2015 IEEE International Conference on Data Mining

Complementary Aspect-based Opinion Mining

Across Asymmetric Collections
Yuan Zuo1 , Junjie Wu2 , Hui Zhang1 , Deqing Wang1 , Hao Lin2 , Fei Wang1 , Ke Xu1
1 School

of Computer Science and Engineering, Beihang University, China

of Economics and Management, Beihang University, China
corresponding author: [email protected]

2 School

environment protection, etc. Aspect-based opinion mining

technique becomes an intuitive candidate to fulll this task.
Moreover, the diverse yet complementary information provided by rich online media of various types brings great
opportunities for public opinion analysis across different collections. Indeed in the literature, there have been quite some
excellent studies on cross-collection topic modeling [21],
[1], [4], [5]. However, they either pay little attention to the
complementarity of aspects across collections [4], or just focus
on topics and aspects without considering the opinions [5].
Therefore, further study is still in great need for building crosscollection aspect-based opinion mining model, based on which
diversity and complementarity in both aspects and opinions
could be leveraged across collections containing substantially
asymmetric information, e.g., the news collection with clear
aspects versus the tweets collection with strong opinions.
To address the above challenge, in this paper, we propose CAMEL (Cross-collection Auto-labeled MaxEnt-LDA),
a novel topic model for complementary aspect-based opinion
mining across asymmetric collections. To our best knowledge,
our work is among the earliest studies in this direction.
CAMEL is essentially a type of cross-collection LDA model,
which models aspect-level opinions and gains complementarity by modeling both common and specic aspects across
different collections. By keeping all the corresponding opinions for both common and specic aspects, CAMEL is also
capable of contrastive opinion analysis. Moreover, as a booster
to CAMEL, we propose AME, an automatic labeling scheme
for maximum entropy model. It helps discriminate aspect and
opinion words without heavy human labeling.
We conducted extensive experiments on synthetic multicollection data sets to evaluate the quality of aspects as well as
opinions induced by CAMEL. Specically, we design a sentence classication experiment to justify that CAMEL can nd
higher-quality aspects than baseline methods, and shows more
robust performances especially with imbalanced collections in
varying degrees. Besides, CAMEL exhibits obvious superiority
in learning more coherent opinions and more relevant aspects
and opinions in terms of the coherence measure. Also, the
AME model for CAMEL indeed outperforms manual labeling
in distinguishing aspect words from opinion ones. Finally, case
study on two public events further demonstrates the practical
value of CAMEL for real-world public opinion analysis.

AbstractAspect-based opinion mining is to nd elaborate

opinions towards an underlying theme, perspective or viewpoint
as to a subject such as a product or an event. Nowadays, with
rapid growing of opinionated text on the Web, mining aspect-level
opinions has become a promising means for online public opinion
analysis. In particular, the booming of various types of online
media provide diverse yet complementary information, bringing
unprecedented opportunities for public opinion analysis across
different populations. Along this line, in this paper, we propose
CAMEL, a novel topic model for complementary aspect-based
opinion mining across asymmetric collections. CAMEL gains
complementarity by modeling both common and specic aspects
across different collections, and keeping all the corresponding
opinions for contrastive study. To further boost CAMEL, we propose AME, an automatic labeling scheme for maximum entropy
model, to help discriminate aspect and opinion words without
heavy human labeling. Extensive experiments on synthetic multicollection data sets demonstrate the superiority of CAMEL to
baseline methods, in leveraging cross-collection complementarity
to nd higher-quality aspects and more coherent opinions as well
as aspect-opinion relationships. This is particularly true when the
collections get seriously imbalanced. Experimental results also
show that the AME model indeed outperforms manual labeling in
suggesting true opinion words. Finally, case study on two public
events further demonstrates the practical value of CAMEL for
real-world public opinion analysis.
Index TermsAspect-based Opinion Mining; Topic Detection
and Tracking; LDA Model; Maximum Entropy Model

I. I NTRODUCTION
With the dramatic growth of opinionated user generated
content on the Web, to automatically understand, extract and
summarize the public opinions expressed in different online
media platforms has therefore become an important research
topic and gained much attention in recent years [13], [24].
Aspect-based opinion mining, a technique proposed originally
for nding elaborate opinions towards a perspective of a
product [19], has become a promising means for mining
aspect-level opinions for online public opinion analysis, where
the concept of an aspect here has been extended to be an
underlying theme, perspective or viewpoint as to a public
event. For instance, for the annual key event Two Sessions
(of the NPC and the CPPCC) 2015 in China, we would like
to know the elaborate public opinions towards a plenty of
relatively focused themes that have generated heated discussions, e.g., the downward pressure on GDP, the opportunities
in Jing-Jin-Ji integration, the Hukou reform, anti-corruption,
1550-4786/15 $31.00 2015 IEEE
DOI 10.1109/ICDM.2015.97

669

II. P ROBLEM D EFINITION

based opinion mining across complementary collections.

Specically, we rst describe the key points of CAMEL
as well as the generative process under CAMEL. We then
introduce the Auto-labeled MaxEnt model (AME for short)
that enables the subsequent joint modeling of aspects and
opinions without expensive human labeling. We nally present
the approximate posterior inference for CAMEL.

Our work in this paper focuses on public opinion mining

across multiple media collections. Specically, we aim to
answer the following interesting questions: 1) What are the
main concerns with a public event for users being active in
different media platforms? Do these concerns share anything
in common or just scatter in different media? 2) What are the
public opinions to these concerns? Do the users in different
media platforms have consistent or diversed opinions?
To answer these questions, we rst need an aspect-based
opinion mining model for capturing public concerns and
opinions simultaneously from text collections. Here aspect
means an underlying theme, perspective or viewpoint as to
a subject like an event or a product, with the assumption that
each sentence is generated by a single aspect. It has been
reported that about 83% sentences in online reviews cover a
single aspect [27], implying this assumption holds particularly
for online social media such as Twitter or Chinese Weibo.
In addition, we need to build a cross-collection framework
for the aspect-based opinion model so as to enable information
integration from different collections. Actually a crosscollection model could benet more. Suppose we want to
mine opinions from both news and micro-blogs collections.
Since micro-blogs are mostly generated by public users,
referred as user generated content (UGC), we could expect
sharper opinions from UGC but less clear concerns or aspects
due to the more emotional and colloquial expression way.
This is in sharp contrast to the news, where the concerns
are usually very clear but the opinions are often monotonous
and implicit. In other words, we have asymmetric collections
or complimentary collections for public opinion analysis.
Therefore, an intuitional way to display both sides respective
advantages is to use news to help tweets identify meaningful
aspects and to use tweets to enrich news by diverse opinions.
This is what we called Complementary Aspect-based Opinion
Mining task dened as follows:

A. Model Description
As illustrated in Sect. II, our main task is to design a
model that can leverage complementary information from
diverse collections to jointly model aspects and opinions for
public opinion analysis. Along this line, two key problems
should be well addressed in the model. The rst one is how
to model aspects and opinions hidden in a collection in a
simultaneous and automatical way, which will be discussed
in detail in Sect. III-B. The second problem is how to capture
the complementarity across multiple collections with possible
severely asymmetric information, which is the main focus of
this subsection.
Let us again consider the case of aspect-based opinion mining from two asymmetric collections, i.e., the news collection
with clear aspects and the tweet collection with sharp opinions.
In the aspect level, to help extract clear aspects from tweets, it
is intuitive to share with the tweets side some similar aspects
found from the news side. This can be done by mining aspects
from the two collections separately, and then linking together
similar aspects from different sides. However, it would be
very difcult, if not impossible, to dene a proper similarity
measure and set a good threshold to it. Therefore, it would be
better to design a cross-collection model that could directly
mine some common aspects shared by different collections. In
the opinion level, however, it would be more interesting to read
public opinions from the tweet side, and compare them with
the opinions from the news side, which are often regarded as
the mainstream opinions from authoritative media. As a result,
our cross-collection model should be able to mine the opinions
separately from different sides for the purpose of comparison.
Based on the above reasoning, we now can describe our
model CAMEL, and give the generative process under it.
CAMEL is essentially a cross-collection LDA model with a
maximum entropy model embedded to determine the priors
for aspect and opinion words switching. CAMEL assumes
that different collections not only share some common aspects
but also have aspects of their own. Hereinafter, we call
aspects shared across collections as common aspects, and
call the aspects only contained in one collection as specic
aspects. CAMEL also assumes that each specic aspect has a
corresponding opinion, and each common aspect has multiple
corresponding opinions, one for each collection. We now describe how to generate a document under CAMEL as follows.
Suppose there are several multinomial word distributions
from a symmetric Dirichlet prior with parameter , including:
KI
K I common aspects {A
z }z=1 shared by all collections, with
O C
C opinions {z,c }c=1 for each A
z , where C is the number
S
S
S
corof collections; K specic aspects {zA }K
z=1 and K

Given multiple text collections about a subject, design a

cross-collection model that can leverage complementary
information from different collections to form aspectsbased opinions for comprehensive and contrastive public opinion analysis.
Remark. There have been some studies on cross-collection
topic modeling in the literature, and to our best knowledge the
ones most related to our task include [4] and [5]. While [4]
also studies contrastive opinion mining problem, it does not
jointly model aspects and opinions and pays little attention
to the complementarity of aspects across collections, which
however is our main focus. The task of [5] is to summarize
text across complementary collections, but it ignores public
opinions totally, which is also the main theme of our study.
III. M ODEL AND I NFERENCE
In this section, we introduce the Cross-collection Autolabeled MaxEnt-Lda model (CAMEL for short) for aspect-

670

KI

TABLE I: Math notations.

Notation
KI
KS
C
D
S
N
A
O
A
O
I
S

O
KS

{O, A}
C

Fig. 1: Plate notation of CAMEL.

S

w
z
x

responding opinions {zO }K

z=1 , one for each collection. All
these are multinomial distributions over the vocabulary, which
we assume has V words in total. Note that we here assume
all collections have the same number of specic aspects for
briefness, which could be relaxed to allow variant easily.
For sentence s in document d, we draw rd,s from a Bernoulli
distribution over {0, 1} parameterized by , which in turn
is drawn from a symmetric Beta(). rd,s is an indicator
of whether sentence s is generated by common or specic
aspects. Specically, when rd,s = 0, we assume a sentence
is generated by a common aspect, otherwise by a collectionspecic aspect. For word n in sentence s, we introduce an
indicator variable yd,s,n for aspect and opinion switching,
which is drawn from a Bernoulli distribution over {0, 1}
parameterized by . Similar to rd,s , we assume a word is
generated by an aspect-word distribution when yd,s,n = 0,
otherwise by an opinion-word distribution. According to some
previous studies [16], [14], topic models that set with
symmetric priors are unable to identify opinion words well.
Therefore, to set for word wd,s,n , we utilize the weights
learned by the maximum entropy component and the feature
vector xd,s,n of w, which give:
p(yd,s,n = l|xd,s,n ) = ld,s,n =

y
r

Bern()
Beta()
M ulti()
Dir()

Description
the # of common aspects in total
the # of specic aspects for each collection
the # of collections
the # of documents in a collection
the # of sentences in a document
the # of words in a sentence
common aspect-word distribution
common opinion-word distribution
specic aspect-word distribution
specic opinion-word distribution
common aspect mixture for a document
collection specic aspect mixture for a document
parameters for common and specic aspect
switching for a sentence
parameters for aspect and opinion switching
for a word
an observed word
aspect index for a sentence
feature vector for the maximum entropy
(MaxEnt) model
weights learned by the MaxEnt model
aspect and opinion switcher for a word
common and specic aspect switcher
for a sentence
Dirichlet prior parameter for
Dirichlet prior parameter for all word distributions
symmetric Beta prior parameters for
Bernoulli distribution with parameter()
Beta distribution with parameter()
Multinomial distribution with parameter()
Dirichlet distribution with parameter()

b) Choose d Dir()
c) Choose d Beta()
d) For each sentence s:
i) Choose rd,s Bern(d )
ii) if rd,s = 0 choose zd,s M ulti(dI )
if rd,s = 1 choose zd,s M ulti(dS )
iii) For each word n:
A) Choose yd,s,n Bern(d,s,n )
B) if rd,s = 0 and yd,s,n = 0 choose
wd,s,n M ulti(A
zd,s )
if rd,s = 0 and yd,s,n = 1 choose
wd,s,n M ulti(O
zd,s ,c )
if rd,s = 1 and yd,s,n = 0 choose
wd,s,n M ulti(zAd,s ,c )
if rd,s = 1 and yd,s,n = 1 choose
wd,s,n M ulti(zOd,s ,c )

exp(l xd,s,n )
,
1l =0 exp(l xd,s,n )

where {0 , 1 } denote the weights learned by the maximum

entropy model upon a set of training data, whose labels are
obtained by an automatic procedure described in the next
subsection.
The plate notation of CAMEL is shown in Fig. 1, with a
summary of math notations provided in Table I. The generative
process is thus described as follows:
1) For each common aspect z:
a) Choose A
z Dir()
2) For each collection c:
a) Choose O
z,c Dir() for each common aspect z
A
O
, z,c
Dir() for each collectionb) Choose z,c
specic aspect z
3) For each document d:
a) Choose a collection indicator c

B. Auto-labeled MaxEnt Model

In order to obtain aspect-specic opinions, we adopt the
MaxEnt-LDA model proposed by Zhao et al. in [27], where
a maximum entropy (MaxEnt) model is trained with Part-OfSpeech (POS) tags of words serving as priors for aspect and
opinion switching. This is motivated by the fact that aspect
and opinion terms normally play different syntactic roles in
a sentence, but it also suffers from the high cost of manual
labeling of word tags.
To address this problem, we propose a procedure to label
training data automatically, and thus form the so-called Autolabeled MaxEnt model (AME). It is motivated by the observa-

671

tion that opinion words usually do not appear nearly to each

other in a sentence. This in other words implies that a word
nearby a known opinion word is likely to be a non-opinion
word. The following gives the details of the procedure:
1) We rst randomly select a set of M opinion words
(O)
{Vm }M
m=1 from a general opinion lexicon, which is
usually publicly available for many languages. Those
selected words are required to have relatively large
document frequency in the target corpus.

A,j,k
is the total number
opinion word to common aspect k. C()
of times any word is assigned as an aspect word to common
O,j,k
is the total number of times any word is
aspect k, and C()
A,j,k
assigned as an opinion word to common aspect k. N(v)
is
the number of times word v is assigned as an aspect word to
O,j,k
aspect k in sentence s of document d, and similarly, N(v)
is the number of times word v is assigned as an opinion word
to aspect k in sentence s of document d. When j = 1, all
d
is the
counts mentioned above refer to specic aspects. C()
number of sentences in document d. Note that all these counts
represented by symbol C exclude sentence s of document d.
With assignments of z and r, we can sample y(d,s,n) for
y(d,s,n) = 0:

2) We then randomly choose a set of sentences S such

that each sentence in S contains at least one opinion
(O)
word in {Vm }M
m=1 .
3) We label a word as an aspect word if it is not in
(O)
{Vm }M
m=1 and appears nearly to a known opinion word
in a sentence contained by S. In this way, we nally
(O)
obtain M opinion words {Vm }M
m=1 and N aspect
(A) N
words {Vn }n=1 .
With the opinion and aspect words extracted via the above
procedure, we can obtain the POS tag features from their
context to train the MaxEnt model, of which we will not go
into the details due to the page limit.

p(y(d,s,n) = 0|z, r, y(d,s,n) , w, x)

A,r

and for y(d,s,n) = 1:

p(y(d,s,n) = 1|z, r, y(d,s,n) , w, x)
O,r

C. Approximate Posterior Inference

It is obvious that exact posterior inference is intractable
in CAMEL, so we turn to a collapsed Gibbs sampling algorithm [6] for approximate posterior inference, which is simple
to derive, comparable in speed to other estimators, and can
approximate a global maximum. Due to the space limit, we
leave out the derivation details and only present the sampling
formulas. Note that the MaxEnt component is trained before
we perform Gibbs sampling, which means {0 , 1 } are xed
during Gibbs sampling.
In CAMEL, we have three sets of latent variables: z, r and
y. Given the assignments of all other hidden variables, we can
jointly sample (zd,s , rd,s ) as a block:

d
+
C(j)

,z

d,s d,s
C(wd,s,n
+
exp(1 xd,s,n )
)

.
1
O,r
,z
d,s
d,s
l =0 exp(l xd,s,n ) C
+V
()

Here counts represented by symbol C indicate that word n in

sentence s of document d has been excluded.
D. Discussions
We here briey discuss the similarities and differences between CAMEL and two most related models, i.e., ccTAM [5]
and CPT [4], from the perspective of model structure. ccTAM
distinguishes common topics from specic ones as CAMEL
did, but it does not separate opinions from topics. CPT
explicitly separates opinions from topics. However, it needs
strict rules to separate opinion and topic words, whereas we
provide a more soft way for opinion-aspect switching over a
word. Besides, CPT dose not distinguish common and specic
topics, which limits its applicability on asymmetric collections.

P (zd,s = k, rd,s = j|z(d,s) , r(d,s) , y, w, x)

,z

d,s d,s
C(wd,s,n
+
exp(0 xd,s,n )
)

,
1
A,r
,z
d,s
d,s
l =0 exp(l xd,s,n ) C
+V
()

d,j
+
C(k)

d + 2
d + Kj
C()
C(j)





A,j,k
V C A,j,k +N A,j,k +
C()
+V

(v)
(v)





A,j,k
A,j,k
A,j,k
C()
+N()
+V v=1
C(v)
+





O,j,k
V C O,j,k +N O,j,k +
+V
C()

(v)
(v)
.




O,j,k
O,j,k
O,j,k
C() +N() +V v=1
C(v) +

IV. E XPERIMENTAL R ESULTS

In this section, we present extensive experimental results
to evaluate CAMEL in both quantitative and qualitative ways.
Hereinafter, we agree to use CAMEL, our method and
ours interchangeably in comparative studies.
A. Experimental Setup
1) Data Sets: Our method is tested on three real world
data sets. One data set is a collection of electronic device
reviews from Amazon1 , which is used to perform quantitative
evaluations of our method. Two real world events data has
been crawled and analysed for the qualitative analysis of our
method.
The online reviews is collected by Jo et al. [12], which
contains electronic device reviews with seven categories. To

We rst consider the case of j = 0. With this condition,

d
is the number of sentences assigned to common aspects
C(j)
d,j
is the number of sentences assigned to
in document d. C(k)
common aspect k in document d. K j is the number of common
A,j,k
is the number of times
aspects, i.e., K I equivalently. C(v)
word v is assigned as an aspect word to common aspect k,
O,j,k
and C(v)
is the number of times word v is assigned as an

1 http://www.amazon.com

672

TABLE II: Statistical description of data sets.

Data
Collection0
Collection1
Collection0&1
APEC News
APEC Tweets
Stampede News
Stampede Tweets

#Documents
3,535
3,659
7,194
9,662
82,366
1,015
13,004

#Sentences
56,053
60,935
116,988
251,023
144,149
42,651
25,425

opinion mining by explicitly separating common and specic

aspects.
3) Evaluation Measures: We briey introduce two sets of
measures used in experiments, one set is macro-averaged precision, recall and f-measure, the other set is opinion coherence
and aspect-opinion coherence.
Since each online review has an category, we can leverage
this supervised information to evaluate aspects learned by our
method as well as baseline methods. As all methods assign
one aspect to each sentence, therefore, we label a sentence
with the category of the review it resides in. By manually
mapping learned aspects to their corresponding category, we
can take a reviews category as ground truth to evaluate them
by performing sentence classication.
To compare the quality of opinions, we choose an automatic
measure called Topic Coherence [17], which has been widely
used in evaluating topics and has been justied in according
with human evaluations. We also slightly modify the coherence
score to measure relevance of aspect and its opinion. Details
please refer to Sect.IV-C1

#Words
315,471
339,192
654,663
3,561,412
1,136,422
262,613
74,353

evaluate the quality of common and specic aspects as well

as their opinions across collections, we create a new data set
based on the reviews. We select reviews under three categories,
namely coffee machine, canister vacuum and MP3 player, to
build a data set with two collections. Intuitively, the reason for
selecting these three categories is that the reviews under these
categories have minimal overlap in contents, which can help to
ensure aspects of common category can be well distinguished
from those of specic category.
To build the new data set, we rst place reviews labeled
as canister vacuum into collection C0 and reviews labeled as
MP3 player into collection C1 , then randomly inject sentences
of coffee machine reviews into reviews in C0 or C1 . This way
of making the new data set is according with characteristics of
real world asymmetric collections, since common aspects are
shared by collections and each collection has its own specic
aspects. Here sentences from coffee machine is regarded as
common part across collections, and canister vacuum or MP3
Player is regarded as specic part only occurs in C0 or C1
respectively. As to real word events, we crawled news and
tweets related to event 2014 Shanghai stampede and 2014
Beijing APEC.
All data sets go through the same preprocessing process:
rst applying POS tagging and then get automatic labeled
training data for Maximum Entropy(MaxEnt) model as illustrated in Sect.III-B. At last, we remove stop words and those
with low document frequency. For tweets, we also remove
URLs and Hashtags. We use Stanford POS Tagger2 to tag
English online reviews and LTP-Cloud3 to tag Chinese news
and tweets. The opinion lexicon used in auto-labeled MaxEnt
model for English corpus is collected by Hu and Liu [9]. As
to Chinese corpus we use an opinion lexicon merged from two
widely used Chinese opinion lexicons. Details of preprocessed
data sets are shown in Table II
2) Baseline Methods: In the quantitative experiments, we
compare our method with three baseline methods. All baseline
methods are MaxEnt-LDA [27] run over different collection
congurations.
1) BL-0: Perform MaxEnt-LDA over collection C0
2) BL-1: Perform MaxEnt-LDA over collection C1
3) BL-2: Perform MaxEnt-LDA over collection C0 and C1 .
Since we adopt the MaxEnt-LDA in our model to obtain
aspect-specic opinions, comparing with these baselines can
give us insights of whether aspects and opinions induced
by our method can benet from complementary aspect-based

B. Aspect Evaluation
We design a sentence classication experiment on reviews
to evaluate the quality of aspects learned by our method.
Note that all baseline methods as well as our method assigns
one aspect to each sentence, and each of induced aspect
often corresponds to one category, thus we can use sentence
classication as the evaluation method for learned aspects.
Specically, better sentence classication results indicate better quality of aspects. In order to illustrate asymmetric aspectbased opinion mining, we created a dataset with common
aspects across collections and specic ones only reside in each
collection as described in Sect.IV-A1.
Each review in our data set has one category in the set
L = {coffee machine, canister vacuum, MP3 player}. We use
the category of the review to label its sentences. Sentences of
coffee machine reviews are injected into reviews in collection
C0 or collection C1 randomly. Therefore, coffee machine sentences exist across collections. Sentences of canister vacuum
or MP3 player are only contained in C0 or C1 . In other words,
we expect aspects about coffee machine as common aspects,
those related to remain two categories as specic aspects.
We run BL-0, BL-1, BL-2 and our method over the data to
get inferred aspect set S A and aspect assignment k of each
sentence. In order to perform sentence classication evaluation, we have to manually map each aspect to one of the three
categories, i.e., we created a mapping functionf (k) : S A L.
Aspects can not be mapped to any category are labeled as
other. Given the mapping, we can get predicted labels of
sentences for each method. Then we evaluate all methods
according to metrics, namely precision, recall and f-measure.
For all methods, we set = 0.1, = 0.01. For our method,
we set = 0.1. In order to keep all methods comparable,
we have to set proper number of aspects for each method.
Specically, we set aspect number K2 of BL-2 equals to the
sum of K0 and K1 , where K0 is the aspect number for BL-0

2 http://nlp.stanford.edu/software/tagger.shtml
3 http://www.ltp-cloud.com/

673

TABLE III: Sentence classication results.

Method
BL-2*
Ours*
BL-0*
BL-2-C0 *
Ours-C0 *
BL-1*
BL-2-C1 *
Ours-C1 *
BL-2
Ours
BL-0
BL-2-C0 
Ours-C0 
BL-1
BL-2-C1 
Ours-C1 
BL-2
Ours
BL-0
BL-2-C0
Ours-C0
BL-1
BL-2-C1
Ours-C1

precision
0.819
0.791
0.866
0.804
0.791
0.891
0.836
0.791
0.825
0.909
0.826
0.889
0.896
0.875
0.935
0.921
0.823
0.869
0.846
0.847
0.843
0.883
0.886
0.856

K=5
recall
0.783
0.833
0.646
0.780
0.819
0.727
0.787
0.847
0.836
0.876
0.943
0.841
0.871
0.954
0.832
0.881
0.819
0.862
0.794
0.811
0.845
0.841
0.809
0.864

f-measure
0.799
0.820
0.734
0.789
0.801
0.800
0.809
0.813
0.827
0.891
0.880
0.862
0.882
0.912
0.880
0.899
0.818
0.863
0.807
0.826
0.842
0.856
0.844
0.856

precision
0.745
0.861
0.856
0.695
0.856
0.858
0.826
0.865
0.797
0.919
0.861
0.882
0.904
0.894
0.935
0.934
0.780
0.869
0.859
0.789
0.880
0.876
0.881
0.899

K = 10
recall
f-measure
0.791
0.765
0.783
0.820
0.736
0.789
0.794
0.733
0.768
0.810
0.780
0.815
0.787
0.806
0.798
0.829
0.761
0.767
0.861
0.889
0.927
0.892
0.689
0.753
0.865
0.884
0.933
0.913
0.833
0.880
0.857
0.894
0.771
0.766
0.860
0.863
0.832
0.841
0.742
0.743
0.817
0.847
0.856
0.864
0.810
0.843
0.828
0.862

and K1 is the aspect number for BL-1. Setting common aspect

number K I and specic aspect number K S in our method
needs to satisfy K I + CK S = K2 , where C is the number
of collections and in this case C = 2. In this experiment,
we keeps K = K0 = K1 , and range K from 5 to 20,
therefore K2 ranges from 10 to 40. We set K I = 2, 4, 6, 8
and K S = 4, 8, 12, 16 for K = 5, 10, 15, 20 respectively.
Here we set K S = 2K I , because we know the approximate
proportion of common aspects in each collection. However,
when dealing with data has no supervised information(or prior
knowledge) as we did in our case study, one might has to try
several congurations of K I and K S to nd one appropriate
setting, or resorts to nonparametric Bayesian inference. For all
methods, we run 1000 iterations of Gibbs sampling, and each
parameter conguration runs 10 samples.
The sentence classication results are reported in Table
III. All results are the average of 10 samples. The marker
indicates metrics are computed for the common category
classication, and the marker  indicates metrics are computed
for specic categories classication, and the results without
marker are macro average of all categories, which reect the
overall performance of each method. The sufx C0 or C1
appends to BL-2 and our method indicates the measures are
computed on the sentences in C0 or C1 respectively. With out
sufx indicates the measures are computed on all sentences in
C0 and C1 .
From the comparison of BL-2 with our method on average
classication results, as well as the common and specic categories classication results, we nd our method consistently
outperforms BL-2. Besides, our method performs quite stable
with different K, however, the performance of BL-2 changes
notably. Since BL-2 performs aspect-opinion mining direct
on combined collections, it fails in leveraging the structure
of common and specic aspects underneath the asymmetric

precision
0.795
0.829
0.864
0.771
0.820
0.879
0.827
0.838
0.810
0.923
0.857
0.886
0.910
0.895
0.940
0.934
0.805
0.868
0.860
0.828
0.865
0.887
0.883
0.886

K = 15
recall
f-measure
0.777
0.785
0.812
0.820
0.725
0.787
0.776
0.770
0.801
0.810
0.780
0.826
0.779
0.802
0.822
0.830
0.814
0.809
0.856
0.888
0.934
0.893
0.798
0.836
0.852
0.880
0.946
0.920
0.830
0.880
0.859
0.895
0.802
0.801
0.859
0.862
0.829
0.840
0.787
0.803
0.827
0.845
0.863
0.873
0.804
0.841
0.841
0.862

precision
0.833
0.819
0.846
0.824
0.816
0.877
0.843
0.821
0.810
0.922
0.855
0.892
0.907
0.890
0.930
0.937
0.818
0.869
0.851
0.858
0.843
0.884
0.886
0.879

K = 20
recall
f-measure
0.747
0.787
0.820
0.818
0.726
0.775
0.743
0.781
0.801
0.808
0.767
0.818
0.751
0.793
0.833
0.827
0.852
0.829
0.858
0.889
0.913
0.880
0.846
0.868
0.860
0.883
0.945
0.917
0.857
0.891
0.856
0.894
0.817
0.815
0.860
0.863
0.819
0.828
0.795
0.824
0.831
0.846
0.856
0.867
0.804
0.842
0.844
0.860

collections that we focused in this paper. Our method explicitly

separates aspects shared by collections and those specic to
each collection, therefore, aspects in one collection serve as
complementary information for aspects extraction in other
collections in an mutual way.
The results of common category classications conrm our
above discussions. From the group of results marked with
, we can see our method still consistently outperforms all
baseline methods on recall and f-measure in almost all cases.
This promising result indicates that complementary aspect
mining can great benet the accuracy of learned common
aspects. Although precision results of our method are better
than most of those of BL-2, we also notice that they are
slightly worse than those of BL-1 in most cases. The results
of specic category classications, i.e., results marked as ,
are in contrary to those in common category classications,
as our method gains relative high precision but low recall as
compared to BL-0 or BL-1.
In order to illustrate this phenomenon, we check the mapping of aspect to category for our method, and nd there is
always an common aspect be mapped to category other, when
K > 5. Then we look into those aspects and nd they are
all related to after-sales service. We give one sample of aftersales service aspect-opinions in TableIV, where opinion0(or
opinion1) means opinions induced from C0 (or C1 ). Note that
after-sale service is independent to review categories and obviously an common aspect across review collections. There is
no label named after-sale service, therefore, sentences related
to after-sale service but originally labeled as other categories
are misclassied by our method. Since BL-0 and BL-1 mixed
after-sales service aspect with other aspects, therefore, no aspect is mapped to other. We guess this might be the reason why
the precisions of our model in common category classications
as well as the recalls in specic category classications are

674

TABLE IV: After-sales service aspect and opinions.

aspect
opinion0
opinion1

TABLE V: Results of opinion and aspect-opinion coherence.

service call customer back return product problem

amazon buy warranty
send free good great ship local wrong long happy fast
send work great free good local defective creative
original easy

Method
BL-0
Ours-C0
BL-1
Ours-C1
BL-2
Ours
Ours
Ours-C0
Ours-C1

less comparative with BL-0s or BL-1s. However, except for

the common aspect we called after-sale service, other aspects
can be clearly mapped to an category, unless it is not ratable.
Therefore, we can conclude from the classication results that
our method can learn more accuracy aspects by performing
complementary aspect-based opinion mining.
In this subsection, we give evaluations of opinions learned
by our method. In most previous works [3], [27], the evaluation
of opinions is highly relied on the judgment of humans, which
is costly and some times bias from the data. Therefore, we
try to use some automatic measures to evaluate opinions.
Recently, an automatic measure named Topic Coherence [17]
has been proposed to evaluate the coherence of distributions
learned by topic models, which has been justied in according
with human evaluations. Therefore we use coherence score to
evaluate opinions, and further propose a new measure based
on it to evaluate the coherence(or relevance) between aspect
and its corresponding opinion.
1) Opinion Coherence: Coherence score measures a single
word distribution by computing the semantic similarity degree
between high probability words in it. Higher score often
indicates better quality. Given T high probability words of
an opinion, the coherence score for the opinion is dened as
t1
T

(k)

log

(k)
D(vl )

t=2 l=1
(k)

(k)

D(vt , vl ) + 

(k)

where V (k) = (v1 , . . . , vT ) is a list of T most probable

words in opinion k. D(v) counts the number of documents
containing the word v, and D(v, v  ) counts the number of
documents containing both v and v  .  is a smoothing variable
used to avoid taking the log of zero for words that never cooccur.
2) Aspect-Opinion Coherence: Coherence score introduced
above can only evaluate the quality of one opinion. It can not
evaluate the coherence between aspect and its opinion. Since
our method learns pairs of aspect and opinion, we propose
a new measure for the evaluation of aspect opinion pairs.
Given T high probability words of an aspect and its opinion
respectively, the coherence score for one pair of aspect and
opinion is dened as
CA,O (k; V A,(k) , V O,(k) ) =

T
T

t=1 l=1

A,(k)

log

D(vt

O,(k)

, vl

A,(k)
D(vt
)

)+

aspect-opinion coherence
T = 10
T = 15
-229.35.3
-572.17.1
-226.26.2
-571.315.7
-246.75.5
-621.111.6
-245.13.2
-613.04.8
-255.23.5
-640.712.1
-240.63.4
-604.06.4
-192.33.0
-483.54.8
-246.66.5
-598.820.1
-245.96.0
-612.110.2

or she already have observed the aspect word v A,(t) in an

document. For both coherence scores, we set  = 112 to
reduce the score for completely unrelated words as suggested
in [22].
We compare our method with baseline methods based on
average opinion coherence score and average aspect-opinion
coherence score of all induced aspect-based opinions. For all
methods, we set = 0.1, = 0.01. For our method, we
set = 0.1. To make all methods comparable, we have to
set proper number of opinions for each method. Note that for
baseline methods, setting number of opinions is the same as
setting number of aspects, details in Sect.IV-B. As our method
learns C opinions for each common aspect, where C is the
number of collections, We have to set C(K I + K S ) equals to
K2 . In this experiment, we set K = K0 = K1 = 15, K2 = 30,
and K I = 5, K S = 10.
The average opinion coherence and aspect-opinion coherence results with T = 10 and T = 15 are illustrated
in Table V. The sufx C0 or C1 appends to our method
indicates we evaluate coherence of common aspect-opinions
and specic ones on C0 or C1 . The marker indicates the
average coherence score for common aspect-opinions only.
From the results, we can see our method outperforms BL2 on both opinion coherence and aspect-opinion coherence
with different T . This result indicates taking complementary
aspect-opinions shared across collections into account can not
only help to nd more accuracy aspects but also improve
the opinion coherence and enhance the relevance of aspect
and opinion. We can also nd the overall performance of our
method surpass BL-0 and BL-1. Since BL-0 and BL-1 run
over single collection, which limits its ability in separating
common aspect-opinions and specic ones, and therefore
hurts its opinion coherence and aspect-opinion relevance. We
also compare the common aspect-opinions of our method on
collection C0 , C1 and both. The result is promising, as it
illustrates those common aspect-opinions are more coherence
on combined collections than on single one, which further
demonstrates the complementary aspect-based opinion mining
is necessary for asymmetric collections.

C. Opinion Evaluation

CO (k; V (k) ) =

opinion coherence
T = 10
T = 15
-121.13.6
-307.67.2
-119.44.3
-307.79.0
-129.83.7
-334.911.9
-127.71.7
-326.75.4
-134.82.7
-345.67.9
-129.82.0
-329.74.1
-103.72.6
-260.93.9
-127.54.6
-309.66.5
-126.04.5
-320.29.6

D. Complementary Aspect Mining Evaluation

We compare performance of baseline methods and ours,
given situations that common aspects across collections are
highly imbalanced. This experiment can give us insights about
how collections lack of common aspects can benet from
complementary informations in other collections. In order to

where V A,(k) is a list of T most probable words in aspect k,

and V O,(k) is a list of T most probable words in opinion k.
A,(k) O,(k)
,v
)
Loosely speaking, the value of D(vD(vA,(k)
estimates the
)
probability one could observe the opinion word v O,(t) if he

675

(a) C0 : recall.

(b) C0 : f-measure.

(c) C0 and C1 : recall.

(d) C0 and C1 : f-measure.

Fig. 2: Common category classication results with different imbalance ratios on C0 or C0 combined with C1 .
simulate the imbalance circumstances, we reform the online
reviews data used in Sect.IV-B by removing sentences labeled
as coffee machine(the common category) from collection C0 ,
while collection C1 remains the same.
To creat different levels of imbalance, we remove coffee
machine sentences according to ratios from 10% to 80%,
which is the proportion of coffee machine sentences we
remove from collection C0 . With reformed data set, we can
perform coffee machine sentences classication on C0 and
C0 combined with C1 to discuss the complementary aspect
mining. We set = 0.1, = 0.01 for all methods, = 0.1
for our method. K = K0 = K1 = 15, K2 = 30, K I = 6 and
K S = 12. The resulting recall and f-measure are illustrated
in Fig. 2.
In Fig.2a and Fig. 2b we can see without complementary
information, the recall and f-measure of BL-0 both drop
rapidly and become extremely unstable when ratio reaches
80%. Since BL-2 and our method both utilize complementary
common aspects in C1 , therefore is relative stable than BL-0.
As we could see in Fig 2, the performance of BL-2 on C0 as
well as on C0 and C1 , also become rather poor and unstable
as ratio reaches 80%. This promising result strongly suggested
the explicit separation of common aspects and specic aspects
is necessary for complementary aspect mining.

TABLE VI: Average P @n of aspect-specic opinion words

with different sizes of training data from MME and AME.
Size
S
S
S
S
S

= 10
= 20
= 30
= 40
= 50

P @5
MME
AME
0.90
0.64
0.80
0.64
0.86
0.84
0.80
0.84
0.82
0.90

P @10
MME
AME
0.82
0.58
0.74
0.56
0.82
0.81
0.75
0.85
0.81
0.83

P @20
MME
AME
0.71
0.51
0.67
0.50
0.70
0.76
0.71
0.78
0.71
0.76

TABLE VII: Complementarity between news & tweets.

Tweets:

News:

!
Thanks to you and those who also shouted fallback, you
are good man!
!
How calm and wise the fallback shouters are, great!

When the event took place, there are hundreds of people

shouting fallback! fallback!

After the event, this group of people are called

fallback shouters by Internet users.

we can set a larger training size without the concern of costly

in time and other resources. Thus, we conclude our AME is
an effective method to avoid manually label training data for
the MaxEnt component in our model.
F. Case Study

E. Validity of Auto-labeled MaxEnt

In this subsection, we apply our method to news and tweets

from two real world events, namely 2014 Shanghai Stampede
and 2014 Beijing APEC, to illustrate its advantages in the
complementary aspect-opinion mining. Hereinafter, we refer
to 2014 Shanghai Stampede as Stampede and refer to 2014
Beijing APEC as APEC. We show some discovered common
aspect-opinions to demonstrate how concrete aspects in news
can help tweets to nd more clear aspects and how rich
opinions resides in tweets supplement to aspects of news. At
last, we give brief discussion about what aspects are those
concerned only in news or tweets.
To illustrate how concrete news aspects supplement to
unclear tweets aspects, as well as how tweets enrich opinions
of news aspects, we list several tweets and sentences in news,
as shown in Table VII.
As we can see, the aspect of above tweets is less clear,
since one may not know who yelled fallback, when and why,
however opinions are very intense. The sentences of news
are just the opposite. Obviously, those tweets and sentences
of news share the same aspect, and their are assigned to an

To justify the validity of our auto-labeled MaxEnt component, we give comparisons of manual labeled MaxEnt(MME
for short) with auto-labeled MaxEnt(AME for short). We compare the P recision@n(P @n for short) of MME with AME
on reviews data with varying number of training sentences
S. Here P @n means how many words are precisely opinion
words other than aspect words given the top n probability
words of an opinion, which is judged by human. For MME,
we select 50 sentences with opinions words and manually label
them. As to AME, we use our automatic procedure to acquire
the same number of training sentences. We increase the size
of training data inputted into the MaxEnt, and compare the
P @5, P @10 and P @20 of MaxEnt-LDA [27] with different
sources of training data. The results are reported in Table VI.
From the results, we can nd that AME is inaccuracy when
training size is small. However, when training size is larger
than 30, AME is even more accuracy than MME. Note that the
training data of AME is obtained fully automatically, therefore,

676

V. R ELATED W ORK

TABLE VIII: Sample specic aspect and opinion from tweets.

Aspect

Opinion

(competence) (improvement) (citizen)

(needs) (countrymen) (safety) (attainment)
(the public) (order) (nation) (awareness)
(improve) (investigate) (objective)
(need) (civilized) (comfort)
(safe) (educate) (accidental) (unordered)

A. Aspect-based Opinion Mining

Two subtasks are usually involved in this problem, namely,
aspect or feature identication and opinion extraction. Most
of the early works on aspect identication are feature-based
approaches [9], [18], e.g., applied frequent itemset mining
to identify product aspects [15], which normally exert some
constraints on high-frequency noun phrases to nd aspects. As
a result, they usually subject to the risk of producing too many
non-aspects and missing low-frequency aspects [7]. Several
early works have applied supervised learning to identify both
aspects and opinions [10], [11], [25], which, however, needs
hand-labeled training sentences and thus is very costly.
In recent years, with the popularity of topic models, more
unsupervised methods are proposed for aspect-based opinion
mining. For instance, Titov and McDonald propose a multigrain topic model to learn both global topics and local topics,
in which local topics correspond to rateable aspects [23].
Another approach to discover aspects is to t a topic model
to sentences instead of documents. For instance, Brody and
Elhadad run the latent dirichlet allocation(LDA) [2] model
over sentences instead of documents to extract aspects [3].
Zhao et al. [27] and Jo et al. [12] assume that all words in a
single sentence are generated from one topic.
Some researchers take approaches that model topic and
sentiment in a unied way. For instance, Lin and He propose
a joint topic-sentiment(JTM) model to detects sentiment and
topic simultaneously from text [14]. The aspect and sentiment
unication model(ASUM) proposed by Jo et al. [12] is similar
to JST, the major difference lays in that ASUM assumes each
single sentence only covers one topic. The above two models
do not explicitly separate topic words and sentiment words.
Mei el al. [16] propose a topic sentiment mixture model, which
represents positive and negative sentiments as language models
separate from topics, but both sentiment language models
capture general opinion words only. Brody et al. [3] take a twostep approach by rst detecting aspects and then identifying
aspect-specic opinion words. Zhao et al. [27] propose a topic
model integrating with a maximum entropy model(MaxEntLDA) to jointly capture both aspects and aspect-specic opinion words within a topic model. [20]gives detailed discussions
about aspect-specic opinion models based on LDA.

TABLE IX: Sample specic aspect and opinion from news.

Aspect

Opinion

(event) (lesson) (Shanghai)

(new year) (accident) (reason)
(government) (report) (survivor) (stairs)
(investigate) (objective) (dispose)
(announce) (strictly investigate) (respect)
(true) (comfort) (care) (serious)

common aspect by our method. We show this common aspect

and its opinions as well as several other samples induced by
our method from Stampede and APEC in Table.X.
The common aspect 0 and 1 are extracted from Stampede,
the rest two are extracted from APEC, and opinion 0 means
the opinion words are extracted from tweets, otherwise from
news. Common aspect 1 is about the fallback shouters that
refers to a bunch of people who yelled fallback to alert the
crowd the break out of stampede, which saved lots of lives.
From the opinion words in tweets, we can nd the public
is mainly touched by fallback shouters and appreciates their
behavior. While there is no obvious opinions from news.
Another case of Stampede is common aspect 2, which is about
the announcement of namelist of victims in the event. Most
people feels sympathize with those who dead young, while
some people criticizes those young people died because of
their ignorance. Common aspect 3 is about event APEC, and
it is about the reform of China lead by Chairman Xi. Opinions
from tweets reect peoples condence and expectation on
this government. Common aspect 4 is about the free-trade
agreement in APEC, which also raises the concern of the
public, since they can benet from this agreement.
From all those common aspect-opinions, we can see the
opinions extracted from tweets are obviously more emotional,
while those extracted from news is monotonous. The common
aspects can be easily recognized, which is less possible if we
directly extract them from tweets. Besides those showed in
this paper, there are also some other common aspects, such as
air pollution, visa-free and APEC holiday etc. in APEC, and
rescue, penalty etc. in Stampede. All of them are concerned
by public and reported frequently by news media. Due to the
limit of pages, we are not going to list them all.
With the induced specic aspect-based opinions from news
and tweets, we can nd aspects concerned by the public while
less reported by news, as well as those widely covered by
news while less focused by the public. We give two examples,
one for tweets and one for news. In the Table.VIII, we can
see the public talks a lot about the populaces cultivation,
and its opinion words show the quality of the nation needs
to be improved and education should be enhanced. While in
Table.IX, we can nd news reports about the government will
draw lessons from the event, however, in tweets people only
cares about the punishment of person in charge.

B. Cross Collection Text Mining

Zhai et al. [26] introduce a task called comparative
text mining and proposed a cross-collection mixture(ccMix)
model based on probabilistic latent semantic index(pLSA) [8].
The goal of the task is to discover the common themes
across all collections and the ones unique to each collection.
Paul et al. [21] extend ccMix to the ccLDA model based on
LDA for cross-culture topic analysis. Gao et al. [5] propose
a cross collection topic aspect model(ccTAM) to perform
event summarization across news and social media steams.
They assume aspects contained only in tweets can server as
a supplementation to those in news. Fang et al. [4] propose

677

TABLE X: A sample of common aspect-opinions induced from APEC and Stampede data.
Common Aspect 1


crowd scene bund policeman
fallback shouters new year event
tourist order voice
opinion 0
opinion 1
(recede)
(above)
(great)
(tumble)
(touched)
(suspected)
(calm)
(at all)
(hope)
(down)
(tender)
(around)
(reduce)
(crowded)
(important)
(stable)
(dangerous)
(noisy)
(thank)
(nearby)

Common Aspect 2


victims namelist event bund
Shanghai identity stampede
square forenoon age
opinion 0
opinion 1
(grieve)
(announce)
(sympathize)
(verify)
(blessing)
(injure)
(regret)
(check)
(perfunctory)
(crowed)
(ignorant)
(happen)
(distressed)
(sympathise)
(appease)
(grieve)
(simple)
(heartfelt)
(fortunately)
(mourn)

a cross-perspective topic model(CPT) to perform contrastive

opinion modeling. They view opinions with the same topic
yet from different news sources as different perspectives, and
learns topics(not aspects) across collections by perform LDA
over aggregated collections, which has no guarantee of nding
shared topics, especially when collections is less comparative
as news versus tweets in our case.

Common Aspect 3


China society diplomacy
reformation world economy advance
international Jinping Xi legislation
opinion 0
opinion 1
(cheer)
(reform)
(progress)
(develop)
(thorough)
(overall)
(great)
(deepen)
(overall)
(innovate)
(happy)
(peaceful)
(civil)
(signicant)
(braw)
(great)
(strong)
(revival)
(fair)
(fair)

Common Aspect 4
APEC

China tariff APEC products
agreement free-trade price
commodity enterprise Australia
opinion 0
opinion 1
(import)
(invest)
(cheap)
(reduce)
(expect)
(cancel)
(bring)
(open)
(develop)
(signicant)
(convenient)
(investigate)
(recover)
(serve)
(share)
(export)
(free)
(promote)
(inuence)
(improve)

[4] Yi Fang, Luo Si, Naveen Somasundaram, and Zhengtao Yu. Mining
contrastive opinions on political texts using cross-perspective topic
model. In WSDM 12, pages 6372, 2012.
[5] Wei Gao, Peng Li, and Kareem Darwish. Joint topic modeling for event
summarization across news and social media streams. In CIKM 12.
[6] T. L. Grifths and M. Steyvers. Finding scientic topics. PNAS04.
[7] Honglei Guo, Huijia Zhu, Zhili Guo, XiaoXun Zhang, and Zhong
Su. Product feature categorization with multilevel latent semantic
association. In CIKM 09, pages 10871096, 2009.
[8] Thomas Hofmann. Probabilistic latent semantic indexing. In SIGIR 99.
[9] Minqing Hu and Bing Liu. Mining and summarizing customer reviews.
In KDD 04, pages 168177, 2004.
[10] Wei Jin and Hung Hay Ho. A novel lexicalized hmm-based learning
framework for web opinion mining. In ICML 09, pages 465472, 2009.
[11] Wei Jin, Hung Hay Ho, and Rohini K. Srihari. Opinionminer: A novel
machine learning system for web opinion mining and extraction. In
KDD 09, pages 11951204, 2009.
[12] Yohan Jo and Alice H. Oh. Aspect and sentiment unication model for
online review analysis. In WSDM 11, pages 815824, 2011.
[13] Kar Wai Lim and Wray Buntine. Twitter opinion topic model: Extracting
product opinions from tweets by leveraging hashtags and sentiment
lexicon. In CIKM 14, pages 13191328. ACM, 2014.
[14] Chenghua Lin and Yulan He. Joint sentiment/topic model for sentiment
analysis. In CIKM 09, pages 375384, 2009.
[15] Bing Liu, Minqing Hu, and Junsheng Cheng. Opinion observer:
Analyzing and comparing opinions on the web. In WWW 05.
[16] Qiaozhu Mei, Xu Ling, Matthew Wondra, Hang Su, and ChengXiang
Zhai. Topic sentiment mixture: Modeling facets and opinions in weblogs.
In WWW 07, pages 171180, 2007.
[17] David Mimno, Hanna M. Wallach, Edmund Talley, Miriam Leenders,
and Andrew McCallum. Optimizing semantic coherence in topic models.
In EMNLP 11, pages 262272, 2011.
[18] Samaneh Moghaddam and Martin Ester. Opinion digger: An unsupervised opinion miner from unstructured product reviews. In CIKM 10.
[19] Samaneh Moghaddam and Martin Ester. Aspect-based opinion mining
from product reviews. In SIGIR 12, pages 11841184, 2012.
[20] Samaneh Moghaddam and Martin Ester. On the design of lda models
for aspect-based opinion mining. In CIKM 12, pages 803812, 2012.
[21] Michael Paul and Roxana Girju. Cross-cultural analysis of blogs and
forums with mixed-collection topic models. In EMNLP 09.
[22] Keith Stevens, Philip Kegelmeyer, David Andrzejewski, and David
Buttler. Exploring topic coherence over many models and many topics.
In EMNLP-CoNLL 12, pages 952961, 2012.
[23] Ivan Titov and Ryan McDonald. Modeling online reviews with multigrain topic models. In WWW 08, pages 111120, 2008.
[24] Yao Wu and Martin Ester. Flame: A probabilistic model combining
aspect based opinion mining and collaborative ltering. In WSDM 15.
[25] Yuanbin Wu, Qi Zhang, Xuanjing Huang, and Lide Wu. Phrase
dependency parsing for opinion mining. In EMNLP 09.
[26] ChengXiang Zhai, Atulya Velivelli, and Bei Yu. A cross-collection
mixture model for comparative text mining. In KDD 04.
[27] Wayne Xin Zhao, Jing Jiang, Hongfei Yan, and Xiaoming Li. Jointly
modeling aspects and opinions with a maxent-lda hybrid. In EMNLP10.

VI. C ONCLUSIONS
In this paper, we propose CAMEL, a novel topic model for
complementary aspect-based opinion mining across asymmetric collections. By modeling both common and specic aspects
and keeping contrastive opinions, CAMEL is capable of integrating complementary information from different collections
in both aspect and opinion levels. An automatically labeling
scheme is also introduced to further boost the applicability
of CAMEL. Extensive experiments and real-world case study
on public events demonstrate the effectiveness of CAMEL
in leveraging complementarity for high-quality aspect and
opinion mining.
ACKNOWLEDGMENT
Deqing Wang was supported by National Natural Science
Foundation of China (NSFC) (71501003) and China Postdoctoral Science Foundation funded project (2014M550591). Junjie Wu was supported by National High Technology Research
and Development Program of China (SS2014AA012303),
NSFC (71322104,71171007,71531001,71471009), National
Center for International Joint Research on E-Business Information Processing (2013B01035), Foundation for the Author
of National Excellent Doctoral Dissertation of China (201189),
and Fundamental Research Funds for Central Universities.
R EFERENCES
[1] Yang Bao, Nigel Collier, and Anindya Datta. A partially supervised
cross-collection topic model for cross-domain text classication. In
CIKM 13, pages 239248. ACM, 2013.
[2] David M. Blei, Andrew Y. Ng, and Michael I. Jordan. Latent dirichlet
allocation. J. Mach. Learn. Res., 3:9931022, 2003.
[3] Samuel Brody and Noemie Elhadad. An unsupervised aspect-sentiment
model for online reviews. In HLT 10, pages 804812, 2010.

678