Product Feature Taxonomy Learning based on User Reviews
Nan Tian
1
, Yue Xu
1
, Yuefeng Li
1
, Ahmad Abdel-Hafez
1
and Audun Josang
2
1
Faculty of Science and Engineering, Queensland University of Technology, Brisbane, Australia
2
Department of Informatics, University of Oslo, Oslo, Norway
Keywords:
Feature Extraction, Opinion Mining, Association Rules, Feature Taxonomy, User Reviews.
Abstract:
In recent years, the Web 2.0 has provided considerable facilities for people to create, share and exchange in-
formation and ideas. Upon this, the user generated content, such as reviews, has exploded. Such data provide
a rich source to exploit in order to identify the information associated with specific reviewed items. Opin-
ion mining has been widely used to identify the significant features of items (e.g., cameras) based upon user
reviews. Feature extraction is the most critical step to identify useful information from texts. Most existing
approaches only find individual features about a product without revealing the structural relationships between
the features which usually exist. In this paper, we propose an approach to extract features and feature rela-
tionships, represented as a tree structure called feature taxonomy, based on frequent patterns and associations
between patterns derived from user reviews. The generated feature taxonomy profiles the product at multi-
ple levels and provides more detailed information about the product. Our experiment results based on some
popularly used review datasets show that our proposed approach is able to capture the product features and
relations effectively.
1 INTRODUCTION
In recent years, the user generated online content ex-
ploded due to the advent of Web 2.0. For instance,
online users write reviews to how they enjoy or dis-
like a product they purchased. This helps to identify
features or characteristics of the product from users’
point of view, which is an important addition to the
product specification. However, to identify the rele-
vant features from users’ subjective review data is ex-
tremely challenging.
Feature-based opinion mining has attracted big at-
tention recently. A significant amount of research
has been proposed to improve the accuracy of feature
generation for products (Hu and Liu, 2004a; Scaffidi
et al., 2007; Hu et al., 2010; Zhang and Zhu, 2013;
Popescu and Etzioni, 2005; Ding et al., 2008). How-
ever, most techniques only extract features; the struc-
tural relationship between product features has been
omitted. For example, “picture resolution” is a com-
mon feature of digital camera in which “resolution”
expresses the specific feature concept to describe the
general feature “picture”. Yet, existing approaches
treat “resolution” and “picture” as two individual
features instead of finding the relationship between
them. Thus, the information derived by existing fea-
ture extraction approaches is not sufficient for gen-
erating a precise product model since all features are
allocated in the same level and independent from each
other.
Association rule mining is a well explored method
in data mining (Pasquier et al., 1999). Based on asso-
ciation rules generated from a collection of item trans-
actions, we can discover the relations between items.
However, the amount of generated association rules
is usually huge and selecting the most useful rules is
challenging (Xu et al., 2011). In our research, we pro-
pose to identify a group of frequent patterns as po-
tential features to assist selecting useful association
rules. The selected rules are used to identify relation-
ships between features. Furthermore, in order to en-
sure that the most useful rules are to be selected, we
also propose to apply statistical topic modelling tech-
nique (Blei et al., 2003) to the selection of association
rules.
Our approach takes advantages of existing feature
extraction approaches and makes two contributions.
Firstly, we present a method to make use of associa-
tion rules to find related features. Secondly, we create
a product model called feature taxonomy which rep-
resents the product more accurately by explicitly rep-
resenting the concrete relationships between general
features and specific features.
184
Tian N., Xu Y., Li Y., Abdel-Hafez A. and Josang A..
Product Feature Taxonomy Learning based on User Reviews.
DOI: 10.5220/0004850201840192
In Proceedings of the 10th International Conference on Web Information Systems and Technologies (WEBIST-2014), pages 184-192
ISBN: 978-989-758-024-6
Copyright
c
 2014 SCITEPRESS (Science and Technology Publications, Lda.)

2 RELATED WORK
Our research aims to extract useful product infor-
mation based on user generated information to cre-
ate a product model. This work is closely related to
feature-based opinion mining which has drawn many
researchers’ attention in recent years. In detail, iden-
tifying features that have been mentioned by users is
considered the most significant step in opinion mining
(Hai et al., 2013). Hu and Liu (2004) first proposed
a feature-based opinion mining method to extract fea-
tures and sentiments from customer reviews. They
use pattern mining to find frequent itemsets (nouns).
These itemsets are pruned and considered frequent
product features. A list of sentiment words (adjec-
tives) that are nearby frequent features in reviews
can be extracted and used to identify those product
features that cannot be identified by pattern mining.
Scaffidi et al. (2007) improved the performance of
feature extraction in their proposed system called Red
Opal. Specifically, they made use of a language model
to find features by comparing the frequency of nouns
in the review and in common use of English. Those
frequent nouns in both reviews and in common use
are considered invalid features. Hu et al. (2010)
make use of SentiWordNet to identify all sentences
that may contain users’ sentiment polarity. Then, the
pattern mining is applied to generate explicit features
based on these opinionated sentences. In addition, a
mapping database has been constructed to find those
implicit features represented by sentiment words(e.g.,
expensive indicates price). To enhance the accuracy
of finding correct features from free text review, Hai
et al (2013) proposed a novel method which evaluates
the domain relevance of a feature by exploiting fea-
tures’ distribution disparities across different corpora
(domain-dependent review corpus such as cellphone
reviews and domain-irrelevant corpus such as culture
article collection). In detail, the intrinsic-domain rel-
evance (IDR) and extrinsic-domain relevance (EDR)
have been proposed to benchmark if a examined fea-
ture is related to a certain domain. The candidate
feature with low IDR and high EDR scores will be
pruned.
Lau et al. (2009) presented an ontology-based
approach to profile the product. In detail, a number
of ontology levels, such as feature level that contains
identified features for a certain product and sentiment
level in which sentiment words that describe a certain
feature are stored, have been constructed (Lau et al.,
2009). This method provides a simple product profile
rather than extracting product features only.
The statistical topic modeling technique has been
used in various fields such as text mining (Blei et al.,
2003; Hofmann, 2001) in recent years. Latent Se-
mantic Analysis (LSA) is first proposed to capture
the most significant features of a document collec-
tion based upon semantic structure of relevant doc-
uments (Lewis, 1992). Then, Probabilistic LSA
(pLSA) (Hofmann, 2001) and Latent Dirichlet Allo-
cation (LDA) (Blei et al., 2003) are proposed to im-
prove the interpretation of results from LSA. These
techniques have been proven more effective on doc-
ument modeling and topic extraction, which are rep-
resented by topic-document and word-topic distribu-
tion, respectively. Particularly, multinomial distribu-
tion over words which is derived based upon word fre-
quency can be generated to represent topics in a given
text collection.
None of aforementioned feature identification ap-
proaches is able to identify the relationships between
the extracted product features. The structural relation-
ships that exist between features can be used to de-
scribe the reviewed product in more depth. However,
how to evaluate and determine the relations between
features is still challenging.
The remainder of the paper is organized as fol-
lows. The next section illustrates the construction
process of our proposed feature taxonomy. Then, the
evaluation of our approach is reported afterwards. Fi-
nally, we conclude and describe future direction of
our research work.
3 THE PROPOSED APPROACH
Our proposed approach consists of two main steps:
product taxonomy construction using association
rules and taxonomy expansion based on reference fea-
tures. The input of our system is a collection of user
reviews for a certain product. The output is a product
feature taxonomy which contains not only all gener-
ated features but also the relationships between them.
3.1 Pre-processing and Transaction File
Generation
First of all, we construct a single document called an
aggregated review document which combines all the
reviews in a collection of reviews, keeping each sen-
tence in the original reviews as one sentence in the
constructed aggregated review document. Three steps
are undertaken to process the review text in order to
extract useful information. Firstly, we generate the
part-of-speech (POS) tag for each word in the aggre-
gated review document to indicate whether the word
is a noun, adjective or adverb etc. For instance, af-
ter the POS tagging,“The flash is very weak.” would
ProductFeatureTaxonomyLearningbasedonUserReviews
185

be transformed to “The/DT flash/NN is/VBZ very/RB
weak/JJ ./.”, where DT, NN, VBZ, RB, and JJ repre-
sent Determiner, Noun, Verb, Adverb and Adjective,
respectively. Secondly, according to the thumb rule
that most product features are nouns or noun phrases
(Hu and Liu, 2004b), we process each sentence in
the aggregated review document to only keep words
that are nouns. All the remaining nouns are also pre-
processed by stemming and spelling correction. Each
sentence in the aggregated review document consists
of all identified nouns of a sentence in the original
reviews. Finally, a transactional dataset is generated
from the aggregated review document. Each sentence
which consists of a sequence of nouns in the aggre-
gated review document is treated as a transaction in
the transactional dataset.
3.2 Potential Features Generation
Our first task is to generate potential product features
that are expressed by those identified nouns or noun
phrases. According to (Hu and Liu, 2004a), signif-
icant product features are discussed extensively by
users in reviews (e.g.,“battery” for cameras). Upon
this, most existing feature extraction approaches
make use of pattern mining techniques to find poten-
tial features. Specifically, an itemset is a set of items
(i.e., words in review text in this paper) that appear
together in one or multiple transactions in a transac-
tional dataset. Given a set of items, I =
{
i
1
,i
2
,...,i
n
}
,
an itemset is defined as X ⊆ I. The support of an
itemset X, denoted as Supp(X), is the percentage of
transactions in the dataset that contain X. All frequent
itemsets from a set of transactions that satisfy a user-
specified minimum support will be extracted as the
potential features. However, not all frequent item-
sets are genuine since some of them may be just fre-
quent but meaningless. We use compactness pruning
method proposed by (Hu and Liu, 2004a) to filter fre-
quent itemsets. After the pruning, we can get a list
of frequent itemsets that are considered potential fea-
tures, denoted as FP.
3.3 Product Feature Taxonomy
Construction
In this step, we propose to utilize association rules
generated from the discovered potential product fea-
tures to identify relations in order to construct a fea-
ture taxonomy.
Association rule mining can be described as fol-
lows: Let I =
{
i
1
,i
2
,...,i
n
}
, be a set of items, and
the dataset consists of a set of transactions D =
{
t
1
,t
2
,...,t
m
}
. Each transaction t contains a subset of
items from I. Therefore, an association rule r repre-
sents an implication relationship between two item-
sets which can be defined as the form X → Y , where
X, Y ⊆ I and X ∩ Y =
/
0. The itemsets X and Y are
called antecedent and consequent of the rule, respec-
tively. To assist selecting useful rules, the support
Supp(X ∪Y ) and the confidence Con f (X → Y ) of the
rule can be used (Xu et al., 2011).
For easily describing our approach, we define
some useful and important concepts as follows:
Definition 1 (Feature Taxonomy): A feature tax-
onomy consists of a set of features and their relation-
ships, denoted as FH =
{
F, L
}
, F is a set of features
where F =
{
f
1
, f
2
,..., f
n
}
and L is a set of relations.
The feature taxonomy has the following constraints:
(1) The relationship between a pair of features is the
sub-feature relationship. For f
i
, f
j
∈ F, if f
j
is
a sub feature of f
i
, then ( f
i
, f
j
) is a link in the
taxonomy and ( f
i
, f
j
) ∈ L, which indicates that f
j
is more specific than f
i
. f
i
is called the parent
feature of f
j
and denoted as P( f
j
).
(2) Except for the root, each feature has only one
parent feature. This means that the taxonomy is
structured as a tree.
(3) The root of the taxonomy represents the product
itself.
Definition 2 (Feature Existence):For a given fea-
ture taxonomy FH =
{
F, L
}
, let W (g) represent a
set of words that appear in a potential feature g, let
ES(g) =
n
a
i
|a
i
∈ 2
w(g)
,a
i
∈ F
o
contain all subsets of
g which exist in the feature taxonomy, ES(g) is called
the existing subsets of g, if
S
a
i
∈ES(g)
W (a
i
) = W (g),
then g is considered exist in FH, denoted as exist(g),
otherwise ¬exist(g).
Opinion mining is also referred as sentiment anal-
ysis (Subrahmanian and Reforgiato, 2008; Abbasi
et al., 2008; Wright, 2009). Adjectives or adverbs
that appear together with product features are consid-
ered as the sentiment words in opinion mining. The
following definition defines the sentiment words that
are related to a product feature.
Definition 3 (Related Sentiments): For a feature
f ∈ F , let RS( f ) denote a set of sentiment words
which appear in the same sentences as f in user re-
views, RS( f ) is defined as the related sentiments of
f .
Definition 4 (Sentiment Sharing): For features
f
1
, f
2
∈ F, the sentiment sharing between f
1
and f
2
is defined as SS ( f
1
, f
2
) = |RS ( f
1
) ∩ RS ( f
2
)|.
For deriving sub features using association rules,
we need to select a set of useful rules rather than using
all the rules. In the next two subsections, we will first
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
186

propose two methods to select rules, one method is
to select rules based on the sentiment sharing among
features and the other method is to select rules by
using the word relatedness derived from the results
generated by using the typical topic model technique
method LDA (Blei et al., 2003); then introduce some
strategies to update the feature taxonomy by adding
sub features using the selected rules.
In order to explain the topic modelling based
method, we first define some related concepts. Let
RE = {r
1
,r
2
,...,r
M
} be a collection of reviews, each
review consists of nouns only, W = {w
1
,w
2
,...,w
n
}
be a set of words appearing in RE, and Z =
{Z
1
,...,Z
v
} be a set of pre-specified hidden topics.
LDA can be used to generate topic models for rep-
resenting the collection as a whole and also for each
review in the collection. At the collection level, the
topic model represents the collection RE using a set
of topics each of which is represented by a probabil-
ity distribution over words (i.e., nouns in the context
of this paper) for topic. In this paper, we will use the
collection level representation to find the relatedness
between words.
At collection level, each topic Z
j
is represented
by a probability distribution over words, φ
j
=
{p(w
1
|Z
j
), p (w
2
|Z
j
),..., p (w
n
|Z
j
)},
∑
n
k=1
ϕ
j,k
= 1,
p(w
k
|Z
j
) is the probability of word w
k
being used
to represent the topic Z
j
. Based on the probability
p(w
k
|Z
j
), we can choose the top words to represent
the topic Z
j
.
Definition 5 (Topic Words): Let φ
j
=
{p(w
1
|Z
j
), p (w
2
|Z
j
),..., p (w
n
|Z
j
)} be the topic
representation for topic Z
j
produced by LDA
and 0 ≤ δ ≤ 1 be a threshold, a set of the topic
words for Z
j
, denoted as TW (Z
j
), is defined as
TW (Z
j
) = {w|w ∈ W, p (w|Z
j
) > δ}.
Definition 6 (Word Relatedness): We use word
relatedness to indicate how likely that two words have
been used to represent a topic together. Let w
i
,w
j
∈
W be two words, the word relatedness between two
words with respect to topic z is defined below:
W R
z
(w
i
,w
j
) =



1 − |p(w
i
|z) − p(w
j
|z)| w
i
∈ TW (z)
and w
j
∈ TW (z)
0 otherwise
(1)
Definition 7 (Feature Topic Representation):
For feature f ∈ F, let W D ( f ) be a set of words ap-
pearing in f and TW (z) be the topic words of topic
z. If W D ( f ) ⊂ TW (z), the feature topic representa-
tion of feature f for topic z is defined as FT P ( f , z) =
{(w, p (w|z))|w ∈ W D( f )}.
Definition 8 (Feature Relatedness): For features
f
i
, f
j
∈ F, if both features appear in a certain topic
z, then the feature relatedness between f
i
and f
j
with
respect to z is defined as:
FR
z
( f
i
, f
j
) = min
w
i
∈WD( f
i
)
w
j
∈WD( f
j
)
{W R
z
(w
i
,w
j
)} (2)
3.3.1 Rule Selection
Let R =
{
r
1
,r
2
,...,r
n
}
be a set of association rules
generated from the frequent itemsets FP, each rule r
in R has the form X
r
→Y
r
, X
r
and Y
r
are the antecedent
and consequent of r, respectively.
Assuming that f
e
is a feature which has already
been in the current feature taxonomy FH, to generate
the sub features for f
e
, we first select a set of can-
didate rules, denoted as R
c
f
e
, which could be used to
generate the sub features:
R
c
f
e
= {X → Y |X → Y ∈ R, X = f
e
,
Supp(X) > (Y )}
(3)
As defined in Equation (3), the rules in R
c
f
e
should
satisfy two constraints. The first constraint, X = f
e
,
specifies that the antecedent of a selected rule must
be the same as the feature f
e
. Sub features repre-
sent specific cases of a feature, they are more spe-
cific compared to the feature. The second constraint
is based on the assumption that more frequent item-
sets usually represent more general concepts, and less
frequent itemsets usually represent more specific con-
cepts. For instance, according to our observation to-
ward features, a general feature (e.g., “picture”, its
frequency is 62) appears more frequently than a spe-
cific feature (e.g., “resolution”, its frequency is 9)
in reviews for the camera 2 in the dataset published
by Liu (Ding et al., 2008). Therefore, only the rules
which can derive more specific features will be se-
lected.
However, not all selected rules represent correct
sub-feature relationship. For instance, mode → auto
is more appropriate for describing a sub-feature rela-
tionship rather than camera → auto. Therefore, the
rule camera → auto should not be considered when
we generate the sub features for “camera”. Upon
this, we aim to prune the unnecessary rules before
generating sub features for each taxonomy feature.
Firstly, a feature and its sub features should share sim-
ilar sentiment words since they describe the same as-
pect of a product at different abstract levels (e.g., vivid
can be use to describe both picture and color). There-
fore, we should select rules whose antecedent (rep-
resenting the feature) and consequent (representing a
possible sub feature) share as many sentiment words
as possible because the more sentiment words they
share, the more possible they are about the same as-
pect of the product. Secondly, based on topic models
ProductFeatureTaxonomyLearningbasedonUserReviews
187

generated from LDA, the more a feature and its po-
tential sub feature appear in the same topics, the more
likely they are related to each other.
Let f
X
, f
Y
be two features and Z
( f
X
, f
Y
)
be a set of
topics that contains both features, the feature related-
ness between f
X
, f
Y
with respect to all topics, denoted
as FR
avg
( f
X
, f
Y
), is defined as the average feature re-
latedness between the two features over Z
( f
X
, f
Y
)
:
FR
avg
( f
X
, f
Y
) =
∑
z∈Z
( f
X
, f
Y
)
FR
z
( f
X
, f
Y
)
|Z
( f
X
, f
Y
)
|
(4)
Based on this view, we propose the following equa-
tion to calculate a score for each candidate rule X → Y
in R
c
f
e
:
Weigh(X → Y ) = α(Supp(Y ) ×Con f (X → Y ))+
β
SS(X,Y )
|RS(X) ∪ RS(Y )|
+ γFR
avg
(X,Y )
(5)
0 < α,β,γ < 1. The value of α,β, and γ is set to
0.8, 0.1, and 0.1, respectively in our experiment de-
scribed in Section 4. There are three parts in Equation
(5). The first part is used to measure the belief to the
consequent Y by using this rule since Con f (X → Y )
measures the confidence to the association between X
and Y and Supp(Y ) measures the popularity of Y . The
second part is the percentage of the shared sentiment
words given by SS(X,Y ) over all the sentiment words
used for either X or Y . Yet, the third part in the equa-
tion is the average feature relatedness between X and
Y . Given a threshold σ, we propose to use the fol-
lowing equation to select the rules from the candidate
rules in R
c
f
e
. The rules in R
f
e
will be used to derive
sub features for the features in FP. R
f
e
is called the
rule set of f
e
.
R
f
e
= {X → Y |X → Y ∈ R
c
f
e
,
Weigh(X → Y ) > σ}
(6)
3.3.2 Feature Taxonomy Construction
Let FH = {F, L} be a feature taxonomy which could
be an empty tree, FP be a set of frequent itemsets
generated from user reviews which are potential fea-
tures, and R be a set of rules generated from user re-
views. This task is to construct a feature taxonomy
if F is empty or update the feature taxonomy if F is
not empty by using the rules in R. Let UF be a set
of features on the tree which need to be processed in
order to construct or update the tree. If F is empty,
the itemset in FP which has the highest support will
be chosen as the root of FH, it will be the only item
in UF at the beginning. If F is not empty, UF will be
F, i.e., UF = F.
Without losing generality, assuming that F is not
empty and the set of features currently on the tree,
UF is the set of features which need to be processed
to update or construct the tree. For each feature in
UF, let f
e
be a feature in U F, i.e., f
e
∈ UF and
X → Y ∈ R
f
e
be a rule with X = f
e
, the next step is
to decide whether or not Y should be added to the
feature taxonomy as a sub feature of f
e
. There are
two possible situations: Y does not exist in the feature
taxonomy, i.e., ¬exist(Y ) and Y does exist in the tax-
onomy, i.e., exist(Y ). In the first situation, the feature
taxonomy will be updated by adding Y as a sub fea-
ture of f
e
, i.e., F = F ∪ {Y }, L = L ∪ ( f
e
,Y ), and Y
should be added to UF for further checking.
In the second situation, i.e., Y already exists in the
taxonomy, i.e., according to Definition 2, there are
two cases, Y /∈ ES(Y ) (i.e., Y is not in the tree) or
Y ∈ ES(Y ) (i.e., Y is in the tree). In the first case,
Y is not considered a sub feature of f
e
and conse-
quently, no change is required to the tree. In the sec-
ond case, ∃ f
y
∈ F, f
y
is the parent feature of Y , i.e.,
P(Y ) = f
y
and ( f
y
,Y ) ∈ L. Now, we need to deter-
mine whether to keep f
y
as the parent feature of Y
or change the parent feature of Y to f
e
. That is, we
need to examine f
y
and f
e
to see which of them is
more suitable to be the parent feature of Y . The ba-
sic strategy is to compare f
y
and f
e
to see which of
them has more sentiment sharing and feature related-
ness with Y . Let f
P
, f
C
be a potential parent feature
and sub feature, respectively. We propose a rank-
ing equation to indicate how likely f
C
is related to
f
P
: Q( f
P
, f
C
) =
SS( f
P
, f
C
)
RS( f
C
)
+ FR
avg
( f
P
, f
C
). Thus, if
Q( f
y
,Y ) < Q( f
e
,Y ), the link ( f
y
,Y ) will be removed
from the taxonomy tree, ( f
e
,Y ) will be added to the
tree, otherwise, no change to the tree and f
y
is still the
parent feature of Y .
3.3.3 Algorithms
The construction of the feature taxonomy is to gener-
ate a feature tree by finding all sub features for each
feature. In this section, we will describe the algo-
rithms to construct the feature taxonomy. As men-
tioned above, if the tree is empty, the feature with the
highest support will be chosen as the root. So, at the
very beginning, F and UF contain at least one item
which is the root. Algorithm 1 describes the method
to construct or update a feature taxonomy.
After the taxonomy construction, some potential
features may be left over in RF and have not been
added to the taxonomy. The main reason is because
these itemsets may not frequently occur in the reviews
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
188

Algorithm 1:Feature Taxonomy Construction.
Input:
R, FH = {F,L}, FP.
Output:
FH, RF //RF is the remaining features which are
not added to FH after the construction
1: if F =
/
0, then root := argmax
f ∈F P
{supp( f )},
F := UF := {root};
2: else U F := F;
3: for each feature f
e
∈ UF
4: if R
f
e
6=
/
0 //the rule set of f
e
is not empty
5: for each rule X → Y ∈ R
f
e
6: if ¬exist(Y) //Y does not exist on the tree
7: F := F ∪ {Y },L := L ∪ ( f
e
,Y ),
UF := UF ∪ {Y },FP := FP − {Y };
8: else //Y exists on the tree
9: if Y ∈ ES(Y ) and Q( f
y
,Y ) < Q( f
e
,Y )
// f
y
is Y
0
s parent feature
10: L := L ∪ ( f
e
,Y ),L := L − ( f
y
,Y );
//add ( f
e
,Y ) and remove ( f
y
,Y )
11: else //Y /∈ ES(Y ), Y is not on the tree
12: FP := FP − {Y };
13: endfor
14: endif
15: UF := UF − { f
e
}; //remove f
e
from UF
16: endfor
17: RF := FP
together with the features that have been added in the
taxonomy. In order to prevent valid features from be-
ing missed out, we check those remaining itemsets in
RF by examining the shared sentiment words and fea-
ture relatedness between the remaining itemsets and
the features in the taxonomy. Let FH = {F,L} be
the constructed feature taxonomy, RF be the set of
remaining potential features, for a potential feature g
in RF, the basic strategy to determine whether g is a
feature or not is to examine the Q ranking between g
and the features in the taxonomy. Let F
g
= { f | f ∈
F, Q( f , g) > 0} be a set of features which are related
to g, if F
g
6=
/
0, g is considered a feature. The most re-
lated feature is defined as f
m
= argmax
f ∈F
g
{Q( f ,g)}.
g will be added to the taxonomy with f
m
as its parent
feature. If there are multiple such features f
m
which
have the highest ranking score with g, the one with the
highest support will be chosen as the parent feature of
g.
Algorithm 2 formally describes the method men-
tioned above to expand the taxonomy by adding the
remaining features.
After the expansion, the features left over in RF
are not considered as features for this product.
Algorithm 2: Feature Taxonomy Expansion.
Input:
FH = {F,L}, RF.
Output:
FH
1: for each feature g ∈ RF
2: if (F
g
:= { f | f ∈ F,Q( f , g) > 0}) 6=
/
0
3: M := {a|a ∈ F
g
and
Q(a,g) = max
f ∈F
g
{Q( f ,g)}}
4: f
m
:= argmax
f ∈M
{supp( f )}
5: F := F ∪ {g}, L := L ∪ ( f
m
,g)
6: RF := RF − {g}
4 EXPERIMENT AND
EVALUATION
We use three datasets in the experiments. Each dataset
contains user reviews for a certain type of digital cam-
eras. One dataset is used in (Hu and Liu, 2004a),
while the other two are used in (Ding et al., 2008).
Each review in the datasets has been manually anno-
tated. In detail, a human examiner read a review sen-
tence by sentence. If a sentence is considered indicat-
ing the user’s opinions, such as positive and negative,
all possible features in the sentence that are modified
by sentiment words are tagged. We take these anno-
tated features as the correct features to evaluate the
performance of our proposed method in feature ex-
traction. The number of reviews and number of anno-
tated features are 51 and 98 for camera 1, 34 and 75
for camera 2, and 45 and 105 for camera 3.
Our proposed feature taxonomy captures both
product features and relations between features.
Therefore, the evaluations are twofold: feature extrac-
tion evaluation and structural relations evaluation.
4.1 Feature Extraction Evaluation
First of all, we evaluate the performance of our ap-
proach by examining the number of accurate features
in user reviews that have been extracted. We use the
feature extraction method (FBS) proposed in (Hu and
Liu, 2004a) as the baseline for comparison. In ad-
dition, in order to examine the effectiveness of using
the sentiment sharing measure, the feature relatedness
measure, and the combination of the two, we conduct
our experiment in four runs:
(1) Rule: construct the feature taxonomy by only uti-
lizing the information of association rules (i.e.,
support and confidence value only) without using
the sentiment sharing and the feature relatedness
measures;
ProductFeatureTaxonomyLearningbasedonUserReviews
189

(2) SS: construct the feature taxonomy by taking the
information of association rules and the sentiment
sharing measure without using the feature related-
ness measure;
(3) FR: construct the feature taxonomy by taking the
information of association rules and the feature
relatedness measure without using the sentiment
sharing measure;
(4) Hybrid: the sentiment sharing and the feature re-
latedness are combined together with the informa-
tion of association rules to construct the feature
taxonomy.
Table 1: Recall Comparison.
Camera 1 Camera 2 Camera 3 Average
FBS 0.57 0.63 0.57 0.59
Rule 0.38 0.52 0.45 0.45
SS 0.56 0.65 0.58 0.60
FR 0.56 0.67 0.58 0.60
Hybrid 0.56 0.68 0.58 0.61
Table 2: Precision Comparison.
Camera 1 Camera 2 Camera 3 Average
FBS 0.45 0.42 0.51 0.46
Rule 0.55 0.57 0.74 0.62
SS 0.62 0.57 0.63 0.61
FR 0.60 0.56 0.63 0.60
Hybrid 0.62 0.59 0.68 0.63
Table 3: F1 Score Comparison.
Camera 1 Camera 2 Camera 3 Average
FBS 0.50 0.50 0.54 0.51
Rule 0.45 0.54 0.56 0.52
SS 0.59 0.61 0.60 0.60
FR 0.58 0.61 0.60 0.60
Hybrid 0.59 0.63 0.63 0.62
Table 1, 2, 3 illustrate the recall, precision, and F1
score results produced in the four runs, respectively.
From the results, we can see that using both the senti-
ment sharing and feature relatedness can obtain better
feature extraction performance than the use of asso-
ciation rule’s information only. In particular, the hy-
brid method, which uses both sentiment sharing and
feature relatedness, achieves the best results in most
cases. However, the size of the review dataset and
the number of annotated features can affect the pre-
cision and recall, which makes the values of the pre-
cision and recall vary in different range for different
datasets. For instance, camera 3 has higher precision
values than camera 2 due to more reviews in camera
3 dataset than that in camera 2 dataset, but camera
3 has lower recall values than camera 2 due to more
manually annotated features in camera 3 dataset.
4.2 Structural Relation Evaluation
The evaluation of the relations requires the standard
taxonomy or knowledge from experts (Tang et al.,
2009). Since there is no existing standard taxonomy
available for comparison, we manually created taxon-
omy for the three cameras according to the product
technical specifications provided online by manufac-
ture organizations
1, 2, 3
. From the product specifica-
tions on these websites, each camera has a number
of attributes such as lens system and shooting modes.
In addition, each attribute may also have several sub
attributes. For instance, the shooting modes of the
camera contains more specific attributes (e.g., intelli-
gent auto and custom). Based upon such information,
we create the product feature taxonomy for three dig-
ital cameras and use the taxonomy as the testing tax-
onomy, called Manual Feature Taxonomy (MFT ), to
evaluate the relations within our proposed feature tax-
onomy.
Due to the difference between the technical spec-
ifications from domain experts and the subjective re-
views from online users, the words used to represent
a feature in user reviews are very often different from
the words for the same feature specified by domain
experts in the product specification. For example, the
feature lens system in the testing taxonomy and the
feature lens in our generated taxonomy should be the
same according to common knowledge even though
they are not exactly matched with each other. Be-
cause of this fact, we will determine the match be-
tween two features based on overlapping of the two
features rather than exact matching.
Let MFT = {F
MFT
,L
MFT
} be the testing taxon-
omy with F
MFT
being a set of standard features given
by domain experts and L
MFT
being a set of links in
the testing taxonomy. For a given link ( f
F p
, f
Fc
) ∈ L
in the constructed product feature taxonomy and two
features f
M p
, f
Mc
∈ F
MFT
in the testing taxonomy, the
link ( f
F p
, f
Fc
) is considered matched with ( f
M p
, f
Mc
)
and therefore represent a correct feature relation if the
following conditions are satisfied:
1. W ( f
M p
) ∩W ( f
F p
) 6=
/
0 and W ( f
Mc
) ∩W ( f
Fc
) 6=
/
0
2. There exists a path in MFT ,
h f
M p
, f
1
, f
2
,..., f
n
, f
Mc
i, ( f
M p
, f
1
), ( f
i
, f
i+1
),
( f
n
, f
Mc
) ∈ L
MFT
, i = 1, ..., n − 1
We examine the testing taxonomy and the con-
structed taxonomy to identify all matched links in the
constructed taxonomy. The traditional measures pre-
cision and recall are used to evaluate the correctness
of the feature relations in the constructed feature tax-
onomy. Let ML(FH) denote the matched links in
1
http://www.canon.com.au/Personal/Products/Camerasand-
Accessories/Digital-Cameras/PowerShot-S100
2
http://www.nikonusa.com/en/Nikon-Products/Product/
Compact-Digital-Cameras/26332/COOLPIX-S4300.html
3
http://www.usa.canon.com/cusa/support/consumer/digital
cameras/powershot g series/powershot g3]Specifications
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
190

Figure 1: Constructed Feature Taxonomy.
Figure 2: Testing Feature Taxonomy.
the constructed taxonomy, the precision and recall are
defined as : Precision = ML(FH)/|L| and Recall =
ML(FH)/|L
MFT
|.
Table 4: Recall and Precision of Relation Evaluation
Relations in MFT Relations in FH Recall Precision
Camera 1 75 97 0.40 0.46
Camera 2 63 97 0.57 0.65
Camera 3 71 102 0.51 0.57
Table 4 illustrates the evaluation results includ-
ing the number of relations within the testing tax-
onomy, the number of relations within our generated
taxonomy, recall and precision for the three different
cameras, respectively. From the results, we can see
that our generated feature taxonomy correctly cap-
ture around 50% of the relationships. Figure 1 and
Figure 2 show a part of the feature taxonomy gen-
erated from our proposed approach and the testing
taxonomy generated based on the product specifica-
tion available online given by domain experts, respec-
tively. From the comparison, our generated feature
taxonomy identifies the relation between picture and
resolution. Although the testing taxonomy uses more
technical terms, which are image sensor instead of
picture; in fact, they refer to the same attribute of the
camera according to common knowledge. Similarly,
the (mode,auto) and (shooting modes,intelligent auto)
indicate the same relationship between two features.
As aforementioned, the online users and manufac-
ture experts may describe the same feature by using
totally different terms or words. This does affect the
performance (both recall and precision) of our pro-
posed approach in feature relationship identification
negatively. For instance, the user may prefer using
“manual” to depict a specific camera mode option.
By contrast, the manufacture experts usually pick the
term “custom” to describe this sub feature which be-
longs to ”shooting modes”. In such a case, the two
relations: (mode, manual) and (shooting modes, cus-
tom) cannot match.
5 CONCLUSION AND FUTURE
WORK
In this paper, we introduced a product feature taxon-
omy learning approach based on frequent patterns and
association rules. The objective is to not only extract
product features mentioned in user reviews but also
identify the relationship between the generated fea-
tures. The results of our experiment indicate that our
proposed approach is effective in both identifying cor-
rect features and structural relationship between them.
Particularly, the feature relationships captured in the
feature taxonomy provide more detailed information
about products. This leads us to represent products
profiles as multi-levels of feature, rather than a single
level as most other methods do.
In the future, we plan to improve and evaluate our
proposed product model by utilizing semantic simi-
larity tools. For instance, the vocabulary mismatch
can be handled by examining the semantic similar-
ity when we undertake the structural relation evalua-
tion. In addition, we plan to develop a review recom-
mender system that makes use of the proposed prod-
uct model in order to identify high quality reviews.
The structural relations of the product model are able
to assist identifying some characteristics of reviews,
such as how a certain feature and its sub features have
been discussed and how many different features have
been covered. Our system will therefore aim at rec-
ommending reviews based upon such criteria to help
users make purchasing decisions.
REFERENCES
Abbasi, A., Chen, H., and Salem, A. (2008). Sentiment
analysis in multiple languages: Feature selection for
opinion classification in web forum. ACM Transac-
tions on Information Systems, 26(3).
Blei, D. M., Ng, A. Y., and Jordan, M. I. (2003). Latent
dirichlet allocation. Journal of Machine Learning Re-
search, 3:993 – 1022.
Ding, X., Liu, B., and Yu, P. S. (2008). A holistic lexicon-
based approach to opinion mining. In Proceedings of
the 2008 International Conference on Web Search and
Data Mining, pages 231 – 240.
Hai, Z., Chang, K., Kim, J., and Yang, C. (2013). Identify-
ing features in opinion mining via intrinsic and extrin-
sic domain relevance. IEEE Transactions on Knowl-
edge and Data Engineering, pages 1 – 1.
ProductFeatureTaxonomyLearningbasedonUserReviews
191

Hofmann, T. (2001). Unsupervised learning by probabilistic
latent semantic analysis. Machine Learning, 42(1 -
2):177 – 196.
Hu, M. and Liu, B. (2004a). Mining and summarizing cus-
tomer reviews. In 10th ACM SIGKDD international
conference on Knowledge discovery and data mining.
Hu, M. and Liu, B. (2004b). Mining opinion features in
customer reviews. In Proceedings of the 19th national
conference on Artifical intelligence.
Hu, W., Gong, Z., and Guo, J. (2010). Mining product fea-
tures from online reviews. In IEEE International Con-
ference on E-Business Engineering, pages 24 – 29.
Lau, R. Y., Lai, C. C., Ma, J., and Li, Y. (2009). Automatic
domain ontology extraction for context-sensitive opin-
ion mining. In Proceedings of the Thirtieth Interna-
tional Conference on Information Systems.
Lewis, D. D. (1992). An evaluation of phrasal and clus-
tered representations on a text categorization task. In
Proceedings of the 15th ACM International Confer-
ence on Research and Development in Information
Retrieval, pages 177 – 196.
Pasquier, N., Bastide, Y., Taouil, R., and Lakhal, L. (1999).
Efficient mining of association rules using closed
itemset lattices. Information Systems, 24(1):25 – 46.
Popescu, A.-M. and Etzioni, O. (2005). Extracting product
features and opinions from reviews. In Proceedings of
the conference on Human Language Technology and
Empirical Methods in Natural Language Processing,
pages 339–346.
Scaffidi, C., Bierhoff, K., Chang, E., Felker, M., Ng, H., and
Jin, C. (2007). Red opal: Product-feature scoring from
reviews. In Proceedings of the 8th ACM conference on
Electronic commerce, number 182 - 191.
Subrahmanian, V. S. and Reforgiato, D. (2008). Ava:
Adjective-verb-adverb combinations for sentiment
analysis. IEEE Intelligent Systems, pages 43 – 50.
Tang, J., Leung, H.-f., Luo, Q., Chen, D., and Gong,
J. (2009). Towards ontology learning from folk-
sonomies. In Proceedings of the 21st international
jont conference on Artifical intelligence, pages 2089 –
2094.
Wright, A. (2009). Our sentiments, exactly. Communica-
tions of the ACM, 52(4):14 – 15.
Xu, Y., Li, Y., and Shaw, G. (2011). Representations for
association rules. Data and Knowledge Engineering,
70(6):237 – 256.
Zhang, Y. and Zhu, W. (2013). Extracting implicit features
in online customer reviews for opinion mining. In
Proceedings of the 22nd international conference on
World Wide Web companion, pages 103 – 104.
WEBIST2014-InternationalConferenceonWebInformationSystemsandTechnologies
192