Natural Language Processing & Information Retrieval

chardfriendlyAI and Robotics

Oct 16, 2013 (4 years and 25 days ago)

108 views






Natural Language Processing &
Information Retrieval


Coursework 1




An investigation into the theory and application of
Sentiment Analysis technology













Student Name

Alan Dent

Student ID

2405942

Course

BSc Computing Studies (year 2)

Unit

Natural Language Processing &
Information Retrival

Tutor

Dave Inman

Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
2

of
22

Table of contents

Abstract:

................................
................................
................................
.........................

3

Section 1. Introduction:

................................
................................
................................
.

3

Section 2. A statistical approach:

................................
................................
...................

5

Section 3. A structural approach:

................................
................................
.................

11

Section 4. Linking opinion to source:

................................
................................
..........

13

Section 5. Applications of Sentiment Analysis:

................................
..........................

14

Section 6. Moral Issues:

................................
................................
...............................

17

References:

................................
................................
................................
...................

19

Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
3

of
22


Abstract:


Sentiment Analysis is an area of study within
Natural Language Processing that is concerned with
identifying the mood or opinion of subjective

elements within a document. With an increasingly
large body of written material available in electronic
format, it is becoming ever more difficult to assess
and make use of this abundance of material. This
paper is intended to establish the rationale be
hind
the study of Sentiment Analysis, introduce some of
the problems that need to be solved in order to
undertake Sentiment Analysis, review some of the
approaches currently being pursued in the topic and
raise some of the moral issues that actual
applicat
ions of Sentiment Analysis might bring
about.


Section 1. Introduction:

It is difficult to find accurate statistics about the number of documents available on
the Internet, and even search engines such as Google are reluctant to publish
information on the

size of their indexes [1], however it seems reasonable to assume
that the number is very large. Beyond the Internet, there are also large numbers of
electronic documents in many forms held by businesses and official bodies. Much of
this material is in t
he form of unstructured documents using natural language.
Material such as e
-
mails, blogs, wikkis, chat rooms, forums and other such informal
documents contain a large amount of data, often reflecting opinions and experiences
that would be of value for ma
ny different types of application. In order to make use
of this potential resource, it will be necessary to develop methods of extracting useful
data from the mass of unstructured material.

Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
4

of
22

The amount of documents involved makes manual analysis impractica
l. However
there are a number of problems that a computer based tool would have to overcome
before it could produce useful data. One such problem is the inherent complexity of
everyday language


As Yi and Niblack note [2],

“Opinions in natural language

are often expressed in subtle and complex ways,
presenting challenges which are not easily addressed by standard text
characterization approaches. Negative reviews may contain apparently
positive phrases even while maintaining a strong negative tone, and

the
opposite is also common.”[2]

Perhaps the most obvious area of research in Sentiment analysis is the attempt to
establish the direction of the sentiment, i.e. whether it is positive or negative. This is
known as the polarity, or semantic orientation o
f the text. Some researchers, such as
Bo Pang and Lillian Lee [3] have attempted to further refine this to measure the
strength of the sentiment, rather than just the polarity. A further area of importance is
establishing which parts of the test under an
alysis are relevant for Sentiment Analysis.
Purely objective text, while it might contain words that have a discernable polarity,
would be intended as a statement of fact and not be explicitly representing the writer’s
opinions. Therefore, in order to un
dertake meaningful analysis, it is necessary to
establish the subjectivity of the piece of writing.

It is also important to establish to what an expressed sentiment is referring. This is
known as
source coference resolution,
and a more detailed explanatio
n of the issues it
addresses is given by Stoyanov and Cardie in the introduction to [4]. This sort of
work can be of use in applying analysis to particular problems, especially where
Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
5

of
22

documents contain opinions on multiple subjects (for example, it might b
e reasonable
to assume that the tone of a film review is applicable only to the reviewed film, but a
reviewers blog may refer to the movie, the cinema, the restaurant they visited
afterwards, the state of the traffic or anything that came to mind in the co
urse of the
writing).

The remainder of this paper is organised as follows. Section 2 will outline some of the
statistical approaches used to determine polarity, and section 3 will present some other
approaches that have a lexical basis. Section 4 will exa
mine methods of linking
opinion to source, section 5 will introduce some applications that make use of
Sentiment Analysis and section 6 will raise potential moral issues with the application
of these theories.

Section 2. A statistical approach:

Supervised
Learning


An important concept in understanding the statistical approach
to Sentiment Analysis is that of Supervised Learning. This is a method of machine
learning whereby a set of data with known values (known as training data) is used to
derive a funct
ion that can produce a desired output. The function can then be used to
predict the desired output from any set of valid inputs, not just those within the
training data (for a more complete definition, see Wikkipedia [5]). An example of
how Supervised Le
arning may be applied is given in the glossary of the statistical
software company StatSoft.com[6].

“… you may have a data set that contains information about who from among
a list of customers targeted for a special promotion responded to that offer.
The
purpose of the classification analysis would be to build a model to predict
Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
6

of
22

who (from a different list of new potential customers) is likely to respond to
the same (or a similar) offer in the future.” [6].

Supervised Learning is used in a number of techniq
ues within the field of Sentiment
Analysis, but all such techniques require a set of training data before they can be
usefully put into practice. Kaji and Kitsugawa [7] raise some of the difficulties
inherent in establishing such a training set

“Since ma
nual construction of tagged corpus is time consuming and
expensive, it is difficult to build large polarity
-
tagged corpus. The method that
relies on review sites can not be applied to domains in which a large amount of
reviews are not available. In addit
ion, the corpus created from reviews is often
noisy”. Kaji and Kitsugawa[7]

They are suggesting that the typical methods for establishing sets of training data are
either by manually establishing the polarity of the input data (i.e. by taking words or
phr
ases and deciding a polarity for each using human judgement) or by using bodies
of documents where a quantative rating is applied to the subjective text (they use the
star rating system of Amazon.com as an example). Their proposed solution to this
shortco
ming is to use linguistic patterns to extract opinion phrases, which can be used
as a corpus of training data. It is interesting to note however that they still needed to
construct a manual lexicon of polarity tagged indicator words.

Naïve Bayes classifie
rs


Baysean statistical methods are based on the concept that
statistical likelihood can be linked to observation, for example if I withdraw at
random 10 fruit from a mixed bag and get 5 apples, the likelihood of the next fruit I
take from the bag being a
n apple is 0.5 (i.e. 5/10, the number of times I have got an
Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
7

of
22

apple divided by the total number of fruit I have taken). This is despite the fact that I
do not know what proportions of fruit the bag contains, and so I cannot make an
absolute statistical inf
erence.

The definition of a Naïve Bayes classifier given in Wikkipedia is

“A
naive Bayes classifier

(also known as
Idiot's Bayes
) is a simple
probabilistic classifier based on applying Bayes' theorem with strong (naive)
independence assumptions.” [8]

A co
rpus of training data is used to identify features which are characteristic of a
particular class, and the classifier then uses this information to infer the most likely
class of an item that has not yet been categorised. One important thing to note about

this type of classifier is the strong independence assumptions. This means that in
assessing whether something is a member of a class based on a number of features,
each feature is assessed independently of the others. More technical descriptions of
Naï
ve Bayes classifiers can be found in the technical notes of the online StatSoft
textbook [9], or in section 2 of Peng, Schuurmans and Wang[10].

Naïve Bayes classifiers are commercially available, and can be easily utilised by
researchers in the field of se
ntiment analysis. Peng, Schuurmans and Wang[10]
propose augmenting established Naïve Bayes Classifiers with other statistical
methods, most specifically by relaxing the independence assumptions in order to
allow for chains of dependence in their models.
They refer to this as a Chain
Augmented Naïve Bayes classifier, and give the result of a number of experimental
analyses of large data sets in English, Greek, Chinese and Japanese. They conclude
that relaxing independence assumptions can be valuable where

there is a large training
Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
8

of
22

data set, although they note, “different classifiers are superior for different languages
and tasks”[10].

Support Vector Machines (SVM)


training data could be represented as a set of co
-
ordinates, mapped on an
n
-
dimensional spa
ce, for example linguistic units (words in
English, Chinese characters etc) could each be given a two numerical values based on
two inherent sentiment properties (for example polarity and strength), which could be
expressed as two dimensional co
-
ordinates.

If these co
-
ordinates are plotted on a
graph, and the graph is divided into two classes by a line (a line is one dimensional


this division is an
n
-
1 dimensional hyperplane), if we find the line that is furthest from
both sets of class data (known as th
e maximum margin hyperplane). Any new data
under analysis could then be classified according to which side of the maximum
margin hyperplane it fell. If
n

properties are identified for each linguistic unit, then
they can be expressed as a vector, which re
presents the co
-
ordinates in
n
dimensional
space. For a fuller description of SVM see [11] and for a detailed description of their
use in language classification see [12].

It may be possible to improve the performance of SVM algorithms by combining
them w
ith other methods. This approach was used by Pang and Lee [13], where they
propose a “meta
-
algorithm” based on metric labelling (that is labelling that explicitly
attempts to ensure that similar items receive similar labels). They use this technique
eval
uating film review sites in order to

“…consider finer grained
scales
: rather than just determine whether a review
is “thumbs up” or not, we attempt to infer the authors numerical rating, such as
“three stars” or “four stars”…[13].

Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
9

of
22

Film review sites were

evaluated because the numerical rating is explicit within the
corpus of data being evaluated.

The main difference between [13] and a simple SVM application is that degrees of
scale are established by “… using explicit item and label similarity informatio
n to
increasingly penalize the initial classifier as it assigns more divergent labels to similar
items”[13]. In simple terms, the premise is that the greater the number of positive or
negative terms used, the more positive or negative the document. They
implement
this experimentally by taking a corpus of training data from a movie review site, and
labelling a significant “snippet” (described in the paper as “a striking extract usually
one sentence long”[13]) with the star rating of the review that contain
ed it. They then
trained a Naïve Bayes classifier on the data set, with which they were able to identify
positive sentences. They then calculated a Positive Sentence Percentage (PSP) that
was defined as the number of positive sentences divided by the tot
al number of
subjective sentences. The PSP was then used to place the review being evaluated on a
graduated scale.

I felt that while this work was interesting, the method was not generally applicable in
that it required a training dataset from a corpus of

documents that were already graded
with a sentiment scale (such as movie reviews). There was nothing in [13] that
suggested that scale values could be inferred between document genres (i.e. that there
was a generally applicable scale for the PSP). Also,

only four authors were evaluated
in the work, raising the question as to whether increasing the number of authors would
distort or smooth the results. On a purely practical note, there was no description of
how the “snippet” was selected

Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
10

of
22

One of the drawb
acks of the methods presented above is that, in order to produce
meaningful results, a large corpus of training data is required. This is because they
are supervised techniques, relying on the application of human judgement to attribute
polarity to each w
ord. Esuli and Sebastiani [14] suggest that online glosses
(glossaries and dictionaries) could be used in a semi
-
supervised learning technique,
which would considerably reduce the amount of training data initially required in
order to determine semantic p
olarity. Their method relies on the definition of a term
within a gloss containing terms with a similar semantic orientation, as they put it “…
the glosses of
honest
and
intrepid

will both contain appreciative glosses, while the
glosses of
disturbing
and
superfluous

will both contain derogative expressions” [14].
The method is semi
-
supervised in that once the semantic polarity of a particular term
has been established, it is added to the training database, thus from a relatively small
“seed” database resu
lts which are either equivalent or superior to those of techniques
using pure supervised learning methods can be achieved.

The application of this idea relies on the ability to extract information from online
glosses in a format that can be processed autom
atically. For this reason, they have
used the WordNet [15] lexicon for their experiments, as it has a structure that
indicates the semantic relationships between words (e.g. Synonyms, Antonyms etc.),
although they also show how the technique could be used

(albeit with less accurate
results) on another online dictionary.

I felt that the approach taken in [14] seemed to promise the most practical application,
for two reasons. Firstly, I felt that the small size of the training data required for the
“seed” c
orpus meant that it could more easily be applied to a particular application,
and that its performance should improve as the application was used. Secondly, I felt
Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
11

of
22

that the method could be extended most easily, for example by using multiple glosses
and co
mparing results in order to achieve greater accuracy and better filtering of noise
inherent in the data.

Section 3. A structural approach:

At first glance I found the work on contextual polarity by Wilson, Wiebe and
Hoffman [16], to be similar to the stati
stical approaches in the previous section, in that
it initially requires a corpus of polarity tagged semantic units (words or phrases) to be
manually created. The words in the document to be analysed by their technique can
then be tagged with their semant
ic polarity. It is the way in which this information is
applied to establish polarity that distinguishes it from the statistical methods in section
2 above. In order to overcome the problem that the strength and polarity of a word or
expression is contex
t dependant, they propose a method of parsing each sentence to
produce a ‘dependency tree’ that can be used to establish the polarity of each part of
the sentence and how that affects the polarity of the next highest level of the tree
structure. Figure 1.

below shows the dependency tree they obtained for the sentence
“The human rights report poses a substantial challenge to the US interpretation of
good and evil”


Figure 1.(taken from [16])

Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
12

of
22

The structure of this approach bears a marked similarity to that
of a rules based
machine transaction, although the function of the parse tree is to establish semantic
polarity rather than syntactic structure. Whilst the method seems very impressive in
the paper [16], I wonder whether it would be subject to similar sho
rtcomings in
dealing with ambiguous polarity as with ambiguous meanings, and whether deeper
trees would prove to be computationally intractable.

The approach taken by Kanayama, Nasukawa and Watanabe [17] even more directly
references machine translation.
They see the objective of sentiment analysis as
analogous to that of translation, except instead of translation to a target language its
objective is translation to sentiment units. Their work concentrates on the analysis of
documents in Japanese, but the
y believe that the technique should be transferable to
any language. Essentially they propose using an existing machine translation system
and “replacing translation patterns and bilingual lexicons with sentiment patterns and
a sentiment polarity lexicon”

[17]. Rather than a binary positive/ negative scheme,
they use four possible tags, favourable, unfavourable, question and request, as they
are assigning tags to all words not just those identified as having semantic polarity
values.

They note that full s
yntactic parsing is necessary in obtaining the correct sentiment, in
order to establish the correct local structures (they give the example “expressions such
as “I don’t think that X is good” and “I hope that X is good” are not favourable
opinions about X,

even though “X is good” appears on the surface” [17]).

The use of existing Machine Translation technology allows the efficient development
of a practical application. However I think that it would be subject to all the
Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
13

of
22

limitations of Machine Translation,

and thus while it may have practical applications
it may be subject to technical limitations dealing with ambiguity and the depth of the
parse tree on more complex syntactic structures.

A broader approach is taken by Argamon et al [18], who are considerin
g a broader
range of semantic functions than simply semantic polarity. While their work is
broader in scope than the other papers introduced here, it is less concerned with the
application of the lexicons, and aims to produce documents containing specific

tags
for lexical items which they define as either words or multi word phrases [18]. Unlike
the approach taken in [17], they believe that locally computable properties are
sufficient in order to extract the desired semantic values.

They use an approach t
o linguistic analysis called Systemic Functional Grammar,
which models words and phrases in terms of their semantic functions. This approach
allows them to evaluate documents based on the lexical and structural features they
exhibit (for more detail see s
ection 3 of [18]).

Section 4. Linking opinion to source:

In establishing the sentiment or semantic polarity of a document, if a finer grained
resolution is required than just that of general tone (i.e. if it is necessary to establish to
what the sentimen
t is attributed. In the case of something like a movie review, it may
be reasonable to assume the subject of the sentiment is the subject of the document,
however establishing the sentiment of a news report or newspaper editorial the subject
may be far le
ss obvious).

Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
14

of
22

The aim of source coference resolution is to establish whether the entity referred to in
a particular Noun Phrase (NP) is the same entity that is the subject of that particular
piece of writing. The problem is particularly apparent in the ana
lysis of English
sentences that contain personal pronouns, and Stoyanov and Cardie [4] have produced
a paper suggesting how NP coference resolution could be extended to tag each NP to
indicate whether or not the NP relates to the opinion source or not.

The
y suggest that there is a three
-
stage process in establishing source coference,
firstly using NLP tools to identify the NP components in a document, secondly
mapping each NP to its source using heuristic methods (they apply a series of rules
that seem to w
ork with the bulk of their data). Once the NPs were mapped to the
sources, they used the corpus under analysis as training data and produced a vector of
57 features which they then used to train a classifier in order to predict whether a
source NP pair we
re actually linked.

Section 5. Applications of Sentiment Analysis:

In this section I aim to introduce some practical applications of the theories of
sentiment analysis as presented above. These will be covered in somewhat less detail
as the objective of t
his section is to illustrate the range of applications that already
exist for the purpose of sentiment analysis. Not all of the work below represents
actually available applications, although several of the papers do cover commercially
available and opera
tional systems, but some of the proposed applications give an
indication of the possible uses for Sentiment Analysis technology.

A large potential commercial application for Sentiment Analysis is in the extraction of
marketing information from the Internet
. Glance et al [19] suggest that there “… is a
Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
15

of
22

wide range of commentary about consumer products. This presents an opportunity to
understand and respond to the consumer by analyzing this unsolicited feedback.” [20].
They present a commercial system (inte
lliseek) that is designed to extract and analyse
this type of information, in a way that is practical for the market.


figure 2. [19]

Figure 2 (above) shows the architecture of their system. The software crawls the web,
extracting data from specific sour
ces such as web logs, message boards and Usenet
forums, which they identify as being the richest sources of consumer data. This data
can then be searched for specific information, and analysed to identify opinions and
trends related to specific products.


The growth in the use of Web logs (or blogs) provides a huge potential source of the
opinions of a wide range of people. These documents are used as online diaries,
providing forums for the expression of opinions and emotions on with few apparent
constr
aints. Mishne [20] presents an initial exploration as to the possibility of using
these as a source of information, suggesting several research areas where the informal
nature of the blog would provide particularly useful information, “such as assisting
b
ehavioural scientists and improving doctor
-
patient interaction.”[20]. He also
identifies several areas which may prove problematic in using blog information, such
Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
16

of
22

as the use of certain characters for emphasis (for example capitalising or
starting/ending w
ords with asterisks to indicate stressed meaning), or the use of
emoticons such as :
-
) to indicate mood.

Mullen and Malouf [21] suggest that the analysis of political opinion would be
another domain for which Sentiment Analysis could be applied with a numb
er of
purposes. They suggest that it could be used to “augment polling data”[21] by
analysing trends in online discussions, it could be used by parties to “target
advertising and communications”[21] both for political campaigning and fund raising
and coul
d be used for identifying bias in apparently unbiased political texts. They
support their theories with an experiment classifying a body of American political
discussion board posts using a Naïve Bayes classifier trained to identify the political
affiliat
ion of the posters. They conclude that the sensitivity of the classification is
influenced by the size of the training corpus, and the number of documents posted by
the writer.

A more detailed description of the architecture of a commercial Sentiment Anal
ysis
product is provided by Yi and Niblack [2], for the IBM WebFountain system.
Components suggested by work in the previous sections (such as a Sentiment Term
dictionary, and the NLP technology to break down sentences and identify Parts of
Speech) are in
tegrated in their Sentiment Miner [2], see figure 3 below. This shows
how information extracted from a number of sources from the Internet and stored in a
data
-
store could be analysed to extract sentiment information. The left side of the
diagram shows t
he NLP tools for breaking down documents and tagging their
syntactic structure. The tagged data is then analysed, in the central area of the
diagram, undertaking source coference analysis within the POS tagging and Feature
Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
17

of
22

Extractor at the bottom of the d
iagram, and using various glosses in order to perform
Sentiment Analysis.


Figure 3. Sentiment Miner in IBM WebFountain[2]

In their paper Yi and Niblack make some impressive claims for the Sentiment Miner
[2], but it should be no
ted that they are employees of the company that is marketing it
as a commercial product, and care should be taken with some of the more extravagant
claims (for example, the general claim of achieving “(~90% accuracy) on various
datasets including online re
view articles and general web pages and news articles.”[2].
However, it does illustrate how components based on research such as that discussed
in the preceding sections, can be integrated to achieve a commercially viable product.

Section 6. Moral Issues:

The preceding sections have concentrated on the academic and practical issues
surrounding the subject of Sentiment Analysis. It is also important to understand that
there is potential, as there is with many technologies, for misuse and abuse. There are
Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
18

of
22

implications for privacy, finance and freedom of speech as well as the positive
opportunities stressed in most of the academic papers.

There is increasing pressure on companies to monitor their staff, for reasons of
compliance and for internal management p
urposes. Sentiment Analysis technology
could be utilised for companies in order to more closely monitor the tone and content
of documents produced by their employees. Whilst this may be regarded as
reasonable practice, the case of a worker sacked for pos
ting negative comments on a
personal blog [22], shows the potential for companies to intrude beyond the scope of
the employees working life.

There is concern in America that work done on Sentiment Analysis may be misused,
especially by the security service
s. In an article on the political news and commentary
site Swans.com, the work of Lillian Lee and Claire Cardie of Cornell University (see
papers [3], [4], [13], [24], [25]) is described as “this "creepy and Orwellian" program,
as it is characterized by L
ucy Dalglish of the Reporters Committee for Freedom of the
Press”[23] a somewhat melodramatic description, although the article highlights other
concerns that the work may be misused by the US administration.

Malouf and Mullen [21] recognise the risk of
misuse in their own work in a footnote
where they comment “It is worth commenting that methods of political sentiment
analysis may also lend themselves to potentially abusive applications… While this is
regrettable the authors believe that the responsibili
ty for the protection of individual
rights lies with an accountable and transparent government answerable to the rule of
law.”[21].

Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
19

of
22

Another area where care is needed in the application of Sentiment Analysis is that of
the financial markets. If this techno
logy is used in order to improve the information
provided for investment decisions it is important that the sources of data remain
clearly identifiable. The following quote from a marketing company web page
illustrates an existing technique for manipulati
ng informal internet channels for
marketing purposes that relies on human interaction with the manipulated source

Viral Marketing


Chats, forums, newsgroups… Internet leaves tremendous room
for discussions, where information circulates, reputations are m
ade or unmade,
opinions are created.” [24]. How much more powerful and potentially misleading
would it be if the manipulated channels were being automatically analysed and
summary information being used as a basis for investment decisions.

Despite these p
otential drawbacks, I believe that it would be a mistake to suggest
limits or otherwise curtail research in the domain of Sentiment Analysis. Even
assuming that such limit or curtailment were practically possible, which clearly it is
not, I feel that it i
s important to separate the implications of the application of
knowledge from the acquisition of that knowledge. I think that the moral implications
of Sentiment Analysis should be evaluated with a full appreciation of this technology,
but if the academic

community does not take the opportunity to initiate discussions, it
will be left to those with less understanding to raise issues in a more sensationalist
manner (as in [23]).

References:

[1]


w
ww.google.com/help/indexsize.html

-

accessed 20th October 2006.

[2]

J. Yi and W. Niblack. 2005.
Sentiment Mining in WebFountain.

ICDE 2005.
Proceedings of 21st International Conference on Data Engineering. Pages 1073
Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
20

of
22



㄰㠳1

x㍝

B⸠K~ng⁡湤ni⸠iee⸠㈰
〵⸠
Seeing Stars: Exploiting class relationships for
sentiment categorization with respect to rating scales.
Proceedings of the 43
rd

Annual meeting of the ACL. Pages 115


ㄲ㐮

x㑝

嘮⁓soy~湯瘠n湤⁃⸠K~牤楥⸠㈰〶⸠
Toward Opinion Summarization: Linking the

Sources.

Association for Computational Linguistics, proceedings of the
Workshop on Sentiment and Subjectivity in Text. Pages 9

14.

[5]

http://en.wikipedia.org/wiki/Supervised_learning

-

ac
cessed 22nd October
2006.

[6]

http://www.statsoft.com/textbook/gloss.html

-

accessed 23rd October 2006.

[7]

N. Kaji and M. Kitsuregawa. 2006.
Automatic Construction of Polarity
-
tagged
Corpus fr
om HTML Documents.

Proceedings of the COLING/ ACL 2006.

[8]

http://en.wikipedia.org/wiki/Naive_bayes

-

accessed 23rd October 2006.

[9]

http://www.statsoft.com/textbook/stnaiveb.html

-

accessed 23rd October 2006.

[10]

F. Peng, D. Schuurmans and S. Wang. 2003.
Augmenting Naïve Bayes
Classifiers with Statistical Language Models.

Information Retrieval, Volume 7,
2004. Pages 317


㌴㔮




桴h瀺p⽥渮/楫i灥摩愮潲g⽷楫i⽓異灯u瑟噥c瑯牟䵡c桩湥

-

~cce獳敤′s牤⁏ct潢o爠
㈰〶O

xㄲ1

吮⁊潡c桩h献′〰㈮s
Learning to Classify Text using Support Vector Machines.
Methods, Theory, and

Algorithms.

Kluwer Academic Publishers,

May 2002. 224 pp.

[13]

B. Pang and L. Lee. 2005.
Seeing Stars: Exploiting class relationships for
sentiment categorization with respect to rating scales.
Proceedings of the 43
rd

Meeting of the ACL, 2005. Pages 11
5
-
124.

[14]

A. Esuli and F. Sebastiani. 2005.
Determining the Semantic Orientation of
Terms through Gloss Classification.
Proceedings of the 14th ACM international
conference on Information and knowledge management . Pages 617


㘲㤮

xㄵ1

桴h瀺p⽷潲/湥琮灲楮te瑯渮敤t

-

䅮湬楮攠汥x楣潮Ⱐ獴牵i瑵牥搠瑯⁲dp牥獥湴⁴he
獥浡湴楣⁲e污瑩潮l桩瀠扥h睥e渠睯牤猠


~cce獳敤″
rd

November 2006.

Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
21

of
22

[16]

T. Wilson, J. Wiebe and P. Hoffmann. 2005.
Recognizing Contextual Polarity
in Phrase level Sentiment Analysis.
Proceedings of joint conference on Human
Language Technology / Conference on Empirical Methods in Natural Language
Processing (HLT/EMNLP'05). Pages 347


㌵㐮

xㄷ1

䠮⁋~湡y~m~Ⱐ吮⁎~獵歡睡⁡湤⁈⸠n~瑡t~扥⸠㈰〴⸠
Deeper
Sentiment
Analysis Using Machine Translation Technology.
Proceedings of Coling 2004.
Pages 494
-
500.

[18]

S.Argamon, C. Whitelaw, P. Chase, S. Dhalwe, S. R. Hota, H. Garg and S.
Levitan. 2005.
Stylistic Text Classification Using Functional Lexical Features
.
Linguistic Cognition Lab, Illinois Institute of Technology.

[19]

N. Glance, M. Hurst, K. Nigam, M. Siegler, R. Stockton and T. Tomokiyo.
2005.
Deriving Marketing Intelligence from Online Discussion.
Proceeding of
the eleventh ACM SIGKDD international co
nference on Knowledge discovery
in data mining. Pages 419


㐲㠮

x㈰O

䜮⁍楳桮攮′〰㔮d
Experiments with Mood Classification in Blog Posts.

Informatics Institute, University of Amsterdam.

[21]

T. Mullen and R. Malouf. 2006.
A Preliminary Investigation int
o Sentiment
Analysis of Informal Political Discourse.
Proceedings of the AAAI
-
2006
Spring Symposium on "Computational Approaches to Analyzing Weblogs".

Pages 159


ㄶ㈮

x㈲O

桴h
瀺p⽢潯歳⹧畡牤楡渮n漮畫⽮敷猯慲瑩c汥猯〬ⰱ㌸l㈹〬〰⹨瑭l

-

乥睳⁡牴楣汥l
牥灯牴楮g渠 ⁷ 牫r爠摩rc楰i楮敤⁦潲⁰o獴s湧 c潭oe湴猠潮⁨楳⁷ 戠汯b


~cce獳敤‶
th

November 2006.

[23]

http://www.sw
ans.com/library/art12/jeb162.html

-

accessed 25th October 2004.

[24]

http://www.visiplus.com/gb/web
-
guerilla
-
marketing.html

-

accessed 6th
November 2004.

[25]

T. Wilson, P. Hoffmann
, S. Somasundaran, J. Kessler, J. Wiebe, Y. Choi, C.
Cardie, E. Riloff and S. Patwardhan. 2005.
OpinonFinder: A system for
subjectivity analysis.

Proceedings of HLT/EMNLP 2005 Demonstration
Abstracts. Pages 34

35.

[26]

L. Lee. 2004.
A Matter of Opinion: S
entiment and Business Intelligence
(position paper).

Presented at the IBM Faculty Summit on the Architecture of
Business on Demand, May 2004.

Alan Dent

Student no: 2405942

NLP & IR Coursework 1

Page
22

of
22

[27]

B. Pang and L. Lee. 2004.
A Sentimental Education: Sentiment Analysis Using
Subjectivity Summarization Base
d on Minimum Cuts.
Proceedings of the ACL
2004, Main Volume, pages 271
-
278.

[28]

Y. Seki, K. Eguchi, N. Kando and M. Aono. 2005.
Multi
-
Document
Summarization Reflecting Information Needs on Subjectiviy.
Proceedings of
NTCIR
-
5 Workshop Meeting, December 20
05.

[29]

M. Vermeij. 2005.
The Orientation of User Opinions Through Adverbs, Verbs
and Nouns.

3rd Twente Student Conference on IT , Enschede June, 2005.

[30]

M. Gamon. 2004.
Sentiment classification on customer feedback data: noisy
data, large feature ve
ctors, and the role of linguistic analysis.
Proceedings of
Coling 2004. Pages 841


㠴㜮

x㌱P

䬮⁁桭慤Ⱐh⹔K獫~y~
-
Te浺楥氬⁄⸠䍨i湧Ⱐi⸠䝩汬~洬⁓⸠䅨浡搬⁈⸠d牡扯畬獩Ⱐ
g⸠乡湫n牶楳⸠㈰〴⸠
Financial Information Grid


an ESRC e
-
Social Science
Pilot.
Proce
edings of the Third UK e
-
Science Programme All Hands Meeting
(AHM 2004).

[32]

J. Yi, T. Nasukawa, R. Bunescu and W. NiBlack. 2003.
Sentiment Analyzer:
Extracting Sentiments about a Given Topic using Natural Language Processing
Techniques.
Proceedings of t
he third IEEE international conference on Data
Mining (ICDM’03). Pages 427


㐳㐮