CROM – Car Review Opinion Miner
Maciej Dabrowski, Thomas Acton
J. E. Cairnes School of Business & Economics, National University of Ireland, Galway, Ireland
Przemyslaw Jarzebowski, Sean O’Riain
Digital Enterprise Research Institute, National University of Ireland, Galway, Ireland
Keywords: Opinion Mining, Sentiment Analysis, e-Commerce, Decision Making.
Abstract: Online shopping is a very goal-oriented activity. Consumers have a set of preferences for a product or
service that is used as criteria for assessment of the available alternatives. However, crucial information
about products is often available as text reviews. Finding a product with specific features is extremely time-
consuming using the typical search functionality found in existing shopping sites. In this work we propose a
method for the seamless integration of unstructured information from product reviews with structured
product descriptions using opinion mining. We demonstrate our method through shopping for a used car
based on 148240 car reviews. Evaluation results using a user study and simulations show that the technique
enables customers to assess more product characteristics and potentially make better decisions.
The increasing availability of product reviews
enables ubiquitous use among customers shopping
online or seeking additional or missing information
about products and services. Gretzel and Yoo (2008)
demonstrate that 97.7% of travel booking decisions
are made after consulting other travellers’ opinions,
of which 77.9% involve the use of customer reviews
as a source of information helping to make a better
decision. In this paper we propose a method for the
aggregation of information about products from
online customer reviews. We deal with the
contextual character of descriptive information using
cost-type and benefit-type attributes (Yang, 2008).
We show how this unstructured information can be
used to complement structured product descriptions
facilitating customer decisions. In particular, we
discuss the impact of the method from a decision-
making perspective.
Seamless integration of the information in product
descriptions with customer reviews requires dealing
with three tasks that have been investigated in the
research literature so far: extraction of feature terms,
opinion mining and sentiment analysis. Approaches
for the extraction of feature terms proposed in
OPINE (Popescu and Etzioni, 2005), in RedOpal
(Scaffidi et al., 2007) and by Hu and Liu (2004)
identify potential features using part-of-speech
(POS) tagging for nouns and nouns phrases. Hu and
Liu (2004) considered extracting neighbour opinion
phrases using a window of a size k on the output of a
noun phrase chunker. OPINE (Popescu and Etzioni,
2005) takes advantage of the syntactic dependencies
computed by the MINIPAR parser. One group of
approaches for opinion mining is based on using
term dictionaries such as WordNet to identify
opinion words in text reviews (Hu and Liu, 2004).
The main disadvantage is the limited set of terms
available in their term dictionaries. Another group of
approaches uses context-aware learned models of
opinion words (Popescu and Etzioni, 2005). These
Dabrowski M., Acton T., Jarzebowski P. and O’Riain S.
A ¸S Car Review Opinion Miner.
DOI: 10.5220/0002808703540357
In Proceedings of the 6th International Conference on Web Information Systems and Technology (WEBIST 2010), page
ISBN: 978-989-674-025-2
2010 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
models handle the limitation of the previous group
but may not generalize well across product
Our method differs from these approaches as
follows. First, we use a hierarchy of product features
based on domain knowledge, in contrast to flat lists.
Second, we use local lists of trigger terms based on
term occurrences for every element of a constructed
feature hierarchy, and use explicit features to
identify potential opinion phrases. Third, we
compute opinion sentiments on three levels: word
sentiment level, chunk sentiment level and context-
dependent chunk sentiment.
There are three generic tasks that an opinion mining
system needs to fulfil: identification of the product
features, discovery of opinion phrases, and
sentiment analysis (see Popescu and Etzioni, 2005,
Scaffidi et al., 2007). In our method, the first of
these tasks is performed using domain knowledge
and data from popular websites offering semi-
structured car reviews. We identified and organized
a list of all features of potential interest to
customers. The list was evaluated in a study with 28
subjects who were interested in buying a car and
organized into a hierarchy based on the domain
knowledge available.
The second task was implemented using a
modified version of the technique presented in Aciar
et al. (2007). The hierarchy was extended with the
list of “trigger terms” (phrases that symbolize
features). As product features are typically nouns or
noun phrases, we eliminated the infrequent and
irrelevant phrases from the set and used association
mining (performed with ARMiner software) to
identify potential bigram features. The set of trigger
phrases was later extended using bigram features,
similarly to Aciar (2007), organized into a hierarchy
and expanded to other parts of speech, for example:
driving (noun) -> drive (verb).
3.1 Opinion Mining
To extract opinion phrases and to select sentences
containing potential opinions, we used term
matching with terms in base morphological form of
a given speech component, using WordNet to
improve accuracy. Sentences containing potential
opinions were annotated with POS tags using the
GATE tagger, due to its accuracy on our corpus in
comparison to other available taggers. We used the
shallow-parsing method based on a set of rules that
extract potential opinions as chunks of text. The
rules are constructed to extract a consistent fragment
of the sentence that contains a feature (opinion head)
and the sentiment about the feature (opinion
content), similar to other common approaches (see
Aciar et al., 2007). The advantage of our method is
that not only nouns and noun phrases are considered
as features and not only adjectives are considered as
opinions. Thus, our method allows dealing with
context examples (e.g. “This car handles like a
dream.”). The method we used is similar to
approaches based on term proximity windows (Hu
and Liu, 2004), involving the computation of
syntactic dependencies (Popescu and Etzioni, 2005).
However, our approach accommodates language
structure, in contrast to Hu and Liu’s (2004)
approach, and is more efficient than that proposed
by Popescu (2005).
3.2 Sentiment Analysis
Our approach deals with sentiment analysis of three
levels, word level, chunk level, and context
dependant chunk level. To assess the sentiment of a
given opinion we used a lexicon-based method first
proposed by Kim and Hovy (2004). The initial list of
sentiment words with known sentiment was enlarged
with synonyms and antonyms based on WordNet.
As proposed by Ding et al. (2008), lists adjectives,
nouns, verbs and adverbs with positive and negative
sentiment were created - word sentiment is based on
the sum of all sentiment values. Using utility theory
(Butler et al., 2001), and to avoid the negative
effects of context, the features were divided into
three classes: Cost-type – features with preference
toward lower values (e.g. fuel consumption);
Benefit-type – higher values are preferred (e.g
reliability); Neutral – the character of a feature is
context-dependent. Similarly, sentiment words were
assigned to cost and benefit categories. Thus,
occurrence of a cost-type sentiment word (e.g.
“low”) with a cost-type feature (e.g. “price”) resulted
in positive sentiment. Conversely, the same word
occurring with a benefit-type feature resulted in
negative sentiment (e.g. “low quality”).
Sentiment of an opinion chunk describing a
feature was computed based on all the sentiment
words identified in the chunk. It is important to note
here that our method dealt with negation by
changing the sentiment of a word to its opposite. To
exploit full potential of our lexicon-based method
we considered chunk context. In our method
dependencies between chunks are assessed. If two
chunks are combined with a sentiment changing
word (e.g. but, however, despite), it is assumed that
two chunks have the opposite sentiment polarity. If
the chunks are connected using a word not from the
list of the sentiment changing words, the same chunk
sentiment is expected.
To evaluate the method proposed here we gathered
148240 car reviews from popular websites. Of these,
12561 were pure text reviews in English available at website, and 135679 were
semi-structured reviews from other websites (eg.
4.1 Feature Extraction
The feature extraction approach proposed here was
evaluated in a user study involving potential car
buyers. First, we listed the features available for
searching for a car at the most popular websites
offering used cars and car reviews. In total, list of 27
features was composed that included both attributes
from car sellers and car reviews. 32 participants
were asked to perform a feature categorization task
using a web application: 29 responded (91%
response rate), with 28 valid cases. There was no
time limit for task completion. Participants did not
report any important car features missing from our
Table 1: Results of the categorization task.
Avg. # of features / cat. 10.52 12 5.48
Std. dev. 3.18 3.28 3.37
Score per feature 2 1 0
Subjects’ responses were consistent, with
standardized Cronbach
= 0,74. The resulting
categorization shows on average 12 fairly important
features (FIF), 10.52 very important features (VIF),
and 5.48 not important features (NIF) per subject.
For convenience, we report a scoring system in
which every category was awarded a score from 0
(least important) to 2 (very important) (see Table 1).
The experiments show high interest of customers
in car features that are available in the reviews (C)
and are not available on typical shopping sites in
structured product descriptions. We note that 60% of
the TOP 10 highly ranked features was available
only in customer reviews (see Table 2).
Table 2: Average score (Score) for TOP 10 features and
number of votes for every category, T indicates source of
the feature (S-seller, C-customer opinion).
Feature Score VIF VIF FIF T
Overall Reliability
1.89 25 3 0
Major problems
1.82 24 3 1
Price 1.75 22 5 1 S
Mileage 1.61 18 9 1 S
1.61 19 7 2
Maintenance cost
1.54 16 11 1
Overall value
1.50 15 12 1
Year 1.43 14 12 2 S
Mechanical quality
1.39 15 9 4
Make 1.39 14 11 3 S
4.2 Opinion Mining
To evaluate the framework we designed a simulation
using a subset of product reviews we gathered from Due to limited resources we annotated
a corpus of 203 reviews (1233 sentences) of Ford
Focus cars, the most popular model in the dataset
based on the number of reviews and number of
online adverts available (2692 adverts, 4,73% of all
cars for sale at For every sentence in the
set an annotation was provided by a group of human
annotators. Every annotation consisted of a list of
featured mentioned explicitly and implicitly in a
sentence together with the expressed sentiment for
the feature using a 5 step scale: -2 (very negative), to
2 (very positive). Annotators negotiated
inconsistencies to avoid a potential negative impact
of subjective opinion on polarity and strength of the
sentiment. We evaluated the performance of our
method with precision and recall metrics for our test
dataset and accuracy of the sentiment analysis
technique (see Table 3).
Table 3: Opinion Mining evaluation results.
Opinion sentence extraction and
Precision Recall Accuracy
76,3% 77,5% 82,6%
4.3 Decision Making Impact
Consumers often face a task to select from a large
set of alternatives, such as choosing a car to buy.
Consumer websites often provide functionality to
search for alternatives, usually by asking a user to
provide his criteria for a desired product. Although
prevalent, both users and retailers can find such
functionality unsatisfying (Hagen et al., 2000). One
of the major reasons users are rarely provided with
WEBIST 2010 - 6th International Conference on Web Information Systems and Technologies
the information they need is that they are often not
able to transform their preferences into requirements
(Viappiani et al., 2008), or simply the information
they are looking for is not available in the
appropriate format.
Our method addresses such drawbacks of
existing websites by extracting opinions about
products and features from product reviews. Our
evaluation shows that reviews provide consumers
with information about products that is valuable to
them, and which is not available on standard
shopping websites. Further, the extracted features
are available together with the existing product
attributes so that no additional action is required
from consumers, decision-making effort is lower and
less time is required to make a decision (Scaffidi et
al., 2007). Moreover the customers can avoid time-
consuming analyses of product reviews. The direct
implication of such approach is the lower decision-
making effort, as less time and information
processing is required. Todd and Benbasat (2000)
point out that decision makers tend to trade off
decision quality for minimization of decision-
making effort: a reduction of the decision-making
effort from using our method can therefore increase
decision quality
We described an opinion mining system that extracts
and integrates opinions about products and features
from very informal, noisy text data (product
reviews) using a hierarchy of features from a
number of websites and domain knowledge. The
major contribution of this paper provides a decision
making perspective on integration of consumer
reviews in customer product selection and
evaluation of customer information needs in the used
car market.
Our method is of value not only to consumer-
based web providers and potential customers but
also to product manufacturers. Without additional
effort, the approach enables consumers to consider
further features of products only available in
customer reviews. Our approach can be of value in
various domains to both customers and product
The work presented in this paper has been funded in
part by Science Foundation Ireland under Grant No.
SFI/08/CE/I1380 (Lion-2)
Aciar, S., Zhang, D., Simoff, S. & Debenham, J. (2007)
Informed recommender: Basing recommendations on
consumer product reviews. Ieee Intelligent Systems,
22, 39-47.
Butler, J., Morrice, D. J. & Mullarkey, P. W. (2001) A
multiple attribute utility theory approach to ranking
and selection. Management Science, 47, 800-816.
Ding, X., Liu, B. & Yu, P. S. (2008) A holistic lexicon-
based approach to opinion mining. WSDM '08:
Proceedings of the international conference on Web
search and web data mining.
Gretzel, U. & Yoo, K. H. (2008) Use and Impact of Online
Travel Reviews Information and Communication
Technologies in Tourism. Innsbruck, Austria.
Hagen, P. R., Manning, H. & Paul, Y. (2000) Must Search
Stink? The Forrester Report
Hu, M. & Liu, B. (2004) Mining and summarizing
customer reviews. Proceedings of the tenth ACM
SIGKDD international conference on Knowledge
discovery and data mining. Seattle, WA, USA, ACM.
Kim, S.-M. & Hovy, E. (2004) Determining the sentiment
of opinions. COLING '04: Proceedings of the 20th
international conference on Computational
Popescu, A.-M. & Etzioni, O. (2005) Extracting product
features and opinions from reviews. Proceedings of
the conference on Human Language Technology and
Empirical Methods in Natural Language Processing.
Vancouver, British Columbia, Canada, Association for
Computational Linguistics.
Scaffidi, C., Bierhoff, K., Chang, E., Felker, M., Ng, H.,
Jin, C. & Acm (2007) Red Opal: Product-Feature
Scoring from Reviews. 8th Annual Conference on
Electronic Commerce (EC 2007). San Diego, CA,
Assoc Computing Machinery.
Todd, P. & Benbasat, I. (2000) Inducing compensatory
information processing through decision aids that
facilitate effort reduction: An experimental
assessment. Journal of Behavioral Decision Making,
13, 91-106.
Viappiani, P., Pu, P. & Faltings, B. (2008) Preference-
based search with adaptive recommendations. AI
Commun., 21, 155-175.
Yang, I. T. (2008) Utility-based decision support system
for schedule optimization. Decision Support Systems,
44, 595-605.