in an HDPBased Rating Regression Model for Online Reviews Zheng Chen 1 Yong Zhang 1 2 Yue Shang 1 Xiaohua Hu 1 1 Drexel University USA 2 China Central Normal University China ID: 586859
Download Presentation The PPT/PDF document "Unifying Topic, Sentiment & Preferen..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Unifying Topic, Sentiment & Preferencein an HDP-Based Rating Regression Model for Online Reviews
Zheng Chen
1
, Yong Zhang
1,
2
, Yue Shang
1
, Xiaohua Hu
1
1
Drexel University, USA
2
China Central Normal University, ChinaSlide2
Overview
TSPRA
Prediction Model
Reviews with Ratings
Product Aspects
Word Sentiments
User-Aspect Preference
Rating RegressionSlide3
Overview
TSPRA
Prediction Model
Reviews with Ratings
Product Aspects
Word Sentiments
User-Aspect Preference
Rating Regression
Topic-Sentiment-Preference Regression AnalysisSlide4
Overview
TSPRA
Prediction Model
Reviews with Ratings
Product Aspects
Word Sentiments
User-Aspect Preference
Rating RegressionSlide5
Overview
TSPRA
Prediction Model
Reviews with Ratings
Product Aspects
Word Sentiments
User-Aspect Preference
Rating Regression
“topics”, a multinomial distribution over vocabulary
a
value representing from negative to positive sentiments
a
value
representing how much a user cares an aspect
a bunch of distributions inferred from input review dataSlide6
Overview
TSPRA
Prediction Model
Reviews with Ratings
Product Aspects
Word Sentiments
User-Aspect Preference
Rating Regression
Aspect Sentiments
Critical Aspects
aggregated word sentiments
aspects with high user preference but low sentimentSlide7
Overview
TSPRA
Prediction Model
Reviews with Ratings
Product Aspects
Word Sentiments
User-Aspect Preference
Reviews without Ratings
Rating Prediction
Rating Regression
Aspect Sentiments
Critical Aspects
The accuracy of rating prediction serves as an indicator of the model performanceSlide8
MotivationConsider preference and sentiment on a product aspect as independent variables.
JMARS (2014)
FLAME (2015)
user preference
sentiment
rating
user preference functions like topic-level sentimentSlide9
MotivationConsider preference and sentiment on a product aspect as independent variables.
user preference
sentiment
rating
In our model preference
and sentiment
are designed as independent variables that co-determine the review rating
Slide10
MotivationAn automatic approach to build up word sentiment resources. Review ratings tend to be a genuine reflection of the review text sentiments.
The is actually an important application of such rating regression model. However, previous publications cited in the paper do not demo their sentiment result.
Employment of a non-parametric model Hierarchical Dirichlet Process (HDP).
Data an parameters together determine the number of topics.Slide11
Model – HDP
The Chinese Restaurant Franchise Representation.
Document – Restaurant
Document-Level Cluster
– TableWord – CustomerTopic – DishTopic Prior – Franchise
The documents are viewed as being generated by the following process:Every word of a doc is generated from a topic associated with a doc-level cluster. There is a chance that the cluster is new, or the associated topic is new.Or figuratively, when a customer comes into a restaurant, one must choose an existing table, or sits on a new table. The table might order an existing dish, or creates a new dish from the franchise. Then the customer enjoys the dish.
The cluster indexes of words , and the topic indexes of clusters
form the state space. During the Gibbs sampling, a word might chose a new cluster, a cluster might change its associated topic. Thus we keep update and
alternately.
infinite
and
priored version of LDA
G
j
G
0
D
z
di
D
Slide12
Model – HDP
G
j
G
0
D
z
di
D
HDPSlide13
Model – Main Design
(drawn from a binomial distribution), the user preference associated with the topic
of the
th
word in doc
authored by user
.
is dependent on the topic of the word and the author of the doc, so
is designed to be drawn from a binomial distribution
indexed
by topics and users.
There are in total
such binomials, with their prior being Dirichlet distribution with concentration parameter
.
Although the
is computed for each word, it is only dependent on topics and authors.
Slide14
Model – Main Design
(drawn from a 3d multinomial distribution), the word sentiment associated with the
the topic
of the
th
word
in doc
.
is dependent on the topic of the word
and the word
itself, so it is designed to be drawn from a multinomial distribution
indexed
by topics and words.
There are in total
such multinomial
s
, with their prior being Dirichlet distribution with concentration parameter
.
Slide15
Model – Main Design
, the rating of word
computed based on a rule.
strong preference, negative sentiment: rating 1Slide16
Model – Main Design
, the rating of word
computed based on a rule.
strong preference,
positive
sentiment: rating 5Slide17
Model – Main Design
, the rating of word
computed based on a rule.
weak preference
negative
sentiment:
the middle value between 1 and the neutral rating
.
p
ositive sentiment
: the middle value between 5 and the neutral rating
.
It is likely that
.
Slide18
Model – Main Design
, the rating of word
computed based on a rule.
Otherwise, the sentiment is neutral, and the rating should be neutral as well.
Reason for this association rule is to reduce the state space and computation complexity. Besides
table indexes and topic indexes that we need to update for HDP, we only add word ratings
to the state space.
Slide19
Model – Main Design
Design
, where
is the mean of non-neutral word ratings (i.e. average of word ratings excluding neutral words), and
is the rating noise.
Run the Gibbs sampling sufficiently many times, then those multinomials
, binomials
, along with the observable words, ratings, can be viewed as a sample from the designed generative process.
Slide20
Prediction Model
Use result from inference and
are treated as known. The prediction process no longer updates them.
becomes unknown, and needs to be summed out during prediction.
We do not present inference formulas because it is a long story. Please refer to original HDP papers and our paper & presentations for details.
http://www.pages.drexel.edu/~zc86/Slide21
Prediction Evaluation
Dataset
Compare with FLAME (WSDM 2015)
https://snap.stanford.edu/data/web-Amazon.htmlSlide22
Prediction EvaluationComparison with FLAME:
absolute error
, correlation, inverted pairs Slide23
Prediction EvaluationComparison with FLAME: absolute error,
correlation
, inverted pairs Slide24
Prediction EvaluationComparison with FLAME: absolute error, correlation,
inverted pairs
Slide25
Word Sentiments
The probability of word
being positive under topic
The probability of word
being negative under topic
Comparison with S
enticNet3, a public general sentiment resourceSlide26
Word SentimentsSlide27
Word Sentiments
quite neutral words in general contextSlide28
Critical Aspects
The probability of topic
being of high concern by user
The probability of topic
being positive under topic
The probability of word
being negative under topic
We also report a weak Pearson's correlation
between user preferences and sentiments
Slide29
Experiments with Parameters
T
he neutral rating is slightly above 3.
Generally
is good for all tested data sets.
Generally
is good for all tested data sets.
Most people are rounding when rating since
.
Slide30
Contribution Summary
Decoupling of “user preference” from sentiments.
Invention of “critical aspects”.
An approach to automatically generate sentiment resources for online reviews.
First attempt to make use of non-parametric topic models for online reviews.
Thank you!