Volume 7, Issue 9, September – 2022 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Aspect Based Sentiment Analysis:

Approaches and Algorithms
SOWMYA B DR. T. JOHN PETER
ASSISTANT PROFESSOR, PROFESSOR & HEAD,
DEPT OF COMPUTER SCIENCE & ENGINEERING, DEPT OF COMPUTER SCIENCE
EAST WEST COLLEGE OF ENGINEERING & ENGINEERING,SAMBHRAM INSTITUTE OF
YELAHANKA NEW TOWN, BANGALORE, TECHNOLOGY M.S.PALYA, BANGALORE,
KARNATAKA 560064, INDIA KARNATAKA 560097, INDIA

Abstract:- The domain of Aspect-based Sentiment analysis is exhibited at various granularities like sentence,
Analysis allows automatic extraction of the fine-grained aspect and the document as a whole.
sentiment information from text documents or sentences.
Aspect based sentiment analysis is one among the three Further Machine learning and Deep learning
main types of sentiment analysis , where aspects are approaches are used to carry out the process of sentiment
extracted , sentiments are analyzed and are evolved over analysis. Hence, in this paper, we mainly focus on
time , is getting much attention with increasing reviews presenting the various tasks of aspect-based sentiment
of customers and peoples on social media. The evolution analysis that have used machine learning and deep learning
of machine learning and deep learning algorithms has approaches for the task. To carryout Aspect Based
made noticeable mark towards aspect level sentiment Sentiment Analysis , aspect identification is the crucial step.
analysis. This survey emphasizes a review of different The aspects are classified into two types , implicit and
works from the recent articles on aspect based sentiment explicit [1]. for example " The cost of room in hotel is high ,
analysis using several machine learning techniques. but the rooms are very clean " . Here "cost" is the aspect for
"hotel" and the polarity is negative towards hotel. In the
Keywords:- Aspect-Based Sentiment Analysis ; Fine-grained above example, cost is an explicit aspect and clean is an
Sentiment Analysis ; Machine Learning ; Deep Learning. implicit aspect. It's not easy to extract implicit and explicit
aspects so some models use Rule Based Methods[9] based
I. INTRODUCTION on machine learning concepts like semantic similarities[2] ,
SVM algorithm[3] , Conditional Random Fields .
Currently, sentiment analysis is a dynamic research
area due to rapid growth of internet and swift participation II. ASPECT BASED SENTIMENT ANALYSIS (ABSA)
for sharing, commenting and discussing over social media,
forums, blogs and shopping websites. Sentiment Analysis Aspect Based Sentiment Analysis is a fine grained
plays an important role for government sectors, business technique that categorizes data by aspect and identifies the
platforms, manufacturers to know the impact of their sentiment of various aspects of an entity within a textual
products. The statistics of the year 2020 unleashed a jaw data. An entity is a single identifiable object , it can be
dropping figure of 700 million tweets in a year which can be anything like a place, movie , hotel , an individual or a
considered as 8000 tweets per second by signifying Twitter product or service. For example, let us consider a review : “I
as an active social platform. Customers and people share am using this laptop since one month , I love this laptop; the
their feelings and opinions in the form of review for a touch display is beautiful, but processing speed is not as
product or service which results in a collection of huge expected .”, laptop is the entity and display and processing
amount of data on the Internet. This unstructured data speed are the aspects. Beautiful , love , not as expected are
contains lots of unused information which can be efficiently the sentiments towards the entity and its aspect. An ABSA
extracted through Sentiment Analysis . A piece of text can for a sentence of this type must give positive sentiment for
easily change the mind-set of a prospective buyer about the the aspect ‘display ' , and negative sentiment for the aspect
product or service. It is highly impossible and unwise idea to 'processing speed'.
process every review comment posted by numerous
customers manually. Sentiment Analysis is one of the fast- Similar to sentiment Analysis , Aspect Based
growing research areas in Natural Language Processing and Sentiment Analysis also follows multistage analysis. Figure
is the perfect solution to analyze the trend purchase 1 shows the working of Aspect Based Sentiment Analysis ,
behavior. Sentiment analysis is the process of extracting in which first step is to preprocess the text data to remove
opinions from a piece of text and classifying them based on extraneous words.
the polarity like positive, negative or neutral. Sentiment

IJISRT22SEP907 www.ijisrt.com 1668

Volume 7, Issue 9, September – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig. 1: Work flow of Aspect Based Sentiment Analysis

In pre-processing , the collected data must be made to grammatical relationship between aspects and opinions , that
a suitable format and then it can be further processed for the can be used to find implicit aspects.
specific task. In Sentiment Analysis , the data will go
through sequence of processing like tokenization , stop word B. Frequency-based or statistical approach
removal , negation handling and so on to clean and convert The implicit feature extraction can be taken place in
the data into a suitable format. The process of converting three different steps. first is co-occurrence matrix , it will
the text into a series of tokens is called tokenization. It show the relationship between opinion words and product
helps in vector formation and eliminating unwanted words aspect. in second step the possible features set can be built
from the text data. Negation handling is one of the important and can check for presence of an implicit feature . In third
step in preprocessing , in which the actual polarity of the step , correct implicit feature can be detected by considering
text will be opposite of the normal outcome , if it is not the scores of possible implicit features which is calculated
examined properly. Punctuations , special characters and by checking opinion words and context information of
numerical tokens are considered as useless and are removed implicit feature. Implicit aspects can be easily identified
, as they do not convey any sentiment or aspect in the textual using a supervised approach[4]. They use review dataset for
data. After tokenization second step to be carried out is products and restaurant. The list and respective frequencies
stemming , in which we tries to find out base word of the of implicit aspects and unique lemmas are generated first
given word in textual data. Word embedding is carried out and then the score is generated i.e.,the ratio of co-occurrence
after the pre processing and is a very critical factor of of each word and the frequency of word for each implicit
sentiment analysis . Word embedding is the process of aspect is evaluated. The score of the aspect which is greater
converting the token of words into a vector format. words than the defined threshold is identified.
like 'weather' and 'whether' sounds similar but there is a
difference in meaning. word embedding's are used to make a A classifier is used to predict the occurrence of
machine understand the difference in meaning and will multiple implicit features which are built using a score
convert the text into some other dimension. further , the function. Basically the score function is based on the
vectors will be sent to a machine learning model to extract number of nouns , adjectives , commas and total number of
aspect and sentiment from the word. In third level the 'and' words. The function parameters are estimated using
aspects of corresponding entity of the given text are logistic regression . By using the prediction of above
identified and further the words that define the sentiment of classifier the feature detection part of an algorithm checks
recognized aspects are identified. In the final step the for one or more implicit features .
polarity of word sentiment is identified accurately. Aspect C. Pattern based approach
Based Sentiment Analysis tasks are further divided into two There is a close association between explicit and implicit
categories , i.e., aspect category sentiment analysis and aspects [10]. Instead of separately examining explicit and
aspect term sentiment analysis [15].Aspect category implicit aspect, all the correlated aspects can be represented
sentiment analysis is the coarse-grained level of extraction in a network-like structure. A framework called Aspect
and the second is the fine-grained level of extraction of the Frame Net is proposed for sentiment analysis. The aspect
aspects. Scalability is one of the main advantages of Aspect pattern in review text can be read by the system and the
Based Sentiment Analysis. Because at fine-grained level the patterns of the aspects and the sentiments for these aspects
textual data can be analyzed easily by Aspect Based are aggregated . Further supervised and unsupervised
Sentiment Analysis. techniques are proposed [16]. In unsupervised approach an
III. APPROACHES FOR ASPECT EXTRACTION association rule mining is applied on co-occurrence
frequency data to find aspect categories.
Many researchers have used different approaches for
aspect extraction that includes supervised, unsupervised, D. Hybrid approach
semi-supervised and hybrid. supervised approach is most Wordnet dictionary based and corpus-based methods for
commonly used in wide variety of domains than extraction of implicit aspect terms can be combined using
unsupervised and semi supervised approaches. Hybrid approach . Implicit Aspect Representation, learning
model enhancement and implicit aspect identification are the
A. Unsupervised rule-based approach different phases where this approach can be worked. In first
Poria.S [9] implemented a technique for both implicit phase all adjectives from the list of words are extracted and
and explicit aspect extraction from opinionated text. A rule then all the relative adjectives for each aspect from training
based approach presented a detection of explicit aspects data are extracted. In second phase a Naïve Bayes classifier
using common-sense knowledge and sentence dependency is trained for detecting implicit aspects which is taken using
trees. Zainuddin, Nurulhuda, Ali Selamat, and Roliana implicit aspect representation from phase 1 . In final phase
Ibrahim [11] used dependency parser to identify

IJISRT22SEP907 www.ijisrt.com 1669

 Volume 7, Issue 9, September – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
the naïve Bayes classification is tested for all terms of
aspects. For aspect identification and separation of opinion-
words we can use combination of LDA topic modeling and
IV. MACHINE LEARNING ALGORITHMS FOR ABSA an unsupervised pre-trained classification model. An Aspect
Sentiment Unification Model (ASUM)[12] is designed
Machine Learning algorithms will work accurately on which is a modified version of LDA that consolidates both
Aspect Based Sentiment Analysis approach. Following are aspect and its respective sentiment.
the different machine learning algorithms and their
suitability for Aspect Based Sentiment analysis. B. Conditional Random Field (CRF)
Conditional Random Field is a discriminative model and
A. Latent Dirichlet Allocation (LDA). is used for predicting sequences. For more accurate
Latent Dirichlet Allocation is a generative statistical predictions in CRF information obtained from the prior label
model that allows a set of observations to be explained as is used. predictive framework is proposed for identifying the
combinations of latent topics. LDA is a topic modeling rating of non-rated reviews from squeal dataset [20]. A
technique which helps in automatically finding the sentiCRF is a variant of CRF which is used for pair term
fundamental topics from a given document. In LDA generation and to find their sentiment scores. Cumulative
documents are considered as mixture of topics with words logic model is introduced to predict the rating of a review
having specific probabilities. Kull-back-Leibler (KL) which takes aspects and their respective sentiment values
divergence [5] method is used for finding the association from the reviews. A heuristic re-sampling algorithm is
between theme model and the given paragraph. LDA model proposed to solve the class imbalance problem during
is used to the blogs theme on the other hand KL divergence sentiment score estimation.
method is used to identify the distance among the themes.
Further a topic modeling tool named Mallet [21] is used in C. Support Vector Machine (SVM)
which LDA is considered to determine the aspect and latent In machine learning Support Vector Machines are a
information from a manually collected reviews. W2VLDA promising supervised machine learning model that analyze
[22] is an unsupervised approach which mainly deals with data for classification and regression analysis. In SVM each
multi-domain and multilingual ABSA. It uses a large and every data item can be plotted on an n-dimensional
amount of unlabelled data and the initial configuration can graph and based on the task of classification or regression a
be made with minimum set of seed words. hyper-plane will be drawn.

Fig. 2: Support Vector Machine Representation

For aspect term extraction on large movie reviews [7] D. Convolutional Neural Network (CNN)
five different feature selection algorithms that includes SVM Convolutional Neural Network is a deep learning
,Naive Bayes are compared and identified that SVM is technique earlier it was used in image processing techniques
giving the best result with the Gini index. The Gini SVM and now which is used in almost all the areas. The given
combines entropy and Kernel Based Model that normalizes input to the CNN model will pass through many
classification margins and conditional probability is given convolutional layers with filters in each layer called kernel .
directly , which is computationally less intensive . In Feature Basically kernel is smaller than image but is more in depth.
Based SVM instead of considering sentiment extraction it is Finally softmax function is used to conclude the final value
better to show the importance of a feature. A hybrid to a probabilistic value, the value must be between 0 and 1.
sentiment classification approach[8] is proposed in which convolutional layer is the very first layer which extracts all
SVM is used in combination with an Association Rule the features from the given input. Non-Linearity problem in
Mining technique. Principal Component Analysis, Latent convolutional neural network (ConvNet) can be tackled
Sentiment Analysis etc , are feature selection methods which using Rectified Linear Unit for a non-linear operation
are applied with a combination of heuristic parts of speech (ReLu). Compared to other classification algorithms the pre-
and is also used for the extraction of aspects. processing required is much lower in a ConvNet. A Deep
Learning method is combined with rule based method that

IJISRT22SEP907 www.ijisrt.com 1670

Volume 7, Issue 9, September – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
enhance the task of Aspect Based Sentiment Analysis and workshop on natural language processing for social
also the performance of sentiment scoring method is media (SocialNLP) (pp. 28-37)
improved. A seven-Layer CNN has been used for the aspect [10.] Chatterji, S., Varshney, N. and Rahul, R.K. (2017)
identification from restaurant reviews. The Multichannel “AspectFrameNet: a frameNet extension for analysis
framework[17] using CNN is proposed in which every of sentiments around product aspects”The Journal of
single input representation is controlled by a single CNN. Supercomputing, 73(3), pp.961-972.
[11.] Zainuddin, Nurulhuda, Ali Selamat, and Roliana
V. CONCLUSION Ibrahim. "Hybrid sentiment classification on twitter
aspect-based sentiment analysis." Applied Intelligence
This survey aims to provide a comprehensive review (2018): 1-15
of the Aspect Based Sentiment Analysis technique , [12.] Y. Jo and A. Oh, “Aspect and sentiment unification
including various tasks , approaches , algorithms and model for online review analysis,” Proc. 4th ACM Int.
potential directions. we first set up the description of ABSA Conf. Web Search Data Mining, WSDM 2011, pp.
research with some elements , the definition and common 815–824, 2011.
modeling approaches. Then we describe each ABSA [13.] S. M. Rezaeinia, R. Rahmani, A. Ghodsi, and H. Veisi,
approaches with a recent advances of the compound ABSA. “Sentiment analysis based on improved pre-trained
It is difficult to detect implicit aspect but it is very important word embeddings,” Expert Syst. Appl., vol. 117, pp.
. Due to advancement in technology , deep learning methods 139–147, 2019.
and other algorithms have given some promising results for [14.] B. Zhang, X. Xu, X. Li, X. Chen, Y. Ye, and Z. Wang,
ABSA. But sometimes results are not up to the expectation “Sentiment analysis through critic learning for
so more methods or algorithms are yet to be described. optimizing convolutional neural networks with rules,”
REFERENCES Neurocomputing, vol. 356, pp. 21–30, 2019.
[15.] H. H. Do, P. W. C. Prasad, A. Maag, and A. Alsadoon,
[1.] Schouten, K. and Frasincar, F. (2016) “Survey on “Deep Learning for Aspect-Based Sentiment Analysis:
aspect-level sentiment analysis” IEEE Transactions on A Comparative Review,” Expert Syst. Appl., vol. 118,
Knowledge and Data Engineering, 28(3), pp.813-830 pp. 272–299, 2019.
[2.] P. D. Turney, “Thumbs up or thumbs down? semantic [16.] Schouten, Kim, et al. "Supervised and unsupervised
orientation applied to unsupervised classification of aspect category detection for sentiment analysis with
reviews,” in ACL, 2002, pp. 417–424. co-occurrence data." IEEE transactions on cybernetics
[3.] M.Sivakumar, Dr.U.Srinivasulu Reddy, “Aspect Based 48.4 (2018): 1263-1275.
Sentiment Analysis of Students Opinion using [17.] D. H. Pham and A. C. Le, “Exploiting multiple word
Machine Learning Techniques”, Proceedings of the embeddings and one-hot character vectors for aspect-
International Conference on Inventive Computing and based sentiment analysis,” Int. J. Approx. Reason., vol.
Informatics (ICICI 2017) IEEE Xplore. 103, pp. 1–10, 2018.
[4.] Schouten, K. and Frasincar, F. (2014) “Finding [18.] X. Ma, J. Zeng, L. Peng, G. Fortino, and Y. Zhang,
implicit features in consumer reviews for sentiment “Modeling multi-aspects within one opinionated
analysis” International Conference on Web sentence simultaneously for aspect-level sentiment
Engineering (pp. 130-144). Springer, Cham analysis,” Futur. Gener. Comput. Syst., vol. 93, pp.
[5.] X. Fu, G. Liu, Y. Guo, and W. Guo, “Multi-aspect 304–311, 2019.
blog sentiment analysis based on LDA topic model [19.] F. Tang, L. Fu, B. Yao, and W. Xu, “Aspect based
and Hownet lexicon,” Lect. Notes Comput. Sci. fine-grained sentiment analysis for online reviews,”
(including Subser. Lect. Notes Artif. Intell. Lect. Notes Inf. Sci. (Ny)., vol. 488, pp. 190–204, 2019.
Bioinformatics), vol. 6988 LNCS, no. PART 2, pp. [20.] J. Qiu, C. Liu, Y. Li, and Z. Lin, “Leveraging
131–138, 2011. sentiment analysis at the aspects level to predict
[6.] Nipuna Upeka Pannala, Chamira Priyamanthi ratings of reviews,” Inf. Sci. (Ny)., vol. 451–452, pp.
Nawarathna, J.T.K.Jayakody, “Supervised Learning 295–309, 2018.
Based Approach to Aspect Based Sentiment [21.] N. Akhtar, N. Zubair, A. Kumar, and T. Ahmad,
Analysis”, 2016 IEEE International Conference on “Aspect based Sentiment Oriented Summarization of
Computer and Information Technology. Hotel Reviews,” Procedia Comput. Sci., vol. 115, pp.
[7.] A. S. Manek, P. D. Shenoy, M. C. Mohan, and K. R. 563–571, 2017.
Venugopal, “Aspect term extraction for sentiment [22.] A. García-Pablos, M. Cuadros, and G. Rigau,
analysis in large movie reviews using Gini Index “W2VLDA: Almost unsupervised system for Aspect
feature selection method and SVM classifier,” World Based Sentiment Analysis,” Expert Syst. Appl., vol.
Wide Web, vol. 20, no. 2, pp. 135–154, 2017. 91, pp. 127–137, 2018.
[8.] N. Zainuddin, A. Selamat, and R. Ibrahim, “Hybrid
sentiment classification on twitter aspect-based
sentiment analysis,” Appl. Intell., vol. 48, no. 5, pp.
1218–1232, 2018.
[9.] Poria, S., Cambria, E., Ku, L.W., Gui, C. and Gelbukh,
A. (2014) “A rule-based approach to aspect extraction
from product reviews” In Proceedings of the second

IJISRT22SEP907 www.ijisrt.com 1671