On the frontiers of Twitter data and sentiment analysis in election prediction: a review

Stage 1: Screening

We collected 250 articles focusing on the elections of the USA 2008, Arab Spring 2010, USA 2012, Pakistan 2013, India 2014, USA 2016, and Pakistan 2018 from various databases such as IEEE, Springer, Emerald Insight, Science Direct, Scopus, and Association for Computing Machinery (ACM). The search criteria were based on keywords such as sentiment analysis, predicting election results and election prediction classification using social media, election prediction using sentiment analysis, election prediction using Twitter data, sentiment analysis using Twitter data, and social network analysis through sentiment analysis. The resultant articles were then analyzed based on the title and abstract of the articles. After the analysis, only those papers that directly correlated with election prediction and had valid digital object identifiers (DOI) were selected. In doing so, 162 articles were selected by the end of the screening process.

Stage 2: Eligibility analysis

After the screening process, publications related to analysis performed with other data sets than Twitter, like Facebook, surveys, and papers whose purpose was not to use Twitter as a predictive system, were excluded from our repository. So, the final number of eligible articles was further narrowed down to 80 papers as the study was based on election prediction using Twitter data. Thus, the resultant repository was quite reasonable for making conclusions and inferences about the impact of SM and different classification approaches performed on elections in various countries.

Stemming

Stemming is a technique that refers to all variations of a word to its root word, such as ‘warming’, ‘warmest’, ‘warmed,’ and ‘warmer’, which will be stemmed from the word ‘warm.’ This method reduces the time and memory space by removing suffixes that have exactly matching meanings and stem. For sentiment analysis on the text data, every word should be represented by the stem rather than the word mentioned in the text (Al-Khafaji & Habeeb, 2017).

Stop word removal

Stop words are the words that are useless within the raw dataset. These words do not provide helpful information in the data set, so they must be removed to save computation time, storage, and space and improve the algorithm’s efficiency. Most stop words are pronouns and helping verbs like is, of, the, to, and/or (Al-Khafaji & Habeeb, 2017).

Tokenization

Tokenization is a method to split the words within a sentence. Each split character, word, or symbol is called a token. It is an appropriate method in text analysis (Al-Khafaji & Habeeb, 2017). Like, [the president has worked well] will be tokenized into [the, president, has, worked, well] (Wongkar & Angdresey, 2019). These tokens help identify a content’s intent which helps in sentiment or text analysis.

Statistical approach

The authors of Ibrahim et al. (2015) presented a new approach for predicting the Indonesian Presidential elections in 2014. Their approach collected data from Twitter and preprocessed it by removing usernames and website links. Furthermore, the Twitter buzzers (Ibrahim et al., 2015) were eliminated using an automatic technique to collect data from real Twitter users and avoid unusual noise in it. The cleaned data was then subdivided into sub-tweets, labeled with a candidate’s name, and their sentiment polarity was computed. They further used mean absolute error (MAE) metric to evaluate the performance, resulting in an MAE of up to 61%.

In another study, Bansal & Srivastava (2018) introduced a novel method called Hybrid Topic Based Sentiment Analysis (HTBSA) for forecasting election results using tweets. The tweets were preprocessed using text formatting techniques, and then the topics were generated using the Biterm Topic Model (BTM). HTBSA was conducted based on the sentiments of topics and tweets, resulting in an 8.4% MAE (Eq. (1)). (1)${\displaystyle MAE=\sum _{i=1}^n|y_i-x_i|/n.}$

Similarly, the authors presented Lexicon-based Twitter sentiment analysis for forecasting elections using emoji and N-gram features in 2019 (Bansal & Srivastava, 2019). Unlike previous studies, sentiment polarity will be analyzed using non-textual data such as emoji(s). The authors gathered data from Twitter while restricting themselves to Uttar Pradesh (UP) geo-location. The data was cleaned from HTML tags, scripts, advertisements, stopwords, punctuations, special symbols, and white spaces. Duplicate tweets were also eliminated in the process. The refined data was then converted to bi-grams and tri-grams, followed by sentiment labeling. Simultaneously, emoji unicode was compared with developed n-grams, and its sentiment was labeled. Consequently, both sentiments were used to calculate election prediction. Another mathematical algorithm was presented in (Nawaz et al., 2022) based on sentiment forecasting for Pakistan democratic elections. The authors manually annotated tweets to avoid implications of spammed data among the datasets. Then, aspects of filtered tweets were extracted, assigning grammatical forms to each word in the sentence. The gathered factors were associated with opinions using the semantic similarity measure RhymeZone (Whitford, 2014). Once the association was done, the Bayesian theorem was applied, which classified tweets with 95% accuracy.

Ontology approach

The authors of Budiharto & Meiliana (2018) forecasted the Indonesian presidential election using tweets from presidential candidates of Indonesia based on a preprocessing algorithm. The tweets were processed with text formatting techniques, including stopwords in the Indonesian language and special character elimination. Once the tweets were refined, top words, favorite lines, and re-tweets were counted. Then the authors calculated the polarity of positive, negative, and neutral reviews. In Salari et al. (2018), researchers proposed text and metadata analysis to predict Iran’s presidential elections in 2017. The text data were gathered in the Persian language from two different platforms: Telegram and Twitter messages. The data was then analyzed using various analyses: sentiment analysis of hashtags, sentiment analysis of posts using Lexipers (Sabeti et al., 2019), time analysis, and several views and users of each message analysis (Telegram). The first two analyses were text analysis, while the others were metadata information analysis. In doing so, the model achieved 97.3% accuracy in predicting the presidential election.

Lexicon based approach

In 2019, Oyebode and Oriji conducted sentiment analysis to forecast Nigeria’s presidential election 2019 (Oyebode & Orji, 2019). The data was extracted from Nairaland (Nelson, Loto & Omojola, 2018) using a web scraping approach, and they were preprocessed with text cleaning techniques. The resultant data were fed to three lexicon-based classifiers (Vader (Hutto & Gilbert, 2014), TextBlob, and Vader-Ext) and to train five machine learning classifiers, namely support vector machine (SVM), logistic regression (LR), multinomial naive Bayes (MNB), stochastic gradient descent (SGD), and random forest (RF). When the classifiers were evaluated, the proposed Vader-Ext outperformed all other classifiers as it resulted in an 81.6% accuracy rate.

Supervised learning approach

Machine learning (ML) is a field of software engineering that uses measurable procedures to enable computer systems to “learn” with information without being programmed explicitly. ML tasks are classified as supervised, unsupervised and deep learning. In 2010, the authors presented an automated method that evaluates sentiments via linguistically analyzed documents (Pak & Paroubek, 2010a). Those documents were trained for the NB classifier and tested using n-gram as a feature. In 2012, the authors of Wang et al. (2012) proposed a real-time election prediction system that analyzed the opinions of various users on Twitter. The opinions were anatomized and later used to train and test the NB classifier. A unique prediction model for Elections held in Pakistan in 2013 was presented in Mahmood et al. (2013). A set of tweets were gathered according to predictive models, which were later cleaned and were used to train CHAID (chi-squared automatic interaction detector) decision tree (DT), SVM, and NB classifiers. When the classifiers were evaluated with test data, the CHAID decision tree dominated compared to SVM and NB (Eq. (2)). (2)${\displaystyle {\chi }^2=\sum \frac{{\left(O_i-E_i\right)}^2}{E_i}\text{where}{\chi }^2=chisquared,O_i=\text{observed value},E_i=\text{expected value}.}$

In 2014, the authors of Razzaq, Qamar & Bilal (2014) proposed a prediction system that evaluated the power of election prediction on the Twitter platform. The authors gathered and preprocessed tweets by eliminating duplicate tweets, URLs, whitespaces, and manual labeling. Furthermore, the Laplace method and Porter Stemming avoided zero values. Processed training data was used to train RF, SVM, and NB classifiers. When the classifiers were tested, NB dominated SVM and RF. Jose & Chooralil (2015) introduced a novel election prediction model using word sense disambiguation. The data was acquired from Twitter which was then cleansed by removing usernames, hashtags, and special characters. The negation handling technique was also applied to enhance the classification accuracy further. Speech tagging and tokenization were done on refined tweets provided by a word sense disambiguation classifier for categorization. The classifier attained a 78.6% accuracy rate. (3)${\displaystyle \sim F1=2\ast \frac{\frac{TN}{TN+FN}\ast \frac{TN}{TN+FP}}{\frac{TN}{TN+FN}+\frac{TN}{TN+FP}}}$

Tunggawan & Soelistio (2016) presented a predictive model for the 2016 US presidential election. They gathered data from Twitter, which went through simple filtration techniques such as URLs and candidate name removal to make the resultant data precise. In doing so, 41% of the data were eliminated. Then the data was labeled manually and fed to the NB classifier (Eq. (3)). The classifier predicted 54.8% accuracy. Sharma and Moh proposed a supervised election prediction method using sentiment analysis on Hindi Twitter data (Sharma & Moh, 2016). In this method, raw Hindi Twitter data underwent a text-cleaning module that removed negated words, stopwords, special characters, emoticons, hashtags, website URLs, and retweet text. 2 supervised (NB, SVM) and one unsupervised (Dictionary based) algorithms were used for classification. The dictionary-based classifier evaluated the tweets with 34% accuracy, whereas NB and SVM classifiers were trained with 80% accuracy of the data. In contrast, the remaining 20% of the data was used for the evaluation purpose, which resulted in 62.1% and 78.4%, respectively.

Ramteke et al. (2016b) presented a two-stage election prediction framework using sentiment analysis using Twitter data and TF-IDF. Further, the data was labeled using hashtag clustering and VADER techniques. 80% of the labeled data was used for training, and the remaining 20% was used for testing the classification algorithm. An accuracy rate of 97% was achieved when the classifier was tested. Ceron-Guzman and Leon-Guzman presented a sentiment analysis approach on Colombia Presidential Election 2014 (Cerón-Guzmán & León-Guzmán, 2016). Twitter data was cleaned and normalized in two stages: basic and advanced pre-processing. The data was stripped from URLs, emails, emoticons, hashtags, and special characters in the basic pre-processing stage. After then, the data was forwarded to the advanced pre-processing step, where lexical normalization and negation handling techniques were applied to refine the data further. Once the text was normalized, it was modified to a feature vector, which was later fed to the classifier. The labeled dataset was split into 80:20 ratios for training and testing classifiers. Overall, the classifier performed with 60.19% accuracy when evaluated on test data. Singh, Sawhney & Kahlon (2017) presented a novel method for forecasting US Presidential elections using sentiment analysis. After collecting data from Twitter, the authors implied a restriction to consider only one tweet per user. All duplicate tweets were removed to avoid interference affecting the method’s performance. Then unwanted HTML tags, web links, and special characters/symbols were removed from the data. The refined data was then used to train the SVM classifier. Once the classifier was trained, it was evaluated with the test data, i.e., to classify the polarity of the data, attaining a 79.3% accuracy rate.

In 2018, Bilal et al. (2018) presented a deep neural network application to forecast the electoral results of Pakistan 2018. They collected 56,000 tweets about the general elections in 2013 and treated them with text-cleaning techniques. The resultant data was then used to train Recurrent Neural Network (RNN). Once the RNN was trained, it was evaluated with test data resulting in an 86.1% prediction rate. In 2019, a new methodology was presented in Joseph (2019), which predicted Indian general elections using a decision tree. Ruling and opposing parties’ data was gathered from Twitter. Stopwords, regular expressions, emojis, Unicode, and punctuation, were pruned from the data. The resultant data was then tokenized and fed to the decision tree classifier for training. Once the classifier was trained, it was evaluated, which resulted in 97.3% accuracy. An efficient method to forecast the Indonesian presidential election using sentiment analysis was presented in Kristiyanti & Umam (2019). The authors collected data from Twitter, tokenized them, and generated Bi-grams(three-letter word combinations). Unlike previous research, the feature selection was made using the particle swarm optimization (PSO) algorithm and genetic algorithm (GA) algorithm separately. Then those features were used to train the SVM classifier. Once the classifier was trained, SVM with PSO performance dominated, against SVM with GA, with 86.2% accuracy rate.

Similarly, Oussous, Boulouard & Zahra (2022) proposed another Arabic sentiment analysis framework that forecasted the Moroccan general election 2021. The data was collected in Arabic from the Hespress website (a Moroccan news website). Then it was treated with text cleaning techniques (tokenization, normalization, and stop words removal) so that the dimensionality and processing time of the framework could be reduced. Term frequency (TF) was used to acquire feature vectors which were then passed on to several ML classification models such as SVM, NB, Adaboost, and LR for training. The classifiers predicted sentiment polarity with 94.35%, 62.02%, 87.55%, and 88.64% accuracy rates.

Deep learning approach

Hidayatullah, Cahyaningtyas & Hakim (2021) conducted sentiment analysis using a neural network to predict the Indonesian presidential election 2019. Two different datasets, i.e., before and after the elections, were collected and labeled using a pseudo-labeling technique. Then they were preprocessed with text-cleaning techniques, including case folding, word normalization, and stemming. This study trained three traditional ML classifiers (SVM, LR, and MNB) and five deep learning classifiers (LSTM, CNN, CNN+LSTM, GRU+LSTM, and bidirectional LSTM). Once the classifiers were ready, they were all evaluated by test data. SVM and bidirectional LSTM ruled better accuracy within their respective categories, but overall, bidirectional LSTM outperformed SVM by gaining an 84.6% accuracy rate. In another study, researchers presented a method for predicting USA presidential elections 2020 using social media activities (Singh et al., 2021). The dataset was pretreated with text formatting techniques plus TF-IDF vectorization. They were then fed into NB, SVM, TextBlob, Vader, and BERT classification models for training and testing purposes. Compared to other classifiers, the BERT classifier prevailed with a 94% precision rate.

In Ali et al. (2022), the authors introduced a deep learning model to forecast Pakistan general elections based on sentiment analysis. The dataset related to Pakistan general elections 2018 was gathered from Twitter, and they were labelled manually. Then the data was preprocessed with data transformation, tokenization, and stemming. Further, the proposed deep learning model was trained and evaluated with training and test data, respectively, resulting in a 92.47% prediction rate.

Era 1 (2010–2017)

This was the era where the research community were onset on innovating and developing election forecasting algorithms before actual election commencement. The prompt goal of their work was to predict and classify sentiments among digital text with optimal accuracy rate.

Pak & Paroubek (2010a) presented a novel method to forecast elections using sentiment analysis. Thus, the method predicted with 60% accuracy. In 2012, Wang et al. (2012) proposed an election prediction system focusing on US Presidential election 2012 data. Unigram features were extracted from 17,000 tweets (training dataset) and were fed to the NB classifier for training. Once the classifier was trained, it was evaluated with a 59% prediction rate. Mahmood et al. (2013) proposed an election prediction method that forecasted Pakistan General Election 2013 by assessing the CHAID decision tree. In doing so, the classifier performed with a 90% prediction rate. Razzaq, Qamar & Bilal (2014) presented a machine learning algorithm that predicted positive and negative sentiments with 70% accuracy. Thus, this method needed to be more consistent due to a biased data set. Ibrahim et al. (2015) proposed a statistical prediction method enthralled on Indonesian Presidential elections 2014. The method performed with 0.61 mean absolute error (MAE). There were several limitations to this method. Firstly, a dataset of all Indonesian voters across all Indonesian provinces should have been considered. Secondly, sentiment analysis (SA) cannot be performed when no keyword is present in candidate-related tweets. In 2015, Jose & Chooralil (2015) proposed an election prediction method using word sense disambiguation. Although the method performed with a 78.6% accuracy rate without the training phase, it was limited to negation handling and manual labeling.

In Tunggawan & Soelistio (2016), the authors innovated a Bayesian election prediction system focusing 2016 US presidential election. Although the system boomed with an exceptional accuracy rate with model test data, it under-performed with a 54.8% accuracy rate when evaluated with test poll data. Similarly, Sharma & Moh (2016) presented an Indian election prediction system using the Hindi dataset. The tweets were preprocessed, and then the polarity of the resultant tweets was calculated. SVM achieved a 78% prediction rate. The system is curtailed with emoticon analysis and extensive training data. A sentiment analysis system was introduced in Cerón-Guzmán & León-Guzmán (2016) predicting Columbia presidential election 2014. It resulted in the lowest MAE of about 4%. In Singh, Sawhney & Kahlon (2017), a sentiment analysis system was presented focusing on US presidential elections 2016. The system got trained with the Twitter processed data, and later, it was evaluated with test data, resulting in a 79% accuracy rate. A separate study (Ceron, Curini & Iacus, 2015a) examined the advantages of supervised aggregated sentiment analysis (SASA) on social media platforms to forecast election outcomes. Analyzing the voting intentions expressed by social media users during several elections in France, Italy, and the United States between 2011 and 2013, they compared 80 electoral forecasts generated through SASA alongside alternative data-mining and sentiment analysis approaches.

Era 2 (2018–2023)

This era is characterized as embarking of deep learning approach as it provided a major breakthrough and expedited state of the art results among its classification algorithms. It gave a new direction towards accuracy improvisation. In doing so, the research community focused on creating and developing new deep learning algorithms as compared to machine learning algorithms. Furthermore, this era also saw a novel sentiment classification proposals relying on statistical, lexicon and ontology approaches as seen in Fig. 3.

Bilal et al. (2018) introduced a deep neural networks (DNN) election prediction model that forecasted Pakistan General Elections 2018 with an 86.1% accuracy rate. Case sensitive tweets of tweets deteriorated the performance of the method. Kristiyanti & Umam (2019) proposed a sentiment analysis method to predict the Indonesian presidential election for 2019–2024. The system utilized particle swarm optimization (PSO) and genetic algorithm (GA) algorithms with SVM to improvise accuracy to 86.2%. Salari et al. (2018) presented an election prediction system for Iran presidential election 2017. Both text and metadata analysis of the tweets were considered to evaluate the system’s performance. The system performed with a 97.3% accuracy rate without the training phase. Another outcome prediction system, based on Indian general elections, was presented by Joseph (2019), which trained and tested the DT classifier resulting in a 97% accuracy rate. Thus, this system works well with tweets in the English language only.

Chaudhry et al. (2021) proposed an election prediction method, mainly focusing on the US election 2020. They collected the Twitter data, preprocessed them, and extracted features using TF-IDF. Features of around 60% of the training dataset were used to train the NB classifier. In contrast, the features of the rest 40% dataset were utilized to evaluate the performance, resulting in a 94.58% accuracy rate. Thus, the authenticity of the dataset (tweets) was not examined, which hurt the method’s performance. Likewise, Xia, Yue & Liu (2021) proposed a sentiment analysis-based election prediction method for the same election campaign. The authors preprocessed the tweets with string replacement and stemming techniques in this method, followed by n-gram feature extraction. Multi-layer perceptron (MLP) classified 27,840 dimension features with an accuracy of 81.53%.

An election prediction method based on a deep learning approach was introduced in Hidayatullah, Cahyaningtyas & Hakim (2021), which forecasted the Indonesian Presidential elections in 2019. The authors trained CNN, long-short term memory (LSTM), gated recurrent unit (GRU), and bidirectional LSTM, SVM, LR, and multinomial NB, from which bidirectional LSTM dominated against the rest by achieving 84.6% prediction rate. It also implied that, in comparison, DNNs attained a better accuracy rate than traditional machine learning algorithms. Similarly, another deep learning approach was implemented to forecast US presidential elections 2020 in Singh et al. (2021). Three machine learning algorithms (SVM, NB, and TextBlob) and one deep learning algorithm (BERT) were trained and evaluated. As a result, the BERT algorithm attained the highest prediction rate of 94%. It denoted that DNN algorithms achieved better accuracy than conventional machine learning algorithms. Ali et al. (2022) introduced another DNN election prediction method focusing mainly on Pakistan General Elections 2018. The data were labeled manually, preprocessed, and later tokenized as usual. The resultant dataset was used to train and evaluate the DNN classifier, resulting in a 92.47% accuracy rate. Thus, the dataset used in the method above needed to be higher, due to which accuracy dropped. Previously, traditional polling data was widely considered the most reliable method for forecasting electoral outcomes. However, recent developments have revealed polling data’s potential incompleteness and inaccuracy. A study was conducted to compare the accuracy of polls with sentiment analysis results obtained from Twitter tweets (Anuta, Churchin & Luo, 2017). The study analyzed a new dataset of 234,697 Twitter tweets related to politics, collected using the Twitter streaming API. The tweets underwent preprocessing, removing hashtags, links, and account names and replacing emotions and symbols with their complete form. The study’s findings indicated that Twitter exhibited a 3.5% higher bias in popular votes and a 2.5% higher bias in state results compared to traditional polls. Consequently, the study concluded that predictions based on Twitter data were inferior to those found on polling data (Anuta, Churchin & Luo, 2017). The researchers highlighted the limitations of previous methods. They recommended incorporating additional techniques, such as POS tagging and sense disambiguation, during preprocessing and considering contextual and linguistic features of words to enhance prediction accuracy (Anuta, Churchin & Luo, 2017).

In the context of the 2016 US elections, traditional techniques like polling were deemed unreliable due to the rapid evolution of technology and the prevalence of social and digital media platforms (Hinch, 2018). A study analyzed slogans used in Twitter tweets during the elections, employing a WordCloud visualization. However, the analysis results could have been more consistent with the actual election outcomes, particularly in predicting Trump’s victory in Michigan and Wisconsin. The researchers emphasized the need to consider qualitative aspects when making electoral predictions, as the approaches employed in the study failed to capture the dynamics accurately.

The relationship between candidates’ social network size and their chances of winning elections was examined in a study that utilized data from Facebook and Twitter (Cameron, Barrett & Stewardson, 2016). The study employed regression analysis and proposed three models, with the number of votes as the dependent variable and the number of Facebook connections and other factors as independent variables. The results indicated a significant correlation between the size of the social network and the likelihood of winning. However, the effect size was small, suggesting that social media data is predictive only in elections with close competition.

A study used social network techniques, such as volumetric analysis and sentiment analysis, to infer electoral results for Pakistan, India, and Malaysia (Jaidka et al., 2019). The study collected approximately 3.4 million tweets using the Twitter streaming API and separated English tweets using a natural language toolkit. Volumetric analysis, measuring the volume of tweets for each party; sentiment analysis assessing positive and negative tweets; and social network analysis determining the centrality score of each party were employed. The study found that Twitter data was ineffective for making election predictions in Malaysia but proved effective and efficient for Pakistan and India. Incorporating multiple techniques, the proposed model was also effective for candidates and parties with fewer votes.

A study conducted in 2016 proposed a predictive model for forecasting the outcome of the US presidential elections based on an NB approach utilizing Twitter data (Tunggawan & Soelistio, 2016). The researchers collected tweets from December to February, covering three months. The collected data underwent simple pre-processing techniques to prepare it for sentiment analysis. The resulting model achieved an impressive accuracy of 95.8% in sentiment prediction. A 10-fold cross-validation technique was employed to assess the model’s robustness. The F1 Score was used to evaluate the model’s accuracy in predicting positive sentiments, while F1 represented the model’s accuracy in classifying negative sentiments. The model’s accuracy in predicting negative sentiments (Tunggawan & Soelistio, 2016). The authors of Heredia, Prusa & Khoshgoftaar (2018) introduced their sentiment analysis model, which classified the data with an accuracy of 98.5%. However, when the model’s predictions were compared with actual polling data, the results indicated an accuracy of 54.8%.