem in literary analysis.
• What features are useful for
discrimination?
• Case study on some books by Jack
London and Mark Twain.
Variables for Literary Analysis
•
Statistical measures
–
–
–
–
•
Vocabulary measures
–
–
–
–
–
•
Syllables per word
Sentence length
Proportions of parts of speech
...
Frequencies of specific words
Type-token ratio
Simpson’s index
Hapax (dis)legomena
...
Syntax measures
Average Sentence Length
Structured Document
Collections
• Multi-dimensional:
• author, title, date, journal, …
• Trees:
• Dewey decimal system
• Graphs:
• web, citations
Citation Networks
ANNOTATIONS
TYPESCRIPT
select source |
|
search on author |
|
search S1 had 7 results |
type S1 format 3 result 1 |
|
|
|
|
|
|
|
first author |
|
|
year, vol, page |
|
|
search for citers |
|
search S2 had 1 result |
?b 434
<16 lines of accounting information removed>
?s au=card sk
S1
?t 1/3/1
7
AU=CARD SK
1/3/1
DIALOG(R)File 434:Scisearch(R)
(C) 1994 Inst For Sci Info. All Rts. Reserv.
12204937
Genuine Article#: KU797
No. Reference...
Title: INFORMATION VISUALIZATION USING 3D INTERAC...
Author(S): ROBERTSON GG; CARD SK; MACKINLAY JD
Corporate Source: XEROX CORP,PALO ALTO RES CTR,33...
ALTO//CA/94304
Journal: COMMUNICATIONS OF THE ACM, 1993, V36, N4...
ISSN: 0001-0782
Language: ENGLISH
Document Type: ARTICLE
?s cr=robertson gg, 1993, v36, p56, ?
S2
1
CR=ROBERTSON GG, 1993, V36, P56, ?
This annotated typescript from a DIALOG session shows a search of the Science
Citation Database for articles that include S. K. Card as an author. Typescripts like this
do not particularly show the structure of a search.
Butterfly Browser - Mackinlay et al (PARC)
Based on four key ideas:
• Visualizations Of References And Citers
– Visualize scholarly articles as user interface objects with two
wings, one wing for listing an article's references and the other
wing for listing the article's citers.
• Link-Generating Queries
– Automatically create link-generating queries that link an article's
record to the corresponding records for the article's references
and citers
• Asynchronous Query Processes
– Uses asynchronous processing for information access so the
user does not have to wait for queries to complete
• Embedded Process Control
– User can explicitly create and terminate query processes
Butterfly Browser
Butterfly:
Left = refs
Right = citers
Yellow = #citers
Blue = visited
3d plot:
date,
Name,
# citers
Unstructured Document
Collections
• Focus on Full Text
• Examples:
• digital libraries, news archives, web pages
• email archives, image galery
• Tasks:
•
•
•
•
•
Search
Browse
Classification, structurization
Statistics, keyword usage, languages
Subjects, themes, coverage
Visualization Strategies
• Cluster Maps
• Keyword Query results
• Relationships
• Reduced representation
• User controlled layout
Cluster Map
• Create a “map” of the document collection
• Similar documents near each other
• Dissimilar documents far apart
• “Library” or “Grocery store” concept
Document Vectors
•
•
•
•
“aardvark”
“banana”
“chris”
…
Doc1
1
2
0
Doc2
2
1
0
Doc3
0
0
3
…
• Now it’s a Multi-D visualization problem?
• Dimensionality reduction:
• Projection: e.g. Principal Components Analysis (PCA)
• Similarity-based methods:
1. Compute “Similarity” between pair of docs
2. Layout documents in 1/2/3-D map by similarity
Similarity Matrix
•
•
•
•
“aardvark”
“banana”
“chris”
…
Doc1
1
2
0
Doc2
2
1
0
Doc3
0
0
3
…
Doc1 Doc2 Doc3 …
Doc1 1
• Similarity metrics?
• dot product
•
0.66 0
Doc2 0.66 1
0
Doc3 0
1
…
0
Layout Mapping
•
•
•
•
Spring model of graph layout
Multi-Dimensional Scaling (MDS)
Self-organizing Map (kohonen map)
Clustering: Partition, hierarchical
• How to label a group?
• …
Cluster Algorithms
• Partition clustering:
Top down
Partition into k subsets
• Pick k seeds
• Iteratively attract nearest neighbor
• Hierarchical clustering:
Dendrogram
• Group nearest-neighbor pair
• Iterate
Bottom up
Landscapes
• Wise et al, “Visualizing the non-visual”
• ThemeScapes, Cartia, IN-SPIRE (PNNL)
• Mountain = topical theme
• Mountain height = number of relevant documents
LandScapes
• Abstract, 3D landscapes of information
• Convey relevant information about topic or
themes without the cognitive load
• Spatial relationships reveal the intricate
interconnection of thems
• Dominant themes are shown in a relief map of
natural terrain.
• Themes are represented by peaks and their
height indicates relative strength within the
document set.
Advantages
• Displays much of the complex content of
the document database
• Utilizes innate human abilities for pattern
recognition and spatial reasoning
• Communicative invariance across levels of
textual scale
• Promotes analysis
ThemeRiver - PNNL
• Displays changes to themes over time
• Helps users identify time-related patterns, trends, and
relationships across a large collection of documents.
• Themes in the collection are represented by a "river" that
flows left to right through time.
• The river widens or narrows to depict changes in the
collective strength of selected themes in the underlying
documents.
• Individual themes are represented as colored "currents"
flowing within the river. The theme currents narrow or
widen to indicate changes in individual theme strength at
any point in time.
ThemeRiver
http://infoviz.pnl.gov/images/ThemeRiver.mov
Galaxies
• Displays cluster and document
interrelatedness
• 2D scatterplot of ‘docupoints’
• Simple point and click exploration
• Sophisticated tools
– Facilitate more in-depth analysis
– Ex) temporal slicer
IN-SPIRE™ - PNNL
• The Galaxy visualization uses the metaphor of
stars in the night sky where each star represents
a document.
• Closely related documents cluster together while
unrelated documents are further apart.
• Galaxies help users to understand what is in a
document collection and allows them to explore
the context of their specific interests.
GalaxyView
Dot = document
Galaxy = cluster
StarLight - PNNL
• Relationships to geography, etc.
The Self-Organizing Map (SOM)
• Data visualization technique invented by
Teuvo Kohonen which reduces the
dimensions of data through the use of selforganizing neural networks.
• SOMs reduce dimensions by producing a
map of 1 or 2 dimensions that plots the
similarities of the data by grouping similar
data items together.
Components of SOM
1. Sample Data
–
e.g. RGB (3 dimensions)
2. Weight vectors
– two components:
•
•
–
The data itself
The data’s natural location
e.g. 2D array of weight vectors
(say, colors at right)
Components of SOM
• Algorithm
Initialize Map
For t from 0 to 1
Randomly select a sample
Get best matching unit
Scale neighbors
Increase t a small amount
End for
Self-organizing Maps
• Xia Lin, “Document Space”
• Kohonen map, http://faculty.cis.drexel.edu/sitemap/index.html