Aspect Miner: Fine-grained, feature-level opinion mining from rated review corpora

Aspect MinerFine-grained feature level opinion mining

from rated review corpora

Stelios KarabasakisDept. of Informatics and Telecommunications National and Kapodistrian University of Athens

in association with the Knowledge Discovery in Databases Laboratory kddlab.di.uoa.gr

MSc Thesis Defense | February 2012

Stelios Karabasakis Feb 2012Aspect Miner: Fine-grained feature-level opinion mining from rated review corpora 2

INTRODUCTION

What is it? The task of recognizing and classifying the opinions and sentiments expressed in unstructured text.

Opinion Mining: an overview

Use cases• product comparison• opinion summarization• opinion-aware recommendation systems• opinion-aware online advertising• reputation management• business intelligence• government intelligence

Opinion sources• news• blogs• reviews• user comments• social networks• forums• discussion groups

Our focus in this work


INTRODUCTION

Reviews

movies books

goods services

restaurantshotels

• Popular form of user generated content» consumers use them to

make informed choices» businesses use them to

gauge and monitor consumer sentiment

• Covering many distinct domains, such as…


INTRODUCTION

Ratings

• Every online review typically carries a rating» picked by the review author» summarizes the sentiment of

the text

• Corpora of rated reviews are» abundant on the web» potentially useful for

supervised opinion mining» largely ignored in the literature!


INTRODUCTION

Not as simple as counting positive vs. negative words

It is pointless to discuss why Hitchcock was a genius.

Distinct opinions about different topics in the same sentenceThe top-notch production values are not enough to distract from a clichéd story that lacks heart and soul.

Semantics of subjective expressions are domain-dependentunpredictable plot twist, gloomy atmosphere (movies) unpredictable service quality, gloomy room (hotels)

Opinion Mining is challenging


INTRODUCTION

classification dimensions• subjectivity: factual vs. subjective statements• polarity: positive vs. negative sentiment• intensity: weak vs. strong sentiment

classification granularity• binary• multiclass

Opinion Mining is a text classification problem

Motivating questionHow can we train a system to distinguish among multiple degrees of sentiment?

?


INTRODUCTION

In “Game of Thrones” (2011), the transition from book to screen is remarkably successful. The carefully chosen location and cast, the top-notch cinematography and the seamless-ness of its narrative come together brilliantly. The new HBO show offers compelling drama, even when rehashing old fantasy themes.

Classification levels

positive

document level


INTRODUCTION



positive

positive

positive

sentence level


INTRODUCTION


adaptation: positive

production: positivecast: positivedirection: positiveplot: positive

serialization: positivesubject: negative


feature level features = domain-specific ratable properties

Motivating questionHow can we identify feature terms and the features they refer to?

?


INTRODUCTION

Produce rich, fine-grained, feature-oriented review summaries

by analyzing reviews at the sentence level and aggregating the results

Problem description

“Avatar” (2009) aggregated summary of 90 reviews

aspect mentions sentiment mean sentiment dispersiondirectio

n 217 9/10 STRONGLY POSITIVE

17%

UNANIMOUS AGREEMENT

story 152 8/10 POSITIVE32%

GENERAL AGREEMENT

acting 177 4/10 WEAKLY NEGATIVE

56%

MIXED REACTION

Sample summary


INTRODUCTION

a sentiment lexiconmulticlass and adapted to the target domain

Solution components

term prior sentiment _masterpiece 10 (very strongly positive)good 8 (positive)mediocre 5 (very weakly negative)terrible 2 (strongly negative)

feature term featureprotagonist CASTperformance CASTdeliver CASTcamera DIRECTIONcinematography DIRECTIONdialogue WRITINGscript WRITING

a feature lexiconfor the target domain

and a set of linguistic rules for sentence classification


INTRODUCTION

The Aspect Miner system

Training subsystem

Term classifier

Feature identifier

Sentence classifier

Lexical Analyzer

Index of terms

Sentiment lexicon

Feature lexicon

Result: Feature-level sentiments

Training corpus(rated reviews)

Text to classify

(a proof-of concept implementation of our approach)

Key features: modular architecture, unsupervised,domain agnostic, configurable granularity


INTRODUCTION

Aspect Miner implementation*

• Implemented in Java with» NekoHTML for scraping» JDBC/MySQL for dataset storage» Lucene as a lexical analysis API and for indexing» Wordnet & JWNL for lemmatization» Stanford Parser for POS-tagging & typed dependency parsing» Mallet’s LDA implementation for topic modeling» GraphViz for visualizations

* source code (MIT-licensed) available from github.com/skarabasakis/ImdbAspectMiner


INTRODUCTION

107.646 movie reviews from IMDB.com, rated 1-10 stars

Training dataset*

*available as an SQL dump from http://db.tt/vAthzJRL

review length (words)

# re

view

s

mean = 291 wordsmedian = 228 words

Sentiment Lexicon ConstructionDesigning a fine-grained term classifier


SENTIMENT LEXICON

A term is a (base form, part of speech) tuple

» part of speech {VERB, NOUN, ADJECTIVE, ADVERB}» a term represents all inflected forms and spellings of a word

e.g. {choose, chooses, chose, chosen, …} [choose VERB]

{localise, localize, …} [localize VERB]

» terms can be compounde.g. [work out VERB] [common sense NOUN]

[meet up with VERB] [as a matter of fact ADVERB]

Terms


SENTIMENT LEXICON

Purpose: to extract terms from texts

» Identifies the base form of words & compounds• Uses Wordnet to look up base forms

» Eliminates non-subjective words• Stop words including very common terms (be,have,…)• Named Entities (i.e. proper nouns)• all articles, pronouns, prepositions etc.

» Eliminates words that would be misleading for sentiment classification• Comparatives & superlatives• Words within a negation scope

Lexical analyzerTraining corpus(rated reviews)

Le

xica

l An

aly

zer

Tokenization

POS tagging

Named Entity identification

Lemmatization

Comparatives annotation

Negation scope resolution

Stop word removal

Bags of terms(one per document)

Open-class word filtering


SENTIMENT LEXICON

Lexical analysis exampleThe most dramatic moment in the Sixth Sense does not occur until the

final minutes and the jaw dropping twist Shyamalan has been building up to.

Eliminate

Lemmatize

Get indexable terms


SENTIMENT LEXICON

Lexicon-based approach• Prior sentiment inferred from lexical associations

(synonyms, antonyms, hypernyms etc.) in a dictionary• High accuracy, limited coverage• Notable example: Sentiwordnet (Esuli & Sebastiani 2006)

Corpus-based approach• Prior sentiment inferred from correlation patterns

(and, or, either…or, but etc.) in a training corpus• Extended coverage, lower accuracy• Notable examples: Hatzivassiloglou & McKeown 1997, Turney & Littman

2003, Popescu & Etzioni 2005, Ding Liu & Yu 2008

Previous approaches to term classification


SENTIMENT LEXICON

• Requires a training set of rated reviews

• Prior sentiment inferred from the distribution of ratings among all the reviews where a term occurs, i.e. the rating histogram of the term

Ratings-based term classification

Our proposal: a ratings-based approachpositive term negative term

neutral term polysemous term


SENTIMENT LEXICON

IMDB dataset: Ratings distribution

rating

# reviews # terms

# reviews# terms

Caution: Ratings are not evenly distributed across the training corpus.


SENTIMENT LEXICON

Why? Weighting is necessary to» eliminate training set biases» make rating frequencies comparable to each other

How? Multiply every rating frequency in a histogram with that rating’s weight wr , calculated as follows:

» cr := cumulative term count of all reviews with rating r

» We pick wr in such a way that wr∙ cr are equal for all r

• Most predominant rating in the dataset has wr =1

• The less frequent the rating, the higher its weight

Rating frequency weighting


SENTIMENT LEXICON

Some sample histogramsextracted from the IMDB dataset


SENTIMENT LEXICON

input: weighted rating histogram for term output: one or more* sets of significant ratings

RC

* if term is polysemous

Designing a term classifier

RC

A weighted mean function can condense RC into a single rating.

71079

8107759

C

C

Rr r

Rr r

C f

rfr

This rating indicates the term’s sentiment.


SENTIMENT LEXICON

For a term to be neutral, its rating histogram must approximate a uniform distribution

Neutrality criterion

thr

rf

rf

tr

tr

)(max

)(min1

where 0 < thr ≤1

max

min


SENTIMENT LEXICON

Picks the histogram’s peak rating as the only significant rating

Term classification schemes

Pros Simplest classifier possible. Useful as a comparison baseline.Surprisingly capable at classifying polarity (almost 2/3 accurate)

Cons Can’t detect polysemyPoor at classifying intensity

Scheme 1: Peak ClassifierRC


SENTIMENT LEXICON


Pros Better at classifying intensityMakes an attempt at detecting polysemy

Cons Weak terms can be mistaken for polysemous

Scheme 2: Positive/Negative Area Classifier (PN) All ratings above a cutoff

frequency are significant Cutoff frequency should

be set a little bit above the frequency average.

Returns separate sets for positive and negative ratings

)(f*. t r11

)(ft r

RC+

RC−


SENTIMENT LEXICON


Pros Avoids detecting false polysemyAvoids biases exhibited by the other classification schemes

Scheme 3: Widest Window Classifier (WW) Looks for windows of

consecutive significant ratings Ratings are added to windows

from most to least frequent Significant rating windows must

satisfy 2 constraints minimum coverage:

W windows must contain at least 1−(2W)-1 of samples

be as wide as possible Returns as many rating classes

as the windows it detects

RC1

RC2


SENTIMENT LEXICON

Classifier evaluation: Ratings Distribution

PEAK

PN

WW

We classified 33.000 terms that appear ≥5 times in the IMDB dataset.

Conclusion: WW classifier distributes rating classes more evenly

Distribution of primary rating classes for each classifier


SENTIMENT LEXICON

Classifier evaluation: Polarity

We evaluate against a reference lexicon of 5272 terms based on the MPQA and General Inquirer subjectivity lexicons.

Accuracy Precision Recall F1-Score

POSITIVE 55.5% 44.2% 49.2% PEAK 63.6%

NEGATIVE 67.3% 65.3% 66.3%

POSITIVE 62.4% 58.4% 60.4% PN 66.2%

NEGATIVE 68.4% 72.3% 70.3%

POSITIVE 70.4% 86.2% 77.5% WW 70.1%

NEGATIVE 69.6% 60.5% 64.8%

POSITIVE 63.6% 61.3% 62.4% SentiWordnet 73.2%

NEGATIVE 83.6% 48.3% 61.3%

WW is the most accurate of the 3 proposed classifiers

But not as accurate than SentiWordnet

However, WW is more accurate for domain-specific terms


SENTIMENT LEXICON

Classifier evaluation: Intensity

We evaluate against a test set of 443 strong + 323 weak terms based on the General Inquirer subjectivity lexicon.

Using the WW classifier to classify intensity:

78% of strong terms are classified 3 and above

83% of weak terms are classified 3 and below

0.0%

10.0%

20.0%

30.0%

40.0%

1 2 3 4 5

Τιμή Έντασης WW

Πο

σο

στό

όρ

ων

WEAK STRONG

Intensity class in WW lexicon

% te

rms

in W

W le

xico

n


SENTIMENT LEXICON

A reusable sentiment lexicon for the movie review domain* downloadable from

github.com/skarabasakis/ImdbAspectMiner/blob/master/imdb_sentiment_lexicon.xls

The Aspect Miner sentiment lexicon*

Feature IdentificationUsing topic models for feature discovery


FEATURE IDENTIFICATION

The traditional approach: discovery through heuristics• frequency: commonly occurring noun phrases are often features

(Hu & Liu 2004)

• co-occurrence: terms commonly found near subjective expressions may be features (Kim & Hovy 2006, Qiu et al. 2011)

• language patterns: in phrases such as 'F of P' or 'P has F‘, P is a product and F is a feature (Popescu & Etzioni 2005)

• background knowledge: user annotations, ontologies, search engine results, Wikipedia data…

An up-and-coming approach: topic modeling

Approaches to feature identification



Probabilistic Topic Models can model the abstract topics that occur in a set of documents

Topic Modeling

documents are mixtures of topics

topics are distributions over words



Probabilistic topic models• require that the user specifies a number of topics

» Topics are just numbers – their semantic interpretation is not the model’s concern

• make an assumption about the probability distribution of topics• define a probabilistic procedure for generating documents from topics

» by inverting this procedure, we can infer topics from documents

A popular topic model: Latent Dirichlet Allocation (LDA)• assumes that topics follow a Dirichlet prior distribution

» i.e. each document is associated with just a small number of topics

Topic Modeling



CARCHASESHOOTVEHICLECOPDRIVEKILLSTREETBULLETROBBERY

WARHEROATTACKGROUPAIRPLANEBUNCHSOLDIERKILLBOMBENEMY

POLICECASEMYSTERYVICTIMSOLVEMURDEROFFICERSUSPECTDETECTIVECRIME

Here are a few sample topics we got from running LDA on the IMDB dataset

Topics vs. Features

Motivating questionFeatures are a form of topics. Can we use topic models to discover features?

?

WARHEROATTACKGROUPAIRPLANEBUNCHSOLDIERKILLBOMBENEMY

POLICECASEMYSTERYVICTIMSOLVEMURDEROFFICERSUSPECTDETECTIVECRIME

ROLEACTORPERFORMANCEPLAYLEADCAST SUPPORTACTRESS SHINESTAR

SCRIPTIDEADIALOGUEWRITEPLOTSCREENPLAYCOME UPCRAFTEXPLAINHOLE

CARCHASESHOOTVEHICLECOPDRIVEKILLSTREETBULLETROBBERY

These topics are features.They are useful to us

These topics are themes.We are not interested in them

ROLEACTORPERFORMANCEPLAYLEADCAST SUPPORTACTRESS SHINESTAR

SCRIPTIDEADIALOGUEWRITEPLOTSCREENPLAYCOME UPCRAFTEXPLAINHOLE



Problem. Topics are global, features are local

Solution. Train topic model on shorter segments (e.g. sentences) rather than full documents.

Problem. Running LDA on such short segments produces noisy topics

Solution. Implement a bootstrap aggregation scheme to filter the noise:

1. Train N topic models from different subsets of dataset

2. Merge similar topics across models to produce a single meta-model» Intuition: Valid feature-topics should occur in >1 models and share many common top

terms. Noisy topics should be isolated to specific models

Feature identification with LDA



Topic Similarity for topics Tm, Tn

» More common terms with higher probabilities higher similarity

Merging topics

COMEDY 0.200JOKE 0.099LAUGH 0.096FUN 0.088FORMULA 0.025

COMEDY 0.180PARODY 0.168SATIRE 0.099JOKE 0.061RIDICULE 0.054

COMEDY 0.380PARODY 0.168JOKE 0.160LAUGH 0.096SATIRE 0.099

+ =

FUN 0.088RIDICULE 0.054FORMULA 0.025

discarded



To merge 2 topic sets • Merge every topic of set A to most similar topic from set B

» but only if that similarity is above average similarity

To merge N topic sets• Merge first two, then merge the result with the third etc.• At the end

» discard topics with a low merging degree» If same term ends up in >1 topics, only keep it in the topic where it

has the highest probability

Merging topic models



Movie feature lexicon

56 topics, manually labeled with 18 labels

Sentence classificationUtilizing language structure for contextual sentiment estimation and feature targeting


SENTENCE-LEVEL ANALYSIS

Sentiment

We define a mapping functionto convert ratings to sentiment classes

(preferably 1:1)

1

2

3

5

6

8

9

10

7

4

-1

+1

mbinary: R10 " S1

1

2

3

5

6

8

9

10

7

4-2

+2

m3: R10 " S3

-3

-1

+1

+3

1

2

3

5

6

8

9

10

7

4

m5: R10 " S5

-5

-4

-3

-1

+1

+3

+4

+5

+2

-2

Sentiment: a (polarity, intensity) tuple, where» polarity {+,−}» intensity {1, 2, …, n} 2n classes



Typed dependencies are binary grammatical relations between word pairs in a sentence(de Marneffe et al., 2006)

Typed Dependencies

Typed dependency trees are • semantically richer than syntax trees• easier to process, because content words are connected directly

rather than through function words

Natalie Portman comes off as very believable, gaining empathy from the audience.

amod(relations, binary)

type governor dependent



Dependency types

Stanford Parser’s representation defines a hierarchy of 48 dependency types



Contextual sentiment estimation

Our model. We empirically developed and formally defined• 6 outcome functions that model types of word interactions• 42 dependency rules that cover all possible dependency patterns

Motivating questionWhat is the contextual sentiment of a dependency, given the prior sentiment of its constituents?

?

Examples

infmod(best/+2, avoid/−4) −4

xcomp(avoid/−4, watching/+2) −2

advmod(disappointing/−2, increasingly/+3) −3

It is best to avoid watching any of the increasingly disappointing sequels.



Outcome functions

Models an interaction whereUNCHANGED base term imposes the sentiment

Ιt seems that they ran out of budget.

STRONGER stronger term imposes the sentiment

a mighty talent wasted in mass produced rom-coms

AVG both terms contribute equally to the sentiment

intelligent and ambitious



Outcome functions

Models an interaction where

INTENSIFY modifier increases the intensity of the base

increasingly disappointing sequels

REFLECT modifier overrides polarity, increases or decreases intensity of base

impossible to enjoy unless you lower your expectations

NEG modifier diminishes or negates the base

not a masterpiece, but not bad either



td(pgov, pdep) outcome_base

Dependency Rules: General form

conj(*,*) AVG_DEP

advmod({n,a,r},*) INTENSIFY_GOV

amod(*,{too}) NEGATIVE_GOV

type label term patterns

A pattern may specify:• a list of allowed parts of

speech• a white list of specific terms

outcome functionone of the following:

UNCHANGED NEGATEDSTRONGER AVGINTENSIFY REFLECTPOSITIVE NEGATIVE

base specifierGOV or DEP

Examples



Aspect Miner dependency rule setgov dep

Td pos wlist pos wlist

outcome base

1. Negation

1.1 neg * * * * NEGATE GOV

1.2.1

1.2.2

1.2.3

1.2.4

1.2.5

1.2.6

det

prt

advmod

dobj

nsubj

dep

* * * negTerms1 NEGATE GOV

1.3 pobj * negTerms1 * * NEGATE DEP

1.4 aux * * * negAux2 NEGATE GOV

1 negTerms = {n't, no, not, never, none, nothing, nobody, noone, nowhere, without, hardly, barely, rarely, seldom, against, minus, sans} 2 negAux = {should, could, would, might, ought}

2. Subjects

2.1.1

2.1.2

nsubj

nsubjpass * * * * INTENSIFY GOV

2.2.1

2.2.2

csubj

csubjpass * * * * REFLECT GOV

3. Objects

3.1.1

3.1.2

dobj

dobj

*

*

negVerbs3

*

*

*

*

*

NEGATE

REFLECT

DEP

GOV

3.2 iobj * * * * UNCHANGED GOV

3.3 pobj * * * * UNCHANGED DEP

3 negVerbs = {avoid, cease, decline , forget, fail, miss , neglect, refrain, refuse, stop}

gov dep td

pos wlist pos wlist outcome base

4. Modifiers

4.1.1

4.1.2

advmod

Amod * * * {enough} POSITIVE GOV

4.2.1

4.2.2

advmod

amod * * * {too} NEGATIVE GOV

4.3 advmod v * * * REFLECT GOV

4.4 advmod n,a,r * * * INTENSIFY GOV

4.5 amod * * * * REFLECT GOV

4.6 infmod a * * * REFLECT GOV

4.7 infmod v,n,r * * * INTENSIFY DEP

4.8 a * * * REFLECT DEP

4.9 partmod

v,n,r * * * STRONGER DEP

4.10 quantmod * * * * INTENSIFY GOV

4.11 prt * * * * STRONGER GOV

4.13 prep * * * {like} UNCHANGED GOV

4.12 prep * * * * REFLECT GOV

5. Clausal Modifiers

5.1 advcl a * * * REFLECT DEP

5.2 advcl v,n,r * * * UNCHANGED DEP

5.3 purpcl * * * * UNCHANGED DEP

6. Clausal complements

6.1.1

6.1.2

6.1.3

ccomp

xcomp

acomp

* * * * REFLECT GOV

6.2.1

6.2.2

6.2.3

conj

appos

parataxis

* * * * AVG GOV

6.3 dep * * * * STRONGER DEP



Initialization • Generate dependency tree from sentence• Annotate subjective terms with prior polarities from sentiment lexicon• Annotate feature terms with labels from feature lexicon

Sentiment estimation• Apply closest matching rule to every dependency relation in the tree

» The sentiment of the dependency replaces previous sentiment of the governor node» Dependencies are processed in reverse postfix order (bottom to top and right to left)

Feature targeting• The scope of a feature term is a subtree that contains the term and goes

» all the way down to the leaves» all the way up to the closest clausal dependency

• the sentiment at the root of the subtree gets assigned to the feature

Sentence classification algorithm



Sentence classification example



Test set: Sentence polarity dataset by Pang & Lee, 2002(5331 positive + 5331 negative sentences from movie reviews)

ResultsPolarity classification is accurate for

71.5% of positive sentences76.9% of negative sentences74.2% of all sentences

Analysis of error causes39.0% inaccurate dependency rule28.5% misclassified term (or we picked the wrong sense)21.5% erroneous sentence parsing 8.5% ambiguous sentence 2.5% dependency rules applied in the wrong order

Sentence polarity evaluation



Comparative evaluationReference Method Accuracy

Linguistic methods

Nakagawa, Irui & Kurohashi, 2005 majority voting 62,9%

Ikeda & Takamura, 2008 majority voting with negations 65.8%

Aspect Miner dependency rules 74.2%

Learning based methods

Andreevskaia & Bergler, 2008 naïve bayes 69.0%

Nakagawa, Irui & Kurohashi, 2005 SVM (bag-of-features) 76.4%

Arora, Mayfield et al., 2010 genetic programming 76.9%

Ikeda & Takamura, 2008 SVM (sentence-wise learning with polarity shifting + ngrams)

77.0%

Nakagawa, Irui & Kurohashi, 2005 dependency tree CRFs 77.3%

Conclusion: Our method fares well among linguistic techniques, but does not match the accuracy of learning based methods

ConclusionsPutting it all together


CONCLUSIONS

Training subsystem

Term classifierTraining corpus(rated reviews)

Lexi

cal A

naly

zer

Tokenization

POS tagging

Named Entity identification

Lemmatization

Comparatives annotation

Negation scope resolution

Stop word removal

Sentence classifier

Dependency tree(s)

Dependency parsing

Bags of terms(one per document)

Feature identifier

Index of terms

Corpus statistics collection

Indexing

Dependency Rule set

Open-class word filtering

Term Histogramgeneration

PEAKclassifier

PNclassifier

WWclassifier

Tra

inin

g s

et

pa

rtiti

on

ing

...

...

...

...

...

partition 1

partition 2

partition N-1

partition N

LDA

Sentiment lexicon

TΜ1 TΜ2 TΜΝ-1 TΜΝ...

Topic models

Assisted labeling

Aggregation

Feature lexicon

Text toclassify

Sentence & Feature Classification

Result:Feature-sentiment

pairs


CONCLUSIONS

• We showed the feasibility of granular prior polarity classification using review ratings» and developed a classifier that

achieved at least 70% accuracy on the training dataset

• We suggested a bagging-inspired meta-algorithm for discovering feature topics with LDA

Summary of contributions

• We developed a reusable sentiment lexicon and feature lexicon for the movie review domain

• We created a set of linguistic rules and developed a methodology that is capable fine-grained feature-level classification of sentences» and achieved 74.2% accuracy

for polarity classification on our test dataset.


CONCLUSIONS

Term classification• Assigning a special class to intensifier terms• Per-feature polysemy resolution

Feature identification• Named entities as features• Applying multi-grain topic models for

discovery of local topics, e.g. MG-LDA (Titov & MacDonald, 2008)

Sentence-level classification• Supervised learning of rules.

Replace manually-made set of rules with a set of rules inferred from frequent dependency patterns.

Suggested Improvements

intensifier term


CONCLUSIONS

B. Liu, “Sentiment analysis and subjectivity,” Handbook of Natural Language Processing,, pp. 978–1420085921, 2010.

B. Pang and L. Lee, “Opinion mining and sentiment analysis,” Foundations and Trends in Information Retrieval, vol. 2, no. 1-2, pp. 1–135, 2008.

A. Esuli and F. Sebastiani, “Sentiwordnet: A publicly available lexical resource for opinion mining,” in Proceedings of LREC, 2006, vol. 6, pp. 417–422.

V. Hatzivassiloglou and K. R. McKeown, “Predicting the semantic orientation of adjectives,” in Proceedings of the eighth conference on European chapter of the Association for Computational Linguistics, 1997, pp. 174–181.

P. Turney, M. L. Littman, and others, “Measuring praise and criticism: Inference of semantic orientation from association,” in ACM Transactions on Information Systems (TOIS), 2003.

A. M. Popescu and O. Etzioni, “Extracting product features and opinions from reviews,” in Proceedings of the conference on Human Language Technology and Empirical Methods in Natural Language Processing, 2005, pp. 339–346.

ReferencesFor a complete list of references, see the full report (in greek)

http://j.mp/AspectMinerM. Hu and B. Liu, “Mining and summarizing customer reviews,” in

Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, 2004, pp. 168–177.

X. Ding, B. Liu, and P. S. Yu, “A holistic lexicon-based approach to opinion mining,” in Proceedings of the international conference on Web search and web data mining, 2008, pp. 231–240.

I. Titov and R. McDonald, “Modeling online reviews with multi-grain topic models,” in Proceeding of the 17th international conference on World Wide Web, 2008, pp. 111–120.

T. Nakagawa, K. Inui, and S. Kurohashi, “Dependency tree-based sentiment classification using CRFs with hidden variables,” in Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the Association for Computational Linguistics, 2010, pp. 786–794.

A. Andreevskaia and S. Bergler, “When specialists and generalists work together: Overcoming domain dependence in sentiment tagging,” ACL-08: HLT, 2008.

D. Ikeda and H. Takamura, “Learning to shift the polarity of words for sentiment classification,” Comp.Intelligence, vol. 25, no. 1, pp. 296–303, 2008.

Aspect Miner: Fine-grained, feature-level opinion mining from rated review corpora

Education

introduction opinion

supervised opinion mining

positive cast

rated review corporafeb

transition positive

introductionopinion

online review

review author