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 Stanford University ¡ Moritz Sudhof Computer Science ¡ Andrés Goméz Emilsson Psychology ¡ Andrew L. Maas Christopher Potts Linguistics ¡ KDD 2014 ¡

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

Moods Corpus 2 million posts 89344 happy 77209 horny 76344 calm 65614 depressed 63035 excited 52975 sad 52097 tired 51590 lonely 37504 hopeful 34998 anxious 33220 annoyed 31609 amused 29850 confused 29235 cheerful 29119 blah 28643 bored 28569 optimistic 26777 stressed 26316 sleepy 25323 alive

Emotion Transition Probabilities

Emotion Transitions amused anxious blah cheerful depressed 0.1 numb log-scale CTP(a,b) chipper blissful chill lonely bouncy 0.06 devastated distressed energetic bewildered flirty artistic curious angry chipper bouncy lazy chill melancholy bitchy 0.04 bored disappointed excited bewildered crushed 0.03

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

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

Review Sequence Transitions positive neutral negative positive neutral negative positive neutral negative positive 0.51 0.44 0.04 positive 0.35 0.47 0.18 positive 0.27 0.56 0.17 neutral 0.21 0.67 0.12 neutral 0.11 0.6 0.28 neutral 0.27 0.56 0.17 negative 0.07 0.4 0.53 negative 0.05 0.3 0.66 0.27 0.56 0.17 negative (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.

Affective Transitions: Polarity MaxEnt CRF 0.83 positive * 0.84 0.75 negative * 0.77 0.79 macro − average * 0.8 0.8 micro − average * 0.81 0.0 0.2 0.4 0.6 0.8 1.0 F1 Score

Affective Transitions: Emotions 0.28 anxious * 0.31 MaxEnt CRF 0.31 stressed * 0.34 0.31 cheerful * 0.36 0.36 satisfied * 0.39 0.51 hopeful * 0.53 0.62 depressed * 0.64 0.4 macro − average * 0.43 0.49 micro − average * 0.51 0.0 0.2 0.4 0.6 0.8 1.0 F1 Score

Social Forces MaxEnt CRF 0.61 positive * 0.63 0.76 neutral * 0.77 0.63 negative * 0.66 0.67 macro − average * 0.69 0.71 micro − average * 0.72 0.0 0.2 0.4 0.6 0.8 1.0 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