SENTIMENTAL

ANALYSIS
UNIT-I
Introduction to Sentiment Analysis Introduction: Sentiment Analysis
Applications - Sentiment Analysis Research - Sentiment Analysis as Mini NLP.
The Problem of Sentiment Analysis: Definition of Opinion - Definition of
Opinion Summary - Affect, Emotion, and Mood - Different Types of Opinions -
Author and Reader Standpoint. Document Sentiment Classification: Supervised
Sentiment Classification - Unsupervised Sentiment Classification - Sentiment
Rating Prediction - Cross-Domain Sentiment Classification - Cross-Language
Sentiment Classification - Emotion Classification of Documents.
Introduction to Sentiment Analysis
Sentiment Analysis Definition: Sentiment analysis, also called opinion mining,
studies people's opinions, sentiments, evaluations, attitudes, and emotions
towards entities like products, services, and topics.
Terminology: It includes various subfields like sentiment mining, subjectivity
analysis, affect analysis, and review mining, all under the broader term "sentiment
analysis."
Industry vs. Academia: The term "sentiment analysis" is more common in
industry, while both "sentiment analysis" and "opinion mining" are frequently
used in academia.
Historical Origins: The term "sentiment analysis" first appeared in 2003
(Nasukawa & Yi), and "opinion mining" also in 2003 (Dave et al.), though related
research began earlier.
Focus on Opinions: The book uses "opinion" to broadly represent sentiment,
evaluation, appraisal, attitude, and emotion but distinguishes them when
necessary.
Positive and Negative Sentiments: Sentiment analysis mainly focuses on opinions
that express positive or negative sentiments.
Early Research Gap: Despite NLP’s long history, little research was done on
opinions and sentiments before 2000.
Rapid Growth Since 2000: Sentiment analysis has become a major research area
due to its vast applications and increasing commercial interest.
Industry Impact: The rise of sentiment analysis is driven by its real-world
applications across multiple domains.
Challenging Research Problems: The field presents unique and complex
challenges that had not been explored before.
Role of social media: The explosion of opinionated data from social media has
fueled sentiment analysis research.
Interdisciplinary Influence: Sentiment analysis impacts NLP, management
sciences, political science, economics, and social sciences.
Earlier Related Work: Prior research focused on metaphor interpretation,
sentiment adjectives, subjectivity, viewpoints, and affects (Hatzivassiloglou &
McKeown, Wiebe, Hearst, etc.).
Purpose of the Book: The book provides an up-to-date, comprehensive
introduction and survey of sentiment analysis, covering key concepts and
techniques.

Sentiment Analysis Applications

Importance of Opinions: Opinions influence human behavior and decision-
making in many aspects of life.
Businesses & Consumers: Companies want to know public opinions about their
products, while consumers check reviews before making purchases.
Traditional Opinion Gathering: Before social media, people relied on friends,
family, surveys, and focus groups for opinions.
Rise of social media: The internet now provides vast amounts of opinions through
reviews, blogs, forums, and social media platforms.
Challenges in Analyzing Opinions: The huge amount of opinionated content
online makes it difficult for individuals and companies to find and summarize
useful insights.
Need for Sentiment Analysis: Automated tools help analyze and summarize
opinions efficiently.
Impact of Social Media Opinions: Social media posts have influenced businesses,
politics, and social movements, such as the Arab Spring in 2011.
Internal & External Data: Organizations collect opinions not only from the
internet but also from internal sources like customer emails and surveys.
Growing Industry Interest: Many companies, from startups to tech giants like
Google and Microsoft, have invested in sentiment analysis technology.
Applications in Various Fields: Sentiment analysis is used in industries like
consumer products, healthcare, finance, social events, and elections.
Research & Practical Use: Many studies have explored sentiment analysis
applications, such as predicting sales, ranking products, and tracking political
opinions.
Social media & Business Insights: Twitter and blog opinions have been used to
predict election results, stock market trends, and box-office revenues.
Sentiment & Emotions in Literature: Some studies have analyzed emotions in
novels, fairy tales, and emails to understand differences in how genders express
emotions.
Investment & Trading Strategies: Researchers have examined how online
sentiment influences trading decisions and stock market performance.
Social Relations & Sentiment: Sentiment analysis has been used to study how
people interact and influence each other online.
Successful Sentiment Analysis Systems: Some systems, like Opinion Parser, have
effectively tracked opinions on movies and accurately predicted box-office
success.

Sentiment Analysis Research

Sentiment analysis is a popular research area not only because of its real-life
applications but also due to its challenging nature in NLP. Before 2000, there was
little research in this field due to the lack of digital opinion text. Since then, it has
rapidly grown into a major topic in NLP, data mining, Web mining, and
information retrieval. It has also expanded beyond computer science into
management sciences and other disciplines.
1. Different Levels of Analysis
Sentiment analysis is studied at three main levels of granularity: document
level, sentence level, and entity/aspect level.

At the document level, the goal is to classify an entire document as

expressing a positive or negative sentiment. This method works well when
the document discusses a single entity but is not effective for documents
that evaluate multiple entities.

At the sentence level, each sentence is analyzed to determine whether it

expresses a positive, negative, or neutral opinion. This level is closely
related to subjectivity classification, which distinguishes factual sentences
from subjective ones. However, even factual sentences can imply opinions,
making this task more complex.

At the entity and aspect level, a more detailed analysis is performed to

identify what specific aspects of an entity are liked or disliked. This method
extracts sentiments directed at specific aspects, such as the battery life or
call quality of a smartphone. This level of analysis is particularly useful for
 structured opinion summaries that enable qualitative and quantitative
insights.

In addition, sentiment analysis must differentiate between regular opinions

and comparative opinions. Regular opinions express sentiments about a
single entity or its aspects (e.g., “Coke tastes very good”), while
comparative opinions evaluate multiple entities based on shared aspects
(e.g., “Coke tastes better than Pepsi”). Comparative opinions introduce
additional challenges, as they involve ranking preferences rather than
simple sentiment classification.

2. Sentiment Lexicon and Its Issues

Sentiment words, also known as opinion words, are key indicators of
sentiment in text. These include positive words like good and amazing and
negative words like bad and terrible. In addition to single words, sentiment
phrases and idioms also play a role. A collection of such words and phrases
is called a sentiment lexicon, which researchers have developed using
various algorithms.

However, relying solely on sentiment words is not enough for accurate

sentiment analysis. One major challenge is that words can have different
meanings in different contexts. For example, suck usually implies a
negative sentiment (this camera sucks), but in some cases, it can be positive
(this vacuum cleaner really sucks).

Another challenge is that some sentences containing sentiment words do

not express any actual sentiment. For instance, questions and conditional
sentences like Can you tell me which Sony camera is good? or If I find a
good camera, I will buy it contain sentiment words but do not convey an
opinion. However, some similar sentences can express sentiments, making
classification difficult.

Sarcasm further complicates sentiment analysis. Sentences like What a

great car! It stopped working in two days use sarcasm to convey negative
sentiment despite including a positive phrase. Sarcasm is particularly
common in political discussions, making political sentiment analysis even
harder.

Additionally, some sentences imply opinions without explicitly using

sentiment words. Objective sentences such as This washer uses a lot of
 water or after sleeping on the mattress for two days, a valley has formed in
the middle express negative sentiments based on factual statements rather
than explicit opinion words.

These challenges highlight the complexity of sentiment analysis, requiring

advanced techniques beyond just sentiment lexicons to achieve accurate
results.

Sentiment Analysis as Mini NLP

3. Natural Language Processing Issues
Sentiment analysis is a complex NLP problem that involves various NLP
challenges such as coreference resolution, negation handling, and word
sense disambiguation. These challenges make sentiment analysis difficult
since they are unsolved problems in NLP. However, sentiment analysis is
also a restricted NLP problem, as it does not require full semantic
understanding but only the detection of sentiments and their targets.

This field provides an excellent platform for NLP researchers to make

progress in various NLP areas while having a practical impact. The book
aims to highlight core problems and state-of-the-art algorithms,
encouraging researchers from other NLP domains to contribute to
sentiment analysis advancements.

Before this book, important works such as Computing Attitude and Affect
in Text (2006) and a survey by Pang and Lee (2008) provided great
insights. However, sentiment analysis has grown significantly in the last
five years, leading to a better understanding of the problem, new models,
and refined methodologies.

Early research focused mainly on classifying sentiment at the document or

sentence level, but real-world applications require more fine-grained
analysis. As the field matures, research has become more structured, and
various approaches are now unified under a clear framework.
The Problem of Sentiment Analysis
The problem of sentiment analysis involves defining an abstraction that structures
the task of analyzing opinions and sentiments in natural language text. This
structured approach helps researchers identify interrelated sub-problems and
develop more robust and accurate techniques. It also provides a framework for
practitioners to understand the key challenges in building a sentiment analysis
system and the expected outputs.

Unlike factual information, opinions are subjective, making it essential to analyze

multiple opinions rather than relying on a single viewpoint. Since the Web
contains vast amounts of opinionated text, summarizing these opinions is
necessary for practical applications. This chapter defines the types of summaries
that may be required and discusses related concepts like subjectivity and emotion.

Although sentiment analysis applies to various forms of text, product reviews are
commonly used as examples because they are highly focused and opinion-rich.
However, other opinion sources, such as news articles, tweets, blogs, and forum
discussions, pose different challenges. Tweets, despite being informal, are easier
to analyze due to their short length and direct expressions. Product reviews are
also relatively straightforward because they contain little irrelevant content.
Forum discussions, however, are more complex due to unstructured
conversations and user interactions.

Additionally, different domains pose varying levels of difficulty. Analyzing

opinions on products and services is usually easier, while social and political
discussions are much harder due to complex sentiments, sarcasm, and irony.
These challenges make sentiment analysis a difficult yet essential field for
understanding public opinion.

Definition of Opinion
Sentiment analysis involves understanding and structuring opinions expressed in
text. A key aspect of this process is identifying the opinion target (what the
opinion is about) and the sentiment (positive, negative, or neutral). For example,
in a product review, opinions may be directed at the product as a whole or at
specific aspects like battery life or picture quality.

Opinions also have opinion holders (who expressed the opinion) and time stamps
(when the opinion was expressed). These elements are crucial for analyzing
sentiment trends over time and understanding different perspectives. A structured
opinion is represented as a quadruple (target, sentiment, holder, time), but in
practical applications, a more detailed quintuple model is used:
(entity, aspect, sentiment, holder, time).

Entities can be anything from products and services to people and events. Each
entity has multiple aspects (attributes or components), making sentiment analysis
more detailed. For example, a camera can have aspects like picture quality,
battery life, and weight. Sentiments can be associated with either the entity as a
whole or specific aspects.

Since natural language is complex, some challenges arise. Sentiments may not
always be explicit, and aspects may not be clearly mentioned in the same
sentence. Sarcasm, context-based opinions, and relationships between entity parts
add further difficulties. For instance, “The ink of this printer is expensive”
expresses an opinion on ink price, not the printer as a whole.

Despite limitations, structuring opinions using the quintuple model allows for
transforming unstructured text into structured data. This enables powerful
quantitative, qualitative, and trend analyses using databases and analytical tools.
Comparative opinions, which express preferences between entities (e.g., "Coke
tastes better than Pepsi"), require a separate framework, discussed in later
sections.

Definition of Opinion Summary - Affect, Emotion, and Mood

Opinion summarization is essential because opinions are subjective, and a single
opinion is usually not enough for decision-making. Most applications require
analyzing opinions from a large number of people, making summaries necessary.
An effective summary should include opinions on different entities and aspects,
along with a quantitative perspective to show overall sentiment trends.

The opinion quintuple (entity, aspect, sentiment, opinion holder, time) provides a
strong foundation for both qualitative and quantitative summaries. A widely used
approach is aspect-based opinion summarization, which organizes opinions by
entity aspects rather than generating traditional text summaries.
For example, a structured summary of reviews for a digital camera might show
that 105 people gave a positive opinion on the camera overall, while 12 people
gave a negative opinion. The picture quality aspect received 95 positive and 10
negative reviews, while the battery life aspect had 50 positive and 9 negative
reviews.

This structured summary allows users to quickly understand customer opinions.

If further details are needed, they can drill down into individual review sentences
for more context. Aspect-based summarization is widely researched and helps in
better decision-making by providing both an overview and detailed insights.

Different Types of Opinions

Opinions can be classified into regular opinions and comparative opinions.
Additionally, opinions can be categorized based on how they are expressed, as
explicit opinions or implicit opinions.

1. Regular and Comparative Opinions

Regular Opinions
Regular opinions are the most common type and can be further divided
into:
Direct opinions: These express sentiments explicitly about an entity or
aspect, e.g., "The picture quality is great."
Indirect opinions: These express opinions indirectly through effects on
related entities. For example, "After taking the drug, my joints felt worse."
Here, the negative sentiment is directed at the drug based on its effect on
the joints. Indirect opinions are more challenging to analyze, especially in
fields like medicine.

Comparative Opinions
Comparative opinions compare two or more entities based on shared
aspects, showing similarities, differences, or preferences. For example:
 - "Coke tastes better than Pepsi." (Comparison between two entities)
- "Coke tastes the best." (Superlative comparison)
Comparative opinions are usually expressed using comparative or
superlative adjectives (better, best) or preference words (prefer). They are
discussed in more detail in later chapters.

2. Explicit and Implicit Opinions

Explicit Opinions
Explicit opinions are subjective statements that clearly express sentiments.
For example:
- "Coke tastes great." (Regular opinion)
- "Coke tastes better than Pepsi." (Comparative opinion)
Explicit opinions are easier to detect and classify, and most research has
focused on them.

Implicit Opinions
Implicit opinions are objective statements that imply a sentiment without
directly stating it. For example:
- "I bought the mattress a week ago, and a valley has formed." (Implies
negative sentiment)
- "The battery life of Nokia phones is longer than Samsung phones."
(Implies a positive sentiment about Nokia's battery life)
Implicit opinions are more difficult to analyze because they require deeper
understanding of context. Some studies have explored how syntactic
choices in sentences, such as news headlines, influence the perception of
sentiment.

Author and Reader Standpoint

Opinions can be viewed from two perspectives: the author (opinion holder) who
expresses the opinion and the reader who interprets it. The same opinion can be
perceived differently based on the reader’s situation.
For example, the statement “The housing price has gone down, which is bad for
the economy” expresses a negative sentiment from the author’s perspective.
However, while sellers may see it as negative, buyers might view it positively.
Similarly, “I am so happy that Google’s share price shot up today” is a positive
statement, but a reader who sold their shares at a loss may not feel the same way.

Currently, this issue is largely ignored in sentiment analysis research. Most

studies assume a fixed standpoint, usually taking the perspective of consumers or
the general public. Opinions from companies or service providers are often
classified as advertisements if explicitly stated or fake reviews if mixed with
consumer opinions. Understanding different viewpoints remains a challenge in
sentiment analysis.

Document Sentiment Classification

Document sentiment classification is a major research area in sentiment analysis
that focuses on determining whether an entire opinion document expresses a
positive or negative sentiment. This approach is also called document-level
sentiment classification because it treats the whole document as a single unit of
information. Most studies in this area focus on classifying online reviews, though
the techniques apply to other forms of opinion text as well.

Problem Definition
The goal is to analyze an opinion document \( d \) and determine the overall
sentiment \( s \) expressed about an entity. This is represented in the quintuple
model as:
(_, GENERAL, s, _, _)
where entity (e), opinion holder (h), and time (t) are either known or considered
irrelevant.

Classification vs. Regression

There are two main approaches based on how sentiment \( s \) is measured:
Classification: If sentiment is labeled as categorical values (e.g., positive or
negative), it is a classification problem.
Regression: If sentiment is represented by numeric scores (e.g., a rating scale
from 1 to 5), it becomes a regression problem.

Assumptions in Sentiment Classification

For this approach to work effectively, research assumes that:
1. The document expresses opinions on only one entity.
2. The document is written by a single opinion holder.

This assumption is valid for product and service reviews, where a user evaluates
a single item. However, it does not hold for forum posts and blog articles, where
the author might discuss multiple entities or compare them.

Classification Techniques
Most document-level sentiment classification techniques use supervised learning,
though some unsupervised methods exist. Sentiment regression is mainly done
using supervised learning.

Recent Advancements
Recent research has introduced cross-domain sentiment classification (adapting
models to different domains) and cross-language sentiment classification
(analyzing sentiment in multiple languages). These topics have become essential
as sentiment analysis expands across different industries and languages.

Sentiment Classification Using Supervised Learning

Sentiment classification is typically formulated as a two-class classification
problem where opinions are categorized as positive or negative. Most research
uses online reviews for training and testing, assigning positive sentiment to
reviews with 4-5 stars and negative sentiment to those with 1-2 stars. The neutral
class (e.g., 3-star reviews) is often ignored to simplify the classification process.
Since sentiment classification is a text classification problem, it shares similarities
with traditional topic-based classification. However, instead of focusing on topic-
related words, it emphasizes sentiment words (e.g., great, excellent, horrible,
worst). Standard supervised learning methods such as Naïve Bayes and Support
Vector Machines (SVM) have been widely used for sentiment classification. The
first major study (Pang, Lee, and Vaithyanathan, 2002) applied these methods to
classify movie reviews and found that using unigrams (bag of words) was
effective.

Key Features for Sentiment Classification

To improve classification performance, researchers use several feature
engineering techniques:
• Term Frequency (TF) and N-grams: Words and phrases with their
frequency counts.
• Part of Speech (POS) Tags: Adjectives are often strong sentiment
indicators.
• Sentiment Words and Phrases: Positive and negative words, including
idioms and expressions.
• Sentiment Shifters: Words like not that change sentiment polarity (e.g., not
good).
• Syntactic Dependency: Analyzing relationships between words using
dependency trees.

Custom Techniques in Sentiment Classification

Instead of using standard machine learning methods, some researchers have
developed custom techniques for sentiment classification. These include score-
based functions, manual domain-specific lexicons, and graph-based approaches.

Advancements and Research Studies

Numerous studies have explored new methods and features for sentiment
classification:
• Graph-based models (Pang & Lee, 2004).
• Syntactic relations (Mullen & Collier, 2004).
 • Context-aware sentiment shifting (Kennedy & Inkpen, 2006).
• Lexicon-based approaches (Qiu et al., 2009).
• Twitter sentiment classification using hashtags and informal language
(Wang et al., 2011).
• Deep learning and word vectors to capture latent word meanings (Maas et
al., 2011).

Supervised learning remains the dominant approach for sentiment classification,

with ongoing research focusing on feature selection, cross-domain adaptation,
and deep learning methods. By leveraging advanced NLP techniques, researchers
continue to enhance sentiment classification accuracy across various
applications.

Sentiment Classification Using Unsupervised Learning

Unsupervised sentiment classification relies on sentiment words and phrases
rather than labeled training data. One of the earliest and most influential
approaches was introduced by Turney (2002), which classifies sentiment using
syntactic patterns and pointwise mutual information (PMI).

Turney’s Unsupervised Sentiment Classification Approach

The algorithm follows three main steps:

1. Phrase Extraction: Extracts two-word phrases using predefined part-of-

speech (POS) patterns (e.g., adjective-noun, adverb-adjective). Words like
JJ (adjective) and RB (adverb) often express opinions, while surrounding
nouns and verbs provide context.

2. Sentiment Orientation (SO) Calculation: Uses PMI to measure a phrase's

association with reference words like "excellent" (positive) and "poor"
(negative). The sentiment score is calculated as:

SO(phrase) = PMI(phrase, “excellent”) - PMI(phrase, “poor”)

 The probabilities are estimated using search engine hit counts to determine
how often a phrase appears near positive or negative words.
3. Review Classification: Computes the average SO of all phrases in a
review. If the score is positive, the review is classified as positive;
otherwise, it is negative. The accuracy ranges from 84% for automobile
reviews to 66% for movie reviews.

Lexicon-Based Sentiment Classification

Another popular unsupervised method is lexicon-based classification, which uses
a dictionary of sentiment words and phrases with predefined polarities and
intensities. This method incorporates intensification and negation to refine
sentiment scoring. It was originally applied to sentence-level and aspect-level
classification (Ding et al., 2008; Hu & Liu, 2004).

Unsupervised methods are useful when labeled data is unavailable. PMI-based

classification and lexicon-based approaches remain widely used, though they
may struggle with contextual sentiment shifts and domain adaptation challenges.

Sentiment Rating Prediction

Sentiment rating prediction focuses on predicting numerical review ratings (e.g.,
1–5 stars) rather than just classifying reviews as positive or negative. This task is
often formulated as a regression problem, as ratings are ordinal values. However,
some researchers have also approached it using classification techniques.

Early Approaches
Pang and Lee (2005) experimented with SVM regression, SVM multiclass
classification (one-vs-all strategy), and a meta-learning method called metric
labeling. They found that one-vs-all classification performed poorly compared to
regression methods, as numerical ratings are not purely categorical.

Goldberg and Zhu (2006) improved rating prediction using a graph-based semi-
supervised learning approach, which used both labeled (with ratings) and
unlabeled (without ratings) reviews. In this approach, each review was
represented as a node in a graph, and links between nodes indicated review
similarity. The algorithm revised initial rating predictions by enforcing
smoothness across the graph, ensuring similar reviews had similar ratings.

Bag-of-Opinions and Regression-Based Approaches

Qu, Ifrim, and Weikum (2010) introduced a bag-of-opinions model instead of the
traditional bag-of-words approach. Each opinion was represented as a triple:
• Sentiment word (good),
• Modifier (very),
• Negator (not).
This method recognized that opinion modifiers and negations significantly impact
rating prediction. A constrained ridge regression model was trained using
sentiment lexicons and ratings from multiple domains. Similarly, Liu and Seneff
(2009) used clause-based sentiment extraction to analyze adverb-adjective-noun
phrases (e.g., very nice car) and assigned sentiment scores based on heuristic rules
rather than machine learning.

Aspect-Based Rating Prediction

Snyder and Barzilay (2007) proposed predicting ratings for individual aspects of
a review rather than an overall score. They introduced:
1. Aspect Model (predicts ratings for each aspect separately).
2. Agreement Model (captures dependencies between aspect ratings).

By combining these models, they improved accuracy in aspect-based rating

prediction.

Long, Zhang, and Zhu (2010) followed a similar approach but used a Bayesian
network classifier. To enhance accuracy, they focused on a selected subset of
reviews that comprehensively evaluated multiple aspects. These reviews
provided more reliable predictions, using Kolmogorov complexity to measure
information content.
Sentiment rating prediction extends beyond simple positive/negative
classification by estimating numerical ratings using machine learning, regression
models, and aspect-based approaches. It remains a challenging task due to
dependencies between aspects, negations, and variations in user opinion
expression.

Cross-Domain Sentiment Classification

Sentiment classification models are highly domain-sensitive, meaning a classifier
trained on one domain (e.g., movie reviews) often performs poorly on another
(e.g., product reviews). This is because opinion expressions, sentiment words,
and language structures differ across domains. Additionally, the same word may
carry different sentiments in different contexts. To address this issue, domain
adaptation or transfer learning is required.

Types of Cross-Domain Adaptation

There are two main approaches:
1. With a small amount of labeled data in the target domain (Aue & Gamon,
2005).
2. Without any labeled data in the target domain (Blitzer et al., 2007; Tan
et al., 2007).
The source domain is where labeled training data is available, while the target
domain is where sentiment classification needs to be performed.

Early Approaches to Domain Adaptation

Aue and Gamon (2005) explored four strategies for transferring sentiment
classifiers:
1. Training on a mixture of labeled reviews from multiple domains.
2. Training on limited features found in the target domain.
3. Using ensembles of classifiers from different domains.
4. Combining small labeled data with large unlabeled data using semi-
supervised learning.
The semi-supervised learning approach performed best as it leveraged both
labeled and unlabeled data.

Yang et al. (2006) proposed a feature selection method to identify domain-

independent features from fully labeled datasets of two domains. Similarly, Tan
et al. (2007) trained a base classifier in the source domain, labeled informative
examples in the target domain, and then retrained a new classifier for the target
domain.

Advanced Techniques for Cross-Domain Adaptation

Blitzer et al. (2007) introduced Structural Correspondence Learning (SCL),
which identifies pivot features—words that frequently appear in both domains
and strongly correlate with sentiment. It then uses Singular Value Decomposition
(SVD) to align features across domains, improving classification in the target
domain.

Pan et al. (2010) developed Spectral Feature Alignment (SFA), which aligns
domain-specific words into unified clusters using domain-independent words as
a bridge. This method applies spectral clustering on a bipartite graph linking
domain-independent and domain-specific words.

Other researchers applied topic modeling (He et al., 2011; Gao & Li, 2011) to
identify shared opinion topics across domains, which were then used as additional
classification features. Bollegala et al. (2011) created a sentiment-sensitive
thesaurus to map words expressing similar sentiments across domains.

Wu et al. (2009) proposed a graph-based approach that propagates sentiment

labels from the source domain to the target domain using document similarity
graphs. Xia & Zong (2011) found that some POS-based features are domain
specific, while others are domain-independent, leading them to develop a POS-
based ensemble model to improve classification.

Cross-domain sentiment classification is essential for building generalizable

sentiment models. Techniques such as SCL, SFA, topic modeling, and graph-
based label propagation help adapt classifiers to new domains, improving
accuracy without requiring large amounts of labeled data in every domain.

Cross-Language Sentiment Classification

Cross-language sentiment classification aims to perform sentiment analysis on
opinion documents in multiple languages. This research is motivated by two main
factors:
1. Many researchers want to build sentiment analysis models for their
native languages, but most sentiment analysis tools and resources are
developed for English.
2. Businesses need sentiment analysis across multiple countries to compare
consumer opinions in different languages. Instead of building models from
scratch, machine translation and existing English sentiment analysis tools
can help create models for other languages.

Early Approaches Using Machine Translation

Wan (2008) used machine translation to translate Chinese reviews into multiple
English versions and applied a lexicon-based sentiment classifier to each
translated version. The final sentiment score was determined by ensemble
methods such as averaging, voting, or weighted scoring. The study showed that
combining multiple translations improved classification accuracy.

Brooke et al. (2009) also used machine translation (English to Spanish) and
applied either lexicon-based or machine learning approaches to classify sentiment
in Spanish documents.

Supervised and Semi-Supervised Learning Approaches

Wan (2009) introduced a co-training method that used an annotated English
corpus to classify Chinese reviews. The approach worked as follows:
1. English reviews were translated into Chinese, and Chinese reviews were
translated into English.
2. English and Chinese versions of each review were treated as two
independent views of the same document.
 3. A co-training algorithm using SVM trained two classifiers, which were
then combined into a final model.
4. Chinese reviews were translated into English before applying the
classifier.
Wei and Pal (2010) proposed a transfer learning approach to reduce errors caused
by machine translation noise. They used Structural Correspondence Learning
(SCL) to identify core sentiment features shared between English and Chinese.
Search engine queries were also used to find additional highly correlated features,
which helped train the classifier more effectively.

Topic Modeling for Multi-Language Sentiment Analysis

Boyd-Graber and Resnik (2010) extended Supervised Latent Dirichlet Allocation
(SLDA) to work with multi-language reviews for sentiment rating prediction.
Their method, MLSLDA, created language-independent sentiment topics by
aligning multilingual WordNets and dictionaries.

Guo et al. (2010) developed a topic modeling method to group aspect expressions
from different languages into common aspect clusters. This approach allowed
businesses to compare sentiment across different countries.

Challenges and Future Directions

Duh et al. (2011) analyzed cross-language sentiment classification and found that
domain mismatch occurs even with perfect machine translation. They argued that
cross-language adaptation is different from monolingual adaptation, and new
algorithms need to be developed for better accuracy.

Cross-language sentiment classification is essential for global sentiment analysis.

Machine translation, co-training, transfer learning, and topic modeling are widely
used to bridge the language gap. However, domain adaptation and translation
errors remain challenges, requiring new adaptation techniques for improved
performance.
UNIT-II
Subjectivity Classification and Challenges: Sentence Subjectivity and Sentiment
Classification: Subjectivity - Sentence Subjectivity Classification - Sentence
Sentiment Classification - Dealing with Conditional Sentences - Dealing with
Sarcastic Sentences - Cross Language Subjectivity and Sentiment Classification
Using Discourse Information for Sentiment Classification - Emotion
Classification of Sentences.
Sentence Subjectivity and Sentiment Classification
Sentence Subjectivity and Sentiment Classification

Document-level sentiment classification is often too broad for many applications.

Instead, sentence-level classification aims to determine the sentiment expressed
in each sentence. However, the fundamental approach remains the same since
sentences are simply shorter documents.

Problem Definition
Given a sentence (x), the goal is to classify whether it expresses a positive,
negative, or neutral (no) opinion. Unlike document-level classification, the
quintuple model (e, a, s, h, t) is not directly used here because sentence-level
classification is often an intermediate step in identifying opinion targets.

Classification Approaches
Sentence sentiment classification can be approached in two ways:
1. Three-class classification (positive, negative, neutral).
2. Two-step classification:
• Step 1: Determine whether a sentence expresses an opinion (subjectivity
classification).
• Step 2: If it does, classify the opinion as positive or negative.

Subjectivity Classification
The first step is to distinguish between opinionated (subjective) sentences and
non-opinionated (objective) sentences. This is known as subjectivity
classification (Hatzivassiloglou & Wiebe, 2000; Riloff & Wiebe, 2003).
Objective sentences are typically treated as neutral, but this can be misleading.

For example, in the review:

*"Then, it stopped working yesterday."*
This is an objective statement but implies a negative sentiment about the phone.
Thus, it is more effective to classify sentences as opinionated or not opinionated,
rather than simply subjective or objective.

Importance of Sentence-Level Classification

While knowing if a sentence is positive or negative is useful, it is not enough
without identifying opinion targets (entities/aspects). However, sentence-level
classification helps in analyzing whether opinions on entities or aspects are
positive or negative, making it a crucial step in fine-grained sentiment analysis.

Sentence Subjectivity Classification

Subjectivity classification is the process of classifying sentences into subjective
(expressing opinions, emotions, or beliefs) and objective (stating factual
information) (Wiebe et al., 1999). Subjective sentences may include evaluations,
judgments, speculations, or stances, but not all of them express sentiment.

Early research treated subjectivity classification as a standalone task, but later

studies integrated it as a first step in sentiment classification to remove objective
sentences, assuming they do not express sentiment. However, this approach has
limitations since some objective sentences can imply opinions, e.g., *"Then, it
stopped working yesterday."*

Supervised Learning Approaches

Most subjectivity classification methods use supervised learning:
• Wiebe et al. (1999) applied a Naïve Bayes classifier with binary features
such as pronouns, adjectives, modal verbs, and adverbs.
• Yu & Hatzivassiloglou (2003) used sentence similarity and Naïve Bayes,
assuming subjective sentences are more similar to each other than to
objective ones.
• Riloff & Wiebe (2003) proposed a bootstrapping approach, using high-
precision classifiers to identify strong subjective and objective sentences.
These were then used to generate patterns for further classification.
Unsupervised Learning and Lexical Approaches
• Wiebe (2000) used subjective expressions as indicators, expanding the
lexicon using distributional similarity. However, this method had low
precision and required filtering.
• Hatzivassiloglou & McKeown (1997) identified gradable adjectives, as
adjective intensity (e.g., *very good, somewhat bad*) is a key subjectivity
marker.
• Pang & Lee (2004) introduced a mincut-based algorithm, leveraging
sentence connectivity in documents to improve classification accuracy.

Advanced Techniques
Wilson et al. (2004) classified clauses within sentences as neutral, low,
medium, or highly subjective instead of treating entire sentences as subjective
or objective.
Benamara et al. (2011) proposed a four-class system:
• S (Subjective Evaluative): Expresses positive/negative sentiment.
• OO (Objective with Implied Opinion): Factual but implies sentiment.
• (Purely Objective): No sentiment or implied opinion.
• SN (Subjective but Non-Evaluative): Expresses emotions or beliefs but
without clear sentiment.

Applications to Different Languages

Subjectivity classification has been studied in multiple languages:
• Arabic (Abdul-Mageed et al., 2011) and Urdu (Mukund & Srihari, 2010)
using language-specific features.
• Barbosa & Feng (2010) applied subjectivity classification to Twitter,
incorporating hashtags, retweets, emoticons, and punctuation.
• Raaijmakers & Kraaij (2008) found that character n-grams (instead of
words) performed well for subjectivity classification.

Conclusion
Subjectivity classification is crucial for sentiment analysis, as filtering out
objective sentences improves sentiment classification accuracy. However, some
objective sentences still imply sentiment, requiring more advanced methods that
consider context and linguistic structure.

Sentence Sentiment Classification

Once a sentence is classified as subjective, the next step is to determine whether
it expresses a positive or negative sentiment. This task is similar to document-
level sentiment classification but focuses on individual sentences. Supervised
learning and lexicon-based methods are commonly used for this classification.

Assumption in Sentence-Level Sentiment Classification

Much of the research assumes that each sentence expresses a single sentiment
from a single opinion holder. This assumption works for simple sentences, e.g.,
*“The picture quality of this camera is amazing.”* However, it fails for
compound and complex sentences, e.g., *“The picture quality is amazing, but the
viewfinder is too small.”* Such sentences contain mixed sentiments, requiring
finer analysis.

Early Approaches
• Yu & Hatzivassiloglou (2003) modified Turney’s (2002) PMI-based
approach by using a large set of seed adjectives and a log-likelihood ratio
to determine sentiment orientation.
• Hu & Liu (2004) proposed a lexicon-based approach, assigning +1 to
positive words and -1 to negative words while considering negation words
(not) and contrast words (but, however).
• Kim & Hovy (2004, 2007) applied multiplicative aggregation of sentiment
scores instead of summation.

Machine Learning and Semi-Supervised Approaches

• Nigam & Hurst (2004) used domain-specific lexicons and shallow NLP
techniques to classify sentence sentiment.
• Gamon et al. (2005) applied semi-supervised learning (Expectation
Maximization with Naïve Bayes) using a small labeled dataset and a large
unlabeled dataset to classify sentences into positive, negative, or mixed/no
sentiment.
 • McDonald et al. (2007) introduced a hierarchical sequence model based on
Conditional Random Fields (CRF) to classify sentiment at both the
sentence and document levels, showing that joint learning improves
accuracy.
• Täckström & McDonald (2011) developed a partially supervised method
that learned from document-level labels but performed both document and
sentence-level classification.

Context-Aware Sentiment Analysis

• Hassan et al. (2010) focused on attitude detection in online discussions
using Markov models. Sentences were classified based on second-person
pronouns, and a dependency tree was used to resolve conflicting sentiment
words.
• Davidov et al. (2010) studied Twitter sentiment classification, where each
tweet is essentially a sentence. They used hashtags, smileys, punctuation,
and frequent patterns as features, proving their effectiveness in sentiment
analysis.

Conclusion
Sentence-level sentiment classification is crucial for fine-grained sentiment
analysis but is challenging due to mixed sentiments in compound sentences.
Lexicon-based and supervised learning approaches remain widely used, while
context-aware and partially supervised models show promising results.
Dealing with Conditional Sentences
Most research in sentence-level sentiment classification assumes a one-size-fits-
all approach, but Narayanan, Liu, and Choudhary (2009) argued that different
types of sentences require different treatments. They proposed a divide-and-
conquer approach, focusing on conditional sentences, which pose unique
challenges for sentiment analysis.

Challenges of Conditional Sentences

Conditional sentences describe hypothetical situations and consequences and
typically consist of two dependent clauses:
1. Condition clause (e.g., *"If someone makes a reliable car,"*).
2. Consequence clause (e.g., *"I will buy it."*).

A major challenge is that sentiment words alone cannot determine whether a

conditional sentence expresses an opinion. For example:
• "If someone makes a reliable car, I will buy it."* → Contains the positive
word "reliable" but does not express an opinion about any specific car.
• "If your Nokia phone is not good, buy this Samsung phone."* → Positive
towards Samsung, but neutral towards Nokia (despite "not good").

Supervised Learning Approach

To address this issue, Narayanan et al. (2009) proposed a supervised learning
method using linguistic features, including:
• Sentiment words/phrases and their positions in the sentence.
• POS tags of sentiment words.
• Tense patterns to identify hypothetical scenarios.
• Conditional connectives (e.g., *if, unless, provided that*).

Need for Specialized Sentence Handling

Another complex sentence type is questions, which are often not opinionated but
sometimes imply sentiment. For example:
• "Can anyone tell me where to buy a good Nokia phone?"* → No sentiment.
• "Can anyone tell me how to fix this lousy Nokia phone?"* → Negative
sentiment towards Nokia.

Conclusion
Handling conditional and question sentences requires customized approaches, as
traditional sentiment classification techniques fail to capture their contextual
meaning. More focused research is needed to improve accuracy in different
sentence types.
Dealing with Sarcastic Sentences
Sarcasm is a complex linguistic phenomenon where the intended meaning is the
opposite of the literal meaning. In sentiment analysis, sarcasm presents a major
challenge because a positive statement may actually imply negative sentiment,
and vice versa. While sarcasm is not very common in product reviews, it is
frequently found in political discussions and online commentaries.

Early Research on Sarcasm Detection

1. Tsur, Davidov & Rappoport (2010) proposed a semi-supervised learning
approach to detect sarcasm.
• Used a small set of labeled sarcastic sentences (seeds) and expanded them
using Web search.
• Assumed that sarcastic sentences tend to co-occur with other sarcastic
sentences.
• Extracted up to 50 Web snippets per sarcastic sentence and added them to
the training set.
• Employed pattern-based features (word sequences) and punctuation-based
features (e.g., exclamation marks, question marks, capitalized words).
• Used a kNN-based classification method but did not classify sentiment
polarity (positive/negative).

2. González-Ibáñez, Muresan & Wacholder (2011) studied sarcasm in Twitter

sentiment analysis:
• Focused on distinguishing sarcastic tweets from directly positive or
negative tweets (neutral tweets were ignored).
• Used SVM and logistic regression with features like unigrams, emoticons,
interjections (ah, oh, yeah), and punctuations (!, ?).
• The classification task was very challenging, with a best accuracy of only
57%.
• Like the previous study, this work only identified sarcasm but did not
classify sentiment polarity.
Challenges and Future Research
• Sarcastic sentiment classification remains difficult, as sarcasm involves
context, tone, and speaker intent.
• Existing models struggle with sarcasm detection, achieving low accuracy.
 • More advanced NLP techniques, such as deep learning and contextual
embeddings, may help improve sarcasm detection and sentiment
classification.

Cross-language Subjectivity and Sentiment Classification

Cross-language subjectivity and sentiment classification aims to analyze
sentiment in multiple languages, leveraging English-language resources and
automated translation. This is crucial for languages with limited sentiment
analysis tools. Researchers have proposed three main strategies:
1. Translate test sentences from the target language into English and classify them
using an English classifier.
2. Translate an English training corpus into the target language and train a
classifier in the target language.
3. Translate an English sentiment lexicon into the target language and build a
lexicon-based classifier.

Early Research and Approaches

1. Kim & Hovy (2006) tested two strategies:
• Translating German emails into English and classifying them with an
English lexicon.
• Translating English sentiment words into German and classifying German
emails.
2. Mihalcea, Banea & Wiebe (2007) experimented with:
• Translating an English subjectivity lexicon into Romanian for
classification.
• Creating a subjectivity-labeled corpus by manually translating a parallel
corpus and projecting sentiment labels from English to Romanian.
• The manual translation method gave better results but required significant
effort.

Automated Translation-Based Approaches

1. Banea et al. (2008, 2010) studied multiple strategies, including:
 • Automatic translation of labeled English text into the target language
for training.
• Classifying translated text in English and mapping the labels back to
the target language.
• Combining translated corpora from multiple languages with English to
improve classification accuracy.
2. Bautin, Vijayarenu & Skiena (2008) translated target language documents
into English and used a lexicon-based approach at the aspect level, similar
to Hu & Liu (2004).

Evaluating Multi-Lingual Sentiment Analysis

• Kim, Li & Lee (2010) introduced multi-lingual comparability, measuring
classification agreement between translated texts and original sentiment
labels using Cohen’s Kappa.
• Lu et al. (2011) proposed a maximum entropy-based EM algorithm to
jointly improve sentiment classification in both the source and target
languages using a parallel corpus.

Conclusion
Cross-language sentiment classification is highly dependent on translation
quality. While translated training data and lexicons have shown promising results,
manually created corpora still perform better. Future research should focus on
improving machine translation and domain adaptation techniques for better
sentiment consistency across languages.

Using Discourse Information for Sentiment Classification

Most sentiment classification methods focus on individual documents or
sentences without considering discourse-level relationships. However, discourse
information, such as sentence-to-sentence or clause-level connections, can
improve sentiment analysis accuracy. Researchers have explored rhetorical
relations and opinion frames to enhance classification.
Discourse-Based Sentiment Annotation
• Asher, Benamara & Mathieu (2008) introduced five rhetorical relations—
Contrast, Correction, Support, Result, and Continuation—to analyze
sentiment at the discourse level.
• Somasundaran, Ruppenhofer & Wiebe (2008) proposed opinion frames,
which represent opinions and relationships between their targets.

Discourse-Based Sentiment Classification

• Somasundaran et al. (2009) applied collective classification, where
sentences are nodes and relations between sentences form edges in a graph.
This method helps classify sentences based on their contextual
relationships rather than treating them in isolation.
• Zhou et al. (2011) used pattern mining to analyze compound sentences and
classify sentiment more accurately. For example, in the sentence:
“Although Fujimori was criticized by the international community, he was
loved by the domestic population because people hated the corrupted ruling
class.”*
The negative words outweigh the positive ones, but the overall sentiment
is positive due to discourse structure.

Segment-Based Classification
Zirn et al. (2011) classified discourse segments, where each segment expresses
one sentiment. They used Markov Logic Networks (MLN) to integrate sentiment
lexicons and neighboring discourse context for better classification.

Conclusion
Discourse relations provide valuable insights for sentiment classification,
particularly in complex sentences and multi-sentence texts. Future research
should further integrate discourse-level features to enhance sentiment analysis
accuracy.
UNIT-III
Sentiment Lexicon generation and Summarization: Sentiment Lexicon
Generation: Dictionary-Based Approach - Corpus-Based Approach - Desirable
and Undesirable Facts. Analysis of Comparative Opinions: Problem Definition -
Identify Comparative Sentences - Identifying the Preferred Entity Set - Special
Types of Comparison - Entity and Aspect Extraction. Opinion Summarization and
Search: Aspect-Based Opinion Summarization - Enhancements to Aspect-Based
Summary - Contrastive View Summarization - Traditional Summarization -
Summarization of Comparative Opinions - Opinion Search - Existing Opinion
Retrieval Techniques. Mining Intentions: Problem of Intention Mining - Intention
Classification - Fine-Grained Mining of Intentions.
Sentiment Lexicon Generation
A sentiment lexicon is a collection of words and phrases that convey positive or
negative sentiments. These words, also known as opinion words, polar words, or
opinion-bearing words, help in sentiment analysis. Positive words describe
desired qualities (e.g., beautiful, amazing), while negative words describe
undesired qualities (e.g., awful, poor). In addition to individual words, sentiment
phrases and idioms, like cost an arm and a leg, are also part of the sentiment
lexicon.

Sentiment words are classified into two types: base type and comparative type.
Base-type sentiment words express direct opinions, while comparative-type
words (e.g., better, worse, best, worst) indicate relative comparisons rather than
absolute opinions. For example, "Pepsi tastes better than Coke" does not state
whether either drink is good or bad, only that Pepsi is preferable in comparison.

There are three main methods for compiling a sentiment lexicon: manual
approach, dictionary-based approach, and corpus-based approach. The manual
approach is accurate but time-consuming, so it is often combined with automated
methods. The dictionary-based approach starts with a small set of words and
expands it using synonyms and antonyms from dictionaries. The corpus-based
approach identifies sentiment words based on patterns and statistical relationships
in large text datasets.

While automated methods are efficient, they can make errors, so manual
verification is often necessary. Additionally, factual statements that imply
opinions pose challenges in sentiment analysis, as they are not always explicitly
positive or negative.

Dictionary-Based Approach
The dictionary-based approach is a widely used method for compiling sentiment
lexicons by leveraging dictionaries like WordNet to expand a set of seed words
based on their synonyms and antonyms. This approach starts with a manually
selected list of positive and negative sentiment words and iteratively grows the
list by finding related words in the dictionary. The process continues until no new
words are discovered. After completion, manual inspection is usually performed
to correct errors and refine the lexicon.

A key improvement to this approach was proposed by Kamps et al. (2004), who
introduced a WordNet distance-based method. They calculated the sentiment
orientation of adjectives based on their relative distance from reference words
like good and bad. Similarly, Blair-Goldensohn et al. (2008) extended the
approach by incorporating a neutral seed set to prevent sentiment propagation
through neutral words. Their method used a directed semantic graph with
weighted edges to assign sentiment scores using a modified label propagation
algorithm.

Graph-based methods have also been explored to enhance sentiment lexicon

generation. Rao and Ravichandran (2009) experimented with three semi-
supervised learning algorithms—Mincut, Randomized Mincut, and label
propagation—to classify sentiment words. Additionally, Hassan and Radev
(2010) used a Markov random walk model over a word-relatedness graph, where
sentiment orientation was determined by the distance from a word to predefined
positive and negative seed sets.

Another widely used technique is the PMI (Pointwise Mutual Information)-based

method by Turney and Littman (2003), which determines a word’s sentiment by
measuring its association strength with positive and negative reference words.
Esuli and Sebastiani (2005, 2006) further refined dictionary-based sentiment
classification by employing supervised learning. They expanded the initial seed
sets using WordNet’s synonyms and antonyms, generated feature vectors for each
term, and trained classifiers to categorize sentiment words.

Some methods also integrated fuzzy set theory to refine sentiment classification.
Andreevskaia and Bergler (2006) employed multiple bootstrapping runs to reduce
errors and normalized sentiment scores within a [0,1] range. Others, like Kaji and
Kitsuregawa (2006, 2007), used heuristics from web page structures (e.g., pros
and cons tables) to extract sentiment words from large corpora.
A significant limitation of the dictionary-based approach is that it primarily
provides general, domain-independent sentiment orientations. Words like quiet
can have different meanings depending on the domain (e.g., negative for a
speakerphone but positive for a car). To address this, corpus-based methods are
often used in combination with dictionary-based approaches to capture context-
dependent sentiment orientations.

Corpus-Based Approach
The corpus-based approach is used to expand a sentiment lexicon or adapt a
general-purpose sentiment lexicon to a specific domain. Unlike the dictionary-
based approach, which provides general sentiment orientations, corpus-based
methods help identify context-dependent sentiment by analyzing text data.
However, sentiment words can have different meanings within the same domain
based on context, making lexicon generation more complex.

One early method was introduced by Hazivassiloglou and McKeown (1997), who
used linguistic rules to determine sentiment words from a corpus. Their sentiment
consistency rule states that adjectives connected by AND usually share the same
sentiment, while conjunctions like BUT indicate a sentiment shift. A graph was
built with words linked by sentiment similarity, and clustering was applied to
classify words as positive or negative.

To improve upon this, Kanayama and Nasukawa (2006) introduced intra-

sentential and inter-sentential sentiment consistency. They analyzed sentiment
propagation within and across sentences, identifying domain-specific sentiment
words. Similarly, Ding, Liu, and Yu (2008) noted that sentiment varies based on
aspect-word pairs. For example, long is positive in “battery life is long” but
negative in “it takes a long time to focus.” Their method classified sentiment
based on both the aspect and the modifying sentiment word.

Other studies have focused on expression-level sentiment classification. Wilson,

Wiebe, and Hoffmann (2005) examined contextual subjectivities, where words
may change sentiment depending on their context. They used supervised learning
to classify expressions as positive, negative, or neutral. Choi and Cardie (2009)
refined this by adapting general lexicons for specific domains using integer linear
programming.

For domain adaptation, Du et al. (2010) developed a method to transfer a

sentiment lexicon from one domain to another. They used mutual reinforcement
learning, where documents with many positive words are classified as positive,
and words frequently appearing in positive documents are also marked as
positive. This method helped align sentiment words between in-domain and out-
of-domain corpora.

Additionally, Wiebe and Mihalcea (2006) explored subjectivity at the word sense
level, showing that word sense disambiguation could improve sentiment analysis.
Another unique approach by Brody and Diakopoulos (2011) analyzed word
lengthening in social media (e.g., slooooow), finding that longer words often
convey stronger emotions.

Overall, corpus-based methods allow sentiment lexicons to be tailored to specific

domains and contexts. However, these methods often require large labeled
datasets and complex linguistic rules, making them more computationally
intensive than dictionary-based approaches.

Desirable and Undesirable Facts

While sentiment words are typically subjective expressions indicating positive or
negative opinions, many objective words can also imply sentiments based on
domain-specific contexts. These words represent desirable or undesirable facts
that influence sentiment analysis.

For example, words like valley and mountain are neutral in general but may carry
negative sentiment in the mattress review domain (e.g., “A valley has formed in
the middle of the mattress”). Identifying such implied sentiments is challenging
but essential for accurate sentiment analysis.
Approach for Identifying Implied Sentiments
Zhang and Liu (2011b) proposed a method to detect nouns and noun phrases that
imply sentiment in a specific domain. The approach is based on the observation
that certain aspects inherently carry only one sentiment—either positive or
negative. For example, the phrase “A bad valley has formed” is uncommon
because valley already has a negative connotation in mattress reviews.

Two-Step Identification Process

1. Candidate Identification
• The algorithm analyzes sentiment contexts around each noun aspect.
• If an aspect appears significantly more in either positive or negative
sentiment contexts, it is inferred to have that sentiment orientation.

2. Pruning Based on Sentiment Consistency

If an aspect is modified by both positive and negative sentiment words, it
is likely not an opinionated aspect.
Two dependency relations were used to analyze the structure of sentences:
• Type 1: A sentiment word directly modifies a noun (e.g., "This TV
has a good picture quality.").
• Type 2: Both the sentiment word and the aspect depend on another
word (e.g., "The springs of the mattress are bad.").

Limitations and Future Work

This method was an initial attempt to solve the problem of implied sentiment
detection, but its accuracy remains low. Further research is needed to improve the
approach, especially for distinguishing context-dependent sentiments more
effectively.
Analysis of Comparative Opinions
Comparative opinions express sentiment by comparing two or more entities
rather than directly stating a positive or negative opinion about one. These
opinions differ from regular opinions in both semantic meaning and syntactic
structure (Jindal and Liu, 2006a, 2006b).

For example, a regular opinion states an entity’s quality, such as "The voice
quality of this phone is amazing." In contrast, a comparative opinion compares
entities without explicitly stating good or bad quality, such as "The voice quality
of Nokia phones is better than that of iPhones."

Types of Comparative Opinions

Comparative opinions can be classified into opinionated and non-opinionated
sentences. A sentence like "Nokia phones have better voice quality than iPhones."
expresses an opinionated comparison, while "iPhone is 1 inch wider than a normal
Nokia phone." is a non-opinionated comparison since it states a fact without
sentiment.

Comparative and Superlative Opinions

There are two main types of opinions based on comparison:
1. Comparative Opinions – Compare two entities (e.g., "Pepsi tastes better than
Coke.").
2. Superlative Opinions – Indicate the highest or lowest degree (e.g., "This is the
best smartphone on the market.").

Both types are often formed using comparative or superlative adjectives and
adverbs, though this is not always the case. Since their semantic meanings and
analysis techniques are similar, they are generally studied together under
comparative opinions.
Importance of Comparative Opinion Analysis
Comparative opinions play a crucial role in sentiment analysis as they help
understand consumer preferences and competitive positioning of products. Due
to their structural and contextual differences from regular opinions, they require
specialized analysis techniques to accurately extract and interpret sentiment.

Analysis of Comparative Opinions: Problem Definition

A comparative sentence expresses similarities or differences between two or more
entities. Comparisons can be classified into gradable and non-gradable
comparisons (Jindal and Liu, 2006a; Kennedy, 2005).

Types of Comparisons

1. Gradable Comparison
Gradable comparisons rank entities based on shared aspects and have three
subtypes:
• Non-equal gradable comparison – Expresses a ranking relationship,
e.g., "Coke tastes better than Pepsi."
• Equative comparison – Indicates equality, e.g., "Coke and Pepsi taste
the same."
• Superlative comparison – Ranks one entity above all others, e.g.,
"Coke tastes the best among soft drinks."

2. Non-Gradable Comparison
Non-gradable comparisons describe relationships without ranking. These
include:
• Similarity or difference, e.g., "Coke tastes differently from Pepsi."
• Different aspects between entities, e.g., "Desktops use external
speakers, but laptops use internal speakers."
• Presence or absence of an aspect, e.g., "Nokia phones come with
earphones, but iPhones do not."
Comparative Keywords

Comparisons are often expressed using comparative and superlative words like
better, best, more, most, less, and least. These words can be regular (e.g., longer
→ longest) or irregular (e.g., better → best). Other words like prefer and superior
also indicate comparisons and are treated similarly.
Comparative keywords in non-equal gradable comparisons are further classified
into:
• Increasing comparative – Indicates a higher degree (e.g., more, longer).
• Decreasing comparative – Indicates a lower degree (e.g., less, fewer).

Objective of Comparative Opinion Mining

The goal of comparative opinion mining (Jindal and Liu, 2006b; Liu, 2010) is to
extract comparative opinion sextuples from a document:

(E1, E2, A, PE, h, t)

Where:
E1, E2 = Entity sets being compared
A = Aspect being compared
PE = Preferred entity
h = Opinion holder
t = Time of opinion

For example, in "Canon’s picture quality is better than those of LG and Sony,"
written by Jim on 9-25-2011, the extracted sextuple is:

(\{Canon\}, \{LG, Sony\}, \{picture\_quality\}, \{Canon\}, Jim, 9-25-2011)

Challenges in Comparative Opinion Mining
Mining comparative opinions involves extracting entities, aspects, opinion
holders, and time. Unlike regular opinions, comparative sentences require
identifying the preferred entity set and distinguishing whether a sentence is truly
comparative. Many comparative keywords do not always indicate comparisons,
making sentence classification an important challenge.

Identify Comparative Sentences

Not all sentences containing comparative and superlative keywords (e.g., better,
best, superior) are genuinely comparative sentences. For example, "I cannot agree
with you more." contains more but does not express a comparison.

Jindal and Liu (2006a) found that almost every comparative sentence includes a
keyword that indicates comparison. Using a set of such keywords, 98% recall was
achieved, though precision was only 32%.

Types of Comparative Keywords

1. Comparative adjectives (JJR) and adverbs (RBR) – e.g., more, less,
better, and words ending in -er.
2. Superlative adjectives (JJS) and adverbs (RBS) – e.g., most, least, best,
and words ending in -est.
3. Other indicative words and phrases – e.g., favor, prefer, beat, win,
superior, number one, outperform.

Since keywords alone provide high recall, they can filter out non-comparative
sentences, leaving only those requiring further precision improvement.

Pattern-Based Learning for Sentence Classification

Comparative sentences exhibit strong syntactic patterns involving comparative

keywords. Jindal and Liu (2006a) used Class Sequential Rule (CSR) mining, a
special type of sequential pattern mining, to discover these patterns. Using Naïve
Bayes, they built a classification model with high-probability CSR patterns as
features.

Yang and Ko (2011) applied a transformation-based learning algorithm to classify

comparative sentences in Korean.

Classification of Comparative Sentences

Once identified, comparative sentences are categorized into four types:
1. Non-equal gradable comparisons (e.g., "Coke is better than Pepsi.").
2. Equative comparisons (e.g., "Coke and Pepsi taste the same.").
3. Superlative comparisons (e.g., "Coke is the best among all soft drinks.").
4. Non-gradable comparisons (e.g., "Coke tastes different from Pepsi.").

Jindal and Liu (2006a) found that SVM classifiers using keywords and
keyphrases performed best for classification.

Identifying Comparative Questions

Li et al. (2010) focused on identifying comparative questions (e.g., "Which city

is better, New York or Chicago?"). Their method used sequential patterns to detect
comparison structures. They applied weakly supervised learning and
bootstrapping to extract compared entities (comparators) directly from text.

Other information extraction techniques can also be applied to improve

comparative sentence identification and classification.

Identifying Preferred Entities

Unlike regular opinions, comparative opinions do not express direct positive or
negative sentiment but rank multiple entities based on their shared aspects.
Instead of sentiment classification, the goal is to identify the preferred entity set
in a comparative opinion sentence. While a comparative sentence can still be
classified as opinionated or non-opinionated, little research has focused on this
aspect.

Lexicon-Based Approach for Identifying Preferences

Ding, Liu, and Zhang (2009) and Ganapathibhotla and Liu (2008) extended the
lexicon-based approach used for regular sentiment classification to comparative
opinions. Comparative opinion words are categorized into two types:

1. General-Purpose Comparative Sentiment Words

These include Type 1 comparatives like better, worse, which generally
indicate domain-independent sentiment.
For Type 2 comparatives (formed using more, less, most, least), the
preference is determined by both the comparative word and the sentiment
word. The following rules apply:
• Comparative Negative = Increasing comparative + Negative word
(e.g., more awful).
• Comparative Positive = Increasing comparative + Positive word
(e.g., more efficient).
• Comparative Negative = Decreasing comparative + Positive word
(e.g., less amazing).
• Comparative Positive = Decreasing comparative + Negative word
(e.g., less terrible).

2. Context-Dependent Comparative Sentiment Words

• Words like higher, lower, longer can have different sentiment orientations
based on context. For example, "Nokia phones have longer battery life than
Motorola phones." implies a positive sentiment for Nokia but a negative
sentiment for Motorola. However, without domain knowledge, it is hard to
determine whether longer is positive or negative for a specific aspect.
• In Type 2 comparatives, the comparative word, adjective/adverb, and
aspect all play a role in determining preference.
Identifying Preferred Entities Using External Data

To determine whether a comparative opinion is positive or negative,

Ganapathibhotla and Liu (2008) used a large corpus of Pros and Cons from
product reviews. The method checks whether an aspect and sentiment word pair
appears more frequently in Pros or Cons. If found more in Pros, the sentiment is
positive; otherwise, it is negative.

Since Pros and Cons rarely use comparative words, these words are first
converted to their base forms using WordNet and English comparative formation
rules. The assumption is that if a base-form adjective or adverb is positive (or
negative), its comparative and superlative forms also carry the same sentiment
(e.g., good → better → best).

Determining the Preferred Entity

Once sentiment words and their orientations are identified, the preferred entity
set is determined using comparison structures:
• If the comparative is positive, the entity before "than" is preferred.
• If the comparative is negative, the entity after "than" is preferred.

This method ensures that sentiment analysis of comparative opinions accurately

reflects preference ranking among entities.

Opinion Summarization and Search: Aspect-Based Opinion Summarization

In sentiment analysis, summarizing opinions from multiple users is essential since
relying on a single opinion may not be sufficient. Aspect-based opinion
summarization (Hu and Liu, 2004) provides a structured way to analyze opinions
by categorizing them based on entities, aspects, and sentiments. This method is
widely used in industry, including by Microsoft Bing and Google Product Search.
Difference from Traditional Summarization
Unlike traditional text summarization, which extracts key sentences from
documents, opinion summarization focuses on entities and aspects, capturing
both qualitative opinions and quantitative data (e.g., number of positive and
negative sentiments). This approach provides a more structured summary than
traditional methods.

Characteristics of Aspect-Based Opinion Summarization

1. Captures opinion targets (entities and aspects) and sentiments – This ensures
that opinions are summarized based on what users care about.
2. Includes quantitative analysis – It provides numbers or percentages of people
expressing positive or negative opinions on each aspect, helping users understand
overall sentiment trends.

Structured Opinion Summaries

Aspect-based summaries organize opinions using opinion quintuples (entity,
aspect, sentiment, opinion holder, time). For example, a digital camera review
summary might include:
GENERAL (Overall camera opinion) → 105 positive, 12 negative reviews
Picture quality → 95 positive, 10 negative reviews
Battery life → 50 positive, 9 negative reviews
These summaries can be visualized using bar charts or other graphical tools,
enabling easy comparison between entities.

Applications and Insights

Opinion summarization has been applied in various domains, such as movies,
service reviews, and Chinese text summarization. It also helps businesses identify
customer segments. For example, young consumers may focus on design and
performance, while families may emphasize back seats and trunk space in car
reviews.

With opinion quintuples, businesses can track sentiment trends over time and
apply data mining techniques for deeper insights, improving decision-making
based on customer preferences.

Improvements to Aspect-based Opinion Summarization

Several researchers have proposed enhancements to the basic aspect-based
opinion summarization to improve organization, readability, and informativeness.
These improvements integrate traditional summarization techniques, ontology-
based structuring, and advanced optimization methods.

Integration with Sentence Selection and Generation

Carenini, Ng, and Pauls (2006) integrated sentence selection and sentence
generation techniques into aspect-based summarization. Their method:
• Extracts aspect expressions from reviews.
• Maps them to an ontology tree and assigns sentiment scores to aspects.
• Uses centroid-based ranking to select representative sentences.

The sentence selection method produced varied language and details, while
sentence generation provided a better sentiment overview.

Ontology-Based Summarization
Tata and Di Eugenio (2010) structured song reviews using ontology trees. They:
 • Selected focused representative sentences mentioning fewer aspects.
• Ordered sentences logically using domain ontologies to maintain
coherence.

Similarly, Lu et al. (2010) structured online reviews by:

• Selecting high-impact aspects using frequency and sentiment coverage.
• Ordering aspects based on their natural appearance in reviews for
coherence.

Blog Opinion Summarization

Ku, Liang, and Chen (2006) summarized blog opinions using:
• Brief summaries – Selecting headlines from documents with the highest
sentiment.
• Detailed summaries – Listing highly opinionated sentences from reviews.

Sentiment-Aware Summarization
Lerman, Blair-Goldensohn, and McDonald (2009) introduced three models for
summarizing product reviews:
1. Sentiment Match (SM) – Extracts sentences to match the average
sentiment score.
2. Sentiment Match + Aspect Coverage (SMAC) – Balances sentiment and
aspect coverage.
3. Sentiment-Aspect Match (SAM) – Ensures both aspect and sentiment
alignment.
User evaluations showed SAM performed best, though not significantly better
than the others.

Optimization-Based Summarization
Nishikawa et al. (2010b) proposed a text summary technique considering
informativeness and readability by:
• Maximizing aspect-sentiment frequency for informativeness.
 • Ensuring natural sentence flow for readability.
• Using integer linear programming to optimize summary structure.

Ganesan, Zhai, and Han (2010) developed a graph-based abstractive

summarization, while Yatani et al. (2011) focused on extracting adjective-noun
pairs for concise summaries.

These advancements refine aspect-based summarization by improving sentence

selection, organization, readability, and sentiment alignment, making opinion
summaries more structured and meaningful.

Contrastive View Summarization

Contrastive view summarization focuses on summarizing opinions by
highlighting opposing viewpoints on the same aspect. For example, one reviewer
may say, “The voice quality of iPhone is really good,” while another states, “The
voice quality of my iPhone is lousy.” Such direct contrasts help readers quickly
understand diverse perspectives.

Extracting Contrastive Sentence Pairs

Kim and Zhai (2009) introduced a method to extract contrastive sentence pairs
from positive and negative sentence sets. A contrastive pair consists of two
sentences about the same aspect but with opposite sentiments. Their approach
used optimization techniques and similarity functions to select k sentence pairs
that best represent both sentiment sets.

Macro and Micro Multi-View Summarization

Paul, Zhai, and Girju (2010) developed an algorithm for multi-view
summarization:
1. Macro Multi-View Summary – Groups sentences into different opinion
sets, showing varied perspectives.
2. Micro Multi-View Summary – Extracts contrastive sentence pairs,
directly juxtaposing opposing opinions.
Their approach used topic modeling to mine aspects and sentiments and applied
a random walk algorithm (similar to PageRank) to rank representative and
contrastive sentences.

Contrastive Summarization in News and Product Reviews

Park, Lee, and Song (2011) applied contrastive summarization to news articles,
identifying opposing viewpoints on current events.

Lerman and McDonald (2009) focused on comparing opinions across two

products by jointly modeling their summaries to emphasize differences in
opinions.

Importance of Contrastive Summarization

Contrastive summarization enhances sentiment analysis by highlighting opposing
views, making it easier to:
• Compare competing products.
• Understand polarized opinions in reviews.
• Identify bias and contrasting viewpoints in news articles.

By using optimization, topic modeling, and ranking methods, contrastive

summarization provides a balanced, structured way to analyze conflicting
opinions.

Traditional Summarization
Traditional opinion summarization focuses on generating short text summaries
without explicitly considering aspects, entities, or sentiments. These summaries
extract important sentences from reviews, conversations, or articles without
structuring them based on opinion targets.
Approaches to Traditional Summarization
• Beineke et al. (2003) proposed a supervised learning method to identify
and extract key sentences from reviews.
• Seki et al. (2006) introduced a paragraph-clustering algorithm to group and
extract important sentences.
• Wang and Liu (2011) applied extractive summarization in conversations,
considering not only sentence ranking but also topic relevance, sentiment,
and dialogue structure.

Limitations of Traditional Summarization

1. Limited consideration of aspects and sentiments – These summaries may
include sentences unrelated to sentiment or key aspects of interest.
2. Lack of quantitative insight – Traditional summarization does not show
the proportion of positive vs. negative opinions, which is crucial in
sentiment analysis. For example, knowing that 5 out of 10 people dislike a
product is more meaningful than just seeing a single negative statement.

Conclusion
While traditional summarization methods help extract key information, they are
less effective for opinion analysis, as they do not structure content based on
entities, aspects, or sentiment distribution. Aspect-based summarization provides
a more informative and quantitative alternative for analyzing opinions.

Opinion Search - Existing Opinion Retrieval Techniques

Opinion Search and Retrieval

As Web search has become an essential service, opinion search is also gaining
importance. Opinion search helps users find public sentiment on various topics,
products, or individuals.
Types of Opinion Search Queries
1. Finding Public Opinions – Users search for customer reviews or public
sentiment about a specific entity or its aspects. Example: "Find opinions
on the picture quality of a digital camera."

2. Finding Opinions of Specific Individuals or Organizations – Users

search for the views of a person or organization on a topic. Example: "Find
Barack Obama’s opinion on abortion." This type is especially useful in
news analysis, where opinion holders are explicitly mentioned.

How Users Formulate Opinion Queries

• For public opinion search, users enter the entity name or an aspect of
interest (e.g., "best battery life smartphones.").
• For opinion holder search, users include the person/organization name
along with the topic (e.g., "Elon Musk's views on AI regulation.").

Conclusion
Opinion search enables users to quickly access public sentiment and expert
opinions, making it a valuable tool for decision-making and analysis in various
fields.

Existing Opinion Retrieval Techniques

Opinion retrieval is typically a two-stage process. First, documents are ranked
based on topical relevance. Then, relevant documents are re-ranked using opinion
scores, which can be obtained using either:
• Machine learning-based classifiers (e.g., SVM).
• Lexicon-based classifiers that analyze sentiment words and their proximity
to query terms.

Some advanced techniques combine topic and sentiment relevance into a single
ranking score for improved accuracy.
Opinion Search System Example
Zhang and Yu (2007) developed an opinion retrieval system, which won the 2007
TREC blog track. The system consists of two components:
1. Retrieval Component – Performs traditional information retrieval (IR)
by considering keywords and concepts (e.g., named entities, Wikipedia
entries). It expands queries using synonyms and pseudo-feedback to
improve search relevance.
2. Opinion Classification Component – Classifies retrieved documents as
opinionated or not using an SVM-based classifier trained on review and
factual data. Opinionated documents are further categorized as positive,
negative, or mixed.

Other Opinion Retrieval Approaches

• Eguchi and Lavrenko (2006) used generative language modeling to
integrate topic relevance and sentiment orientation, improving retrieval
accuracy.
• Huang and Croft (2009) combined topic and opinion relevance models to
rank documents.
• Na et al. (2009) introduced a lexicon-based approach with query-specific
sentiment lexicons, adapting to different domains.
• Li et al. (2010) used a graph-based approach (HITS algorithm) to rank
documents by linking topic-sentiment word pairs with the documents that
contain them.

Re-Ranking in Opinion Retrieval

Pang and Lee (2008) proposed a simple re-ranking method for review search,
using an idiosyncrasy measure that prioritizes rare terms in search results. This
method does not rely on a sentiment lexicon.

Conclusion
Opinion retrieval techniques have evolved from basic document ranking to
advanced models integrating topic and sentiment relevance. These methods
enhance search accuracy, helping users find relevant, sentiment-rich content more
effectively.
UNIT-IV
Identifying intention, fake and quality of opinion: Detecting Fake or Deceptive
Opinions: Different Types of Spam - Supervised Fake Review. Detection -
Supervised Yelp Data Experiment - Automated Discovery of Abnormal Patterns
– Model Based Behavioral Analysis - Group Spam Detection - Identifying
Reviewers with Multiple User ids - Exploiting Business in Reviews - Some
Future Research Directions. Quality of Reviews: Quality Prediction as a
Regression Problem - Other Methods - Some New Frontiers.
Opinion Spam Detection
Opinions from social media influence purchase decisions, elections, marketing,
and product design. Since positive opinions can lead to profits and fame, some
individuals and organizations attempt to manipulate public perception by posting
fake reviews or opinions without revealing their true motives. This deceptive
practice is known as opinion spamming, and those who engage in it are called
opinion spammers. Opinion spam can be particularly dangerous when it
influences social and political issues, as it can mislead the public and create false
narratives.

Opinion spam detection is a major challenge because, unlike other types of spam
(such as email or web spam), fake reviews and opinions are difficult to identify
just by reading them. Unlike link spam, which manipulates hyperlinks, or content
spam, which stuffs irrelevant keywords, opinion spam involves fabricated or
misleading statements that appear legitimate. Additionally, since opinion spam
often blends with genuine reviews, detecting it requires analysis beyond just the
text itself.

One of the key challenges in opinion spam detection is the lack of a clear way to
distinguish between real and fake reviews by simply reading them. For example,
a positive review written for a good restaurant can be copied and posted as a fake
review for a bad restaurant, making it impossible to determine its authenticity
without external data. This means that effective opinion spam detection must rely
on additional information such as user behavior, review patterns, and metadata
rather than just textual analysis.

While much of the existing research focuses on opinion spam in consumer

reviews, there is limited study of spam detection in other forms of social media.
As opinion spamming becomes more widespread and sophisticated, developing
advanced detection techniques is crucial to maintaining the integrity of online
opinions and preventing the spread of misinformation.

Types of Spam and Spamming

Opinion spam in online reviews can be categorized into three main types: fake
reviews, brand-only reviews, and non-reviews. These types were identified by
Jindal and Liu (2008) in their study on spam detection. While some forms of spam
are easier to identify, fake reviews pose the greatest challenge.

Type 1 (Fake Reviews): Fake reviews are deceptive opinions written with hidden
motives rather than genuine experiences. They are often used to promote certain
products or services by posting undeservedly positive reviews or to harm
competitors by posting false negative opinions. This type of spam is the most
harmful as it misleads consumers and manipulates public perception.

Type 2 (Reviews About Brands Only): These reviews focus on a brand or

manufacturer rather than the specific product or service being reviewed. While
they may be genuine, they are considered spam because they do not provide
useful product-specific feedback and are often biased. For example, a review of
an HP printer that simply states, “I hate HP. I never buy any of their products” is
not a useful product review.

Type 3 (Non-Reviews): These include advertisements, irrelevant content, or

random discussions that do not provide an actual opinion about the product or
service. Since they do not contribute to consumer decision-making, they are
categorized as spam, though they are relatively easy to detect.

Harmful Fake Reviews

Not all fake reviews are equally harmful. Some are written to promote a product,
while others aim to damage its reputation. Fake reviews that exaggerate the
quality of a bad or average product (regions 3, 4, and 5 in Table 10.1) are
particularly harmful, as they mislead consumers into making poor decisions.
Similarly, negative fake reviews that unfairly damage the reputation of a good or
average product (regions 2 and 4) can harm businesses. Detection algorithms
should focus on identifying these harmful types of fake reviews.

Individual and Group Spamming

Fake reviews can be written by individuals or groups. Individual spammers work
alone, using a single user ID, such as an author writing fake reviews for their own
book. Group spammers, however, coordinate their efforts to manipulate sentiment
on a product. They may be a network of people working together or a single
person using multiple accounts (sock puppeting). Group spamming is particularly
damaging as it can completely control the perception of a product, especially
during its launch.

Types of Data, Features, and Detection

Opinion spam detection relies on three main types of data:
1. Review Content: Linguistic features such as word usage and sentence
structure can help detect fake reviews, but they are not always sufficient
since fake reviews can be carefully crafted to resemble genuine ones.
2. Meta-Data: Information such as review timestamps, user activity, IP
addresses, and location data can reveal suspicious patterns. For example, if
multiple positive reviews come from the same computer or location, they
may be fake.
3. Product Information: Comparing review sentiment with product sales data
can reveal inconsistencies. For instance, if a poorly selling product has an
unusually high number of positive reviews, it raises suspicion.
Effective spam detection aims to identify fake reviews, fake reviewers, and
coordinated spammer groups. Since these are interconnected, detecting one type
can aid in identifying the others. While detecting individual fake reviews is
challenging, analyzing reviewer behavior and group patterns can improve
detection accuracy.

Supervised Spam Detection

Opinion spam detection can be approached as a classification problem, where
reviews are categorized as either fake or non-fake. Supervised learning is a
natural method for this task, but the lack of reliable labeled data poses a
significant challenge. Since fake reviews can be carefully crafted to resemble
genuine ones, manually identifying them is difficult, making it hard to train
machine learning models effectively.

Use of Duplicate Reviews

Jindal and Liu (2008) addressed this challenge by analyzing duplicate and near-
duplicate reviews. In a study of 5.8 million reviews from Amazon, they found
four types of duplicates:
 1. Same user-ID, same product (possibly accidental)
2. Different user-IDs, same product (likely fake)
3. Same user-ID, different products (likely fake)
4. Different user-IDs, different products (likely fake)
Since the last three types of duplicates were highly suspicious, they were used as
labeled fake reviews for training. Machine learning models were then built using
review-centric, reviewer-centric, and product-centric features. Logistic
regression was applied, leading to several key findings:
• Negative outlier reviews (extremely low ratings) were more likely to be
fake.
• Reviews that were the only ones for a product were often fake.
• Highly ranked reviewers were sometimes spammers, as some wrote
thousands of reviews.
• Fake reviews could receive positive feedback, misleading readers.
• Lower-selling products were more frequently targeted by spam.
Despite these findings, the results were tentative, as not all duplicate reviews are
necessarily fake, and many fake reviews are not duplicates.

Alternative Supervised Approaches

Li et al. (2011) used manually labeled reviews from Epinions, where users rated
reviews and left comments. They applied Naïve Bayes classification using
various features such as subjectivity, positivity, negativity, reviewer profiles, and
authority scores. Their semi-supervised approach also considered the tendency of
spammers to write multiple fake reviews.

Ott et al. (2011) used Amazon Mechanical Turk to crowdsource fake hotel
reviews. Turkers were asked to write promotional reviews as if they worked for
the hotels, while truthful reviews were taken from TripAdvisor. They
experimented with several classification techniques and found that unigram and
bigram-based text classification performed best. However, their dataset had
limitations, as Turkers were not actual spammers, and the 50/50 class distribution
did not reflect real-world conditions.
Overall, supervised spam detection shows promise, but the lack of high-quality
labeled data remains a challenge. Many studies rely on imperfect datasets, which
can affect the accuracy of detection models.

Group Spam Detection

Group spam detection focuses on identifying groups of reviewers who collaborate
to manipulate product ratings. Mukherjee et al. (2011) proposed an initial
algorithm, which was later improved by Mukherjee, Liu, and Glance (2012).
Their approach detects spammer groups in two main steps:

1. Frequent Pattern Mining: This step analyzes review data to identify

patterns of reviewers who frequently review the same set of products. Since
professional spammers often write multiple reviews for various products
to maximize profit, this method helps uncover collusion among them. A
group of reviewers who have only collaborated once is harder to detect, but
those who repeatedly work together across multiple products can be
flagged.

2. Ranking Groups Based on Spam Indicators: Not all identified groups are
actual spammers. To refine the results, various behavioral indicators are
used, such as reviewing products within a short time frame, writing reviews
right after a product launch, having similar review content, and showing
rating deviations. A relational model called GSRank (Group Spam Rank)
was developed to rank candidate spammer groups based on these
indicators. An iterative algorithm was then applied to improve accuracy.

Strengths and Limitations

The method was evaluated using manually labeled spammer groups and showed
promising results. However, one limitation is that the frequency threshold used in
pattern mining may miss groups that have worked together only a few times (less
than three).
This approach is unsupervised, meaning it does not rely on pre-labeled training
data. Although supervised learning methods were tested, they were found to be
less effective in this context. The group detection method remains a useful tool
for identifying coordinated spam activities in online reviews.
Quality of Reviews
The quality of reviews is an important aspect of opinion analysis but differs from
spam detection. Low-quality reviews may not necessarily be spam, and fake
reviews may still be perceived as high-quality if they are well-crafted. The goal
of this research is to determine the usefulness, helpfulness, or utility of reviews,
allowing platforms to rank and display the most informative ones first. Many
review sites, like Amazon, allow users to rate the helpfulness of reviews, but
automating this process is valuable since feedback accumulation takes time.

Quality as a Regression Problem

Review quality is often treated as a regression problem, where a model predicts
a quality score for ranking reviews. User feedback on helpfulness is commonly
used as ground truth data. Researchers have explored various features to train
regression models:

• Structural features: Review length, sentence count, presence of questions

or exclamations, and HTML formatting tags.
• Lexical features: Unigrams and bigrams weighted using tf-idf.
• Syntactic features: Percentage of nouns, verbs, adjectives, and adverbs in
the review.
• Semantic features: Sentiment words and product aspect mentions.
• Meta-data features: Review ratings (star scores).

Different studies have used variations of these features, with some also
incorporating reviewer expertise, timeliness of reviews, and writing style to
improve accuracy.

Social Context in Review Quality

Lu et al. (2010) introduced the idea of social context in determining review
quality. Their model incorporated:
• Author consistency: A single reviewer tends to produce reviews of similar
quality.
• Trust consistency: If a reviewer trusts another, the latter’s reviews should
be of equal or higher quality.
 • Co-citation consistency: If two reviewers are trusted by the same person,
their quality should be similar.
• Link consistency: Connected users in a social network likely produce
reviews of similar quality.

This approach was tested on Ciao, a social review platform where users can trust
and follow others, making it an effective method for sites with social networks.

Other Approaches
Several alternative methods have been proposed:
• Classification-based methods: Some studies classify reviews as "helpful"
or "non-helpful" using content, social, and sentiment features.
• Addressing biases in helpfulness votes: Liu et al. (2007) argued that early
reviews and highly ranked reviews receive more votes, which may not
always reflect quality. They manually labeled reviews into categories and
trained an SVM classifier using features like informativeness, subjectivity,
and readability.
• Unsupervised ranking: Tsur and Rappoport (2009) introduced a method
where reviews are compared to a virtual "core review" and ranked based
on similarity.
• Personalized review quality prediction: Moghaddam et al. (2012) proposed
factorization models, arguing that review helpfulness varies by user
preference.
• Diversity in review selection: Some researchers suggested that top-ranked
reviews should not be redundant and should cover different aspects and
viewpoints of a product.

Conclusion
Determining review quality is a complex task that combines text analysis, user
feedback, social interactions, and ranking algorithms. While many approaches
have been explored, challenges remain, such as handling bias in user votes and
ensuring diverse, informative reviews are displayed first.