User Behavior Analysis for Sentiment Classification

User Modeling with Neural Network for Review Rating Prediction.

Duyu Tang, Bing Qin, Ting Liu. IJCAI 2015, full paper.

Learning Semantic Representations of Users and Products for Document Level Sentiment Classification. Duyu Tang, Bing Qin, Ting Liu. ACL 2015, full paper.

1

Sentiment Classification

• Given a piece of text, the task aims to determine its polarity as • Positive / Negative

• 1-5 stars

• The task can be at• Word/phrase level, sentence level, document level

• We target at document-level sentiment classification in this work

2

Standard Supervised Learning Pipeline

TrainingData

Learning Algorithm

FeatureRepresentation

SentimentClassifier

3

Feature Learning Pipeline

TrainingData

Learning Algorithm

FeatureRepresentation

SentimentClassifier

Learn text representation/feature from data!

4

Deep Learning Pipeline

TrainingData

Learning Algorithm

FeatureRepresentation

SentimentClassifier

Word Representation

Words

Semantic Composition

w1 w2 …… wn−1wn

5

User Behavior is important for SA

• From a sentiment analysis perspective , users have different habits to• Assign ratings on IMDB, Yelp …

• Use diverse sentiment words to express one’s sentiment

User 1

User 2

Rating Histories Frequently used words

good, ok, just soso, disgusting…

awesome, amazing, excellent, not bad…

6

7

Softmax gold rating = 2

w1

h1 h2 hn

Lookup

Linear

……

Convolution

Poolingvd

Tanh

w2 wn

• Text semantics

The Approach

wi: word

The Approach

8

Softmax gold rating = 2

w1

h1

Uk

h2 hn

Lookup

Linear

……

Convolution

Poolingvd

Tanh

×

w2Uk

×

wnUk

×

• Text semantics• +User-Text Associations

Uk: user wi: word

The model in IJCAI 2015

The Approach

9

Softmax gold rating = 2

w1

h1

Uk

h2 hn

Lookup

Linear

……

Convolution

Poolingvd

Tanh

×

w2Uk

×

wnUk

×

• Text semantics• +User-Text Associations

• +User-Sentiment Associations

Uk: user wi: word

uk

The Approach

10

Softmax gold rating = 2

w1

h1

Uk Pj

h2 hn

Lookup

Linear

……

Convolution

Poolinguk pjvd

Tanh

w1

× ×

w2Uk Pj w2

× ×

wnUk Pj wn

× ×

• Text semantics• +User-Text Associations +Product-Text Associations

• +User-Sentiment Associations +Product-Sentiment

Pj: product Uk: user wi: word

The model in ACL 2015

Experiments

• We conduct supervised learning on three datasets

• Some details• We build the datasets by ourselves.

• Split each corpus into training, developmentand testing sets with a 80/10/10 split

• IMBD is from Diao et al. 2014

• Yelp 2013 & 2014 come from Yelp Dataset Challenge11

Experimental Results

• On IMDB dataset

12

Our Approach

Experimental Results

• On Yelp 2014 dataset

13

Our Approach

Experimental Results

• On Yelp 2013 dataset

14

Our Approach

In Summary

• We encode users (and products) in semantic vector space, and apply them to sentiment classification.

• User and product representations can improve the sentiment classification accuracy.

15

Thanks!

codes and resources will be publicly available at: http://ir.hit.edu.cn/~dytang

16