CSE 512 - Data Visualization

Text Visualization

Jeffrey Heer University of Washington

Why Visualize Text?
Why Visualize Text?

Understanding – get the “gist” of a document

Grouping – cluster for overview or classification

Comparison – compare document collections, or

inspect evolution of collection over time

Correlation – compare patterns in text to those in

other data, e.g., correlate with social network
Text as Data
Documents
Articles, books and novels
E-mails, web pages, blogs
Tags, comments
Computer programs, logs

Collections of Documents
Messages (e-mail, blogs, tags, comments)
Social networks (personal profiles)
Academic collaborations (publications)
 Example:
Health Care Reform
Example: Health Care Reform
Recent History
Initiatives by President Clinton
Overhaul by President Obama

Text Data
News articles
Speech transcriptions
Legal documents

What questions might you want to answer?

What visualizations might help?
A Concrete Example
Tag Clouds: Word Count
President Obama’s Health Care Speech to Congress [NY Times]

economix.blogs.nytimes.com/2009/09/09/obama-in-09-vs-clinton-in-93
 Bill Clinton 1993

Barack Obama 2009

economix.blogs.nytimes.com/2009/09/09/obama-in-09-vs-clinton-in-93
Word Tree: Word Sequences
Gulfs of Evaluation
Many text visualizations do not represent the text
directly. They represent the output of a language
model (word counts, word sequences, etc.).

· Can you interpret the visualization? How well

does it convey the properties of the model?
· Do you trust the model? How does the model
enable us to reason about the text?
Text Visualization Challenges
High Dimensionality
Where possible use text to represent text…
… which terms are the most descriptive?

Context & Semantics

Provide relevant context to aid understanding.
Show (or provide access to) the source text.

Modeling Abstraction
Determine your analysis task.
Understand abstraction of your language models.
Match analysis task with appropriate tools and models.
Topics

Text as Data
Visualizing Document Content
Evolving Documents
Visualizing Conversation
Document Collections
Text as Data
Words as nominal data?

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
1. Tokenization
Segment text into terms.
Remove stop words? a, an, the, of, to, be
Numbers and symbols? #gocard, @stanfordfball, Beat Cal!!!!!!!!
Entities? San Francisco, O’Connor, U.S.A.

2. Stemming
Group together different forms of a word.
Porter stemmer? visualization(s), visualize(s), visually -> visual
Lemmatization? goes, went, gone -> go

3. Ordered list of terms

Tokenization & Stemming
Well-formed text to support stemming?
txt u l8r!

Word meaning or entities?

#berkeley -> #berkelei

Reverse stems for presentation.

Ha appl made programm cool?
Has Apple made programmers cool?
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-term matrix

· Document vector space model
 Document-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
WordCounts (Harris ’04)

http://wordcount.org
Tag Clouds
Strengths
Can help with gisting and initial query formation.

Weaknesses
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
Keyword Weighting
Term Frequency
tftd = count(t) in d
Can take log frequency: log(1 + tftd)
Can normalize to show proportion: tftd / Σt tftd
Keyword Weighting
Term Frequency
tftd = count(t) in d

TF.IDF: Term Freq by Inverse Document Freq

tf.idftd = log(1 + tftd) × log(N/dft)
dft = # docs containing t; N = # of docs
Keyword Weighting
Term Frequency
tftd = count(t) in d

TF.IDF: Term Freq by Inverse Document Freq

tf.idftd = log(1 + tftd) × log(N/dft)
dft = # docs containing t; N = # of docs
2
G : Probability of different word frequency
E1 = |d| × (tftd + tft(C-d)) / |C|
E2 = |C-d| × (tftd + tft(C-d)) / |C|
2
G = 2 × (tftd log(tftd/E1) + tft(C-d) log(tft(C-d)/E2))
Limitations of Freq. Statistics
Typically focus on unigrams (single terms)

Often favors frequent (TF) or rare (IDF) terms

Not clear that these provide best description

A “bag of words” ignores additional information

Grammar / part-of-speech
Position within document
Recognizable entities
How do people describe text?
We asked 69 subjects (graduate students) to read
and describe dissertation abstracts.

Students were given 3 documents in sequence;

they then described the collection as a whole.

Students were matched to both familiar and

unfamiliar topics; topical diversity within a
collection was varied systematically.
[Chuang, Manning & Heer, 2012]
Bigrams (phrases of 2 words)
are the most common.
Phrase length declines with
more docs & more diversity.
Term Commonness

log(tfw) / log(tfthe)

The normalized term frequency relative to the

most frequent n-gram, e.g., the word “the”.
Measured across a corpus or across the entire
English language (using Google n-grams)
Selected descriptive terms
have medium commonness.
Judges avoid both rare and
common words.
Commonness increases with
more docs & more diversity.
Scoring Terms with Freq, Grammar & Position
G2 Regression Model
Yelp Review Spotlight (Yatani 2011)
Yelp Review Spotlight (Yatani 2011)
Tips: Descriptive Phrases
Understand the limitations of your language model.
Bag of words:
Easy to compute
Single words
Loss of word ordering

Select appropriate model and visualization

Generate longer, more meaningful phrases
Adjective-noun word pairs for reviews
Show keyphrases within source text
Document Content
Information Retrieval
Search for documents
Match query string with documents
Visualization to contextualize results
TileBars [Hearst]
SeeSoft [Eick]
New York Times
Visual Thesaurus [ThinkMap]
Concordance
What is the common local context of a term?
Word Tree [Wattenberg et al.]
Filter Infrequent Runs
Recurrent Themes in Speeches
Glimpses of Structure…
Concordances show local, repeated structure
But what about other types of patterns?

Lexical: <A> at <B>

Syntactic: <Noun> <Verb> <Object>
Phrase Nets [van Ham et al.]

Look for specific linking patterns in the text:

‘A and B’, ‘A at B’, ‘A of B’, etc
Could be output of regexp or parser.
Visualize patterns in a node-link view
Occurrences -> Node size
Pattern position -> Edge direction
Portrait of the Artist as a Young Man
X and Y
Node Grouping
The Bible
X begat Y
Pride & Prejudice
X at Y
Lexical Parser, < 1sec running time
Pride & Prejudice
X at Y
Syntactic Parser, > 24 hours running time
18th & 19th Century Novels
X’s Y
X of Y
X of Y
Document Content
Understand Your Analysis Task
Visually: Word position, browsing, brush & link
Semantically: Word sequence, hierarchy, clustering
Both: Spatial layout reflects semantic relationships
The Role of Interaction
Language model supports visual analysis cycles
Allow modifications to the model: custom patterns
for expressing contextual or domain knowledge
Evolving Documents
Visualizing Revision History

How to depict contributions over time?

Example: Wikipedia history log
Animated Traces [Ben Fry]

http://benfry.com/traces/
Diff
History Flow [Viegas et al.]
Wikipedia History Flow (IBM)
Conversations
Visualizing Conversation
Many dimensions to consider:
Who (senders, receivers)
What (the content of communication)
When (temporal patterns)

Interesting cross-products:
What x When -> Topic “Zeitgeist”
Who x Who -> Social network
Who x Who x What x When -> Information flow
Usenet Visualization [Viegas & Smith]
Show correspondence patterns in text forums
Initiate vs. reply; size and duration of discussion
Newsgroup crowds / Authorlines
Email Mountain [Viegas]
Conversation by person over time (who x when).
Themail [Viegas]

One person over time, TF.IDF weighted terms

Enron E-Mail Corpus
[Heer]
Washington Lobbyist
?
Document Collections
NewsMap: Google News Treemap [Weskamp]
Named Entity Recognition
Label named entities in text:
John Smith -> PERSON
Soviet Union -> COUNTRY
353 Serra St -> ADDRESS
(555) 721-4312 -> PHONE NUMBER

Entity relations: how do the entities relate?

Simple approach: do they co-occur in a small
window of text?
Similarity & Clustering
Compute vector distance among docs
For TF.IDF, typically cosine distance
Similarity measure can be used to cluster

Topic modeling
Assume documents are a mixture of topics
Topics are (roughly) a set of co-occurring terms
Latent Semantic Analysis (LSA): reduce term matrix
Latent Dirichlet Allocation (LDA): statistical model
Parallel Tag Clouds [Collins et al.]
Theme River [Havre et al.]
History of Comp. Ling. [Hall et al.]
Tiara [Wei et al.]
Stanford Dissertation Browser
with Jason Chuang, Dan Ramage & Christopher Manning
Oh, the humanities!
 Computer Science Music

“Word Borrowing” via Labeled LDA

Summary
High Dimensionality
Where possible use text to represent text…
… which terms are the most descriptive?

Context & Semantics

Provide relevant context to aid understanding.
Show (or provide access to) the source text.

Modeling Abstraction
Determine your analysis task.
Understand abstraction of your language models.
Match analysis task with appropriate tools and models.