An Introduction to

Sentiment Analysis

Ashish Katrekar
AVP, Big Data Analytics

Sentiment analysis and opinion mining have

become an integral part of the product
marketing and user experience as both
businesses and consumers turn to online
resources for feedback on products and
services. This white paper explores the
evolution and challenges of sentiment anaysis,
as well as how to best leverage it.

www.globallogic.com
An Introduction to Sentiment Analysis Ashish Katrekar, AVP, Big Data Analytics

Table of Contents
Introduction .................................................................................................................................................................................................................. 3

Sentiment Classification Levels ........................................................................................................................................................................... 3

Document Level Classification ........................................................................................................................................................... 3
Sentence Level Classification ............................................................................................................................................................ 3

Converting Unstructured Text into Structured Opinions .............................................................................................................................. 4

Aspect Based Sentimental Analysis .................................................................................................................................................................... 4

Aspect Extraction ..................................................................................................................................................................................... 5
Sentiment Classification ........................................................................................................................................................................ 5
Finding Sentiments for Aspects ........................................................................................................................................................ 5

Conclusion .................................................................................................................................................................................................................... 6

References ................................................................................................................................................................................................................... 6

GlobalLogic Inc. www.globallogic.com 2

An Introduction to Sentiment Analysis Ashish Katrekar, AVP, Big Data Analytics

Introduction Sentiment Classification Levels

The opinions of others have a significant influence in Sentiment analysis can occur at different levels:
our daily decision-making process. These decisions document level, sentence level or aspect/feature level.
range from buying a product such as a smart phone to
making investments to choosing a school—all decisions Document Level Classification
that affect various aspects of our daily life. Before the
Internet, people would seek opinions on products and In this process, sentiment is extracted from the entire
services from sources such as friends, relatives, or review, and a whole opinion is classified based on the
consumer reports. overall sentiment of the opinion holder. The goal is to
classify a review as positive, negative, or neutral.
However, in the Internet era, it is much easier to collect
diverse opinions from different people around the world. Example
People look to review sites (e.g., CNET, Epinions.com), “I bought an iPhone a few days ago. It is such a nice
e-commerce sites (e.g., Amazon, eBay), online opinion phone, although a little large. The touch screen is cool.
sites (e.g., TripAdvisor, Rotten Tomatoes, Yelp) and social The voice quality is clear too. I simply love it!”
media (e.g., Facebook, Twitter) to get feedback on how
a particular product or service may be perceived in the Is the review classification positive or negative?
market. Document level classification works best when the
document is written by a single person and expresses an
Similarly, organizations use surveys, opinion polls, opinion/sentiment on a single entity.
and social media as a mechanism to obtain feedback
on their products and services. sentiment analysis or Sentence Level Classification
opinion mining is the computational study of opinions,
sentiments, and emotions expressed in text. The use of This process usually involves two steps:
sentiment analysis is becoming more widely leveraged
because the information it yields can result in the • Subjectivity classification of a sentence into one of
monetization of products and services. two classes: objective and subjective
For example, by obtaining consumer feedback on a • Sentiment classification of subjective sentences into
marketing campaign, an organization can measure the two classes: positive and negative
campaign’s success or learn how to adjust it for greater
success. Product feedback is also helpful in building An objective sentence presents some factual
better products, which can have a direct impact on information, while a subjective sentence expresses
revenue, as well as comparing competitor offerings. personal feelings, views, emotions, or beliefs. Subjective
sentence identification can be achieved through different
This white paper will describe the various types of methods such as Naïve Bayesian classification. However,
sentiment classification, explore how to convert just knowing that sentences have a positive or negative
unstructured text into structured opinions, and address opinion is not sufficient. This is an intermediate step that
the current challenges in the field. helps filter out sentences with no opinions and helps
determine to an extent if sentiments about entities and
their aspects are positive or negative. A subjective
sentence may contain multiple opinions and subjective
and factual clauses.

GlobalLogic Inc. www.globallogic.com 3

An Introduction to Sentiment Analysis Ashish Katrekar, AVP, Big Data Analytics

Example Example
“iPhone sales are doing well in this bad economy.” Review from XYZ on 7/8/2013 - “I purchased a Galaxy 5S
phone. It is a great phone overall. The screen resolution
Sentiment classification at both the document and is cool and has a good battery life.”
sentence levels are useful, but they do not find what
people like or dislike, nor do they identify opinion targets. This opinion can be expressed by the following
quintuples:
Aspect/Feature Level Classification (Galaxy 5S, screen, +, XYZ, 7/8/2013)
(Galaxy 5S, battery, +, XYZ, 7/8/2013)
In this process, the goal is to identify and extract object
features that have been commented on by the opinion Quintuples form the basis for opinion summarization. The
holder and determine whether the opinion is positive, objective of the quintuple is to convert the unstructured
negative, or neutral. Feature synonyms are grouped, data to a more structured form that can be used
and a feature-based summary of multiple reviews is for further analysis. The initial step is to discover all
produced. quintuples and find the five attributes required by the
quintuplet. Once the data is in a more structured form,
it is much easier to analyze and perform sentiment
Converting Unstructured Text into analysis. Once the quintuples have been extracted, they
Structured Opinions can be fed to visualization and analysis tools.

Consumer sentiments are mainly expressed in an Aspect Based Sentimental Analysis

unstructured format. Text analysis involves taking
unstructured data and finding ways to convert it into a (ABSA)
more structured format that facilitates analysis of data
and getting deeper insights into the sentiments. There ABSA is based on identifying aspects of given target
are various techniques that can be used to convert entities and estimating the sentiment polarity for each
unstructured data into a structured format. mentioned aspect. This can be decomposed into
two tasks: aspect extraction and aspect sentiment
One such way is to express an opinion as a quintuple (ej, classification.
ajk, soijkl, hi, tl), where:
ej is a target entity Aspect extraction pertains to recognizing aspects of the
ajk is an aspect/feature of the entity ej entity, and more generally can be seen as an information
soijkl is the sentiment value of the opinion from extraction task. Aspect sentiment classification
the opinion holder determines whether the opinions on different aspects
soijkl is +ve, -ve, or neutral, or more granular are positive, negative or neutral.
ratings
hi is an opinion holder While lexicon-based approaches use a list of aspect-
tl is the time when the opinion is expressed related sentiment phrases as the main resource, the key
issue for learning methods is to determine the scope of
An entity e is a product, person, event, organization, or each sentiment expression, if it covers the aspect in the
topic and is represented as an hierarchy of components, sentence.
sub-components, and so on. Each node is represented
as a component and is associated with a set of attributes
for the component.

GlobalLogic Inc. www.globallogic.com 4

An Introduction to Sentiment Analysis Ashish Katrekar, AVP, Big Data Analytics

Aspect Extraction

To identify all the aspect terms present in a sentence, all MPQA Subjective Cues Lexicon (http://www.cs.pitt.edu/
highly frequent phrases across reviews (e.g. food) should mpqa/subj_lexicon.html)
be found and filtered by rules like “occurs right after
sentiment word” (e.g. great food). Then a set of phrases Bing Liu’s Opinion Lexicon (http://www.cs.uic.edu/~liub/
that occur frequently can be built. Another approach is FBS/opinion-lexicon-English.rar)
to determine all the aspects in advance and find them in
the reviews. For a restaurant, the aspects could be: food, SentiWordNet (http://sentiwordnet.isti.cnr.it/)
service, value, décor.
The online lexicons may not be sufficient because
Example they may not have enough words or do not pertain to
“The food was great, but the service was slow.” the domain or topic of discussion. In such a case, a
Aspects: food, service new sentiment lexicon can be built. This can be done
using techniques such as a semi-supervised lexicon,
Sentiment Classification which leverages a small amount of information (e.g., a
few labeled examples or a few hand-built patterns) to
Words express various kinds of sentiments that may bootstrap a complete lexicon through the learning of
be positive, negative, strong, or weak. To perform lexicons. In a bootstrapping approach, a high-precision
sentiment analysis, it is important to understand the classifier is first used to identify some subjective and
polarity of words and classify sentiments into categories objective sentences. A set of patterns is learned from
such as positive, negative, or neutral. This task can be these identified subjective and objective sentences.
accomplished through the use of sentiment lexicons. The learned patterns are then used to extract more
There are different types of sentiment lexicons available subjective and objective sentences. This process can
that have words classified as having positive or negative then be repeated until the desired lexicon is built.
sentiments.
Finding Sentiments for Aspects
Examples
The General Inquirer (http://www.wjh.harvard. For this scenario, let’s analyze reviews for a particular
edu/~inquirer) restaurant. One customer posted the review, “The food
was great, but the service was slow.” The figure below
LIWC (Linguistic Inquiry and word count) (http://www. shows system components for finding sentiments for
liwc.net) aspects.

Fig 1. System Overview: Finding Sentiments for Aspects

Sentences &
Sentences Sentences & Final
Reviews Phrases with
& Phrases Phrases with Summary
Sentiments &
Sentiments
Aspects Labeled

Text Sentiment Aspect Aggregator

Extractor Classifier Extractor & Summarizer

GlobalLogic Inc. www.globallogic.com 5

An Introduction to Sentiment Analysis Ashish Katrekar, AVP, Big Data Analytics

Finding sentiments for aspects can be divided into four

subtasks:
Conclusion
Sentiment analysis is an evolving field with a variety of
1. Text extraction: extract sentences and phrases from use applications. Although sentiment analysis tasks are
the reviews challenging due to their natural language processing
origins, much progress has been made over the last
2. Sentiment classification: run a sentiment classifier on few years due to the high demand for it. Not only do
each extracted sentence and phrase to determine if companies want to know how their products and
it is positive, negative, or neutral services are perceived by consumers (and compare to
competitors), but consumers want to know the opinions
3. Aspect extractor: perform aspect term extraction for of others before making buying decisions.
the ones that express a sentiment and determine the
polarity for each of the aspects identified (e.g., food, The growing need for product insights – and the
service) technical challenges currently facing the field –will
keep sentiment analysis and opinion mining relevant
4. Aspect polarity aggregation: group the sentiments for for the foreseeable future. Next-generation opinion
aspects together and produce a final summary (e.g., mining systems need a deeper bind between complete
food: positive, service: negative) knowledge bases with reasoning methods inspired by
human thought and psychology. This will lead to a better
understanding of natural language opinions and will
Challenges in Sentiment Analysis more efficiently bridge the gap between unstructured
information in the form of human thoughts and
However, there are still some challenges to overcome
structured data that can be analyzed and processed by a
before sentiment analysis can become a more perfect
machine.
tool. For example, human judgment is still far more
accurate as a gauge in sentiment analysis. Automated
The result: intelligent opinion mining systems capable
systems cannot differentiate sarcasm from sincere
of handling semantic knowledge, making analogies,
text, nor can they always correctly analyze the specific
continuous learning, and detecting emotions — leading
contextual meaning of a word. Use of acronyms like
to highly efficient sentiment analysis.
“lol” or word abbreviations also pose interpretation
challenges.
References
Furthermore, mixed opinions such as “I like the printer
quality, but the size is too big,” can be difficult to classify, 1. Sentiment Analysis and Opinion Mining by Bing Liu
as is identifying genuine feedback in a general comment (http://www.cs.uic.edu/~liub/FBS/SentimentAnalysis-
like, “I surprised my wife on her birthday with this new and-OpinionMining.html)
Samsung phone, and she just freaked out!” It’s also
unlikely that an automated system could identify biased 2. Sentiment Analysis by Professor Dan Jurafsky
or fake reviews on a product or service. (https://web.standford.edu/class/cs124/lec/
sentiment.pdf)

3. S. Blair-Goldensohn, Hannan, McDonald, Neylon,

Reis and Reynar 2008 – Building a Sentiment
Summarizer for Local Service Reviews (http://www.
ryanmcd.com/papers/local_service_summ.pdf)

GlobalLogic Inc. www.globallogic.com 6

About GlobalLogic Inc.
GlobalLogic is a full-lifecycle product development services leader
that combines deep domain expertise and cross-industry experience
to connect makers with markets worldwide.Using insight gained
from working on innovative products and disruptive technologies,
we collaborate with customers to show them how strategic research
and development can become a tool for managing their future. We
build partnerships with market-defining business and technology
leaders who want to make amazing products, discover new revenue
opportunities, and accelerate time to market.

For more information, visit www.globallogic.com

Contact

Emily Gunn
Marketing Communications
[email protected]