Knowledge-Based Systems 135 (2017) 9–17
Contents lists available at ScienceDirect
Knowle dge-Base d Systems
journal homepage: www.elsevier.com/locate/knosys
FREERL: Fusion relation emb e dde d representation learning framework
for aspect extraction
Jian Liao
a
, Suge Wang
a , b , ∗
, Deyu Li
a , b
, Xiaoli Li
c
a
School of Computer & Information Technology, Shanxi University, China
b
Key Laboratory of Computational Intelligence and Chinese Information Processing of Ministry of Education, Shanxi University, China
c
Institute for Infocomm Research, A
∗
STAR, Singapore
a r t i c l e i n f o
Article history:
Received 5 February 2017
Revised 15 June 2017
Accepted 13 July 2017
Available online 14 July 2017
Keywords:
Fusion learning
Structure-based embedding
Language expression feature
Entity representation
a b s t r a c t
Opinion object-attribute extraction is one of the fundamental tasks of fine-grained sentiment analysis. It
is accomplished by identifying opinion aspect entities (including object entities and attribute entities) and
then aligning object entities to attribute entities. Recent studies on knowledge graphs have shown that by
adding the embeddings of semantic structures between opinion aspect entities, structure-based learning
models can achieve better performance in link-prediction than traditional methods. The studies, however,
focused only on learning semantic structures between aspect entities, did not take language expression
features into account. In this paper, we propose the Fusion RElation Embedded Representation Learning
(FREERL) framework, by which, one can fuse semantic structures and language expression features such
as statistical co-occurrence or dependency syntax, into the embeddings of object entities and attribute
entities. The obtained embeddings are then used to align object-attribute pairs and to predict new pairs
in a zero-shot scenario. Experimental results on the datasets of COAE2014 and COAE2015 show that the
best results in our framework achieve 12.1% and 32.1% improvements over the baselines, respectively.
© 2017 Elsevier B.V. All rights reserved.
1. Introduction
In the past decade, opinion mining has become a hot research
topic due to its potential to solve many challenging problems
[11] , such as affective com puting [3] , emotion recognition [24] and
data mining in social network [18,32] . According to the survey of
[33] , opinion mining can be studied at three levels of granular-
ity, namely, the document level, the sentence level, and the aspect
level. In the aspect level, opinion mining focuses on extracting as-
pect entities and analyzing their polarities. This task is similar to
the semantic role labeling problem [20] , the words or phrases in
sentences need to be identified in the sense of what is the object
and what is the attribute, but there are some differences between
them. Although opinion aspect is generally regarded as a whole
expression [12,25] , it is composed of opinion object and opinion
attribute, and thus provides more detailed and hierarchical infor-
mation for sentiment analysis.
For most fine-grained sentiment analysis tasks, opinion aspect
can be briefly represented as a triplet < object, attribute, polar-
ity > . This representation is more suitable for short texts in social
∗
Corresponding author.
E-mail address: wsg@sxu.edu.cn (S. Wang).
media such as Twitter and Weibo, which are generally produced
with only a few sentences and thus suffer from sparsity in the bag-
of-words model. In such short texts, opinion aspects are usually
nouns or noun phrases. For example, the weibo text “ ipad4
”(The new Apple ipad4’s battery is cheating.)
is represented as a triplet < ipad4, (battery), negative > ,
and the complete opinion target is “ipad4 ipad4 ” (ipad4’s
battery), where “ipad4” is the object and “ ” (battery) is
the attribute. It is obvious that the triple can be also represented
as a quadruple < object, attribute, expression, polarity > , i.e., <
ipad4, (battery), (cheating), negative > , where the
opinion expression is usually omitted because most opinion min-
ing tasks are only interested in the polarities instead of the opin-
ion expressions. From the perspective, it is more appropriate to re-
gard opinion mining as a two-stage process, namely, extracting the
pairs < object, attribute > and then analyzing the polarities of ex-
pressions. The two-stage process will produce more distinguishing
and unambiguous representations of opinions, especially in the cir-
cumstance where more than one expression can be extracted from
one opinion target. Another benefit of two-stage process is that it
is good at dealing with implicit opinion expressions whose sen-
timent polarity cannot be simply derived from explicit sentiment
words. These implicit opinion expressions are usually affected by
http://dx.doi.org/10.1016/j.knosys.2017.07.015
0950-7051/© 2017 Elsevier B.V. All rights reserved.