![](https://csdnimg.cn/release/download_crawler_static/18824203/bg1.jpg)
OPINION SUMMARIZATION OF CUSTOMER REVIEWS
Dongmei Zhang
1,2,
*, Hong’an Dong
3
, Sheng’en Li
1
, Jing Yi
1
, Ling Song
1
1
School of Computer Science & Technology
Shandong Jianzhu University
Jinan , China
2
School of Computer Science & Technology
Shandong University
Jinan, China
3
Department of computer science
Binzhou Polytechnic College
Binzhou, China
*E-mail zhandongmei@sdjzu.edu.cn
Fax (86)531-86366705
Keywords: Opinion Summarization; Automatic
Summarization; Latent Dirichlet Allocation; Semantic
Orientation
Abstract
This paper focuses on the problem of opinion summarization
of customer reviews, which has not been solved completely in
how to get a summary that overlaps the most important topics
of review text and to extract the opinion on these topics. With
the intuition that both factual text and evaluative text should
be considered in opinion summarization, this paper deals with
sentiment and topics of review text simultaneously and
proposes a multi-document opinion summarization approach
based on Latent Dirichlet Allocation model. Experiments on
product reviews show that the approach is effective.
1 Introduction
Recently, the Web has become a rich source of various
opinions in the form of customer reviews, consumer
complaints, movie reviews, issue discussions and so on.
Many product designers and manufacturers are facing with
the task of summarizing these evaluative text data and there
has been much interest in the task of multi-document opinion
summarization. For the website that collects many online
customer reviews of varies products, summaries of these
reviews could be of great useful to other consumers and be of
great value to consumers, product designers and
manufacturers. Thus, the corresponding summarization
systems must have the power of understanding the
sentiment/opinion expressed in such evaluated text. In other
words, these systems must deal with both the topic and
sentiment information in multiple documents rather than only
deal with the topics of documents.
This paper attempts to deal with the above problem. In order
to solve this problem, an intuitionist method is to classify the
sentences of reviews into subjective or objective categories
and summarize the opinion sentences separately. The flaw of
this kind of method is that it may be confounded by errors in
the objective and subjective classification phase. Some works
summarize product reviews through extracting product
features and opinion words [14], but it is difficult to extract
the feature and the opinion word accurately. This paper
considers sentiment and mixture of topics of a sentence
simultaneously in summarization phrase and avoids the above
problems.
The latter part of this paper is structured as follows. First, a
brief review of multi-document summarization and Latent
Dirichlet Allocation (LDA) model is given in Section 2. Then,
the methodology used in this paper in section 3 and the
experiments and results in detail in section 4. At last, section
5 concludes this paper.
2 Related Works
Automatic multi-document summarization receives more and
more attention as the number of web documents increases
dramatically. Multi-document summarization can provide
users with a clear and concise overview and can help users to
understand the tremendous documents quickly.
A representative multiple document summarization system is
MEAD
[1]
. Through treating words involved in most of the
text as the centroid, the MEAD system sorts the sentences by
the correlation between a sentence and the centroid and
extracts sentences to generate abstracts. This paper deals with
multi-document summarization that extracts several sentences
from multiple documents to form the abstract without any
modification to these sentences.
Latent Dirichlet Allocation (LDA) [2] model is a kind of topic
model that simulate a document’s generation process and is
used in this paper to capture the topics in text document
collection. Being a generative probabilistic model, LDA
assumes that a document is composed of mixture of latent
topics, a latent topic is represented as a probability
distribution over a vocabulary and the words of a document
are generated by mixture of latent topics using a probability
distribution.
Figure 1 [1] illustrates the LDA model where K represents the
topic number, M represents the document number and N
m
represents the length of a document m. W
m,n
and Z
m,n
,
respectively, is the nth word and its corresponding topic of
document m. ĭ
k
represents the probability distribution of
lexical items in topic k, and ș
m
represents the topic probability
1476