2018 2nd IEEE Advanced Information Management,Communicates,Electronic and Automation Control Conference(IMCEC 2018)

A Sentiment Analysis of Online Reviews Based on

the Word Alignment Model: A Product Improvement
Perspective
Shugang Li1, Yueming Li2
1. School of management, Shanghai University, PR China, 200444
2. School of management, Shanghai University, PR China
[email protected].

Abstract—Online reviews have become the main way for Traditionally, companies have relied on market surveys to
customers to publish their product usage perception. How to better understand consumers’ requirements. But the
extract valuable real-time information from online comments information obtained through the questionnaire is inaccurate
and make business decision with more intelligence is the main and incomplete. In this way, the quality of the survey results is
task of big data business. In this study, we put forward a word often not guaranteed. Online reviews, however, provide more
alignment model to analyze the customer's emotional response to comprehensive information than questionnaires. The number of
product attributes in online comments, from the perspective of online comments is large, so the amount of information is more,
product designers. Then, the KANO method is applied to analyze
and the information is more real, because the comments are the
online comments and develop appropriate product improvement
opinions and emotions that the customer wants to express. In
strategies. Our research shows that the combination of online
reviews and classic management patterns can bring success to big
addition, online reviews offer richer information on less time
data businesses. and lower costs than offline or paper-and-pencil surveys
because respondents are willing to participate independently
Keywords—Online comments; Product improvement; KANO ; [13,14]. However, because the amount of online reviews is too
Word alignment model large, it is impractical to observe the comments by artificial
means. Therefore, this study presents a new approach to
I. INTRODUCTION improve product design using online reviews. The main
research questions are: (1) how to get customers' opinions and
With the development of the Internet, it provides a new emotions on product attributes from online reviews, and (2)
platform for people to express their attitudes, opinions and how to use the opinions and emotions to improve the product.
feelings. At the same time, the explosion of online reviews,
presents a challenge that how people discover useful To answer these two questions, first we propose an
information and make use of it effectively in various situations. improved word alignment model to obtain customers' opinions
Previous studies have found that online reviews are much more and emotions on product attributes from online reviews. We
reliable than the information extracted from other sources. We verified the validity of our model by comparing it with the
can get more detailed information from online comments. For artificial discriminant method. Then, the product improvement
example, based on online reviews, companies can understand strategy is developed by combining the collocation model
what customers are satisfied with the product and what they are based on word alignment and KANO model. We use data from
dissatisfied with. Therefore, Sentiment analysis starts to play JD.com, one of China's largest online marketplaces, to test our
an important role in extract the opinions or sentiments of approach.
reviews, which is to identify and extract subjective information The rest of this paper is arranged as follows. Section 2
in text materials, such as opinions and feelings [1]. So, online provides a brief introduction to related work. Section 3
reviews have become an important source of information for introduces the data preprocessing. Section 4 describes the
consumers that significantly influence consumer choices and sentiment analysis approach with the word alignment model. In
product sales [2瀑11]. And sentiment analysis of online reviews section 5, product improvement strategies are appropriately
provides manufacturers with the opportunity to make use of produced by combining the collocation model based on word
online reviews for product designs. More and more companies alignment with the traditional KANO model. Finally, this
are paying close attention to the comments online reviews in paper discusses the management meaning and future research
the e-commerce environment. The manufacturers can be direction.
inspired by online reviews, which can be targeted for a series
of activities such as product quality improvement, new product
development and customer relationship management. This will II. LITERATURE REVIEW
improve the sales and market share of its products, which will Product manufacturers always want to know how their
bring greater economic benefits to the manufacturers [12]. products are used to optimize their designs to meet actual needs

978-1-5386-1803-5/18/$31.00 ©2018 IEEE 2226

and use, to provide a better user experience and outperform Through the notebook online review, we get the unique
their competitors. Therefore, many scholars have studied properties of the notebook, which we divide into 16 categories.
product improvement strategies from various aspects Then, for each attribute, find out what positive and negative
comments the customer has on each attribute. We divide it into
To exploit competitive information of mature-period (MP) two categories: positive and negative. Finally, find out the
products, Li Y L, Tang J F, Chin K S, et al. proposed a method, emotional words that are expressed when the client evaluates
which is based on the maximal deviation-based approach and negatively evaluates one of the attributes. We divide our
(MDBA) for dealing with the performance estimations of emotional words (satisfaction level) into five categories, which
engineering characteristics (ECs) to determine the TPs of ECs. is very satisfied, satisfied, general, dissatisfied and very
Then, AHP and scale method were integrated for acquiring the dissatisfied. This part of work is mainly used for the word
priority rating of achieving the improvement goal of alignment model of part 4 and part 5.
performance evaluation (IGPE) of each EC [1濈]. Bueno M R
and Borsato M proposed a decision support approach for
IV. MINING EMOTIONAL RELATIONSHIPS BETWEEN
identifying product improvement opportunities based on data
extracted from the assurance records. Their method uses field COMBINATION WORDS AND EMOTIONAL WORDS USING
data, which are analyzed under six different perspectives of THE WORD ALIGNMENT MODEL
failure modes [1濉]. However, these previous researches mainly This section is designed to recognize potential emotional
focus on the technology of product improvement. The data relationships in the sentence and assess the relationship
used is the objective use of products, such as product failure between combination words and emotional words. We
information, product competition information and so on. There temporarily undertake combination words and emotional words
is very little research on the use of emotions from a subjective respectively to be nouns/noun phrases and adjectives, which
perspective. JAD Santos and P Harland proposed a customer have been widely used in former work [濄濌瀑濅濅]. Therefore, our
satisfaction index for the product improvement project. Their purpose is to find the latent emotional relationship between
indicator makes it possible to simulate Customer Satisfaction in combination words and emotional words in the sentence and
the future and can verify how much the technical calculate the relation between them. As stated above, we regard
characteristics of a product contribute to Customer Satisfaction emotional relationship identification as a word alignment job.
[濄濊]. Qi J, Zhang Z, Jeon S, et al. proposed an automatic We adjust the bilingual word alignment model to the
filtering model to predict the helpfulness of online comments monolingual word alignment濁 In our way, every sentence is
from product designers ' perspective. The KANO method, copied to generate a similar corpus, and the bilingual word
which is based on classic conjoint analysis model, is then alignment model is applied to the monolingual situations, to
innovatively applied to analyze online comments to develop align a combination word with its emotional word.
appropriate product improvement strategies [濄濋]. Nevertheless,
Assumed a sentence with m words
customer satisfaction is a broad concept. The product designer
needs to further understand the customer's expressed emotions , the word alignment
when different attributes of the product show different qualities. , can be attained by
In this way, the product designer can better understand which maximizing the word alignment possibility of the sentence as
attributes of the product are more closely watched by the follows.
customer, thus making more accurate improvements to each
attribute of the product.
In this paper, we propose a word alignment model to
determine the weight of product attributes through attribute Where means that a combination word at
recognition and sentiment analysis. Then, the product position is aligned with an emotional word at position .
improvement strategy is developed by combining the
collocation model based on word alignment and KANO model. We employ the bilingual word alignment model, IBM
Model 1[23], to monolingual word alignment and to identify
emotional relations. The possibility of the alignment sequence
III. DATA ANALYSIS is premeditated using Eq. (2).
In this paper, we want to get the emotional (satisfaction
level) expressed by the customer when the customer has
positive evaluation and negative evaluation of the product
attributes. Therefore, we need to find out the unique attributes
of the product, and what positive evaluation terms and negative
evaluation words for each attribute. And find out the emotional Word collocation possibility describes the
words that are expressed when the client evaluates and
negatively evaluates one of the attributes. possibility of aligning with . For example, in a
In this article, we take a notebook as the research object. So, notebook comment, "satisfaction" often co-occurs with "large
what we need to do in the first step is to build up the memory", so " satisfaction " has high association with " large
characteristic word library that are unique to the laptop, the memory ".
evaluation word library and the emotional word library.

2227
 Using constrained monolingual word alignment model, we combination word of attribute words and negative evaluative
can create a series of aligned word pairs on the reviews. After words instead of the product having a certain attribute.
gettogether altogether word pairs from the comment sentences,
we compute the rate of recurrence of the two words aligned on This paper divides the factors that affect customer
sentiment (customer satisfaction) into five categories, including:
the reviews, signified as . Then, we can (1) Basic requirements (B), also known as essential
evaluation the alignment possibilities between combination requirements, are the basic requirements of customers for
words and emotional words as shown in Eq. (3) and (4). products or services provided by enterprises.
(2) Expected demand (E), also known as the willingness
type demand, refers to the requirement that the satisfaction of
the customer is proportional to the level of satisfaction of the
requirements.
(3) Attractive demand (A), also known as an excited
demand, is the need not to be overestimated by customers. As
for the attractive demand, customer satisfaction increases
sharply with the increase of meeting customer expectation. But
where incomes the alignment probabilities
once satisfied, even if the performance is not perfect, the
from emotional words to combination words. Similarly,
customer's performance satisfaction is very high. On the other
hand, even when expectations are not met, customers don't
means the alignment probabilities from show an obvious dissatisfaction.
combination words to emotional words. Therefore, like the
method of Liu et al [24], the sentiment association (4) Indifference demand (I): regardless of the performance
of such requirements, there is no impact on the user experience.
between a combination word and an emotional (5) Reverse demand (R): a quality feature that leads to
word is estimated as follows. strong dissatisfaction because not all consumers have similar
preferences. For example, some customers prefer high-tech
products while others prefer ordinary products. Excessive extra
functions can lead to customer dissatisfaction [29].
where is the choral factor to combine these two In the fourth section, we used the monolingual word
alignment possibilities. In this article, we set . alignment method to calculate the sentiment association (SA)
between combination words and emotional words. With the SA,
V. PRODUCT IMPROVEMENT STRATEGIES ANALYSIS USING THE we can classify the product attributes into the five categories
IMPROVED KANO MODEL described above according to the KANO model. Then, we can
get a two-dimensional table: the satisfaction degree of the
According to the monolingual word alignment
g model, positive evaluation and the satisfaction degree of the negative
we get sentiment association between evaluation. By the sentiment association of the positive/
negative combination words and satisfaction degrees, one
combination words and emotional words, that is, the sentiment attribute can be classified as a certain type of demand attribute.
(satisfaction degree) of the commenter on each attribute of the For example, when the customer has a positive evaluation of
product. With the sentiment association , we the battery, his emotions are general. But when customers have
can classify product attributes into different categories a negative evaluation of the battery, his emotions are
according to the KANO model. Here, we try to provide the unsatisfied. So, the battery is a basic attribute. This is shown in
mapping rules. table1.

The KANO model analysis method is based on the KANO

model for customer demand segmentation principle, developed
a set of structural questionnaire and analysis method.
In this paper, we combine emotional analysis with the
traditional KANO model. Unlike traditional questionnaire
survey, we use the word alignment method in the foreword to
find out the customer satisfaction degree from the commenter's
online comment. As mentioned in the third section, because we
want to obtain fine-grained customers’ emotion (customer
satisfaction) when they have different evaluations of product
attributes. We replace that the product does not have a certain
attribute in traditional KANO model with a combination word
of attribute-word and positive evaluation words. Similarly, the

2228
 TABLE I. ATTRIBUTE CLASSIFICATION RULES TABLE

Attribute Negative evaluation

Probability(SA)
Very Satisfied General Unsatisfied very
satisfied dissatisfied
Very Suspicious Indifference Attractive Expected E
Positive Probabilit satisfied results attribute attribute attribute
evaluation y(SA) Satisfied Reverse I A E E
attribute
General R I I Basic B
attribute
Unsatisfied R R I I B
very R R R R Suspicious
dissatisfied results
crawler to gather all online comments of a laptop on JD.com.
VI. EXPERIMENT Due to the boundaries of the anti-crawling, 4021 online
comments were eventually captured from JD.com.
A. Date collection and pre-processing
The first stage is to recognize the product attributes. We put
In this sector, we will select data resources and study
on POS (part-of-Speech) tagging, the Latent Dirichlet
objects. And then apply natural language processing (NLP)
Allocation and Page Rank. A whole of 3404 words were
techniques to infer emotional positioning of attributes.
removed as candidate attributes. We then separated the filtered
JD.com is a well-known e-commerce site in China. JD.com properties into 16 categories. To classify the attribute terms in
primarily emphases on digital products, and more and more online comments, 137 words have been built. The classification
Chinese are buying laptops at JD.com. Therefore, we choose attributes and relation quantities of terms are shown in table 2.
JD.com as the online comment resource. We established a data

TABLE II. CLASSIFICATION ATTRIBUTES AND RELATION QUANTITIES OF TERMS.

Attribute Number of Attribute Number of terms

terms
Feeling 17 Appearance 15
Boot speed 3 Function 13
Display 13 Temperature 3
System 5 Logistics 9
Price 4 Quality 11
Peripheral 4 CPU 18
Service 6 Battery 4
Memory 3 Rest 9
As the mood of the notebook is detailed, an existing unsatisfied, very dissatisfied. The emotional lexicon, including
emotional lexicon, for example ICTCLAS (Institute of 198 words, is established.
Computing Technology, Chinese Lexical Analysis System) is
not fit for this schoolwork. So, the building of emotional
lexicon is like the attribute lexicon such as the following table
3. To better apply the KANO model, we divide our emotional
words into five levels, very satisfied, satisfied, general,

2229
TABLE III. CLASSIFIED SENTIMENT AND RELATIVE NUMBER OF TERMS. TABLE IV. CLASSIFIED EVALUATION AND RELATIVE NUMBER
OF TERMS.
Sentiment Number of terms
Evaluation Number of terms
Very satisfied 45
Positive evaluation 133
Satisfied 50
Negative evaluation 161
General 14
Unsatisfied 47 B. Results
We used the word alignment model in the fourth part to
very dissatisfied 42 calculate the sentiment association (SA) between the
Because we want to obtain fine-grained customers’ combination words and emotional words in the 4021 comments.
emotion (customer satisfaction) when they have different The results are shown in table 5. According to the sentiment
evaluations of product attributes., we need to get combination association (SA) in table 5 and the attribute classification rules
words of combination words and evaluation words, which table (table 1), we can get the classification results for the
reviewers' comments on attributes. So, we also need to find out attributes, as shown in table 6.
the reviewers' evaluation of product attributes, which are
different from emotional words. For example, high-grade is the C. Performance analysis
evaluation word, happiness is the emotional word. The We verified the validity of our model by comparing it with
establishment of the evaluation word library is the same as the the artificial discriminant method. We invited three field
above methods. The evaluation lexicon includes 294 words, experts to read 4000 comments and determine which emotions
and the result is displayed in table 4. the clients expressed about the same product attributes. Then,
based on the judgment of three experts, the first three emotion
words of the customer were obtained. The results show that our
method is effective.

TABLE V. SENTIMENT ASSOCIATION TABLE BETWEEN THE COMBINATION WORDS AND EMOTIONAL WORDS(A PART OF THE RESULTS).

Very Satisfied General Unsatisfied very

satisfied dissatisfied
Pos-felling 0.131937 0.355084 0.289545 0.140542 0.082891
Neg-felling 0.093643 0.276122 0.307154 0.213511 0.10957
Pos-boot speed 0.15551 0.495366 0.252317 0.072091 0.024717
Neg-boot speed 0.12894 0.297994 0.336676 0.159026 0.077364
Pos-display 0.113019 0.386266 0.300429 0.135908 0.064378
Neg-display 0.090943 0.267569 0.370913 0.169485 0.10109
Pos-system 0.138584 0.371884 0.326022 0.117647 0.045862
Neg-system 0.069887 0.258993 0.366906 0.212744 0.09147
Pos-price 0.152406 0.385027 0.295455 0.11631 0.050802
Neg-price 0.074074 0.222222 0.382716 0.160494 0.160494
… … … … … …

TABLE VI. THE TABLE OF CLASSIFICATION RESULTS.

Attributes felling boot speed display system price peripheral

Classification A A A A A A

Attributes service memory appearance function temperature

Classification A I A A A

Attributes logistics quality CPU battery rest

Classification A A A A A

2230
 [4] C. Dellarocas, X. Zhang, N.F. Awad, Exploring the value of online
VII. DISCUSSION AND CONCLUSIONS product reviews in forecasting sales: the case of motion pictures, J.
Interact. Mark. 21(4) (2007) 23–45.
We use the perspective of product design to develop online [5] W. Duan, B. Gu, A.B. Whinston, Do online reviews matter?—An
reviews to improve products. We combine word collocation, empirical investigation of panel data, Decis. Support Syst. 45(4) (2008)
word alignment model and KANO model. Our work has a big 1007–1016.
impact on big data. [6] C. Forman, A. Ghose, B. Wiesenfeld, Examining the relationship
between reviews and sales: the role of reviewer identity disclosure in
electronic markets, Soc. Sci. Electron. Publ. 19 (3) (2008) 291–313.
A. Theoretical implications
[7] X. Li, L.M. Hitt, Self selection and information role of online product
Big data businesses depend on large amounts of data to reviews, Inform. Syst. Res. 19 (4) (2007) 456–474.
collect valuable real-time insights that drive smarter, more [8] Y. Liu, Word of mouth for movies: its dynamics and impact on box
lucrative business decisions. It is important to know how to office revenue, J. Mark. 70 (3) (2006) 74–89.
abstract value from big data for big data business. Based on [9] E. Toby Stuart, S. Olav, Social networks and entrepreneurship,
online comments, we considered the KANO model, and we Handbook of Entrepreneurship Research, Springer US, 2005, pp. 233–
displayed that big data analytics can get support for large data 252.
businesses from a knowledge-based perspective. [10] M. Sun, How does the variance of product ratings matter? Manage. Sci.
58 (4) (2012) 696–707.
If we can syndicate big data with classic management [11] F. Zhu, X. Zhang, Impact of online consumer reviews on sales: the
models, we can double the effort with half the effort. This moderating role of product and consumer characteristics, J. Mart. Q.
means that if we associate big data with classic management Publ. Am. Mark. Assoc. 74 (2) (2013) 133–148.
models, we can get more efficient and reliable insights from [12] Zhang qian, guo xiaolin. Research on the impact of online commodity
big data. reviews on manufacturers' earnings [J]. Business era, 2016(11):72-74.
[13] E.D.D. Leeuw, W.D. Heer, Longitudinal and international comparison,
Chin. J. Electron. 41 (3) (2002) 515–518.
B. Practical implications
[14] R.M. Groves, Nonresponse rates and nonresponse bias in household
The e-commerce industry can get the following surveys, Public Opin. Q. 70 (5) (2006) 646–675 (30).
enlightenment from our research. First, our research shows that [15] Li Y L, Tang J F, Chin K S, et al. Estimating the final priority ratings of
business decisions can depend on the combination of big data engineering characteristics in mature-period product improvement by
and knowledge. Our research suggests a way to derive value MDBA and AHP[J]. International Journal of Production Economics,
2011, 131(2):575-586.
from big data based on knowledge. Secondly, our research
[16] Bueno M R, Borsato M. A Method for Identifying Product Improvement
provides practical application for improving the improvement Opportunities through Warranty Data[C]// Moving Integrated Product
of new products. These results are useful for processing big Development To Service Clouds in the Global Economy: Proceedings of
data. Our findings are also helpful for those seeking to improve the, Ispe International Conference on Concurrent Engineering.
product design. 2014:122-131.
[17] Santos J A D, Harland P. Customer satisfaction as a quantifiable
Although we use laptops as research objects, our model is criterion for the product improvement process[C]// International Ice
universal and can easily be extended to other industries, such Conference on Engineering, Technology and Innovation. IEEE, 2012:1-
as other digital products. 8.
[18] Qi J, Zhang Z, Jeon S, et al. Mining Customer Requirements from
Online Reviews: A Product Improvement Perspective[J]. Information &
ACKNOWLEDGMENT Management, 2016, 53(8):951-963.
This work was supported by the Chinese National Natural [19] Hu M, Liu B. Mining and summarizing customer reviews[C]// Tenth
ACM SIGKDD International Conference on Knowledge Discovery and
Science Foundation (No. 71272232), and by Shanghai Pujiang Data Mining, Seattle, Washington, Usa, August. DBLP, 2004:168-177.
Program(No.15PJC049).
[20] Qiu G, Liu B, Bu J, et al. Opinion word expansion and target extraction
through double propagation[J]. Computational Linguistics, 2011,
37(1):9-27.
REFERENCES [21] Wang B, Wang H. Bootstrapping Both Product Features and Opinion
[1] Wang H, Liu L, Song W, et al. Feature-based Sentiment Analysis Words from Chinese Customer Reviews with Cross-Inducing[J]. 2009.
Approach for Product Reviews[J]. Journal of Software, 2014, 9(2). [22] Ding X, Liu B, Yu P S. A holistic lexicon-based approach to opinion
[2] J. Berger, A.T. Sorensen, S.J. Rasmussen, Positive effects of negative mining[C]// International Conference on Web Search and Data Mining.
publicity: when negative reviews increase sales, Soc. Sci. Electron. Publ. ACM, 2008:231-240.
29(5)(2007) 815–827. [23] Brown P F, Pietra V J D, Pietra S A D, et al. The mathematics of
[3] P.K. Chintagunta, S. Gopinath, S. Venkataraman, The effects of online statistical machine translation: parameter estimation[J]. Computational
user reviews on movie box-office performance: accounting for Linguistics, 1993, 19(2):263-311濁
sequential rollout and aggregation across local markets, Market. Sci.
29(5) (2010) 944–957.

2231