What is Web Opinion Mining?

1. Web Opinion Mining
Dupré Marc-Antoine, Alexander Patronas, Erhard Dinhobl, Ksenija Ivekovic,
Martin Trenkwalder
TU Wien, Wintersemester 2009/10
marcantoine.dupre@gmail.com, e0425487@student.tuwien.ac.at,
e0525938@student.tuwien.ac.at, xenia.ivekovic@gmail.com, trenkwaldermartin@gmail.com
Abstract. This paper covers an overview about the topic Web Opinion Mining,
which includes the structure of an opinion, several different approaches,
opinion spam and analysis and existing tools using sentiment analysis
techniques to gather the opinions from different sources.
Web 2.0 has dramatically changed the way in which people communicate with
each other. People are writing their point of view about every topic you can
imagine of on the web. For example there are opinions about people, a product,
website or a specific service. The need for good opinion mining is increasing in
a very fast way. Market analysis or companies capitalize on those techniques. A
very interesting aspect for those companies is the knowledge what people,
respectively the market, is thinking currently about a new product they just
released. Of course for individuals gathering opinions from several product
reviews is also very useful.
Keywords: Data Mining, Opinion Mining, Sentiment Analysis, Opinion Mining
Tools, Sentiment Analysis Tools
Introduction
Think about everything is posted on blogs, facebook-feeds, twitter and so on. Users
express there what they think, their opinions and also maybe their political, religious
point of view. There are also websites like wikipedia, research-information-sites or
something like that which describe facts. So we can distinguish between opinions and
facts on the web [14]. Data what is read and declared as a fact must be assumed that is
it is true. Currently search engines search and index facts. They can be associated
with keywords, tags and can be grouped by topics [14]. But opinions underlie a more
complex situation. Usually they are out of a question like, what do people think of
Motorola Cell phones, or, what do people in America think about Barack Obama [14].
Todays search algorithms are not designed to receive opinions. In most cases it is also
very difficult to determine such data and also user opinion data is mostly part of the
deep web [14] (Bing Liu defines this as user generated content, but this is exactly
what it is, namely the deep web mostly, but there is also other content). It is not part
of the global scope of the web but more on one’s circle of friends. Most data lies in

2. review sites, forums, blogs, black boards and so on. This type of information is also
called “Word-of-mouth”. To mine opinions expressed in such content needs some
kind of artificial intelligence algorithm [14]. This is not easy. But practically it would
be very useful, for example in market intelligence for organisations and companies to
serve better product and service advertising. Maybe persons are interested in other
opinions when purchasing products or discussion about political topics. It is also
interesting for overall search functions like “Opinions: Motorola cell phones” or
“BMW vs. Porsche”. Due that data types there crystallize out two types of opinions
namely direct opinions and comparisons. The former is some kind of expression on an
object like products, events, persons and so on. The latter describes a relation between
objects, usually an ordering of them like “product x is more expensive than y” [14].
These relations can be objective like prices but also subjective.
Opinion mining concept
To get a realizable way to opinion mining the process must be formalized. The basic
components of an opinion are [14]:
• Opinion holder: the person or organization which has written an opinion
on the web
• Object: object on which the opinion holder expressed the opinion
• Opinion: the content on the object from the opinion holder
Model
An object is an entity like a product, event or so and represents a hierarch of
components and each component is associated with attributes [14]. O is the root node.
There also exist sub-events or sub-topics. The represent the whole component tree we
describe this as “feature”. Therefore expressing an opinion on a feature makes it
easier not to determine between components attributes. In that sense the object is also
a feature. So the object O is defined by a finite set of feature F= {f1, f2, f3, …, fn}.
Every feature fi F defines a set of words or phrases Wi as synonyms. Wi W. W =
{W1, W2, W3, …., Wn}.
Now the opinion holder is j and comments on a subset of features Sj of F of O.
Now feature fk Sj is commented by j by a word or phrase from Wk to determine the
feature and a positive, negative or neutral opinion on fk.
Task
The opinion mining task seen as the sentiment classification is done on three levels.
[14] First it is done on document level. There is one assumption namely that one
document focuses only on a single opinion from a single opinion holder. In many

3. cases like forums or something like that this is not true therefore the document must
be separated. In this level the opinion is given the class it belongs to like positive,
negative or neutral. This level is too coarse grained for most applications. The second
one is mining at sentence level. In this level are two tasks. First one is to determine
the sentence type like objective or subjective. The second task is to determine the
sentence class to which it belongs to like positive, negative or neutral. The
assumption is that sentence contains only one opinion but this is not very easy to
match. Therefore clauses or phrases may be useful and focuses on identifying
subjective sentences. The last and third level of the mining task is done at the feature
level. In overall task focus is on sentiment words like great, excellent, horrible, bad,
worst and so. In topic-based classification topic words are important.
Summary-List
1. document level - class determining (1 opinion from 1 opinion holder)
2. sentence level (one opinion)
a. sentence type determining (objective or subjective)
b. sentence class determining (neutral, positive, negative)
3. feature level – determining words and phrases
Words and Phrases
The basic question is how to determine the sentiment classification on document and
sentence level [14]. Negative sentiment doesn’t mean that the opinion holder dislikes
the feature of the product or the whole product and a positive one that he/she likes
everything. There is more! Sentiment words are often context dependent, for example
long. Long runtime of a benchmark on a graphic card would be very bad but long
runtime of a battery would be very nice. To get such word and phrase lists there are
three approaches:
1. manual approach: manual creation of the list, one time effort
2. corpus-based approach
text is analyzed by co-occurrence patterns and is domain dependent
3. dictionary-based approach
Using constraints on connectives of words to identify opinion words, for
example “This camera is beautiful AND spacious” where and gives the same
orientation. This constraint using can also be applied to OR, BUT, EITHER-
OR and NEITHER-OR. For this learning approach there exists a database
which contained 21 million words in 1987. There is a good online resource
called “WordNet”.
Document-level sentiment analysis
In order to analyse the general opinion of documents most of the research studies
use classifiers. A classifier is an algorithm or the program based on it. Given a set of

4. documents, a sentiment classifier classifies each document in two classes : positive or
negative (the class neutral is seldom used). A document classified in the positive class
expresses a general opinion which is positive. And a document classified as negative
expresses a general negative opinion. Such a classifier is unable to determine who are
the holders of the opinions or what are the objects targeted by the opinions. Thus the
set of documents has to be chosen wisely, the topic of all the documents could be a
single object for example. It is assumed that a single document only expresses the
opinion of a single holder. Several approaches exist to perform sentiment
classification at a document level, we describe three of them below [14, 30].
Classification based on sentiment phrases
This approach is a research field of Tuney [28]. It can be divided into three steps.
First the document is tagged using the Part-of-speech (POS) method [30]. It
basically replaced each word by a linguistic category according to its syntactic or
morphological behavior. For instances, JJ means adjective and VBN means verb in
past participle. It has be proven [29] that, for sentiment classification purposes, the
adjectives are the most relevant words. Nevertheless an adjective may have several
semantic orientation depending of the context. "unpredictable" might be negative in a
automotive review but be positive in a movie review [29]. That is why, thanks to the
POS tagging, pairs of words are extracted depending on precise patterns in order to
determine precisely the semantic orientation of the adjectives. The following table
contains some of the patterns used for extracting two-words phrases.
First word Second word Third word
(not extracted)
JJ NN anything
RB JJ not NN
JJ JJ not NN
NN JJ not NN
RB VB anything
The table above presents a simple version of the extraction patterns. NN are nouns,
RB adverbs, VB vers and JJ adjectives. For example, in the sentence "This camera
produces beautiful pictures", "beautiful pictures" will be extracted (first pattern : NN
+ JJ).
The second step is based on a measure called the pointwise mutual information
(PMI). The concept is to search if a given phrase is more likely to co-occur with the
word "excellent" or with the word "poor" on the web.

5. Pr(term1 ^ term2) is the probability that term1 and term2 co-occur.
Pr(term1)Pr(term2) is the probability that term1 and term2 co-occur if they are
statistically independant. Thus the ratio gives an information about the statistical
dependence of those two terms. Tuney proposes to compute a value of the semantic
orientation of a phrase by the following way :
Then by using the number of hits on a search engine it is possible to estimate the
probabilities and the SO equation becomes
The last step of Turney's algorithm is, given a review, to compute the average SO
of all phrases in the review. If it is greater than null then the review expresses a
positive opinion. Otherwise it expresses a negative opinion.
Final classification results on reviews from various domains are from 84% for
automobile review to 66% for movie reviews [29, 30].
Classification using text classification methods
Sentiment classification can be tackled as a topic-based text classification problem.
All the usual text classification algorithms can be used, e.g., naïve Bayes, SVM, kNN,
etc. This approach was experimented by Pang et al. [31].
They have classified 1400 movies reviews from IMDb.com with a random-choice
baseline of 50%. They used the three following algorithms, SVM, naïve Bayes and
Maximum Entropy. Each of those algorithm usually produces good results on text
classification problems.
With various pre-processing options and a 3-fold cross-validation, the results
spread from 72.8% to 82.9%. The best result is achieved by SVM algorithm on
unigrams data. All the results are above the random-choice baseline and the human
bag-of-words experiences (58% and 64%). They are superior to the PMI-IR algorithm
from Turney on movies review (66%).

6. Still the three used algorithms are expected to get results around 90% on topic-
based text classification problems. Thus sentiment classification is a more difficult
task because of the various semantic values and uses of sentiment phrases.
Classification using a score function
Another approach by Dave et al. [32] is by using a score function. The first step is
to score each term of the learning set with the following score function
the score number is between -1 and 1, it indicates toward which class, C or C', the
term is more likely to belong to. A learning set is a set of reviews which have been
labeled manually. So it is possible to compute statistics such as Pr(t|C): probability
that the term t appears in a review belonging to class C. Then a document is classified
according to the sum of the scores of all its terms. On a large set of reviews from the
web (more than 13000) and by working with bigrams and trigrams, the classification
rate is between 84.6% and 88.3%.
Sentence-level sentiment analysis
The sentiment classification at the document-level is the most important field of
web opinion mining. However, for most applications, the document-level is too
coarse. Therefore it is possible to perform finer analysis at the sentence-level. The
research studies in this field mostly focus on a classification of the sentences wether
they hold a objective or a subjective speech, the aim is to recognise subjective
sentences in news articles and not to extract them. The sentiment classification as it
has been described in the document-level part still exists at the sentence-level, the
same approaches as the Turney's algorithm are used, based on likelihood ratios.
Because this approach has already been described in this paper, this part focuses on
the objective/subjective sentences classification and presents two methods to tackle
this issue.
The first method is based on a bootstrapping approach using learned patterns. It
means that this method is self-improving and is based on phrases patterns which are
learned automatically. This method comes from the study of Wiebe & Riloff [33], the
following schema helps to understand the bootstrapping

7. process.
The input of this method is known subjective vocabulary and a collection of
unannotated texts.
• The high-precision classifiers find wether the sentences are objective or
subjective based on the input vocabulary. High-precision means their
behaviours are stable and reproductible. They are not able to classify all the
sentences but they make almost no errors.
• Then the phrase patterns which are supposed to represent a subjective
sentence are extracted and used on the sentences the HP classifiers have let
unlabeled.
• The system is self-improving as the new subjective sentences or patterns are
used in a loop on the unlabeled data.
This algorithm was able to recognise 40% of the subjective sentences in a test set
of 2197 sentences (59% are subjective) with a 90% precision. In order to compare, the
HP subjective classifier alone recognises 33% of the subjective sentences with a 91%
precision.
Along this original method, more classical data mining algorithm are used such as
the naïve bayes classifier in the research studies of Yu & Hatzivassiloglou [34]. The
naïves bayes is a supervised learning method which is simple and efficient, especially
for text classification problems (i.e. when the number of attributes is huge). To cope
with an important and unavoidable approximation about their training data to avoid
human labelisation on enormous data set, they use a multiple naïve bayes classifiers
method. The general concept is to split each sentence in features -- such as presence
of words, presence of n-grams, heuristics from other studies in the field -- and to use

8. the statistics of the training data set about those features to classify new sentences.
Their results show that the more features, the better. They achieved at best a 80-90%
recall and precision classification for subjective/opinions sentences and a 50% recall
and precision classification for objective/facts sentences.
The sentence-level sentiment classification methods are improving, this results
from research studies in 2003 show that they were already quite efficient then and that
the task is possible.
Feature-based opinion mining
Main objective of feature-based opinion mining is to find what reviewers (opinion
holders) like and dislike about observed object. This process consists of following
tasks:
1. extract object features that have been commented on in each review
2. determine whether the opinions on the features are positive, negative or
neutral
3. group feature synonyms
4. produce a feature-based opinion summary
There are three main review formats on the Web which may need different techniques
to perform the above tasks:
1. Format 1 – Pros and Cons: The reviewer is asked to describe Pros and Cons
separately. Example: C|net.com
2. Format 2 – Pros, Cons and detailed review: The reviewer is asked to
describe Pros and Cons separately and also write a detailed review. Example:
Epinions.com
3. Format 3 – free format: The reviewer can write freely, there is no separation
of Pros and Cons. Example: Amazon.com
Analysing reviews of formats 1 and 3:
The summarization is performed in three main steps:
1) mining product features that have been commented on by customers:
• part-of-speech tagging: Product features are usually nouns or noun phrases
in review sentences. Each review text is segmented into sentences and part-
of-speech tag is produced for each word. Each sentence is saved in the
review database along with the POS tag information of each word in the
sentence.
Example of sentence with POS tags:

9. <S> <NG><W C='PRP' L='SS' T='w' S='Y'> I </W> </NG>
<VG> <W C='VBP'> am </W><W C='RB'> absolutely
</W></VG> <W C='IN'> in </W> <NG> <W C='NN'> awe
</W> </NG> <W C='IN'> of </W> <NG> <W C='DT'> this
</W> <W C='NN'> camera </W></NG><W C='.'> .
</W></S>
• frequent feature identification: Frequent features are those features that are
talked about by many customers. To identify them, association mining is
used. However, not all candidate frequent features generated by association
mining are genuine features. Two types of pruning are used to remove those
unlikely features. Compactness pruning checks features that contain at least
two words, which we call feature phrases, and remove those that are likely
to be meaningless. In redundancy pruning, redundant features that contain
single words are removed. Redundant features are described with concept of
p-support (pure support). p-support of feature ftr is the number of sentences
that ftr appears in as a noun or noun phrase, and these sentences must contain
no feature phrase that is a superset of ftr. Minimum p-support value is used
to prune those redundant features.
• infrequent feature generation: For generating infrequent features following
algorithm is applied:
for each sentence in the review database
if (it contains no frequent feature but one or more opinion words)
{ find the nearest noun/noun phrase around the opinion word. The
noun/noun phrase is stored in the feature
set as an infrequent feature. }
2) identify orientation of an opinion sentence
To determine the orientation of the sentence, dominant orientation of the opinion
words (e.g. adjectives) in the sentence is used. If positive opinion prevails, the
opinion sentence is regarded as a positive and opposite.
3) Summarizing the results
The following picture shows an example summary for the feature picture of a
digital camera.
Feature: picture
Positive: 12
•
Overall this is a good camera with a really good
picture clarity.
•
The pictures are absolutely amazing - the camera

10. captures the minutest of details.
•
After nearly 800 pictures I have found that this camera
takes incredible pictures.
…
Negative: 2
•
The pictures come out hazy if your hands shake even
for a moment during the entire process of taking a
picture.
•
Focusing on a display rack about 20 feet away in a
brightly lit room during day time, pictures produced by
this camera were blurry and in a shade of orange.
Analysing reviews of format 2:
Features extracted based on the principle that each sentence segment contains at most
one product feature. Sentence segments are separated by ‘,’, ‘.’, ‘and’, and ‘but'.
For extracting product features, suprevised rule discovery is used. First, training
dataset has to be prepaired. The steps are following:
• perform part-of-speech tagging
e.g.
<N> Battery <N> usage
<V> included <N> MB <V>is <Adj> stingy
• replace actual feature words in a sentence with [feature]
e.g.
<N> [feature] <N> usage
<V> included <N> [feature] <V> is <Adj> stingy
• use n-gram to produce shorter segments from long ones
e.g.
<V> included <N> [feature] <V> is
<N> [feature] <V> is <Adj> stingy
After these steps, rule generation can be performed – definition of extraction patterns.
e.g. of extraction pattern:
<JJ> <NN> [feature]
easy to <VB> [feature]
The resulting patterns are used to match and identify features from new reviews.
Sometimes mistakes made during extraction have to be corrected. E.g. when there are

11. two or more candidate features in one sentence segment or there is a feature in the
sentence segment but not extracted by any pattern. First problem can be solved by
implementing an iterative algorithm to deal with the problem by remembering
occurrence counts.
Orientation (positive or negative) of extracted features is easily to define as we know
if the feature is from Pros or Cons of a review.
These features are usually used to make comparison of consumer’s opinions of
different products.
Opinion Spam and Analysis
The web has dramatically changed the way that people express themselves and
interact with others. They are now able to post reviews of products at merchant sites
and interact with others via blogs and forums. Reviews contain rich user opinions on
products and services. They are used to by potential customers to find opinions of
existing users before deciding to purchase a product and they are also helpful for
product manufacturers to identify product problems and to find marketing intelligence
information about their competitors. Due to the fact that there is no quality control,
anyone can write anything on the Web. This results in many low quality reviews and
review spam.

12. It is now very common for people to read opinions on the Web for many reasons. For
example, if someone wants to buy a product and sees that the reviews of the product
are mostly positive, one is very likely to buy the product. If the reviews are mostly
negative, one is very likely to choose another product. There are generally three types
of spam reviews:
1. Untruthful opinions: Those that the reviewer is giving an unjustly positive
review to a product or an object in order to promote the object (hyper spam)
or when the reviewer is giving some wrongly negative comment to some
object in order to damage that product (defaming spam)
2. Reviews on brands only: Those are comments given by a reviewer only for
the brands, the seller or the manufactures but not for the specific product or
object. In some cases it is useful, but it is considered as spam because it
focuses not to the specific product.
3. Non-Reviews: Those are comments that are not related to the product, for
example advertisements, questions, answers and random texts.
In general, spam detection can be regarded as a classification problem with two
classes, spam and non-spam. However, due to the specific nature of different types of
spam, we have to deal with them differently. For spam reviews of type 2 and type 3,
we can detect them based on traditional classification learning using manually labeled
spam and non-spam reviews because these two types of spam reviews are
recognizable manually. Quite a lot of reviews of this two types are duplicates and
easy to detect. To detect the remaining spam reviews it is necessary to create a model
containing the following model:
• The content of the review:
i.e. number of helpful feedbacks, length of the review title, length of the
review body, position of the review, textual features etc.
• The previewer who wrote the review:
i.e. number of reviews of the reviewer, average rating given by the reviewer,
standard deviation in rating
• The product being reviewed:
i.e. price of the product, average rating, standard deviation in ratings
Using the just discussed model with logistic regression, that produces a probability
estimate of each review being a spam. It was evaluated on 470 spam reviews searched
on amazon.com and it got the following result:
Spam Type Num AUC AUC – text AUC – w/o
reviews features only feedbacks
Types 2 & 3 470 98,7 % 90 % 98%
Types 2 only 221 98,5 % 88 % 98 %
Types 3 only 249 99,00 % 92 % 98 %

13. To perform the logistic regression it was used the statistical package R (http://www.r-
project.org/). The AUC (Area under ROC curve) is a standard measure used in
machine learning for assessing the model quality.
Without using feedback features it is reached quite the same result as include them in
the evaluation. This is important because feedbacks can be spammed too.
However for the first type of spam, manual labeling by simply reading the reviews is
quite impossible. The point is to distinguish the untruth review of a spammer from a
innocent review. The only way is to create a logistic regression model using
duplicates as positive training examples and the rest of the reviews as negative
training examples. The model was evaluated on totally 223.002 reviews, of them there
was 4488 duplicate spam reviews and 218514 other reviews.
Features used AUC
All features 78 %
Only review features 75 %
Only reviewer features 72,5 %
Without feedback features 77 %
Only text features 63 %
The table shows that review centric features are most helpful. Using only text features
gives only 63 % AUC, which demonstrates that it is very difficult to identify spam
reviews using text content alone. Combining all the features gives the best result.
Opinion mining Tools
Following there will be a categorized list of several different tools, which can be used
for opinion mining. A short review is enclosed for each mentioned tool.
APIs
Evri
[15] Evri is a semantic search engine. It automatically reads web content in a similar
way humans do. It performs a deep linguistic analysis of many millions documents
which is then build up to a large set of semantic relationships expressing grammatical
Subject-Verb-Object style clause level relationships.
Evri offers a complex API for developers, with it, it is easy to automatically, cost
effectively and in a fully scalable manner, analyze text, get recommendations,
discover relationships, mine facts and get popularity data.

14. Further it is possible to get Widgets with different usage, one of those is using the
sentiment aspect. An example of the sentiment widget displays the positive and
negative aspects in a percentage bar of the opinion on the new mobile operating
system running on the Linux Kernel, called “Android”. [16]
OpenDover
[17] OpenDover is a Java based webservice that allows to easily integrate sementic
features within your blog, content management system, website or application.
Basically how it works is that your content is sent through a webservice to their
servers (OpenDover), which process and analyze the content by using the
implemented linguistic processing technologies. After processing the content is sent
back, emotion tagged along with an indicating value how positive or negative the
content is.
Without any effort it is possible to test this service at a live-demo site on their website
[17].
As an example i chose an arbitrary review on a camera from amazon.com:
„...the L20 is unisex and it's absolutely right in line with the jeweled quality of Nikon.
I was able to use the camera right out of the box without having to read the
instruction manual, it's that easy to use....
The camera feels good in my hands and the controls are easy to find without having
to take your eyes off your subject...
The Nikon L20 comes with a one year manufactures warranty - "Not that you would
need a warranty for a Nikon camera" - Impressive warranty details, I was amazed
that any camera manufacturer would offer a one year on a point and shoot but Nikon
has such a good reputation and so I doubt very much that you would even need to use
it.
In a nutshell, I love this camera so much that I would recommend this Nikon L20 to
my friends, family and anyone else looking to buy. It's a real beauty!“
The first BaseTag was set to “camera”, the second to “Nikon L20”, which the product
review was about. The Mode was set to “Accurate” and the selected subject domain
was “camera”.
The output is then the emotion tagged text. It recognizes positive, negative words and
the object. The result of their algorithm is good, for example positive words like “easy
to use”, “good”, “impressive” and “love” are marked with green colour.
Twitter/Blogsphere
RankSpeed
[18] RankSpeed is a sentimant search tool for the blogosphere / twittersphere.
It finds the best websites, the most useful web apps, the most secure web services and
other topics with the help of sentiment analysis.
It is possible to search for any website category using tags and rank them by any

15. desired criteria. Criterias like good, useful, easy and secure.
A statistical analysis computes the percentage of bloggers/users who correspond to
the desired criteria. The given result, which is a list of links from the source, is then
sorted in an descending order by the given percentage.
Twittratr
[19] Twittratr is a simple search tool for answers to questions like "Are Tweets about
Obama generally positve or negative?". The functionality is kept simple. It is based
on a list of positive and negative keywords. Twitter is searched for these keywords
and the results are crossreferenced against their adjective lists, then displayed
accordingly.
TwitterSentiment
[20] "Twitter Sentiment is a graduate school project from Stanford University. It
started as a Natural Language Processing class project in Spring 2009 and will
continue as a Natural Language Understanding CS224U project in Winter 2010."
Twitter Sentiment was created by three Computer Science graduate students at
Stanford University: Alec Go, Richa Bhayani, Lei Huang. It is an academic project.
As it is obvious the are doing sentiment analysis on Tweets from Twitter.
[27] The approach they are working with is different from other sentiment analysis
sites due to following reasons:
• Use of classifiers built from machine learning algorithms. Other sites tend to
use a keyword-based approach which is much simpler, it may have higher
precision, although lower recall.
• Transparent in how classification on individual tweets is done. Other sites
often do not display the classification of individual tweets. There are only
showing aggregated numbers, which makes it almost impossible to assess
how accurate the classifiers are.
WE Twendz Pro
[21] Waggener Edstorm twendz pro service is a Twitter monitoring and analytics web
application. It enables the user to easily measure the impact of a specific message
within the key audiences.
It uses a keyword-based approach to determine general emotion. Meaningful words
are compared against a dictionary of thousands of words which are associated with
positive or negative emotion.
Each word has a specific score, combined with the other scored words it results in an
educated guess at the overall emotion.

16. Newspaper
Newssift
[22] Newssift is a sentiment search tool on Newspapers and a product from Financial
Times. It indexes content from major news and business sources. The query, for
example brands, legal risks and environmental impact, is matched in regards to the
business topics. This gives you information about changing issues across time for a
company or product.
Applications
LingPipe
[23] "LingPipe is a state-of-the-art suite of natural language processing tools written
in Java that performs tokenization, sentence detection, named entity detection,
coreference resolution, classification, clustering, part-of-speech tagging, general
chunking, fuzzy dictionary matching. These general tools support a range of
applications."
The idea on how sentiment analysis is done using LingPipe's language classification
framework is to make two classification tasks:
• separating subjective from objective sentences
• separating positive from negative reviews
A tutorial is online at their website [23] which describes how to use LingPipe for
sentiment analysis.
Radian6
[24] Radian6 is a commercial social media monitoring application. It has much
functionality, like working with dashboards, widgets. Radian6 gathers from blogs,
comments, multimedia and forums and communities like Twitter the discussions and
opinions and gives businesses the ability to analyze, manage, track and report on their
social media engagement and monitoring efforts.
RapidMiner
[25] RapidMiner is an open-source system, at least the Community Editon, for data
mining and machine learning. It is available as a stand-alone application for data
analysis and as a data-mining engine for the integration into own products. Sentiment
Analysis is also supported. It is used for both real-world data mining and in the
research area.

17. Web Opinion Mining 17
References / Further Readings
1. Liu, Bing, Mining Opinion Features in Customer Reviews, Department of
Computer Science, University of Illinois at Chicago
2. Liu, Bing, Mining and Summarizing Opinions on the Web, Department of
Computer Science, University of Illinois at Chicago
3. Liu, Bing, From Web Content Mining to Natural Language Processing,
Department of Computer Science, University of Illinois Chicago
4. Liu, Bing, Mining and Searching Opinions in User-Generated Contents,
Department of Computer Science, University of Illinois Chicago
5. Hu, Minquing, Liu, Bing, Mining and Summarizing Customer Reviews,
Department of Computer Science, University of Illinois Chicago
6. Ding, Xiaowen, Liu, Bing, Zhang, Lei, Entity Discovery and Assignment for
Opinion Mining Applicatinos, Department of Computer Science, University
of Illinois Chicago
7. Liu, Bing, Opinion Mining, Department of Computer Science, University of
Illinois Chicago
8. Liu, Bing, Opinion Mining and Search, Department of Computer Science,
University of Illinois Chicago
9. Ding, Xiaowen, Liu, Bing, Yu, Philip S., A Holistic Lexicon-Based
Approach to Opinion Mining, Department of Computer Science, University
of Illinois Chicago
10. Liu, Bing, Opinion Mining & Summerazation – Sentment Analysis,
Department of Computer Science, University of Illinois Chicago
11. Jindal, Nitin, Liu, Bing, Opinion Spam and Analysis, Department of
Computer Science, University of Illinois Chicago
12. Liu, Bing, Web Content Mining, Department of Computer Science,
University of Illinois Chicago
13. Liu, Bing, Hu, Minqing, Cheng, Junsheng, Opinion Observer: Analyzing and
Comparing Opinions on the Web, Department of Computer Science,
University of Illinois Chicago
14. Liu, Bing Web Data Mining – Exploring Hyperlinks, Contents and Usage
Data – Lecture Sides, Springer, Dec. 2006
15. Evri, Semantic Web Search Engine; [cited 2010 Jan 19].
<http://www.evri.com/>.
16. Evri, Widget Sentiment Analysis Example on “Android”; [cited 2010 Jan
19].
<http://www.evri.com/widget_gallery/single_subject?widget=sentiment&ent
ity_uri=/product/android-0xf14fe&entity_name=Android>.
17. OpenDover, Sentiment Analysis Webservice; [cited 2010 Jan 19].
<http://www.opendover.nl/>.
18. RankSpeed, Sentiment Analysis on Blogosphere and Twittersphere; [cited
2010 Jan 19].
<http://www.rankspeed.com/>.

18. 18 Dupré Marc-Antoine, Alexander Patronas, Erhard Dinhobl, Ksenija Ivekovic,
Martin Trenkwalder
19. Twittratr; [cited 2010 Jan 19]. <http://twitrratr.com/>.
20. Twitter Sentiment, a sentiment analysis tool; [cited 2010 Jan 19].
<http://twittersentiment.appspot.com/>.
21. WE twendz pro service, influence analytics for twitter; [cited 2010 Jan 19].
<https://wexview.waggeneredstrom.com/twendzpro/default.aspx>.
22. Newssift, sentiment analysis based on Newspapers; [cited 2010 Jan 19].
<http://www.newssift.com/>.
23. LingPipe, Java libraries for the linguistic analysis of human language; [cited
2010 Jan 19].
<http://alias-i.com/lingpipe/index.html>.
24. Radian6, social media monitoring and engagement; [cited 2010 Jan 19].
<http://www.radian6.com/>.
25. Sysomos, Business Intellegence for Social Media; [cited 2010 Jan 19].
<http://sysomos.com/>.
26. RapidMiner, environment for machine learning and data mining
experiments; [cited 2010 Jan 19]. <http://rapid-i.com/>.
27. Go, Alec, Bhayani, Richa, Huang, Lei, Twitter Sentiment Classifciation
using Distand Supervision, Stanford University; [cited 2010 Jan 19].
Available From:
<http://www.stanford.edu/~alecmgo/papers/TwitterDistantSupervision09.pdf
>.
28. Turney, P. Thumbs Up or Thumbs Down? Semantic Orientation Applied to
Unsupervised Classification of Reviews. In Proc. of the 15th Intl. Conf. on
World Wide Web (WWW'06), 2006.
29. Liu, Bing Web Data Mining – Exploring Hyperlinks, Contents and Usage
Data, Springer, 2007.
30. Santorini, B. Part-of-speech Tagging Guidelines for the Penn Treebank
Project. Technical Report MS-CIS-90-47, Department of Computer and
Information Science, University of Pennsylvania, 1990.
31. Pang, B., Lee, L., Vaithyanathan, S. Thumbs Up? Sentiment Classification
Using Machine Learning Techniques. In Proc. of the EMNLP'02, 2002.
32. Dave, K., Lawrence, S., Pennock, D. Mining the Peanut Gallery : Opinion
Extraction and Semantic Classification of Product Reviews. In WWW'03,
2003.
33. Wiebe, J., Riloff, E. Learning Extraction Patterns for Subjective Expressions.
34. Yu, H., Hatzivassiloglou, V. Towards Answering Opinion Questions:
Separating Facts from Opinions and Identifying the Polarity of Opinion
Sentences. Proceedings of the 2003 Conference on Empirical Methods in
Natural Language Processing, pp. 129-136, 2003.