sentiment analysis research papers 2022

sentiment analysis Recently Published Documents

Total documents.

• Latest Documents
• Most Cited Documents
• Contributed Authors
• Related Sources
• Related Keywords

Aspect-based Sentiment Analysis using Dependency Parsing

In this paper, an aspect-based Sentiment Analysis (SA) system for Hindi is presented. The proposed system assigns a separate sentiment towards the different aspects of a sentence as well as it evaluates the overall sentiment expressed in a sentence. In this work, Hindi Dependency Parser (HDP) is used to determine the association between an aspect word and a sentiment word (using Hindi SentiWordNet) and works on the idea that closely connected words come together to express a sentiment about a certain aspect. By generating a dependency graph, the system assigns the sentiment to an aspect having a minimum distance between them and computes the overall polarity of the sentence. The system achieves an accuracy of 83.2% on a corpus of movie reviews and its results are compared with baselines as well as existing works on SA. From the results, it has been observed that the proposed system has the potential to be used in emerging applications like SA of product reviews, social media analysis, etc.

Sentiment Analysis Applied to News from the Brazilian Stock Market

Trg-datt: the target relational graph and double attention network based sentiment analysis and prediction for supporting decision making.

The management of public opinion and the use of big data monitoring to accurately judge and verify all kinds of information are valuable aspects in the enterprise management decision-making process. The sentiment analysis of reviews is a key decision-making tool for e-commerce development. Most existing review sentiment analysis methods involve sequential modeling but do not focus on the semantic relationships. However, Chinese semantics are different from English semantics in terms of the sentence structure. Irrelevant contextual words may be incorrectly identified as cues for sentiment prediction. The influence of the target words in reviews must be considered. Thus, this paper proposes the TRG-DAtt model for sentiment analysis based on target relational graph (TRG) and double attention network (DAtt) to analyze the emotional information to support decision making. First, dependency tree-based TRG is introduced to independently and fully mine the semantic relationships. We redefine and constrain the dependency and use it as the edges to connect the target and context words. Second, we design dependency graph attention network (DGAT) and interactive attention network (IAT) to form the DAtt and obtain the emotional features of the target words and reviews. DGAT models the dependency of the TRG by aggregating the semantic information. Next, the target emotional enhancement features obtained by the DGAT are input to the IAT. The influence of each target word on the review can be obtained through the interaction. Finally, the target emotional enhancement features are weighted by the impact factor to generate the review's emotional features. In this study, extensive experiments were conducted on the car and Meituan review data sets, which contain consumer reviews on cars and stores, respectively. The results demonstrate that the proposed model outperforms the existing models.

A Comprehensive Guideline for Bengali Sentiment Annotation

Sentiment Analysis (SA) is a Natural Language Processing (NLP) and an Information Extraction (IE) task that primarily aims to obtain the writer’s feelings expressed in positive or negative by analyzing a large number of documents. SA is also widely studied in the fields of data mining, web mining, text mining, and information retrieval. The fundamental task in sentiment analysis is to classify the polarity of a given content as Positive, Negative, or Neutral . Although extensive research has been conducted in this area of computational linguistics, most of the research work has been carried out in the context of English language. However, Bengali sentiment expression has varying degree of sentiment labels, which can be plausibly distinct from English language. Therefore, sentiment assessment of Bengali language is undeniably important to be developed and executed properly. In sentiment analysis, the prediction potential of an automatic modeling is completely dependent on the quality of dataset annotation. Bengali sentiment annotation is a challenging task due to diversified structures (syntax) of the language and its different degrees of innate sentiments (i.e., weakly and strongly positive/negative sentiments). Thus, in this article, we propose a novel and precise guideline for the researchers, linguistic experts, and referees to annotate Bengali sentences immaculately with a view to building effective datasets for automatic sentiment prediction efficiently.

Employee Sentiment Analysis Towards Remote Work during COVID-19 Using Twitter Data

Topic modelling and sentiment analysis of global warming tweets.

With the increasing extreme weather events and various disasters, people are paying more attention to environmental issues than ever, particularly global warming. Public debate on it has grown on various platforms, including newspapers and social media. This paper examines the topics and sentiments of the discussion of global warming on Twitter over a span of 18 months using two big data analytics techniques—topic modelling and sentiment analysis. There are seven main topics concerning global warming frequently debated on Twitter: factors causing global warming, consequences of global warming, actions necessary to stop global warming, relations between global warming and Covid-19; global warming’s relation with politics, global warming as a hoax, and global warming as a reality. The sentiment analysis shows that most people express positive emotions about global warming, though the most evoked emotion found across the data is fear, followed by trust. The study provides a general and critical view of the public’s principal concerns and their feelings about global warming on Twitter.

Transparent Aspect-Level Sentiment Analysis Based on Dependency Syntax Analysis and Its Application on COVID-19

Aspect-level sentiment analysis identifies fine-grained emotion for target words. There are three major issues in current models of aspect-level sentiment analysis. First, few models consider the natural language semantic characteristics of the texts. Second, many models consider the location characteristics of the target words, but ignore the relationships among the target words and among the overall sentences. Third, many models lack transparency in data collection, data processing, and results generating in sentiment analysis. In order to resolve these issues, we propose an aspect-level sentiment analysis model that combines a bidirectional Long Short-Term Memory (LSTM) network and a Graph Convolutional Network (GCN) based on Dependency syntax analysis (Bi-LSTM-DGCN). Our model integrates the dependency syntax analysis of the texts, and explicitly considers the natural language semantic characteristics of the texts. It further fuses the target words and overall sentences. Extensive experiments are conducted on four benchmark datasets, i.e., Restaurant14, Laptop, Restaurant16, and Twitter. The experimental results demonstrate that our model outperforms other models like Target-Dependent LSTM (TD-LSTM), Attention-based LSTM with Aspect Embedding (ATAE-LSTM), LSTM+SynATT+TarRep and Convolution over a Dependency Tree (CDT). Our model is further applied to aspect-level sentiment analysis on “government” and “lockdown” of 1,658,250 tweets about “#COVID-19” that we collected from March 1, 2020 to July 1, 2020. The experimental results show that Twitter users’ positive and negative sentiments fluctuated over time. Through the transparency analysis in data collection, data processing, and results generating, we discuss the reasons for the evolution of users’ emotions over time based on the tweets and on our models.

Aspect Based Sentiment Analysis of Unlabeled Reviews Using Linguistic Rule Based LDA

In this digital era, people are very keen to share their feedback about any product, services, or current issues on social networks and other platforms. A fine analysis of these feedbacks can give a clear picture of what people think about a particular topic. This work proposed an almost unsupervised Aspect Based Sentiment Analysis approach for textual reviews. Latent Dirichlet Allocation, along with linguistic rules, is used for aspect extraction. Aspects are ranked based on their probability distribution values and then clustered into predefined categories using frequent terms with domain knowledge. SentiWordNet lexicon uses for sentiment scoring and classification. The experiment with two popular datasets shows the superiority of our strategy as compared to existing methods. It shows the 85% average accuracy when tested on manually labeled data.

Aspect Based Sentiment Analysis of Unlabeled Reviews using Linguistic Rule Based LDA

Measuring citizen satisfaction with e-government services by using sentiment analysis technology, export citation format, share document.

ORIGINAL RESEARCH article

Text sentiment analysis based on transformer and augmentation.

• 1 School of Computer Science and Technology, Soochow University, Suzhou, China
• 2 School of Computer Science and Engineering, Changshu Institute of Technology, Suzhou, China

With the development of Internet technology, social media platforms have become an indispensable part of people’s lives, and social media have been integrated into people’s life, study, and work. On various forums, such as Taobao and Weibo, a large number of people’s footprints are left all the time. It is these chats, comments, and other remarks with people’s emotional evaluations that make up part of public opinion. Analysis of this network public opinion is conducive to maintaining the peaceful development of society. Therefore, sentiment analysis has become a hot research field and has made great strides as one of the hot topics in the field of natural language processing. Currently, the BERT model and its variants have achieved excellent results in the field of NLP. However, these models cannot be widely used due to huge demands on computing resources. Therefore, this paper proposes a model based on the transformer mechanism, which mainly includes two parts: knowledge distillation and text augmentation. The former is mainly used to reduce the number of parameters of the model, reducing the computational cost and training time of the model, and the latter is mainly used to expand the task text so that the model can achieve excellent results in the few-sample sentiment analysis task. Experiments show that our model achieves competitive results.

Introduction

In today’s era of extremely complex social networking, the number of Internet users in China has increased dramatically. As of 2021, the number of domestic Internet users will reach 989 million ( Cnnic, 2021 ). Such a huge scale that it involves a wide range of fields, such as shopping, games, communications, and video. In all these areas, there is a huge amount of data and information left behind, and these data and information have formed a certain tendency of online opinion on the Internet, which is both positive, negative, and neutral. Many companies can recommend products based on users’ comments and preferences, thereby increasing product sales; while some undesirable emotional messages can cause social unrest, such as donation scams, which can lead to heated debates among internet users. For the textual information formed by some remarks on the Internet, if sentiment analysis can be carried out in a certain way, the chaotic information can be extracted by certain means, and then differentiated according to different categories to form useful information for the society, which can ultimately guide the normal development of society is a direction worthy of study ( Zhang et al., 2020 ).

To solve the above problems, deep learning plays an important role. Deep learning ( LeCun et al., 2015 ) is to learn the inherent laws and representation levels of sample data, and the information obtained during these learning processes is of great help to the interpretation of data, such as text, images, and sounds. Its ultimate goal is to enable machines to have the ability to analyze and learn like humans, and to recognize data, such as words, images, and sounds. Deep learning is a complex machine learning algorithm that far outperforms previous related technologies in speech and image recognition.

Deep learning requires a large amount of labeled data to get excellent results, and the need for large amounts of labeled data is one of its most fundamental shortcomings. When the labeled data are limited, supervised deep learning models are tended to suffer from overfitting ( Hawkins, 2004 ). The strong reliance on labeled data limits the practical use of deep learning models, due to the need for a large amount of time and money to obtain enough labeled data. As a result, semi-supervised learning (SSL; Van Engelen and Hoos, 2020 ) has received much attention, because it is one of the most promising paradigms of leveraging unlabeled data to address this weakness ( Chawla and Karakoulas, 2005 ).

The current NLP technology is witnessing a revolution in pre-trained self-supervised models. These models often have hundreds of millions of parameters. Among these models, BERT ( Devlin et al., 2018 ) is one of the most representatives. These models show substantial improvement in results on previous tasks. However, as one of the largest models in the history of natural language processing, the BERT model has a large number of parameters, large size, and high latency. These drawbacks prevent the model from running on devices with limited computing power or taking too long to run. In some real-time scenarios, the long response times would outweigh the benefits in comparison to the degree of improvement in accuracy. Therefore, we investigate reducing the number of parameters of the model while retaining its accuracy, thereby reducing the computational cost and response time required to run the model. This paper investigates how to reduce the parameters of the model. It is found that good results can be achieved using knowledge distillation and text augmentation for text tasks. Experimental results show that the accuracy of text sentiment classification can be improved with a small number of samples.

The contributions of this paper are summarized as follows:

1. We propose a text sentiment classification model based on the Transformer mechanism, which combines knowledge distillation and text augmentation methods to improve the accuracy of sentiment classification in the few-shot labeling task.

2. The model reduces the number of parameters based on the Transformer mechanism by means of knowledge distillation and uses the method of text augmentation to solve the problem of low accuracy for the few-sample task.

3. We conduct experiments on two public corpora and compare with different state-of-the-art methods to verify the performance of the model.

Related Work

Neural language model for nlp.

Sentiment analysis is the classification of emotions and attitudes in subjective texts, and the main methods are machine learning and deep learning. From the perspective of machine learning, the text sentiment analysis method based on machine learning ( Lin and Luo, 2020 ) needs to use a corpus to train a classification model. For example, Yan et al. (2020) used the bag-of-words model to classify text data and considered Word2Vec to establish a new word vector; in the ensemble algorithm Naive Bayes and SVM ( Tiwari and Mallick, 2016 ), both the precision and recall rate were improved. From the perspective of deep learning, in recent years, deep learning technology has made great progress in processing text information-related tasks. The neural network structure has achieved remarkable results in text classification. The neural network structure is used to construct different neural network algorithms, such as CNN, bidirectional LSTM, text CNN, and so on. Chen (2015) first proposed to apply CNN to text orientation analysis, and obtained good results; on the basis of Chen’s analysis, Conneau et al. (2016) proposed a VDCNN model by using a deep convolutional network method. The subsequent rise of pre-trained models, which have led to major breakthroughs in the field of NLP, The BERT model proposed by ( Vaswani et al., 2017 ) is a better improvement in terms of evaluation indicators. Devlin et al. (2018) mainly introduced the practical value of the BERT model, and obtained good research results on 11 natural language processing tasks.

One of the major breakthroughs in the history of natural language processing is the attention mechanism. The attention mechanism is proposed to solve long-term dependency problems of models that use a single context vector compressing every input from previous time steps. The attention mechanism makes it possible to capture the global semantics in the texts. Because the attention mechanism allows the model to take hidden states from multiple time steps as input and calculate the importance of the input regarding the current time step. Transformer serving as the core architecture of the attention mechanism has been widely used for language models ( Han et al., 2021 ) to capture complex linguistic patterns. These pre-trained language models work well on various NLP tasks; however, it is accepted that these pre-trained models rely on large-scale data training and have high requirements for the quality of the data set.

Pre-training and Fine-Tuning Framework

Neural network methods generally start with random initialization of model parameters and then train the model parameters using optimization algorithms, such as back propagation and gradient descent. Before the advent of pre-training techniques, the application of neural network-based deep learning in NLP faced the following problems: firstly, deep learning models at this time were not complex enough, and simply stacking neural network layers did not bring more performance gains; secondly, data-driven deep learning models lacked large-scale annotated data, and manual annotation was too expensive to drive complex models. Therefore, pre-training techniques based on knowledge augmentation, migration learning, and multi-task learning are gradually being given more attention by scholars.

The pre-training and fine-tuning framework have achieved great success in recent years, especially in the NLP field, and has been applied to a series of NLP tasks. Howard and Ruder (2018) proposed to pre-train a language model on a large corpus and fine-tune it on the target task. Such a model uses some novel techniques like gradual unfreezing and slanted triangular learning rates. Encouraged by the good performance of the pre-trained models, researchers get excellent performance even with small amounts of labeled data. Pre-trained models are often applied to different objectives, such as language modeling and masked language modeling. With the increase in the training data size, the performance of pre-trained models is also improved ( Araci, 2019 ; Rezaeinia et al., 2019 ).

Data Augmentation for Language Data

Data augmentation is a technique that can increase the size of a data set. Many researchers work on data augmentation, mainly in the field of computer vision, speech, etc. Compared to these, there is less research on data augmentation in the field of text and there is no standard method yet. The commonly used method is to use a dictionary or thesaurus or database of synonyms to make word replacements. In a situation where there is no dictionary, the other way is to use distributed word representation for finding similar words. A method belonging to this is called synonym augmentation. In fact, the best way to augment is to artificially change the wording of the language, but the cost of this method is too expensive. Therefore, the most option in data augmentation for most research is to replace words or phrases with their synonyms. For example, the most popular open-source lexical database for the English language is WordNet ( Fellbaum, 2010 ). Another method is called semantic similarity augmentation, it uses distributed word representation, namely, word embedding, one can identify semantically similar words. This approach requires either pre-trained word embedding models for the language at hand or enough data from the target application to be able to build the embedding model. Its advantage is that no additional dictionaries are required to find synonyms. And the last method is back translation, its process is translating words, phrases, or texts into another language, namely, forward translation, then translating the results back into the original language, this is called back translation ( Edunov et al., 2018 ).

Mixup augmentation ( Guo, 2020 ) creates new training examples by drawing samples from the original data and combining them convexly, usually, the samples are two or even more, it combines the data both in terms of the input and output. It takes pair of samples from the initial data set and sums both the input and output. The main idea of the mixup is a sample. Given two labeled data ( x i , y i ) and ( x j , y j ) , where x is the input vector and y is the one-hot encoded labels, the algorithm creates training samples by linear interpolations:

Where λ ∈ [ 0 , 1 ] , and λ ∼ B e t a ( α , α ) λ = max ( λ , 1 − λ ) in which α is the hyper-parameter to control the distribution of λ . It is best suited for learning models that use the cross-entropy loss and change the input. Augmentation by mixup can be done on text representation for text problems. as such we can use mixup with bag-of-words models, word embeddings, and language models.

Knowledge Distillation

Deep learning has achieved incredible performance in numerous fields, including computer vision, speech recognition, natural language processing, and more. However, most models are too computationally expensive to run on mobile or embedded devices. Therefore, the model needs to be compressed, and knowledge distillation ( Gou et al., 2021 ) is one of the important techniques in model compression. Knowledge distillation adopts the Teacher–Student mode: the complex and large model is used as the teacher, the student model structure is relatively simple, and the teacher is used to assist the training of the student model. The teacher has strong learning ability and can transfer the knowledge it has learned to relatively weak learning ability. The student model, in order to enhance the generalization ability of the student model. The complex and cumbersome but effective Teacher model is not online, it is simply a mentor role, and the flexible and lightweight Student model is really deployed online for prediction tasks.

Vanilla Knowledge Distillation is simply the learning of lightweight student models from the soft targets output by the teacher model. However, when the teacher model becomes deeper, learning the soft targets alone is not enough. Therefore, we need to acquire not only the knowledge output from the teacher model, but also other knowledge that is implicit in the teacher model, such as output feature knowledge, intermediate feature knowledge, relational feature knowledge, and structural feature knowledge. The four forms of knowledge distilled from the student’s problem-solving perspective can be compared to the following: output knowledge provides the answer to a problem, intermediate knowledge provides the process of solving a problem, relational knowledge provides the method of solving a problem, and structural knowledge provides the complete body of knowledge.

Materials and Methods

In this section, we have made improvements to the Mixup method. Mixup can be interpreted in different ways. On the one hand, the Mixup method can be viewed as a data augmentation approach that creates new data based on the original training set. On the other hand, it enhances the regularization of the model. Mixup was first designed for images tasks; thus, mixup was demonstrated to work well on continuous image data; however, it is challenging to extend mixup from image to text, since it is infeasible to compute the interpolation of discrete tokens. To overcome this challenge, we propose a novel method that expands on the original text and interpolates in the hidden space. Inspired by the excellent performance of the pre-trained models like Bidirectional Encoder Representations from Transformers (BERT), we use a multi-layer model to encode the given text to get the semantic representations, in this process; we expand the original text and apply interpolations within hidden space as a data augment method for text. For an encoder with L layers, we choose to mix up the hidden representation at layer K.

As shown in Figure 1 , we first use text augmentation techniques to expand the labeled data size, such as EDA, back translation, and Word2vec-based (learned semantic similarity) augmentation, and then, we compute the hidden representation in the bottom layers and mixup the hidden representation at layer m, and feed the interpolated hidden representations to the upper layers. In mathematical expressions, we denote the layer m in the encoder network as m i x u p m ( . ; θ ) , thus the hidden representation of the layer m can be computed as h i d d e n m = m i x u p m ( h i d d e n m − 1 ; θ ) . For the text samples X i and its augmentation X j define the layer 0 as the embedding layer, for example, h 0 i = W E X i and the hidden representation of the two samples from the lower layers and the mixup at layer k and forward passing to upper layers are defined as follows:

Figure 1 . Mixup augmentation text.

We use an encoder model m i x u p ( . ; θ ) , Equation (3) represents the mixup method before the fusion layer m , where θ represents the remaining parameters in the model. Equation (4) expresses the calculation method in the mixed layer. Equations ( 5 , 6 ), respectively, express the operation above the mixed layer and the whole mixing method. MATEXT combines the original text and its simple augmentation text as input and interpolates textual semantic hidden representations as a type of data augmentation. Compared to the mixup ( Zhang et al., 2017 ) defined in the augmentation on text representation. In the experiment, we sample the mix parameter λ from a Beta distribution for every batch to perform the interpolation:

In the equation, α is the hyper-parameter to control the distribution of λ . In MATEXT, different from the original mixup, we share the label of the original text and its augmentation text.

Data Augmentation

We use Easy Data Augmentation (EDA) technique ( Wei and Zou, 2019 ) which consists of four different text editing operations: (1) Synonym replacement: N-words are randomly selected from the sentence and replaced with one of its synonyms chosen at random. (2) Random noise injection: N-words are randomly selected from the sentence and a single character of each word is replaced with a random alphabetical character. (3) Random swap: we randomly choose two words in the sentence and swap their positions and repeat N times. (4) Random deletion: N-words from the sentence are randomly chosen and removed from the sentence.

The value N is depended on the length of each sentence. In our experiment, the p is set 0.1 and calculated p × l e n ( s e n t e n c e ) , where the number words in the sentence is used as a length of the sentence. Rounded up value of p × l e n ( s e n t e n c e ) is used as the value of N.

In addition, Back translations are also a common data augmentation technique and can generate diverse paraphrases while preserving the semantics of the original sentences. And Word2vec is another robust augmentation method that uses a word embedding model trained on the public data set to find the most similar words for a given input word, which is called Word2vec-based (learned semantic similarity) augmentation ( Budhkar and Rudzicz, 2018 ). Table 1 shows some examples of text augmentation.

Table 1 . Example of text augmentation.

The main purpose of this chapter is to reduce the parameters of the current mainstream BERT model through the method of knowledge distillation. We study the knowledge distillation under a limited amount of labeled data and a large amount of unlabeled data. With sufficiently accurate teacher models and large amounts of distilled unlabeled data, using Bi-LSTM and RNN as encoders as student models can greatly reduce the parameter space and perform on par with large pre-trained models. The distillation technique is mainly divided into two parts, first using a fine-tuned teacher model trained with labeled text data to automatically label a large amount of unlabeled data and then using the augmented data to train a student model with a supervised cross-entropy loss. In the second part, we used the logarithmic and internal representations from the transformer mechanism to generate training student models on unlabeled data, using different training schedules to optimize different loss functions, and experiments proved that this partial distillation can further regularize these models to improve performance. The specific steps are shown in Figure 2 .

Figure 2 . Knowledge distillation steps diagram.

The definition of the loss function is shown in following formula:

The soft loss function is for the teacher model to label the unlabeled data and then train the student model, while the hard loss function is for the student model to predict the label and take the cross-entropy of the labeled data, because there may be different probability distributions of different categories in the training data. Balanced, a reasonable distribution of soft labels and hard labels, that is, the T value in the formula can improve the generalization ability of the model.

Among them, T is the temperature coefficient, which is used to control the softening degree of the output probability. It is easy to see that Equation (11) represents the class probability of the network output Softmax when T  = 1. When T is positive infinity, Hinton et al. (2015) show in their paper that Equation (11) represents the logical unit of the network output at this point. v i and z i are both generated by the neural network using the softmax layer to generate class probabilities. The output layer converts the logarithm of z i calculated by each class into probabilities q i and pi by comparing z i and other logical probabilities. From Jiao et al. (2019) , it is known from experiments that knowledge distillation from the transformer to the non-transformer framework will limit the knowledge learnable from the teacher model due to the incomparability of the parameters of the intermediate layers, due to the transfer of knowledge to smaller models, the effect is not good, and the knowledge distillation transferred to the same transformer framework, due to the interoperability of the middle layer, the parameters can be used, and the effect is better than the former because the reduction of the parameter amount will also reduce the running time in the actual processing process and computing power costs.

The model structure we designed is shown in Figure 3 . The input text data are expanded and then mixed in the transformer layer before training and classification. In addition to the traditional modules, the model is mainly divided into four parts, namely, the data augmentation part, the mixed layer, the linear layer, and the stacked module.

Figure 3 . Model structure diagram.

The mixup and text augmentation modules have been shown in the two vignettes above, and the linear layer in Figure 3 can also be referred to as the bottleneck layer. We have introduced this linear layer into the model for two reasons: on the one hand, this bottleneck layer allows for the scaling of the input and output sizes to 512, facilitating subsequent knowledge transfer and model training; on the other hand, the choice of a linear layer over an activation function prevents the non-linearity from corrupting too much information, while the activation function may filter out much useful information.

The addition of the bottleneck structure to the transformer layer causes the balance between the multi-head attention mechanism and the feedforward neural network to be broken, that is, the feature mapping of the input and output through the latter two will be different through the bottleneck layer, which can cause many problems. The overall structure of the multi-head attention mechanism allows the model to jointly process information in different subspaces, while the feedforward neural network increases the non-linearity of the model. In the original BERT model, the ratio of the multi-head attention mechanism to the feedforward neural network is 1:2. However, with the addition of the bottleneck structure, the input to the multi-head attention mechanism will come from a wider feature map, that is, the size of the modules between each other, while the input to the feedforward neural network will come from a narrower bottleneck, that is, the size of the modules. These two changes lead to the inclusion of more parameters in the multi-headed attention mechanism. To address this issue, stacked layers, that is, stacked feedforward neural networks, were introduced to rebalance the relative relationship between the multi-headed attention mechanism and the feedforward neural network. Size, and in this chapter, we use four stacked feedforward neural networks for balancing.

Data Set Introduction

This chapter uses two data sets to analyze the model from multiple perspectives. These data sets are AG News Corpus, The Stanford Sentiment Treebank (SST) data set, and AG News Corpus with 496,835 items in four categories and more than 2,000 News articles from a news source, the data set has title and description fields, and each category has 30,000 training samples and 1,900 test samples, of which we randomly select 10,000 samples for training. The SST data set is a sentiment analysis data set released by Stanford University. Its essence is a sentiment classification task, mainly for the sentiment classification of movie reviews. Therefore, SST belongs to single-sentence text classification. We selected SST-2 for the two classification task.

Model Settings

In order to compare the prediction effect of the model, we compared the results of multiple models on multiple tasks and compared the accuracy of multiple models on the classification task, in which a BERT-based-uncased tagger was used to tag the text, and use this model for text encoding. In order to reduce the parameters of the original BERT model, we designed the above model to change the effect of the student model by adjusting the ratio of the multi-head attention mechanism and the feedforward neural network, as shown by the experiments in Mobilebert. The model performance will peak when the ratio between the two is 0.4 ~ 0.6, so we choose a ratio of 0.5. In addition, the maximum sentence length is 256. For sentences that exceed the length, we keep the first 256 characters, the learning rate is set to 5e-5, and the batch size is 32. Due to the difference in the number of training samples, we set the saving step in the different numbers of samples. When the number of trains is the maximum value, please refer to Table 2 for detailed parameters.

Table 2 . Experimental parameter table.

We tried to compare a number of different classification models. The following are the model names and their brief introductions.

This model is one of the current mainstream models; we use the already trained BERT-based-uncased model and fine-tune it for classification tasks.

This model is a simplified version of the BERT model. It is transformed on the basis of the BERT model and reduces a lot of parameters, which greatly improves the speed of model training and model prediction, and the effect of the model will only be a slight decrease ( Lan et al., 2019 ).

This model is a compressed version of the BERT model ( Sun et al., 2020 ). It uses knowledge distillation to improve based on BERT. Its scale is reduced by 3–4 times and the speed is increased by 4–5 times. Compared with the original knowledge distillation model, the model prunes parameters while retaining most of the accuracy.

This chapter conducts comparative experiments on three data sets, namely, the SST data set and the AG news corpus. Based on these three data sets, experiments are conducted with different data scales to test the effectiveness of the model; the results are shown in Table 3 .

Table 3 . Performance [test accuracy(%)] comparison with baselines.

As can be seen from the results in the table, we have cut the data sets and cut different scales on the different corpus. It can be seen that when the amount of data is sufficient, BERT The results obtained by the model are the best, but since they are all transformer-based models, the difference between the final results is not large. It is easy to see that the method given in this paper can achieve better results in the case of a small number of samples. For example, in the case where the number of labeled labels in the AG News data set is less than 1,000, the method given in this paper has achieved relatively good results. The excellent results, especially when there are only 100 labeled data are greatly improved compared to the other methods. When MobileBERT does not introduce the mixing layer and the data enhancement layer, it can be seen from the above table that the results achieved by the same type of data set and the same data scale are lower than those achieved by other models.

Figures 4 , 5 can clearly show that when the sentiment analysis data set and the news data set are classified, especially when the number of labeled data is less than 100, our method has an improvement of about 20% compared with the other methods. This is due to the small amount of training data and the extreme demands of each transformer-based model for training data, both of which lead to problems, such as overfitting of the trained model, resulting in poor results. The method proposed in this paper is when the data scale is small, the method of text expansion is provided, which increases the scale of training data to a certain extent and thus has better results than other models. With the increase of data scale, the results of each model have a tendency to approach each other. When the data size is greater than 1,000, the difference between the results of each model is small. Similar to the principle of a small amount of training data, when the training data reach a certain size, the difference between the training results of each model is not large. The advantage of the method proposed in this paper is that our results are excellent when used in a small amount of labeled data, and also have good results in large-scale data.

Figure 4 . Performance [test accuracy (%)] on AG News.

Figure 5 . Performance [test accuracy (%)] on Stanford Sentiment Treebank (SST).

The purpose of text sentiment analysis is to analyze and study the large amount of comment information on the web and provide a relevant basis for the authorities to rectify the online environment. To alleviate the reliance on annotated data, a text sentiment classification method that combines data augmentation techniques and Transformer is proposed. Through experiments on two benchmark text classification data sets, the reliability and efficiency of text sentiment classification are fully weighed, and the reliability and efficiency of text sentiment classification are greatly improved, which also has great advantages compared with other state-of-the-art research. We plan to explore the effect on tasks outside the experimental data set, such as other real-world scenarios with limited labeled data.

Data Availability Statement

Publicly available data sets were analyzed in this study. These data can be found at: https://gluebenchmark.com/tasks .

Ethics Statement

The individual(s) provided their written informed consent for the publication of any identifiable images or data presented in this article.

Author Contributions

XG was responsible for designing the framework of the entire manuscript, from topic selection to solution to experimental verification. All authors contributed to the article and approved the submitted version.

This work was supported by the Humanities and Social Sciences Foundation of the Ministry of Education under Grant 18YJCZH229, and in part by the 13th Five-Year Plan Project of Educational Science in Jiangsu Province under Grant X-a/2018/10, and the National Natural Science Foundation of China (61972059).

Conflict of Interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher’s Note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Araci, D. (2019). Finbert: financial sentiment analysis with pre-trained language models. arXiv [Preprint]. doi: 10.48550/arXiv.1908.10063

CrossRef Full Text | Google Scholar

Budhkar, A., and Rudzicz, F. (2018). Augmenting word2vec with latent Dirichlet allocation within a clinical application. arXiv [Preprint]. doi: 10.48550/arXiv.1808.03967

Chawla, N. V., and Karakoulas, G. (2005). Learning from labeled and unlabeled data: an empirical study across techniques and domains. J. Artif. Intell. Res. 23, 331–366. doi: 10.1613/jair.1509

Chen, Y. (2015). Convolutional Neural Network for Sentence Classification. Master's thesis. University of Waterloo

Google Scholar

Cnnic (2021). The 45th Statistical Report on China’s Internet Development.

Conneau, A., Schwenk, H., Barrault, L., and Lecun, Y. (2016). Very deep convolutional networks for text classification. arXiv [Preprint]. doi: 10.48550/arXiv.1606.01781

Devlin, J., Chang, M. W., Lee, K., and Toutanova, K. (2018). Bert: pre-training of deep bidirectional transformers for language understanding. arXiv [Preprint]. doi: 10.48550/arXiv.1810.04805

Edunov, S., Ott, M., Auli, M., and Grangier, D. (2018). Understanding back-translation at scale. arXiv [Preprint]. doi: 10.48550/arXiv.1808.09381

Fellbaum, C. (2010). “WordNet,” in Theory and Applications of Ontology: Computer Applications . eds. R. Poli, M. Healy, and A. Kameas (Dordrecht: Springer), 231–243.

Gou, J., Yu, B., Maybank, S. J., and Tao, D. (2021). Knowledge distillation: a survey. Int. J. Comput. Vis. 129, 1789–1819. doi: 10.1007/s11263-021-01453-z

Guo, H. (2020). “Nonlinear mixup: Out-of-Manifold Data Augmentation for Text Classification.” in Proceedings of the AAAI Conference on Artificial Intelligence ; April 03, 2020 Vol. 34, 4044–4051.

Han, K., Xiao, A., Wu, E., Guo, J., Xu, C., and Wang, Y. (2021). “Transformer in transformer,” in Advances in Neural Information Processing Systems . eds. M. Ranzato, A. Beygelzimer, Y. Dauphin, P. S. Liang and J. Wortman Vaughan (MIT Press), 34.

Hawkins, D. M. (2004). The problem of overfitting. J. Chem. Inf. Comput. Sci. 44, 1–12. doi: 10.1021/ci0342472

PubMed Abstract | CrossRef Full Text | Google Scholar

Hinton, G., Vinyals, O., and Dean, J. (2015). Distilling the knowledge in a neural network. arXiv [Preprint]. doi: 10.48550/arXiv.1503.02531

Howard, J., and Ruder, S. (2018). Universal language model fine-tuning for text classification. arXiv [Preprint]. doi: 10.48550/arXiv.1801.06146

Jiao, X., Yin, Y., Shang, L., Jiang, X., Chen, X., Li, L., et al. (2019). Tinybert: Distilling bert for natural language understanding. arXiv [Preprint]. doi: 10.48550/arXiv.1909.10351

Lan, Z., Chen, M., Goodman, S., Gimpel, K., Sharma, P., and Soricut, R. (2019). Albert: A lite bert for self-supervised learning of language representations. arXiv [Preprint]. doi: 10.48550/arXiv.1909.11942

LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature 521, 436–444. doi: 10.1038/nature14539

Lin, P., and Luo, X. (2020). “A Survey of Sentiment Analysis Based on Machine Learning.” in CCF International Conference on Natural Language Processing and Chinese Computing . Springer, Cham. 372–387.

Rezaeinia, S. M., Rahmani, R., Ghodsi, A., and Veisi, H. (2019). Sentiment analysis based on improved pre-trained word embeddings. Expert Syst. Appl. 117, 139–147. doi: 10.1016/j.eswa.2018.08.044

Sun, Z., Yu, H., Song, X., Liu, R., Yang, Y., and Zhou, D. (2020). Mobilebert: a compact task-agnostic bert for resource-limited devices. arXiv [Preprint]. doi: 10.48550/arXiv.2004.02984

Tiwari, D., and Mallick, B. (2016). SVM and naïve bayes network traffic classification using correlation information. Int. J. Comput. Appl. 147, 1–5. doi: 10.5120/ijca2016911010

Van Engelen, J. E., and Hoos, H. H. (2020). A survey on semi-supervised learning. Mach. Learn. 109, 373–440. doi: 10.1007/s10994-019-05855-6

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., et al. (2017). Attention is all you need. Adv. Neural Inf. Proces. Syst. 30:15. doi: 10.48550/arXiv.1706.03762

Wei, J., and Zou, K. (2019). Eda: easy data augmentation techniques for boosting performance on text classification tasks. arXiv [Preprint]. doi: 10.48550/arXiv.1901.11196

Yan, D., Li, K., Gu, S., and Yang, L. (2020). Network-based bag-of-words model for text classification. IEEE Access 8, 82641–82652. doi: 10.1109/ACCESS.2020.2991074

Zhang, H., Cisse, M., Dauphin, Y. N., and Lopez-Paz, D. (2017). mixup: Beyond empirical risk minimization. arXiv [Preprint]. doi: 10.48550/arXiv.1710.09412

Zhang, L., Wei, J., and Boncella, R. J. (2020). Emotional communication analysis of emergency microblog based on the evolution life cycle of public opinion. Inf. Discov. Deliv. 48, 151–163. doi: 10.1108/IDD-10-2019-0074

Keywords: sentiment analysis, social media, transformer, knowledge distillation, text augmentation

Citation: Gong X, Ying W, Zhong S and Gong S (2022) Text Sentiment Analysis Based on Transformer and Augmentation. Front. Psychol . 13:906061. doi: 10.3389/fpsyg.2022.906061

Received: 28 March 2022; Accepted: 25 April 2022; Published: 13 May 2022.

Reviewed by:

Copyright © 2022 Gong, Ying, Zhong and Gong. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY) . The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.

*Correspondence: Wenhao Ying, [email protected]

Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.

S em E val 2022 Task 10: Structured Sentiment Analysis

Jeremy Barnes , Laura Oberlaender , Enrica Troiano , Andrey Kutuzov , Jan Buchmann , Rodrigo Agerri , Lilja Øvrelid , Erik Velldal

Export citation

• Preformatted

Markdown (Informal)

[SemEval 2022 Task 10: Structured Sentiment Analysis](https://aclanthology.org/2022.semeval-1.180) (Barnes et al., SemEval 2022)

• SemEval 2022 Task 10: Structured Sentiment Analysis (Barnes et al., SemEval 2022)
• Jeremy Barnes, Laura Oberlaender, Enrica Troiano, Andrey Kutuzov, Jan Buchmann, Rodrigo Agerri, Lilja Øvrelid, and Erik Velldal. 2022. SemEval 2022 Task 10: Structured Sentiment Analysis . In Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022) , pages 1280–1295, Seattle, United States. Association for Computational Linguistics.

A Survey on Sentiment Analysis

Ieee account.

• Change Username/Password
• Update Address

Purchase Details

• Payment Options
• Order History
• View Purchased Documents

Profile Information

• Communications Preferences
• Profession and Education
• Technical Interests
• US & Canada: +1 800 678 4333
• Worldwide: +1 732 981 0060
• Contact & Support
• About IEEE Xplore
• Accessibility
• Terms of Use
• Nondiscrimination Policy
• Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

• Publications
• Account settings

Preview improvements coming to the PMC website in October 2024. Learn More or Try it out now .

• Advanced Search
• Journal List
• Springer Nature - PMC COVID-19 Collection

Survey on sentiment analysis: evolution of research methods and topics

Jingfeng cui.

1 Institute of High Performance Computing, A*STAR, 1 Fusionopolis Way, Singapore, 138632 Singapore

2 School of Information Management, Nanjing Agricultural University, 1 Weigang, Nanjing, 210095 China

Zhaoxia Wang

3 School of Computing and Information Systems, Singapore Management University, 80 Stamford Rd, Singapore, 178902 Singapore

Seng-Beng Ho

Erik cambria.

4 School of Computer Science and Engineering, Nanyang Technological University, 50 Nanyang Avenue, Singapore, 639798 Singapore

Associated Data

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Sentiment analysis, one of the research hotspots in the natural language processing field, has attracted the attention of researchers, and research papers on the field are increasingly published. Many literature reviews on sentiment analysis involving techniques, methods, and applications have been produced using different survey methodologies and tools, but there has not been a survey dedicated to the evolution of research methods and topics of sentiment analysis. There have also been few survey works leveraging keyword co-occurrence on sentiment analysis. Therefore, this study presents a survey of sentiment analysis focusing on the evolution of research methods and topics. It incorporates keyword co-occurrence analysis with a community detection algorithm. This survey not only compares and analyzes the connections between research methods and topics over the past two decades but also uncovers the hotspots and trends over time, thus providing guidance for researchers. Furthermore, this paper presents broad practical insights into the methods and topics of sentiment analysis, while also identifying technical directions, limitations, and future work.

Introduction

Web 2.0 has driven the proliferation of user-generated content on the Internet. This content is closely related to the lives, emotions, and opinions of users. Therefore, analysis of this user-generated data is beneficial for monitoring public opinion and assisting in making decisions. Sentiment analysis, as one of the most popular applications of text-based analytics, can be used to mine people’s attitudes, emotions, appraisals, and opinions about issues, entities, topics, events, and products (Cambria et al. 2022a , b , c , d ; Injadat et al. 2016 ; Jiang et al. 2017 ; Liang et al. 2022 ; Oueslati et al. 2020 ; Piryani et al. 2017 ). Sentiment analysis can help us interpret emotions in unstructured texts as positive, negative, or neutral, and even calculate how strong or weak the emotions are. Today, sentiment analysis is widely used in various fields, such as business, finance, politics, education, and services. This analytical technique has gained broad acceptance not only among researchers but also among governments, institutions, and companies (Khatua et al. 2020 ; Liu et al. 2012 ; Sánchez-Rada and Iglesias 2019 ; Wang et al. 2020b ). It helps policy leaders, businessmen, and service people make better decisions.

The majority of user-generated content data is unstructured text, which increases the great difficulty of sentiment analysis. Since 2000, researchers have been exploring techniques and methods to enhance the accuracy of such analysis. The popularity of social media platforms has brought people around the world closer together. With the continuous advancement of technology, the research topics, application fields, and core methods and technologies of sentiment analysis are also constantly changing.

Comparing and analyzing papers from specific disciplines can help researchers gain a comprehensive understanding of the field. There have been many surveys on sentiment analysis (Nair et al. 2019 ; Obiedat et al. 2021 ; Raghuvanshi and Patil 2016 ). However, there is a lack of adequate discussion on the connections between research methods and topics in the field, as well as on their evolution over time. In 1983, Callon et al. proposed co-word analysis (Callon et al. 1983 ). It can effectively reflect the correlation strength of information items in text data. Co-word analysis based on the frequency of co-occurrence of keywords used to describe papers can reveal the core contents of the research in specific fields. An evolutionary analysis of the associations between core contents is helpful for a comprehensive understanding of the research hotspots and frontiers in the field (Deng et al. 2021 ). It can provide guidance for researchers, especially those who are new to the field, and help them determine research directions, avoid repetitive research, and better discover and grasp the research trends in this field (Wang et al. 2012 ). To fill in the gap in existing research, we conduct keyword co-occurrence analysis and evolution analysis with informetric tools to explore the research hotspots and trends of sentiment analysis.

The main contributions of this survey are as follows:

• Using keyword co-occurrence analysis and the informetric tools, the paper presents a survey on sentiment analysis, explores and discovers useful information.
• A keyword co-occurrence network is constructed by combining the paper title, abstract, and author keywords. Through the keyword co-occurrence network and community detection algorithm, the research methods and topics in the field of sentiment analysis, along with their evolution in the past two decades, are discussed.
• The paper summarizes the research hotspots and trends in sentiment analysis. It also highlights practical implications and technical directions.

The remainder of this paper is organized as follows: In Sect.  2 , we summarize and analyze the existing surveys on sentiment analysis and present the research purpose and methodologies of this paper. Section  3 details the survey methodology, including the collection and processing of scientific publications, visualization, and analysis using different methods and tools. In Sect.  4 , we analyze the results obtained from the keyword co-occurrence analysis and evolution analysis, along with the research hotspots and trends in sentiment analysis identified through the analysis results. Finally, in Sect.  5 , we summarize the research conclusions as well as the practical implications and technical directions of sentiment analysis. We also clarify the limitations of this paper and make suggestions for future work.

Existing surveys on sentiment analysis

Sentiment analysis is a concept encompassing many tasks, such as sentiment extraction, sentiment classification, opinion summarization, review analysis, sarcasm detection or emotion detection, etc. Since the 2000s, sentiment analysis has become a popular research field in natural language processing (Hussein 2018 ). In the existing surveys, the researchers mainly conducted specific analyses of the tasks, technologies, methods, analysis granularity, and application fields involved in the sentiment analysis process.

Surveys on contents and topics of sentiment analysis

When research on sentiment analysis was still in its infancy, the contents and topics of surveys mainly focused on sentiment analysis tasks, analysis granularity, and application areas. Kumer et al. reviewed the basic terms, tasks, and levels of granularity related to sentiment analysis (Kumar and Sebastian 2012 ). They also discussed some key feature selection techniques and the applications of sentiment analysis in business, politics, recommender systems and other fields. Nassirtoussi et al. explored the application of sentiment analysis in market prediction (Nassirtoussi et al. 2014 ). Medhat et al. analyzed the improvement of the algorithms proposed in 2010–2013 and their application fields (Medhat et al. 2014 ). Ravi et al. analyzed the papers related to opinion mining and sentiment analysis from 2002 to 2015. Their study mainly discussed the necessary tasks, methods, applications, and unsolved problems in the field of sentiment analysis (Ravi and Ravi 2015 ).

Existing surveys of the applications of sentiment analysis have focused more on the domains of market research, medicine, and social media in recent years. Rambocas et al. examined the application of sentiment analysis in marketing research from three main perspectives, including the unit of analysis, sampling design, and methods used in sentiment detection and statistical analysis (Rambocas and Pacheco 2018 ). Cheng et al. summarized techniques based on semantic, sentiment, and event extraction, as well as hybrid methods employed in stock forecasting (Cheng et al. 2022 ). Yue et al. categorized and compared a large number of techniques and approaches in the social media domain. That study also introduced different types of data and advanced research tools, and discussed their limitations (Yue et al. 2019 ). In the context of the COVID-19 epidemic, Alamoodi et al. reviewed and analyzed articles on the occurrence of different types of infectious diseases in the past 10 years. They reviewed the applications of sentiment analysis from the identified 28 articles, summarizing the adopted techniques such as dictionary-based models, machine learning models, and mixed models (Alamoodi et al. 2021b ); Alamoodi et al. also conducted a review of the applications of sentiment analysis for vaccine hesitancy (Alamoodi et al. 2021a ). Researchers also reviewed the application of sentiment analysis in the fields of election prediction (Brito et al. 2021 ), education (Kastrati et al. 2021 ; Zhou and Ye 2020 ) and service industries (Adak et al. 2022 ).

Quite a number of research works investigated sentiment analysis works in non-English languages. Sentiment analysis in Chinese (Peng et al. 2017 ), Arabic (Al-Ayyoub et al. 2019 ; Boudad et al. 2018 ; Nassif et al. 2021 ; Oueslati et al. 2020 ), Urdu (Khattak et al. 2021 ), Spanish (Angel et al. 2021 ), and Portuguese (Pereira 2021 ) were conducted. They mainly reviewed the classification frameworks of the sentiment analysis process, supported language resources (dictionaries, natural language processing tools, corpora, ontologies, etc.), and deep learning models used (CNN, RNN, and transfer learning) for each of the languages involved.

Surveys on methods of sentiment analysis

Before machine learning technology became mature, researchers were particularly concerned about feature extraction methods. For example, Feldman summarized methods for extracting preferred entities from indirect opinions and methods for dictionary acquisition (Feldman 2013 ). Asghar et al. reviewed the natural language processing techniques for extracting features based on part of speech and term position; statistical techniques for extracting features based on word frequency and decision tree model; and techniques for combining part of speech tagging, syntactic feature analysis, and dictionaries (Asghar et al. 2014 ). Koto et al. discussed the best features for Twitter sentiment analysis prior to 2014 by comparing 9 feature sets (Koto and Adriani 2015 ). They found that the current best features for sentiment analysis of Twitter texts are AFINN (a list of English terms used for sentiment analysis manually rated by Finn Årup Nielsen) (Nielsen 2011 ) and Senti-Strength (Thelwall et al. 2012 ). Taboada sorted out the characteristics of words, phrases, and sentence patterns in sentiment analysis from the perspective of linguistics (Taboada 2016 ). Besides, Schouten and Frasinar conducted a comprehensive and in-depth critical evaluation of 15 sentiment analysis web tools (Schouten and Frasincar 2015 ). Medhat et al. ( 2014 ) and Ravi et al. (Ravi and Ravi 2015 ) also analyzed the early algorithms for sentiment analysis.

In the study by Schouten et al., the authors focused on aspect-level sentiment analysis, combing the techniques of aspect-level sentiment analysis before 2014, such as frequency-based, syntax-based, supervised machine learning, unsupervised machine learning, and hybrid approaches. They concluded that the latest technology was moving beyond the early stages (Schouten and Frasincar 2015 ). As research into sentiment analysis became more and more popular and there was important progress made in the development of deep learning technologies, researchers started to pay more attention to the techniques and methods of sentiment analysis. Deep learning methods in particular became the focus of discussions among researchers.

Prabha et al. analyzed various deep learning methods used in different applications at the level of sentence and aspect/object sentiment analysis, including Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long Short-term Memory (LSTM) (Prabha and Srikanth 2019 ). They discussed the advantages and disadvantages of these methods and their performance parameters. Ain et al. introduced deep learning techniques such as Deep Neural Network (DNN), CNN and Deep Belief Network (DBN) to solve sentiment analysis tasks like sentiment classification, cross-lingual problems, and product review analysis (Ain et al. 2017 ). Zhang et al. investigated deep learning and machine learning techniques for sentiment analysis in the contexts of aspect extraction and categorization, opinion expression extraction, opinion holder extraction, sarcasm analysis, multimodal data, etc. (Zhang et al. 2018 ). Habimana et al. compared the performance of deep learning methods on specific datasets and proposed that performance could be improved using models including Bidirectional Encoder Representations from Transformers (BERT), sentiment-specific word embedding models, cognitive-based attention models, and commonsense knowledge (Habimana et al. 2020 ). Wang et al. reviewed and discussed existing analytical models for sentiment classification and proposed a computational emotion-sensing model (Wang et al. 2020b ).

Some researchers also discussed web tools (Zucco et al. 2020 ), fuzzy logic algorithms (Serrano-Guerrero et al. 2021 ), transformer models (Acheampong et al. 2021 ), and sequential transfer learning (Chan et al. 2022 ) for sentiment analysis.

Overall survey methodology

With the increase in the popularity of sentiment analysis research, more related research results began to accumulate. Researchers needed to systematically organize and analyze results from a large number of publications to perform literature reviews. They used different survey methodologies to conduct surveys of a large number of papers.

Content analysis is a powerful approach to characterizing the contents of each study by carefully reading its content and manually identifying, coding, and organizing key information in it. A literature review is formed as a result of the repeated use of this approach (Elo and Kyngäs 2008 ; Stemler 2000 ). Content analysis has been used for different studies and systematic reviews (Qazi et al. 2015 , 2017 ). For example, Birjali et al. have studied the most commonly used classification techniques in sentiment analysis from a large amount of literature and introduced the application areas and sentiment classification processes, including preprocessing and feature selection (Birjali et al. 2021 ). They conducted a comprehensive analysis of the papers, discovering that supervised machine learning algorithms are the most commonly used techniques in the field. A complete review of methods and evaluation for sentiment analysis tasks and their applications was conducted by Wankhade et al. ( 2022 ). They compared the strengths and weaknesses of the methods, and discussed the future challenges of sentiment analysis in terms of both the methods and the forms of the data. Although this method can review the research contents and penetrate into the cores of the papers most systematically, it requires a considerable amount of manpower and time for in-depth literature reading.

The systematic literature review guideline proposed by Kitchenham and Charters has gradually attracted the attention of researchers (Kitchenham 2004 ; Kitchenham and Charters 2007 ; Sarsam et al. 2020 ). This review process is divided into six stages: research question definition, search strategy formulation, inclusion and exclusion criteria definition, quality assessment, data extraction, and data synthesis. Researchers can eliminate a large number of retrieved papers by using this standard process and finally conducting further analysis and research on a small number of papers. Kumar et al. reviewed context-based sentiment analysis in social multimedia between 2006 and 2018. From the 573 papers retrieved in the initial search, they finally selected 37 papers to use in discussing sentiment analysis techniques (Kumar and Garg 2020 ). This approach was also used by Kumar et al. in their research on sentiment analysis on Twitter using soft computing techniques. They selected 60 articles out of 502 for follow-up analysis (Kumar and Jaiswal 2020 ). Zunic et al. selected 86 papers from 299 papers retrieved in the period 2011–2019 to discuss the application of sentiment analysis techniques in the field of health and well-being (Zunic et al. 2020 ); Ligthart et al. followed Kitchenham’s guideline and identified 14 secondary studies. They provided an overview of specific sentiment analysis tasks and of the features and methods required for different tasks (Ligthart et al. 2021 ). Obiedat (Obiedat et al. 2021 ), Angel (Angel et al. 2021 ) and Lin (Lin et al. 2022 ) also all followed this guideline to select literature for further analysis. This method can reduce the amount of literature that requires in-depth reading, but in the case of a large amount of literature, more effort is still required to search and screen the material than in traditional literature review methods (Kitchenham and Charters 2007 ).

There are also a few authors who have used informetric methods to review papers. Piryani et al. conducted an informetric analysis of research on opinion mining and sentiment analysis from 2000 to 2015 (Piryani et al. 2017 ). The authors used social network analysis, literature co-citation analysis, and other methods in the paper. They analyzed publication growth rates; the most productive countries, institutions, journals, and authors; and topic density maps and keyword bursts, among other elements. To a certain extent, they interpreted core authors, core papers, areas of research focus in this field, and the current state of national cooperation. In order to explore the application of sentiment analysis in building smart societies, Verma collected 353 papers published between 2010 and 2021 (Verma 2022 ). Using a topic analysis perspective combined with the Louvain algorithm, the author identified four sub-topics in the research field. Similarly, Mantyla et al. employed LDA techniques and manual classification to explore the topic structures of sentiment analysis articles (Mäntylä et al. 2018 ). The informetric methods use natural language processing technologies to intuitively conduct topic mining and analysis of a large number of papers. Through topic clustering, the literature is organized and analyzed, which reduces the time researchers spend on reading the literature in depth. These methods are suitable for exploring research topics and trends in the field.

Summary of advantages and disadvantages of the existing surveys

In the following, we discuss the advantages and disadvantages of the existing surveys from a number of different points of view.

From the point of view of the contents and topics of sentiment analysis

As summarized in Table ​ Table1, 1 , the researchers organized the literature and conducted depth investigations of the contents and topics of sentiment analysis. They reviewed the tasks of sentiment analysis (e.g., different text granularity, opinion mining, spam review detection, and emotion detection), the application areas of sentiment analysis (e.g. market, medicine, social media, and election prediction), and different languages for sentiment analysis, such as Chinese, Spanish, and Arabic (Adak et al. 2022 ; Al-Ayyoub et al. 2019 ; Alamoodi et al. ( 2021a , b ); Alonso et al. 2021 ; Angel et al. 2021 ; Boudad et al. 2018 ; Brito et al. 2021 ; Cheng et al. 2022 ; Hussain et al. 2019 ; Kastrati et al. 2021 ; Khattak et al. 2021 ; Koto and Adriani 2015 ; Kumar and Sebastian 2012 ; Ligthart et al. 2021 ; Medhat et al. 2014 ; Nassif et al. 2021 ; Nassirtoussi et al. 2014 ; Oueslati et al. 2020 ; Peng et al. 2017 ; Pereira 2021 ; Rambocas and Pacheco 2018 ; Ravi and Ravi 2015 ; Schouten and Frasincar 2015 ; Sharma and Jain 2020 ; Yue et al. 2019 ; Zhou and Ye 2020 ). They summarized the methods and application prospects of sentiment analysis under different contents and topics. As the field has grown, new topics have emerged, and knowledge from other fields has been gradually integrated into it. In recent years, the popularity of social media has aroused increasing interest in sentiment analysis research, and the number of papers published, especially those related to different topics of sentiment analysis, has grown rapidly. However, the existing surveys cover a short time range, and there has not been a survey dedicated to the evolution of research contents or topics of sentiment analysis. There have also been few survey works analyzing the connections between topics and methods, or their evolution (e.g., how the contents and topics of sentiment analysis have changed over time).

Advantages and disadvantages of the existing surveys

From the point of view of the methods of sentiment analysis

Some researchers reviewed different techniques and methods of sentiment analysis in different application areas and tasks. They analyzed and discussed sentiment analysis methods based on lexicons, rules, part of speech, term position, statistical techniques, supervised and unsupervised machine learning methods, as well as deep learning methods like LSTM, CNN, RNN, DNN, DBN, BERT, and other hybrid approaches (Acheampong et al. 2021 ; Ain et al. 2017 ; Alamoodi et al. 2021b ; Asghar et al. 2014 ; Chan et al. 2022 ; Cheng et al. 2022 ; Feldman 2013 ; Habimana et al. 2020 ; Koto and Adriani 2015 ; Kumar, Akshi and Sebastian 2012 ; Medhat et al. 2014 ; Prabha and Srikanth 2019 ; Ravi and Ravi 2015 ; Schouten and Frasincar 2015 ; Serrano-Guerrero et al. 2021 ; Taboada 2016 ; Wang et al. 2020b ; Yue et al. 2019 ; Zhang et al. 2018 ; Zucco et al. 2020 ). These researchers also compared the advantages and disadvantages of each method. As summarized in Table ​ Table1, 1 , even though existing surveys analyze the techniques and methods of sentiment analysis, providing good insights, there has not been a survey that analyzes the evolution of research methods over time. There have also been few survey works that focuses on the connections between topics and methods of sentiment analysis, and their evolution over time.

From the point of view of the overall survey methodology

The survey methods used have mainly been the content analysis method, Kitchenham and Charters' guideline, and the informetric methods. As summarized in Table ​ Table1, 1 , the content analysis method can effectively analyze the contents of research papers in depth, but it does not address the issue of the evolution of the research methods and topics (Bengtsson 2016 ; Birjali et al. 2021 ; Elo and Kyngäs 2008 ; Krippendorff 2018 ; Qazi et al. 2015 , 2017 ; Wankhade et al. 2022 ). Although the number of papers that need to be read in depth can be reduced by following Kitchenham and Charters' guideline, more effort is needed to search and screen literature than in traditional literature review methods (Angel et al. 2021 ; Kitchenham 2004 ; Kitchenham and Charters 2007 ; Kumar and Garg 2020 ; Ligthart et al. 2021 ; Lin et al. 2022 ; Obiedat et al. 2021 ; Sarsam et al. 2020 ; Zunic et al. 2020 ). The informetric methods are best suited to investigating the research methods and topics of sentiment analysis (Bar-Ilan 2008 ; Mäntylä et al. 2018 ; Piryani et al. 2017 ; Santos et al. 2019 ; Verma 2022 ). There are three surveys using informetric techniques and tools that are well suited for analysis of a large number of papers over many years (Mäntylä et al. 2018 ; Piryani et al. 2017 ; Verma 2022 ). However, the evolution of research methods and topics of sentiment analysis over time has not been studied with informetric methods. There have also been few survey works that leverages keyword co-occurrence analysis and community detection to analyze the connections between research methods and topics, and their evolution over time.

Therefore, to address the gaps in the existing surveys, this study presents a survey on the research methods and topics, and their evolution over time. It combines keyword co-occurrence analysis and informetric analysis tools to reveal the methods and topics of sentiment analysis and their evolution in this field from 2002 to 2022.

The following section, Sect.  3 , describes our proposed survey methodology in detail.

The proposed survey methodology

This section describes our proposed survey methodology, including collection of scientific publications, processing of scientific publications, as well as visualization and analysis using different methods and tools. The overall scheme of this survey (Fig.  2 ) is also presented in the end of Sect.  3 to better visualize and summarize the proposed survey methodology in this research.

Graphical representation of the overall scheme of this survey. Module A: Collection of scientific publications; Module B: Processing of scientific publications; Module C: Visualization and analysis using different methods and tools; Module D: Result analysis and discussions considering various aspects

Collection of scientific publications

We collected research data from the Web of Science platform. We used keywords such as "sentiment analysis," "sentiment mining," and "sentiment classification" to search for relevant papers as data samples. In examining the retrieved papers, we found that some paper topics, paper types, and publication journals were not related to sentiment analysis, so we excluded them. The papers we included were mainly related to the sentiment analysis of texts. We excluded papers on sentiment analysis related to image processing, video processing, speech processing, biological signal processing, etc. Therefore, the retrieval strategy was as follows:

Topic Search (TS) = ("sentiment analy*" or "sentiment mining" or "sentiment classification") And Abstract (AB) = "sentiment" NOT TS = ("face image*" or "speech recognition" or "speech emotion" or "physiological signal*" or "music emotion*" or "facial feature extraction" or "video emotion" or "electroencephalography " or "biosignal*" or "image process*") NOT Title = ("facial" or "speech" or "sound*" or "face" or "dance" or "temperature" or "image*" or "spoken" or "electroencephalography" or "EEG" or "biosignal*" or "voice*" not AB = "facial."

The results in conferences are given the same relevance as journal papers. We chose four databases in the Web of Science: two conference citation databases (Conference Proceedings Citation Index—Social Sciences & Humanities [CPCI-SSH], and Conference Proceedings Citation Index—Science [CPCI-S]), and two journal citation databases (Science Citation Index Expanded [SCI-Expanded] and Social Sciences Citation Index [SSCI]). Given the various forms of words such as "analyzing" and "analysis," a truncated search technique (marked with an asterisk) was used to prevent the omission of relevant papers. The time frame of the retrieved papers was from January 2002 to January 2022, and the publication types of the papers included "article," "conference paper," "review," and "edited material." A total of 9,714 papers were obtained from the four databases above. These included 3,809 articles, 5,633 proceeding papers, 267 reviews, and 5 pieces of editorial material from 2002 to 2022. Overall, there were 104 papers from January 2022. The number of papers each year from 2002 to 2021 is shown in Fig.  1 .

The number of papers each year from 2002 to 2021

Processing of scientific publications

In this process, our purpose was to extract the key contents of the papers, which are used to analyze the research methods and topics in the field of sentiment analysis. Due to their limited number, the author keywords in each paper often cannot fully represent the key content of the paper. We found that combining the title and abstract could better reflect the core information. Therefore, we synthesized the title, abstract, and author keywords of each paper to extract keywords that represented the main research method and topic of the paper involved using KeyBERT 1 . KeyBERT is a keyword extraction technique that uses BERT embedding to create keywords and key phrases that most closely resemble document content (Grootendorst and Warmerdam 2021 ). The specific keyword extraction process was as follows:

First, we used KeyBERT to extract 8 keywords and eliminated keywords with a weight lower than 0.3. We then combined the extracted keywords with the author keywords and removed duplicates. After that, we standardized the whole collection of keywords and merged synonyms. Finally, we counted the number of keywords and removed meaningless terms like "sentiment analysis," "sentiment classification," and "sentiment mining."

After statistical analysis, we obtained 41,827 keywords with a total word frequency of 88,104. As there were 9,714 papers and 41,827 keywords, we found that most of the keywords with word frequency below 10 were not representative of the research contents of sentiment analysis. As a result, a total of 685 representative keywords were reserved for subsequent analysis. These keywords appeared a total of 30,801 times. Table ​ Table2 2 shows the keywords with word frequency in the top 50.

Keywords with word frequency in the top 50

High-frequency keywords generally represent research hotspots. We therefore extracted high-frequency keywords to serve as the basis for the subsequent analysis. We found that most of the keywords with word frequency 18 and lower, such as "ranking," "mask," "experience," "affect," "online forum," and so on, were not relevant to sentiment analysis. Therefore, the keywords with a word frequency higher than 18 were reserved for analysis. These keywords appeared 25,429 times in the collected data, accounting for close to 83% of all the keywords. We obtained 275 keywords, which were used to analyze the main methods and topics of sentiment analysis.

Visualization and analysis using different methods and tools

Analytical methods.

Keywords are the core natural language vocabulary to express the subject, content, ideas, and research methods of the literature (You et al. 2021 ). Keywords represent the topics of the domain, and cluster analysis of these words can reflect the structure and association of topics. Keyword co-occurrence analysis counts the number of occurrences of a set of keywords in the same document. The strength and number of associations between research contents can be obtained through keyword co-occurrence analysis. Dividing research methods and topics into sub-communities helps researchers to analyze hotspots and trends in methods and topics, as well as to obtain sub-fields of sentiment analysis research (Ding et al. 2001 ).

Visualization and analysis tools

BibExcel 2 is a software tool for analyzing bibliographic data or any text-based data formatted in a similar way (Persson 2017 ). The tool generates structured data files that can be read by Excel for subsequent processing (Persson et al. 2009 ). Our processing steps are as follows. First, we imported the standardized bibliographic data into BibExcel. This tool can help structure the data. Second, we checked and corrected the data and used BibExcel to count the number of co-occurrences of keywords.

We then used Pajek 3 software to visualize the keyword co-occurrence network and divided the sub-communities. Pajek is a large and complex network analysis tool (Batagelj and Andrej 2022 ; Batagelj and Mrvar 1998 ). It can calculate certain indicators to reveal the state and properties of the network involved. In addition, Pajek’s Louvain community detection algorithm can help divide the keyword co-occurrence network into sub-communities, which represent sub-fields of sentiment analysis (Blondel et al. 2008 ; Leydesdorff et al. 2014 ; Rotta and Noack 2011 ). The Louvain community-detection algorithm unfolds a complete hierarchical community structure for the network. It has an advantage in subdividing different areas of study: multiple knowledge structures and details can be shown in one network (Deng et al. 2021 ).

After that, we applied VOSviewer 4 to optimize the visualization of sub-communities (Van Eck and Waltman 2010 ; VOSviewer 2021 ; Perianes-Rodriguez et al. 2016 ; Waltman and Van Eck 2013 ; Waltman et al. 2010 ). VOSviewer can help display the core keywords in each sub-community and the correlation between keywords. It can also reflect the closeness of the association between sub-communities. Finally, we used Excel to count the frequency of keywords for each year and to map the evolution of research methods and topics in the field of sentiment analysis.

Graphical representation of the overall scheme of this survey

This paper proposes and conducts a new research survey on sentiment analysis. The graphical representation of the overall scheme of this survey is shown in Fig.  2 . The main scheme includes four modules: Module A, Collection of scientific publications; Module B, Processing of scientific publications; Module C, Visualization and analysis through different methods and tools, and Module D, Result analysis and discussions based on various aspects.

In Module A, scientific publications are collected from the Web of Science (WOS) platform, as has been detailed in Sect.  3.1 Collection of scientific publications above. Module B, Processing of scientific publications, has been detailed in Sect.  3.2 above. It performs a data processing procedure to obtain key information, which includes all the representative keywords and high-frequency keywords. The title, abstract and keywords of the papers are used to extract such key information using KeyBERT (Grootendorst and Warmerdam 2021 ). Such key information is analyzed and visualized through different methods, including different visualization tools, as introduced in Sect.  3.3 (Module C), Visualization and analysis using different methods and tools, above.

In Module C, the number of co-occurrences of keywords is obtained using BibExcel (Persson 2017 ), the co-occurrences of keywords are analyzed and visualized using Pajek (Blondel et al. 2008 ; Leydesdorff et al. 2014 ; Rotta and Noack 2011 ) and VOSviewer (Van Eck and Waltman 2010 ; VOSviewer 2021 ; Perianes-Rodriguez et al. 2016 ; Waltman and Van Eck 2013 ; Waltman et al. 2010 ). The keyword community network and the keyword community evolution are analyzed and visualized using these tools, as described in Sect.  3.3 (Module C), Visualization and analysis using different methods and tools. According to the visualization and analysis results obtained in Module C, Module D, Result analysis and discussions, will be detailed in Sect.  4 .

In the following section, Sect.  4 (Module D), results are analyzed and discussed considering various aspects, including the research methods and topics of sentiment analysis in each community, the evolution of research methods and topics along with the research hotspots and trends over time.

Results and analysis through various aspects

Research methods and topics of sentiment analysis, overall characteristic analysis.

The high-frequency keywords were presented in Table ​ Table2. 2 . These keywords can be regarded as the main research contents in the field of sentiment analysis. "Twitter" ranks at the top. It is followed by "opinion mining," "natural language processing," "machine learning," and so on. The high-frequency keywords cover the topics of the studies, the contents of the studies, and the techniques and methods used. Based on these keywords, we used Pajek’s Louvain method to construct a keyword co-occurrence network to represent the research methods and topics as shown in Fig.  3 . The keyword co-occurrence network is divided into six communities. The research methods and topics of the six communities include social media platforms (C1), machine learning methods (C2), natural language processing and deep learning methods (C3), opinion mining and text mining (C4), Arabic sentiment analysis (C5), and others, such as domain sentiment analysis and transfer learning, etc. (C6).

Keyword community network

In Fig.  3 , the size of the node represents the number of keywords. The thickness of the line between the nodes represents the number of collaborations between keywords. The top 20 keywords in each community are sorted in descending order, as shown in Table ​ Table3. 3 . The keyword co-occurrence network features of the six sub-communities are described in Table ​ Table4. 4 . The number of nodes shows the number of keywords in each community, and the number of links shows the correlations between the keywords.

The top 20 keywords in each community

Global network characteristics of sub-communities

As shown in Table ​ Table4, 4 , we can see from the number of links between sub-communities that there is a strong correlation between them, especially the link between C3 and C4, which has 1306 lines. The reason may be that the research methods of C4 focus on "opinion mining" and "text mining," while those of C3 focus on "natural language processing" and "deep learning," and C3 provides more technical support for C4 research. In C5 and C6, the research methods and topics are scattered. Their internal links are also low, but the connections with C3 and C4 are relatively high. The contents of C5 and C6 may include some emerging research methods and topics. We will present a specific analysis on the methods and topics of each sub-community in the next subsection.

Analysis on research methods and topics of sub-communities

Analysis on research methods and topics of the c1 community.

Figure  4 shows the keyword co-occurrence network of the C1 community. The research methods and topics of the C1 community focus on three areas: "social media," "topic models," and "covid-19." In the context of big data, web 2.0 technology provides users with a way to express reviews and opinions of services, events, and people. Various social media platforms, such as Twitter, YouTube, and Weibo, have a large amount of users’ emotional data (Momtazi 2012 ). Compared to traditional news media, information on social media spreads more quickly, and people are able to express their feelings more freely. It is important to analyze the emotions generated by the information shared and published on social media (Abdullah and Zolkepli 2017 ; Wang et al. 2014 ). Researchers have been extracting text data from social media platforms for years to detect unexpected events (Bai and Yu 2016 ; Preethi et al. 2015 ), improve the quality of products (Abrahams et al. 2012 ; Isah et al. 2014 ; Myslin et al. 2013 ), understand the direction of public opinion (Fink et al. 2013 ; Groshek and Al-Rawi 2013 ), and so on.

The keyword co-occurrence network for the C1 community

Users’ sentiments are often associated with the topics, and the accuracy of sentiment analysis can be improved through the introduction of topic models (Li et al. 2010 ). Among them, the Latent Dirichlet Allocation (LDA) method is cited most frequently. Previous studies found that the LDA method can be effective in subdividing topics and identifying the sentiments of the contents. This method is quite general, and there are also many improved models based on this one that can be applied to any type of web text, helping to enhance the accuracy of sentiment polarity calculation (Chen et al. 2019 ; Liu et al. 2020 ).

As the COVID-19 pandemic has unfolded, a large number of individuals, media and governments have been publishing news and opinions about the COVID-19 crisis on social media platforms. This has resulted in a lot of sentiment analysis studies focusing on COVID-19-related texts exploring the impact of the epidemic on people’s lives (Sari and Ruldeviyani 2020 ; Wang, T. et al. 2020a ), physical health (Berkovic et al. 2020 ; Binkheder et al. 2021 ) and mental health (Yin et al. 2020 ), and so on. Therefore, we can see many related keywords, such as "infodemiology," "healthcare," and "mental health."

Analysis on research methods and topics of the C2 community

The contents of the C2 community mainly focus on "machine learning," "text classification," "feature extraction," and "stock market" (see Fig.  5 ). Most keywords are related to the research methods of sentiment analysis. Machine learning approaches have expanded from topic recognition to more challenging tasks such as sentiment classification. It is very important to explore and compare machine learning methods applied to sentiment classification (Li and Sun 2007 ). Methods like Support Vector Machine (SVM) and Naive Bayes models are widely used (Altrabsheh et al. 2013 ; Dereli et al. 2021 ; Shofiya and Abidi 2021 ; Tan et al. 2009 ; Wang and Lin 2020 ) and are used as benchmarks for the comparisons of models proposed by many researchers (Kumar et al. 2021 ; Sadamitsu et al. 2008 ; Waila et al. 2012 ; Zhang et al. 2019 ). Many algorithms, such as random forest (Al Amrani et al. 2018 ; Fitri et al. 2019 ; Sutoyo et al. 2022 ), tf-idf (Arafin Mahtab et al. 2018 ; Awan et al. 2021 ; Dey et al. 2017 ), logistic regression (Prabhat and Khullar 2017 ; Qasem et al. 2015 ; Sutoyo et al. 2022 ), and n-gram (Ikram and Afzal 2019 ; Singh and Kumari 2016 ; Xiong et al. 2021 ) are used to enhance the accuracy of machine learning, as shown in Fig.  5 .

The keyword co-occurrence network for the C2 community

The trading volume and asset prices of financial commodities or financial instruments are influenced by a variety of factors in the online environment. Machine learning and sentiment analysis are powerful tools that can help gather vast amounts of useful information to predict financial risk effectively (Li et al. 2009 ). Research on the relationship between public sentiment and stock prices has always been the focus of many scholars (Smailović et al. 2014 ; Xing et al. 2018 ). They have used machine learning methods to explore the influence of sentiments on stock prices through sentiment analysis of news articles, and then predicted the trend changes in the stock market (Ahuja et al. 2015 ; Januário et al. 2022 ; Maqsood et al. 2020 ; Picasso et al. 2019 ).

Analysis on research methods and topics of the C3 community

The contents of the C3 community also mainly focus on the methods for sentiment analysis, like "natural language processing", "deep learning," "aspect-based sentiment analysis," and "task analysis" (Fig.  6 ). Sentiment analysis is a sub-field of natural language processing (Nicholls and Song 2010 ), and natural language processing techniques have been widely used in sentiment analysis. Using natural language processing technology can help to better parse text features, such as part-of-speech tagging, word sense disambiguation, keyword extraction, inter-word dependency recognition, semantic parsing, and dictionary construction (Abbasi et al. 2011 ; Syed et al. 2010 ; Trilla and Alías 2009 ). With the rise of deep learning technology, researchers began to introduce it to sentiment analysis. Neural network models like LSTM (Al-Dabet et al. 2021 ; Al-Smadi et al. 2019 ; Li and Qian 2016 ; Schuller et al. 2015 ; Tai et al. 2015 ), CNN (Cai and Xia 2015 ; Jia and Wang 2022 ; Ouyang et al. 2015 ), RNN (Hassan and Mahmood 2017 ; Tembhurne and Diwan 2021 ; You et al. 2016 ), and some combination of these, as well as other models (An and Moon 2022 ; Li et al. 2022 ; Liu et al. 2020a ; Salur and Aydin 2020 ; Zhao et al. 2021 ), have received significant attention.

The keyword co-occurrence network for the C3 community

Sentiment analysis granularity is subdivided into document level, sentence level, and aspect level. Document-level sentiment analysis takes the entire document as a unit, but the premise is that the document needs to have a clear attitude orientation—that is, the point of view needs to be clear (Shirsat et al. 2018 ; Wang and Wan 2011 ). Sentence-level sentiment analysis is intended to perform sentiment analysis of the sentences in the document alone (Arulmurugan et al. 2019 ; Liu et al. 2009 ; Nejat et al. 2017 ). Aspect-based analysis is a fundamental and significant task in sentiment analysis. The aim of aspect-level sentiment analysis is to separately summarize positive and negative views about different aspects of a product or entity, although overall sentiment toward a product or entity may tend to be positive or negative (Rao et al. 2021 ; Thet et al. 2010 ). Aspect-level sentiment analysis facilitates a more finely-grained analysis of sentiment than either document or sentence-level analysis (Liang et al. 2022 ; Wang et al. 2020c ). The traditional levels of analysis, such as sentence-level analysis can only calculate the comprehensive sentiment polarity of paragraphs or sentences (Wang et al. 2016 ; Zhang et al. 2021 ). In recent years, the aspect level has become more and more popular, and with the application of deep learning technology, it has become better at capturing the semantic relationship between aspect terms and words in a more quantifiable way (Huang et al. 2018 ). The process of sentiment analysis involves the coordination of multiple tasks, and the subtasks include feature extraction (Bouktif et al. 2020 ; Lin et al. 2020 ), context analysis (Yu et al. 2019 ; Zuo et al. 2020 ), and the application of some analytical models (Tan et al. 2020 ).

Analysis on research methods and topics of the C4 community

The C4 community mainly shows keywords related to the research methods and topics of "opinion mining" and "user review," which is the largest of the six sub-communities (Fig.  7 ). With the popularity of platforms like online review sites and personal blogs on the Internet, opinions and user reviews are readily available on the web. Opinion mining has always been a hot field of research (Khan et al. 2009 ; Poria et al. 2016 ). From Table ​ Table4, 4 , we can see that the link between C3 and C4 has 1306 lines. In opinion mining, researchers use many text mining methods to discover users’ opinions on goods or services, and then help improve the quality of corresponding products or services (Da’u et al. 2020 ; Lo and Potdar 2009 ; Martinez-Camara et al. 2011 ). In addition, scholars have found that the consideration of user opinions can help improve the overall quality of recommender systems (Artemenko et al. 2020 ; Da’u et al. 2020 ; Garg 2021 ; Malandri et al. 2022 ). Therefore, "recommendation system" has a strong correlation with "opinion mining."

The keyword co-occurrence network for C4 community

Evaluation metrics for quantifying the existing approaches are also a popular topic related to opinion mining. There is a keyword named "performance sentiment" in the C4 community. Precision, recall, accuracy and F1-score are the most commonly used evaluation metrics (Dangi et al. 2022 ; Jain et al. 2022 ; JayaLakshmi and Kishore 2022 ; Li et al. 2017 ; Wang et al. 2021 ; Yi and Niblack 2005 ). Some researchers have also used runtimes to calculate the model efficiency (Abo et al. 2021 ; Ferilli et al. 2015 ), p-value to statistically evaluate the relationship or difference between two samples of classification results (JayaLakshmi and Kishore 2022 ; Salur and Aydin 2020 ), paired sample t-tests to verify that the results are not obtained by chance (Nhlabano and Lutu 2018 ), and standard deviation to measure the stability of the model (Chang et al. 2020 ). There have also been researchers who have used G-mean (Wang et al. 2021 ), Pearson Correlation Coefficient (Corr) (Yang et al. 2022 ), Mean Absolute Error (MAE) (Yang et al. 2022 ), Normalized Information Transfer (NIT) and Entropy-Modified Accuracy (EMA) (Valverde-Albacete et al. 2013 ), Mean Squared Error (MSE) (Mao et al. 2022 ), Hamming loss (Liu and Chen 2015 ), Area Under the Curve (AUC) (Abo et al. 2021 ), sensitivity and specificity (Thakur and Deshpande 2019 ), etc.

Analysis on research methods and topics of the C5 & C6 communities

Both sub-communities C5 (Fig.  8 ) and C6 (Fig.  9 ) are small in size. The C5 community has 25 nodes and the C6 community has 41 nodes. The core content of the C5 community is "Arabic sentiment analysis." Before 2011, most resources and systems built in the field of sentiment analysis were tailored to English and other Indo-European languages. It is increasingly necessary to design sentiment analysis systems for other languages (Korayem et al. 2012 ), and researchers are increasingly interested in the study of tweets and texts in the Arabic language (Heikal et al. 2018 ; Khasawneh et al. 2013 ; Oueslati et al. 2020 ). They use technologies such as named entity recognition (Al-Laith and Shahbaz 2021 ), deep learning (Al-Ayyoub et al. 2018 ; Heikal et al. 2018 ), and corpus construction (Alayba et al. 2018 ) to enhance the accuracy of sentiment analysis.

The keyword co-occurrence network for the C5 community

The keyword co-occurrence network for the C6 community

The contents of the C6 community are not very concentrated. From the size of the circle, we can see that the keywords "domain adaptation"(Blitzer et al. 2007 ; Glorot et al. 2011 ), "domain sentiment," and "cross-domain" appear more frequently. Cross-domain sentiment classification is intended to address the lack of mass labeling data (Du et al. 2020a ). It has attracted much attention (Du et al. 2020b ; Hao et al. 2019 ; Yang et al. 2020b ). Advances in communication technology have provided valuable interactive resources for people in different regions, and the processing of multilingual user comments has gradually become a key challenge in natural language processing (Martinez-Garcia et al. 2021 ). Therefore, some keywords related to "lingual" have appeared. Other keywords, such as "transfer learning," "active learning," and "semi-supervised learning," are mainly related to sentiment analysis technologies.

Evolution of research methods and topics of sentiment analysis

Overall evolution analysis.

Annual changes in keyword frequency in sentiment analysis research can reflect the evolution of research methods and topics in this field. Based on the keyword community network (Fig.  3 ), we counted the frequency of keywords in each sub-community for each year. The keyword community evolution diagram is shown in Fig.  10 . Since there were fewer papers published before 2006, we combined the occurrences of keywords from 2002 to 2006. We can see that the C1 community and the C3 community have shown a significant growth trend. The C2 community was in a state of growth until 2019, and the frequency of keywords decreased year by year after 2019. The frequency of C4 community keywords continued to increase until 2018 and declined after 2018. The number of keywords in the C5 community and in the C6 community both had a slow growth trend, but the trend was not obvious.

Keyword community evolution diagram

Evolution analysis of sub-communities

We selected the high-frequency keywords under each category and plotted the change of word frequency in each year, as shown in Figs.  11 and ​ and12. 12 . In the C1 community, "social medium," "Twitter," "social network," "covid-19," "Latent Dirichlet Allocation," "topic model," and "text analysis" all had significant increases in word frequency, and the growth trend in 2021 was obvious. "Covid-19" appears in 2020, and the word frequency increased rapidly in 2021. Social media platforms have always been the focus of researchers’ attention. Under the influence of COVID-19, more people express their emotions, stress, and thoughts through social media platforms. Sentiment analysis on data from social media platforms related to COVID-19 has become a hot topic (Boon-Itt and Skunkan 2020 ). We believe that due to the impact of COVID-19, the widespread use of social platforms in 2020–2021 has led to a surge in the number of C1-related keywords.

C1, C2, C5, C6 communities: High-frequency keyword evolution diagram

C3, C4 communities: High-frequency keyword evolution diagram

The C2 community focuses on the method of "machine learning," and the C3 community focuses on the methods of "deep learning" and "natural language processing." The keywords in the two communities are mainly related to the techniques and methods of sentiment analysis. We have found that before 2016 (Fig.  10 ), the frequency of keywords in the C2 community was higher than that in the C3 community, and in 2016 and later, the frequency of keywords in the C3 community gradually accounted for a larger proportion of the total. This reflects the fact that deep learning-related technologies and methods have become a research hotspot, and the attention given to SVM, Naive Bayes, supervised learning, and other technologies in machine learning has declined. In addition to deep learning models such as Bi-LSTM, Long Short-term Memory, and recurrent neural network in the C3 community, the number of "aspect based" and "feature extraction" keywords have also been growing, which shows that researchers now pay more attention to the aspect level of text granularity in the field of sentiment analysis.

Among the keywords found in the C4 community, the word frequency of the "opinion mining" keyword has decreased since 2018. This shows that in the field of sentiment analysis, researchers have begun to reduce the attention they give to sentiment analysis of opinions on product or service quality, while still maintaining a certain degree of attention to "user review" and "online review." In addition, the number of keywords for "sentiment lexicon" and "lexicon-based" has declined. It may be because, in the context of the widespread application of deep learning technology in recent years, the lexicon-based method requires more time and higher labor costs (Kaity and Balakrishnan 2020 ). However, its accuracy still attracts attention due to the high involvement of experts, especially in non-English languages (Bakar et al. 2019 ; Kydros et al. 2021 ; Piryani et al. 2020 ; Tammina 2020 ; Xing et al. 2019 ; Yurtalan et al. 2019 ).

The high-frequency keywords in the C5 and C6 communities are "Arabic language," "Arabic sentiment analysis," and "transfer learning." Arabic has 30 variants, including the official Modern Standard Arabic (MSA) (ISO 639–3 2017). Arabic dialects are becoming increasingly popular as the language of informal communication on blogs, forums, and social media networks (Lulu and Elnagar 2018 ). This makes them challenging languages for natural language processing and sentiment analysis (Alali et al. 2019 ; Elshakankery and Ahmed 2019 ; Sayed et al. 2020 ). Transfer learning can solve the problem by leveraging knowledge obtained from a large-scale source domain to enhance the classification performance of target domains (Heaton 2018 ). In recent years, based on the success of deep learning technology, this method has gradually attracted attention.

Research hotspots and trends

Through the analysis in Sects.  4.1 and 4.2 , we found that the research methods and topics of sentiment analysis are constantly changing. The keyword topic heat map is shown in Fig.  13 . From this map, we can see that in the past two decades, research hotspots have included social media platforms (such as "social medium," "social network," and "Twitter"); sentiment analysis techniques and methods (such as "machine learning," "svm," "natural language processing," "deep learning," "aspect-based," "text mining," and "sentiment lexicon"), mining of user comments or opinions (e.g., "opinion mining," "user review," and "online review"), and sentiment analysis for non-English languages (e.g., "Arabic sentiment analysis" and "Arabic language").

Keyword topic heat map

With the popularity of digitization, a large amount of user-generated content has appeared on the Internet, where users express their opinions and comments on different topics such as the news, events, activities, products, services, etc. through social media. This is especially so in the case of the Twitter mobile platform, launched in 2006, which has become the most popular social channel (Kumar and Jaiswal 2020 ). However, online text data is mostly unstructured. In order to accurately analyze users’ sentiments, the research methods for sentiment analysis, such as natural language processing technology, and automatic sentiment analysis models have become the focus of researchers’ works. From Fig.  11 , we can see that early technologies and methods are dominated by machine learning and that SVM and Naive Bayes have always been favored by researchers. This has also been confirmed in studies by Neha Raghuvanshi (Raghuvanshi and Patil 2016 ), Harpreet Kaur (Kaur et al. 2017 ), and Marouane Birjali (Birjali et al. 2021 ). With the improvement of neural network and artificial intelligence technology, deep learning technology has been widely used in sentiment analysis, and has resulted in good outcomes (Basiri et al. 2021 ; Ma et al. 2018 ; Prabha and Srikanth 2019 ; Yuan et al. 2020 ). However, deep learning technology still has room for improvement, and the hybrid methods combining sentiment dictionary and semantic analysis are gradually becoming a trend (Prabha and Srikanth 2019 ; Yang et al. 2020a ).

The granularity of sentiment analysis ranges from the early text level to the sentence level and finally to the aspect level, which is currently gaining strong attention. The granularity of sentiment analysis is gradually being refined, but the method is immature at present, and further research work in the future is needed (Agüero-Torales et al. 2021 ; Li et al. 2020 ; Trisna and Jie 2022 ).

Early sentiment analysis was mainly in the English language. In recent years, non-English languages such as Chinese (Lai et al. 2020 ; Peng et al. 2018 ), French (Apidianaki et al. 2016 ; Pecore and Villaneau 2019 ), Spanish (Chaturvedi et al. 2016 ; Plaza-del-Arco et al. 2020 ), Russian (Smetanin 2020 ), and Arabic (Alhumoud and Al Wazrah 2022 ; Ombabi et al. 2020 ) have attracted more and more attention. Furthermore, cross-domain sentiment analysis technology is in urgent need of research and discussion by researchers (Liu et al. 2019 ; Singh et al. 2021 ).

Conclusion and future work

Judging from the increasing number of papers related to sentiment analysis research every year, sentiment analysis has been on the rise. Although there are many surveys on sentiment analysis research, there has not been a survey dedicated to the evolution of research methods and topics of sentiment analysis. This paper has used keyword co-occurrence analysis and the informetric tools to enrich the perspectives and methods of previous studies. Its aims have been to outline the evolution of the research methods and tools, research hotspots and trends and to provide research guidance for researchers.

By adopting keyword co-occurrence analysis and community detection methods, we analyzed the research methods and topics of sentiment analysis, as well as their connections and evolution trends, and summarized the research hotspots and trends in sentiment analysis. We found that research hotspots include social media platforms, sentiment analysis techniques and methods, mining of user comments or opinions, and sentiment analysis for non-English languages. Moreover, deep learning technology, with its hybrid methods combining sentiment dictionary and semantic analysis, fine-grained sentiment analysis methods, and non-English language analysis methods, and cross-domain sentiment analysis techniques have gradually become the research trends.

Practical implications and technical directions of sentiment analysis

Sentiment analysis has a wide range of application targets, such as e-commerce platforms, social platforms, public opinion platforms, and customer service platforms. Years of development have led to many related tasks in sentiment analysis, such as sentiment analysis of different text granularity, sentiment recognition, opinion mining, dialogue sentiment analysis, irony recognition, false information detection, etc. Such analysis can help structure user reviews, support product improvement decisions, discover public opinion hotspots, identify public positions, investigate user satisfaction with products, and so on. As long as user-generated content is involved, sentiment analysis technology can be used to mine the emotions of human actors associated with the content. The improvement of sentiment analysis technology can help machines better understand the thoughts and opinions of users, make machines more intelligent, and make better decisions for policy leaders, businessmen, and service people. However, most of the current sentiment analysis methods are based on sentiment dictionaries, sentiment rules, statistics-based machine learning models, neural network-based deep learning models, and pre-training models, and have yet to achieve true language understanding in the sense of comprehension at the deep semantic level, though this does not prevent them from being useful in certain practical applications.

As an important task in natural language understanding, sentiment analysis has received extensive attention from academia and industry. Coarse-grained sentiment analysis is increasingly unable to meet people's decision-making needs, and for aspect-level sentiment analysis and complex tasks, pure machine learning is still unable to flexibly achieve true language understanding. Once the scene or domain changes, problems such as the domain incompatibility of the sentiment dictionary and the low transfer effect of the model involved keep appearing. At present, the accuracy of sentiment analysis provided by machines is far less than that of humans. To achieve human-like performance for machines, we believe that it is necessary to incorporate human commonsense knowledge and domain knowledge, as well as grounded definitions of concepts, in order for machines to understand natural language at a deeper level. These, combined with rules for affective reasoning to supplement interpretable information, will be effective in improving the performance of sentiment analysis. Future research in this direction can be strengthened to achieve true language understanding in machines.

Limitations and future work

There are some research limitations in this paper. First, we only studied papers written in English and searched from the Web of Science platform. We believe there are papers in other languages or other databases (e.g., Scopus, PubMed, Sci-hub, etc.) that also involve sentiment analysis but that were not included in our study. In addition, the keywords we chose to search in the Web of Science were mainly "sentiment analysis," "sentiment mining," and "sentiment classification." There may be papers related to our research topic that do not have these keywords. To track developments in sentiment analysis research, future studies could replicate this work by employing more precise keywords and using different literature databases.

Second, we selected the main high-frequency keywords for analysis, and some important low-frequency keywords may have been ignored. In future work, we can analyze the changes in each keyword in detail from the perspective of time and obtain more comprehensive analysis results.

Third, the results show that the themes of sentiment analysis cover many fields, such as computer science, linguistics, and electrical engineering, which indicates the trend of interdisciplinary research. Therefore, future work should apply co-citation and diversity measures to explore the interdisciplinary nature of sentiment analysis research.

Acknowledgements

The authors would like to thank the China Scholarship Council (CSC No. 202106850069) for its support for the visiting study.

This work has not received any funding.

Data availability

Declarations.

The authors declare that they have no conflict of interest or competing interest in this article.

This article does not contain any studies with human participants or animals performed by any of the authors.

1 https://github.com/MaartenGr/KeyBERT .

2 https://homepage.univie.ac.at/juan.gorraiz/bibexcel/ .

3 http://mrvar.fdv.uni-lj.si/pajek/ .

4 https://www.vosviewer.com/ .

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Contributor Information

Jingfeng Cui, Email: nc.ude.uajn@5004129102 .

Zhaoxia Wang, Email: gs.ude.ums@gnawxz .

Seng-Beng Ho, Email: gs.ude.rats-a.cphi@bsoh .

Erik Cambria, Email: gs.ude.utn@airbmac .

• Abbasi A, France S, Zhang Z, Chen H. Selecting attributes for sentiment classification using feature relation networks. IEEE Trans Knowl Data Eng. 2011; 23 (3):447–462. doi: 10.1109/TKDE.2010.110. [ CrossRef ] [ Google Scholar ]
• Abdullah NSD, Zolkepli IA (2017) Sentiment analysis of online crowd input towards Brand Provocation in Facebook, Twitter, and Instagram. In: Proceedings of the international conference on big data and internet of thing, association for computing machinery, pp 67–74. 10.1145/3175684.3175689
• Abo MEM, Idris N, Mahmud R, Qazi A, Hashem IAT, Maitama JZ, et al. A multi-criteria approach for Arabic dialect sentiment analysis for online reviews: exploiting optimal machine learning algorithm selection. Sustainability. 2021; 13 (18):10018. doi: 10.3390/su131810018. [ CrossRef ] [ Google Scholar ]
• Abrahams AS, Jiao J, Wang GA, Fan W. Vehicle defect discovery from social media. Decis Support Syst. 2012; 54 (1):87–97. doi: 10.1016/j.dss.2012.04.005. [ CrossRef ] [ Google Scholar ]
• Acheampong FA, Nunoo-Mensah H, Chen W. Transformer models for text-based emotion detection: a review of BERT-based approaches. Artif Intell Rev. 2021; 54 (8):5789–5829. doi: 10.1007/s10462-021-09958-2. [ CrossRef ] [ Google Scholar ]
• Adak A, Pradhan B, Shukla N. Sentiment analysis of customer reviews of food delivery services using deep learning and explainable artificial intelligence: systematic review. Foods. 2022; 11 (10):1500. doi: 10.3390/foods11101500. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Agüero-Torales MM, Salas JIA, López-Herrera AG. Deep learning and multilingual sentiment analysis on social media data: an overview. Appl Soft Comput. 2021; 107 :107373. doi: 10.1016/j.asoc.2021.107373. [ CrossRef ] [ Google Scholar ]
• Ahuja R, Rastogi H, Choudhuri A, Garg B (2015) Stock market forecast using sentiment analysis. In: 2015 2nd International conference on computing for sustainable global development, INDIACom 2015, Bharati Vidyapeeth, New Delhi, pp 1008–1010. 10.48550/arXiv.2204.05783
• Ain QT, Ali M, Riaz A, Noureen A, Kamranz M, Hayat B, et al. Sentiment analysis using deep learning techniques: a review. Int J Adv Comput Sci Appl. 2017; 8 (6):424–433. doi: 10.14569/ijacsa.2017.080657. [ CrossRef ] [ Google Scholar ]
• Al-Ayyoub M, Nuseir A, Alsmearat K, Jararweh Y, Gupta B. Deep learning for Arabic NLP: a survey. J Comput Sci. 2018; 26 :522–531. doi: 10.1016/j.jocs.2017.11.011. [ CrossRef ] [ Google Scholar ]
• Al-Ayyoub M, Khamaiseh AA, Jararweh Y, Al-Kabi MN. A comprehensive survey of Arabic sentiment analysis. Inf Process Manag. 2019; 56 (2):320–342. doi: 10.1016/j.ipm.2018.07.006. [ CrossRef ] [ Google Scholar ]
• Al-Dabet S, Tedmori S, AL-Smadi M. Enhancing Arabic aspect-based sentiment analysis using deep learning models. Comput Speech Lang. 2021; 69 :1224. doi: 10.1016/j.csl.2021.101224. [ CrossRef ] [ Google Scholar ]
• Al-Laith A, Shahbaz M. Tracking sentiment towards news entities from Arabic news on social media. Futur Gener Comput Syst. 2021; 118 :467–484. doi: 10.1016/j.future.2021.01.015. [ CrossRef ] [ Google Scholar ]
• Al-Smadi M, Talafha B, Al-Ayyoub M, Jararweh Y. Using long short-term memory deep neural networks for aspect-based sentiment analysis of Arabic reviews. Int J Mach Learn Cybern. 2019; 10 (8):2163–2175. doi: 10.1007/s13042-018-0799-4. [ CrossRef ] [ Google Scholar ]
• Alali M, Sharef NM, Murad MAA, Hamdan H, Husin NA. Narrow convolutional neural network for Arabic dialects polarity classification. IEEE Access. 2019; 7 :96272–96283. doi: 10.1109/ACCESS.2019.2929208. [ CrossRef ] [ Google Scholar ]
• Alamoodi AH, Zaidan BB, Al-Masawa M, Taresh SM, Noman S, Ahmaro IYY, et al. Multi-perspectives systematic review on the applications of sentiment analysis for vaccine hesitancy. Comput Biol Med. 2021; 139 :4957. doi: 10.1016/j.compbiomed.2021.104957. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Alamoodi AH, Zaidan BB, Zaidan AA, Albahri OS, Mohammed KI, Malik RQ, et al. Sentiment analysis and its applications in fighting COVID-19 and infectious diseases: a systematic review. Expert Syst Appl. 2021; 167 :114155. doi: 10.1016/j.eswa.2020.114155. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Alayba AM, Palade V, England M, Iqbal R (2018) Improving sentiment analysis in arabic using word representation. In: 2018 IEEE 2nd International Workshop on Arabic and Derived Script Analysis and Recognition (ASAR), IEEE, pp 13–18. 10.1109/ASAR.2018.8480191
• Alhumoud SO, Al Wazrah AA. Arabic sentiment analysis using recurrent neural networks: a review. Artif Intell Rev. 2022; 55 (1):707–748. doi: 10.1007/s10462-021-09989-9. [ CrossRef ] [ Google Scholar ]
• Alonso MA, Vilares D, Gómez-Rodríguez C, Vilares J. Sentiment analysis for fake news detection. Electronics. 2021; 10 (11):1348. doi: 10.3390/electronics10111348. [ CrossRef ] [ Google Scholar ]
• Altrabsheh N, Gaber MM, Cocea M (2013) SA-E: sentiment analysis for education. In: The 5th KES International Conference on Intelligent Decision Technologies (KES-IDT), Sesimbra, Portugal, pp 353–362. 10.3233/978-1-61499-264-6-353
• Al Amrani Y, Lazaar M, El Kadirp KE. Random forest and support vector machine based hybrid approach to sentiment analysis. Procedia Comput Sci. 2018; 127 :511–520. doi: 10.1016/j.procs.2018.01.150. [ CrossRef ] [ Google Scholar ]
• An H, Moon N. Design of recommendation system for tourist spot using sentiment analysis based on CNN-LSTM. J Ambient Intell Hum Comput. 2022; 13 :1653–1663. doi: 10.1007/s12652-019-01521-w. [ CrossRef ] [ Google Scholar ]
• Angel SO, Negron APP, Espinoza-Valdez A. Systematic literature review of sentiment analysis in the spanish language. Data Technol Appl. 2021; 55 (4):461–479. doi: 10.1108/DTA-09-2020-0200. [ CrossRef ] [ Google Scholar ]
• Apidianaki M, Tannier X, Richart C (2016) Datasets for aspect-based sentiment analysis in French. In: Proceedings of the tenth international conference on language resources and evaluation (LREC’16), Portorož, Slovenia: European Language Resources Association (ELRA), pp 1122–1126. https://aclanthology.org/L16-1179
• Arafin Mahtab S, Islam N, Mahfuzur Rahaman M (2018) Sentiment analysis on Bangladesh cricket with support vector machine. In: 2018 International conference on Bangla Speech and language processing (ICBSLP), IEEE, pp 1–4. 10.1109/ICBSLP.2018.8554585
• Artemenko O, Pasichnyk V, Kunanets N, Shunevych K (2020) Using sentiment text analysis of user reviews in social media for E-Tourism mobile recommender systems. In: COLINS, CEUR-WS, Aachen, pp 259–271. http://ceur-ws.org/Vol-2604/paper20.pdf
• Arulmurugan R, Sabarmathi KR, Anandakumar H. Classification of sentence level sentiment analysis using cloud machine learning techniques. Clust Comput. 2019; 22 (1):1199–1209. doi: 10.1007/s10586-017-1200-1. [ CrossRef ] [ Google Scholar ]
• Asghar MZ, Khan A, Ahmad S, Kundi FM. A review of feature selection techniques in sentiment analysis. J Basic Appl Sci Res. 2014; 4 (3):181–186. doi: 10.3233/IDA-173763. [ CrossRef ] [ Google Scholar ]
• Awan MJ, Yasin A, Nobanee H, Ali AA, Shahzad Z, Nabeel M, et al. Fake news data exploration and analytics. Electronics. 2021; 10 (19):2326. doi: 10.3390/electronics10192326. [ CrossRef ] [ Google Scholar ]
• Bai H, Yu G. A Weibo-based approach to disaster informatics: incidents monitor in post-disaster situation via weibo text negative sentiment analysis. Nat Hazards. 2016; 83 (2):1177–1196. doi: 10.1007/s11069-016-2370-5. [ CrossRef ] [ Google Scholar ]
• Bakar MFRA, Idris N, Shuib L (2019) An enhancement of Malay social media text normalization for Lexicon-based sentiment analysis. In: 2019 International conference on Asian language processing (IALP), IEEE, pp 211–215. 10.1109/IALP48816.2019.9037700
• Bar-Ilan J. Informetrics at the beginning of the 21st century—a review. J Informet. 2008; 2 (1):1–52. doi: 10.1016/j.joi.2007.11.001. [ CrossRef ] [ Google Scholar ]
• Basiri ME, Nemati S, Abdar M, Cambria E, Acharya UR. ABCDM: an attention-based bidirectional CNN-RNN deep model for sentiment analysis. Futur Gener Comput Syst. 2021; 115 :279–294. doi: 10.1016/j.future.2020.08.005. [ CrossRef ] [ Google Scholar ]
• Batagelj V, Andrej M (2022) Pajek [Software]. http://mrvar.fdv.uni-lj.si/pajek/
• Batagelj V, Mrvar A (1998) Pajek-program for large network analysis eds. M. Jünger and P Mutzel. Connections 21(2): 47–57. http://vlado.fmf.uni-lj.si/pub/networks/doc/pajek.pdf
• Bengtsson M. How to plan and perform a qualitative study using content analysis. NursingPlus Open. 2016; 2 :8–14. doi: 10.1016/j.npls.2016.01.001. [ CrossRef ] [ Google Scholar ]
• Berkovic D, Ackerman IN, Briggs AM, Ayton D. Tweets by people with arthritis during the COVID-19 pandemic: content and sentiment analysis. J Med Internet Res. 2020; 22 (12):e24550. doi: 10.2196/24550. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Binkheder S, Aldekhyyel RN, Almogbel A, Al-Twairesh N, Alhumaid N, Aldekhyyel SN, et al. Public perceptions around Mhealth applications during Covid-19 pandemic: a network and sentiment analysis of tweets in Saudi Arabia. Int J Environ Res Public Health. 2021; 18 (24):1–22. doi: 10.3390/ijerph182413388. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Birjali M, Kasri M, Beni-Hssane A. A comprehensive survey on sentiment analysis: approaches, challenges and trends. Knowl Based Syst. 2021; 226 :107134. doi: 10.1016/j.knosys.2021.107134. [ CrossRef ] [ Google Scholar ]
• Blitzer J, Dredze M, Pereira F (2007) Biographies, bollywood, boom-boxes and blenders: domain adaptation for sentiment classification. In: 45th Annual Meeting of the association of computational linguistics, association for computational linguistics, pp 440–447. 10.1287/ijoc.2013.0585
• Blondel VD, Guillaume JL, Lambiotte R, Lefebvre E. Fast unfolding of communities in large networks. J Stat Mech Theory Exp. 2008; 2008 (10):P10008. doi: 10.1088/1742-5468/2008/10/P10008. [ CrossRef ] [ Google Scholar ]
• Boon-Itt S, Skunkan Y. Public perception of the COVID-19 pandemic on Twitter: sentiment analysis and topic modeling study. JMIR Public Health Surv. 2020; 6 (4):1978. doi: 10.2196/21978. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Boudad N, Faizi R, Thami ROH, Chiheb R. Sentiment analysis in Arabic: a review of the literature. Ain Shams Eng J. 2018; 9 (4):2479–2490. doi: 10.1016/j.asej.2017.04.007. [ CrossRef ] [ Google Scholar ]
• Bouktif S, Fiaz A, Awad M. Augmented textual features-based stock market prediction. IEEE Access. 2020; 8 :40269–40282. doi: 10.1109/ACCESS.2020.2976725. [ CrossRef ] [ Google Scholar ]
• Brito KDS, Filho RLCS, Adeodato PJL. A systematic review of predicting elections based on social media data: research challenges and future directions. IEEE Trans Comput Soc Syst. 2021; 8 (4):819–843. doi: 10.1109/TCSS.2021.3063660. [ CrossRef ] [ Google Scholar ]
• Cai G, Xia B (2015) Convolutional neural networks for multimedia sentiment analysis. In: Natural Language Processing and Chinese Computing, Springer, Cham, p 159–167. 10.1007/978-3-319-25207-0_14
• Callon M, Courtial J-P, Turner WA, Bauin S. From translations to problematic networks: an introduction to co-word analysis. Soc Sci Inf. 1983; 22 (2):191–235. doi: 10.1177/053901883022002003. [ CrossRef ] [ Google Scholar ]
• Cambria E, Liu Q, Decherchi S, Xing F, Kwok K (2022a) SenticNet 7: a commonsense-based neurosymbolic AI Framework for Explainable Sentiment Analysis. In: LREC, Marseille: European Language Resources Association (ELRA), pp 3829–3839. https://sentic.net/senticnet-7.pdf
• Cambria E, Dragoni M, Kessler B, Donadello I. Ontosenticnet 2: enhancing reasoning within sentiment analysis. IEEE Intell Syst. 2022; 37 (2):103–110. doi: 10.1109/MIS.2021.3093659. [ CrossRef ] [ Google Scholar ]
• Cambria E, Kumar A, Al-Ayyoub M, Howard N. Guest editorial: explainable artificial intelligence for sentiment analysis. Knowl Based Syst. 2022; 238 (3):107920. doi: 10.1016/j.knosys.2021.107920. [ CrossRef ] [ Google Scholar ]
• Cambria E, Xing F, Thelwall M, Welsch R. Sentiment analysis as a multidisciplinary research area. IEEE Trans Artif Intell. 2022; 3 (2):1–4. [ Google Scholar ]
• Chan JYL, Bea KT, Leow SMH, Phoong SW, Cheng WK. State of the art: a review of sentiment analysis based on sequential transfer learning. Artif Intell Rev. 2022 doi: 10.1007/s10462-022-10183-8. [ CrossRef ] [ Google Scholar ]
• Chang J-R, Liang H-Y, Chen L-S, Chang C-W. Novel feature selection approaches for improving the performance of sentiment classification. J Ambient Intell Hum Comput. 2020 doi: 10.1007/s12652-020-02468-z. [ CrossRef ] [ Google Scholar ]
• Chaturvedi I, Cambria E, Vilares D (2016) Lyapunov filtering of objectivity for Spanish sentiment model. In: Proceedings of the International Joint Conference on Neural Networks (IJCNN), IEEE, pp 4474–4481. 10.1109/IJCNN.2016.7727785
• Chen Z, Teng S, Zhang W, Tang H, Zhang Z, He J, et al (2019) LSTM sentiment polarity analysis based on LDA clustering. In: Communications in Computer and Information Science, Springer, Singapore, pp 342–355. 10.1007/978-981-13-3044-5_25
• Cheng WK, Bea KT, Leow SMH, Chan JY-L, Hong Z-W, Chen Y-L. A review of sentiment, semantic and event-extraction-based approaches in stock forecasting. Mathematics. 2022; 10 (14):2437. doi: 10.3390/math10142437. [ CrossRef ] [ Google Scholar ]
• Da’u A, Salim N, Rabiu I, Osman A. Recommendation System Exploiting Aspect-Based Opinion Mining with Deep Learning Method. Inf Sci. 2020; 512 :1279–1292. doi: 10.1016/j.ins.2019.10.038. [ CrossRef ] [ Google Scholar ]
• Dangi D, Bhagat A, Dixit DK. Sentiment analysis of social media data based on chaotic coyote optimization algorithm based time weight-adaboost support vector machine approach. Concurr Comput. 2022; 34 (3):6581. doi: 10.1002/cpe.6581. [ CrossRef ] [ Google Scholar ]
• Deng S, Xia S, Hu J, Li H, Liu Y. Exploring the topic structure and evolution of associations in information behavior research through co-word analysis. J Librariansh Inf Sci. 2021; 53 (2):280–297. doi: 10.1177/0961000620938120. [ CrossRef ] [ Google Scholar ]
• Dereli T, Eligüzel N, Çetinkaya C. Content analyses of the international federation of Red Cross and Red Crescent Societies (Ifrc) based on machine learning techniques through Twitter. Nat Hazards. 2021; 106 (3):2025–2045. doi: 10.1007/s11069-021-04527-w. [ CrossRef ] [ Google Scholar ]
• Dey A, Jenamani M, Thakkar JJ (2017) Lexical Tf-Idf: An n-Gram Feature Space for Cross-Domain Classification of Sentiment Reviews. In: International Conference on Pattern Recognition and Machine Intelligence, Springer, Cham, pp 380–386. 10.1007/978-3-319-69900-4_48
• Ding Y, Chowdhury GG, Foo S. Bibliometric cartography of information retrieval research by using co-word analysis. Inf Process Manag. 2001; 37 (6):817–842. doi: 10.1016/S0306-4573(00)00051-0. [ CrossRef ] [ Google Scholar ]
• Du C, Sun H, Wang J, Qi Q, Liao J (2020a) Adversarial and domain-aware BERT for cross-domain sentiment analysis. In: Proceedings of the 58th Annual meeting of the association for computational linguistics, association for computational linguistics, p 4019–4028. 10.18653/v1/2020a.acl-main.370
• Du Y, He M, Wang L, Zhang H. Wasserstein based transfer network for cross-domain sentiment classification. Knowl Based Syst. 2020; 204 :6162. doi: 10.1016/j.knosys.2020.106162. [ CrossRef ] [ Google Scholar ]
• Elo S, Kyngäs H. The qualitative content analysis process. J Adv Nurs. 2008; 62 (1):107–115. doi: 10.1111/j.1365-2648.2007.04569.x. [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Elshakankery K, Ahmed MF. HILATSA: a hybrid incremental learning approach for arabic tweets sentiment analysis. Egypt Inform J. 2019; 20 (3):163–171. doi: 10.1016/j.eij.2019.03.002. [ CrossRef ] [ Google Scholar ]
• Feldman R. Techniques and applications for sentiment analysis. Commun ACM. 2013; 56 (4):82–89. doi: 10.1145/2436256.2436274. [ CrossRef ] [ Google Scholar ]
• Ferilli S, De Carolis B, Esposito F, Redavid D (2015) Sentiment analysis as a text categorization task: a study on feature and algorithm selection for Italian language. In: 2015 IEEE International Conference on Data Science and Advanced Analytics (DSAA), IEEE, pp 1–10. 10.1109/DSAA.2015.7344882
• Fink C, Bos N, Perrone A, Liu E, Kopecky J (2013) Twitter, public opinion, and the 2011 Nigerian Presidential Election. In: 2013 International conference on social computing, IEEE, pp 311–320. 10.1109/SocialCom.2013.50
• Fitri VA, Andreswari R, Hasibuan MA. Sentiment analysis of social media Twitter with case of anti-LGBT campaign in Indonesia using Naïve Bayes, Decision Tree, and Random Forest Algorithm. Procedia Comput Sci. 2019; 161 :765–772. doi: 10.1016/j.procs.2019.11.181. [ CrossRef ] [ Google Scholar ]
• Garg S (2021) Drug recommendation system based on sentiment analysis of drug reviews using machine learning. In: 2021 11th International conference on cloud computing, data science & engineering (confluence), IEEE, pp 175–181. 10.1109/Confluence51648.2021.9377188
• Glorot X, Bordes A, Bengio Y (2011) Domain adaptation for large-scale sentiment classification: a deep learning approach. In: 28th International Conference on Machine Learning, International Machine Learning Society (IMLS), pp 513–520. https://dl.acm.org/doi/10.5555/3104482.3104547
• Grootendorst M, Warmerdam VD (2021) MaartenGr/KeyBERT (Version 0.5) [Computer program]. 10.5281/ZENODO.5534341.
• Groshek J, Al-Rawi A. Public sentiment and critical framing in social media content during the 2012 US Presidential Campaign. Soc Sci Comput Rev. 2013; 31 (5):563–576. doi: 10.1177/0894439313490401. [ CrossRef ] [ Google Scholar ]
• Habimana O, Li Y, Li R, Gu X, Yu G. Sentiment analysis using deep learning approaches: an overview. Sci China Inf Sci. 2020; 63 (1):1–36. doi: 10.1007/s11432-018-9941-6. [ CrossRef ] [ Google Scholar ]
• Hao Y, Mu T, Hong R, Wang M, Liu X, Goulermas JY. Cross-domain sentiment encoding through stochastic word embedding. IEEE Trans Knowl Data Eng. 2019; 32 (10):1909–1922. doi: 10.1109/TKDE.2019.2913379. [ CrossRef ] [ Google Scholar ]
• Hassan A, Mahmood A (2017) Efficient deep learning model for text classification based on recurrent and convolutional layers. In: 2017 16th IEEE International Conference on Machine Learning and Applications (ICMLA), IEEE, pp 1108–1113. 10.1109/ICMLA.2017.00009
• Heaton J (2018). Ian Goodfellow, Yoshua Bengio, and Aaron Courville: Deep Learning. Genetic Programming and Evolvable Machines 19: 305–307. 10.1007/s10710-017-9314-z
• Heikal M, Torki M, El-Makky N. Sentiment analysis of Arabic tweets using deep learning. Procedia Comput Sci. 2018; 142 :114–122. doi: 10.1016/j.procs.2018.10.466. [ CrossRef ] [ Google Scholar ]
• Huang B, Ou Y, Carley KM (2018) Aspect level sentiment classification with attention-over-attention neural networks. In: International Conference on Social Computing, Behavioral-Cultural Modeling and Prediction and Behavior Representation in Modeling and Simulation, Springer, Cham, pp 197–206. 10.1007/978-3-319-93372-6_22
• Hussain N, Mirza HT, Rasool G, Hussain I, Kaleem M. Spam review detection techniques: a systematic literature review. Appl Sci. 2019; 9 (5):987. doi: 10.3390/app9050987. [ CrossRef ] [ Google Scholar ]
• Hussein DMEDM. A survey on sentiment analysis challenges. J King Saud Univ. 2018; 30 (4):330–338. doi: 10.1016/j.jksues.2016.04.002. [ CrossRef ] [ Google Scholar ]
• Ikram MT, Afzal MT. Aspect based citation sentiment analysis using linguistic patterns for better comprehension of scientific knowledge. Scientometrics. 2019; 119 (1):73–95. doi: 10.1007/s11192-019-03028-9. [ CrossRef ] [ Google Scholar ]
• Injadat MN, Salo F, Nassif AB. Data mining techniques in social media: a survey. Neurocomputing. 2016; 214 :654–670. doi: 10.1016/j.neucom.2016.06.045. [ CrossRef ] [ Google Scholar ]
• Isah H, Trundle P, Neagu D (2014) Social media analysis for product safety using text mining and sentiment analysis. In: 2014 14th UK Workshop on Computational Intelligence (UKCI), IEEE, pp 1–7. 10.1109/UKCI.2014.6930158
• ISO 639-3 (2017) Registration Authority. https://iso639-3.sil.org/
• Jain DK, Boyapati P, Venkatesh J, Prakash M. An intelligent cognitive-inspired computing with big data analytics framework for sentiment analysis and classification. Inf Process Manag. 2022; 59 (1):2758. doi: 10.1016/j.ipm.2021.102758. [ CrossRef ] [ Google Scholar ]
• Januário BA, de Carosia AEO, da Silva AEA, Coelho GP. Sentiment analysis applied to news from the Brazilian stock market. IEEE Latin Am Trans. 2022; 20 (3):512–518. doi: 10.1109/TLA.2022.9667151. [ CrossRef ] [ Google Scholar ]
• JayaLakshmi ANM, Kishore KVK. Performance evaluation of DNN with other machine learning techniques in a cluster using apache spark and MLlib. J King Saud Univ. 2022; 34 (1):1311–1319. doi: 10.1016/j.jksuci.2018.09.022. [ CrossRef ] [ Google Scholar ]
• Jia X, Wang L. Attention enhanced capsule network for text classification by encoding syntactic dependency trees with graph convolutional neural network. PeerJ Comput Sci. 2022; 7 :e831. doi: 10.7717/PEERJ-CS.831. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Jiang D, Luo X, Xuan J, Xu Z. Sentiment computing for the news event based on the social media big data. IEEE Access. 2017; 5 :2373–2382. doi: 10.1109/ACCESS.2016.2607218. [ CrossRef ] [ Google Scholar ]
• Kaity M, Balakrishnan V. Sentiment Lexicons and non-English languages: a survey. Knowl Inf Syst. 2020; 62 (12):4445–4480. doi: 10.1007/s10115-020-01497-6. [ CrossRef ] [ Google Scholar ]
• Kastrati Z, Dalipi F, Imran AS, Nuci KP, Wani MA. Sentiment analysis of students’ feedback with Nlp and deep learning: a systematic mapping study. Appl Sci. 2021; 11 (9):3986. doi: 10.3390/app11093986. [ CrossRef ] [ Google Scholar ]
• Kaur H, Mangat V, Nidhi (2017) A survey of sentiment analysis techniques. In: 2017 International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud)(I-SMAC), IEEE, pp 921–925. 10.1109/I-SMAC.2017.8058315
• Khan K, Baharudin BB, Khan A (2009) Mining opinion from text documents: a survey. In: 2009 3rd IEEE International conference on digital ecosystems and technologies, IEEE, pp 217–222. 10.4304/jetwi.5.4.343-353
• Khasawneh RT, Wahsheh HA, Al-Kabi MN, Alsmadi IM (2013) Sentiment analysis of Arabic social media content: a comparative study. In: 8th International Conference for Internet Technology and Secured Transactions (ICITST-2013), IEEE, pp 101–106. 10.1109/ICITST.2013.6750171
• Khattak A, Asghar MZ, Saeed A, Hameed IA, Asif Hassan S, Ahmad S. A survey on sentiment analysis in Urdu: a resource-poor language. Egypt Inform J. 2021; 22 (1):53–74. doi: 10.1016/j.eij.2020.04.003. [ CrossRef ] [ Google Scholar ]
• Khatua A, Khatua A, Cambria E. Predicting political sentiments of voters from Twitter in multi-party contexts. Appl Soft Comput J. 2020; 97 :106743. doi: 10.1016/j.asoc.2020.106743. [ CrossRef ] [ Google Scholar ]
• Kitchenham B. Procedures for performing systematic reviews, version 1.0. Empir Softw Eng. 2004; 33 (2004):1–26. [ Google Scholar ]
• Kitchenham B, Charters SM. Guidelines for performing systematic literature reviews in software engineering. Tech Rep. 2007; 5 :1–57. [ Google Scholar ]
• Korayem M, Crandall D, Abdul-Mageed M (2012) Subjectivity and sentiment analysis of Arabic: a survey. In: International conference on advanced machine learning technologies and applications, Springer, Berlin, Heidelberg, p 128–139. 10.1007/978-3-642-35326-0_14
• Koto F, Adriani M (2015) A comparative study on Twitter sentiment analysis: Which Features Are Good? In: International conference on applications of natural language to information systems, Springer, Cham, p 453–457. 10.1007/978-3-319-19581-0_46
• Krippendorff K (2018) Content analysis: an introduction to its methodology. Sage publications.
• Kumar A, Garg G. Systematic literature review on context-based sentiment analysis in social multimedia. Multimed Tools Appl. 2020; 79 (21):15349–15380. doi: 10.1007/s11042-019-7346-5. [ CrossRef ] [ Google Scholar ]
• Kumar A, Jaiswal A. Systematic literature review of sentiment analysis on twitter using soft computing techniques. Concurr Comput. 2020; 32 (1):e5107. doi: 10.1002/cpe.5107. [ CrossRef ] [ Google Scholar ]
• Kumar A, Sebastian TM. Sentiment analysis: a perspective on its past, present and future. Int J Intell Syst Appl. 2012; 4 (10):1–14. doi: 10.5815/ijisa.2012.10.01. [ CrossRef ] [ Google Scholar ]
• Kumar A, Narapareddy VT, Gupta P, Srikanth VA, Neti LB, Malapati A (2021) Adversarial and auxiliary features-aware BERT for sarcasm detection. In: 8th ACM IKDD CODS and 26th COMAD, association for computing machinery, p 163–170. 10.1145/3430984.3431024
• Kydros D, Argyropoulou M, Vrana V. A content and sentiment analysis of Greek tweets during the pandemic. Sustainability (switzerland) 2021; 13 (11):6150. doi: 10.3390/su13116150. [ CrossRef ] [ Google Scholar ]
• Lai Y, Zhang L, Han D, Zhou R, Wang G. Fine-grained emotion classification of chinese microblogs based on graph convolution networks. World Wide Web. 2020; 23 (5):2771–2787. doi: 10.1007/s11280-020-00803-0. [ CrossRef ] [ Google Scholar ]
• Leiden University's Centre for Science and Technology Studies (CWTS) (2021) VOSviewer (Version 1.6.17)[Software]. https://www.vosviewer.com/
• Leydesdorff L, Park HW, Wagner C. International co-authorship relations in the social science citation index: is internationalization leading the network? J Assoc Inf Sci Technol. 2014; 65 (10):2111–2126. doi: 10.48550/arXiv.1305.4242. [ CrossRef ] [ Google Scholar ]
• Li D, Qian J (2016) Text sentiment analysis based on long short-term memory. In: 2016 First IEEE International Conference on Computer Communication and the Internet (ICCCI), IEEE, pp 471–475. 10.1109/CCI.2016.7778967
• Li F, Huang M, Zhu X (2010) Sentiment analysis with global topics and local dependency. In: Proceedings of the AAAI Conference on Artificial Intelligence, Atlanta, Georgia, USA: AAAI Press, Palo Alto, California USA, pp 1371–1376. 10.1609/aaai.v24i1.7523
• Li J, Sun M (2007) Experimental study on sentiment classification of chinese review using machine learning techniques. In: 2007 International Conference on Natural Language Processing and Knowledge Engineering, IEEE, pp 393–400. 10.1109/NLPKE.2007.4368061
• Li N, Liang X, Li X, Wang C, Wu DD. Network environment and financial risk using machine learning and sentiment analysis. Hum Ecol Risk Assess. 2009; 15 (2):227–252. doi: 10.1080/10807030902761056. [ CrossRef ] [ Google Scholar ]
• Li W, Zhu L, Shi Y, Guo K, Cambria E. User reviews: sentiment analysis using Lexicon integrated two-channel CNN–LSTM family models. Appl Soft Comput J. 2020; 94 :6435. doi: 10.1016/j.asoc.2020.106435. [ CrossRef ] [ Google Scholar ]
• Li W, Shao W, Ji S, Cambria E. BiERU: bidirectional emotional recurrent unit for conversational sentiment analysis. Neurocomputing. 2022; 467 :73–82. doi: 10.1016/j.neucom.2021.09.057. [ CrossRef ] [ Google Scholar ]
• Li Y, Pan Q, Yang T, Wang S, Tang J, Cambria E. Learning word representations for sentiment analysis. Cogn Comput. 2017; 9 (6):843–851. doi: 10.1007/s12559-017-9492-2. [ CrossRef ] [ Google Scholar ]
• Liang B, Su H, Gui L, Cambria E, Xu R. Aspect-based sentiment analysis via affective knowledge enhanced graph convolutional networks. Knowl Based Syst. 2022; 235 :107643. doi: 10.1016/j.knosys.2021.107643. [ CrossRef ] [ Google Scholar ]
• Ligthart A, Catal C, Tekinerdogan B. Systematic reviews in sentiment analysis: a tertiary study. Artif Intell Rev. 2021; 54 (7):4997–5053. doi: 10.1007/s10462-021-09973-3. [ CrossRef ] [ Google Scholar ]
• Lin B, Cassee N, Serebrenik A, Bavota G, Novielli N, Lanza M. Opinion mining for software development: a systematic literature review. ACM Trans Softw Eng Methodol. 2022; 31 (3):1–41. doi: 10.1145/3490388. [ CrossRef ] [ Google Scholar ]
• Lin Y, Li J, Yang L, Xu K, Lin H. Sentiment analysis with comparison enhanced deep neural network. IEEE Access. 2020; 8 :78378–78384. doi: 10.1109/ACCESS.2020.2989424. [ CrossRef ] [ Google Scholar ]
• Liu F, Zheng J, Zheng L, Chen C. Combining attention-based bidirectional gated recurrent neural network and two-dimensional convolutional neural network for document-level sentiment classification. Neurocomputing. 2020; 371 :39–50. doi: 10.1016/j.neucom.2019.09.012. [ CrossRef ] [ Google Scholar ]
• Liu L, Nie X, Wang H (2012) Toward a fuzzy domain sentiment ontology tree for sentiment analysis. In: 2012 5th International congress on image and signal processing, IEEE, pp 1620–1624. 10.1109/CISP.2012.6469930
• Liu R, Shi Y, Ji C, Jia M. A survey of sentiment analysis based on transfer learning. IEEE Access. 2019; 7 :85401–85412. doi: 10.1109/ACCESS.2019.2925059. [ CrossRef ] [ Google Scholar ]
• Liu S, Lee K, Lee I. Document-level multi-topic sentiment classification of email data with BiLSTM and data augmentation. Knowl Based Syst. 2020; 197 :105918. doi: 10.1016/j.knosys.2020.105918. [ CrossRef ] [ Google Scholar ]
• Liu SM, Chen JH. A multi-label classification based approach for sentiment classification. Expert Syst Appl. 2015; 42 (3):1083–1093. doi: 10.1016/j.eswa.2014.08.036. [ CrossRef ] [ Google Scholar ]
• Liu X, Zeng D, Li J, Wang F-Y, Zuo W. Sentiment analysis of Chinese documents: from sentence to document level. J Am Soc Inform Sci Technol. 2009; 60 (12):2474–2487. doi: 10.1002/asi.21206. [ CrossRef ] [ Google Scholar ]
• Lo YW, Potdar V (2009) A review of opinion mining and sentiment classification framework in social networks. In: 2009 3rd IEEE International conference on digital ecosystems and technologies, IEEE, pp 396–401. 10.1109/DEST.2009.5276705
• Lulu L, Elnagar A. Automatic arabic dialect classification using deep learning models. Procedia Comput Sci. 2018; 142 :262–269. doi: 10.1016/j.procs.2018.10.489. [ CrossRef ] [ Google Scholar ]
• Ma Y, Peng H, Cambria E (2018) Targeted aspect-based sentiment analysis via embedding commonsense knowledge into an attentive LSTM. In: 32nd AAAI conference on artificial intelligence, New Orleans, Louisiana, USA: AAAI Press, Palo Alto, California USA, pp 5876–5883. 10.1609/aaai.v32i1.12048
• Malandri L, Porcel C, Xing F, Serrano-Guerrero J, Cambria E. Soft computing for recommender systems and sentiment analysis. Appl Soft Comput. 2022 doi: 10.1016/j.asoc.2021.108246. [ CrossRef ] [ Google Scholar ]
• Mäntylä MV, Graziotin D, Kuutila M. The evolution of sentiment analysis-a review of research topics, venues and top cited papers. Comput Sci Rev. 2018; 27 :16–32. doi: 10.1016/j.cosrev.2017.10.002. [ CrossRef ] [ Google Scholar ]
• Mao Y, Zhang Y, Jiao L, Zhang H. Document-level sentiment analysis using attention-based bi-directional long short-term memory network and two-dimensional convolutional neural network. Electronics. 2022; 11 (12):1906. doi: 10.3390/electronics11121906. [ CrossRef ] [ Google Scholar ]
• Maqsood H, Mehmood I, Maqsood M, Yasir M, Afzal S, Aadil F, et al. A local and global event sentiment based efficient stock exchange forecasting using deep learning. Int J Inf Manag. 2020; 50 :432–451. doi: 10.1016/j.ijinfomgt.2019.07.011. [ CrossRef ] [ Google Scholar ]
• Martinez-Camara E, Martin-Valdivia MT, Urena-Lopez LA (2011) Opinion classification techniques applied to a Spanish Corpus. In: International conference on application of natural language to information systems, Springer, Berlin, Heidelberg, pp 169–176. 10.1007/978-3-642-22327-3_17
• Martinez-Garcia A, Badia T, Barnes J (2021) Evaluating morphological typology in zero-shot cross-lingual transfer. In: Proceedings of the 59th annual meeting of the association for computational linguistics and the 11th international joint conference on natural language processing, association for computational linguistics, pp 3136–3153. 10.18653/v1/2021.acl-long.244
• Medhat W, Hassan A, Korashy H. Sentiment analysis algorithms and applications: a survey. Ain Shams Eng J. 2014; 5 (4):1093–1113. doi: 10.1016/j.asej.2014.04.011. [ CrossRef ] [ Google Scholar ]
• Momtazi S (2012) Fine-grained German sentiment analysis on social media. In: Proceedings of the 8th International conference on language resources and evaluation (LREC’12), European Language Resources Association (ELRA), pp 1215–1220. http://www.lrec-conf.org/proceedings/lrec2012/pdf/999_Paper.pdf
• Myslin M, Zhu SH, Chapman W, Conway M. Using Twitter to examine smoking behavior and perceptions of emerging tobacco products. J Med Int Res. 2013; 15 (8):174. doi: 10.2196/jmir.2534. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Nair RR, Mathew J, Muraleedharan V, Deepa Kanmani S (2019) Study of machine learning techniques for sentiment analysis. In: 2019 3rd International Conference on Computing Methodologies and Communication (ICCMC), IEEE, pp 978–984. 10.1109/ICCMC.2019.8819763
• Nassif AB, Elnagar A, Shahin I, Henno S. Deep learning for Arabic subjective sentiment analysis: challenges and research opportunities. Appl Soft Comput. 2021; 98 :6836. doi: 10.1016/j.asoc.2020.106836. [ CrossRef ] [ Google Scholar ]
• Nassirtoussi AK, Aghabozorgi S, Wah TY, Ngo DCL. Text mining for market prediction: a systematic review. Expert Syst Appl. 2014; 41 (16):7653–7670. doi: 10.1016/j.eswa.2014.06.009. [ CrossRef ] [ Google Scholar ]
• Nejat B, Carenini G, Ng R (2017) Exploring joint neural model for sentence level discourse parsing and sentiment analysis. In: Proceedings of the 18th annual sigdial meeting on discourse and dialogue, association for computational linguistics, pp 289–298. 10.18653/v1/w17-5535
• Nhlabano VV, Lutu PEN (2018). Impact of text pre-processing on the performance of sentiment analysis models for social media data. In: 2018 International Conference on Advances in Big Data, Computing and Data Communication Systems (IcABCD), IEEE, pp 1–6. 10.1109/ICABCD.2018.8465135
• Nicholls C, Song F (2010) Comparison of feature selection methods for sentiment analysis. In: Canadian conference on artificial intelligence, Springer, Berlin, Heidelberg, pp 286–289. 10.1007/978-3-319-96292-4_21
• Nielsen FA (2011) A New ANEW: Evaluation of a Word List for Sentiment Analysis in Microblogs. In: Proceedings of the ESWC2011 workshop on “Making Sense of Microposts”: big things come in small packages, Heraklion, Crete, Greece: CEUR-WS, Aachen, pp 93–98. 10.48550/arXiv.1103.2903
• Obiedat R, Al-Darras D, Alzaghoul E, Harfoushi O. Arabic aspect-based sentiment analysis: a systematic literature review. IEEE Access. 2021; 9 :152628–152645. doi: 10.1109/ACCESS.2021.3127140. [ CrossRef ] [ Google Scholar ]
• Ombabi AH, Ouarda W, Alimi AM. Deep learning CNN–LSTM framework for Arabic sentiment analysis using textual information shared in social networks. Soc Netw Anal Min. 2020; 10 (1):1–13. doi: 10.1007/s13278-020-00668-1. [ CrossRef ] [ Google Scholar ]
• Oueslati O, Cambria E, Ben HM, Ounelli H. A review of sentiment analysis research in Arabic language. Futur Gener Comput Syst. 2020; 112 :408–430. doi: 10.1016/j.future.2020.05.034. [ CrossRef ] [ Google Scholar ]
• Ouyang X, Zhou P, Li CH, Liu L (2015) Sentiment Analysis Using Convolutional Neural Network. In: 2015 IEEE International conference on computer and information technology; ubiquitous computing and communications; dependable, autonomic and secure computing; pervasive intelligence and computing, IEEE, p 2359–2364. 10.1109/CIT/IUCC/DASC/PICOM.2015.349
• Pecore S, Villaneau J (2019) Complex and Precise Movie and Book Annotations in French Language for Aspect Based Sentiment Analysis. In: LREC 2018—11th International conference on language resources and evaluation, European Language Resources Association (ELRA), p 2647–2652. https://aclanthology.org/L18-1419
• Peng H, Cambria E, Hussain A. A review of sentiment analysis research in Chinese language. Cogn Comput. 2017; 9 (4):423–435. doi: 10.1007/s12559-017-9470-8. [ CrossRef ] [ Google Scholar ]
• Peng H, Ma Y, Li Y, Cambria E. Learning multi-grained aspect target sequence for Chinese sentiment analysis. Knowl Based Syst. 2018; 148 :167–176. doi: 10.1016/j.knosys.2018.02.034. [ CrossRef ] [ Google Scholar ]
• Pereira DA. A survey of sentiment analysis in the Portuguese language. Artif Intell Rev. 2021; 54 (2):1087–1115. doi: 10.1007/s10462-020-09870-1. [ CrossRef ] [ Google Scholar ]
• Perianes-Rodriguez A, Waltman L, van Eck NJ. Constructing bibliometric networks: a comparison between full and fractional counting. J Informetr. 2016; 10 (4):1178–1195. doi: 10.1016/j.joi.2016.10.006. [ CrossRef ] [ Google Scholar ]
• Persson O (2017) BibExcel [Software]. Available from https://homepage.univie.ac.at/juan.gorraiz/bibexcel/
• Persson O, Danell R, Schneider JW (2009) How to Use Bibexcel for Various Types of Bibliometric Analysis. In: Celebrating scholarly communication studies: a festschrift for Olle Persson at his 60th birthday, ed. J. Schneider F. Åström, R. Danell, B. Larsen. Leuven, Belgium: International Society for Scientometrics and Informetrics, pp 9–24
• Picasso A, Merello S, Ma Y, Oneto L, Cambria E. Technical analysis and sentiment embeddings for market trend prediction. Expert Syst Appl. 2019; 135 :60–70. doi: 10.1016/j.eswa.2019.06.014. [ CrossRef ] [ Google Scholar ]
• Piryani R, Madhavi D, Singh VK. Analytical mapping of opinion mining and sentiment analysis research during 2000–2015. Inf Process Manag. 2017; 53 (1):122–150. doi: 10.1016/j.ipm.2016.07.001. [ CrossRef ] [ Google Scholar ]
• Piryani R, Piryani B, Singh VK, Pinto D. Sentiment analysis in Nepali: exploring machine learning and lexicon-based approaches. J Intell Fuzzy Syst. 2020; 39 (2):2201–2212. doi: 10.3233/JIFS-179884. [ CrossRef ] [ Google Scholar ]
• Plaza-del-Arco FM, Martín-Valdivia MT, Ureña-López LA, Mitkov R. Improved emotion recognition in spanish social media through incorporation of lexical knowledge. Futur Gener Comput Syst. 2020; 110 :1000–1008. doi: 10.1016/j.future.2019.09.034. [ CrossRef ] [ Google Scholar ]
• Poria S, Cambria E, Gelbukh A. Aspect extraction for opinion mining with a deep convolutional neural network. Knowl Based Syst. 2016; 108 :42–49. doi: 10.1016/j.knosys.2016.06.009. [ CrossRef ] [ Google Scholar ]
• Prabha MI, Srikanth GU (2019). Survey of Sentiment Analysis Using Deep Learning Techniques. In: 2019 1st International Conference on Innovations in Information and Communication Technology (ICIICT), IEEE, p 1–9. 10.1109/ICIICT1.2019.8741438
• Prabhat A, Khullar V (2017). Sentiment Classification on Big Data Using Naïve Bayes and Logistic Regression. In: 2017 International Conference on Computer Communication and Informatics (ICCCI), IEEE, p 1–5. 10.1109/ICCCI.2017.8117734
• Preethi PG, Uma V, Kumar A. Temporal sentiment analysis and causal rules extraction from Tweets for event prediction. Procedia Comput Sci. 2015; 48 :84–89. doi: 10.1016/j.procs.2015.04.154. [ CrossRef ] [ Google Scholar ]
• Qasem M, Thulasiram R, Thulasiram P (2015) Twitter Sentiment Classification Using Machine Learning Techniques for Stock Markets. In: 2015 International Conference on Advances in Computing, Communications and Informatics (ICACCI), IEEE, p 834–840. 10.1109/ICACCI.2015.7275714
• Qazi A, Fayaz H, Wadi A, Raj RG, Rahim NA, Khan WA. The artificial neural network for solar radiation prediction and designing solar systems: a systematic literature review. J Clean Prod. 2015; 104 :1–12. doi: 10.1016/j.jclepro.2015.04.041. [ CrossRef ] [ Google Scholar ]
• Qazi A, Raj RG, Hardaker G, Standing C. A systematic literature review on opinion types and sentiment analysis techniques: tasks and challenges. Internet Res. 2017; 27 (3):608–630. doi: 10.1108/IntR-04-2016-0086. [ CrossRef ] [ Google Scholar ]
• Raghuvanshi N, Patil JM (2016) A Brief Review on Sentiment Analysis. In: 2016 International Conference on Electrical, Electronics, and Optimization Techniques (ICEEOT), IEEE, p 2827–2831. 10.1109/ICEEOT.2016.7755213
• Rambocas M, Pacheco BG. Online sentiment analysis in marketing research: a review. J Res Interact Mark. 2018; 12 (2):146–163. doi: 10.1108/JRIM-05-2017-0030. [ CrossRef ] [ Google Scholar ]
• Rao G, Gu X, Feng Z, Cong Q, Zhang L (2021) A Novel Joint Model with Second-Order Features and Matching Attention for Aspect-Based Sentiment Analysis. In: 2021 International Joint Conference on Neural Networks (IJCNN), IEEE, p 1–8. 10.1109/IJCNN52387.2021.9534321
• Ravi K, Ravi V. A survey on opinion mining and sentiment analysis: tasks, approaches and applications. Knowl Based Syst. 2015; 89 :14–46. doi: 10.1016/j.knosys.2015.06.015. [ CrossRef ] [ Google Scholar ]
• Rotta R, Noack A. Multilevel local search algorithms for modularity clustering. ACM J Exp Algorithmics. 2011; 16 (2):1–27. doi: 10.1145/1963190.1970376. [ CrossRef ] [ Google Scholar ]
• Sadamitsu K, Sekine S, Yamamoto M (2008) Sentiment Analysis Based on Probabilistic Models Using Inter-Sentence Information. In: Proceedings of the sixth international conference on language resources and evaluation (LREC’08), European Language Resources Association (ELRA), p 2892–2896. http://www.lrec-conf.org/proceedings/lrec2008/pdf/736_paper.pdf
• Salur MU, Aydin I. A novel hybrid deep learning model for sentiment classification. IEEE Access. 2020; 8 :58080–58093. doi: 10.1109/ACCESS.2020.2982538. [ CrossRef ] [ Google Scholar ]
• Sánchez-Rada JF, Iglesias CA. Social context in sentiment analysis: formal definition, overview of current trends and framework for comparison. Inf Fusion. 2019; 52 :344–356. doi: 10.1016/j.inffus.2019.05.003. [ CrossRef ] [ Google Scholar ]
• Santos R, Costa AA, Silvestre JD, Pyl L. Informetric analysis and review of literature on the role of BIM in sustainable construction. Autom Constr. 2019; 103 :221–234. doi: 10.1016/j.autcon.2019.02.022. [ CrossRef ] [ Google Scholar ]
• Sari IC, Ruldeviyani Y (2020) Sentiment Analysis of the Covid-19 Virus Infection in Indonesian Public Transportation on Twitter Data: A Case Study of Commuter Line Passengers. In: 2020 International Workshop on Big Data and Information Security (IWBIS), IEEE, pp 23–28. 10.1109/IWBIS50925.2020.9255531
• Sarsam SM, Al-Samarraie H, Alzahrani AI, Wright B. Sarcasm detection using machine learning algorithms in Twitter: a systematic review. Int J Mark Res. 2020; 62 (5):578–598. doi: 10.1177/1470785320921779. [ CrossRef ] [ Google Scholar ]
• Sayed AA, Elgeldawi E, Zaki AM, Galal AR (2020) Sentiment Analysis for Arabic Reviews Using Machine Learning Classification Algorithms. In: 2020 International Conference on Innovative Trends in Communication and Computer Engineering (ITCE), IEEE, p 56–63. 10.1109/ITCE48509.2020.9047822
• Schouten K, Frasincar F. Survey on aspect-level sentiment analysis. IEEE Trans Knowl Data Eng. 2015; 28 (3):813–830. doi: 10.1109/TKDE.2015.2485209. [ CrossRef ] [ Google Scholar ]
• Schuller B, Mousa AED, Vryniotis V. Sentiment analysis and opinion mining: on optimal parameters and performances. Wiley Interdiscip Rev. 2015; 5 (5):255–263. doi: 10.1002/widm.1159. [ CrossRef ] [ Google Scholar ]
• Serrano-Guerrero J, Romero FP, Olivas JA. Fuzzy logic applied to opinion mining: a review. Knowl Based Syst. 2021; 222 :107018. doi: 10.1016/j.knosys.2021.107018. [ CrossRef ] [ Google Scholar ]
• Sharma S, Jain A. Role of sentiment analysis in social media security and analytics. Wiley Interdiscip Rev. 2020; 10 (5):e1366. doi: 10.1002/widm.1366. [ CrossRef ] [ Google Scholar ]
• Shirsat VS, Jagdale RS, Deshmukh SN (2018) Document Level Sentiment Analysis from News Articles. In: 2017 International Conference on Computing, Communication, Control and Automation (ICCUBEA), IEEE, pp 1–4. 10.1109/ICCUBEA.2017.8463638
• Shofiya C, Abidi S. Sentiment analysis on Covid-19-related social distancing in Canada using Twitter data. Int J Environ Res Public Health. 2021; 18 (11):5993. doi: 10.3390/ijerph18115993. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Singh RK, Sachan MK, Patel RB. 360 Degree view of cross-domain opinion classification: a survey. Artif Intell Rev. 2021; 54 (2):1385–1506. doi: 10.1007/s10462-020-09884-9. [ CrossRef ] [ Google Scholar ]
• Singh T, Kumari M. Role of text pre-processing in twitter sentiment analysis. Procedia Comput Sci. 2016; 89 :549–554. doi: 10.1016/j.procs.2016.06.095. [ CrossRef ] [ Google Scholar ]
• Smailović J, Grčar M, Lavrač N, Žnidaršič M. Stream-based active learning for sentiment analysis in the financial domain. Inf Sci. 2014; 285 (1):181–203. doi: 10.1016/j.ins.2014.04.034. [ CrossRef ] [ Google Scholar ]
• Smetanin S. The applications of sentiment analysis for Russian language texts: current challenges and future perspectives. IEEE Access. 2020; 8 :110693–110719. doi: 10.1109/ACCESS.2020.3002215. [ CrossRef ] [ Google Scholar ]
• Stemler S. An overview of content analysis. Pract Assess Res Eval. 2000; 7 (1):1–16. doi: 10.1362/146934703771910080. [ CrossRef ] [ Google Scholar ]
• Sutoyo E, Rifai AP, Risnumawan A, Saputra M. A comparison of text weighting schemes on sentiment analysis of government policies: a case study of replacement of national examinations. Multimed Tools Appl. 2022; 81 (5):6413–6431. doi: 10.1007/s11042-022-11900-9. [ CrossRef ] [ Google Scholar ]
• Syed AZ, Aslam M, Martinez-Enriquez AM (2010) Lexicon Based Sentiment Analysis of Urdu Text Using SentiUnits. In: Mexican international conference on artificial intelligence, Springer, Berlin, Heidelberg, pp 32–43. 10.1007/978-3-642-16761-4_4
• Taboada M. Sentiment analysis: an overview from linguistics. Annu Rev Linguist. 2016; 2 :325–347. doi: 10.1146/annurev-linguistics-011415-040518. [ CrossRef ] [ Google Scholar ]
• Tai KS, Socher R, Manning CD (2015) Improved Semantic Representations from Tree-Structured Long Short-Term Memory Networks. In: Proceedings of the 53rd Annual meeting of the association for computational linguistics and the 7th international joint conference on natural language processing, association for computational linguistics, pp 1556–1566. 10.3115/v1/p15-1150
• Tammina S (2020) A Hybrid Learning Approach for Sentiment Classification in Telugu Language. In: 2020 International conference on Artificial Intelligence and Signal Processing (AISP), IEEE, p 1–6. 10.1109/AISP48273.2020.9073109
• Tan S, Cheng X, Wang Y, Xu H (2009) Adapting Naive Bayes to Domain Adaptation for Sentiment Analysis. In: European Conference on Information Retrieval, Springer, Berlin, Heidelberg, p 337–349. 10.1007/978-3-642-00958-7_31
• Tan X, Cai Y, Xu J, Leung H-F, Chen W, Li Q. Improving aspect-based sentiment analysis via aligning aspect embedding. Neurocomputing. 2020; 383 :336–347. doi: 10.1016/j.neucom.2019.12.035. [ CrossRef ] [ Google Scholar ]
• Tembhurne JV, Diwan T. Sentiment analysis in textual, visual and multimodal inputs using recurrent neural networks. Multimed Tools Appl. 2021; 80 (5):6871–6910. doi: 10.1007/s11042-020-10037-x. [ CrossRef ] [ Google Scholar ]
• Thakur RK, Deshpande MV. Kernel optimized-support vector machine and mapreduce framework for sentiment classification of train reviews. Int J Uncertain Fuzziness Knowl Based Syst. 2019; 27 (6):1025–1050. doi: 10.1142/S0218488519500454. [ CrossRef ] [ Google Scholar ]
• Thelwall M, Buckley K, Paltoglou G. Sentiment strength detection for the social web. J Am Soc Inform Sci Technol. 2012; 63 (1):163–173. doi: 10.1002/asi.21662. [ CrossRef ] [ Google Scholar ]
• Thet TT, Na JC, Khoo CSG. Aspect-based sentiment analysis of movie reviews on discussion boards. J Inf Sci. 2010; 36 (6):823–848. doi: 10.1177/0165551510388123. [ CrossRef ] [ Google Scholar ]
• Trilla A, Alías F (2009) Sentiment Classification in English from Sentence-Level Annotations of Emotions Regarding Models of Affect. In: 10th Annual Conference of the International Speech Communication Association, International Speech Communication Association (ISCA), p 516–519. 10.21437/interspeech.2009-189
• Trisna KW, Jie HJ. Deep learning approach for aspect-based sentiment classification: a comparative review. Appl Artif Intell. 2022 doi: 10.1080/08839514.2021.2014186. [ CrossRef ] [ Google Scholar ]
• Valverde-Albacete FJ, Carrillo-de-Albornoz J, Peláez-Moreno C (2013) A Proposal for New Evaluation Metrics and Result Visualization Technique for Sentiment Analysis Tasks. In: International conference of the cross-language evaluation forum for European languages, Springer, Berlin, Heidelberg, p 41–52. 10.1007/978-3-642-40802-1_5
• Van Eck NJ, Waltman L. Software survey: VOSviewer, a computer program for bibliometric mapping. Scientometrics. 2010; 84 (2):523–538. doi: 10.1007/s11192-009-0146-3. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Verma S. Sentiment analysis of public services for smart society: literature review and future research directions. Gov Inf Quart. 2022; 39 (3):101708. doi: 10.1016/j.giq.2022.101708. [ CrossRef ] [ Google Scholar ]
• Waila P, Marisha S, Singh VK, Singh MK (2012) Evaluating Machine Learning and Unsupervised Semantic Orientation Approaches for Sentiment Analysis of Textual Reviews. In: 2012 IEEE International conference on computational intelligence and computing research, IEEE, pp 1–6. 10.1109/ICCIC.2012.6510235
• Waltman L, Van Eck NJ. A smart local moving algorithm for large-scale modularity-based community detection. Eur Phys J B. 2013; 86 (11):1–33. doi: 10.1140/epjb/e2013-40829-0. [ CrossRef ] [ Google Scholar ]
• Waltman L, Van Eck NJ, Noyons ECM. A unified approach to mapping and clustering of bibliometric networks. J Inform. 2010; 4 (4):629–635. doi: 10.1016/j.joi.2010.07.002. [ CrossRef ] [ Google Scholar ]
• Wang C, Yang X, Ding L. Deep learning sentiment classification based on weak tagging information. IEEE Access. 2021; 9 :66509–66518. doi: 10.1109/ACCESS.2021.3077059. [ CrossRef ] [ Google Scholar ]
• Wang L, Wan Y (2011) Sentiment Classification of Documents Based on Latent Semantic Analysis. In: International conference on computer education, simulation and modeling, Springer, Berlin, Heidelberg, p 356–361. 10.1007/978-3-642-21802-6_57
• Wang T, Lu K, Chow KP, Zhu Q. COVID-19 sensing: negative sentiment analysis on social media in China via BERT model. IEEE Access. 2020; 8 :138162–138169. doi: 10.1109/ACCESS.2020.3012595. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Wang Z, Chong CS, Lan L, Yang Y, Ho S-B, Tong JC (2016) Fine-Grained Sentiment Analysis of Social Media with Emotion Sensing. In: 2016 Future Technologies Conference (FTC), IEEE, pp 1361–1364. 10.1109/FTC.2016.7821783
• Wang Z, Ho S-B, Cambria E. A review of emotion sensing: categorization models and algorithms. Multimed Tools Appl. 2020; 79 (47):35553–35582. doi: 10.1007/s11042-019-08328-z. [ CrossRef ] [ Google Scholar ]
• Wang Z, Ho S-B, Cambria E. Multi-level fine-scaled sentiment sensing with ambivalence handling. Int J Uncertain Fuzziness Knowl-Based Syst. 2020; 28 (4):683–697. doi: 10.1142/S0218488520500294. [ CrossRef ] [ Google Scholar ]
• Wang Z, Lin Z. Optimal feature selection for learning-based algorithms for sentiment classification. Cogn Comput. 2020; 12 (1):238–248. doi: 10.1007/s12559-019-09669-5. [ CrossRef ] [ Google Scholar ]
• Wang Z, Tong VJC, Chan D (2014) Issues of Social Data Analytics with a New Method for Sentiment Analysis of Social Media Data. In: 2014 IEEE 6th International conference on cloud computing technology and science, IEEE, pp 899–904. 10.1109/CloudCom.2014.40
• Wang ZY, Li G, Li CY, Li A. Research on the semantic-based co-word analysis. Scientometrics. 2012; 90 (3):855–875. doi: 10.1007/s11192-011-0563-y. [ CrossRef ] [ Google Scholar ]
• Wankhade M, Rao ACS, Kulkarni C. A survey on sentiment analysis methods, applications, and challenges. Artif Intell Rev. 2022; 55 :5731–5780. doi: 10.1007/s10462-022-10144-1. [ CrossRef ] [ Google Scholar ]
• Xing FZ, Cambria E, Welsch RE. Natural language based financial forecasting: a survey. Artif Intell Rev. 2018; 50 (1):49–73. doi: 10.1007/s10462-017-9588-9. [ CrossRef ] [ Google Scholar ]
• Xing FZ, Pallucchini F, Cambria E. Cognitive-inspired domain adaptation of sentiment lexicons. Inf Process Manage. 2019; 56 (3):554–564. doi: 10.1016/j.ipm.2018.11.002. [ CrossRef ] [ Google Scholar ]
• Xiong Z, Qin K, Yang H, Luo G. Learning Chinese word representation better by cascade morphological N-Gram. Neural Comput Appl. 2021; 33 (8):3757–3768. doi: 10.1007/s00521-020-05198-7. [ CrossRef ] [ Google Scholar ]
• Yang B, Shao B, Wu L, Lin X. Multimodal sentiment analysis with unidirectional modality translation. Neurocomputing. 2022; 467 :130–137. doi: 10.1016/j.neucom.2021.09.041. [ CrossRef ] [ Google Scholar ]
• Yang L, Li Y, Wang J, Sherratt RS. Sentiment analysis for E-commerce product reviews in Chinese based on sentiment lexicon and deep learning. IEEE Access. 2020; 8 :23522–23530. doi: 10.1109/ACCESS.2020.2969854. [ CrossRef ] [ Google Scholar ]
• Yang M, Qu Q, Shen Y, Lei K, Zhu J. Cross-domain aspect/sentiment-aware abstractive review summarization by combining topic modeling and deep reinforcement learning. Neural Comput Appl. 2020; 32 (11):6421–6433. doi: 10.1007/s00521-018-3825-2. [ CrossRef ] [ Google Scholar ]
• Yi J, Niblack W (2005) Sentiment Mining in WebFountain. In: 21st International Conference on Data Engineering (ICDE’05), IEEE, p 1073–1083. 10.1109/ICDE.2005.132
• Yin H, Yang S, Li J (2020) Detecting Topic and Sentiment Dynamics Due to COVID-19 Pandemic Using Social Media. In: International conference on advanced data mining and applications, Springer, Cham, p 610–623. 10.1007/978-3-030-65390-3_46
• You L, Li Y, Wang Y, Zhang J, Yang Y (2016) A deep learning-based RNNs model for automatic security audit of short messages. In: 2016 16th International Symposium on Communications and Information Technologies (ISCIT), IEEE, p 225–229. 10.1109/ISCIT.2016.7751626
• You T, Yoon J, Kwon O-H, Jung W-S. Tracing the evolution of physics with a keyword co-occurrence network. J Korean Phys Soc. 2021; 78 (3):236–243. doi: 10.1007/s40042-020-00051-5. [ CrossRef ] [ Google Scholar ]
• Yu J, Jiang J, Xia R. Entity-sensitive attention and fusion network for entity-level multimodal sentiment classification. IEEE/ACM Trans Audio Speech Lang Process. 2019; 28 :429–439. doi: 10.1109/TASLP.2019.2957872. [ CrossRef ] [ Google Scholar ]
• Yuan JH, Wu Y, Lu X, Zhao YY, Qin B, Liu T. Recent advances in deep learning based sentiment analysis. Sci China Technol Sci. 2020; 63 (10):1947–1970. doi: 10.1007/s11431-020-1634-3. [ CrossRef ] [ Google Scholar ]
• Yue L, Chen W, Li X, Zuo W, Yin M. A survey of sentiment analysis in social media. Knowl Inf Syst. 2019; 60 (2):617–663. doi: 10.1007/s10115-018-1236-4. [ CrossRef ] [ Google Scholar ]
• Yurtalan G, Koyuncu M, Turhan Ç. A polarity calculation approach for lexicon-based Turkish sentiment analysis. Turk J Electr Eng Comput Sci. 2019; 27 (2):1325–1339. doi: 10.3906/elk-1803-92. [ CrossRef ] [ Google Scholar ]
• Zhang L, Wang S, Liu B. Deep learning for sentiment analysis: a survey. Wiley Interdiscip Rev. 2018; 8 (4):e1253. doi: 10.1002/widm.1253. [ CrossRef ] [ Google Scholar ]
• Zhang Yin, Du J, Ma X, Wen H, Fortino G. Aspect-based sentiment analysis for user reviews. Cogn Comput. 2021; 13 (5):1114–1127. doi: 10.1007/s12559-021-09855-4. [ CrossRef ] [ Google Scholar ]
• Zhang Y, Zhang Z, Miao D, Wang J. Three-way enhanced convolutional neural networks for sentence-level sentiment classification. Inf Sci. 2019; 477 :55–64. doi: 10.1016/j.ins.2018.10.030. [ CrossRef ] [ Google Scholar ]
• Zhao N, Gao H, Wen X, Li H. Combination of convolutional neural network and gated recurrent unit for aspect-based sentiment analysis. IEEE Access. 2021; 9 :15561–15569. doi: 10.1109/ACCESS.2021.3052937. [ CrossRef ] [ Google Scholar ]
• Zhou J, Ye J. Sentiment analysis in education research: a review of journal publications. Interact Learn Environ. 2020 doi: 10.1080/10494820.2020.1826985. [ CrossRef ] [ Google Scholar ]
• Zucco C, Calabrese B, Agapito G, Guzzi PH, Cannataro M. Sentiment analysis for mining texts and social networks data: methods and tools. Wiley Interdiscip Rev. 2020; 10 (1):e1333. doi: 10.1002/widm.1333. [ CrossRef ] [ Google Scholar ]
• Zunic A, Corcoran P, Spasic I. Sentiment analysis in health and well-being: systematic review. JMIR Med Inform. 2020; 8 (1):e16023. doi: 10.2196/16023. [ PMC free article ] [ PubMed ] [ CrossRef ] [ Google Scholar ]
• Zuo E, Zhao H, Chen B, Chen Q. Context-specific heterogeneous graph convolutional network for implicit sentiment analysis. IEEE Access. 2020; 8 :37967–37975. doi: 10.1109/ACCESS.2020.2975244. [ CrossRef ] [ Google Scholar ]

Help | Advanced Search

Computer Science > Machine Learning

Title: the muse 2022 multimodal sentiment analysis challenge: humor, emotional reactions, and stress.

Abstract: The Multimodal Sentiment Analysis Challenge (MuSe) 2022 is dedicated to multimodal sentiment and emotion recognition. For this year's challenge, we feature three datasets: (i) the Passau Spontaneous Football Coach Humor (Passau-SFCH) dataset that contains audio-visual recordings of German football coaches, labelled for the presence of humour; (ii) the Hume-Reaction dataset in which reactions of individuals to emotional stimuli have been annotated with respect to seven emotional expression intensities, and (iii) the Ulm-Trier Social Stress Test (Ulm-TSST) dataset comprising of audio-visual data labelled with continuous emotion values (arousal and valence) of people in stressful dispositions. Using the introduced datasets, MuSe 2022 2022 addresses three contemporary affective computing problems: in the Humor Detection Sub-Challenge (MuSe-Humor), spontaneous humour has to be recognised; in the Emotional Reactions Sub-Challenge (MuSe-Reaction), seven fine-grained `in-the-wild' emotions have to be predicted; and in the Emotional Stress Sub-Challenge (MuSe-Stress), a continuous prediction of stressed emotion values is featured. The challenge is designed to attract different research communities, encouraging a fusion of their disciplines. Mainly, MuSe 2022 targets the communities of audio-visual emotion recognition, health informatics, and symbolic sentiment analysis. This baseline paper describes the datasets as well as the feature sets extracted from them. A recurrent neural network with LSTM cells is used to set competitive baseline results on the test partitions for each sub-challenge. We report an Area Under the Curve (AUC) of .8480 for MuSe-Humor; .2801 mean (from 7-classes) Pearson's Correlations Coefficient for MuSe-Reaction, as well as .4931 Concordance Correlation Coefficient (CCC) and .4761 for valence and arousal in MuSe-Stress, respectively.

Submission history

Access paper:.

• Other Formats

References & Citations

• Google Scholar
• Semantic Scholar

BibTeX formatted citation

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

• Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

Subscribe to the PwC Newsletter

Join the community, add a new evaluation result row, sentiment analysis.

1283 papers with code • 43 benchmarks • 91 datasets

Sentiment Analysis is the task of classifying the polarity of a given text. For instance, a text-based tweet can be categorized into either "positive", "negative", or "neutral". Given the text and accompanying labels, a model can be trained to predict the correct sentiment.

Sentiment Analysis techniques can be categorized into machine learning approaches, lexicon-based approaches, and even hybrid methods. Some subcategories of research in sentiment analysis include: multimodal sentiment analysis, aspect-based sentiment analysis, fine-grained opinion analysis, language specific sentiment analysis.

More recently, deep learning techniques, such as RoBERTa and T5, are used to train high-performing sentiment classifiers that are evaluated using metrics like F1, recall, and precision. To evaluate sentiment analysis systems, benchmark datasets like SST, GLUE, and IMDB movie reviews are used.

Further readings:

• Sentiment Analysis Based on Deep Learning: A Comparative Study

Benchmarks Add a Result

Latest papers with no code

M2sa: multimodal and multilingual model for sentiment analysis of tweets.

no code yet • 2 Apr 2024

In recent years, multimodal natural language processing, aimed at learning from diverse data types, has garnered significant attention.

Two Heads are Better than One: Nested PoE for Robust Defense Against Multi-Backdoors

In this paper, we propose Nested Product of Experts(NPoE) defense framework, which involves a mixture of experts (MoE) as a trigger-only ensemble within the PoE defense framework to simultaneously defend against multiple trigger types.

Creating emoji lexica from unsupervised sentiment analysis of their descriptions

no code yet • 1 Apr 2024

Online media, such as blogs and social networking sites, generate massive volumes of unstructured data of great interest to analyze the opinions and sentiments of individuals and organizations.

From attention to profit: quantitative trading strategy based on transformer

no code yet • 30 Mar 2024

In traditional quantitative trading practice, navigating the complicated and dynamic financial market presents a persistent challenge.

TRABSA: Interpretable Sentiment Analysis of Tweets using Attention-based BiLSTM and Twitter-RoBERTa

Sentiment analysis is crucial for understanding public opinion and consumer behavior.

Entertainment chatbot for the digital inclusion of elderly people without abstraction capabilities

no code yet • 29 Mar 2024

Many of them relieve their loneliness with traditional media such as TV and radio, which are known to create a feeling of companionship.

A Hybrid Approach To Aspect Based Sentiment Analysis Using Transfer Learning

no code yet • 25 Mar 2024

The approach focuses on generating weakly-supervised annotations by exploiting the strengths of both large language models (LLM) and traditional syntactic dependencies.

Large language models for crowd decision making based on prompt design strategies using ChatGPT: models, analysis and challenges

no code yet • 22 Mar 2024

This paper analyzes the use of ChatGPT based on prompt design strategies to assist in CDM processes to extract opinions and make decisions.

Extracting Emotion Phrases from Tweets using BART

no code yet • 21 Mar 2024

Sentiment analysis is a natural language processing task that aims to identify and extract the emotional aspects of a text.

On Prompt Sensitivity of ChatGPT in Affective Computing

no code yet • 20 Mar 2024

Recent studies have demonstrated the emerging capabilities of foundation models like ChatGPT in several fields, including affective computing.