Emotions from text: machine learning for text-based emotion prediction

1
Emotions from text machine learning for
text-based emotion prediction

• Cecilia Alm, Dan Roth, Richard Sproat
• UIUC, Illinois
• HLT/EMPNLP 2005

2
Objective

• Classify the emotional affinity of sentences in
  the narrative domain of childrens fairy tales
• From the perspective of the story characters

3
Application

• Text-to-Speech synthesis of fairy tales

4
Classification Task

• Experiment 1 Classify a sentence into Emotional
  or Neutral classes
• Experiment 2 Classify a sentence into Neutral,
  Positive Emotion or Negative Emotion classes

5
Corpus

• 1580 manually-annotated sentences from fairy
  tales
• Positive Emotions Happy, Surprised
• Negative Emotions rest
• 90 training, 10 testing

6
Corpus Statistics
7
Classification Method

• SNoW classifier
• 10-fold cross-validation to tweak the parameters

8
Sentence Features

1. First sentence in story
2. Combinations of features (711)
3. Direct speech (quote)
4. Thematic story type (e.g. animal tales)
5. Special punctuation (e.g. ! and ?)
6. Complete upper-case word
7. Sentence length in words
8. Ranges of story progress (e.g. 90-100)
9. Percentages of JJ, N, V, RB
10. Verb count in sentence
11. Positive and negative word counts (Di Cico et
    al.)
12. WordNet emotion words (Fellbaum)
13. Interjections and affective words (Johnson-Laird
    and Oatley)
14. Content BOW N, V, JJ, RB words by POS

9
Experiment 1 Neutral and Emotional

• P(Netural) always predict neutral
• Sequencing use the correct emotion classes of
  adjacent sentences as features
• Columns two sets of paramters

10
Experiment 2 Neutral, Positive Emotion and
Negative Emotion

• Positive Emotions Happy, Surprised
• Negative Emotions Angry, Disgusted, Fearful,
  Sad, -Surprised

11
Cumulative Removal of Feature Groups
12
Conclusion

• Text-based emotion prediction