Sentiment Expression Conditioned by Affective Transitions and Social - - PowerPoint PPT Presentation

Sentiment Expression Conditioned by Affective Transitions and Social Forces

Moritz Sudhof Andrés Goméz Emilsson Andrew L. Maas Christopher Potts

KDD 2014 ¡

Sentiment Expression Conditioned by Affective Transitions and Social Forces

Moritz Sudhof Andrés Goméz Emilsson Andrew L. Maas Christopher Potts

KDD 2014 ¡

Stanford University ¡ Psychology ¡ Linguistics ¡ Computer Science ¡

Sentiment Expression Conditioned by Affective Transitions and Social Forces

Human emotional states are not independent, random, or isolated. They proceed along systematic paths and are conditioned by the states of others in our communities. We can improve sentiment classification by incorporating information about emotion sequences using CRFs.

Affective Transitions

Emotions aren’t random. Your current emotional state is heavily influenced by your previous emotional state.

Affective Transitions ¡

Studying Affective Transitions

Experience Project mood status corpus

alive sleepy stressed

• ptimistic

bored blah cheerful confused amused annoyed anxious hopeful lonely tired sad excited depressed calm horny happy 25323 26316 26777 28569 28643 29119 29235 29850 31609 33220 34998 37504 51590 52097 52975 63035 65614 76344 77209 89344

Moods Corpus

2 million posts

Emotion Transition Probabilities

Emotion Transitions

log-scale CTP(a,b)

bewildered artistic curious bouncy chill distressed disappointed excited bewildered crushed chill chipper lazy bitchy bored chipper blissful bouncy energetic flirty numb lonely devastated angry melancholy amused anxious blah cheerful depressed

0.03 0.04 0.06 0.1

Emotion Transitions

Worried

Hopeful

Blessed

Social Forces

We are social animals. Our opinions and expressions are influenced by the opinions and expressions of people in our community.

Social Forces ¡

Studying Social Forces

Review Ratings

Rating Reviews 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0 0.3K 1.3K 2.5K 3.3K 4.8K 7.9K negative (18% of reviews) positive (25% of reviews) neutral (58% of reviews)

2.9 million reviews

Review Sequence Transitions

negative neutral positive positive neutral negative 0.07 0.21 0.51 0.4 0.67 0.44 0.53 0.12 0.04 (a) All product sequences

negative neutral positive positive neutral negative 0.27 0.27 0.27 0.56 0.56 0.56 0.17 0.17 0.17 negative neutral positive positive neutral negative 0.05 0.11 0.35 0.3 0.6 0.47 0.66 0.28 0.18

Review Sequence Transitions

negative neutral positive positive neutral negative 0.07 0.21 0.51 0.4 0.67 0.44 0.53 0.12 0.04

(a) All product sequences (b) High-variance sequences (c) Randomized sequences

Classification Experiments

Conditional Random Field (CRF) vs. MaxEnt

Experimental Setup

• CRF and MaxEnt
• Simple unigram features
• L2 regularization
• For each model, choose the regularization

penalty by cross-validating over possible values.

• Run 20 trials: randomly split the data 80/20,

train, test, rinse repeat.

• Non-parametric Wilcoxon rank-sums test

measures the significance of the difference between CRF and MaxEnt performance.

micro−average macro−average negative positive

MaxEnt CRF

0.0 0.2 0.4 0.6 0.8 1.0

0.81 0.8 0.77 0.84 0.8 0.79 0.75 0.83

* * * *

Affective Transitions: Polarity

F1 Score

Affective Transitions: Emotions

F1 Score

micro−average macro−average depressed hopeful satisfied cheerful stressed anxious

MaxEnt CRF

0.0 0.2 0.4 0.6 0.8 1.0

0.51 0.43 0.64 0.53 0.39 0.36 0.34 0.31 0.49 0.4 0.62 0.51 0.36 0.31 0.31 0.28

* * * * * * * *

micro−average macro−average negative neutral positive

MaxEnt CRF

0.0 0.2 0.4 0.6 0.8 1.0

0.72 0.69 0.66 0.77 0.63 0.71 0.67 0.63 0.76 0.61

* * * * *

Social Forces

F1 Score

Conclusion

Social and psychological forces influence our expression. If sentiment models are sensitive to these contextual forces, we can make better decisions.

Thank You

sudhof@stanford.edu