A survey of sentiment analysis methods based on graph neural network

doi:10.21203/rs.3.rs-3173515/v1

Download PDF

Research Article

A survey of sentiment analysis methods based on graph neural network

https://doi.org/10.21203/rs.3.rs-3173515/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Sentiment analysis is an active research field as one of the most popular tasks of natural language processing, which aims to extract valuable information from various social platforms and extensive online texts to process and find people's attitudes in business and advertising, government, economic fields, and even political orientations. Hence, researchers have made many efforts in this field, which mainly refer to traditional approaches based on dictionaries, machine learning, and deep learning models. Graphs as a robust and interpretable data structure have been considered for applications of artificial intelligence models such as machine vision and natural language processing which are used for learning non-structured data like text or images. Although deep learning methods have achieved promising results in this field, due to problems such as assigning indecisive weights and high dimensions in feature extraction stages, they are still a “black box.” Meanwhile, graph neural networks (GNNs) are a particular type of deep neural network that are interpretable and flexible. Their adaptability in solving complex problems in data analysis with a graph structure has made them one of the most efficient methods in the last decade. Considering the large amount of textual information in social media and various online platforms, sentiment analysis or opinion mining aims to help marketing strategies for business owners and awareness of the attitude of public opinion in governments has become one of the crucial issues in today's modern societies. This comprehensive review focuses on GNN-based approaches in sentiment analysis and summarizes the recent state-of-the-art in this area. Also, we discussed their weaknesses and strengths, and challenges on specific datasets. Our goal is to show the development process and the potential of GNN-based approaches in different problems of sentiment analysis compared to previous methods and to help find more effective directions for researchers interested in this field.

Natural language processing

Deep learning

Graph neural networks

Sentiment analysis

The increase of users in social networks and the creation of a large amount of data, especially opinions, and sentiments that they can freely share on these online platforms, make it necessary to be aware of these opinions and public attitudes for various economic, political, and management fields. Therefore, sentiment analysis or opinion mining as a subset of text classification tasks has become one of the most attractive topics in recent years. Many efforts have been made in this field with different techniques that can be categorized into three general groups containing rule-based, automated, and hybrid methods.

Rule-based or knowledge-based methods known as a lexicon-based approach (Trinh et al. 2016) score a pre-prepared text by sentiment lexicon. In these methods, the text can be analyzed without training and only by relying on a set of rules that are influential and unambiguous lexicons. For example, these explicit words are “good,” “happy,” “sad,” and “boring.” Some knowledge-based methods allocate arbitrary words a possible dependency on specific emotions in addition to the provide a list of apparent words.
On the other hand, in automated or Statistical methods, sentiment analysis performs by elements of machine learning algorithms such as support vector machine (SVM), Random Forest (RF), Naïve Bayes (NB), Bag of Words (BOW), word embedding models, and deep learning approaches.
Hybrid methods are efficient and widely used techniques that are the combination of both above approaches that sentiment analysis has archived high accuracy and stable results by them. Despite the considerable results of these methods, there are various limitations and shortcomings which are mainly related to the capability of taking the global long-range dependent and should be improved. Recent progress in deep neural networks led to the extension of the GNN model, which can extract multi-scale features and overcome the convolution neural networks (CNN) model's limitations. They provided a generalized form from Euclidean space to Non-Euclidean data space (LeCun et al. 2015). In the last decade, GNNs have been a powerful tool for representation and processing graph data and it is worthy of more research attention as new geometric deep learning in various fields of NLP like sentiment analysis.

In our work, there are many specialized terms that we briefly describe. In Table 1, we have prepared a list of these abbreviations with their description. In the following, the structure, tasks, and applications of graph neural networks are discussed, especially in natural language processing tasks.

Table 1 A list of specialized abbreviations used in the article and their description

Abbreviation	Description
ABSA	Aspect-Based Sentiment Analysis
TTL	Two-Phase Transfer learning
BERT	Bidirectional Encoder Representations from Transformers
Bi-LSTM	Bi-directional Long Short-Term Memory
BOW	Bag of Words
CLSA	cross-lingual sentiment analysis
DBN	Deep Belief Networks
DNN	Deep Neural Network
DT	Decision Tree
ELMo	Embedding Language Models
GAT	Graph Attention Network
GRU	Gated recurrent unit
GCN	Graph Convolution Network
Glove	Global Vectors
GNN	Graph Neural Network
GPT	Generative Pre-trained Transformer
HAGNN	heterogeneous aspect graph neural network model
H-GAT	Heterogeneous Graph Attention
HGNN	Hypergraphs Graph Neural Network
KNN	K nearest Neighbors
LR	Logistic Regression
LSTM	Long Short-Term Memory
ME	Maximum Entropy
MLSA	Multilingual sentiment analysis
MN	Memory Network
MPNN	Message-Passing Neural Networks
MSA	Multimodal Sentiment Analysis
MSA	Multimodal Sentiment Analysis
NB	Naive Bayes
NLP	Natural Language Processing
RecNN	Recursive Neural Networks
RF	Random Forest
RGAT	Relational graph attention network
RGNN	Recurrent Graph neural network
RNN	Recurrent Neural Network
SG	Skip Gram
SGCN	Signed Graph Convolution Network
SAGAT	Syntax-Aware Graph Attention Network
SVM	Support Vector Machine
TF-IDF	Term Frequency-Inverse Document Frequency
ReMemNN	Recurrent Memory Neural Network
MNHMA	Memory Network hierarchical multi-head attention
HFV	Hybrid Feature Vector
RRF	Review-Related Features
ARF	Aspect-Related Features
ATE	Aspect Terms Extraction
ACD	Aspect Categories Detection
OTE	Opinion Term Extraction
ASC	Aspect Sentiment Classification
LDA	Latent Dirichlet Allocation
TD	Target Dependency
CDT	Convolution Dependency Tree
WGCN	Weighted Graph Convolutional Network
LTP	Language Technology Platform
MHA	Multi-Head Attention Mechanism

1.1 GNN structures

GNNs are a typical class of neural network that can be performed to graph-structure data. This type of data describes a set of entities and their relationship by nodes and edges, respectively. A form of GNN named message-passing neural networks (MPNN) was introduced by Gilmer et al. (2017) as a spatial-based graph filter.

This type of GNN is the principal design element and an effective method for designing graph neural networks and getting good generalizations because of its ability to learn local structures and permutation equivariance. However, this model has limitations, such as learning topological features and representation power. Vignac et al. (2020) proposed a powerful message-passing structure on the ZINC dataset which propagates a one-hot encoding of the nodes that lead to learning a local context matrix that include great local information near each node and a way for the parameterization of the message that led to effective results on molecular graph regression.

Different GNN architects based on the message-passing model have been proposed recently. However, researchers are considering other probable GNN structured “going beyond” message passing as an open research question. Furthermore, Graphsage, introduced by Hamilton et al. (2017), is another spatial-based representation learning model for dynamic graphs. Several stacked layers them can make semantic and structural features level. The heterogeneous or multi-relational graphs are another variant of graphs applied by more GNN models using meta-path (Wang et al. 2019c; Fu et al. 2020). Linmei et al. (2019) proposed encoding a heterogeneous graph attention (H-GAT) that edges represent connectivity. An et al. (2023) proposed a heterogeneous aspect graph neural network model (HAGNN) with three types of nodes: word, aspect, and sentence. They take internal relationships between sentences and aspects by heterogeneous graphs to learn semantical and structural knowledge. There are other variants of graphs in some literature like signed graph convolution network as SGCN (Derr et al. 2018) and hypergraphs graph neural network as HGNN (Feng et al. 2019). Zhou et al. (2020) reviewed variants of the GNN model and general design pipeline and systematically categorized the applications into structural and non-structural.

Generally, there are some common types of GNN in the majority of literature, which generally can be categorized into the following groups: Recurrent Graph Neural Network (RGNN), Spatial Convolutional Network, Spectral Convolutional Network, which was first introduced by Bruna et al. (2014), and Graph Attention Network (GAT) which is part of spatial approaches. Wu et al. (2019b) provide a comprehensive overview of GNN models.

Recurrent Graph neural network (RGNN)

After looking at the expanding deep learning approaches such as RNN and CNN, the success of GNN as a powerful data structure has attracted many researchers to use GNN instead of traditional methods in NLP. Meanwhile, RGNN as a specific case of RNN has been introduced. Ioannidis et al. (2019) have proposed a new RGNN structure for scalable semi-supervised learning to use multi-relational graphs that be able to detect complex relationships and non-linear data correlations, combine them, and finally scale based on graph size that has good performance. Compared to previous models. Chen et al. (2020b) proposed an RGNN model for processing a sequence of passage graphs. The most widely used multi-relational GNN models are RGCN (Schlichtkrull et al. 2018) and relational GGNN (R-GGNN) (Beck et al. 2018).

Spatial and spectral graph convolutional network (GCN)

The spatial convolutional network approach is inspired by CNN, which is directly applied to graphs. However, spectral-based graph ﬁlters have a powerful mathematical basis based on graph signal processing theory (Shuman et al. 2013). Generally, both of them started with different foundations but they have similar propagation rules. Graph convolutional networks (GCN) are a common typical sample of spectral-based graphs (Kipf and Welling 2016; Zhang et al. 2019b). In recent years GCNs have been developed in many NLP applications (Yao et al. 2019)

Graph attention network (GAT)

The attention-based approach is an impressive idea in the field of NLP. Some developed models such as Bidirectional Encoder Representations from Transformers (BERT) (Devlin et al. 2019; Huang and Carley 2019; Chen et al. 2020a) and GPT (Ethayarajh 2019) as transfer learning or pre-training models rely on attention techniques. Lee et al. (2018a) have prepared a review on graph attention models. GAT models pay attention to different parts of the context by specific values, Unlike GCNs which perform all the contexts similarly. In other words, GAT models utilize the multi-head attention mechanism inspired by applying this technique in the Transformer model (Velickovic et al et al. 2018; Vaswani et al. 2017). Because GNN models are not able to adjust edge connection information during the learning process but the GAT model as a special case of MPNN is allocated different weights for each neighbor (higher score for important neighboring nodes) to get better results and reduce noise. Therefore, by attention mechanism, the similarity between the target node and the neighbor is considered easily. Other types of GATS are the Relational graph attention network (RGAT) and gated attention network (GaAN) (Wang et al. 2020a; Zhang et al. 2018a).

1.2 Tasks of GNN

Depending on what kind of task we want to apply in a graph as structured data, there are three tasks: node, edge, and graph levels. At the node level, which contains node clustering, classification, and regression, focus on nodes. In edge-level, which refers to link prediction or edge classification tasks, the model predicts an edge between two nodes or classifies edge type and graph level, including graph matching, graph classification, etc. Some applications in node classification which focus on predicting and assigning the label of nodes include YouTube, citation, finding friends relationships, etc. Cetoli et al. (2017) proposed decency trees significantly impact entity detection by GCN for improving Bi-LSTM. Yao et al. (2019) proposed text graph convolutional networks by heterogeneous graphs that convert document classification to node classification problems. The purpose of link prediction is to predict the relationship of entities in a graph (Zhang and Chen 2018c; Rossi et al 2021). Graph classification can be applied in many practical domains like NLP tasks such as text classification (Peng et al, 2018; Huang et al 2019b; Zhang et al. 2020b; Lee et al. 2018b). For graph learning, we have a variety of training, such as Supervised, Semi-supervised, transductive, and Unsupervised.

1.3 Some applications of GNN

In this section, we introduce some of the top applications of GNN in the real world and artificial intelligence domain depending on a data structure.

1.3.1 Computer vision (CV)

In recent years GNNs have been developed in many fields of CV, such as image retrieval (Kan et al. 2022), image generation, understanding, reasoning (Pradhyumna et al, 2021), social relationship (Wang et al. 2018), visual question answering (Wang et al. 2018; Narasimhan et al. 2018), and interaction detection (Qi et al. 2018), As a generalization form of neural networks, CNN has achieved outstanding performance in many problems, such as object detection (Hu et al 2018). However, they need help identifying semantic relations between them. GNNs are powerful tools for analyzing graphs, and this new area needs more attention to solving many applications in the computer vision domain.

1.3.2 Natural language processing (NLP)

As mentioned earlier, GNNs can be applied to non-structured data such as images and text in various problems. In addition to GNN applications in CV, they are new methods for NLP tasks such as text classification (Yao et al. 2019; Hu et al. 2019), machine reading comprehension (MRC), and question answering (Chen et al. 2020b; Qiu et al. 2019; Tu et al. 2019; Zeng et al. 2021), Relational Reasoning (Palm et al. 2018; Zhang et al. 2020a), interaction detection (Qi et al. 2018), relation extraction (Sahu et al. 2019; Zeng et al. 2020) and sentiment analysis. GNN utilizes the internal relationship of words in the text and builds a syntactic model by considering different parts of sentences. In contrast, RNN or LSTM consider only sequential form, so they approach the problem differently. Wu et al. (2021) provide a literature survey regarding the GNN-based approach for different NLP tasks which gives an overview to researchers in this field. They organized GNN-based methods into three sections: construction of graph, graph representation, and encoder-decoder based on graphs.

1.3.3 Other applications

The other applications of GNNs refer to Structural scenarios such as Physics (Sanchez et al. 2018; Kipf et al. 2018), Chemistry (Do et al. 2019b), Biology regarding disease classiﬁcation (Rhee et al. 2018), Combinatorial Optimization (Zheng et al. 2020b; Sato et al. 2019) traffic prediction (Zheng et al. 2020a; Yu et al. 2018) knowledge graph (Xu et al. 2019a; Shang et al. 2019), recommendation systems (Wu et al. 2019a; Wang et al. 2020b), graph generation (Shi et al. 2020), etc.

NLP aims to understand a text to accomplish various tasks. Since word embedding by deep learning-based approaches is not able to cover all the semantic and syntactic structures of the text, researchers utilize the hybrid of deep learning models and graph-structured representations for various NLP tasks. Sentiment analysis as a subtask of NLP due to the huge amount of data in online platforms and the need of the owners of industries and governments to evaluate the opinions of users in various fields of marketing and politics has been extremely noticed by researchers. As mentioned earlier, there are various approaches to sentiment analysis which are categorized into three groups: lexicon-based, automated, and hybrid approaches. The traditional model relies on text sequence representation and makes weak feature space and ignores syntactic ingredients. Deep learning models are currently popular and are widely used at different levels such as document-level, sentence-level, and aspect levels but they suffer some limitations such as high computational cost, overfitting, and fast convergence due to high learning time speed and poor interpretability. In recent years, researchers have become increasingly interested in GNN-based methods in sentiment analysis tasks such as node classification and graph representation. They utilize dependency graphs for the representation of information between features in a parser tree in which the set of vertices represents the words and the edges represent the better relationship between two words in the sentence and record the long distance between aspect and opinion word and provide a differentiated syntactic path. In recent years, a relatively large number of review articles on deep learning-based sentiment analysis have been published, but few review articles have mentioned approaches based on graph neural networks. In a review article, Luo et al. (2021) dealt with aspect-based sentiment analysis (ABSA) methods. They believe that ABSA methods can be classified into two categories: transformer-based and GNN-based methods. They also reviewed common frameworks, datasets, and evaluation metrics based on the proposed approaches. Pham et al. (2023) mentioned the recent developments in the combination of text representation and graph-based approaches named TG-GNN to overcome some limitations of deep neural network methods. In this survey article, we only focus on one of the most popular natural language processing tasks, namely sentiment analysis based on graph neural networks, to demonstrate the potential of this new approach compared to previous tradition and deep learning methods in this field and managing complex features and essential words and extraction of more detailed semantic and structural information.

The main contribution of this survey article can be summarized as follows:

We first introduced the structure of GNN and their application in the field of NLP, especially sentiment analysis to achieve to general view and acceptable knowledge in this relatively new field for interested researchers
Our most significant goal is to show the potential GNN-based approaches to sentiment analysis tasks compared to traditional machine learning methods and recent deep learning approaches and overcome the restrictions and problems due to them and design more effective sentiment analysis systems.
We examine the recent process of developing the approach of emotional analysis based on the neural network of graphs.
We have tried to cover recent state-of-the-art sentiment analysis based on GNN approaches at different levels.
We also explained the performance of each approach, its strengths and weaknesses, experimental results based on different databases, and some future directions.

The literature survey paper is organized in section 2 background and related work on sentiment analysis approaches from traditional models to developed deep learning-based methods have been investigated. In section 3, we present a list of the recent state-of-the-art GNN-based methods in sentiment analysis at different levels. Also, we investigate the weaknesses and strengths, and challenges of GNN-based models in detail. Finally, in section 4, we introduce the most popular datasets for researchers interested in this area.

2.1 Approaches

Sentiment analysis as a subfield of NLP gets a lot of attention from researchers, governments, business owners, and the industry. Liu (2012) has a definition that describes each opinion containing a target or a feature of an entity or mood that almost is classified (positive, negative, or neutral). Sentiment analysis problems have various challenges that are generally categorized into three groups: structured, semi-structured, and unstructured. Wankhade et al. (2022) investigated some of the significant challenges in sentiment analysis data based on the structure of data. Some of the most important challenges of textual data are their incorrect spelling. Also, the presence of slang and ambiguous words, tone, polarity, sarcasm, Emoticons, negation, and idioms are a variety of limitations that lead to incorrect analysis.

Numerous efforts and methods have been done in the sentiment analysis area. Traditional approaches such as lexicon-based (Trinh et al. 2016) and machine learning-based (Alhajj and Rokne 2014) have achieved a high accuracy but because of their time-consuming and feature extraction, recent researches necessitate complex and reliable methods for getting more accurate results. Deep learning techniques have impressive performance and outperform traditional approaches in sentiment analysis. In the following section, various sentiment analysis approaches have been presented.

2.1.1 lexicon-based approaches

There are two fundamental approaches to making sentiment lexicons: dictionary-based and corpus-based (Bittar et al. 2021).

Dictionary-based methods include a predefined set of opinion words that are collected manually and then this list is expanded by checking synonyms and antonyms (Kaity and Balakrishnan 2020; Singh et al. 2017). Some advantages and disadvantages of these methods are as follows: They don't need to train data and achieve good outcomes in a limited range. Also, they have quick access to the dictionary of word meanings but they are domain orientated and unable to find opinion terms, so words in one field can only be used in the same field (Moreo et al. 2012). Some examples of a lexicon in sentiment analysis are WordNet-Affect (Fellbaum 2017), MPQA (Alshari et al. 2018), Bing Liu's Sentiment Lexicon, LingPipe, etc. Developing the lexicon method is a crucial challenging domain because of the mass of unstructured data generated in social media, and traditional methods such as Bag of Words need to distinguish Adjectives and adverbs from verbs or propositions. For example, “like” in two sentences, “I like you” and “I am like you,” has different meanings, and part of speech in the first case is a verb with positive polarity and in the following case is just a neutral proposition. Finding opinion words particular for each domain as the polarity is one of the other disadvantages of dictionary-based methods (Hajek et al. 2020). To overcome these limitations, corpus approaches use semantic and syntactic structures to detect the sentiment of a sentence and provide a more accurate result. Corpus-based methods contain collections of texts that are collected based on specific criteria. They include statistical and semantic approaches that try to adopt new data types.

Statistical approaches found co-occurrence patterns by statistical analysis and are mostly used in sentiment analysis applications. LSA is one statistical model for analyzing and finding semantic qualities (Cao et al. 2011).

Semantic approaches that depend on relies on sentiment-rich words estimate the similarity score between tokens in text sentiment analysis (Hershcovich and Donatelli 2021).

Corpus-based methods face various challenges such as negation, high-dimensionally, abbreviations, and complex structural and cultural textual data. Hence, the development of machine learning approaches compared to dictionary-based methods, led to improved performance and increased accuracy. In the following, various machine learning approaches are discussed.

2.1.2 Machine learning approaches

This approach uses linguistic and/or syntactic elements for sentiment analysis. Furthermore, they can be trained for complex cases such as perceiving contextual, negation, sarcasm, misapplied words, etc. They can be categorized into supervised, semi-supervised, and unsupervised learning methods. There are numerous models based on machine learning for sentiment analysis problems, such as logistic regression (LR), Naive Bayes (NB) (Hajek et al. 2020), Support vector machine (SVM) (Imamah et al. 2020), K‑nearest neighbors (KNN) Decision tree (Revathy and Lawrance 2017; Patel and Prajapati 2018), Maximum Entropy (ME) (Bergsma et al. 2012), neural network, Semi-supervised learning (Janjua et al. 2021). Van de Camp and Van den Bosch (2012) proposed a method with a single-layer neural network and SVM and got improved results. Moraes et al (2013) in their empirical analysis of the comparison of SVM and neural networks in sentiment analysis, concluded that neural networks have no attention. In another investigation, Ravi and Ravi (2015) showed that neural network has better performance than SVM, but their main disadvantage is related to computational cost. For the efficient performance of systems, Al Amrani et al. (2018) proposed a hybrid approach based on machine learning, including SVM and RF. Wankhade et al. (2022) published a survey and presented a list of various methods for sentiment analysis and sentiment classification and multimodal at different levels based on machine learning and deep learning from 2010 to 2021. They stated that SVM and NB algorithms compete with newly proposed methods. Also, they comprehensively compared machine learning methods and some deep learning methods based on their advantages and disadvantages which we suggest to study for better understanding of researchers. Although hybrid approaches in machine learning improve the performance of lexicons-based and individual models, deep learning algorithms have shown their superiority over more traditional techniques.

2.1.3 Deep learning approaches

Deep learning is the subfield of machine learning approaches presented by deep neural network models. The principal layer of data processing in sentiment analysis systems based on deep learning methods is word embedding, which expresses words in vector space. The first idea returns to the 1960s and developed models such as Word2vec, which has two models, continuous bag of words (CBOW) and SG (Skip Gram) (Mikolov et al. 2013; Alami et al. 2019). Global Vectors (Glove2), and Embedding Language Models (ELMo) for improving traditional models such as Term Frequency-Inverse Document Frequency (TF-IDF) and n-grams have been presented. Convolution neural networks (CNN) as feedforward neural networks have gotten good results in the NLP area, especially sentiment analysis, because of their ability to contextual characteristics, but they have long-term connection problems. To overcome these constraints, RNNs models (recurrent neural networks) were proposed, such as LSTM (Long Short-Term Memory) and GRU (Gated recurrent unit) (Cheng et al. 2020). Hao et al. (2018) proposed Bidirectional RNN with better results because it can simultaneously feature extraction and sentiment analysis.

The other developing RNN models followed as (Bi-LSTM) (Abid et al. 2019), Aspect Fusion LSTM (AF-LSTM) (Tay et al. 2018), Aspect-Aware LSTM (AA-LSTM) (Xing et al. 2019). Recursive neural networks (RecNNs) are a generalization type of RNN that can discover hierarchical patterns and get structured predictions recursively while RNNs can process sequential data. Sadr et al. (2019) proposed a robust sentiment analysis model based on the combination of CNN and RecNNs that used RecNNs instead of a pooling layer to reduce some limitations in CNN such as loss of local information and long-term dependency. Also, Aydin and Güngör (2020) proposed a combination of RNN and RecNNs for ABSA using inter-aspect relations. They first used RecNNs for building and training the recursive neural trees, then used their output as input in the RNN model and achieved significant results compared with baseline methods.

The attention mechanism in deep learning was first introduced by Bahdanau et al. (2014) where the decoder would have finite availability to the information prepared by the input. In other words, the attention mechanism permits the encoder to utilize the most relevant vectors by focusing on important parts of a sequence. This technique in NLP is one of the most valuable progresses in deep learning areas which caused many improvements in transformer architecture and Google’s BERT. Kardakis et al. (2021) prepared a comparative analysis that shows the effect of attention-based models built on RNN in recognizing opinions and emotions in movie reviews. The results indicated that attention-based models got a 3.5% improvement in their accuracy compared with the baseline models without the use of attention. In the last decade, attention-based models in sentiment analysis are expanding. Wang et al. (2019b) for avoiding dependency on manual annotation and reducing errors proposed a hybrid model based on an attention mechanism and Bi-directional Long-Short-Term Memory (Bi-LSTM) for Stanford Sentiment Treebank (SST) dataset. They showed in their research that the accuracy increased by 2.787% and 1.946% for SST-1 and SST-2, respectively rather than Conv-RNN methods. From the attention mechanism view, RNN models because of learning long-term dependencies have better results in sentiment analysis than CNN-based models while CNNs have their superiority. Usama et al. (2020) proposed a model with the merits of both architectures simultaneously. In this model first, CNN learns high-level features of input text with an attention score and finally used these features in the RNN model to process them consecutively. In this way, by using the attention scheme their proposed model focused on significant features and achieved good results on three benchmark datasets.

Deep learning models have been used for sentiment analysis and achieved significant results. In contrast, some problems like losing aspect information in pre-trained word embedding and poor interaction between remarkable aspects and context in the attention-based models led to sentiment analysis depending on limited aspects. There are other active areas of research in deep learning-based methods such as deep belief networks (DBN) and Memory Networks (MN) which have provided new opportunities for future research (Xiao et al. 2019; Shen et al. 2020). Liu and Shen (2020) proposed a novel method termed Recurrent Memory Neural Network (ReMemNN). They used adjustment word embeddings and a multielement attention mechanism to resolve the weakness of word embedding and interaction between context and aspect. Chen et al (2021) proposed a comprehensive context representation for aspect-based sentiment analysis (ABSA) based on a memory network model with hierarchical multi-head attention (MNHMA). Extracting semantic information mechanism by rotational unit with hierarchical multi-head attention for keeping aspect information and fully connected layers in each attention layer are prominent features of this method.

Several survey articles have been published on deep learning-based sentiment analysis methods. In Table 2, we list some of the authentic papers that were published in recent years. Also, in this section, we present a comparative discussion of the performance of different sentiment analysis articles based on deep learning for further understanding.

Ain et al. (2017) published a survey on the sentiment analysis by various deep learning models such as CNN, RNN, RecNNs, Deep Belief Networks (DBN), and many more for solving various problems like sentiment classification, visual and textual analysis, cross-lingual and product review analysis Separately. Finally, they summarized all the best studies discussed in this review and concluded that deep neural networks are better than normal neural networks and SVM due to the presence of hidden layers and more accurate training and automatic feature extraction.

Zhang et al. (2018b) appropriated the deep learning approaches into three sentiment analysis levels: document-level, sentence-level, and aspect-level, and explained the applied algorithms per analysis level. According to their study, the most common algorithms for document-levels are CNN, LSTM, Bi-LSTM, GRU, and MN, and for sentence-levels are dynamic CNN, LSTM, RecNNs, CNN-LSTM, Bi-LSTM, Recurrent Random Walk Network, and hybrid models and finally for Aspect-levels are Attention-based mechanisms such as Attention-based LSTM, Interactive Attention Network, Adaptive RecNNs Neural Network, MN, RNN Attention Networks and hybrid models.

Prabha and Umarani Srikanth (2019) proposed a detailed review of sentiment analysis techniques based on the most popular deep-learning models containing CNN, RNN, and LSTM independently or in a combination form of them in sentence and aspect levels. They also discussed the advantages and disadvantages of methods based on their performance parameters.

Do et al. (2019a) reviewed More than 40 articles about deep learning methods for aspect-based sentiment analysis (ABSA) According to their study, ABSA with deep learning models is in the primary stages and has some challenges like linguistic complications, and multi-lingual and labeled data.

Yadav and Vishwakarma (2020) published a survey article and investigated 130 papers that had used deep learning models such as CNN, RNN, RecNNs, LSTM and GRU, Bi-RNN, Deep Belief Networks (DBN) which LSTM had the best result among them while these methods depend on a large number of datasets.

Lin et al. (2020) published a survey of sentiment analysis methods based on deep learning models containing CNN, RNN, LSTM, deep neural network (DNN), DBN, and MN in ABSA and multimodal sentiment analysis. They compared these methods based on their advantages and disadvantages of them. they concluded that the future direction can include having high-quality datasets, reducing training time and increasing accuracy, optimizing feature extraction, using deep learning-based methods with fuzzy rules, and considering the reinforcement learning methods and pre-training models in sentiment analysis tasks. Also, they stated that the results of coarse-grained multimodal sentiment analysis are mostly better than fine-grained multimodal sentiment analysis due to the lack of labeled data. They believe multi-label prediction is an interesting research area.

Dang et al. (2020) reviewed 32 papers that applied used most widely deep learning algorithms such as CNN, simple RNN, Recurrent Neural Tensor Network (RNTN), LSTM, and its developed form like tree-LSTM, discourse-LSTM, Cognition attention-LSTM, Bi-LSTM, and GRU on eight datasets. According to their research, RNN models got the best efficiency.

Habimana et al. (2020) in their survey article categorized deep learning methods based on several sentiment analysis levels as follows: CNN, RNN, Attention-based, and Adversarial Network Models CNN, RNN, Deep Reinforcement Learning (DRL), RNN, and RNN with Cognition attention-based models in sentence levels and CNN, RNN, RNN models with attention memory, Attention-based models and hybrid models.

Minaee et al. (2020) published a comprehensive study of over 150 papers on various NLP tasks such as sentiment analysis, natural language inference, topic analysis, news categorization, question answering by deep learning algorithms containing RNN, CNN, Transformers, GNN Capsule Neural Networks, attention-based mechanism, Memory augmented networks, Hybrid models, DRL, Autoencoders, and Siamese Neural Networks that showed the considerable effect of deep learning models in text classification.

Joseph et al. (2022) published a survey on deep learning-based sentiment analysis that investigated various Deep learning methodologies namely CNN, RNN, LSTM, GRU, and their variants on different datasets. They concluded that deep learning models perform better than machine learning models like SVM due to the handling of huge data sets and existing hidden layers. LSTM and GRU both are better than RNN, because of their capability to catch long-term dependency.

Mercha et al. (2023) in their study presented an overview of the methods used to perform sentiment analysis across languages based on multilingual (MSA) and cross-lingual sentiment analysis (CLSA) approaches. They covered most of the machine learning and deep learning approaches to provide novel directions. They believe that deep learning techniques have good results but the scarcity of multilingual annotated data limits the development and comparison of multilingual and cross-lingual methods. Generally, CLSA methods refer to using transfer learning between a source and target languages based on (Two-Phase Transfer learning) TTL and Cross-Language Record Linkage (Hettiarachchi et al. 2023) while MSA methods tend to use language-agnostic and translation-free systems and provide predictions independent of the source language.

In other survey articles published in the field of sentiment analysis, other different opinions about the performance of deep learning approaches have been expressed which can help the active researchers in this field to understand more. For instance, Ligthart et al. (2021) have provided a tertiary study. They investigated 112 articles published in 2020 on sentiment analysis based on deep learning methods. According to their study, the most applied algorithm is LSTM (35.53%) Second algorithm is CNN (33.33%) and the next common algorithm is GRU (8.77%). Other deep learning methods such as RNN, BERT, DNN, RecNN, GCN, and hybrid models got lower ranks in that list that indicated most of the deep learning approaches followed the supervised learning machine learning algorithms. Colón-Ruiz and Segura-Bedmar (2020) in interesting research, presented a benchmark comparison of deep learning methods (different architecture of CNN, LSTM, and hybrid of these models) in sentiment analysis in the pharmaceutical field to obtain information about the effectiveness and side effects of drugs. They conducted experiments on datasets from the Drugs.com website in 10 classes (0 to 9 scores for each review) based on the patient’s degree of satisfaction with the drug. They found CNN models with static word embedding has poor efficiency for low training data while CNN requires less training time than hybrid and Bi-LSTM models. Considering the recent popularity of the BERT method in the field of sentiment analysis, they utilized the combination of BERT and Bi-LSTM models and concluded that BERT models improve performance despite the high computational cost. In the latest survey paper published by Cui et al. (2023a) the evolution of the research methods and tools in sentiment analysis tasks have been presented. they found that deep learning and hybrid methods have gradually become research trends. For further investigation of methods of this research field and getting a comparative view, we recommend reading this article.

2.1.4 Transfer learning

Transfer learning approaches are a kind of pre-train model which uses the acquired knowledge of the trained model to transfer to the new model. It is obvious that in this model we need less training time because of doesn’t require any explicit training data. In recent years, transfer learning models have revolutionized machine vision and natural language processing such as sentiment analysis due to their computational power. Bartusiak et al. (2015) used traditional models like N-gram and Bi-gram for encoding complex words and phrases for document-level sentiment analysis. They used two different datasets to transfer knowledge from one domain to another Meng et al. (2019) proposed a multiple-layer CNN with a transfer learning approach that used the features from a pre-trained model and fine-tuned weights of fully connected layers. Since 2018, the BART model has made a big change in the field of natural language processing, so most researchers have used it as an alternative to the LSTM and Bi-LSTM models Han et al. (2021). The main advantage of BERT models is that in addition to being less expensive and faster for training, they overcome the exploding gradient issue. BERT utilizes transformers and consists of encoders-based transformers and trainers in two stages including pre-training and fine-tuning which can be done as per the different tasks such as sentiment analysis (Singh et al. 2021), emotion detection (Acheampong et al. 2021) and aspect detection (Li et al. 2019)

2.1.5 Hybrid approaches

Sentiment analysis is a combination of statistical and knowledge-based methods and hybrid models as one of the approaches of this area refers to the combination of both lexicon and machine learning-based approaches to achieve optimal results. Most researchers believe that hybrid models increase accuracy and get an impressive performance in the NLP area especially sentiment analysis tasks (Shoukry and rafea 2015). Some of the hybrid models structured are a combination of CNN-RNN (Basiri et al. 2021), CNN-LSTM (Rehman et al. 2019; Jing et al. 2021), CNN-Bi-LSTM (Minaee et al. 2019), GRU-RNN (Shrestha and Nasoz 2019), etc. hybrid approaches which use both lexicon-based and automated learning models are still an interesting topic for researchers in this field and need more research. Hassonah et al. (2020) proposed a novel method based on a hybrid filter and evolutionary wrapper approach for sentiment analysis on Twitter. they found that LSTM-based RNN + Glove performs better than individual models and hybrid models with appropriate design and correct selection of hyperparameters have efficient performance. Kaur et al. (2023) developed a deep learning-based model by hybrid feature extraction including review-related (RRF) and aspect-related features (ARF) for making a hybrid feature vector (HFV) for each review. Sentiment classification is performed by LSTM. They evaluated the proposed model by three groups of datasets including SemEval-2014, STS-Gold, and Sentiment140 and the HFV + LSTM model achieved 94.46%, 91.63%, and 92.81% F1-Score values, respectively.

Table 2

Survey articles of sentiment analysis based on deep learning models
References	Title	Retrieved from	Year
Ain et al. (2017)	Sentiment analysis using deep learning techniques: a review	IJACSA	2017
Zhang et al. (2018b)	Deep learning for sentiment analysis: A survey	WIREs	2018
Prabha and Umarani Srikanth (2019)	Survey of Sentiment Analysis Using Deep Learning Techniques	IEEE	2019
Do et al. (2019a)	Deep learning for aspect-based sentiment analysis: a comparative review	Science Direct	2019
Yadav and Vishwakarma (2020)	Sentiment analysis using deep learning architectures: a review	Springer	2020
Lin et al. (2020)	A Survey of Sentiment Analysis Based on Deep Learning	WASET	2020
Dang et al. (2020)	Sentiment analysis based on deep learning: A comparative study	MDPI	2020
Habimana et al. (2020)	Sentiment analysis using deep learning approaches: an overview	Springer	2020
Liu et al. (2020)	Aspect-Based Sentiment Analysis: A Survey of Deep Learning Methods	IEEE	2020
Minaee et al. (2020)	Deep learning-based text classification: A comprehensive review	Arxiv.org	2020
Joseph et al. (2022)	A survey on deep learning-based sentiment analysis `	Science Direct	2022
Mercha et al. (2023)	Machine learning and deep learning for sentiment analysis across languages: A survey	Science Direct	2023

2.2 Multimodal sentiment analysis (MSA)

MSA is a new and rapidly expanding field of research that emphasizes creating a new level in addition to standard sentiment analysis based on text. Because other types of data like pictures, audio, and video have a lot of valuable words, the combinations of these multimodal inputs can be created (MSA) systems with bimodal or trimodal (Soleymani et al. 2017; Stappen et al. 2020) which uses various source of information containing text, audio, acoustic, and visual models. In recent years using multimodal representation for sentiment analysis problem have become a challenge (Baltrusaitis et al. 2019). Xu et al. (2019b) proposed aspect-based sentiment analysis ABSA by multimodal techniques that utilize a combination of aspect of text and image and their interactions. Cai and Xia (2016) proposed a CNN graphic fusion for sentiment analysis. Lai and Yan (2022) proposed a Multimodal sentiment Analysis model with asymmetric window multi-attention that displays the weights of contexts at a specific timestamp. Huang et al. (2019a) proposed deep multimodal attention fusion of text and image for sentiment analysis by Getty image and Twitter and Flicker datasets which achieved accuracy equal to 0.869, 0.763, and 0.859, respectively. Luo et al. (2022) summarized multimodal sentiment analysis approaches based on GNN and RNN models. Also, Gandhi et al. (2022) systematically reviewed newly released multimodal sentiment analysis and their challenges and provided solutions for future directions. Several other attempts have been made on text, audio, and video for sentiment analysis problems by various models (Kim and Lee 2020; Harish et al. 2020; Chen et al. 2020c).

2.3 Sentiment analysis levels

Sentiment analysis has been considered in multi-levels; Document-Level, Sentence-Level, Phrase-Level, and Aspect-Level.

Document-level applies to the whole document and is assigned single polarity to it (Saunders 2021), sentence-level is considered independent polarity for each sentence (Ferrari and Esuli 2019), and phrase-level focuses on sentiment analysis by mining at phrase because of existing single or multiple aspects in each phrase and is one of the interesting topics for researchers (Flek 2020) and aspect-level that assigns a point to all detected aspects in sentences for final aggregating sentiment by these scores (Schouten and Frasincar 2015; Vanaja et al. 2018) is known as features-level or entity-level.

Aspect-based sentiment analysis (ABSA)

In the last decade, aspect-based sentiment analysis (ABSA) as a fine-grained model has attracted significant interest in various literature while coarse-grained sentiment analysis models only focus on the polarity of emotions. Generally, the principal task in ABSA is identifying four sentiment elements containing aspect terms extraction (ATE), aspect categories detection (ACD), opinion term extraction (OTE) as sentiment polarities or Aspect Sentiment Classification (ASC) (Zhang et al. 2021). For instance, in this sentence “The food is delicious.”, “pizza” is the aspect term, “food” is the aspect category, “delicious” is the opinion term, and sentiment polarity is positive. Recently, various efforts have been performed to apply deep learning to ABSA tasks (Wang et al. 2021). Poria et al. (2016) proposed an improved model of the LDA algorithm for clustering by semantic aspect-base sentiment analysis instead of syntactical analysis. Zhang et al. (2022) provided a comprehensive survey on pre-trained language models for ABSA and sorted current efforts on cross-domain/lingual transfer to achieve better performance in this field. Also, they refer to GNN-based methods as novel techniques in sentiment analysis tasks which explicitly leverage the syntactic information. Generally, identifying aspects is a complex task; hence ABSA has enormous challenges. Therefore, multiple complex models (LSTM, Bi-LSTM, GRU, CNN, etc.) and attention-based approaches (BERT, GPT2, GPT3) have been used (Devlin et al. 2019; Brown et al. 2020 ).In recent years attention-based models have been performed in ABSA (Xu et al. 2020; Liu et al. 2019). In recent years attention-based models have been performed in ABSA (Xu et al. 2020; Liu et al. 2019 ). In recent years, many attention-based models have been performed in ABSA with high performance (Xu et al. 2020; Liu et al. 2019a; Yadav et al. 2021).

In this section, we investigate the recent state-of-the-art sentiment analysis based on GNN techniques and have covered most of the impressive papers in this area. Also, we explain their strengths and weaknesses separately and show the results of each on different databases. As mentioned above GNNs have become widely graph analysis methods for NLP areas such as speech recognition (Alsobhani et al. 2021), text classification (Hu et al. 2019; Xie et al. 2021; Ma et al. 2021), and sentiment analysis (Niu et al. 2021; ). GNNs extract spatial features and unlike CNN can perform on non-Euclidean data in multi scales that these abilities lead to make expressive representation. Unlike the label propagation algorithm (LPA), which is a traditional graph-based algorithm (Goldberg and Zhu 2006), GNNs operate by several neural layers and transform, propagate, and aggregate attributes of nodes or edges; due to these features, they can be expressed best representation by a graph structure. This new area, known as geometric deep learning, has received enormous efforts based on it. Sentiment analysis is a classification problem that identifies and categorizes sentimental tone (positive, negative, or neutral) in a user's opinion in part of the text or based on specific aspects and in multimodal approaches can be expanded by visual type of data such as image and video (Yu et al. 2016; Liu et al. 2019b).

Sentiment analysis methods based on GNN usually operate by making a graph representation (e.g., Adjacency matrix or dependency tree) from the text that these representation techniques are different between various approaches; then GNN models apply for learning word embedding to predict sentiment polarity. For graph construction, most kinds of literature on sentiment analysis at the sentence level use a dependency tree for converting text to graph (Zhang et al. 2019a; Wang et al. 2020a). Furthermore, some approaches use syntactic and global lexical graphs (Zhang and Qian 2020; Liao et al. 2021)

For graph representation learning, we have a wide variety of models that are almost referred to as graph embedding, which converts graphs (node, edge, and their features) into vector space with low dimension by considering both graph structure and information (Tang et al. 2020; Chen et al. 2020a; Pouran Ben Veyseh et al. 2020). In the sentiment analysis model based on GNN, the design of GNN and initial graph embedding are critical factors in achieving high performance. In recent years, the most popular GNN models such as GCN (Zhang and Qian 2020; Xu et al. 2023), graph Transformer (Tang et al. 2020), GAT (Chen et al. 2020a), multi-relational graphs models like R-GAT (Busbridge et al. 2019; Wang et al. 2020a), R-GCN models (Ghosal et al. 2020) and other various hybrid models based on GNN have been used in sentiment analysis problems and achieved decent results compared to previous deep learning models and traditional methods. In Table 3, we summarized different sentiment analysis approaches from three perspectives including traditional machine learning, deep learning-based, and GNN-based techniques, and their advantages and disadvantages.

Generally, there are various challenges of GNNs for NLP tasks. One of the critical factors that

affect overall efficiency is graph construction which often refers to art rather than science. Efforts in terms of graph construction for heterogeneous graphs are in priority because they can carry more valuable information. In addition, by observing the results of the research, it can be seen that the construction of dynamic graphs in some natural language processing problems is not enough and it is better to use a combination of static and dynamic graphs to achieve higher performance. As mentioned in previous sections transformers models are used in sentiment analysis that are special types of GNNs that operate on dynamic graphs considering attention mechanism. GNNs unlike transformers operate on both graph structures like knowledge graphs and non-structured ones like text, images etc. Hence, the combination of GNNs with transformers and expanding pre-training GNN models on a large scale are the most attractive trends in sentiment analysis problems that can be able to explore another aspect of original input information on a graph except the attention mechanism. Developing an efficient design for multi-relational GNN which have been achieved impressive progress in NLP tasks, especially sentiment analysis is another challenge in this area; due to various relations in the graph there is an over-parameterization problem that affects the expression of a model which transformers help to exploit multi-relational graphs but how to handle this exchange is still a challenge. In Table 4, we summarize recent state-of-the-art sentiment analyses that have used GNN-based approaches based on different datasets at various levels. Also, we discussed the performance, challenges, results, and future directions from each of the approaches in detail.

Table 3

Summary of different sentiment analysis approaches and their advantages and disadvantages
Approach	Advantage	Disadvantage
Machine learning-based	Supervised learning: Easily identifies trends and patterns, simply implement in most algorithms, less training time for low data Unsupervised learning: No human intervention and labeled data are needed Semi-supervised learning: it can leverage the benefits of both supervised and unsupervised learning, Overcoming problems such as data scarcity and quality	Supervised learning: Labeled data is required human work and linguistic knowledge are required, and poor interpretation of the results accurately Unsupervised learning: Not having enough capacity to do this semi-supervised: design and implementing a sufficient model might be challenging if unlabeled data are noisy
Deep learning-based	RNN: The ability to model the sequential information of sentences, gain long-term affiliations, Ability to display sentences, and Improve prediction accuracy with sequence learning LSTM and Bi-LSTM: More efficient than RNN, overcoming long-term dependency problems in RNN models, Suitable for large datasets, Finding bidirectional dependencies in GRU: Less complex than LSTM, Suitable for small datasets. CNN: More efficient than RNN, Fast training, Ability to capture local features and aspect information, High accuracy Transformer: Efficient Parallel Processing than traditional sequential models, Effective Handling of Long-Term Dependencies by attention mechanism, Increased Model Capacity, Flexibility with variable-length Sequences	RNN: High training time and computational cost, Complex structure, and gradient disappearance problem LSTM: High training time and complex model, overfitting problem, Lack sensitivity for some words, and no outstanding efficiency in SA Bi-LSTM: High computational cost and slow training time GRU: Complexity, high training time, overfitting problem due to more parameters and layers, slow convergence, and low learning efficiency CNN: unable to establish a semantic relationship between aspect and context, time-consuming implementation Transformer: High Computational Cost, A significant amount of data is required to train effectively, Overfitting Vulnerability especially in small datasets.
GNN-based	General benefits of GNN: High performance and accuracy in intricate relationships, ability to handle complex graph-structured data, capture non-linear relationships between nodes, High transparency and interpretability, scalability and adaptability GCN: Transudative model, suited to model syntactic dependency graphs, learning to represent nodes, Getting the local position and latent feature of the nodes, Identifying syntactic relationships by useful information GAT: Inductive model with shared edge-wise mechanism, different levels have different attention weights, better choice in handling unseen nodes than GCN, does not depend on the global graph structure	General limitations of GNN: are not robust to noise, less handling of edges of graphs, Leaning and updating of hidden states of edges is a big problem, Limited to a fixed number of points, using the same parameters in the iteration unlike deep learning models which used different parameters in different layers as a hierarchical feature extraction GCN: Incorrectly correlate aspects with irrelevant words by iterating on graph convolution propagation GAT: Ignoring some important word by assigning lower attention weight, Noise problem

Zhang et al. (2019a) proposed the first Aspect-specific Graph Convolutional Networks (ASGCN) model for aspect-based sentiment classification for five datasets containing Twitter Dong et al. (2014), LAP14, REST14, REST15, and REST16 respectively (Pontiki et al. 2014; Pontiki et al. 2015; Pontiki et al. 2016). They aimed to tackle the limitation of CNN with an attention mechanism that lacks a mechanism to record syntactic information and long-range dependencies. In their study first, after the word embedding step for finding word orders, the Bi-LSTM model is constructed to produce hidden state vectors. Then for obtaining Aspect-oriented features, they applied multi-layer GCN over dependency trees of sentences with an aspect-specific masking layer on its top for filtering non-aspect words and keeping high-level aspect-specific features. Since dependency trees are directed graphs, while GCN networks do not consider direction therefore they proposed two kinds of ASGCNs named un-directional ASGCN-DG models on dependency and directional ASGCN-DT on dependency trees. After Aspect-specific Masking they used the Aspect-aware Attention mechanism. The idea behind this mechanism is to retrieve important features that are semantically related to aspect words from hidden state vectors. In this way, they determined a retrieval-based attention weight for each text word. They compare the proposed model with the baseline models such as SVM, LSTM, MN, AOA, identity-aware network (IAN), TNet-LF, and ASCNN. The result showed ASGCN-DG outperformed ASGCN-DT and all baseline models on LAP14 and REST15 and comparable results on TNet-LF and REST14.

Sun et al. (2019) proposed CDT (convolution over a dependency tree) with a Bi-LSTM model for aspect-Level Sentiment Analysis to refine Bi-LSTM embeddings and extract embedding with considering both textual and dependency information owing to the Bi-LSTM and the GCN respectively. They also presented two restricted versions denoted as ASP Bi-LSTM and ASPGCN on Rest14, Laptop, Twitter, and Rest16 datasets. They compare the proposed model with the baseline models CNN + Position, LSTM + Position, CNN + ATT, and some other baseline models. They found that CDT outperforms all models for the different datasets. Also, they concluded the performance of the proposed models, depends on the number of layers of the GCN. In the proposed model ASPGCN, ASP-Bi-LSTM after the 6-th layer converged and overfitting occurred.

Huang et al. (2019b) propose a novel GNN-based model that makes text-level graphs by global parameter sharing. Although GNN-based models achieved good results in performing complex structures and keeping global information, they don’t support online testing and high memory consumption. To tackle these limitations their model builds graphs for each input text depending on its context instead of a single graph for the whole corpus which extracts more local features and reduces memory consumption. Hence, the scale of nodes and edges has been greatly reduced. According to their results on R8, R52, and Ohsumed datasets, they found their GNN-based model has better performance than traditional models like CNN, LSTM, and fastTest. Also, they concluded proposed model outperforms Graph-CNN models. Graph- CNN connects word nodes using by BOW model and isn’t able to distinguish the importance between different words while the proposed model uses a global trainable edge. Furthermore, they found the proposed model achieved better results than Text-GCN models, due to expressive edges and differences in representation learning models. Text-GCN models use corpus-level cooccurrence while the proposed model used contextual window.

Huang and Carley (2019) proposed the first aspect-based sentiment classification using a target-dependency graph attention network (TD-GAT) model without converting its structure on two widely used datasets from SemEval 2014 Task 4 containing Laptop and Restaurant. They used a dependency graph instead of a word sequence by Stanford neural parser (Chen and Manning, 2014) by two embedding methods including GloVe and BERT representation. They compared the proposed model to the following baseline methods: Feature-based SVM, TD-LSTM, AT-LSTM, MN, IAN, PG-CNN, AOA-LSTM, BERT-AVG, BERT-CLS, and results showed TD-GAT-GloVe outperforms all baseline methods. Also, they found using BERT got excellent performance but observed such fine-tuning in some trials unable to converge. They believe in a lot of potential progress could be made in this area. For future directions, they suggested using an attention mechanism that can focus on important words in the aspect. They stated since our proposed model ignores kinds of relations in the graph, we will consider dependency relation types. Also, they plan to combine their model with a sequence-based model for avoiding possible noise.

Chen et al. (2020a) proposed a Cooperative Graph Attention Networks (Co-GAN) model for Aspect Sentiment Classification (ASC). Since most literature largely ignores the document-level sentiment preference information, they explored two kinds of sentiment preference information such as intra-aspect sentiment consistency and inter-aspect sentiment tendency for cooperatively learning the aspect-related sentence representation. They conducted their experiments on four datasets restaurant15 and laptop15 (from SemEval-2015 Task 12) and restaurant16 and laptop16 (from SemEval-2016 Task 50. They designed the proposed model in five following blocks: 1) Encoding Block by BERT-based model to encode aspect and sentence; 2) Intra-Aspect Consistency Modeling Block by a consistency-aware GAN; 3) Inter-Aspect Tendency Modeling Block by leveraging a tendency-aware GAN; 4) Interaction Block by two strategies to learn the sentence representation as Pyramid Layers and Adaptive Layer-Fusion respectively and 5) SoftMax Decoding Block. They compared their proposed model with baseline models such as TC-LSTM, ATAE-LSTM, RAM, IAN, Clause-Level ATT, LSTM + synATT + TarRep, BERT, CADMN, IMN, and BERT-QA and concluded CoGAN outperforms all the baseline approaches and improved 11.6% (Accuracy), 14.3% (Macro- F1) on both RES15,16 and 9.1% (Accuracy), 12.6%(Macro-F1) on Lap15,16.

Wang et al. (2020a) proposed a novel GAT model as a relational graph attention network (R-GAT) to encode comprehensive syntax information for aspect-based sentiment analysis. The aim of this novel approach helping to attention-based model for establishing implicitly the connections between aspects and opinion words. Because of the complexity of language and the existence of multiple aspects, attention-based models confused the relations. The results of their research on the Laptop and Restaurant database (from SemEval 2014) and Twitter datasets confirmed R-GAT model performance. They constructed an aspect-oriented dependency tree from an ordinary dependency tree. This aspect-oriented structure had two strengths. First, each aspect had its dependency tree therefore it could be less influenced by unrelated relations second advantage refers to the aggregation of dependency relationships in that aspect when an aspect includes more than one word. They used a few baseline models in three groups including Syntax aware models, Attention-based models, and Other recent methods for comparison. Experimental results showed that the R-GAT model outperforms most of the baseline models and the performance of the GAT model improved when incorporated with relational heads in their aspect-oriented dependency tree structure. Furthermore, they concluded basic BERT is better than all the existing ABSA approaches that indicating the potency of this pre-trained model. After the combination of R-GAT + BERT, they observed a strong and more effective model rather than the proposed RGAT which according to the results, accuracy and Macro-F1 improved for all three data groups, i.e., Restaurant Laptop Twitter.

Tang et al. (2020) proposed a novel model based on GCN and Transformer named (DGEDT) which is a dependency graph-enhanced dual transformer network by considering the connections in the dependency tree as a supplementary GCN module for dual-transformer structure. Although attention-based methods and GCN are used in ABSA to express the relationship between aspects and related emotional words, their progress is limited due to the noise and instability of dependency trees. For overcoming these problems, they proposed a reinforced dependency graph. First, they utilized Bi-LSTM from BERT as an aspect-based encoder. Then, after obtaining the contextual hidden representations from the encoder, they developed a dual-transformer structure including a multi-layer Transformer and a multi-layer BiGCN, then an attention mechanism to identify relevant words and subsequently a masking mechanism to avoid assigning too high a weight to aspects. Experimental results were conducted based on their proposed DGEDT with Bi-LSTM and DGEDT + BERT with BERT on five datasets, containing Twitter, Lap14, Rest 14, Rest 15, and Rest 16 that demonstrated DGEDT using by transformer obtains better performance than DGEDT(BiGCN) compared to baseline methods. Also, the DGEDT-BERT model outperformed all five datasets.

Pouran Ben Veyseh et al. (2020) proposed a novel graph-based deep learning model based on Gated Graph Convolutional Networks and Syntax-based Regulation. According to the literature deep learning models are widely used in ABSA and in recent years, the syntactic dependency trees have been integrated into them as graph-based deep learning models, but these models have two major problems that should be addressed to improve performance. The first problem is the representation vectors in hidden layers of current graph-based models are not customized for ABSA. This problem might lead to suboptimal representation vectors, while in the ideal state in ABSA, representation vectors mainly involve the most important and related information. The second problem is current graph-based models focus only on syntactically neighboring words and lose the use of overall word contextual scores from the dependency tree for ABSA. For overcoming these limitations, they proposed a gate vector for each layer of the graph-based model, then this layer was applied over the hidden vectors to make customized hidden vectors for ABSA. They compared their proposed model with three groups of baseline models in the following order: the feature-based model and SVM the deep learning models, and the graph-based models on three benchmark datasets such as Restaurant and Laptop from the SemEval 2014 Task 4 and MAMS (introduced by Jiang et al. 2019). They demonstrated the effectiveness of the proposed by achieving 87.2%, 82.8%, and 88.2% accuracy for Rest. Laptop MAMS respectively.

Zhang and Qian. (2020) proposed a novel architecture that convoluted over hierarchical syntactic and lexical graphs (named BiGCN model) for ABSA to overcome two limitations in graph-based models; ignoring the corpus-level word co-occurrence information and not recognizing different types of syntactic dependency. They utilized a global lexical graph to capture the global word co-occurrence information in the training corpus. In other words, BiGCN takes this graph as the input to get the initial sentence representation. Finally, they designed a HierAgg module to refine the sentence representation and let the lexical and syntactic graphs work together. They used five benchmark datasets including Twitter, Lap14, Rest14, Rest15, and Rest16, and compared the proposed model with the eight baselines containing typical neural structures like attention, LSTM, CNN, memory, and RNN; AF-LSTM model and finally graph-based and syntax integrated models. Experimental results demonstrated that their proposed BiGCN achieved the best macro-F1 on all datasets (an improvement of 3.12, 2.77, and 1.36 F1-score for Rest16, Rest15, and Twitter respectively). Furthermore, the graph-based and syntax-integrated models got good results that can be concluded that dependency relationships are effective in identifying polarity. Also, they found that the AF-LSTM method has not shown any progress compared to classical methods which can be inferred that explicit integration of word association by attention mechanism is not enough.

Ghosal et al. (2020) introduced a new framework as Knowledge-Guided Domain Adaptation (KinGDOM) for Sentiment Analysis that demonstrates a novel view of external commonsense knowledge (KB). Augmenting neural models with KB has multiple advantages in the range of NLP applications and are popular method but most of these efforts such as domain-dependent word embeddings (K Sarma et al. 2019) and co-occurrences of domain-specific with domain-independent terms (Sharma et al. 2018) have been sporadic. To this end, they proposed a domain-adversarial framework for unsupervised domain adaptation by external KB (Concept-Net) to combat the domain gap in sentiment analysis applications. KinGDOM aims to improve the DANN model by ConceptNet unlike semantic knowledge graphs (e.g., WordNet) and traditional word embeddings. Because this knowledge base includes both domain-specific and domain-general knowledge. They conducted their experiments on Amazon-reviews benchmark datasets consisting of Books, DVDs, electronics, and kitchens for domain adaptation in sentiment analysis. They compared their proposed model with several baseline and state-of-the-art models that demonstrated the effectiveness of the proposed model for the task of cross-domain sentiment analysis.

Huang et al. (2020) proposed a novel model named Syntax-Aware Graph Attention Network (SAGAT) which uses syntactic awareness by graph attention network on the dependency tree and BERT which can obtain more accurate representations of words by graph attention to overcome long-distance dependence problems for ABSA. They evaluated their proposed model on Twitter and Restaurant, Laptop, datasets (from SemEval 2014) and compared them with the various baseline models. The proposed model achieved high performance in Restaurant and Twitter datasets and slightly worse than SDGCN-BERT on the Laptop dataset. According to their research, they concluded the following in general: Graph-based models would show additional information rather than traditional approaches. SDGCN model which is a GNN-based model (Zhao et al. 2019) considers the sentiment dependencies between aspects to build two kinds of graph. CDT (Sun et al. 2019) builds graphs by dependency parsing. Also, it can shorten the distance from keywords to aspect words but uses GCN to propagate graphs instead of an attention mechanism.

Meng et al. (2020) proposed a novel architecture GCN named weighted graph convolutional network (WGCN) over dependency tree to tackle the problems of GCN by taking advantage of all the syntactic information obtained from the dependency parsing. The proposed model can extract rich syntactic information based on the feature combination and uses pre-trained language models (BERT) instead of Bi-LSTM to make an alignment method to retain word-level dependencies. According to their experiment on five aspect-based sentiment analysis datasets and three sentiment analysis datasets including SEM14 (LAP), SEM14 (Rest), Rest15, Rest16, Twitter for ABSA tasks, and SST2, SST5, SE13 for sentiment analysis indicated BERT-WGCN outperforms most of the compared baseline models TWITTER, REST15, and REST16, and achieved competitive results on SEM14 (LAP) and SEM14 (REST).

Hou et al. (2021) proposed an effective graph ensemble technique named GraphMerge to improve the performance of previous graph-based methods. Although graph-based methods have led to improved performance in the field of sentiment analysis in recent years, these methods are vulnerable to parsing errors. To this end, they proposed a simple yet effective GraphMerge model to combine the different dependency trees before applying representation learners such as GNNs to construct an ensemble graph. This model has several advantages. First, the GNN models can be exposed to multiple parsing hypotheses and allow the model to learn to use more efficient edges. Second, because applying GNNs to a single graph with the same number of nodes, the proposed model doesn’t require extra computational cost, and finally GraphMerg model avoids overfitting due to limiting over-parameterization and subsequently reduces the diameter of graphs. According to their experimental result on three datasets: REST14, Laptop14 (from SemEval 2014 Task 4), and ACL 14 (Twitter), the proposed model outperformed all baselines (BERT-baseline; GAT; RGAT; Label and feature ensemble by at least 1.42 accuracy and 2.34 Macro-F1 respectively.

Liao et al. (2021) proposed a novel multi-level GNN (MLGNN) by scaled dot-product attention mechanism as a message-passing mechanism that will be able to focus on local and global features by connecting windows with different levels and sizes. Experimental results on different datasets such as SST-binary (Socher et al. 2013); Sentube-A and Sentube-T (Uryupina et al. 2014) confirmed the efficiency of the proposed model among baseline models (BOW, AVE, LSTM, Bi-LSTM, CNN, Huang et al. 2019b).

AlBadani et al. (2022) proposed a novel Sentiment Transformer Graph Convolutional Network (ST-GCN) which was the first study to model sentiment corpus as a heterogeneous graph. This model was able to word embeddings and identify sufficient connections between nodes that are not directly connected. Also, they utilized Laplacian eigenvectors to fuse node positional information for graph datasets inspired by the transformer models as positional encoding in NLP tasks. They used the BERT model as document node embeddings for overcoming vector representation limitations and the multi-head attention allows a simple interpretation of the model. According to their experimental results on SemEval SST-B IMDB Yelp 2014 datasets, ST-GCN achieved high performance among all baseline models and some interesting future directions are applying ST-GCN for link predictions and graph classification because they utilize dynamic neighborhood aggregation operators for improving classification tasks.

Li and Li. (2022b) proposed a GNN-based model for sentiment analysis of comments on the Weibo platform which extracts semantic and structural features. They stated that traditional models due to relying on text sequence representation and ignoring syntactic components and poor interpretability of feature space are not suitable for unspecified sentence analysis. They utilized the Language Technology Platform (LTP) for semantic graph construction. Then, they used a spatial graph filter for heterogeneous semantic graphs based on the MPNN framework and LSTM as a state updater to filter node noise, simultaneously. For a better extracting feature, 14 dependency encodings are used as edge feature weights. Due to the complex graph structure, the syntactic tree had noise information that to address this problem, the filter should preserve serialized feature analysis. GNN-LSTM structure consists of three convolution layers with LN and RELU and a global pool function for global features. The experimental results on Weibo_senti_100k, online_shopping_10_cats, and book review dataset demonstrated superior performance by achieving 95.25% accuracy and 95.22% F1 score.

Li et al. (2022a) proposed a hierarchical multi-head attention mechanism and a graph convolutional network (MHAGCN) to avoid the loss of important information. Due to the limitation of the attention mechanism and GCN models in ignoring the syntactic relationships between aspects and the corresponding contextual words, they proposed a model that captures effectively and makes entire use of syntactic information well and ignores contextual words that are not related to aspects words. hierarchical multi-head attention helps the model to focus on the interaction. Also, they used two pre-trained models for embedding, GloVe and BERT to obtain a fixed word embedding for each word. The experimental results were conducted on Sem-Eval 2014 Task4 (restaurant and Laptop) and the ACL2014 Twitter dataset showed MHAGCN(BERT) model achieved an accuracy of 79.06, 82.57, 74.53%, and Macro-F175.70, 75.83, and 73.75% for Laptop Restaurant Twitter respectively among baseline models.

Bie et al. (2023) proposed a novel end-to-end ABSA model, namely, SSi-LSi with stronger interpretability to overcome the limitations of pipeline and end-to-end methods that ignore the specific kinds of dependency relationships, and this issue lead to insufficient sentiment analysis. Their proposed model fuses the syntactic and semantic information by two network branches, then combines this information by attention mechanism to get higher quality results. They compare their proposed model with baseline and pipeline approaches on three benchmark datasets and observed SSi-LSi model outperforms the MNN model by 2.34%, 3.57%, and 2.26% improvement, and the best INABSA model by 0.39%, 0.68%, and 0.41% in three datasets.

Cui et al. (2023b) proposed a hybrid model based on affective-knowledge-enhanced GCN and multi-head attention mechanism (MHA) named MHAKE-GCN which external sentiment has incorporated in GCN and semantic interaction has performed by MHA. They utilized Bi-LSTM for encoding and constructed a GCN for each input sentence over the dependency tree. Then, they used an affective knowledge-enhanced GCN model by enhancing the representation of the adjacency matrix for more efficiency and to enhance the dependency relationship between context and aspect. They conducted experiments on four public datasets including Restaurant14, Laptop14, Restaurant15, and Restaurant16, and compare the proposed model with twelve baseline models than the MHAKE-GCN that was constructed by the traditional dependency achieved 2.4% and 1.57% F1-Score behind R-GAT model and their BERT-based model achieved accuracies 1.05%, 1.37%, 0.78%, and 2.19% higher than the best BERT-based results for all databases.

Table 4

Summaries of recent sentiment analysis or classification based on GNN-based approaches
Reference	Title	Approach	Dataset
Zhang et al. (2019a)	Aspect-based Sentiment Classification with Aspect-specific Graph Convolutional Networks	ASGCN-DT ASGCN-DG	Twitter LAP14 REST14 REST15 REST16
Sun et al. (2019)	Aspect-Level Sentiment Analysis Via Convolution over Dependency Tree	ASP-Bi-LSTM ASP-GCN CDT	Rest14 Laptop14 REST16 Twitter
Huang et al. (2019b)	Text Level Graph Neural Network for Text Classification	GNN	R8 R52 Ohsumed
Huang and Carley (2019)	Syntax-Aware Aspect Level Sentiment Classification with Graph Attention Networks	TD-GAT-GloVe TD-GAT-BERT	REST14 Laptop14
Chen et al. (2020a)	Aspect sentiment classification with document-level sentiment preference modeling	(Cooperative Graph Attention Networks (CoGAN)	REST15 laptop15 REST16 laptop16
Wang et al. (2020a)	Relational Graph Attention Network for Aspect-based Sentiment Analysis	R-GAT R-GAT-BERT	Rest14 Laptop14 Twitter
Tang et al. (2020)	Dependency graph enhanced dual-transformer structure for aspect-based sentiment classification	dependency graph enhanced dual-transformer network (DGEDT-BERT)	Twitter Lap14 REST 14 REST 15 Rest16
Pouran Ben Veyseh et al. (2020)	Improving Aspect-based Sentiment Analysis with Gated Graph Convolutional Networks and Syntax based Regulation	Gated graph convolutional network (GGCN)	Rest14 Laptop14 MAMS
Zhang and Qian (2020)	Convolution over hierarchical syntactic and lexical graphs for aspect-level sentiment analysis	BiGCN	Twitter SemEval
Ghosal et al. (2020)	Knowledge Guided Domain Adaptation for Sentiment Analysis	Knowledge Guided Domain adaptation (KinGDOM) with R-GCN as graph encoder network	Amazon-reviews benchmark datasets include Books, DVDs, Electronics, and Kitchen appliances.
Huang et al. (2020)	Syntax-Aware Graph Attention Network for Aspect-Level Sentiment Classification	SAGAT-BERT	RES14 Laptop14 ACL14 (Twitter)
Meng et al. (2020)	Sentiment Analysis with Weighted Graph Convolutional Networks	BERT-WGCN	LAP14 REST14 SEM14(AVG) REST15 REST16 Twitter
Hou et al. (2021)	Graph Ensemble Learning over Multiple Dependency Trees for Aspect-level Sentiment Classification	Graph Merge over multiple dependency trees	REST14 Laptop14 ACL 14 (Twitter)
Liao et al. (2021)	Multi-level graph neural network for text sentiment analysis	multi-level graph neural network (MLGNN)	SST-binary Sentube-A Sentube-T
AlBadani et al. (2022)	Transformer-Based Graph Convolutional Network for Sentiment Analysis	Sentiment Transformer Graph Convolutional Network (ST-GCN)	SemEval SST2 IMDB Yelp 2014 (Restaurant)
Li and Li (2022b)	Sentiment Analysis of Weibo Comments Based on Graph Neural Network	GNN-LSTM	Weibo_senti_100k online_shopping_10_cats & 20000 book review
Li et al. (2022a)	Graph convolutional networks with hierarchical multi-head attention for aspect-level sentiment classification	multi-head attention mechanism and a graph convolutional network (MHAGCN)	Sem-Eval-2014 Task4 ACL2014 (Twitter)
Bie et al. (2023)	Fusing Syntactic Structure Information and Lexical Semantic Information for End-to-End Aspect-Based Sentiment Analysis	SSi-LSi SSi-LSi BERT	D_L D_R D_T
Cui et al. (2023b)	Affective-Knowledge-Enhanced Graph Convolutional Networks for Aspect-Based Sentiment Analysis with Multi-Head Attention	MHAKE-GCN (GCN + MHA)	REST14 Laptop14 REST15 REST16

Data collection is one of the critical steps in sentiment analysis. It is an essential task for data collection to find the relevant datasets instead of finding a large volume that can cover a wide use case. We will deal with textual, audio, image, and video data depending on the problem. Some common data sources are social media posts, e-commerce websites (Jain et al. 2019), blogs (Annett and Kondrak 2008), and forums (Korkontzelos et al. 2016). Table 5 shows some of the common and popular resources of datasets in this area.

Table 5

Some of the common datasets in sentiment analysis tasks
Dataset	Description
SemEval-datasets	SemEval is a set of international NLP workshops to create high-quality annotated datasets. Annually, the workshops presented shared tasks and are compared by the computational semantic analysis system.
Amazon product data (Haque et al. 2018)	The data set includes 320 million customer reviews from 1996 to July 2018. This metadata contains reviews about brands, prices, and other useful information about products that are labeled based on their positive, negative, and neutral emotional tone
Twitter US Airline Sentiment (Rane and Kumar 2018)	Consist of nearly 15000 tweets in terms of six various commercial airlines which categorized into positive, neutral, and negative classes
OpinRank Review Dataset for hotels and cars	A complete review includes hotel and automotive industries with 2,59,000 hotel reviews (10 cities with 80–700 hotels) and 42,230 car reviews (between 2007 and 2009.)
Sentiment Polarity Lexicons For 81 Languages	The dataset contains sentiment lexicons for 81 languages that are built by graph propagation based on English lexicons and knowledge graphs.
Cornell Movie Review Dataset:	Includes around 2,000 positive and negatively tagged reviews and more than 10,000 positive and negative tagged sentence texts
First GOP Debate Twitter Sentiment	Includes 14,000 labeled tweets that are positive, neutral, and negative in the first GOP debate in 2016
Yelp Dataset (AlBadani et al. 2022)	Contains 5.2 million Yelp reviews for the academic challenge in North America that includes star ratings, reviews, businesses, and user data.
Multi-Domain Sentiment Analysis Dataset	It Contains Amazon product positive and negative reviews such as books and DVDs and musical instruments
Paper Reviews	The dataset includes English and Spanish language reviews about computing and informatics conferences
Opinion Lexicon	This dataset contains nearly 7000 positive and negative English sentiment or opinion words
Lexicoder Sentiment Dictionary	This dataset uses Lexicoder for sentiment analysis, this dictionary contains more than 2800 negative and 1709 positive sentiment word
IMDB Reviews Dataset (Amulya et al. 2022)	Large Movie Review Dataset (IMDB) contains 50K movie reviews (25000 training set and 25000 testing set) for binary sentiment classification based on deep learning approaches.
Financial Phrasebank	This dataset contains almost 5000 English sentences from financial news in three classes positive, negative, or neutral
Stanford Sentiment Treebank (Wang et al. 2019a)	Datasets include more than 10000 pieces of Stanford data which are rated (between 1–25) and are collected by HTML files of Rotten Tomatoes based on the sentence structure.
SPOT (Sentiment Polarity Annotations Dataset) (Angelidis and Lapata. 2018)	Consist of 197 reviews from Yelp’13 and IMDB which annotates by polarity labels (positive, neutral, and negative) for sentiment analysis problems based on fine-grain and segment-level

In the past few years, the explosion of data information in social networks has created the need to improve sentiment analysis methods as one of the natural language processing tasks. There are various traditional machine learning and deep learning approaches with their strengths and weaknesses in this field. In recent years, deep learning methods and their hybrid models have been able to achieve acceptable results, however, they still face problems such as weak interpretability of feature space, ignoring syntactic structure, and relying too much on text sequences. GNNs as a powerful and popular method for managing structured graph data are a rapidly evolving field with many exciting developments that have attracted much attention from the AI research community. Their principal ability refers to expressing power, flexibility, complex structure, interpretability, learning algorithms, and analysis power. Unlike many machine learning algorithms that cannot exploit syntactic information, GNNs convince systems to think more logically and naturally. In this review article, we focus on the latest GNN-based sentiment analysis approaches and their strengths and weaknesses. We also discuss and introduce GNN structures from two different perspectives, structured and unstructured. Also, we showed the outstanding challenges of each approach based on different datasets to provide a way to improve future directions and show the significant potential of GNN in this new field. According to this study, the following future directions can be considered: the need for new labeled and unlabeled data for challenging tasks such as multilingual and cross-lingual embedding and emotion analysis by considering language complications and domain adaptation techniques. Dynamic sentiment analysis and heterogeneous graphs should be more considered by interpretable deep learning-based like GNN-based. Development of cross-domain and cross-lingual ABSA, Fuzzy parallel models, reinforcement learning, generative adversarial networks, optimization methods such as ensemble learning, and, memory network models in both standard and multimodal sentiment analysis should be further investigated. In the future, we intend to provide an overview of GNN-based methods and their related architectures in the field of multimodal sentiment analysis.

Funding / Competing interests

No funding was received to assist with the preparation of this manuscript.

The authors have no competing interests to declare that are relevant to the content of this article.

Abid F, Alam M, Yasir M, Li C (2019) Sentiment analysis through recurrent variants latterly on convolutional neural network of Twitter. Futur Gener Comput Syst 95, pp 292–308
Acheampong FA, Nunoo-Mensah H, Chen W (2021) Transformer models for text-based emotion detection: a review of BERT-based approaches. Artif Intell Rev 54, pp 5789–5829
Agarwal, B., Mittal, N. (2016). Semantic Orientation-Based Approach for Sentiment Analysis. In: Prominent Feature Extraction for Sentiment Analysis. Socio-Affective Computing. Springer, Cham. https://doi.org/10.1007/978-3-319-25343-5_6
Ain QT, Ali M, Riaz A, Noureen A, Kamranz M, Hayat B, et al (2017) Sentiment analysis using deep learning techniques: a review. Int J Adv Comput Sci Appl 8(6), pp 424–433. https://doi.org/10.14569/ijacsa.2017.080657
Al Amrani Y, Lazaar M, El Kadiri KE (2018) Random Forest and support vector machine-based hybrid approach to sentiment analysis. Procedia Comput Sci 127, pp 511–520
Alami N, Meknassi M, En-nahnahi N (2019) Enhancing unsupervised neural networks-based text summarization with word embedding and ensemble learning. Expert Systems with Applications 123, pp 195–211
AlBadani B, Shi R, Dong J, Al-Sabri R, Moctard OB (2022) Transformer-Based Graph Convolutional Network for Sentiment Analysis. Appl. Sci 12(3):1316
Alhajj R, Rokne J (2014) Encyclopedia of social network analysis and mining. Springer
Alshari EM, Azman A, Doraisamy S, Mustapha N, Alkeshr M 2018 Effective method for sentiment lexical dictionary enrichment based on Word2Vec for sentiment analysis. In Fourth International Conference on Information Retrieval and Knowledge Management (CAMP), pp 1–5
Alsobhani A, ALabboodi H.M.A, Mahdi H (2021) Speech Recognition using Convolution Deep Neural Networks. Conf. Ser. 1973 012166, 3rd International Scientific Conference of Engineering Sciences and Advances Technologies (IICESAT)
Amulya K, Swathi SB, Kamakshi P Bhavani Y (2022) Sentiment Analysis on IMDB Movie Reviews Using Machine Learning and Deep Learning Algorithms. IEEE 4th International Conference on Smart Systems and Inventive Technology (ICSSIT), pp 814–819
An W, Tian F, Chen P, Zheng Q (2023) "Aspect-Based Sentiment Analysis With Heterogeneous Graph Neural Network," in IEEE Transactions on Computational Social Systems, vol. 10, no. 1, pp. 403–412
Angelidis S, Lapata M (2018) Multiple Instance Learning Networks for Fine-Grained Sentiment Analysis. Transactions of the Association for Computational Linguistics (TACL), Vol 6, pp 17–31
Annett M, Kondrak G (2008) A comparison of sentiment analysis techniques: Polarizing movie blogs. In: Conference of the Canadian Society for Computational Studies of Intelligence. Springer, pp 25–35. https://doi.org/10.1007/978-3-540-68825-93
Aydin C.R, Güngör T, Combination of Recursive and Recurrent Neural Networks for Aspect-Based Sentiment Analysis Using Inter-Aspect Relations (2020) IEEE Access, vol. 8, pp 77820–77832
Baltrusaitis T, Ahuja C, Morency L (2019) Multimodal machine learning: A survey and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 41, no. 2, pp 423–443
Bartusiak R, Augustyniak L, Kajdanowicz T, Kazienko P (2015) Sentiment analysis for Polish using transfer learning approach. In: 2015 second European network intelligence conference. IEEE, pp 53–59
Basiri ME, Nemati S, Abdar M, Cambria E, Acharya UR (2021) ABCDM: an attention-based bidirectional CNN-RNN deep model for sentiment analysis. Futur Gener Comput Syst 115, pp 279–294
Beck D, Haffari G, Cohn T (2018) Graph-to-sequence learning using gated graph neural networks. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics Vol.1, pp 273–283
Bergsma S, McNamee P, Bagdouri M, Fink C, Wilson T (2012) Language identification for creating language-specific Twitter collections. In: Proceedings of the second workshop on language in social media, pp 65–74
Bie Y, Yang Y, Zhang Y (2023) Fusing Syntactic Structure Information and Lexical Semantic Information for End-to-End Aspect-Based Sentiment Analysis in Tsinghua Science and Technology, vol. 28, no. 2, pp 230–243
Bittar A, Velupillai S, Roberts A, Dutta A (2021) Using General-purpose Sentiment Lexicons for Suicide Risk Assessment in Electronic Health Records: Corpus-Based Analysis. JMIR Medical Informatics 9(4)
Brown TB, Mann B, Ryder N, Subbiah M, et al (2020) Language models are few-shot learners. arXiv preprint arXiv:2005.14165
Busbridge D, Sherburn D, Cavallo P, Hammerla NY (2019) Relational graph attention networks. arXiv preprint arXiv:1904.05811
Cai GY, Xia BB (2016) Multimedia sentiment analysis based on convolutional neural network. Journal of Computer Applications, vol. 36, no. 2, pp 428–431
Cao Q, Duan W, Gan Q (2011) Exploring determinants of voting for the “helpfulness’’ of online user reviews: a text mining approach. Decis Support Syst 50(2), pp511–521
Cetoli A, Bragaglia S, O'Harney AD, Sloan M (2017) Graph Convolutional Networks for Named Entity Recognition. arXiv:1709.10053 .16th International Workshop on Treebanks and Linguistic Theories
Chen X, Lu GM, Yan JJ (2020c) Multimodal sentiment analysis based on multi-head attention mechanism. In Proceedings of the 4th International Conference on Machine Learning and Soft Computing, pp 34–39
Chen X, Sun C, Wang J, Li S, Si L, Zhang M, Zhou G (2020a) Aspect sentiment classiﬁcation with document-level sentiment preference modeling. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics pp 3667–3677
Chen Y, Wu L, and Zaki MJ (2020b) Graphﬂow: Exploiting conversation ﬂow with graph neural networks for conversational machine comprehension. In Proceedings of the Twenty-Ninth International Joint Conference on Artiﬁcial Intelligence, IJCAI2020, pp 1230–1236
Cheng Y, Yao L, Xiang G, Zhang G, Tang T, Zhong L (2020) Text sentiment orientation analysis based on multi-channel CNN and bidirectional GRU with attention mechanism. IEEE Access 8:134964–134975
Chen Y, Zhuang T, Guo K (2021) Memory network with hierarchical multi-head attention for aspect-based sentiment analysis. Appl Intell 51, pp 4287–4304. https://doi.org/10.1007/s10489-020-02069-5
Colón-Ruiz C, Segura-Bedmar I, (2020) Comparing deep learning architectures for sentiment analysis on drug reviews, Journal of Biomedical Informatics, Vol110, ISSN 1532-0464
Cui J, Wang Z, Ho SB, Cambria E (2023a) Survey on sentiment analysis: evolution of research methods and topics. Artif Intell Rev 56, pp 8469–8510
Cui X, Tao W, Cui X (2023b) Affective-Knowledge-Enhanced Graph Convolutional Networks for Aspect-Based Sentiment Analysis with Multi-Head Attention. Appl. Sci 13(7):4458
Dang NC, Moreno-Garcia MN, De la Prieta F (2020) Sentiment analysis based on deep learning: a comparative study. Electronics 9(3):483
Derr T, Ma Y, Tang J (2018) Signed graph convolutional networks. Proceedings of ICDM, pp 929–934
Do HH, Prasad PWC, Maag A, Alsadoon A (2019a) Deep learning for aspect-based sentiment analysis: a comparative review. Expert Syst Appl 118, pp 272–299
Do K, Tran T, Venkatesh S (2019b) Graph transformation policy network for chemical reaction prediction. Proceedings of SIGKDD, pp 750–760
Dong L, Wei F, Tan CH, Tang D, Zhou M, Xu K (2014) Adaptive recursive neural network for target-dependent Twitter sentiment classification. In Proceedings of the 52nd annual meeting of the Association for computational linguistics, Vol 2, pp 49–54
Devlin J, Chang MW, Lee K, Toutanova K (2019) Bert: pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv: 18100 4805
Ethayarajh K (2019) How contextual are contextualized word representations? Comparing the geometry of BERT, ELMO, and GPT-2 embedding. arXiv preprint arXiv: 19090 0512
Fellbaum C (2017) WordNet: An electronic lexical resource. The Oxford Handbook of Cognitive Science, pp 301–314
Feng Y, You H, Zhang Z, Ji R, Gao Y (2019) Hypergraph neural networks. In: Proceedings of AAAI, vol. 33, pp 3558–3565
Ferrari A, Esuli A (2019) An NLP approach for cross-domain ambiguity detection in requirements engineering. Autom Softw Eng 26(3), pp 559–598
Flek L (2020) Returning the N to NLP: towards contextually personalized classification models. In: Proceedings of the 58th annual meeting of the Association for computational linguistics, pp 7828–783
Fu X, Zhang J, Meng Z, King I (2020) Met-path aggregated graph neural network for heterogeneous graph embedding. Proceedings of WWW '20 2331–2341
Gandhi A, Adhvaryu K, Poria S, Cambria E, Hussain A (2023) Multimodal sentiment analysis: A systematic review of history, datasets, multimodal fusion methods, applications, challenges, and future directions. Information fusion, Volume 91,2023, pp 424 – 444. https://doi.org/10.1016/j.inffus.2022.09.025
Ghosal D, Hazarika D, Roy A, Majumder N, Mihalcea R, Poria S (2020) KinGDOM: Knowledge Guided Domain Adaptation for Sentiment Analysis. In Proceedings of the58th Annual Meeting of the Association for Computational Linguistics, pp 3198–3210
Gilmer J, Schoenholz SS, Riley PF, Vinyals O, Dahl GE (2017) Neural Message Passing for Quantum Chemistry. International Conference on Machine Learning (ICML). https://doi.org/10.48550/arXiv.1704.01212
Goldberg AB, Zhu X (2006) Seeing stars when there aren’t many stars: Graph-based semi-supervised learning for sentiment categorization. In Proceedings of Text Graphs: The ﬁrst workshop on graph-based methods for natural language processing, pp 45–52
Habimana O, Li Y, Li R, Gu X, Yu G (2020) Sentiment analysis using deep learning approaches: an overview. Sci China Inf Sci 63(1), pp 1–36
Hajek P, Barushka A, Munk M (2020) Fake consumer review detection using deep neural networks integrating word embeddings and emotion mining. Neural Comput Appl 32(23), pp 17259–17274
Hamilton W, Ying Z, Leskovec J (2017) Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems, pp 1024–1034
Han K, Xiao A, Wu E, Guo J, Xu C, Wang Y (2021) Transformer in transformer. arXiv preprint arXiv: 21030 0112
Hao ZF, Huang H, Cai RC, Wen W (2018) Fine-grained opinion analysis based on multi-feature fusion and bidirectional RNN. Computer Engineering, vol. 44, no. 7, pp 199–2049
Haque TU, Saber NN Shah FM (2018) Sentiment analysis on large scale Amazon product reviews. IEEE International Conference on Innovative Research and Development (ICIRD), pp 1–6
Harish A, Sadat F (2020) Trimodal attention module for multimodal sentiment analysis. In Proceedings of the Thirty-Fourth AAAI Conference on Artiﬁcial Intelligence, pp 13803–13804
Hassonah MA, Al-Sayyed R, Rodan A, Ala’M AZ, Aljarah I, Faris H (2020) An efficient hybrid filter and evolutionary wrapper approach for sentiment analysis of various topics on twitter. Knowl-Based Syst 192:105353
Hershcovich D, Donatelli L (2021) It’s the meaning that counts: the state of the art in NLP and semantics. KI-Künstliche Intelligenz pp 1–16
Hettiarachchi H, Adedoyin-Olowe M, Bhogal J, et al (2023) TTL: transformer-based two-phase transfer learning for cross-lingual news event detection. Int. J. Mach. Learn. & Cyber. 14, pp 2739–2760. https://doi.org/10.1007/s13042-023-01795-9
Hou X, Qi P, Wang G, Ying R, Huang J, He X, Zhou B (2021) Graph Ensemble Learning over Multiple Dependency Trees for Aspect-level Sentiment Classification. arXiv:2103.11794
Hu H, Gu J, Zhang Z, Dai J, Wei Y (2018) Relation Networks for Object Detection. IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp 3588–3597
Hu L, Yang T, Shi CH, Ji H, Li X (2019) Heterogeneous Graph Attention Networks for Semi-Supervised Short Text Classification. EMNLP
Huang B, Carley K (2019) Syntax-Aware Aspect Level Sentiment Classiﬁcation with Graph Attention 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), pp 5469–5477
Huang FR, Zhang XM, Zhao ZH, Xu J, Li ZH (2019a) Image-text sentiment analysis via deep multimodal attentive fusion. Knowledge-Based Systems, vol. 167, pp 26–37
Huang L, Ma D, Li S, Zhang X, and Wang H (2019b) Text Level Graph Neural Network for Text Classiﬁcation. arXiv: 1910.02356
Huang L, Sun X, Li S, Zhang L Wang H (2020) Syntax-Aware Graph Attention Network for Aspect-Level Sentiment Classification. COLING
Imamah, Husni, Rachman EM, Suzanti IO, Mufarroha FA (2020) Text Mining and Support Vector Machine for Sentiment Analysis of Tourist Reviews in Bangkalan Regency. J. Phys.: Conf. Ser. 1477 022023
Ioannidis VN, Marques AG, Giannakis G. B (2019) A Recurrent Graph Neural Network for Multi-Relational Data. ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2019, pp 8157-8161
Jain PK, Pamula R, Ansari S, Sharma D, Maddala L (2019) Airline recommendation prediction using customer generated feedback data. In: 2019 4th International Conference on Information Systems and Computer Networks (ISCON). IEEE, pp 376–379. https://doi.org/10.1109/ISCON47742.2019.9036251
Janjua F, Masood A, Abbas H, Rashid I, Khan MMZM (2021) Textual analysis of traitor-based dataset through semi-supervised machine learning. Futur Gener Comput Syst 125, pp 652–660
Jiang Q, Chen L, Xu R, Ao X, Yang M. (2019) A challenge dataset and effective models for aspect-based sentiment analysis. In ACL
Jing N, Wo ZH, Wang H (2021) A hybrid model integrating deep learning with investor sentiment analysis for stock price prediction. Expert Systems with Applications, Vol.178, ISSN 0957-4174
Joseph J, Vineetha S, Sobhana NV (2022) A survey on deep learning based sentiment analysis. Materials Today: Proceedings, Vol 58, pp 456-460. https://doi.org/10.1016/j.matpr.2022.02.483
K Sarma P, Liang Y, Sethares W (2019) Shallow Domain Adaptive Embeddings for Sentiment Analysis. 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), pp 5549–5558
Kan S.H, Cen Y, Li Y, Vladimir M, He Z (2022) Local Semantic Correlation Modeling Over Graph Neural Networks for Deep Feature Embedding and Image Retrieval. IEEE Transactions on Image Processing. pp 2988 – 3003
Kardakis S, Perikos I, Grivokostopoulou F, Hatzilygeroudis I (2021) Examining Attention Mechanism in Deep Learning Models for Sentiment Analysis. Appl. Sci,11(9):3883. https://doi.org/10.3390/app11093883
Kaity M, Balakrishnan V (2020) Sentiment lexicons and non-english languages: a survey. Knowl Inf Syst 1–36
Kaur G, Sharma A (2023) A deep learning-based model using hybrid feature extraction approach for consumer sentiment analysis. J Big Data 10
Kim T, Lee B (2020) Multi-attention multimodal sentiment analysis. In Proceedings of the 2020 International Conference on Multimedia Retrieval, pp 436–441
Kipf TN, Fetaya E, Wang K, Welling M, Zemel RS (2018) Neural relational inference for interacting systems. In: Proceedings of ICML. PMLR, pp 2688–2697
Korkontzelos I, Nikfarjam A, Shardlow M, Sarker A, Ananiadou S, Gonzalez GH (2016) Analysis of the effect of sentiment analysis on extracting adverse drug reactions from tweets and forum posts. J Biomed Inform 62:148–158. https://doi.org/10.1016/j.jbi.2016.06.007
Lai H, Yan X (2022) Multimodal sentiment analysis with asymmetric window multi-attentions. Multimed Tools Appl 81, pp 19415–19428. https://doi.org/10.1007/s11042-021-11234-y
LeCun, Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521, pp 436–444.
Lee JB, Rossi RA, Kim S, Ahmed NK, Koh E (2018a) Attention Models in Graph s: A Survey. arXiv:1807.07984
Lee JB, Rossi R, Kong X (2018b) Graph Classification using Structural Attention. KDD 2018, pp 19–23
Liao W, Zeng B, Liu J, Wie P, Cheng X, Zhang W (2021) Multi-level graph neural network for text sentiment analysis. Computer & Electrical Engineering Volume 92, ISSN 00457906. https://doi.org/10.1016/j.compeleceng.2021.107096
Ligthart A, Catal C, Tekinerdogan B (2021) Systematic reviews in sentiment analysis: a tertiary study. Artif. Intell. Rev 54 pp 4997–5053. https://doi.org/10.1007/s10462-021-09973-3
Linmei H, Yang T, Shi CH, Ji H, and Li X (2019) Heterogeneous Graph Attention Networks for Semi-Supervised Short Text Classification. 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), pp 4821–4830
Li X, Bing L, Zhang W, Lam W (2019) Exploiting BERT for end-to-end aspect-based sentiment analysis. arXiv preprint
Li X, Lu R, Liu P, et al (2022a) Graph convolutional networks with hierarchical multi-head attention for aspect-level sentiment classification. J Supercomput 78, pp 14846–14865
Li Y, Li N (2022b) Sentiment Analysis of Weibo Comments Based on Graph Neural Network. In IEEE Access, vol.10, pp 23497–23510
Lin P, Luo X, Fan Y (2020) A Survey of Sentiment Analysis Based on Deep Learning. World Academy of Science, Engineering, and Technology, International Journal of Computer and Information Engineering Vol:14, No:12, 2020
Liu B (2012) Sentiment analysis and opinion mining. Synthesis Lectures on Human Language Technologies 5(1), 1–167. https://doi.org/10.2200/S00416ED1V01Y201204HLT016
Liu H, Chatterjee I, Zhou M, Lu XS, Abusorrah A (2020) Aspect-Based Sentiment Analysis: A Survey of Deep Learning Methods. In IEEE Transactions on Computational Social Systems, vol. 7, no. 6, pp 1358–1375. 10.1109/TCSS.2020.3033302
Liu N, Shen B (2020) A novel memory neural network for powerful interaction in aspect-based sentiment analysis. Neurocomputing, Vol 395, pp 66–77, ISSN 0925-2312. https://doi.org/10.1016/j.neucom.2020.02.018
Liu N, Shen B, Zhang Z, Zhang Z, Mi K (2019a) Attention-based Sentiment Reasoner for aspect-based sentiment analysis. Hum. Cent. Comput. Inf. Sci. 9, 35. https://doi.org/10.1186/s13673-019-0196-3
Liu QY, Zhang D, Wu LQ, Li SS (2019b) Multi-modal sentiment analysis with context-augmented LSTM. Computer Science, vol. 46, no. 11, pp 181–185. https://doi.org/10.1109/ICETCI55101.2022.9832378
Luo W, Zhang W, Zhao Y (2021) A Survey of Transformer and GNN for Aspect-based Sentiment Analysis. International Conference on Computer Information Science and Artificial Intelligence (CISAI), Kunming, China, pp 353-357
Luo Y, Wang Y, Xiong Y, Yu C (2022) Deep Learning Techniques for Multimodal Sentiment Analysis. IEEE 2nd International Conference on Electronic Technology, Communication and Information (ICETCI), pp 839–844
Ma Q, Yuan CH, Zhou W, Hu S (2021) Label-Specific Dual Graph Neural Network for Multi-Label Text Classification. ACL
Meng F, Feng J, Yin D, Chen S, Hu M (2020) Sentiment Analysis with weighted Graph Convolutional Networks
Meng J, Long Y, Yu Y, Zhao D, Liu S (2019) Cross-domain text sentiment analysis based on CNN-FT method. Information 10(5):162
Mercha EM, Benbrahim H (2023) Machine learning and deep learning for sentiment analysis across languages: A survey. Neurocomputing, Vol 531, pp 195–216. https://doi.org/10.1016/j.neucom.2023.02.015
Mikolov T, Chen K, Corrado G, Dean J (2013) Efficient estimation of word representations in vector space. arXiv preprint arXiv: 13013 781
Minaee S, Azimi E, Abdolrashidi A (2019) Deep-sentiment: sentiment analysis using ensemble of CNN and bi-LSTM models. http://arxiv .org/abs/arXiv:1904.04206
Minaee S, Kalchbrenner N, Cambria E, Nikzad N, Chenaghlu M, Gao J (2020) Deep learning-based text classification: a comprehensive review 1(1), pp 1–43.
Moraes R, Valiati JF, Neto WPG (2013) Document-level sentiment classification: an empirical comparison between SVM and ANN. Expert Syst Appl 40(2), pp 621–633
Moreo A, Romero M, Castro J, Zurita JM (2012) Lexicon-based comments-oriented news sentiment analyzer system. Expert Syst Appl 39(10), pp 9166–9180
Narasimhan M, Lazebnik S, SchwingA G, (2018) Out of the box: reasoning with graph convolution nets for factual visual question answering. In: Proceedings of NeurIPS, pp 2654–2665
Niu L, Zheng Q, Zhang L (2021) Enhance gated graph neural network with syntactic for sentiment analysis. 2021 IEEE International Conference on Advances in Electrical Engineering and Computer Applications (AEECA), Dalian, China, 2021, pp 1055-1060
Palm R, Paquet U, Winther O (2018) Recurrent relational networks. Proceedings of NeurIPS, pp 3368–3378
Patel HH, Prajapati P (2018) Study and analysis of decision tree-based classification algorithms. Int J Comput Sci Eng 6(10), pp 74–78
Peng H, Li J, He Y, Liu Y, Bao M, Wang L, Song Y, Yang Q (2018) Large-scale hierarchical text classiﬁcation with recursively regularized deep graph-CNN. In Proceedings of the 2018 World Wide Web Conference, pp 1063–1072
Pham P, Nguyen LTT, Pedrycz W, et al (2023) Deep learning, graph-based text representation and classification: a survey, perspectives and challenges. Artif Intell Rev 56, pp 4893–4927. https://doi.org/10.1007/s10462-022-10265-7
Poria S, Chaturvedi I, Cambria E, Bisio F (2016) Sentic LDA: ımproving on LDA with semantic similarity for aspect-based sentiment analysis. In: 2016 international joint conference on neural networks (IJCNN). IEEE, pp 4465–4473
Pouran Ben Veyseh A, Nouri N, Dernoncourt F, Tran QH, Dou D, Nguyen TH (2020) Improving Aspect-based Sentiment Analysis with Gated Graph Convolutional Networks and Syntax based Regulation. In Findings of the Association for Computational Linguistics: EMNLP 2020, pp 4543–4548
Prabha MI, Umarani Srikanth G (2019) Survey of Sentiment Analysis Using Deep Learning Techniques. 1st International Conference on Innovations in Information and Communication Technology (ICIICT), Chennai, India, 2019, pp. 1-9. Doi: 10.1109/ICIICT1.2019.8741438
Pradhyumna P, Shreya GP, Mohana (2021) Graph Neural Network (GNN) in Image and Video Understanding Using Deep Learning for Computer Vision Applications. IEEE Second International Conference on Electronics and Sustainable Communication Systems (ICESC)
Qi S, Wang W, Jia B, Shen J, Zhu SC (2018) Learning human-object interactions by graph parsing neural networks. In: Proceedings of ECCV, pp 401–417
Qiu L, Xiao Y, Qu Y, Zhou H, Li L, Zhang W, Yu Y (2019) Dynamically fused graph network for multi-hop reasoning. In: Proceedings of ACL, pp 6140–6150
Rane A, Kuma A (2018) Sentiment Classification System of Twitter Data for US Airline Service Analysis. IEEE 42nd Annual Computer Software and Applications Conference (COMPSAC, pp 769–773
Ravi K, Ravi V (2015) A survey on opinion mining and sentiment analysis: tasks, approaches, and applications. Knowl-Based Syst 89, pp 14–46
Rehman A, Malik A, Raza B, and Ali W (2019) A hybrid CNN-LSTM model for improving accuracy of movie reviews sentiment analysis. Multimedia Tools and Applications, vol. 78, no. 18, pp 26597–26613
Revathy R, Lawrance R (2017) Comparative analysis of c4. 5 and c5. 0 algorithms on crop pest data. Int J Innovative Res Comput Commun Eng 5(1), pp 50–58
Rhee S, Seo S, Kim S (2018) Hybrid Approach of Relation Network and Localized Graph Convolutional Filtering for Breast Cancer Subtype Classiﬁcation. In: Proceedings of IJCAI, pp 3527–3534
Rossi A, Barbosa D, Firmani D, Matinata A, Merialdo P (2021) Knowledge Graph Embedding for Link Prediction. A Comparative Analysis ACM Transaction on Knowledge Discovery from DataV15, pp 1–49
Sadr H, Pedram M.M, Teshnehlab M.A (2019) Robust Sentiment Analysis Method Based on Sequential Combination of Convolutional and Recursive Neural Networks. Neural Process Lett 50, pp 2745–2761. https://doi.org/10.1007/s11063-019-10049-1
Sahu SK, Christopoulou F, Miwa M, Ananiadou S (2019) Inter-sentence relation extraction with document-level graph convolutional neural network. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp 4309–4316
Sanchez A, Heess N, Springenberg J.T, Merel J, Hadsell R, Riedmiller M.A, Battaglia P (2018) Graph networks as learnable physics engines for inference and control. In: Proceedings of ICML, pp 4470–4479
Sato R, Yamada M, Kashima H, (2019) Approximation ratios of graph neural networks for combinatorial problems. Proceedings of NeurIPS, pp 4081–4090
Saunders D (2021) Domain adaptation for neural machine translation. Ph.D. thesis, University of Cambridge
Schlichtkrull M, Kipf TN, Bloem P, Van Den Berg R, Titov I, Welling. Modeling M (2018) relational data with graph convolutional networks. In European Semantic Web Conference Springer, pp 593–607
Schouten K, Frasincar F (2015) Survey on aspect-level sentiment analysis. IEEE Trans Knowl Data Eng 2
Shang C, Tang Y, Huang J, Bi J, He X, Zhou B (2019) End-to-end structure-aware convolutional networks for knowledge base completion. In Proceedings of the AAAI Conference on Artiﬁcial Intelligence, vol 33, pp 3060–3067
Sharma R, Bhattacharyya P, Dandapat S, Sharad Bhatt H (2018) Identifying transferable information across domains for cross-domain sentiment classification. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, ACL 2018, Vol 1: Long Papers, pp 968–978
Shen J, Ma M, Xiang R, Lu Q, Vallejos EP, Xu G, Huang CR, Long Y (2020) Dual memory network model for sentiment analysis of review text. Knowledge-Based Systems, vol. 188, 105004
Shi C, Xu M, Zhu Z, Zhang W, Zhang M, Tang J (2020) GraphAF: A Flow-Based Autoregressive Model for Molecular Graph Generation. Proceedings of ICLR
Shoukry A, Rafea A (2015) A hybrid approach for sentiment classification of Egyptian dialect tweets. First International Conference on Arabic Computational Linguistics (ACLing)
Shrestha N, Nasoz F (2019) Deep learning sentiment analysis of amazon.com reviews and ratings. International Journal on Soft Computing, Artiﬁcial Intelligence, and Applications, vol. 8, no. 1, pp 1–15
Shuman DI, Narang SK, Frossard P, Ortega A, Vandergheynst P (2013) The emerging ﬁeld of signal processing on graphs: extending high-dimensional data analysis to networks and other irregular domains. IEEE SPM 30, pp 83–98
Singh JP, Irani S, Rana NP, Dwivedi YK, Saumya S, Roy PK (2017) Predicting the “helpfulness’’ of online consumer reviews. J Bus Res 70, pp 346–355
Singh M, Jakhar AK, Pandey S (2021) Sentiment analysis on the impact of coronavirus in social life using the BERT model. Soc Netw Anal Min 11(1), pp 1–11
Soleymani M, Garcia D, Jou B, Schuller B, Chang SF, Pantic M (2017) A survey of multimodal sentiment analysis. Image Vis Comput 65, pp 3–14
Stappen L, Schuller B, Lefter I, Cambria E, Kompatsiaris I (2020) Summary of Muse 2020: multimodal sentiment analysis, emotion-target engagement, and trustworthiness detection in real-life media. In: Proceedings of the 28th ACM international conference on Multimedia, pp 4769–4770
Sun K, Zhang R, Mensah S, Mao Y, Liu X (2019) 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), pp 5679–568
Tang H, Ji D, Li C, Zhou Q (2020) Dependency graph enhanced dual-transformer structure for aspect-based sentiment classiﬁcation. In Proceedings of the 58thAnnualMeetingof the Association for Computational Linguistics, pp 6578–6588
Tay Y, Tuan LA, and Hui SC (2018) Learning to attend via word-aspect associative fusion for aspect-based sentiment analysis. in Proceedings of the Thirty-Second AAAI Conference on Artiﬁcial Intelligence, vol. 18, pp 5956–5963
Trinh S, Nguyen L, Vo M, Do P (2016) Lexicon-based sentiment analysis of Facebook comments in Vietnamese language. Springer, pp 263–276
Tu M, Wang G, Huang J, Tang Y, He X, Zhou B (2019) Multi-hop reading comprehension across multiple documents by reasoning over heterogeneous graphs. In: Proceedings of ACL, pp 2704–2713
Usama M, Ahmad B, Song E, Shamim Hossain M, Alrashoud M, Muhammad GH (2020) Attention-based sentiment analysis using convolutional and recurrent neural network. Future Generation Computer Systems, Vol 113, pp 571-578, ISSN 0167-739X, https://doi.org/10.1016/j.future.2020.07.022
Van de Camp M, Van den Bosch A (2012) The socialist network. Decis Support Syst 53(4), pp 761–769
Vanaja S, Belwal M (2018) Aspect-Level Sentiment Analysis on E-Commerce Data. 2018 International Conference on Inventive Research in Computing Applications (ICIRCA), pp 1275–1279
Vaswani A Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez AN, Kaiser L, Polosukhin I (2017) Attention is all you need. In NIPS
Velickovic P, Cucurull G, Casanova A, Romero A, Li'o P, Bengio Y (2018) Graph attention networks. In the 6th International Conference on Learning Representations
Vignac C, Loukas A, Frossard P (2020) Building powerful and equivariant graph neural networks with structural message-passing, 34th conference on Neural Information Processing Systems (NeurIPS 2020)
Wang J, Yu LC, Robert Lai K, Zhang X (2019a) Investigating Dynamic Routing in Tree-Structured LSTM for Sentiment Analysis. Association for Computational Linguistics, 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), pp 3432–3437
Wang J, Xu B, Zu Y (2021) Deep learning for Aspect-based Sentiment Analysis. International Conference on Machine Learning and Intelligent Systems Engineering (MLISE), Chongqing, China, 2021, pp 267–271
Wang K, Shen W, Yang Y, Quan X, Wang R (2020a) Relational Graph Attention Network for Aspect-based Sentiment Analysis. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp 3229–3238
Wang M, Zhu Y, Liu SH, Song CH, Wang ZH, Wang P, Qin X (2019b) Sentiment Analysis Based on Attention Mechanisms and Bi-Directional LSTM Fusion Model. IEEE SmartWorld. Ubiquitous Intelligence & Computing, Advanced & Trusted Computing, Scalable Computing & Communication, Cloud & Big Data Computing, Internet of People and Smart City Innovation, pp 865–868
Wang SH, Hu L, Wang Y, He X, Sheng QZ, et al (2020) Graph learning approaches to recommender systems: A review. arXiv:2004.11718, pp 4644–4652
Wang X, Ji H, Shi C, Wang B, Ye Y, Cui P, Yu PS (2019c) Heterogeneous graph attention network. In: Proceedings of WWW, pp 2022–2032
Wang Z, Chen T, Ren J.S.J, Yu W, Cheng H, Lin L, (2018) Deep reasoning with knowledge graph for social relationship understanding. Proceedings of IJCAI, pp 1021–1028
Wankhade M, Rao ACS, Kulkarni C (2022) A survey on sentiment analysis methods, applications, and challenges. Artif Intell Rev 55, pp 5731–5780
Wu L, Chen Y, Shen K, Guo X, Gao H, Li S.H, Pei J, Long B (2021) Graph Neural Networks for Natural Language Processing: A Survey. ArXiv: 2106.06090
Wu Q, Zhang H, Gao X, He P, Weng P, Gao H, Chen G, (2019a). Dual graph attention networks for deep latent representation of multifaceted social effects in recommender systems. In: Proceedings of WWW, pp 2091–2102
Wu Z, Pan S, Chen F, Long G, Zhang C, Yu P.S (2019b) A Comprehensive Survey on Graph Neural Networks arXiv preprint arXiv:1901.00596
Xiao Y, Wang DY, Hou LG (2019) Unsupervised emotion recognition algorithm based on improved deep belief model in combination with probabilistic linear discriminant analysis. Personal and Ubiquitous Computing, pp 553–562
Xie Q, Huang J, Du P, Peng M, Nie J.Y (2021) Inductive Topic Variational Graph Auto-Encoder for Text Classification. NAACL
Xing BW, Liao LJ, Song DD, Wang JG, Zhang FZ, Huang HY (2019) Earlier attention? aspect-aware LSTM for aspect-based sentiment analysis. in Proceedings of the Twenty-Eighth International Joint Conference on Artiﬁcial Intelligence, 2019, pp 5313–5319
Xu K, Wang L, Yu M, Feng Y, Song Y, Wang Z, Yu D (2019a) Cross-lingual knowledge graph alignment via graph matching neural network. In: Proceedings of ACL. Association for Computational Linguistics, pp 3156–3161
Xu L, Pang X, Wu J, Cai M, Peng J (2023) Learn from structural scope: Improving aspect-level sentiment analysis with hybrid graph convolutional networks. Neurocomputing, Vol 518, pp 373–383
Xu N, Mao WJ, Chen GD (2019b) multi-interactive memory network for aspect-based multimodal sentiment analysis. In Proceedings of the Thirty-Third AAAI Conference on Artiﬁcial Intelligence, pp 371–378
Xu Q, Zhu L, Dai T, Yan CH (2020) Aspect-based sentiment classification with multi-attention network. Neurocomputing, Vol 388, pp 135–143, ISSN 0925-2312. https://doi.org/10.1016/j.neucom.2020.01.024
Yadav A, Vishwakarma D.K (2020) Sentiment analysis using deep learning architectures: a review. Artif Intell Rev 53, pp 4335–4385 https://doi.org/10.1007/s10462-019-09794-5
Yadav RK, Jiao L, Goodvin M, Granmo O-CH (2021) Positionless aspect-based sentiment analysis using attention mechanism. Knowledge-Based Systems, Vol 226, ISSN 0950-7051
Yao L, Mao, C, Luo Y (2019) Graph convolutional networks for text classiﬁcation. Proceedings of AAAI 33, pp 7370–7377
Yu B, Yin H, Zhu Z (2018) Spatio-temporal Graph Convolutional Networks: A Deep Learning Framework for Trafﬁc Forecasting. Proceedings of IJCAI, pp 3634–3640
Yu Y, Lin H, Meng J (2016) Visual and textual sentiment analysis of a microblog using deep convolutional neural networks. Algorithms, vol.9, no. 2, pp 41–51
Zeng J, Liu T, Jia W, Zhou J (2021) Fine-grained Question-Answer sentiment classification with hierarchical graph attention network. Neurocomputing, pp 214–224
Zeng S, Xu R, Chang B, Li L (2020) Double graph-based reasoning for document-level relation extraction. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp1630–1640
Zhang C, Li Q, Song D (2019a) 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), pp 4568–457
Zhang J, Shi X, Xie J, Ma H, King I, Yeung DY (2018a) GaAN: gated attention networks for learning on large and spatiotemporal graphs. In: Proceedings of UAI
Zhang L, Wang S, Liu B (2018b) Deep learning for sentiment analysis: a survey. Wiley Interdiscip Rev Data Min Knowl Discov 8(4): e1253
Zhang M, Chen Y (2018c) Link Prediction Based on Graph Neural Networks. 32nd Conference on Neural Information Processing Systems (NIPS 2018), arXiv:1802.09691v3
Zhang M, Qian T (2020) 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), pp 3540–3549
Zhang S, Tong H, Xu J, Maciejewski R (2019b) Graph convolutional networks: a comprehensive review. Computational Social Networks 6 (1), 1–23
Zhang W, Li X, Deng Y, Bing L, Lam W (2022) A Survey on Aspect-Based Sentiment Analysis: Tasks, Methods, and Challenges. In IEEE Transactions on Knowledge and Data Engineering, pp 1-20.10.1109/TKDE.2022.3230975
Zhang W, Li X, Deng Y, Bing L, Lam W (2021) Aspect sentiment quad prediction as paraphrase generation. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp 9209–9219
Zhang Y, Song D, Li X, Zhang P, Wang P, Rong L, Yu G, Wang B (2020a) A quantum-like multimodal network framework for modeling interaction dynamics in multiparty conversational sentiment analysis. Information Fusion, vol. 62, pp 14–31
Zhang Y, Yu X, Cui Z, Wu S, Wen Z, Wang L (2020b) Every Document Owns Its Structure: Inductive Text Classification via Graph Neural Networks. arXiv:2004.13826. https://doi.org/10.48550/arXiv.2004.13826
Zheng C, Fan X, Wang C, Qi J, Gman (2020a) A graph multi-attention network for trafﬁc prediction. Proceedings of AAAI 34, pp 1234–1241
Zheng X, Dan C, Aragam B, Ravikumar P, Xing E (2020b) Learning sparse nonparametric dags. In: Proceedings of AISTATS. PMLR, pp 3414–3425
Zhou J, Cui G, Hu SH, Zhang ZH, Yang CH, Liu ZH, Wang L, Li CH, Sun M (2022) Graph neural networks: A review of methods and applications. AI Open, Vol 1, pp 57–81. https://doi.org/10.1016/j.aiopen.2021.01.001

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

A survey of sentiment analysis methods based on graph neural network

Status:

Version 1

Abstract

1. Introduction

2. background and related works

2.1 Approaches

2.1.1 lexicon-based approaches

2.1.2 Machine learning approaches

2.1.3 Deep learning approaches

2.1.4 Transfer learning

2.1.5 Hybrid approaches

2.2 Multimodal sentiment analysis (MSA)

2.3 Sentiment analysis levels

3. GNN-based approaches in sentiment analysis

4. Datasets

5. Conclusion and future works

Declarations

References

Additional Declarations

Status:

Version 1