Sentimental Analysis is a process of computing and categorizing the expressed opinions of people about certain event, subject or product as positive, negative or neutral. The major objective of Sentimental Analysis is to help data-driven decisions using insights from replies in social media, surveys and product reviews. Sentiment Analysis can be done with words, sentences, documents, features or aspects, concepts, phrases, links, clauses and implications. Recently, there has been a lot of attention on sentiment analysis especially from researchers in the fields of text mining and natural language processing. But due to extreme absence of annotated datasets which are used to train models in various domains, the accuracy of sentiment analysis has been hindered. Many types of research have been done to confront the challenge and enhance sentiment analysis classification. Sentiment analysis is important as it helps in identifying the emotional and attitude states of people. Positive or negative feelings of people can be expressed in different ways. This research article talks about, in subtle terms, the different ways to deal with sentiment analysis mostly in Machine Learning, Lexicon-based, Hybrid and Ontology-based approaches. This research article gives point by point perspective of the distinctive applications and challenges of Sentiment Analysis.
Sentimental Analysis is a process of determining whether the text, document and media content as positive, negative or neutral. The foremost objective of Sentimental Analysis is to help data-driven decisions using perceptions from responses in social media, surveys and product reviews. Sentiment Analysis can be done with words, sentences, documents, features or aspects, concepts, phrases, links, clauses and implications. The bag of words recovered from word cloud is used for word-level Sentiment Analysis. At the sentence level Sentiment Analysis is done with sentences in reviews and comments written by users. A document-level analysis is done by classifying the opinions expressed in an entire document into different sentiments. Features can be used likewise for opinion mining and Sentiment Analysis. Feature level classification is done by identifying and extracting the different product features from raw source data. The feature analysis is done when a desired sentimental aspect or feature is to be got from a review. This article is structured as follows: section-2.presents background information related to the survey. Section-3.presents the related works conducted on various aspects of sentimental analysis. Section-4.presents the methodology of the survey conducted. Detailed discussion on open issues and challenges of sentiment analysis is presented in section-5. Finally, this article is concluded in section-6.
Machine Learning (ML) techniques have a variety of features that help in the construction of classifiers for the expressed sentiments in texts. ML approaches make use of ML algorithms for Sentimental Analysis to extract the linguistic and syntactic features and form standard classifications of those features. Another popular method used for sentimental analysis is Deep-Learning which makes use of neural network techniques.
Supervised learning: Supervised learning is a technique that makes use of labelled datasets to create a model. A variety of supervised models can be created depending on the type of procedure that is employed to create the model.
Decision tree classifiers: Decision tree classifiers prepare the information space in different levels of decreasing order that distinguishes data based on the estimation of their credit. A recursive procedure is followed while creating the data division space that will result in the leaf hubs having the base quantities of the records used as the end goal of characterization.
Linear classification: Linear classifiers are of different types. The Sustenance Vector machines are one kind of linear classifiers that make use of distinct direct separators for classification. Support Vector Machines (SVM) which is used in neural networks makes use of supervised learning method to create decision planes that provide the decision boundaries specification. Decision planes consist of sets of objects that have different class membership. A linear classifier is a line that separates the sets of objects classified according to their corresponding spheres. If the partitioning is more than two dimensions, the classifier is a curve that is known as the hyperplane classifier. The SVM produces 0 or 1 which is positive or negative from the organized input in a vector space using its portrayed information. The content that is archived in a unique space cannot be used for learning. It has to be configured for calculation using ML algorithms. The prepreparation of the content archive involves the transformation of every word to measurement and indistinguishable words will relate to the same measurement. For content classification, SVM has been proved to be useful for viable learning calculations.
Rule-based Technique When an ‘if-then’ rule is applied to a relationship that consists of an antecedent and its corresponding consequent, then it is a Rule-based Technique.
An antecedent pronounces a state and can be symbolized as a single token or several tokens that are connected by the “∧” operator. A token can be either “?” which symbolizes a proper noun or a word or the token can be “#” that denotes the consequence of the state described by the antecedent.
{token1 ∧ token2 ∧ ... ∧ token} ⇒ {+|−} The three simple rules A, B and C relates to words describing three sentiments, each denoting an antecedent.
{Good} ⇒ {Positive}… … (A) {Bad} ⇒ {Negative}…..… (B) {Ok} ⇒ {Neutral}……..… (C)
Probabilistic classifiers: Probabilistic classifiers also known as the generative classifier makes use of a mix of models that generatively examines each section for a particular term using the mix and classifies the sections accordingly. The three probabilistic classifiers are Naïve Byes, Maximum Entropy and Bayesian Network.
Naive Bayes Classifier (NB): NB is a probabilistic classifier that relies on upon Bayes hypothesis with solid and innocent freedom suppositions. It is a champion among the most principal content order procedures with different applications such as archive classification, email spam discovery, individual email sorting, and slant recognition and dialect identification. It performs better in numerous perplexing genuine issues. The Naive Bayes Variations are Bernoulli Naive Bayes, Binarized Multinomial Naive Bayes and the Multinomial Naive Bayes. Each framework passes on a very surprising result since they use a unique model. When the different events of the words matter a great deal in the characterization issue multinomial naive bayes is utilized. Binarized multinomial naive bayes is used when the regularities of the words don't accept a key part in the arrangement.
Bayesian Network (BN): The Naïve Bayes classifier is the independence of the features. Assumption of Naïve Bayes is to expect that every one of the components is completely dependent. This prompts to the BN which shows a coordinated acyclic graph and whose nodes correspond random variables, and edges that represent conditional dependencies. BN is viewed as an entire factor with their association. In this way, an entire joint probability distribution over each one of the elements is resolved for a model. The computation complexity of the BN is exceptionally costly in text mining. So, BN is used very little. BN is utilized to consider a true issue.
Maximum Entropy (MaxEnt) Classifier: MaxEnt classifiers are feature-based models that are most preferred when it is required to uniformly satisfy a given constraint. Unlike the Naïve Bayes model, MaxEnt does not make any autonomous assumption of the features. So MaxEnt allows the addition of features like bigrams and phrases without the problem of feature overlapping. The MaxEnt principle is explicitly important in the case of information that needs to be examined to determine if a given distribution is consistent with the testable information. MaxEnt or its variations are useful because of their accuracy. Consistency results of the algorithm are another major advantage of it. The ability of the classifier is to work with a huge amount of data describes its performance or its efficiency. When it comes to handling different types of data in a single platform and classify them according to it, the flexibility of the classifier is almost perfect.
Unsupervised Learning: Unsupervised learning technique applies a comparison formula between sentiment values in a lexicon with the components of the text at hand. The words in the lexicon have predefined values and they are applied to similar words in the text. Two most commonly used unsupervised techniques are hierarchical clustering and partial clustering.
Semi-Supervised Learning: Semisupervised learning involves the use of both the unsupervised and supervised techniques in its classification models. The semi-supervised learning makes use of both the labelled and some unlabelled data in its training sets. So the goal of semi-supervised learning could be to either to predict the values of the unlabelled data in the training set in which case it would be called transductive semi-supervised learning or to find the values of the test sets in which case it would be called the inductive semisupervised learning.
This approach depends on sentiment Lexicons. Lexicon is considered as an important indicator for sentiment, which is called opinion word. Lexicon can be separated into the dictionary approach and corpus-based approach
Dictionary-based approach: An arrangement of sentiment word is gathered manually with known instructions. The conclusion set is created by looking in the prominent repository WordNet for their proportionate word and antonyms. The next iteration starts when the words are added into the seed list. In the absence of finding any new words, the iterative process stops. After the system completes, a manual appraisal can be done to evacuate or amend errors. Dictionary approach has an inconvenience, that is, the inability to discover feeling words with space and setting specific introductions.
Corpus-based approach: Corpus-based method deals with the issue of finding sensitivity words with setting specific presentation. Corpus construct techniques depend on syntactic illustrations or examples that are found together with a semblance of assessment words to find other inferencing words in a broad corpus. The objectives are for conjunctions like AND, OR, BUT, EITHER EXOR. The conjunction in AND case that are conjoined descriptive words, for the most part, have the comparative introduction. This contemplation is called notion consistency, which is for the most part not solid for all predictable practicality. There are also adversative expressions on account of BUT they are shown as opinion changes to figure out whether two conjoined descriptors are of the same or distinctive introductions. The corpusbased approach uses factual or semantic techniques to discover assumption extremity.
A hybrid approach is a combination of ML
and Lexicon approaches used in the
Sentimental Analysis. Although the hybrid
approach may not be used frequently, they are
known to produce results better than the
approaches mentioned earlier. [
Ontology applies the hierarchy of concepts
to represent domain knowledge. An ontology
can be defined as an unambiguous,
machinereadable representation of a common
formulation of concepts. An Ontology is used
to represent knowledge formally by modelling
the terms of a particular domain and by
capturing the semantic relationship between the
terms. The relationship between terms is
particularly important for aspect-level
sentiment analysis more specifically in product
reviews since product reviews are usually
qualified by their aspects. Ontological
approaches are generally used to capture the
relation between the product and their
properties in a hierarchical order. Such ordering
is done using definitions of fundamental
Entities and interrelationship between the
entities that can be interpreted by machines. An
ontology model incorporates the entities,
concepts, classes or objects, their properties and
the relation between them. For the sharing of
knowledge among researchers about a
particular domain, the ontology provides
common vocabulary definitions. Such
ontological definitions not only help to share a
common knowledge understanding among
people and software agents but also allows the
reusability of domain knowledge. Storage of
knowledge in an ontology model is done using
either OWL/XML or RDF/XML format. The
merging of two ontology models which have
the same domain knowledge is possible.
Information retrieval from an ontology model is
done by querying it. [
In the concept of sentiment analysis portraits different levels, various approaches, many classifications and evaluation methods are displayed in figure 2.1.
Various sentimental approaches are
surveyed [
The tools for Sentiment Analysis that are
used in different domains were also studied
analyzing its features and performance.
SentiWordNet, LIWC, EMOTIONS,
SenticNet, Happiness Index, AFINN,
PANASt, Sentiment140 Dataset, NRC, EWGA, FRN
[
Twenty four articles are compared in the
survey to study sentiment analysis for its
importance, effects and the challenges in
sentiment evaluation. The sentiment review
structure and the challenges in sentiment
analysis are initially compared. The
domaindependence of the sentiment challenges is an
essential feature that is revealed by the
comparison. The popularity of the negation
challenges in the structure of all the reviews
types differed in its implicit or explicit
meaning. The result of such comparison
provided a capacity to measure the effects of
each sentiment challenge based on the structure
of the review types. The topic, nature and the
structure of the review regulate the appropriate
challenges for the assessment of reviews on
sentiment analysis. The next comparison that is
done between the relevance of the challenges in
sentiment analysis relevant with the accuracy
rate. From this comparison, two things were
evident, namely, the status of the sentiment
challenges for evaluation of the sentiments and
the method of finding the fitting challenge to
improve accuracy. The relation between the
amount of theoretical and technical use of
sentiment techniques to resolve sentiment
challenges is also found. It was also established
Table 4.1
List of research articles searched in different databases.
and explained that the theoretical type of
sentiment challenges is the growing area of
research. By proving that with the growth in
research in a sentiment analysis the average
accuracy has reduced, the inference from the
average of accuracy based on the number of
researches in each challenge has been
established. The compassion circle could be
extended to new research literature for further
work. [
The Preferred Reporting Items for
Systematic Reviews and Meta-Analyses
(PRISMA) are a method of reporting a
minimum set of items grounded on shreds of
evidence in systematic reviews and
metaanalyses. The PRISMA flow diagram in figure
4.2 exhibits how information moves along the
different phases in systematic reviews. The
number of records identified, included and
excluded, and the reasons for exclusions are
charted out in the PRISMA flow diagram. [
The following procedures of the systematic literature review were carried out to survey the current study and research in and around the subject of sentiment analysis.
The search process included the search for pertinent research literature on the internet through websites, online libraries and databases like Springer, IEEE Xplore, Web of Science, Scopus, Science Direct, ACM Digital Library, Elsevier and Google Scholar, etc.. (Listed in Table 4.1). The vital terms or strings that were used for search are ‘‘survey on sentiment analysis”, ‘‘survey on opinion mining”, and “Art of survey on sentimental analysis”. It was found that searching the references of the research articles of the various studies with the three key terms yielded more articles.
Databases
IEEE Xplore
Web of Science Scopus
Scholar /Proceedings ACM Digital Elsevier Library 4 2 2 2 2004 - 2010 2011 - 2013
The number of literature available on sentimental analysis is quite vast. To set the number of literature to that which can be managed or those with a focus, certain criteria were devised in the selection of the literature for review.
i) Criterion for Including – Only those literature on sentiment analysis or on tools and techniques for sentiment analysis that was published during the period 2005–2019 (mentioned in figure 4.2) are considered. If the same or similar content is published in more than one journal or conference proceeding, the most complete version of the literature was chosen.
ii) Criteria for Excluding – Works published in unknown conferences or journals, research articles that were not relevant to above the keyword search or not relevant to the survey and ‘white’ articles as mentioned in table 4.2.
The search process described above was
used to identify the relevant literature and these
were then checked by their title and abstract
using the criteria of including or excluding
literature. Once all the relevant research articles
that were identified for the review process, they
were downloaded extracting their data and to
analyse them further. In Figure 4.2 is a
PRISMA flowchart to illustrate the process of
the search that was done for the selection for
literature for review [
and Spams - Language Dependency - Detection of Fake Reviews - Lexicon
and Spams - Language Dependency - Detection of Fake Reviews - Lexicon
and Spams - Limitations in Classification
Filtering - Domain Dependency - Domain Dependency - Lexicon - Comparisons - Negations - Sarcasm - Detection of Fake Reviews - Lexicon
and Spams - Language Dependency - Detection of Fake Reviews - Lexicon
and Spams - Language Dependency - Detection of Fake Reviews - Lexicon
and Spams - Domain Dependency - Sarcasm - Interrogative Sentences - Sentiment without sentiment
words - Conditional sentences - Detection of Fake Reviews - Lexicon
and Spams - Language Dependency - Detection of Fake Reviews - Lexicon
and Spams - Language Dependency - Detection of Fake Reviews - Lexicon
and Spams - Language Dependency
Having studied the contents of the literature on
Sentimental Analysis, the following issues and
challenges were evident:
Finding Fake Reviews and Spams: Internet
contents especially in social media comprises
both authentic and spam contents. Therefore, it
is necessary to remove the spam and fake
contents before pre-processing. [
Machine Learning Approach – different approaches proposed for sentiment analysis (like Lexicon based, Machine learning based and Ontology-based), but no implementation based on multimedia content using machine learning.
There has been a lot of attention recently on sentiment analysis especially from researchers from the fields of text mining and natural language processing. But due to extreme absence of annotated datasets which is used to train models in various domains, the accuracy of sentiment analysis has been hindered. Several kinds of research have been done out to confront the challenge and enhance sentiment analysis classification. Sentiment analysis is important as it helps in identifying the emotional and attitude states of people. Positive or negative feelings of people can be expressed in different ways. This research article talks about, in subtle terms, the different ways to deal with sentiment analysis mostly in Machine Learning, Lexicon-based, Hybrid and Ontology-based approaches. This research article gives a point by point perspective of the distinctive applications and challenges of Sentiment Analysis.