IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 2, June 2024, pp. 1668~1678

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i2.pp1668-1678  1668

Aspect based sentiment analysis using a novel ensemble deep

network

Abraham Rajan, Manohar Manur

Department of Computer Science and Engineering, Christ University, Bangalore, India

Article Info ABSTRACT

Article history: Aspect-based sentiment analysis (ABSA) is a fine-grained task in natural
language processing, which aims to predict the sentiment polarity of several
Received Apr 5, 2023 parts of a sentence or document. The essential aspect of sentiment polarity and
Revised Nov 11, 2023 global context have deep relationships that have not received enough
Accepted Dec 1, 2023 attention. This research work design and develops a novel ensemble deep
network (EDN) which comprises the various network and integrated to
enhance the model performance. In the proposed work the words of the input
Keywords: sentence are converted into word vectors using the optimised bidirectional
encoder representations from transformers (BERT) model and an optimised
Aspect-based sentiment BERT-graph neural networks (GNN) model with convolutions is built that
analysis analyses the ABSA of the input sentence. The optimised GNN model with
Ensemble deep network convolutions for context-based word representations is developed for the
Graph neural network word-vector embedding. We propose a novel EDN for an ABSA model for
Sentiment analysis optimised BERT over GNN with convolutions. The proposed ensemble deep
network proposed system (EDN-PS) is evaluated with various existing
techniques and results are plotted in terms of metrics for accuracy and F1-
score, concluding that the proposed EDN-PS ensures better performance in
comparison with the existing model.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Abraham Rajan
Department of Computer Science and Engineering, Christ University
Bangalore, India
Email: [email protected]

1. INTRODUCTION
Through blogs, discussion forums, social networking sites, and e-commerce websites, users have
expressed their opinions on a wide range of topics, including products, individuals, and organizations, during
the past several years. This swapping of ideas is made possible by the fast development of web applications
and the broad availability of inexpensive internet connections, both of which generate massive amounts of
online data. Sentiment analysis uses this large quantity of data that is used to analyses major challenges. Before
the widespread use of the internet as a resource for decision-making, due to the widespread use of the internet
and the world wide web, it is now possible to read the opinions and perspectives of anonymous individuals
from a range of cultures and places. Many users' shopping opinions are impacted by internet reviews by various
other shoppers [1].
Systems that analyses user sentiment will automatically summarize user reviews, which may help a
customer, make an opinion. Major advantages of sentiment analysis systems include scalability, the ability to
summarize massive volumes of text, real-time analysis, and the capability to deliver fast assessments, and
consistent criteria, which is the result of automation and lacks bias compared to user judgment. Organizations,
whether public or private, require a sentiment analysis system; it is feasible to supplement standard
recommendation systems using sentiment analysis. It assists the owners of the organization in understanding

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938  1669

the sentiments of customers towards their items, however, this is suitable for market research and competition
analysis, politics, public policy development, and legal research are further uses [2]. Although sentiment
analysis may be undertaken at various levels, aspect-level sentiment analysis is the most useful. Researchers
have conducted an extensive study on the identification of explicit sentiments that result in the generation of
many alternatives.
Due to its complexity, embedded characteristic recognition receives less attention; this is
simultaneously a substantial portion of a sentence, which has various parameters, to detect sentiment analysis
in understanding the concept. Sentiment analysis research focuses on making predictions at the phrase or
document level by assessing the overall sentiment towards the entire phrase or content [2]–[4]. The opinion is
based on the perhaps inaccurate assumption that the available data determines a single opinion towards a
specific issue. In consequence, the need for aspect-based sentiment analysis (ABSA) is a method for
differentiating more complex aspect-level thoughts and feelings, which has surged over the past decade [5],
[6]. The target achieved for opinion mining on ABSA shifts its focus from being a phrase or document on a
whole to being an entity or a particular characteristic of an entity. For example, an entity may be an
online-sold product, with its features including its price and size. Refer to the following example review to
understand the words used in aspect-level sentiment analysis: "the display quality of the Samsung TV is
outstanding." The term outstanding in this context denotes a favorable opinion of the Samsung TV's display,
which is the element of the Samsung TV that is being discussed. As illustrated in Figure 1, the three methods
that may be employed to perform aspect-level sentiment analysis are aspect identification, sentiment evaluation
concerning the specific aspect, and sentiment aggregation.

Figure 1. Aspect level sentiment analysis

Context-based methods and syntax-based methods are the two main areas of focus in the existing
research for ABSA based on deep neural networks [5], [6]. To gather aspect word representations in a phrase,
context-based approaches often employ convolutional neural networks (CNN) or recurrent neural networks in
conjunction with attention processes. Graph convolutional networks (GCN) are frequently used to represent
phrases in syntax-based models. Recent research on ABSA has relied widely on deep neural models due to
their capability to capture the semantic features of specific aspects [1]–[4]. In the existing research on ABSA
that uses deep neural networks, context-based techniques [5], [6] and syntax-based methods [7]–[9] are
distinguished into two groups. Context-based algorithms frequently combine CNN or recurrent neural networks
with attention mechanisms to collect aspect word representations in a sentence, in syntax-based models, GCN
are typically employed to represent sentences.
Finding several aspect-level sentiment components, including sentiment polarities, opinion terms,
aspect words, and aspect categories, is the focus of this research work. Present ABSA models that have been
developed to achieve better performance in a variety of tasks, which focuses on training and testing data, which
is developed from the same distribution (e.g., the same domain or the same language). When the data
distribution changes, it is usually important to retrain the ABSA model to maintain its performance. Huge
quantities of labelled data are often too expensive or even impossible to collect for ABSA applications that
require aspect-level annotations. Incorporating various other methods for building ABSA-based models, which
generalize across various domains and languages, is cross-domain transfer or cross-lingual transfer, which
requires adapting the trained model to unknown domains or unknown languages.
− Convert the words of the input sentence into word-vectors using the optimized bidirectional encoder
representations from transformers (BERT) model.
− To extract the context-based word representation using the optimized graph neural network (GNN) model
with convolutions over the word-vector embedding.
− To develop a BERT-GNN model that analyses the ABSA of the input sentence.
Aspect based sentiment analysis using a novel ensemble deep network (Abraham Rajan)
1670  ISSN: 2252-8938

− To develop a proposed novel ensemble deep network (EDN) for an ABSA model for optimized BERT over
GNN with convolution.
This particular research is organized as follows: first section of the research starts with background
work and significance of sentiment analysis in natural language processing (NLP); section concludes with
research motivation and contribution. Second section of the research discuss various existing deep learning
protocol along with its shortcoming; third section presents mathematical modelling of EDN. Thereafter, EDN
is evaluated in fourth section.

2. RELATED WORK
ABSA has been one of the key area of research in sentiment analysis, thus there has been plenty of
research in past decade. Traditional approach lacks the deep understanding of the common sense among the
sentences. This section discuss several deep learning mechanism for sentiment analysis. Huang and Carley [10]
have developed a novel, reliable CNN for aspect-level sentiment analysis, and integrating this aspect data into
a CNN with the aspect-level features results in developing a high-accuracy model. Long short-term memory
(LSTM) networks are difficult to analyses and have low temporal efficiency, whereas CNN can only allocate
local semantic data. The sparse attention-based convolutional neural network (SA-SDCCN) proposed by [11]
deals with semantic and sentiment embedding that is added to an embedding tensor at the multichannel
embedding layer, which leads to accurate representations of the input sequence.
According to Wang et al. [12], a novel model known as unified position-aware convolutional neural
network (UP-CNN), which initially proposed an aspect detection network with prior knowledge to
accommodate the absence of aspect positions in ACSA, and has proposed an aspect mask to construct
aspect-aware contextual representations to accommodate CNN in ABSA. Consequently, several NLP
applications have successfully integrated GCNs, by using grammatical links in aspect-level sentiment analysis
tasks, Zhang et al. [7] have developed the first aspect-specific graph convolutional network (ASGCN). Initially,
the context information is collected using a network with bidirectional long short-term memory (Bi-LSTM),
then aspect-based word information is collected using the masked approach. The ABSA approach developed
by [13] is based on the convolution of dependency trees.
Zhang and Qian [14] introduced a bidirectional GCN model consisting of a hierarchical syntactic
network and a lexical graph for each phrase based on syntactic links and word co-occurrence data. Based on a
dependency network, Tang et al. [15] suggested a modified double-transformer structure for aspect sentiment
categorization that focuses on essential sentence components. Using relational graph-based attention networks,
Wang et al. [16] have modified the dependent trees. The previously published GCN-based models perform
well on the ABSA; however, they lacked sentiment data across context-related factors. In addition to this the
interactive sentiment correlation between aspects recovered by GCNs on the dependency tree with external
sentiment-based knowledge and integrate with MHSA to generate semantic information between context and
aspects.
According to Zhang et al. [17], transformer-based semantic-primary knowledge transferring (TSPKT)
network for aspect-term sentiment analysis across domains makes use of semantic-primary information to
facilitate the transfer of knowledge across multiple domains. Before semantic-primary information is collected
from the S-Graph, it is generated using external semantic lexicons. The second postulate is that Aoa graphormer
learns syntactically meaningful aspect-term phrases. Thirdly, the standard by LSTM classifier is developed by
adding a knowledge-aware memory unit (KAMU) that holds the whole corpus of semantic primary input.
According to Hu et al. [18], this approach focuses on the above-mentioned two issues by combining aspect and
task-level regularizations. Aspect-level regularizations limit the attention-based weights to reduce noise. Sparse
regularization is more suited for sentences with a single aspect than orthogonal regularization for sentences
with various aspects. Here they have proposed a task-level regularization by adding aspect-level category
identification as an orthogonal auxiliary task to generate sentiment-dominant features.

3. PROPOSED METHODOLOGY
The proposed novel EDN for ABSA model for optimized BERT over GNN with convolutions.
Figure 2 shows the proposed architecture. The pooling layer here is responsible to extract features of similar
size by selecting the maximum number from each feature vector. Further, this connected layer fine-tunes the
model considering the sentiments taken from the previous layer. Finally determining the sentiments from the
softmax layer.
− Transfer the words of the input sentence into words using the optimized BERT model.
− To extract the context-based word representation using the optimized GNN model over the word-vector
embedding.

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1668-1678

Int J Artif Intell ISSN: 2252-8938  1671

− Build a BERT-GNN model with convolutions that analyses the ABSA of the input sentence.
Consider a given set of sentiments D = {d1 , d2, … . . , dn} for l ∈ L, let
Sen = {sen1 , sen2, … . . , senn} be a set of sentences for k sentiments. For sen ∈ Sen be the set of words
sen = {l1 , l2, … . . , ln} for one sentence denoted assen. In the proposed architecture for a novel EDN for
aspect-based sentiment, analysis the input is given in the form of sentiments, which is then converted into a
sentence–word-vector. This is further fed into the optimized BERT model where it is converted into word
embedding by the optimized BERT model. The optimized LSTM model here extract the context-word
representations, the context-based words are fed into the GNN with convolution, which converts the node
matrices into the Attention mechanism wherein the further bias is reduced while, processing the GNN model
with convolutions.

Figure 2. Proposed architecture

3.1. Preliminaries and problem definition

A fixed set of c sentiments D = {d1 , d2, … . . , dn} in context with the target entity, for a target
sentiment dk . Let fxy be the y − th aspect for the given entity, g xy is the essential feature relevant to the aspect
fxy and Sxy determines the sentiment of the aspect fxy . The main aim here is to develop an optimized GNN
model for aspect-level sentiment analysis. The main objective here is subjected to the mapping function as
mentioned in (1). The proposed model incorporates the following steps, to transform the words of the input
provided into word vectors along with the optimized-BERT model.

1, if Sxy is positive
G: (g xy ) = { 0, if Sxy is neutral (1)
−1, if Sxy is negative

3.1.1. Word-vector generation

The word vector model automatically inserts a cls at the start of the sentence as well as two separate
symbols in the initial phase and ending phase of the term to determine the sentiment. The word sequence here
is determined by converting a specific sentence into a separate form v = [cls] + v + [diff] + fx + [diff]. This
sequence of words is then fed into this model to obtain the vector for each word Oi as shown in (2). Here the
Opt − BERT model (Oi ) is the word vector extracted via the pre-trained model, as well as dimo is the dimension
for the word vector. The input sequence is determined from the input sequence through c, the sequence obtained
through the G = {g1 , g 2 , … , g k }.

Aspect based sentiment analysis using a novel ensemble deep network (Abraham Rajan)
1672  ISSN: 2252-8938

g x = Opt − BERT(Oi ) ∈ H dimo ∗1 (2)

3.1.2. Word–sentence–vector converter

The word representations are depicted as each word converted into an aggregated form through the
entire vector that is condensed into the entire input vector sequence. This study deploys an optimized LSTM
model over the word-vector embedding for word-vector context basis word representation. Optimised LSTM
by learning the context information through understanding the meaning of words by reading the input given in
two different ways. Context-based representations are made according to the following steps.
− Step 1: first step is dealing with input layer; the word-vector sequence is represented as G = {g1 , g 2 , … , g k }
given as input to the optimised LSTM model.
− Step 2: second step is dealing with optimized LSTM this layer is responsible to co-ordinate the context-
based information by the remaining words in the direction of the sentence in two different ways for the
specific-target word. The optimized LSTM model incorporates a dual-LSTM model, left-to-right model
(→ Opt − LSTM ) model, and right-to-left model (← Opt − LSTM), which is responsible to encode the
sentence from left-to-right as (→ jx (x = [1, k])) and right-to-left (← jx (x = [1, k])) as shown in (3).

→ jx = (→ Opt − LSTM(Lg→ j g x + g → j → j → j x−1 + e→ j ) ∈ Npj (3)

← jx = (→ Opt − LSTM(Lg← j g x + g ← j ← j → j x−1 + e← j ) ∈ Npj (4)

jx = (L→ j j → jx + L→ j j ← jx + ej ) (5)

Here L is the weight matrix, by considering the Lg→ j amidst the input given and the forward hidden vectors, jx
for representing the context. jx+1 is the hidden vector represented for jx and jk+1 = 0. jx−1 for the previously
hidden vector for jx and j0 = 0.
− Step 3: the context-based word-vector matrix denoted as G = {g1 , g 2 , … , g k }∈ Nk∗pj , here pj depicts the
size of the vector in the matrix G.

3.2. Building a graph neural network with convolutions

− Step 1: build a sentence GNN
A sentence-based GNN with convolution denoted as SGNN = ( N, V, AM) including a set of nodes in
correspondence with the k words in a given sentence, a set E of edges denotes the dependencies of the adjacent
node pairs for syntactic dependency as well as an adjacency matrix AM∈Z |N|∗|N| denoting the node relations
denoted as shown in (6):

1, if Nx , Ny ∈ N, and vxy ∈ V
Fxy = { 1, if Nx = Ny , (6)
0, else

Additionally, this graph SGNN has a feature matrix embedded with nodes denoted as Y = [G] ∈ N|N|∗pj , where
each row is denoted as Gx which is responsible for the context-based representation of word node Nx ∈ V.
− Step 2: generation of node-embedding
Representing the nodes: the matrices are generated as P ∈ N|N|∗pj and F ∈ N |N|∗pj are given as input into a
GNN network with convolution shown to generate the nodes represented as (7):

H = G (t) = ϑ(PG (b−1) L(b) + q(b) (7)

Here b depicts the number of layers in the GNN network with convolutions layer, P ∈ N |N|∗ph ; G (0) = Y, here
ϑ is a non-linear activation function as ReLU. L(b) ∈ N|N |∗ph for the transformation layer in the b − th layer.
q(b) is the bias associated with GNN layers associated with convolution, respectively:

P = K −1/2 TAM −1/2 (8)

P is the symmetric matrix of AM where T is the degree of the matrix AM whereas;

K xx = ∑y Fxy (9)

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1668-1678

Int J Artif Intell ISSN: 2252-8938  1673

− Step 3: GNN with convolution transformation

In this step, the noise is reduced and the bias is associated with processing the GNN network. A position
attention mechanism, which captures the necessary parts in the sentence irrespective of the sentiment as shown
in (10):

|N|
Ix = ∑x=1 ϑx Hx (10)

Here,

exponential(CF tanh (LF [Hx ⊕wx ]+qF

ϑx = |N| (11)
∑h=1 exponential(CF tanh (LF [Hh ⊕wh ]+qF

Here LF ∈ N( pj + pv ) ∗ (pj + pv ), qF ∈ N( pj + pv ), and C F ∈ N pj + pv for learning matrix. ⊕ is the

concatenation operator. wx ∈ Npv is the weight associated with the x − th word formulated as in (12):

|x−initiate|
, x ≤ initiate
k
wx = {0, initiate ≤ x ≤ terminate (12)
|x−terminate|
, x > terminate
k

Here initiate and terminate determines the starting and ending phase of the sentiment in the sentence.

3.3. Constructing the classifier

The dimension of the node embedding is reduced by the convolution layer by generating the feature
vectors ux , which determines the sliding of the filter R ∈ N r∗ph with length r from x to x + r − 1 to extract
the necessary information shown as (13):

ux = ReLU(RꭈMx:x+|N|−1 + q) (13)

Here x = [1, |N|] determines the order of the node representation M; ꭈ is the convolution operator, the
activation function is denoted by ReLU and q is the bias term. The feature vectors are constructed by node
representations as shown in (14):

u = [u1, u2, … … , u|N| ] (14)

− Step 1: max-pooling layer

The pooling layer builds the feature vectors for the same size by picking the maximum number of each vector
uk, the main aim here is to determine the size of feature vectors uk, which depends on the size of each matrix
M and R. The size will differ if the length of the sentence and the filter sizes are different, new features such as
u, are defined as in (15):

u = [u1, u2, … … , u|N| ] (15)

− Step 2: fine-tuning of the layers

This layer is responsible for fine-tuning the characteristics of the previous layer, which necessarily determines
the sentiments as shown in (16):

Y = SM(H M . u + q) (16)

Here H M ∈ Lb∗|N| and q ∈ H b is the weight matrix and bias of the layer. b is the number of sentiment phases.
Training the optimized GNN model with convolution to reduce the entropy loss values.

3.4. Optimizing loss of ensemble deep network

To train the CNN model over optimized LSTM model integrated with GNN network with
convolutions to reduce the loss function denoted as shown in (17)
q
Loss = − ∑k yk logyk ^ + δ||γ||2 (17)

Aspect based sentiment analysis using a novel ensemble deep network (Abraham Rajan)
1674  ISSN: 2252-8938

Here q denotes the sentiment classes for the training model, yk denotes the distribution of true sentiment for
the k − th class and yk ^ denotes the distribution of the sentiment to be predicted for the k − th class, δ denotes
the regulation co-efficient, and γ is the parameter set from the previous layer.

4. PERFORMANCE EVALUATION
The task of solving ABSA is more complex than conventional text-level sentiment analysis. It focuses
on recognizing the characteristics or features of an item described in a text and the perspective has taken on
each characteristic. It has been proven that applying optimized BERT to ABSA as a sentence-pair classification
task and developing an auxiliary sentence yields better results than the existing state-of-art models and baseline
models. Further, we have evaluated our proposed ensemble deep network proposed system (EDN-PS) model
with existing state-of-art methods and baseline models in terms of accuracy and F1-score.

4.1. Dataset details

To evaluate the efficiency of the proposed model to test the performance of the association for
computational linguistics (ACL) twitter social dataset [19]. The sentiment polarity is determined by dividing
the dataset into three categories: positive, negative and neutral. Table 1 shows the dataset details. The details
of the various methodologies used for comparison with the proposed EDN-PS model.

Table 1. Dataset details

Dataset Positive Negative Neutral Total
Training Testing Training Testing Training Testing Training Testing
Twitter 1,561 173 1,560 173 3,127 346 3,127 346

4.2. Compared models

To comprehensively evaluate our proposed model, we compared it to certain baseline models and
state-of-the-art (SOTA) models from different times. The eight baseline models are evaluated are mentioned
here. The seventeen models listed below, including the existing ones, have always performed at SOTA level.

4.2.1. Baseline models:

− IAN [4]: constructs attentive representations of the target context and engages in interactive attentive
learning of the target and context, respectively.
− MemNet [20]: determines the importance of various context words for a single feature and utilizes this
information as the semantic representation of the sentence to distinguish the most essential context
information of a particular trait.
− ATSM-S [21]: classifies the sentiment of extremely complex and coarsely written language by specifically
modelling the desired qualities at three distinct granularity levels, followed by categorizing the sentiment.
− AOA [22]: aspects and sentences are represented in a decentralized manner, with explicit capture of the
relationship between context phrases and aspects. The AOA module simultaneously teaches the model
aspects and sentence representation, with an emphasis on the main parts of the phrase.
− ASGCN-DG [7]: possesses a structure of undirected dependency. The adjacency matrix is produced using
DSPT words to avoid the limitations of attention processes and CNN-based models and depends on an
ordered dependency tree. Two networks are distinguished by their adjacency matrices.
− MAN [23]: a multi-attention network (MAN) was developed to alleviate the problem of information loss
when a coarse-level attention mechanism is employed to simulate aspects.
− UP-CNN-BERT [12]: this model proposes a novel unified location-aware CNN that constructs an
aspect-aware contextual representation for ABSA using an aspect mask to match the CNN, to compensate
for the absence of aspect locations in ACSA.

4.2.2. State-of-art models at different times

− ATAE-LSTM [3]: by using LSTM in conjunction with the attention mechanism to do sentence-level
semantic modelling to address the issue of aspect-level sentiment analysis. This is accomplished by
evaluating the significance of diverse contextual information to a certain attribute.
− GANN [24]: the most important local components of the overall emotional signals are recovered using
convolution and pooling algorithms. This is done while monitoring the sentiment signal semantic linkages,
the distance between each context word and aspect target, and the sequence information.

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1668-1678

Int J Artif Intell ISSN: 2252-8938  1675

− RAM [25]: it increases resistance to irrelevant data by using a method called multiple attention to identify
sentiment features from a distance.
− MGAN [26]: employs a combination of coarse- and fine-grained attention to capture the interactions
between aspects and contexts at the word level and an aspect alignment loss to explain the aspect-level
interactions between aspects that share a shared context.
− CDT [13]: the BiLSTM-learned embedding are amplified using GCN after learning the feature
representation of sentences with BiLSTM. GCN directly adjusts a phrase's dependency tree.
− DMMN-SDCM [27]: integrates information from semantic analysis into the memory network and assigns
an auxiliary task to learn the sentiment distribution of the entire phrase to offer the necessary background
knowledge for sentiment classification of the target aspect.
− TD-BERT [28]: a max-pooling layer and an FCL are added to the BERT model to create the necessary
sentiment classification model.
− AEN-BERT [29]: uses attention-based encoding to model the link between context and target, examines
the issue of inaccurate labels, and suggests label smoothing regularization. The idea also applies to Bert's
prior condition.
− BERT-PT [30]: after training BERT, a technique is introduced to compensate for this lack of domain and
task-related knowledge and to improve the training's results by utilizing the remaining quantity of data.
− BERT-BASE [31]: in all stages, attempts are made to simultaneously condition on both left and right
contexts to pre-train deep bidirectional representations from the unlabeled text. The pre-trained BERT
model may be enhanced to produce existing models for a variety of purposes with just one additional output
layer. We employ ABSA for multilingual learning.
− RepWalk [32]: finds the context words that have a great impact on the ability of aspect words to predict the
sentiment by randomly searching the grammatical graph. To determine the importance of each word in a
phrase using grammatical knowledge.
− SK-GCN (BERT) [33]: combines two modelling strategies for both GCN's syntactic dependency trees and
common sense information to specifically enhance phrase representation.
− SAGAT-BERT [34]: the BERT language model's external pre-training data and understanding of syntax
perception are applied to the dependent tree structure using the graph attention network, which aids in
replicating the interaction between the context and aspects.
− ASGCN-BERT [7]: uses the syntactic dependent structure of the phrase to solve the long-standing multi-
word reliability issue of ABSA.
− R-GAT+BERT [16]: by using this methodology and pruning the traditional dependency parse tree, a unified
aspect-oriented dependency tree structure depending on the desired aspect is produced. It is also
recommended to encode a novel tree structure for sentiment prediction using a relational graph attention
network (R-GAT).
− LCFS-BERT-CDM [1]: this method recommends employing syntactic relative distance to reduce the
impact of unrelated words with unstable syntactic ties to aspects to improve the aspect-based sentiment
classifier’s accuracy. To improve the efficiency of the aspect extractor, it also incorporates part of speech
embedding, syntactic relation embedding, and context embedding (e.g., BERT and Roberta).

4.3. Experimental analyses

The proposed EDN-ABSA is compared with various state-of-art models and baseline models. Graph
is plotted for accuracy and F1-score for ACL Twitter Dataset. the Baseline models used for comparison here
are IAN, MemNet, ATMS-S, AOA, ASGCN-DG, ASGCN-DT, MAN, and UP-CNN-BERT and state-of-art
models at different times are ATAE-LSTM, GANN, RAM, MGAN, CDT, DMMN-SDCM, TD-BERT,
AEN-BERT, BERT-PT, BERT-BASE, RepWalk, SK-GCN(BERT), SAGAT-BERT, ASGCN-BERT,
R-GAT-BERT, LCFS-BERT-CDM, and LCFS-BERT-CDW.

4.3.1 Accuracy
The accuracy is evaluated for various existing state-of-art models and baseline models and a graph is
plotted against the proposed model we can conclude that the proposed EDN-ABSA model performs better in
comparison with the existing system. Figure 3 shows the accuracy comparison graphically. Table 2 shows the
accuracy comparison. The accuracy value for the baseline method for ASGCN-DT is 71.53 and whereas for
IAN is 72.5 and the maximum value among the baseline methods is MAN methodology, which gives an
accuracy value of 76.56. Whereas the accuracy value for the state-of-art-model for a methodology for
ATAE-LSTM is 68.64 and for the RepWalk methodology, it is 74.4 and the maximum value for the
state-of-art-model at different times is TD-BERT value is 76.7, whereas the proposed EDN-PS model
outperforms other existing state-of-art and baseline models with an accuracy value of 78.5.

Aspect based sentiment analysis using a novel ensemble deep network (Abraham Rajan)
1676  ISSN: 2252-8938

4.4. F1-score
The F1-score is evaluated for various existing state-of-art models and baseline models and a graph is
plotted against the proposed model we can conclude that the proposed EDN-PS model performs better in
comparison with the existing models. Figure 4 shows the F1-score comparison graphically. Table 3 shows the
F1-score comparison. The F1-score value for the baseline method for ASGCN-DT is 69.68, which is the least,
and whereas for MemNet is 69.9 and the maximum value among the baseline methods is R-GAT-BERT
methodology, which gives an accuracy value of 74.88. Whereas the F1-score value for the state-of-art-model
at different times for methodology for MemNet is 69.9, and for SAGAT-BERT methodology it is 74.17 and
the maximum value for the state-of-art-model at different times is R-GAT-BERT value is 74.88, whereas the
proposed EDN-PS model outperforms other existing state-of-art and baseline models with an F1-score value
of 76.6.

Figure 3. Accuracy comparison Figure 4. F1-score comparison

Table 2. Accuracy comparison Table 3. F1-score comparison

Methodology Accuracy Methodology F1-score
ATAE-LSTM 68.64 ATAE-LSTM 66.6
RAM 69.36 RAM 67.3
UP-CNN-BERT 70.95 ASGCN-DT 69.68
MemNet 71.48 MemNet 69.9
ASGCN-DT 71.53 AOA 70.2
ASGCN-DG 72.15 ASGCN-DG 70.4
AOA 72.3 MGAN 70.81
IAN 72.5 IAN 70.81
MGAN 72.54 GANN 72.01
ASGCN-BERT 73.19 MAN 72.19
GANN 74.25 ASGCN-BERT 72.54
RepWalk 74.4 RepWalk 72.6
CDT 74.66 SK-GCN-BERT 73.01
AEN-BERT 74.71 AEN-BERT 73.13
CFS-BERT-CDM 74.82 CFS-BERT-CDM 73.5
SK-GCN-BERT 75 CDT 73.66
SAGAT-BERT 75.4 SAGAT-BERT 74.17
LGCF-CDM 75.69 LGCF-CDM 74.31
GCF-CDM-CDW 75.69 CFS-BERT-CDW 74.4
CFS-BERT-CDW 75.83 LGCF-CDW 74.53
R-GAT-BERT 76.15 GCF-CDM-CDW 74.77
LGCF-CDW 76.27 R-GAT-BERT 74.88
TD-BERT 76.7 EDN-ABSA 76.6
MAN 76.56
EDN-ABSA 78.56

5. CONCLUSION
ABSA, a method of text analysis, divides data into groups and evaluates the sentiment connected to
each group. ABSA may assess customer feedback by linking different sentiments with specific qualities of an
item or service. The existing research is focused on the correlation between the sentiment polarity of various
aspects and the local context. The essential deep correlations between the global context and aspects are not
focused on attention. The proposed EDN-ABSA is compared with various state-of-art models and baseline
models and a graph is plotted for accuracy and F1-score for ACL Twitter Dataset whereas the proposed

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1668-1678

Int J Artif Intell ISSN: 2252-8938  1677

EDN-PS model outperforms other existing state-of-art and baseline models with an accuracy value of 78.5 and
F1-score value of 76.6.

REFERENCES
[1] M. H. Phan and P. O. Ogunbona, “Modelling context and syntactical features for aspect-based sentiment analysis,” in Proceedings
of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 3211–3220, doi: 10.18653/v1/2020.acl-
main.293.
[2] N. Majumder, S. Poria, A. Gelbukh, M. S. Akhtar, E. Cambria, and A. Ekbal, “IARM: inter-aspect relation modeling with memory
networks in aspect-based sentiment analysis,” in Proceedings of the 2018 Conference on Empirical Methods in Natural Language
Processing, 2018, pp. 3402–3411, doi: 10.18653/v1/D18-1377.
[3] Y. Wang, M. Huang, X. Zhu, and L. Zhao, “Attention-based LSTM for aspect level sentiment classification,” in Proceedings of the
2016 Conference on Empirical Methods in Natural Language Processing, 2016, pp. 606–615, doi: 10.18653/v1/D16-1058.
[4] D. Ma, S. Li, X. Zhang, and H. Wang, “Interactive attention networks for aspect-level sentiment classification,” in Proceedings of
the Twenty-Sixth International Joint Conference on Artificial Intelligence, Aug. 2017, pp. 4068–4074, doi: 10.24963/ijcai.2017/568.
[5] B. Xu, X. Wang, B. Yang, and Z. Kang, “Target embedding and position attention with LSTM for aspect based sentiment analysis,”
in Proceedings of the 2020 5th International Conference on Mathematics and Artificial Intelligence, Apr. 2020, pp. 93–97, doi:
10.1145/3395260.3395280.
[6] L. Xu, L. Bing, W. Lu, and F. Huang, “Aspect sentiment classification with aspect-specific opinion spans,” in Proceedings of the
2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020, pp. 3561–3567, doi:
10.18653/v1/2020.emnlp-main.288.
[7] C. Zhang, Q. Li, and D. Song, “Aspect-based sentiment classification with aspect-specific graph convolutional networks,” in
Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint
Conference on Natural Language Processing (EMNLP-IJCNLP), 2019, pp. 4568–4578, doi: 10.18653/v1/D19-1464.
[8] X. Hou et al., “Graph ensemble learning over multiple dependency trees for aspect-level sentiment classification,” in Proceedings
of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language
Technologies, 2021, pp. 2884–2894, doi: 10.18653/v1/2021.naacl-main.229.
[9] R. Li, H. Chen, F. Feng, Z. Ma, X. Wang, and E. Hovy, “Dual graph convolutional networks for aspect-based sentiment analysis,”
in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint
Conference on Natural Language Processing (Volume 1: Long Papers), 2021, pp. 6319–6329, doi: 10.18653/v1/2021.acl-long.494.
[10] B. Huang and K. Carley, “Parameterized convolutional neural networks for aspect level sentiment classification,” in Proceedings
of the 2018 Conference on Empirical Methods in Natural Language Processing, 2018, pp. 1091–1096, doi: 10.18653/v1/D18-1136.
[11] C. Gan, L. Wang, Z. Zhang, and Z. Wang, “Sparse attention based separable dilated convolutional neural network for targeted
sentiment analysis,” Knowledge-Based Systems, vol. 188, p. 104827, Jan. 2020, doi: 10.1016/j.knosys.2019.06.035.
[12] X. Wang, F. Li, Z. Zhang, G. Xu, J. Zhang, and X. Sun, “A unified position-aware convolutional neural network for aspect based
sentiment analysis,” Neurocomputing, vol. 450, pp. 91–103, Aug. 2021, doi: 10.1016/j.neucom.2021.03.092.
[13] K. Sun, R. Zhang, S. Mensah, Y. Mao, and X. Liu, “Aspect-level sentiment analysis via convolution over dependency tree,” in
Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint
Conference on Natural Language Processing (EMNLP-IJCNLP), 2019, pp. 5679–5688, doi: 10.18653/v1/D19-1569.
[14] M. Zhang and T. Qian, “Convolution over hierarchical syntactic and lexical graphs for aspect level sentiment analysis,” in
Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020, pp. 3540–3549, doi:
10.18653/v1/2020.emnlp-main.286.
[15] H. Tang, D. Ji, C. Li, and Q. Zhou, “Dependency graph enhanced dual-transformer structure for aspect-based sentiment
classification,” in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 6578–6588,
doi: 10.18653/v1/2020.acl-main.588.
[16] K. Wang, W. Shen, Y. Yang, X. Quan, and R. Wang, “Relational graph attention network for aspect-based sentiment analysis,” in
Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 3229–3238, doi:
10.18653/v1/2020.acl-main.295.
[17] B. Zhang, X. Fu, C. Luo, Y. Ye, X. Li, and L. Jing, “Cross-domain aspect-based sentiment classification by exploiting domain-
invariant semantic-primary feature,” IEEE Transactions on Affective Computing, vol. 14, no. 4, pp. 3106–3119, Oct. 2023, doi:
10.1109/TAFFC.2023.3239540.
[18] M. Hu, S. Zhao, H. Guo, and Z. Su, “Hybrid regularizations for multi-aspect category sentiment analysis,” IEEE Transactions on
Affective Computing, vol. 14, no. 4, pp. 3294–3304, Oct. 2023, doi: 10.1109/TAFFC.2023.3236948.
[19] L. Dong, F. Wei, C. Tan, D. Tang, M. Zhou, and K. Xu, “Adaptive recursive neural network for target-dependent Twitter sentiment
classification,” in Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short
Papers), 2014, vol. 2, pp. 49–54, doi: 10.3115/v1/P14-2009.
[20] D. Tang, B. Qin, and T. Liu, “Aspect level sentiment classification with deep memory network,” in Proceedings of the 2016
Conference on Empirical Methods in Natural Language Processing, 2016, pp. 214–224, doi: 10.18653/v1/D16-1021.
[21] H. Peng, Y. Ma, Y. Li, and E. Cambria, “Learning multi-grained aspect target sequence for Chinese sentiment analysis,” Knowledge-
Based Systems, vol. 148, pp. 167–176, May 2018, doi: 10.1016/j.knosys.2018.02.034.
[22] B. Huang, Y. Ou, and K. M. Carley, “Aspect level sentiment classification with attention-over-attention neural networks,” in Social,
Cultural, and Behavioral Modeling, Cham: Springer, 2018, pp. 197–206, doi: 10.1007/978-3-319-93372-6_22.
[23] Q. Xu, L. Zhu, T. Dai, and C. Yan, “Aspect-based sentiment classification with multi-attention network,” Neurocomputing,
vol. 388, pp. 135–143, May 2020, doi: 10.1016/j.neucom.2020.01.024.
[24] N. Liu and B. Shen, “Aspect-based sentiment analysis with gated alternate neural network,” Knowledge-Based Systems, vol. 188,
p. 105010, Jan. 2020, doi: 10.1016/j.knosys.2019.105010.
[25] P. Chen, Z. Sun, L. Bing, and W. Yang, “Recurrent attention network on memory for aspect sentiment analysis,” in Proceedings of
the 2017 Conference on Empirical Methods in Natural Language Processing, 2017, pp. 452–461, doi: 10.18653/v1/D17-1047.
[26] F. Fan, Y. Feng, and D. Zhao, “Multi-grained attention network for aspectlevel sentiment classification,” in Proceedings of the 2018
Conference on Empirical Methods in Natural Language Processing, 2018, pp. 3433–3442, doi: 10.18653/v1/D18-1380.
[27] P. Lin, M. Yang, and J. Lai, “Deep mask memory network with semantic dependency and context moment for aspect level sentiment
classification,” in Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, Aug. 2019,
pp. 5088–5094, doi: 10.24963/ijcai.2019/707.

Aspect based sentiment analysis using a novel ensemble deep network (Abraham Rajan)
1678  ISSN: 2252-8938

[28] Z. Gao, A. Feng, X. Song, and X. Wu, “Target-dependent sentiment classification with BERT,” IEEE Access, vol. 7,
pp. 154290–154299, 2019, doi: 10.1109/ACCESS.2019.2946594.
[29] Y. Song, J. Wang, T. Jiang, Z. Liu, and Y. Rao, “Targeted sentiment classification with attentional encoder network,” in Artificial
Neural Networks and Machine Learning – ICANN 2019: Text and Time Series, Cham: Springer, 2019, pp. 93–103, doi:
10.1007/978-3-030-30490-4_9.
[30] H. Xu, B. Liu, L. Shu, and P. S. Yu, “BERT post-training for review reading comprehension and aspect-based sentiment analysis,”
in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human
Language Technologies, 2019, pp. 2324–2335.
[31] J. Devlin, M. W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language
understanding,” Arxiv-Computer Science, pp. 1–16, 2019.
[32] Y. Zheng, R. Zhang, S. Mensah, and Y. Mao, “Replicate, Walk, and Stop on Syntax: An Effective Neural Network Model for
Aspect-Level Sentiment Classification,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 34, no. 5,
pp. 9685–9692, Apr. 2020, doi: 10.1609/aaai.v34i05.6517.
[33] J. Zhou, J. X. Huang, Q. V. Hu, and L. He, “SK-GCN: modeling syntax and knowledge via graph convolutional network for aspect-
level sentiment classification,” Knowledge-Based Systems, vol. 205, p. 106292, Oct. 2020, doi: 10.1016/j.knosys.2020.106292.
[34] L. Huang, X. Sun, S. Li, L. Zhang, and H. Wang, “Syntax-aware graph attention network for aspect-level sentiment classification,”
in Proceedings of the 28th International Conference on Computational Linguistics, 2020, pp. 799–810, doi:
10.18653/v1/2020.coling-main.69.

BIOGRAPHIES OF AUTHORS

Abraham Rajan earned his Bachelor's of Engineering BE degree in CSE from

VTU, Belagavi in 2016. He has obtained his master's degree in M.Tech (CSE) from Reva
University in 2018. Currently he is a research scholar at Christ University doing his Ph.D. in
Computer Science and Engineering and also working as assistant professor in Nagarjuna
College of Engineering and Technology. He has attended many workshops and induction
programs conducted by various universities. His areas of interest are big data analytics and
cloud computing. He can be contacted at email: [email protected].

Manohara Manur is an associate professor in the Department of Computer

Science and Engineering at School of Engineering and Technology of Christ University,
Bangalore with an experience of 22 years in Teaching. He is qualified in Bachelor and Master
Degrees in Computer Science and Engineering, and Ph.D. in Computer Science and
Engineering in the area of data mining and big data. His areas of interest are data mining,
computer vision, machine learning, artificial intelligence, internet of things, and image
processing. He can be contacted at this email: [email protected].

Int J Artif Intell, Vol. 13, No. 2, June 2024: 1668-1678