PowerPoint-Pr

1
Extraction and Visualisation of Emotion from News
Articles
Eva Hanser, Paul Mc Kevitt
School of Computing Intelligent Systems Faculty
of Computing Engineering University of Ulster,
Magee hanser-e_at_email.ulster.ac.uk,
p.mckevitt_at_ulster.ac.uk
2
News Visualisation Emotion Extraction
1 Introduction What is NewsViz?
2 Background Related Projects
3 Design Implementation The NewsViz
Application
4 Prototype Demonstration
6 Testing
7 Relation to Other Work
8 Conclusion Future Work
3
News Visualisation Emotion Extraction
What is NewsViz?
From natural language to visual
presentation NewsViz automatically produces
animations from text
Input
Online News Article
NewsViz System
Output
Animation
4
News Visualisation Emotion Extraction
What is NewsViz?
Aim Animation embedded into news
website Objectives Entertainment Quick
overview Emotional aspects
gtgt view website
5
News Visualisation Emotion Extraction
What is NewsViz?

• The Challenges
• 1. Natural Language Processing
• (computational interpretation of meaning of
  text)
• 2. Automatic creation of animations
• A manageable project
• Prototype limited to one topic football news
• Focus on determining emotional aspects
• Reduced to background visualisation

6
News Visualisation Emotion Extraction
Related Projects
Qtag

• Syntactic Analysis (based on grammar)
• Part-of-Speech Tagging (e.g. Qtag)
• identifying word types such as nouns,
  adjectives, verbs,
• 95-97 correct

Tag-list
Tagged text
Bayern_VB Munich_NP stretched_VBD their_DPS
lead_NN at_PRP the_AT top_NN as_CJS Hamburg_NP
suffered_VBD a_AT tragic _JJ surprise_NN home_NN
loss_NN ._.
PRP preposition JJ adjective, general NN noun,
common singular NNS noun, common plural NP noun,
proper singular VB verb, base from VBD verb, past
tense . . .
http//www.english.bham.ac.uk/staff/omason/softwar
e/qtag.html
7
News Visualisation Emotion Extraction
Related Projects
WordsEye

WordsEye Creates static 3D scenes from text input
http//www.wordseye.com
8
News Visualisation Emotion Extraction
Related Projects
WordsEye

WordsEye Description and Rendered Image
The skiff is on the ocean. The grassy mountain
is 20 feet behind the boat. The dog is in the
boat. The fishing pole is two feet in front of
the dog. The bottom of the palm tree is below
the bottom of the mountain. It is 20 feet behind
the boat.
http//www.wordseye.com
9
News Visualisation Emotion Extraction
Related Projects
WordsEye

More Syntax Analysis Structure of
Sentences Dependency Parser (e.g. used in
WordsEye) Finding relations between words and
phrases Dependency rules
http//www.wordseye.com
10
News Visualisation Emotion Extraction
Related Projects
WordsEye

• Graphical Database in WordsEye
• 3D objects, their attributes (colour, size,
  surface)

http//www.wordseye.com
11
News Visualisation Emotion Extraction
Related Projects
WordNet

Semantic Analysis (based on meaning) Lexical
Knowledgebase (e.g. WordNet) sets of synonymous
words and basic semantic relations
http//wordnet.princeton.edu/
12
News Visualisation Emotion Extraction
Related Projects
Story Picturing Engine

• The Story Picturing Engine
• matching keywords image regions

step 1 filtering out common words (a, the,
of, ) step 2 identification of proper words
(places and people involved) step 3
similarity count of remaining keywords (words
with too many meanings are too vague)
further steps for image processing
13
News Visualisation Emotion Extraction
Related Projects
Story Picturing Engine

Example text on walk through ParisH highest
ranked images, L Lowest ranked images
14
News Visualisation Emotion Extraction
NewsViz
NewsViz Architecture
15
News Visualisation Emotion Extraction
NewsViz
Emotion Visualiser
16
News Visualisation Emotion Extraction
NewsViz
Graphics Database Abstract Visuals for 4
Emotions
2 - boring
1 - sad
4 - happy
3 - tense
17
News Visualisation Emotion Extraction
NewsViz
Word Lexicon with Emotion Indices
ltwordgt ltnamegtchallengeslt/namegt ltmoodgt3lt/moodgt
lt! tension ? ltintensitygt3lt/intensitygt ltsynonyms
gtlt/ synonyms gt lt/wordgt ltwordgt ltnamegthomelt/name
gt ltmoodgt4lt/moodgt lt! happy ? ltintensitygt1lt/inte
nsitygt lt/wordgt ltwordgt ltnamegtgapslt/namegt ltmoodgt1lt
/moodgt lt! sad ? ltintensitygt2lt/intensitygt lt/wordgt
18
News Visualisation Emotion Extraction
NewsViz
Summarization Options
19
News Visualisation Emotion Extraction
Demonstration
20
News Visualisation Emotion Extraction
Demonstration
21
News Visualisation Emotion Extraction
Demonstration
22
News Visualisation Emotion Extraction
Demonstration
23
News Visualisation Emotion Extraction
Demonstration
24
News Visualisation Emotion Extraction
Demonstration
25
News Visualisation Emotion Extraction
Demonstration
26
News Visualisation Emotion Extraction
Demonstration
27
News Visualisation Emotion Extraction
Demonstration
28
News Visualisation Emotion Extraction
Testing
Procedure NewViz performance evaluated
against human interpretation 1.
General mood course (3-5 emotions per text) 2.
1-2 Emotions per sentence types of emotion
extraction error Falsely detected emotion 0
points Missing emotion points depending on
significance Overall feeling represented, 2-3
points Similar emotion 4 points Exact
emotion 5 points
29
News Visualisation Emotion Extraction
Testing
Results Course of moods mostly identified
correctly Word-by-Word method highest
correctness but too fine grained for
animation Best results with both (adjective
and nouns)
Method Method Word by Word by Sentence Sentence Threshold Threshold average
Word based based 2 3
Word type Word type correct grain correct grain correct grain correct grain
adjectives adjectives adjectives 3.125 12 3.25 7.5 2.375 5 1.25 2.3 2.5
nouns nouns 3.875 31 2.625 9.3 2.875 14 2 4.8 2.844
both both 4 33 2.75 9.5 3.5 18 1.5 10 2.938
average average 3.667 25 2.875 8.8 2.917 12 1.583 5.7
30
News Visualisation Emotion Extraction
Conclusion Future Work
Summary Emotional interpretation of online
news articles Course of moods depicted in
abstract 2D animations Different methods of
language processing Satisfactory
outcome User Evaluation Appreciation of
animations as quick overviews Future Work
Extension of knowledge bases Inclusion of
different topics Improvement of
summarisation, e.g dependency parser