Visualizing Text: Seth Redmore at the 2015 Smart Data Conference

33
Visualizing Text Smart Data Week Seth Redmore; CMO, Lexalytics, Inc. @sredmore

Transcript of Visualizing Text: Seth Redmore at the 2015 Smart Data Conference

© 2015 Lexalytics Inc. All rights reserved

Visualizing TextSmart Data Week

Seth Redmore; CMO, Lexalytics, Inc.@sredmore

© 2015 Lexalytics Inc. All rights reserved

Agenda

The Word Cloud Vectors to Visualize Ways to group/count Manipulating the words (stemming/lemmatization/etc) Line/Bubble/Pie Treemaps Heatmaps Clusters Graphs

2

© 2015 Lexalytics Inc. All rights reserved

The Word Cloud

3

© 2015 Lexalytics Inc. All rights reserved

Which word is gone?

4

© 2015 Lexalytics Inc. All rights reserved

How about now?

• stem 86•word 53• algorithm 49• rule 36• suffix 27• strip 23• approach 21• form 21

• language 20• edit 20• example 18• root 18• apply 14• search 13• inflect 12• english 10

• stem 86•word 53• algorithm 49• rule 36• strip 23• approach 21• form 21• language 20

• edit 20• example 18• root 18• apply 14• search 13• inflect 12• English 10• part 10

5

© 2015 Lexalytics Inc. All rights reserved

Visualization vectorsContent Derived Associated Metadata

• Stemmed Words/Words/Phrases• Part-of-Speech• Extracted Features

– Entities– Themes– Topics– Intentions

• Sentiment/Emotions

• Language• Geography• Time• Publication/Author/@handle• Socioeconomic• Social associations

6

© 2015 Lexalytics Inc. All rights reserved

Ways to group or count

• Weighting Factors– Counts– “Importance”

• Similarity• Co-occurrence

– Categories– Other words

7

© 2015 Lexalytics Inc. All rights reserved

Pies (one axis)

Positive: 28.65%Negative: 9.16%Neutral: 62.20%

For any more than 3 data points pie charts become increasingly hard to read.If you have 3 or fewer data points, why not just use a table?

8

28.65%

9.16%

62.20%

© 2015 Lexalytics Inc. All rights reserved

What is the “true” Sentiment?

-0.1 to +0.1 is neutral-0.2 to +0.2 is neutral

Positive: 28.65%Negative: 9.16%Neutral: 62.20%

Positive: 29.77%Negative: 9.99%Neutral: 60.24%

9

28.65%

9.16%

62.20%

29.77%

9.99%

60.24%

© 2015 Lexalytics Inc. All rights reserved

Lines (2 axes)

10

© 2015 Lexalytics Inc. All rights reserved

Bars

11

© 2015 Lexalytics Inc. All rights reserved

Bubbles (4 axes)

Courtesy of Provalis Research

12

© 2015 Lexalytics Inc. All rights reserved

Stemmed Words vs. Words vs. Word Phrases vs. Relationships

• I was greatly satisfied with my dinner.• Greatly satisfied• Greatly • Great• I hate the cracked screen on my phone.• Cracked screen• Crack

Satisfied(x1.5) dinner

Cracked Screen phone

13

© 2015 Lexalytics Inc. All rights reserved

LemmatizationStemming

Walking WalkBetter BetterI am meeting him tomorrow

Meeting MeetIn our last meeting, we…

Meeting Meet

Walking WalkBetter GoodI am meeting him tomorrow

Meeting MeetIn our last meeting, we…

Meeting Meeting

Stemming vs. Lemmatization

Examples from Wikipedia

14

© 2015 Lexalytics Inc. All rights reserved

Top themes from Samsung Galaxy® Announcement

Themes are contextually scored noun-phrases.

15

© 2015 Lexalytics Inc. All rights reserved

Top themes + relative occurrence

16

© 2015 Lexalytics Inc. All rights reserved

Plus Sentiment

17

© 2015 Lexalytics Inc. All rights reserved

+Time

18

© 2015 Lexalytics Inc. All rights reserved

+Sentiment

19

© 2015 Lexalytics Inc. All rights reserved

+Gender (too much!)

20

© 2015 Lexalytics Inc. All rights reserved

GenderThemeSentiment

21

Important to consider how you can get the structured data in there with the unstructured data.

© 2015 Lexalytics Inc. All rights reserved

Word Cloud

22

© 2015 Lexalytics Inc. All rights reserved

Treemap

23

© 2015 Lexalytics Inc. All rights reserved

Treemap Comparison

24

© 2015 Lexalytics Inc. All rights reserved

Usenet Treemap

Treemaps are good for data that has hierarchy

25

© 2015 Lexalytics Inc. All rights reserved

Force-directed Graphs

Courtesy of Bottlenose

http://www.d3noob.org/2013/03/d3js-force-directed-graph-examples.html

26

© 2015 Lexalytics Inc. All rights reserved

Clustering

Courtesy of Quid

27

© 2015 Lexalytics Inc. All rights reserved

Clustering Zoom

Courtesy of Quid

© 2015 Lexalytics Inc. All rights reserved

Heatmaps

Courtesy of Provalis Research

29

© 2015 Lexalytics Inc. All rights reserved

CodeNo-Code

• Datawrapper– Built for news orgs, better with

structured data• Charted

– Input CSV or google spreadsheet• Tableau Public• Google Charts

• D3– Hugely powerful, many relevant chart

types for text– https://github.com/mbostock/d3/

wiki/Gallery• R

– Full blown stats + visualization

Open Source/Free Tools and Toolsets

30

© 2015 Lexalytics Inc. All rights reserved

Full Analytics Systems (with content)Graphing/Charting

• Tableau• Jreport• Domo• Qlik• Tibco Spotfire• Wordstat/Simstat • SAS• SPSS

Many of them. We work with lots of them, so, I can’t list them all here.

Commercial Toolkits

31

© 2015 Lexalytics Inc. All rights reserved

Summary• Don’t use pie charts, use tables instead.• Don’t use word clouds if you can avoid them.• Really don’t use word clouds for any sort of comparison over time.• If you’re going to use word clouds

– use intelligent colors

– use them either as a user-interface

– or use them when you’ve already done a bunch of filtering

• Many other chart types have the visual appeal of word clouds while providing more information.– Time-series charts

– Treemaps

– Force Directed Graphs

– Clusters

– Heatmaps

32

And check this out…http://www.ted.com/talks/hans_rosling_shows_the_best_stats_you_ve_ever_seen?language=en

© 2015 Lexalytics Inc. All rights reserved