Social Network Analysis and Mining (2024) 14:6

https://doi.org/10.1007/s13278-023-01166-w

ORIGINAL ARTICLE

Knowledge‑enhanced graph convolutional networks for Arabic aspect

sentiment classification
Rajae Bensoltane1 · Taher Zaki1

Received: 20 July 2022 / Revised: 2 November 2023 / Accepted: 6 November 2023

© The Author(s), under exclusive licence to Springer-Verlag GmbH Austria, part of Springer Nature 2023

Abstract
Aspect sentiment classification (ASC) is a sub-task of aspect-based sentiment analysis (ABSA) that aims at identifying the
sentiment polarity toward a specific aspect in a given text or sentence. Most existing research on Arabic ABSA adopted
rule-based or machine learning-based methods, with little attention to deep learning techniques. Additionally, the major-
ity of these deep learning-based models relied on attention mechanisms to capture the interaction between the context and
aspect words. However, attention-based methods are generally inefficient in extracting the syntactic dependencies between
contextual tokens and aspects. Therefore, we introduce a combined model that incorporates an Arabic BERT model with
graph convolutional network and local context focus layers to capture syntactic dependencies relevant to a specific aspect
while emphasizing the contribution of semantic-related tokens related to this aspect. We also integrate affective common-
sense knowledge into the graph networks to capture the sentiment-related dependencies between contextual words and the
specific aspect. The experimental results on an Arabic hotel dataset show that the proposed method outperforms the baseline
and related work models and achieves a state-of-the-art accuracy score of 92.77% in Arabic ASC. The achieved results show
the effectiveness of the proposed model in enhancing the aspect-specific sentiment representations, which can be promising
for future research in this field.

Keywords Aspect sentiment classification · BERT · Graph convolutional networks · Affective commonsense knowledge ·
Local context focus · Arabic

1 Introduction or a sentence, which is insufficient when different opinions

are expressed toward multiple aspects or entities in the
In today’s digital age, the daily exchange of vast amounts of same text. Therefore, the aspect level, also referred to as
data via a multitude of websites and social media platforms aspect-based sentiment analysis (ABSA), has become an
has ignited significant interest in various research areas, with active research task of sentiment analysis, as it enables the
sentiment analysis standing out as a notable example. As a extraction of aspect terms and the corresponding sentiment
subfield of natural language processing, sentiment analysis polarities.
allows the automated extraction of opinions, thoughts, and ABSA can be divided into two main sub-tasks: aspect
attitudes from a given text, finding widespread applications term extraction (ATE) and aspect sentiment classifica-
across numerous domains, such as social networks (Li et al. tion (ASC). The purpose of the ATE task is to extract the
2023; Rodríguez-Ibánez et al. 2023) and business (Costola discussed aspect terms, while the ASC task aims at iden-
et al. 2023; Huang et al. 2023a, b). tifying the sentiment polarity toward each aspect. Given
Sentiment analysis can be performed on three levels: this sentence: “the pasta was delicious, but the waiter was
document level, sentence level, and aspect level. The first unfriendly,” the discussed aspects are pasta and waiter, and
two levels identify the sentiment of the whole document the corresponding sentiments are positive and negative,
respectively. The ASC task allows us to gain deep insights
* Rajae Bensoltane into users’ preferences for different aspects of a given entity
[email protected] in various domains, such as education (Kastrati et al. 2020),
healthcare (Zhao et al. 2023), box office predictors (Yang
1
Laboratory of Innovation in Mathematics and Intelligent et al. 2023), and brand monitoring (Xu et al. 2022).
Systems, Faculty of Applied Sciences, Agadir, Morocco

13
Vol.:(0123456789)
 6 Page 2 of 14 Social Network Analysis and Mining (2024) 14:6

Arabic is one of the most spoken languages in the world, The reminder of the paper is structured as follows: Sect. 2
with more than 422 million speakers (Omar and Abd El- overviews related work to Arabic ASC, GCN, and Sentic-
Hafeez 2023). According to the Internet World Stats, it is Net; Sect. 3 describes the proposed model in detail; Sect. 4
ranked as the fourth most used language on the internet. presents the experimental setup; Sect. 5 discusses the eval-
The Arabic online content is growing rapidly (Alqurashi uation results; finally, the conclusion and future research
2023), enabling Arabic sentiment analysis (ASA) to get lots directions are provided in Sect. 6.
of attention from researchers in recent years. However, most
existing ASA work has focused on the document and sen-
tence levels, with little attention to the ABSA tasks (Ben-
2 Related work
soltane and Zaki 2023a).
Moreover, most existing Arabic ABSA papers have
This section is divided into three main subsections. The first
adopted rule-based or traditional machine learning classifi-
subsection overviews related work to Arabic ASC using
ers to handle these tasks with limited work on deep learning
deep learning models. The second subsection introduces
models (Bensoltane and Zaki 2023b). Besides, most of the
graph-based studies for ASC. Since the lack of Arabic ASC
proposed Arabic ABSA deep learning-based methods used
methods using GCN, papers that targeted the English lan-
traditional word embedding models for word vector repre-
guage are overviewed. Finally, the third subsection discusses
sentations, which provide static embedding vectors for each
studies that have used SenticNet knowledge base, primarily
word independently of the context in which it occurs. Fur-
for handling the ASC task.
thermore, most of these deep learning-based methods were
implemented based on attention mechanisms, which are
inefficient in extracting syntactical dependencies between 2.1 Arabic aspect sentiment classification
contextual tokens and aspect words (Huang et al. 2023a, b).
Therefore, the aim of this paper is to overcome the afore- Most existing ASA work focused mainly on the document or
mentioned shortcomings by proposing a combined model sentence levels with limited studies on ABSA. Furthermore,
that incorporates an Arabic BERT model, called AraBERT deep learning models are under-discovered in this task for
(Antoun et al. 2020), along with graph convolutional net- the Arabic language (Bensoltane and Zaki 2023a).
work (GCN) and local context focus (LCF) layers to cap- One of the earlier studies adopting a deep learning
ture the syntactically related words to the specific aspect method for Arabic ASC is Ruder et al. (2016). The sys-
based on the syntactic dependency tree of the sentence tem was submitted to the SemEval workshop (Pontiki et al.
while focusing on local context features that are relevant 2016) and was ranked 1st in Arabic. The model fed both
to the given aspect term. Moreover, we explore the use of the aspect embedding and the input sequence embeddings
an affective commonsense knowledge base to enhance the into a convolutional neural network (CNN) model for sen-
dependency graphs of the sentence. The main contributions timent polarity identification. Different variants of adding
of this study are: the aspect vector to the model were examined, and the best
results were achieved by concatenating each aspect vector
• We introduce a combined model that incorporates the with the word vector before the convolution. The proposed
strengths of the AraBERT model with GCN and LCF model achieved an overall enhancement of 11.8% over the
layers to further enhance the task of ASC on Arabic data- baseline model. Nevertheless, one main limitation of this
set. To our knowledge, this is the first work to use this model is the use of random word embeddings to represent
combination to handle this task in Arabic. input text, which negatively affects its ability in encoding
• We exploit the syntax-related context information using context and semantic features.
GCN over the dependency tree of the sentence to enhance The authors in Al-Smadi et al. (2019) implemented their
this task, which is, to our knowledge, the first time to use model based on bidirectional long short-term memory (BiL-
GCN for the Arabic ASC task. STM) to enhance the ASC task on the same Arabic hotel
• We investigate the use of an affective commonsense dataset from SemEval 2016 task 5. They followed the work
knowledge to capture the contextual affective dependen- of (Wang et al. 2016) by encoding the opinions targets as
cies related to the specific aspect. As far as we know, features to force the attention layer to focus on the impor-
this is the first time to examine the use of an affective tant parts of the sentence. Experiments showed that the pro-
knowledge base for this task in Arabic. posed method obtained similar results to the previous model
• Extensive experiments on an Arabic reference dataset (accuracy = 82.6% vs. 82.7%). Although the proposed model
show that the proposed model outperforms the baseline utilizes an attention mechanism to capture the importance of
and related work models and achieves a state-of-the-art context words to a specific aspect term, the context vectors
accuracy score. must encode both the aspect and sentiment information and

13
Social Network Analysis and Mining (2024) 14:6 Page 3 of 14 6

the alignment scores are uniformly applied across all feature of 15% over the baseline model. Nevertheless, it is unable to
dimensions, irrespective of their differences. capture syntax information and long-range word dependen-
An attempt to enhance the previous results was proposed cies, which are crucial for the ASC task.
in Abdelgwad et al. (2021). The authors followed the work A recent study of Fadel et al. (2023) adapted the LCF-
of Ma et al. (2017) to implement an interactive attention ATEPC model (Yang et al. 2021) to handle extract aspect
network (called IAN-BGRU) but using bidirectional gated terms and identify their sentiment polarities at the same
recurrent units (BiGRU) instead of BiLSTM. Two BiGRU time. The authors employed an Arabic BERT model as a
layers were first applied to extract both aspect and context shared layer for Arabic contextual text representation. They
representations, followed by two attention mechanisms to then used multi-head self-attention and LCF mechanism to
compute attention vectors. The proposed model achieved an model the interactive information between the context and
overall improvement over the model of Ruder et al. (2016). a specific aspect word. The proposed model achieved an
However, the main limitation of this model is the use of a accuracy score of 91.5% in ASC and an F1 score of 75.94%.
coarse-grained attention mechanism since they rely on aver- This method considers the word counts between two words
age vectors to determine attention weights for contexts, and as their semantic-relative distance, disregarding the mutual
as a result, they cannot capture word-level interactions. syntactic associations. Nonetheless, crucial sentiment words
Al-Dabet et al. (2021) have also sought to enhance the may be linked to the target aspect terms through grammatical
previous results by exploiting the success of memory net- rules, even if they have a considerable semantic separation.
works and attention mechanisms. They first implemented a
memory module using a stacked bidirectional independent 2.2 Graph convolutional network
long short-term memory network (IndyLSTM) with three
layers, followed by a recurrent attention module to help GCN has recently been successively applied in many NLP
the model focus on the sentence’s key parts. The proposed tasks, such as named entity recognition (Madan et al.
model achieved more than 3% improvement compared to 2023), text classification (Zhao et al. 2022a, b), and ques-
the IAN-BGRU model. Although this model achieved bet- tion answering (Yusuf et al. 2022). In addition, GCN-based
ter results than other attention-based models thanks to the models have enabled the detection of syntactically relevant
combination of multiple attentions with IndyLSTM layers, words to the given aspect by capturing long-term syntactic
it primarily focuses on the semantic information of the sen- dependencies through convolutional operation, which has
tence and ignores the syntactic structure knowledge, which shown promising results in ASC. One of the first papers
may increase the noise introduced by attention mechanisms. investigating the use of GCN for the ASC task is Zhang
Unlike the previous models that exploited context-inde- et al. (2019). The authors implemented a GCN layer over
pendent word embeddings to initialize the neural network the dependency tree of the sentence to capture syntactical
models, the authors in Abdelgwad et al. (2022) investigated information and long-range multi-word dependencies. The
using contextualized word embeddings from the BERT authors in Zhao et al. (2020) used a GCN-based model to
model to enhance the Arabic ASC task. They used BERT’s learn the sentiment dependencies between multiple aspects
sentence pair input architecture to handle this task by accept- in one sentence. Zhang and Qian (2020) first used a global
ing the sentence as the first input and the aspect as the sec- lexical graph for encoding word co-occurrence informa-
ond input. The BERT-based model was then fine-tuned on tion at the corpus level. They then build a concept hier-
the downstream task with a linear classification layer for the archy on each of the syntactic and lexical graphs for dif-
sentiment polarity prediction. The proposed model achieved ferentiating various types of dependency relations or word
enhanced results compared with previous models on the co-occurrence relations. Zhao et al. (2022a, b) introduced
Arabic hotel reviews (accuracy = 89.51%). Despite BERT’s a GCN model with multiple weight mechanisms, called
effectiveness in capturing contextual and semantic features MWM-GCN. Unlike conventional methods, MWM-GCN
from text, it does not possess explicit capabilities for mod- utilizes a dynamic weight alignment mechanism to ensure
eling sentence syntactic structures and dependencies. BERT’s WordPiece unit aligns with word-level dependen-
Al-Smadi et al. (2023) has exploited a multilingual uni- cies, improving its performance. An aspect-aware weight
versal sentence encoder (MUSE) with a gated recurrent unit mechanism regulates the flow of information to the aspect
model to improve the previous results for the Arabic ASC during graph convolution. Additionally, it incorporates an
task. Instead of word or character-level embeddings, the aspect-oriented dynamic loading layer and multi-head self-
MUSE model provides sentence-level embeddings. Besides, attention to enhance the model’s ability to capture syntax
The BiGRU layers were concatenated using global max and semantics for classification. A recent work of Huang
pooling and global average pooling to reduce the dimensions et al. (2023a, b) employed a conditional random field (CRF)
of the previous layers while maintaining the important fea- chain to identify the opinion span of a word specific to a
tures. The proposed model achieved an overall enhancement particular aspect. It then incorporates the context within this

13
 6 Page 4 of 14 Social Network Analysis and Mining (2024) 14:6

opinion span into a multilayer graph convolutional network that impact sentiment polarity, including negations, modi-
by utilizing an enhanced position decay function. Finally, it fiers, and other related terms.
computes vector expressions of global nodes to predict senti- Despite the success of the previous GCN-based models in
ment polarity labels that are specific to the aspect. handling the ASC task, this method is under-discovered in
Arabic ABSA. Therefore, the aim of this study is to enhance
the Arabic ASC task by exploiting the syntactic information
2.3 SenticNet knowledge base using GCN over the dependency tree of the sentence. We
also investigate the enhancement of the graph dependencies
SenticNet is one of the most used affective knowledge bases, using affective knowledge from BabelSenticNet. Moreover,
which is a publicly available lexical resource that provides our proposed model integrates an LCF layer to reduce the
several thousand commonsense concepts along with their influence of irrelevant words to a specific aspect.
sentiment scores (Bisio et al. 2017). Different versions have
been released for SenticNet, primarily focused on Eng-
lish, such as SenticNet 2 (Cambria et al. 2012), SenticNet 3 Research methodology
3 (Cambria et al. 2014), and SenticNet 4 (Cambria et al.
2016). In addition, a multilingual version, called BabelSen- 3.1 Problem definition
ticNet, has been provided to handle 40 languages (i.e., Ara-
bic), which we use in this study. SenticNet has been utilized The aim of the ASC task is to determine the sentiment
in many tasks such as sarcasm detection (Du et al. 2022), polarity of a specific aspect word in a given text by captur-
emotion recognition (Chen et al. 2023), and text classifica- ing aspect-related sentiment information from the context,
tion (Wu 2023). Recently, many studies have used the affec- which can be formulated as follows:
tive knowledge from SenticNet to enhance the sentiment Given a sentence S = [W1, W2,…, Wn] of n words and
features representation for handling the ASC task. Zhou an aspect term x = [x1, x2,…, xt], which contains t words
et al. (2020) implemented two strategies to model the syn- (1 ≤ t ≤ n). The aim of this task is to predict the sentiment
tactic dependency tree and commonsense knowledge graph polarity (positive, negative, or neutral) regarding the aspect
to enrich the representation of a sentence toward a specific term x for S.
aspect. In the first strategy, two GCN models were employed
to independently encode the syntactic graph and the knowl- 3.2 Model overview
edge graph. In the second strategy, a combined syntactic and
knowledge GCN model was constructed to represent both As illustrated in Fig. 1, our combined model is composed of
the syntactic and knowledge graphs jointly. The authors in different components: two independent AraBERT layers for
Liang et al. (2022) proposed a novel solution for construct- local and global contextual features encoding, an LCF layer
ing the graph by enhancing the dependency graph of the to focus on local contextual words, an enhanced GCN based
sentence using affective knowledge from SenticNet. The on BabelSenticNet to exploit the affective dependencies
novel affective-enhanced graph model takes into account between contextual words and specific aspects, an interac-
both the relationships among contextual words and aspect tive learning layer to learn the profound correlation between
words, as well as the affective information between opinion context and targeted aspects, and finally an output layer to
words and the aspect. Liu et al. (2023) introduced a dual- predict the sentiment polarity toward the aspect word.
gated graph convolutional network (DGGCN) model that
integrates GCN into the gating mechanism to amplify the 3.2.1 AraBERT layers
interaction between the context and the aspect word while
consolidating the affective attributes linked to the specified Pre-trained language models (PTMs) have achieved state-
aspect. Moreover, including contextual affective knowledge of-the-art in many NLP tasks, including ABSA (Abas et al.
in the graph networks helps DGGCN better understand sen- 2020; Zhao and Yu 2021). Many PTMs models were pro-
timent-related data. Gu et al. (2023) proposed a GCN model vided in the last few years, such as BERT (Devlin et al.
(called EK-GCN) that incorporates external knowledge like 2019) and XLNET (Yang et al. 2019). BERT is a deeply
SenticNet and part-of-speech information. It employed a bidirectional model based on transformers. Unlike tradi-
Part-of-Speech Matrix to capture the influence of words tional word embedding models such GloVe (Pennington
like negations and modifiers, a Sentiment Lexicon to assign et al. 2014) and Word2Vec (Mikolov et al. 2013), BERT
sentiment scores to words, and a Word–Sentence Interaction can provide different vector representations to the same word
Network (WSIN) to filter relevant sentence information. The based on the context in which it appears.
proposed model can effectively overcome the challenge of In this paper, two independent AraBERT layers are
capturing edge labels and enhance the significance of words employed to provide two types of contextual features. One

13
Social Network Analysis and Mining (2024) 14:6 Page 5 of 14 6

Fig. 1  The overall architecture

of the proposed model

{
layer generates global context features by accepting a sen- Z, SRDi ≤ 𝛼
tence pair as input (Xg = [CLS]sentence [SEP]aspect term Ki =
O, SRDi > 𝛼 (4)
[SEP]) to help detect the semantic relationship between the
contextual sentence and the aspect. The other layer provides [ ]
local context features by taking the context sentence as input M = K1m , K2m , K3m , … Knm (5)
(Xl = [CLS]sentence [SEP]). The outputs of the AraBERT
layers are annotated as follows: BlCDM = Bl ⋅ M (6)
( )
Bg = AraBERTglobal Xg (1) M is a feature masking matrix and Ki denotes the mask
vectors for each token. m is the input sequence length,
( ) whereas α is the SRD threshold. O and Z are the zeros and
Bl = AraBERTlocal Xl (2)
the ones vectors, respectively. “·” denotes the vector dot
product operation.
3.2.2 LCF layer Unlike CDM that totally ignores the features of non-local
context, CDW keeps less-semantic-relative context features
Following the work of Zeng et al. (2019), we integrated but de-emphasizes them according to their SRD to the aspect
an LCF layer to help the model focus on the sentimental term:
information contained in the neighbor words to the aspect {
term. This layer uses the semantic-relative distance (SRD) Z SRDi ≤ 𝛼
Ki = SRDi −𝛼 (7)
to detect the local context of the targeted aspect. SRD is n
⋅ Z SRDi > 𝛼
measured by counting words between context tokens and
aspect terms as follows: [ ]
Q = K1w , K2w , K3w , … Knw (8)
⌊ ⌋
l
SRDi = ||i − Pa || − (3)
2
BlCDW = Bl ⋅ Q (9)
where i represents the location of the context word, Pa is
the average position of the aspect term, and l represents the where Q is the weight matrix and Ki denotes the weight vec-
length of the aspect. S­ RDi denotes the SRD between a spe- tor for each less-semantic-relative contextual word.
cific aspect and the i-th context word.
Two architectures are evaluated in this paper to focus 3.2.3 Knowledge‑enhanced GCN layers
on local contexts, namely, context-features dynamic mask
(CDM) and context-features dynamic weighting (CDW). For the ASC task, it is crucial to capture syntactical infor-
CDM masks out less-semantic context features learned mation and mine word dependencies between words from a
by the AraBERT layer for which the SRD to aspect words given text. Therefore, we applied a multilayer GCN over the
is greater than a pre-defined threshold. The local context syntactic dependency tree of the sentence. First, the graph
output is calculated as follows: of convolutional networks over the dependency tree of each

13
 6 Page 6 of 14 Social Network Analysis and Mining (2024) 14:6

layer. F is a position-aware transformation function (Zhang

et al. 2019). Ẽi is a symmetric normalized adjacency matrix:

Ẽi = Ei ∕(Ui + 1) (16)

Fig. 2  Example of a dependency tree of an Arabic sentence n

∑
Ui = Eij (17)
i=1

Table 1  Example of affective Word Score Ui is the degree of the i-th token in the tree.
words and their corresponding
scores from BabelSenticNet ‫( سيء‬Bad) − 0.605
3.2.4 Interactive learning layer
‫( رائع‬Amazing) 0.793
‫( صعب‬Difficult) − 0.539
In this layer, a multi-head self-attention (MHSA) is applied
‫( جميل‬Beautiful) 0.721
on the concatenation (Bconc) of local context, global context,
‫( مقزز‬Disgusting) − 0.439
and the final output of GCN layers to improve the model’s
ability to learn the profound correlation between context and
aspect words:
sentence is constructed. Figure 2 illustrates an example of
a dependency tree. Then, the adjacency matrix of the graph Bconc
FCL
= W conc ⋅ Bconc + bconc (18)
is derived as follows:
{ ( )
1, if wi , wj contains dependency BILconc = MHSA Bconc
FCL (19)
Ai,j = (10)
0, 0 otherwise
Wconc and bconc are the weight and bias vectors of the fully
connected layer (FCL), respectively.
Inspired by the work of Liang et al. (2022), the represen-
tation of the adjacency matrix is enhanced by incorporating 3.2.5 Output layer
the affective information between context tokens and aspect
words using BabelSenticNet as follows: The sentiment polarity is predicted by pooling out the output
( ) of the interactive layer and feed it into a SoftMax layer:
Ei,j = Ai,j × Si,j + Ri,j + 1 (11) ( )
Bconc
pool
= POOL BILconc (20)
( ) ( )
Si,j = BabelSenticNet wi + BabelSenticNet wj (12) � �
exp Bconc
pool
{
1, if wi or wj is an aspect term Y=∑ � � (21)
Ri,j = (13) C
exp Bconc
0, 0 otherwise j=1 pool

where BabelSenticNet (wi) and BabelSenticNet (wj) are where C denotes the number of sentiment classes and Y is
the affective scores from BabelSenticNet database of the the predicted value of aspect sentiment polarity.
words wi and wj, respectively. Examples of affective scores
of some Arabic words are illustrated in Table 1.
The graph is then fed into the GCN layers to learn the 4 Experimental setup
affective dependencies related to a given aspect. Each node
in the l-th GCN layer is updated based on its neighboring 4.1 Dataset
nodes:
( ) The experiments were conducted on an Arabic dataset
vli = relu Ẽi gl−1
i
wl
+ b l
(14) from the hotel domain (Al-Smadi et al. 2016; Pontiki et al.
2016). The data was selected from hotel reviews collected
( ) by ElSahar and El-Beltagy (2015). The annotation scheme
gl−1
i
= F vl−1
i (15) followed the SemEval 2016 task 5 guidelines. Figure 3
illustrates an example of an annotated review from the
where vli denotes the product of current GCN layer. gl−1 is
i used dataset. Each review consists of one or multiple
the representation of i-th token evolved from the previous
sentences, and each sentence consists of a list of tuples

13
Social Network Analysis and Mining (2024) 14:6 Page 7 of 14 6

Fig. 3  Example of an annotated

review from the Arabic hotel
dataset

Table 2  Distribution size of the Arabic hotel dataset in terms of train- Table 4  Experimental hyper-parameters
ing and testing sets
Hyper-parameter Value
Reviews Sentences Tuples
Batch size 16
Training set 1839 4802 10,509 Max sequence length 128
Testing set 452 1227 2604 Dropout 0.2
Optimizer Adam
L2 regularization 1e−5
Table 3  Number of aspects per sentiment class in the training and Learning rate for BERT-based models 2e−5
testing sets Learning rate for non-BERT models 1e−3
Positive Negative Neutral Loss function Cross-entropy

Training set 6197 3629 683

Testing set 1508 927 169
4.3 Evaluation metrics

The accuracy score is the primary used metric in previ-

containing three elements: category, target, and polarity. ous Arabic ASC papers. We also adopt the macro aver-
The distribution size of the dataset is shown in Table 2, age of F1 score, as it is more suitable when the dataset is
whereas Table 3 illustrates the distribution of aspects per unbalanced.
sentiment class.
T
Accuracy = (22)
N
4.2 Experimental settings where T is the number of correctly predicted samples and N
is the total number of samples.
The models were implemented using PyTorch and Python
3.6 and trained on NVIDEA Tesla K80. For non-BERT 2 × MP × MR
Macro F1 = (23)
models, we used a Twitter-based 300-dimensional Ara- MP + MR
Vec model (Soliman et al. 2017) to initialize word vectors.
The BERT-based models were implemented using the base
C
1∑
architecture of the second version of the AraBERT model, MP = MPj (24)
C j
named AraBERTv02, with a hidden size of 768 and 12 lay-
ers of transformers with a number of self-attention heads
C
of 12. Besides, the dependency graph was constructed with 1∑
MR = MRj (25)
un-directional. For the knowledge base, the BabelSenticNet C j
database was utilized to handle the Arabic language. Table 4
illustrates the adopted hyper-parameters in this study, which where C denotes the number of sentiment classes. M ­ Pj and
were selected using a grid search technique. ­MRj are the precision and recall for class j, respectively.

13
 6 Page 8 of 14 Social Network Analysis and Mining (2024) 14:6

Table 5  Main experimental Type of representation Model Accuracy (%) Macro F1

results score (%)

– Baseline (Pontiki et al. 2016) 76.4† –

AraVec-based AB-LSTM-PC (Al-Smadi et al. 2019) 82.60† –
IAN-GRU (M.Abdelgwad et al. 2021) 83.98† –
MBRA (Al-Dabet et al. 2021) 87.31† –
AOA 85.74* 66.51*
ASGCN-DG 87.53* 67.27*
SenticGCN 88.52* 68.23*
AraBERT-based BERT-Linear-pair (Abdelgwad et al. 2022) 89.51† –
LCF-BERT 90.97* 79.83*
ASGCN-BERT 90.03 72.23
SenticGCN-BERT 90.84* 79.4*
ArBLCSNGCN (CDW) (ours) 91.85 78.45
ArBLCSNGCN (CDM) (ours) 92.77 81.59

The results with “†” are retrieved from the original papers. The results with “*” are obtained by imple-
menting the models based on the published source codes using the same Arabic dataset. The best results
are marked in bold and the unreported values are represented with “–”

4.4 Baseline and related work LCF-BERT (Zeng et al. 2019): Uses a local context focus
mechanism with an SRD threshold to help the model focus
In order to comprehensively evaluate the performance of on local context words and alleviate the negative influence
our model, we compare it with the following baseline and of unrelated sentiment tokens.
related work models: ASGCN-BERT: The BERT version of ASGCN-DG.
Baseline (Pontiki et al. 2016): Is an SVM-based classifier SenticGCN-BERT (Liang et al. 2022): The BERT version
provided by the SemEval competition. N-gram features were of SenticGCN.
used to train an SVM classifier with a linear kernel. It is worth mentioning that AOA, LCF-BERT, and
AB-LSTM-PC (Al-Smadi et al. 2019): The aspect embed- ASGCN-DG were implemented based on a publicly avail-
ding is combined with each word embedding, and then, an able repository,1 while SenticGCN was implemented using
attention mechanism is applied to help the model to focus the official published source code.2
on context words related to the aspect term.
IAN-GRU (Abdelgwad et al. 2021): Employs BiGRU to
provide the hidden states of targets and context, followed by 5 Results and discussion
two associated attention networks to learn the interactions
of aspects and their related context words. 5.1 Comparative analysis
MBRA (Al-Dabet et al. 2021): A memory module is
implemented using a stack of bidirectional independent The experimental results are illustrated in Table 5. They
long short-term memory with three layers, followed by a show that the AraBERT-based models achieve better results
recurrent attention mechanism to handle long and complex than AraVec-based models. This can be justified thanks to
sentences. the ability of AraBERT model to represent semantic fea-
AOA (Huang et al. 2018): This model adopts an atten- tures better than context-independent word embeddings.
tion-over-attention module to capture the interaction rela- Additionally, AraBERT uses a WordPiece tokenizer that
tionships between aspects and context sentences explicitly. splits unknown words into a set of sub-words, allowing bet-
ASGCN-DG (Zhang et al. 2019): This model uses GCN ter handling of out of vocabulary (OOV) issues, especially
on the dependency tree of a sentence to leverage the syn- in morphologically rich languages like Arabic.
tactical information and word dependencies to resolve the Besides, the graph-based (i.e., ASGCN-DG and Sen-
problem of multi-word dependency for the ASC task. ticGCN) methods have outperformed attention-based models
SenticGCN (Liang et al. 2022): Similar to ASGCN, but (i.e., AB-LSTM-PC, IAN, AOA, and MBRA). This proves
enhances the graph of the sentence using affective informa-
tion from BabelSenticNet.
BERT-Linear-pair (Abdelgwad et al. 2022): Fine-tuned 1
https://​github.​com/​songy​ouwei/​ABSA-​PyTor​ch
BERT with a sentence pair input for ASC task. 2
https://​github.​com/​BinLi​ang-​NLP/​Sentic-​GCN

13
Social Network Analysis and Mining (2024) 14:6 Page 9 of 14 6

Table 6  Evaluation results of ablation study context words is more beneficial than down-weighting their
Model Accuracy (%) Macro F1 signals for the ASC task.
score (%)

ArBLCSNGCN-wo-LCF 90.37 75.17

5.2 Ablation study
ArBLCSNGCN-wo-global 90.24 75.88
ArBLCSNGCN-wo-BabelSenticNet 91.04 79.87
To further evaluate the impact of each component of
ArBLCSNGCN (ours) 92.77 81.59
our proposed model, an ablation study is performed on
ArBLCSNGCN model. The results are illustrated in Table 6.
The best results are marked in bold Removing the LCF layer leads to performance drop, which
indicates the importance of alleviating the negative influence

Fig. 4  Accuracy and macro F1

scores of our proposed model
using different numbers of GCN
layers

that extracting syntactical dependencies between contextual

of less-semantic-relative contextual words. Besides, ignoring
tokens and aspect words is beneficial for detecting the senti-
the global context features has decreased both accuracy and
ment polarities of aspects. Besides, the GCN-based models
F1 scores by 2.53% and 5.71%, respectively. This illustrates
are better at detecting long-range word dependencies com-
the importance of incorporating the AraBERT encoded
pared to attention-based models.
semantic features between context words and a given aspect
Furthermore, the SenticGCN model achieves better
for the ASC task. On the other hand, constructing the graph
results than ASGCN. This proves the efficiency of integrat-
for the sentence without using SenticNet has dropped the
ing affective knowledge from BabelSenticNet in enhancing
performance by ~ 1.73% for the accuracy score. This shows
word dependencies of the sentence, which is compatible
that integrating the affective information provided by Sen-
with the findings of previous studies (Liang et al. 2022).
ticNet between contextual opinion words and aspects is ben-
On the other hand, our proposed model (ArBLCSNGCN)
eficial in enhancing the prediction of sentiment polarities for
obtains the best results over all the baseline and related work
specific aspects.
models. It outperforms SenticGCN-BERT and LCF-BERT
by more than 1.8% for the accuracy score. This shows that
combining the global context features with the sentiment- 5.3 Impact of GCN layers
related dependencies between contextual words and specific
aspects while emphasizing the contribution of local context To investigate the impact of the number of GCN layers (L)
tokens that are relevant to target words has yielded better on the performance of the proposed model, we conducted
performance. a set of experiments on the ArBLCSNGCN model using
Additionally, integrating the CDM layer (ArBLCSNGCN different L values and checked the corresponding F1 and
(CDM)) leads to a better performance than using CDW. This accuracy scores. The number of layers varies from 1 to 7. As
indicates that masking features of less-semantic-relative shown in Fig. 4, the best results are achieved when L equals

13
 6 Page 10 of 14 Social Network Analysis and Mining (2024) 14:6

Fig. 5  Accuracy and macro

F1 scores of our model based
on different values of SRD
threshold

3. Nevertheless, the performance generally gets dropped 5.6 Case study

with the increase in layers’ number. This can be justified
because when the number of layers increases, the number of We provide a case study with four different testing examples
parameters augments, and the model becomes too difficult to better understand how our model works. The predictions
to train and hard to be generalized. Hence, L is set to 3 in of these examples using BERT-Linear-pair, LCF-BERT,
our experiments. ASGCN-BERT, SenticGCN-BERT, and ArBLCSNGCN
are illustrated in Table 7.
5.4 Impact of SRD Given the first example, i.e., ‫المبنى جميل والغرف ال بأس بها‬
(The building is beautiful and the rooms are okay),
The impact of the dynamic masking technique on our pro- BERT-Linear-pair model fails to identify the correct sen-
posed model is validated by investigating the sensitivity of timent compared with the other models, including
the SRD thresholds on the model’s performance. ASGCN-BERT. This can be justified because the BERT-
Figure 5 shows the experimental results of our best model pair model focuses on the adjective ‫( جميل‬beautiful),
when using different SRD values that range from 0 to 7. It which is related to the aspect ‫( المبنى‬building). Hence, this
can be seen that considering only the aspect word (SRD = 0) shows the effectiveness of supplementing the semantic
and ignoring other related local context tokens is insuffi-
cient for the ASC. On the other hand, the ArBLCSNGCN
model achieves the best accuracy and F1 scores when SRD
is equal to 2. Therefore, the value of SRD is set to 2 in our
experiments.

5.5 Impact of fine‑tuning

We examined the effect of fine-tuning on the overall perfor-

mance of our best model by freezing the AraBERT’s param-
eters during training. Figure 6 illustrates the comparison
between the performance of the proposed model with and
without fine-tuning. The results show that the performance
has dropped significantly when setting the parameters fixed.
Thus, this indicated that fine-tuning on task-specific datasets
is indispensable to release the true power of AraBERT, espe- Fig. 6  The accuracy and macro F1 scores of our proposed model with
cially in low-resource settings. and without fine-tuning the AraBERT model

13
Social Network Analysis and Mining (2024) 14:6 Page 11 of 14 6

Table 7  Results of case study based on four different sentences, randomly selected from the test dataset
Reviews Aspect Models Predicted Label True label

AR: ‫المبنى جميل والغرف ال بأس بها‬ ‫( الغرف‬rooms) BERT-Linear-pair Positive Neutral
EN: The building is beautiful and the rooms are okay LCF-BERT Neutral
ASGCN-BERT Neutral
SenticGCN-BERT Neutral
ArBLCSNGCN (ours) Neutral
AR: ‫إذا كنت تود التعرض للسرقة فعليك بهذا المكان‬ ‫( المكان‬location) BERT-Linear-pair Positive Negative
EN: If you want to be robbed, you should visit this place LCF-BERT Positive
ASGCN-BERT Negative
SenticGCN-BERT Negative
ArBLCSNGCN (ours) Negative
AR: ‫ الغداء افضل بكثير‬,‫ و رغم أن الفطور عادي‬,‫االنترنت عادة جيدة‬ ‫الغداء‬ BERT-Linear-pair Neutral Positive
EN: The internet is usually good, and although breakfast is normal, (lunch) LCF-BERT Positive
lunch is much better
ASGCN-BERT Neutral
SenticGCN-BERT Positive
ArBLCSNGCN (ours) Positive
ً ‫ اإلفطار بطيء‬،‫ والتكييف يعمل‬،‫الفندق جيد! الغرف نظيفة ومريحة‬
AR: ‫قليال ولكن جيد‬ ‫اإلفطار‬ BERT-Linear-pair Positive Neutral
EN: The hotel is good! The rooms are clean and comfortable, the air- (breakfast) LCF-BERT Negative
conditioning works, the breakfast is a little slow but good
ASGCN-BERT Positive
SenticGCN-BERT Positive
ArBLCSNGCN (ours) Neutral

features extracted by AraBERT with the syntactic infor- 5.7 Limitations

mation and word dependency relations from GCN.
For the second example, i.e., ‫إذا كنت تود التعرض للسرقة‬ Introducing a method’s limitations is essential for transpar-
‫( فعليك بهذا المكان‬If you want to be robbed, you should visit ency and understanding its potential shortcomings. In this
this place), LCF-BERT model fails to capture the correct section, we highlight the main limitations of our proposed
label compared to the other GCN-based models. This can model.
be due to the local context expression ‫( عليك ب‬you should)
that it is widely used for recommendation. Instead, GCN- • Dataset limited size Arabic is a low-resource language
based models can capture syntactic dependencies to iden- that encounters challenges due to the limited availability
tify the negative sentiment conveyed by ‫( التعرض للسرقة‬be of corpora and their small sizes, especially in the context
robbed). of ABSA. Consequently, additional efforts are necessary
In the third example, the ASGCN-BERT model pre- to create large-scale corpora in this field and augment the
dicts a wrong label for the aspect ‫( الغداء‬lunch). This can existing ones through the collection of domain-specific
be justified by the fact that this model wrongly empha- data.
sizes the opinion word ‫( عادي‬normal), which is related to • Data imbalance across classes The distribution of aspect
the aspect breakfast. On the other hand, SenticGCN and words across sentiment classes is highly imbalanced, as
our model can focus on the key affective words (i.e., ‫افضل‬ illustrated in Table 3. This can lead to biased predictions
(better)) related to this aspect. and reduced accuracy, particularly in minority classes.
For the last example, both ASGCN-BERT and Sen- To address the imbalance issue, techniques like oversam-
ticGCN-BERT predicted a wrong label for the aspect term pling, undersampling, and SMOTE (synthetic minority
‫( اإلفطار‬breakfast). This can be justified because both oversampling technique) can help balance the dataset. In
models emphasize the opinion word ‫( جيد‬good) and ignore addition, data augmentation methods like back-transla-
other local contextual words. In contrast, with the asso- tion can be explored to further increase the samples for
ciation between local and global context information, our minority classes.
model detects the ambivalence in this sentence based on • Coverage limitations Knowledge bases like BabelSen-
the negative word ‫( بطيء‬slow). ticNet are limited by the data they contain. They may

13
 6 Page 12 of 14 Social Network Analysis and Mining (2024) 14:6

lack coverage for domain-specific or emerging con- Declarations

cepts. This limitation can lead to incomplete or inac-
curate results when analyzing text that falls outside Competing interests The authors declare no competing interests.
the knowledge base’s scope. Therefore, there is a need
to explore appropriate methods for expanding these
knowledge bases by incorporating domain-specific
References
terms and addressing missing words.
• Generic BERT model The AraBERT model used in this Abas AR, El-Henawy I, Mohamed H, Abdellatif A (2020) Deep learn-
study was pre-trained on generic corpora such as Wiki- ing model for fine-grained aspect-based opinion mining. IEEE
pedia content and news articles. While this model can Access 8:128845–128855. https://d​ oi.o​ rg/1​ 0.1​ 109/A
​ CCESS.2​ 020.​
30088​24
capture general language knowledge, it may lack spe-
Abdelgwad MM, Soliman TH, Taloba AI, Farghaly MF (2021) Arabic
cific domain-related terminology, context, and nuances. aspect based sentiment analysis using bidirectional GRU based
To bridge the gap between the pre-trained and fine- models. J King Saud Univ Comput Inf Sci. https://​doi.​org/​10.​
tuned data, we can explore further pre-training of the 1016/j.​jksuci.​2021.​08.​030
Abdelgwad MM, Soliman THA, Taloba AI (2022) Arabic aspect sen-
AraBERT model on in-domain datasets and assess its
timent polarity classification using BERT. J Big Data 9(1):115.
impact on the overall performance of our model. https://​doi.​org/​10.​1186/​s40537-​022-​00656-6
Al-Dabet S, Tedmori S, Mohammad A-S (2021) Enhancing Arabic
aspect-based sentiment analysis using deep learning models.
Comput Speech Lang 69:101224
Alqurashi T (2023) Arabic sentiment analysis for twitter data: a sys-
6 Conclusion tematic literature review. Eng Technol Appl Sci Res 13(2):10292–
10300. https://​doi.​org/​10.​48084/​etasr.​5662
In this paper, we proposed a combined method incorpo- Al-Smadi M, Qawasmeh O, Talafha B, Al-Ayyoub M, Jararweh Y,
Benkhelifa E (2016) An enhanced framework for aspect-based
rating the AraBERT model with GCN and LCF layers to
sentiment analysis of Hotels’ reviews: Arabic reviews case study.
capture the syntactic dependencies relevant to a specific In: 2016 11th international conference for internet technology and
aspect while emphasizing the contribution of local con- secured transactions (ICITST), pp 98–103
text words that are related to this aspect. In addition, we Al-Smadi M, Talafha B, Al-Ayyoub M, Jararweh Y (2019) Using
long short-term memory deep neural networks for aspect-based
explored enhancing the dependency graphs of the sentence
sentiment analysis of Arabic reviews. Int J Mach Learn Cybern
using the affective knowledge form BabelSenticNet data- 10(8):2163–2175
base. Experiments conducted on a reference Arabic hotel Al-Smadi M, Hammad MM, Al-Zboon SA, Al-Tawalbeh S, Cambria E
dataset showed that the proposed model outperformed the (2023) Gated recurrent unit with multilingual universal sentence
encoder for Arabic aspect-based sentiment analysis. Knowl-Based
baseline and related work models, achieving an enhanced
Syst 261:107540. https://​doi.​org/​10.​1016/j.​knosys.​2021.​107540
accuracy score of 92.77%. Additionally, an ablation study Antoun W, Baly F, Hajj H (2020) AraBERT: transformer-based model
was conducted to prove the effectiveness of each com- for arabic language understanding. In: Proceedings of the 4th
ponent of our proposed model. Moreover, we provided workshop on open-source Arabic corpora and processing tools,
with a shared task on offensive language detection
a case study to validate our model’s performance in real
Bensoltane R, Zaki T (2023a) Aspect-based sentiment analysis:
cases. Future work includes integrating additional senti- an overview in the use of Arabic language. Artif Intell Rev
ment features, particularly subjectivity, to evaluate their 56(3):2325–2363. https://​doi.​org/​10.​1007/​s10462-​022-​10215-3
impact on the overall performance of the proposed model. Bensoltane R, Zaki T (2023b) Combining BERT with TCN-BiGRU for
enhancing Arabic aspect category detection. J Intell Fuzzy Syst
In addition, other strategies for building weights of the
44:4123–4136. https://​doi.​org/​10.​3233/​JIFS-​221214
adjacency matrix will be explored. Moreover, we plan to Bisio F, Meda C, Gastaldo P, Zunino R, Cambria E (2017) Concept-
adapt this model to handle other ABSA tasks, including level sentiment analysis with SenticNet. In: Cambria E, Das D,
aspect term extraction, and aspect category sentiment anal- Bandyopadhyay S, Feraco A (eds) A practical guide to senti-
ment analysis. Socio-affective computing, vol 5. Springer, Cham.
ysis. Besides, we intend to modify the proposed model to
https://​doi.​org/​10.​1007/​978-3-​319-​55394-8_9
jointly identify aspect terms and the corresponding senti- Cambria E, Havasi C, Hussain A (2012) Senticnet 2: a semantic and
ment polarities. affective resource for opinion mining and sentiment analysis.
Twenty-fifth international FLAIRS conference
Cambria E, Olsher D, Rajagopal D (2014) SenticNet 3: a common and
Author contributions Rajae Bensoltane: Conceptualization, Methodol- common-sense knowledge base for cognition-driven sentiment
ogy, Software, Validation, Investigation, Writing - original draft, Writ- analysis. Twenty-eighth AAAI conference on artificial intelligence
ing - review & editing. Taher Zaki: Conceptualization, Methodology, Cambria E, Poria S, Bajpai R, Schuller B (2016) SenticNet 4: a seman-
Investigation, Writing - review & editing, Supervision. tic resource for sentiment analysis based on conceptual primitives.
In: Proceedings of COLING 2016, the 26th international confer-
Funding The authors did not receive support from any organization ence on computational linguistics, Technical papers
for the submitted work. Chen J, Yang T, Huang Z, Wang K, Liu M, Lyu C (2023) Incor-
porating structured emotion commonsense knowledge and

13
Social Network Analysis and Mining (2024) 14:6 Page 13 of 14 6

interpersonal relation into context-aware emotion recognition. Pontiki M, Galanis D, Papageorgiou H, Androutsopoulos I, Manand-
Appl Intell 53(4):4201–4217 har S, Al-Smadi M, Al-Ayyoub M, Zhao Y, Qin B, De Clercq O
Costola M, Hinz O, Nofer M, Pelizzon L (2023) Machine learning (2016). Semeval-2016 task 5: aspect based sentiment analysis.
sentiment analysis, COVID-19 news and stock market reactions. International workshop on semantic evaluation
Res Int Bus Finance 64:101881 Rodríguez-Ibánez M, Casánez-Ventura A, Castejón-Mateos F,
Devlin J, Chang M-W, Lee K, Toutanova K (2019) BERT: pre-train- Cuenca-Jiménez P-M (2023) A review on sentiment analysis
ing of deep bidirectional transformers for language understand- from social media platforms. Expert Syst Appl 119862
ing. In: Proceedings of the 2019 conference of the north Amer- Ruder S, Ghaffari P, Breslin JG (2016) INSIGHT-1 at SemEval-2016
ican chapter of the association for computational linguistics: task 5: deep learning for multilingual aspect-based sentiment
human language technologies, vol 1 (Long and Short Papers) analysis. Proc SemEval 330–336
Du Y, Li T, Pathan MS, Teklehaimanot HK, Yang Z (2022) An effec- Soliman AB, Eissa K, El-Beltagy SR (2017) AraVec: a set of Ara-
tive sarcasm detection approach based on sentimental context bic word embedding models for use in Arabic NLP. Procedia
and individual expression habits. Cognit Comput 1–13 Comput Sci 117:256–265. https://​doi.​org/​10.​1016/j.​procs.​2017.​
ElSahar H, El-Beltagy SR (2015) Building large arabic multi-domain 10.​117
resources for sentiment analysis Wang Y, Huang M, Zhu X, Zhao L (2016) Attention-based LSTM
Fadel AS, Abulnaja OA, Saleh ME (2023) Multi-task learning model for aspect-level sentiment classification. In: Proceedings of
with data augmentation for arabic aspect-based sentiment analy- the 2016 conference on empirical methods in natural language
sis. Comput Mater Contin 75(2) processing
Gu T, Zhao H, He Z, Li M, Ying D (2023) Integrating external Wu M (2023) Commonsense knowledge powered heterogeneous
knowledge into aspect-based sentiment analysis using graph graph attention networks for semi-supervised short text clas-
neural network. Knowl-Based Syst 259:110025. https://​doi.​org/​ sification. Expert Syst Appl 120800
10.​1016/j.​knosys.​2022.​110025 Xu C, Luo X, Wang D (2022) MCPR: a Chinese product review data-
Huang B, Ou Y, Carley KM (2018) Aspect level sentiment classifi- set for multimodal aspect-based sentiment analysis. Cognitive
cation with attention-over-attention neural networks. In: Inter- Computing—ICCC 2022, Cham
national conference on social computing, behavioral-cultural Yang Z, Dai Z, Yang Y, Carbonell J, Salakhutdinov RR, Le QV
modeling and prediction and behavior representation in mod- (2019) XLNet: generalized autoregressive pretraining
eling and simulation for language understanding. Adv Neural Inf Process Syst
Huang B, Zhang J, Ju J, Guo R, Fujita H, Liu J (2023a) CRF-GCN: 32:5753–5763
an effective syntactic dependency model for aspect-level senti- Yang H, Zeng B, Yang J, Song Y, Xu R (2021) A multi-task learning
ment analysis. Knowl-Based Syst 260:110125. https://​doi.​org/​ model for Chinese-oriented aspect polarity classification and
10.​1016/j.​knosys.​2022.​110125 aspect term extraction. Neurocomputing 419:344–356
Huang H, Asemi A, Mustafa MB (2023) Sentiment analysis in Yang G, Xu Y, Tu L (2023) An intelligent box office predictor based
E-commerce platforms: a review of current techniques and on aspect-level sentiment analysis of movie review. Wirel Netw
future directions. IEEE Access 29(7):3039–3049
Kastrati Z, Arifaj B, Lubishtani A, Gashi F, Nishliu E (2020) Aspect- Yusuf AA, Chong F, Xianling M (2022) Evaluation of graph convo-
based opinion mining of students’ reviews on online courses. lutional networks performance for visual question answering on
In: Proceedings of the 2020 6th international conference on reasoning datasets. Multimed Tools Appl 81(28):40361–40370
computing and artificial intelligence Zeng B, Yang H, Xu R, Zhou W, Han X (2019) LCF: a local context
Li X, Zhang J, Du Y, Zhu J, Fan Y, Chen X (2023) A novel deep focus mechanism for aspect-based sentiment classification. Appl
learning-based sentiment analysis method enhanced with Emo- Sci 9(16):3389
jis in microblog social networks. Enterp Inf Syst 17(5):2037160 Zhang M, Qian T (2020) Convolution over hierarchical syntactic and
Liang B, Su H, Gui L, Cambria E, Xu R (2022) Aspect-based senti- lexical graphs for aspect level sentiment analysis. In: Proceed-
ment analysis via affective knowledge enhanced graph convo- ings of the 2020 conference on empirical methods in natural
lutional networks. Knowl-Based Syst 235:107643 language processing (EMNLP)
Liu H, Wu Y, Li Q, Lu W, Li X, Wei J, Liu X, Feng J (2023) Enhanc- Zhang C, Li Q, Song D (2019) Aspect-based sentiment classification
ing aspect-based sentiment analysis using a dual-gated graph with aspect-specific graph convolutional networks. In: Proceed-
convolutional network via contextual affective knowledge. Neu- ings of the 2019 conference on empirical methods in natural
rocomputing 553:126526. https://​d oi.​o rg/​1 0.​1 016/j.​n eucom.​ language processing and the 9th international joint conference
2023.​126526 on natural language processing (EMNLP-IJCNLP)
Ma D, Li S, Zhang X, Wang H (2017) Interactive attention networks Zhao A, Yu Y (2021) Knowledge-enabled BERT for aspect-based
for aspect-level sentiment classification. IJCAI sentiment analysis. Knowl-Based Syst 107220
Madan M, Rani A, Bhateja N (2023) Applications of named entity Zhao P, Hou L, Wu O (2020) Modeling sentiment dependencies
recognition using graph convolution network. SN Comput Sci with graph convolutional networks for aspect-level sentiment
4(3):266. https://​doi.​org/​10.​1007/​s42979-​023-​01739-8 classification. Knowl-Based Syst 193:105443
Mikolov T, Chen K, Corrado G, Dean J (2013) Efficient estimation Zhou J, Huang JX, Hu QV, He L (2020) SK-GCN: modeling syntax
of word representations in vector space. Preprint http://​arxiv.​ and knowledge via graph convolutional network for aspect-level
org/​abs/​1301.​3781 sentiment classification. Knowl-Based Syst 205:106292
Omar A, Abd El-Hafeez T (2023) Quantum computing and machine Zhao H, Xie J, Wang H (2022a) Graph convolutional network based
learning for Arabic language sentiment classification in on multi-head pooling for short text classification. IEEE Access
social media. Sci Rep 13(1):17305. https://​d oi.​o rg/​1 0.​1 038/​ 10:11947–11956
s41598-​023-​44113-7 Zhao Z, Tang M, Tang W, Wang C, Chen X (2022b) Graph convo-
Pennington J, Socher R, Manning CD (2014) Glove: global vectors lutional network with multiple weight mechanisms for aspect-
for word representation. In: Proceedings of the 2014 conference based sentiment analysis. Neurocomputing 500:124–134.
on empirical methods in natural language processing (EMNLP) https://​doi.​org/​10.​1016/j.​neucom.​2022.​05.​045

13
 6 Page 14 of 14 Social Network Analysis and Mining (2024) 14:6

Zhao Y, Zhang L, Zeng C, Lu W, Chen Y, Fan T (2023) Construction Springer Nature or its licensor (e.g. a society or other partner) holds
of an aspect-level sentiment analysis model for online medical exclusive rights to this article under a publishing agreement with the
reviews. Inf Process Manage 60(6):103513 author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of
Publisher's Note Springer Nature remains neutral with regard to such publishing agreement and applicable law.
jurisdictional claims in published maps and institutional affiliations.

13