key: cord-1051250-2b3mrkcy authors: Vashishtha, Srishti; Susan, Seba title: Neuro-fuzzy network incorporating multiple lexicons for social sentiment analysis date: 2021-11-29 journal: Soft comput DOI: 10.1007/s00500-021-06528-0 sha: 8cc8c775819b05f4313d20f28c06b643c32f3cd7 doc_id: 1051250 cord_uid: 2b3mrkcy We have proposed MultiLexANFIS which is an adaptive neuro-fuzzy inference system (ANFIS) that incorporates inputs from multiple lexicons to perform sentiment analysis of social media posts. We classify tweets into two classes: neutral and non-neutral; the latter class includes both positive and negative polarity. This type of classification will be considered for applications that aim to test the neutrality of content posted by the users in social media platforms. In our proposed model, features are extracted by integrating natural language processing with fuzzy logic; hence, it is able to deal with the fuzziness of natural language in a very efficient and automatic manner. We have proposed a novel set of 64 rules for the proposed neuro-fuzzy network that can classify tweets correctly by working on fuzzy features fetched from VADER, AFINN and SentiWordNet lexicons. The proposed novel rules are domain independent, i.e., we can extend these rules for any textual data that employs lexicons. The antecedent and consequent parameters of the ANFIS are optimized by gradient descent and least squares estimate algorithms, respectively, in an iterative manner. The key contributions of this paper are: (1) a novel neuro-fuzzy system: MultiLexANFIS that takes as its input the positive and negative sentiment scores of tweets computed from multiple lexicons—VADER, AFINN and SentiWordNet, in order to classify the tweets into neutral and non-neutral content, (2) a novel set of 64 rules for the Sugeno-type fuzzy inference system—MultiLexANFIS, (3) single-lexicon-based ANFIS variants to classify tweets when multiple lexicons are not available and (4) comparison of MultiLexANFIS with different fuzzy, non-fuzzy and deep learning state of the art on various benchmark datasets revealing the superiority of our proposed neuro-fuzzy system for social sentiment analysis. Social sentiment analysis has a huge impact on daily life, in modern times, due to the widespread use of social media for expressing public opinion. People prefer to communicate on the online social network such as Twitter, Facebook, WhatsApp, Instagram and YouTube; among them, Twitter is the most popular microblogging forum for expressing public sentiment. Sentiment analysis (SA) is used to inspect the polarity or intention of tweets about any entity. The content of tweets could be movie reviews (Vashishtha and Susan 2021; Parveen and Pandey 2016; Rahman et al. 2016; Anagha et al. 2015; Kulkarni et al. 2014; Ghiassi et al. 2013) , product reviews (Ghiassi et al. 2013; Indhuja and Reghu 2014) , stock market trends and political events (Khan et al. 2020) , political views (Katta and Hegde 2019), COVID-19 (Nemes and Kiss 2021) or about any other topic that attracts public opinion. Tweets can be classified as positive, negative or neutral using either supervised or unsupervised machine learning algorithms. Generally, the twitter data contains a major portion of neutral tweets compared to positive and negative tweets. It has been proved that neutral samples are vital for learning polarity in order to obtain remarkable classification results Schler 2005, 2006) . In this research work, classification of tweets is not the stereotype-positive, negative or neutral; rather, it is neutral and & Srishti Vashishtha srishtidtu@gmail.com Seba Susan seba_406@yahoo.in non-neutral, and the latter class includes both positive and negative polarity. The purpose of this type of classification will be beneficial for such systems that check neutral content posted by the users. There are many intelligence agencies that aim to track an individual's social media activities for national security reasons, in search of antinationalists or terrorist activities, etc. Social media is widely used by people, and they tend to post, like and share content of their interest on a regular basis. These agencies analyze the content posted by individuals to determine whether the content is neutral or biased toward any political event or leader, or generally indicative of any malicious intent. Hence, such tasks need to deploy SA on online social network in order to classify the social media posts into neutral and non-neutral category related to an individual or group. To the best of our knowledge, none of the previous works on SA have concentrated on solely classifying tweets into neutral and non-neutral and none of them have targeted to build models that would aid security applications based on sentiment scores. Also, mostly the authors use one lexicon at a time and have not simultaneously incorporated multiple lexicons into their learning framework as proposed in our work. Two types of ANFIS variants-single lexicon and multiple lexicons, are presented in this work. The architecture of the proposed MultiLexANFIS model consists of fuzzy-inspired feature extraction phase where the input features are fetched from multiple lexicons-VADER (Hutto and Gilbert 2014) , AFINN (Nielsen 2011) and SentiWordNet (Baccianella et al. 2010) , and fuzzy inference phase is constructed using a novel set of 64 rules devised specifically for these three lexicons. The rules can be easily extended for adaptation to a greater number of lexicons. The major contributions of this research are: (1) development of a novel neuro-fuzzy system: MultiLexANFIS that integrates sentiment scores from multiple lexicons to classify tweets into neutral and non-neutral. These sentiment scores are computed from pre-processed textual data and are interpreted as a fuzzy membership values. (2) A novel set of 64 rules for Sugeno-type fuzzy inference system (FIS)-MultiLexANFIS, is constructed. Each lexicon provides two sentiment scores-positive and negative, so a total of six sentiment scores are derived from the VADER, AFINN and SentiWordNet lexicons for each tweet. Each of these six inputs is assigned two membership functions (MFs): Low and High. (3) We have also constructed single-lexicon-based ANFIS variants with nine rules each to classify tweets. Three variants that we call as VADER, AFINN and SentiWordNet-ANFIS consider the two sentiment scores of positive and negative as input at a time. Each of these inputs is assigned three membership functions (MFs): Low, Medium and High. These variants can be deployed when multiple lexicons are not available. The experiments are conducted on MultiLexANFIS and all three single-lexicon ANFIS variants; the results reveal that MultiLexANFIS performs better than all single-lexicon variants. (4) A comparative analysis of proposed Mul-tiLexANFIS with fuzzy approaches and non-fuzzy approaches including deep learning models on various benchmark datasets reveals the supremacy of our proposed neuro-fuzzy system. The performance of all models and the state of the art is measured in terms of RMSE (root mean square error). This paper is organized as follows. Section 2 describes several recent works related to ANFIS and sentiment analysis. The proposed MultiLexANFIS is explained in Sect. 3. The experimental setup and implementation details are discussed in Sect. 4. The results are demonstrated in Sect. 5. Finally, Sect. 6 draws the overall conclusion. Sentiment analysis is a process that is involved in the determination and classification of opinions or feelings expressed in text, through the use of computing machines. The major challenge is to extract information from unstructured data. Most methodologies in this area focus on polarity or opinion classification into positive or negative classes (Siddiqua et al. 2016; Vashishtha and Susan 2020b; Jefferson et al. 2017; Vashishtha and Susan 2020a, b) ; some researchers also include a neutral class (Parveen and Pandey 2016; Anagha et al. 2015; Indhuja and Reghu 2014; Susan 2018, 2019a) . Others detect neutral opinions and filter out them to enhance binary polarity classification (Valdivia et al. 2017 (Valdivia et al. , 2018 . Sentiment analysis models utilizing part-of-speech tags and n-gram feature extraction are gaining popularity. These models deploy either unsupervised algorithms (Vashishtha and Susan 2021) or supervised machine learning algorithms involving SVM, naïve Bayes, etc. (Tripathy et al. 2016) . One notable supervised technique is the combination of hidden Markov model for feature extraction and SVM and ANN for the classification process (Kalarani and Brunda 2019) . Machine learning classifier models have proved to perform well for sentiment classification, and some of the popular algorithms are naïve Bayes (Parveen and Pandey 2016) , support vector machines (SVM) (Katta and Hegde 2019), (Valdivia et al. 2017 (Valdivia et al. , 2018 , long short-term memory (LSTM) (Rahman et al. 2016; Vashishtha and Susan 2019) , recurrent neural network (RNN) (Nemes and Kiss 2021) , maximum entropy with probabilistic latent semantic analysis (Xie et al. 2019 ) and ANFIS. Out of all these options, neuro-fuzzy networks perform generally well in classification tasks involving ambiguous and uncertain systems since they combine learning with reasoning which is a human trait (Susan et al. 2013 ). In recent years, several fuzzy approaches have started to emerge for text processing. Fuzzy logic is a concept used in computing based on the ''degree of truth'' instead of true or false labels (0 or 1). It uses fuzzy sets, where each element belongs to the set with a certain degree; this degree is determined by a particular membership function. It has been meticulously exploited as a methodology useful both for representing information associated with sentiment and for executing various sentiment analysis tasks. The exploitation of fuzzy concepts in the field of SA can be incorporated in feature extraction, in fuzzy-based model architecture or in both (Serrano-Guerrero et al. 2021 ). An unsupervised fuzzy logic-based method for sentiment classification of textual reviews was proposed in Vashishtha and Susan (2020b) . The authors formulated fuzzy cardinality, as the measure, from fuzzy sets for the evaluation of word polarity scores, and the proposed model has two versions based on the sentiment lexicon utilized in the model. Many research works have implemented fuzzy membership functions (MFs): triangular (Vashishtha and Susan 2019a) , trapezoidal Jefferson et al. 2017) , Gaussian and generalized bell MFs (Gupta et al. 2014; Talpur et al. 2017) . These MFs for a basic atomic term can be modified using linguistic Hegdes, which are adverbs or adjectives. Linguistic Hegdes with fuzzy logic have been successfully used to compute the sentiment of text Susan 2018, 2021; Khattak et al. 2020) . A combination of fuzzy entropy, k-means clustering and sentiment lexicon shortlists words that help in sentiment cognition in Vashishtha and Susan (2019b) . In the field of social media SA, there is also a shift toward multimodal social web, where users post their opinion in the form of text in Facebook and Twitter, audio or video clips in YouTube. The social media reviews are either classified using only fuzzy logic-based text classification (Wu et al. 2017) or can be classified by multimodal sentiment classification using a fusion of fuzzy logic with audio and text features (Vashishtha and Susan 2020a) . Sentiment classification can be deployed using rule-based systems, a special type of expert system which uses a set of pre-defined rules for the classification task. Rule-based classifier can be supervised (Siddiqua et al. 2016) or may use fuzzy rules constructed in an unsupervised manner (Anagha et al. 2015; Liu and Cocea 2017; Jefferson et al. 2017; Vashishtha and Susan 2019a; Talpur et al. 2017 ). The rules can be further applied to supervised machine learning algorithms: naïve Bayes (Siddiqua et al. 2016 ) and SVM (Katta and Hegde 2019) and can also be learned in ANFIS (Jang 1993) . Ahuja et al. (2015) implemented a SA model on public opinion data from Twitter to find a relationship between the stock market and public mood, using a self-organizing fuzzy neural network. An optimized fuzzy technique-shuffled frog leaping algorithm (SFLA), developed by Madhusudhanan et al. (2019) is deployed for improving the SA process. The SFLA is an optimization technique which imitates the mimetic evolutionary activity of frogs searching for a place with most quantity of food; along with SFLA, a local search algorithm is implemented for performing SA on book reviews. Phan et al. (2019) proposed a method that detects polarity of only specific kind of tweets that contain fuzzy sentiment phrases (FSPs). The first step of this method is to detect FSPs and then extract a set of features related to FSPs based on the syntactic, lexical, semantic and polarity sentiment of the words. The second step is to classify tweets using multilayer perceptron model. In a recent research work (Yakubu et al. 2021 ), customer satisfaction model based on online reviews has been developed, based on the sentiment scores derived by applying a novel multigene genetic programming-based fuzzy regression (MGGP-FR) approach. To address the fuzziness of customer opinions, the sentiment scores are transformed into asymmetrical fuzzy numbers. Then, the defuzzified sentiment scores are applied to MGGP; further, Tanaka's fuzzy regression is employed to determine whether the customer is satisfied or not. In 1993, Jang proposed ANFIS, a fuzzy inference system implemented in the framework of adaptive networks (Jang 1993) . It uses a hybrid learning procedure to construct an input-output mapping based on both human knowledge (in the form of fuzzy if-then rules) and designated inputoutput data pairs. ANFIS has been applied for classification (Ž unić et al. 2016) , prediction (Cosma and Acampora 2016; Kulkarni et al. 2012 Kulkarni et al. , 2014 Mazumder et al. 2018) , subjectivity analysis ( Kulkarni et al. (2012) developed an ANFIS based on sentiment information from movie blogs and reviews, for predicting sales performance in movie domain. The proposed predictive model is built to incorporate the auto-regressive method and ANFIS technique to handle the fuzzy demand with incomplete information. Another work by Kulkarni et al. (2012) proposed neuro-fuzzy methodology-ANFIS, for predicting sales performance of movies by analyzing sentiments in online reviews. The ANFIS model here has three inputs: review ratings, revenue budget and type of the movies. Twitter data has been analyzed for political issues: using a hybrid system of ANFIS and SVM, where the SVM classifier optimizes the fuzzy rules (Katta and Hegde 2019), and using principal component analysis with ANFIS to predict political sentiments (Mazumder et al. 2018) . Another hybrid classifier of ANFIS with genetic algorithm (GA) has shown good results in twitter SA (Padmaja and Hegde 2019; Acampora and Cosma 2014) . GA extracts the important and semantic features from each tweet and optimizes the fuzzy principles in ANFIS. The authors in Asimuzzaman et al. (2017) implemented ANFIS to predict the polarity of Bangla tweets. The ANFIS utilizes semantic rules which are simple yet have a greatly affect the actual polarity of the text. Since the model is designed for classifying tweets into positive and negative polarity, the neutral tweets were ignored in pre-processing phase. Each and every word of Bangla text is translated into English text, so that the polarity value of each word can be extracted from SentiWordNet lexicon. The experiments of this model are conducted on a small twitter samples training set, and it has not been compared with recent state of the art for SA. Kamil et al. (2018) applied ANFIS for executing three different classification problems: (1) sentence-level subjectivity detection, (2) sentiment analysis of texts and (3) detecting user intention in natural language call routing system. In their work, the main goal is to avoid using human annotation or lexical knowledge; the membership degree of each term is calculated by pruned ICF (inverse-class frequency). The fuzzification operations applied involve the computation of the maximum of membership degree with respect to the classes for every term and mean of maxima for all classes. A recent study compared supervised soft computing approaches: fuzzy time series (FTS), artificial neural network (ANN)-based FTS and ANFIS to estimate the human emotional states expressed in tweets on social media networks (Ç akıt et al. 2019) . Combination of ANFIS and hidden Markov model has been deployed for SA in (Rustamov et al. 2013a; b) and subjectivity analysis in (Rustamov 2018) . During the preparatory investigation in this study, it was found that triangular, Gaussian and generalized bell MFs have lower root mean square error (RMSE) and this motivated us to use these MFs in our proposed MultiLexANFIS. A multimodal convolutional neuro-fuzzy framework has been developed by Nguyen et al. (2019) that integrates the concepts of fuzzy logic with deep neural network for emotion classification of videos. Convolutional neural network is deployed to extract features from text, audio and visual modalities; these are sent as input to ANFIS that generates new rules to correctly classify emotions. An improved ANFIS (IANFIS) model has been proposed by Sasikala and Sheela (2020) aimed for future prediction of online products. The authors have also developed a deep learning modified neural network (DLMNN) for SA of online products. A food review dataset being analyzed, initially the dataset is divided into three groups: grade based, content based and collaboration based. For the grade-based data, polarity scores are determined and the review is classified as positive, negative or neutral using DLMNN, ANN and hybrid dragonfly-genetic algorithm (HDF-GA). Social media platform provides abundant and huge amount of data that can be analyzed for different purposes. A hybrid machine learning model: bi-objective optimization and LAN 2 FIS (logistic adaptive network based on neuro-fuzzy inference system), proposed by Nagamanjula et al. (2020) aims to utilize and handle this vast amount of data for sentiment classification of tweets. The bi-objective optimization implements minimum redundancy and maximum relevancy for selecting optimal set of features from tweets. These features are fed as input to logistic regression-based ANFIS for initially classifying tweets into five classes: positive, negative, strongly positive, strongly negative and neutral; further, the classes are reduced to positive, negative and neutral. 3 Proposed ANFIS architecture for social sentiment analysis In this paper, we have proposed a novel ANFIS-based system: MultiLexANFIS that integrates inputs from multiple sentiment lexicons to compute the sentiment of social media posts. We have also proposed single-lexicon-based ANFIS variants that require input from one sentiment lexicon at a time. These variants can be deployed when multiple lexicons are not available. The datasets chosen in this research work comprise tweets that are categorized into two classes: neutral and non-neutral. We have merged the positive and negative polarities into one class: nonneutral. In the last decade, fuzzy systems and their hybrid derivations have proved that they can emulate the typical human reasoning ability in a computationally efficient way. Fuzzy models have been used to construct cognitive models based on rules, expert systems, classifiers and universal approximators. Adaptive neuro-fuzzy inference system (ANFIS) is a neuro-fuzzy system that incorporates the power of artificial neural network and fuzzy logic. Our ANFIS model combines natural language processing (NLP) with fuzzy logic, which boosts the performance of SA. The importance of fuzziness comes into play while dealing with natural language due to the presence of ambiguity in word senses and the semantic interpretation of sentences. The amalgamation of fuzzy logic concepts into SA helps in dealing with the fuzziness of natural language in a very efficient and automatic manner. Neuro-fuzzy systems have been applied, before, in the field of SA or opinion mining, but these systems were not able to handle the uncertainty and vagueness present in the natural language. Moreover, these systems usually classified the data into positive or negative class; our goal is to classify data into neutral on non-neutral class that would aid organizations to test for the neutrality of content posted by users in social media platforms about persons, events, services, etc. To the best of our knowledge, this novel sentiment classification problem-segregation of neutral and non-neutral classes, has not been explicitly handled by any research work. We have designed two types of ANFIS variants: for single lexicon and multiple lexicons. The single-lexicon variant is based on the type of sentiment lexicon chosen. The experiments have been designed to figure out which lexicon is best suited for ANFIS-based sentiment classification in the field of social SA. Since single-lexicon variant proved to be inconclusive about which fuzzy membership is best suited for ANFIS-based SA, we designed a multiplelexicon ANFIS variant-MultiLexANFIS. To solve the above confusion of choosing the best membership function (MF) for a given lexicon and dataset combination, the MultiLexANFIS architecture has been proposed. This system integrates all the lexicons into a single unit and produces lower error compared to other models, and it takes less execution time in contrast to single-lexiconbased ANFIS. In our work, the proposed novel rules are domain independent. We can apply these rules for any textual data that employs lexicons to extract positive and negative score for each textual document or sentence or review. The motivation behind segregating neutral and non-neutral classes is that in some applications, the neutral content carries more importance than others. It has been demonstrated that a mix of training samples containing neutral samples is better as compared to a training set of the same size that includes only positives and negatives samples Schler 2005, 2006) . The use of neutral training samples in learning not only promotes better distinction between positive and negative samples but also increases the classification accuracy (Koppel and Schler 2006) . There exists a bias in social media content posted by users. Social theories claim that bias is a deep-seated feature of human behavior, which is described by a lack of proper balance and neutrality; one of the instances of bias is when someone supports one side too strongly or too often (Guerra et al. 2011) . We have designed an ANFIS-based architecture that can be used by organizations which hunt for neutral and unbiased content. Currently, there is a new trend; tasks like personality profiling and behavior profiling are deployed by organizations based on the neutrality of content. Organizations which deal with visa applications may require the candidates to have neutral behavior for their country; those candidates who are too biased can pose a threat for that country. Such systems can also be used by military services and intelligence agencies for tracking anti-nationalists or terrorist activities. Katta and Hegde (2019) proposed a hybrid ANFIS that integrates nonlinear SVM for SA of political twitter data. In feature extraction phase, unigrams and bigrams models are used. Only the words or pair of words are fed as input to ANFIS; whereas in our proposed model we evaluate the positive and negative sentiment scores of each tweet and send them as input to ANFIS. This is expected to yield better results because sentiment scores extracted from lexicons would provide better information for sentiment classification. In Katta and Hegde (2019), authors used only one fuzzy MF: generalized bell, while in our proposed model we have explored three fuzzy MFs: triangular (trimf), generalized bell (gbellmf) and Gaussian (gaussmf), in order to select the best performing MF. The authors in Katta and Hegde (2019) have conducted experiments and compared their model's different hybrid ANFIS-SVM variants but not with other state-of-the-art methods, whereas we have conducted experiments and compared with one fuzzy-rule-based classifier and four non-fuzzy state-of-the-art methods. Kamil et al., (2018) applied ANFIS for executing three different classification problems. In the pre-processing phase, the text is converted into an array of words which is sorted in increasing order of alphabets. Each word is assigned a unique code. The words are not lemmatized; hence, neither verbs in different tenses nor nouns as singular or plural are combined. In our approach, the words are lemmatized. The advantage of lemmatization is that it takes into consideration the context of the word to determine which is the intended meaning the user is looking for. This process allows to decrease noise and speed up the user's task. The authors in Kamil et al. (2018) do not use lexical knowledge, whereas we use lexicon to extract the positive and negative polarity scores that contribute to SA. Sentiment lexicons permit the creation of important and high-quality features for supervised SA. Moreover, they can be applied to diverse domains. The incorporation of sentiment lexicon makes our model domain independent. In Kamil et al. (2018) , the membership degree of each term is calculated by pruned ICF (inverse-class frequency), whereas in our model, three popular fuzzy MFs are applied for fuzzification of sentiment scores. Padmaja and Hegde (2019) developed a hybrid classifier based on ANFIS and genetic algorithm (GA) to classify the tweets into binary sentiment classes-positive or negative class, while our proposed model classifies tweets into neutral or non-neutral sentiment class. The ANFIS classifier integrates the benefits of both artificial neural network and fuzzy logic for generating the fuzzy rules, and the GA optimizes the fuzzy principles in the ANFIS classifier. After pre-processing of raw tweets, features are extracted; here, positive word count, negative word count and tag count are considered as features. The positive and negative words are attained by considering only adjectives, whereas in our proposed model, the positive and negative scores of tweets are computed by considering adjectives, nouns, verbs and adverbs. The inclusion of other part of speech enhances the sentiment evaluation process. Moreover, they have not specified how the words are classified as positive or negative. In our ANFIS model, we have deployed sentiment lexicons to compute the positive and negative scores of tweets. The authors in Padmaja and Hegde (2019) have compared with only one existing methodology latent Dirichlet allocation possibilistic fuzzy C-means (LDA-PFCM) and conducted experiments on only one dataset of 479 tweets. Our proposed model has been compared with eight existing SA approaches, and the experiments are conducted on ten benchmark datasets. In this section, we will discuss the overall procedure of tweet processing. Initially, the twitter data is collected; many publicly available datasets consist of positive, negative and neutral tweets. We have used a total of ten publicly available datasets: Apple Twitter Dataset (2019) Table 3 . We have combined the positive and negative tweets into one class: non-neutral; our goal is to classify the tweets by ANFIS into two classes: neutral and non-neutral. We change the labels of positive and negative tweets to nonneutral; thus, we have now two types of classes: neutral and non-neutral. The labels, 1: neutral and 0: non-neutral, are the class labels in ANFIS. The various steps in tweet processing are outlined in the following. The first step is data pre-processing, which is very essential because raw tweets are in unstructured form. In data preprocessing, we eliminate URLs, ''@'' symbol and hashtag ''#'' symbol. We have re-phrased commonly used phrases (like ''can't'') with their grammatical form (''can not''). This has been re-phrased because words like ''not'' contain polarity scores, if these phrases are not replaced, these words would be ignored which can lead to incorrect results. We have to compute the sentiment of this pre-processed tweet which is required to be in a numerical format so that it can be used as an input to ANFIS. Sentiment lexicons are employed for this purpose, and these lexicons assign scores to either words or the whole tweet. The three sentiment lexicons are VADER (Hutto and Gilbert 2014) , AFINN (Nielsen 2011) and SentiWordNet (Baccianella et al. 2010) . These sentiment lexicons are popularly used in SA applications. VADER (Valence Aware Dictionary for sEntiment Reasoning) is a sentiment lexicon that is especially attuned to sentiment in microblog-like contexts (Hutto and Gilbert 2014) . It works fast and yields high accuracy. It computes the scores for the whole tweet s using VADER lexicon's polarity_scores(s) method and gives the positive (Tweet-Pos) and negative (TweetNeg) scores of a tweet as the output. AFINN lexicon consists a list of English terms manually rated for valence with an integer between -5 (negative) and ? 5 (positive) (Nielsen 2011) . This list also contains slang and obscene words; hence, it can deal with modern-day tweets very well. AFINN lexicon employs the AFINN method to fetch the score of each word a as represented in Eq. (1). If the score is greater than 0, it is a positive word (Eq. 2). If the score is less than 0, it is a negative word (Eq. 3). The positive score of the tweet (TweetPos) is calculated by adding up the scores of all the positive words in the tweet as shown in Eq. (4). The negative score of the tweet (TweetNeg) is calculated by adding up the scores of all the negative words in the tweet as shown in Eq. (5). where a is a word in a tweet and m is the number of selected words. SentiWordNet lexicon is one of the most popular sentiment lexicons used by researchers. It assigns three numerical scores to each word: positive, negative and objective; these scores are in the range of 0.0 to 1.0 (Nielsen 2011) . The tweets are processed for this lexicon in the following manner: We remove the stopwords and punctuations, perform lemmatization, part-of-speech (POS) tagging and word sense disambiguation (WSD) by Lesk (Banerjee and Pedersen 2002) . WSD is the method of determining which ''sense'' (meaning) of a word is activated by the use of the word in a particular context. The SentiWordNet method acquires the scores of each word a from this lexicon using syn.pos_score () and syn.neg_score (). Each word has positive (Eq. (8)) and negative score (Eq. (9)) computed using WSD, which can be interpreted as a fuzzy membership belonging to the fuzzy sets Pos (Eq. (6)) and Neg (Eq. (7)). The words, which are having higher positive score than negative score in a tweet, are summed up to evaluate the positive score (TweetPos) of the tweet (Eq. (10)). Similarly, words which are having higher negative score than positive score in a tweet are summed up to evaluate the negative score (TweetNeg) of the tweet (Eq. (11)). These scores are computed for all tweets. Pos ¼ fða; l Pos ðaÞg; a 2 Xi ð6Þ where a is the word in a tweet, m is the number of selected words and X i is the set of total words. The computed TweetPos and TweetNeg of each tweet represent the numerical score of that tweet; these scores are given as input in ANFIS. ANFIS was introduced by Jang in (1993); it takes crisp inputs and fuzzifies these inputs using membership functions. ANFIS is a method of artificial intelligence that uses a hybrid way of learning by connecting fuzzy logic and neural networks into a single unit. It is based on specific ''human features,'' because fuzzy logic has a human subjectivity feature, while the neural network has a feature of human way of learning from past experience. ANFIS network is appropriate for solving function problems in various fields. Generally, the problem to be solved is the approximation of an unknown function. We next explain the working of a general ANFIS network constituted by M rules of Sugeno first-order type (Sugeno and Kang 1988) . The kth rule, k = 1…M, has the following form: where x ¼ ½x 1 x 2 . . .x n is the input pattern and y (k) is the output associated with the kth rule. Each rule is characterized by the membership functions (MFs) l B ðkÞ j ðx j Þ of the fuzzy input variables B j (k) , j = 1…n, and by the coefficients a j (k) , j = 1…n, of the crisp output. There are numerous options available for the fuzzification of crisp inputs, the composition of input MFs, and the way rule outputs are combined. The next step is to find the AND product of the fuzzy memberships of all input variables in a rule. This product is a MF l B ðkÞ ðxÞ, also called the firing strength (or weight) of the rule. The M-rule fuzzy inference system for ANFIS networks can be mathematically represented as: whereỹ is the estimate of the actual value y = f(x), for a given input x. We have proposed two ANFIS-based architectures: One is the single-lexicon-based ANFIS and the other, Mul-tiLexANFIS that combines all the lexicons into a single integrated unit. Single-lexicon-based ANFIS takes two inputs, TweetPos and TweetNeg, and one output, sentiment label (1: neutral or 0: non-neutral), for one sentiment lexicon-dataset combination at a time. MultiLexANFIS has six inputs; it includes the positive and negative score of a tweet from all the three lexicons: VADER, AFINN and SentiWordNet, and one sentiment label as output for each dataset. The fuzzy inference system (FIS) is of Sugeno first-order type with two inputs TweetPos and TweetNeg and one output label sentiment. A common rule set with two inputs and three MFs for each input would yield nine rules (3 2 = 9). The MFs for the input have linguistic labels: Low, Medium and High. The labels Low, Medium and High are two fuzzy sets created using different fuzzy membership for input variables: positive score (TweetPos) and negative score (TweetNeg). Mathematically, we can denote the inputs as x and y, MFs for input x as A i , MFs for input y as B i , where i = 1, 2, 3. The fuzzy if-then rules are then defined as follows: Rule2 : If x is A 2 and y is B 1 then Rule3 : If x is A 3 and y is B 1 then Rule4 Rule5 : If x is A 2 and y is B 2 then Rule6 : If x is A 3 and y is B 2 then f 6 ¼ p 6 x þ q 6 y þ r 6 Rule7 : If x is A 1 and y is B 3 then f 7 ¼ p 7 x þ q 7 y þ r 7 Rule8 : If x is A 2 and y is B 3 then Rule9 : If x is A 3 and y is B 3 then f 9 ¼ p 9 x þ q 9 y þ r 9 : The antecedent part of these rules is the same as in the recently published previous work of authors on unsupervised fuzzy rule-based sentiment analysis system (Vashishtha and Susan 2019a) . In Mamdani FIS, a fuzzy system with two non-interactive inputs A and B (antecedents) and a single output C (consequent) is expressed by a number of r linguistic IF-THEN propositions in the Mamdani form: where A 1j and B 1j are the fuzzy sets representing the jth antecedent or premise pairs and C 1j is the fuzzy set representing the jth consequent. In this work for single-lexicon ANFIS, we transform the set of nine Mamdani fuzzy rules in (Vashishtha and Susan 2019a) to parametrized Sugeno form where the parameters are learnt by ANFIS. Mamdani FIS entails a substantial computational burden and is well suited to human input, while Sugeno FIS is computationally efficient, has flexibility in system design and is well appropriate to mathematical analysis. Hence, our proposed system in this work applies Sugeno FIS for better classification results. The ANFIS architecture comprises five layers (Jang et al. 1997 ). Our single-lexicon ANFIS architecture is shown in Fig. 1 . We denote the output of the ith node in layer l as O l,i . Each layer is described next. Layer 1 Every node i in this layer is an adaptive node with a node function where O l,i is the membership grade of a fuzzy set A (= A 1 , A 2 or A 3 ) or B (= B 1 , B 2 or B 3 ). Here, the MF for A and B can be any parametrized MF: triangular (trimf), generalized bell (gbellmf) and Gaussian (gaussmf) described in Eqs. (26), (27) and (28) respectively. l t ðxÞ ¼ 0; x d ðx À dÞ=ðe À dÞ; d\x e ðf À xÞ=ðf À eÞ; e\x f 0; x A brief explanation of these MFs ensues. The triangular membership function has the simplest shape among others. It is defined by three parameters: d and f for feet and e for the tip of the curve. The mathematical formula of the function is given in Eq. (26). Gbellmf is a symmetrical shape similar to a bell. As expressed by Eq. (27), this function employs three parameters: u determines the width of the bell like curve, v is a positive integer, while w sets the center of the curve in the universe of discourse. Gaussmf also has a smooth curve. However, as compared to all MFs mentioned above, it utilizes only two parameters: c for locating center and r for determining the width of the curve as expressed mathematically by Eq. (28) . The values of these parameters are attuned in the learning phase of ANFIS. The parameters in this layer are referred as premise parameters. Layer 2 Every node in this layer is a fixed node labeled as G, whose output is the product of all the incoming signals shown in Eq. (29). Each node output represents the firing strength of a rule, where i = 1, 2,…, 9. Layer 3 Every node in this layer is a fixed node labeled N. The ith node calculates the ratio of the ith rule's firing strength to the sum of all rules' firing strengths in Eq. (30). Outputs of this layer are called the normalized firing strengths. Layer 4 Every node i in this layer is an adaptive node with a node function: where w i is the normalized firing strength from layer 3 and {p i , q i , r i } is the parameter set of this node. Parameters in this layer are referred to as consequent parameters. Layer 5 The single node in this layer is a fixed node labeled P , which computes the overall output as the summation of all incoming signals: MultiLexANFIS combines all the three lexicons into a single integrated unit. MultiLexANFIS has six inputs; it includes positive and negative scores of a tweet from all the three lexicons, and one sentiment label as output for each dataset. The six inputs are: vpos, vneg, apos, aneg, spos and sneg. vpos and vneg are positive and negative score of a tweet computed using VADER lexicon (Hutto and Gilbert 2014) ; similarly, apos and aneg are evaluated using AFINN lexicon (Nielsen 2011) and spos and sneg are calculated using SentiWordNet lexicon (Baccianella et al. 2010) for the same tweet. The output is the sentiment label, which has value 1 for neutral tweet and 0 for non-neutral tweet. For ease of computations and for reducing the number of fuzzy rules, each input has been assigned two membership functions (MFs): Low and High. The labels Low and High are two fuzzy sets created using different fuzzy membership for input variables: vpos, vneg, apos, aneg, spos and sneg. Hence, there are a total of 64 rules (2^6 = 64). These rules are depicted in Tables 1 and 2. These proposed novel rules are domain independent. We can apply these rules for any textual data that employs lexicons to extract positive and negative scores for each textual document or sentence or review. MultiLexANFIS architecture has five layers similar to the single-lexicon ANFIS discussed in Sect. 3.3.1, except for the fact that the number of inputs has increased from two to six and the number of MFs for each input is two. MultiLexANFIS architecture is depicted in Fig. 2 . ANFIS learns by updating tunable parameters which are the premise parameters and the consequent parameters. The first and fourth layers have adaptive nodes, so they are trainable, whereas the nodes of rest of the layers are fixed. The optimization method adopted by ANFIS is, hybrid learning algorithm, composed of two passes. The first pass is known as forward pass; here, ANFIS computes node outputs until the fourth layer where it uses least square estimate method to update the consequent parameters before calculating the final output. In the second pass or backward pass, the error is propagated Fig. 1 Single-lexicon-based ANFIS architecture with nine rules Neuro-fuzzy network incorporating multiple lexicons for social sentiment analysis backward until the first layer where ANFIS employs gradient descent to tune the premise parameters. In forward pass, the antecedent part is fixed, while in backward pass the consequent part is fixed. The cost of training and computational complexity of the network is ascertained by the apt choice of the number and shape of membership functions. The performance of ANFIS is measured using root mean square error (RMSE). It is the most popularly used measure. It is a measure of the differences between values predicted by the model and the observed values, as shown in Eq. (33). In other words, it measures the quality of the fit between the actual data and the predicted model. where actual(t) and predicted(t) are the actual and estimated values, respectively, and n is the total number of records in the testing data, t = 1,2,3… p. The overall process flow of our proposed ANFIS for social SA is summarized in Fig. 3 . In the experimental phase, the proposed ANFIS was simulated in MATLAB 2020b version with Intel Core i5 processor, 64-bit operating system and 8 GB RAM. We have used a total of ten publicly available datasets: Apple Twitter Dataset (2019) Table 3 . We have combined the positive and negative tweets into one class: non-neutral; our goal is to classify the tweets by ANFIS into two classes: neutral and non-neutral. The training and testing sets of these datasets are constructed using 70-30% split ratio. Our proposed model uses inputs derived from sentiment lexicons. The three sentiment lexicons used in this study are: VADER (Hutto and Gilbert 2014) , AFINN (Nielsen 2011) and SentiWordNet (Baccianella et al. 2010 ). Our proposed model is compared with eight state-of-the-art methods that include both fuzzy and non-fuzzy approaches; their description is mentioned in the next section. In our single-lexicon-based ANFIS variants, there are two inputs: positive and negative scores, and one output for sentiment. Class 1 indicates neutral tweets, and Class 0 indicates non-neutral tweets. The number of fuzzy rules in grid partitioning is number of MFs^number of input variables. In single-lexicon-based ANFIS variant, since Fig. 3 Overall process of our proposed ANFIS for social SA Neuro-fuzzy network incorporating multiple lexicons for social sentiment analysis each input has three membership functions (MFs), so in total, ANFIS has nine rules (3^2 = 9). In MultiLexANFIS, there are six inputs (two from each of the three lexicons) and the same output variable for sentiment, and each input has two MFs. Hence, it has a total of 64 rules (2^6 = 64). Optimal number of inputs and MFs are selected for both the cases. In MultiLexANFIS, we have considered two MFs compared to three MFs in single-lexicon ANFIS variants in order to reduce the curse of dimensionality. The ''curse of dimensionality'' refers to such situation where the number of fuzzy rules, when the grid partitioning is used, increases exponentially. Grid partitioning generates rules by enumerating all possible combinations of membership functions of all inputs; this leads to an exponential explosion even when the number of inputs is moderately large. ANFIS with grid partitioning was used to test different MFs: triangular, Gaussian and generalized bell. The optimization method used to train the ANFIS is hybrid learning, which combines the gradient descent method and the least squares estimate (LSE) to learn the parameters. For all the variants, the error tolerance is set to 0. While conducting experiments, it was observed that training error in single-lexicon ANFIS variants remains constant after 500 epochs, while in MultiLexANFIS, the training error remains constant after 100 epochs. Hence, the number of epochs required by single lexicon is 500 epochs, while MultiLexANFIS requires 100 epochs during training phase. In this paper, we have proposed two ANFIS for sentiment analysis of social media data: single-lexicon-based ANFIS and MultiLexANFIS, that incorporates inputs from multiple lexicons. These systems have been implemented on ten datasets. Root mean square error (RMSE) is the mostly used metric in most of the papers involving ANFIS models. Thus, we have used this as the performance metric in all experiments. All the RMSE values depicted in this work are test errors. Our ANFIS has been compared with eight state-of-theart methods that include various fuzzy, non-fuzzy and deep learning models. The eight state-of-the-art methods are: (1) A deep learning model-recurrent neural network (RNN), developed by (Nemes and Kiss 2021) for classifying tweets. (2) A deep learning model-long short-term memory (LSTM) model, developed by (Rahman et al. 2016 ) for text classification. (3) An n-gram model developed by (Tripathy et al. 2016 ) that converts text reviews into numeric matrices using count vectorizer and TF-IDF, which are then given as input to machine learning algorithm-support vector machine (SVM) for classification. (4) An unsupervised fuzzy rule-based system for sentiment analysis is a published previous work of authors (Vashishtha and Susan 2019a) . This fuzzy rule-based classifier has nine rules implemented in a Mamdani fuzzy inference system. (5) VADER SA by Hutto and Gilbert (2014) ; this approach yields positive, negative, neutral scores for a given text by deploying a number of rules. (6) AFINN SA by Nielsen (2011) approach uses AFINN dictionary to assign scores to each word; the sentiment is computed for text by summing up all scores. (7) SA using SentiWordNet lexicon by Cavalcanti et al. (2011) approach uses SentiWordNet to assign scores to selected wordsadjectives, verbs, adverbs and nouns, and then, final sentiment is computed by considering the average score of all words. (8) SA using SentiWordNet lexicon by Ortega et al. (2013) approach uses SentiWordNet to check the polarity of words, assigns fixed values for each polarity and then computes the sentiment by summing up the values. In this research work, we have considered three sentiment lexicons, so there are three single-lexicon-based ANFIS: VADER (Hutto and Gilbert 2014) , AFINN (Nielsen 2011) and SentiWordNet (Baccianella et al. 2010 ) single-lexiconbased ANFISs. These systems have nine rules. We have performed experiments with each of these ANFISs for all the datasets and three MFs; the aim is to find out which MF is best suitable for the given ANFIS. Table 4 depicts the outcomes for VADER lexicon-based ANFIS, and the bold values highlight the lowest RMSE achieved for each dataset. Both trimf and gaussmf membership functions Positive 519 10 423 2375 5157 4377 182 2363 72,249 15,829 Negative 572 19 1219 3972 1225 1745 177 9178 35,509 8278 Neutral 2333 161 2162 5937 2667 5593 139 3099 55,212 12,909 Total 3424 190 3804 12,284 9049 11,715 498 14,640 162,970 37,014 S. Vashishtha, S. Susan (MFs) have lower RMSE in a greater number of datasets compared to gbellmf. The least RMSE, 0.33174, is scored in nuclear dataset under gaussmf, while in the case of AFINN-based ANFIS in Table 5 , gbellmf has performed better than trimf and gaussmf; it has the lowest RMSE for three datasets. The least RMSE, 0.3873, is scored in Airline Twitter dataset under gbellmf. In Table 6 , we can observe the results for SentiWordNet-based ANFIS; here, trimf and gbellmf have more cases of lower RMSE compared to gaussmf. Again, the Nuclear dataset has least RMSE of 0.3635 under the trimf. Hence, we cannot identify a single MF which works well for all the three variants. It can be deduced that trimf and gbell MFs are better as in majority cases they yield lower RMSE. Comparison of VADER lexicon-based ANFIS with VADER-specific approaches is shown in Table 7 . It has been compared with two approaches-fuzzy rule approach (Vashishtha and Susan 2019) and the non-fuzzy approach (Hutto and Gilbert 2014) . It can be clearly observed the our ANFIS outperforms all methods for all datasets. For instance, in Apple dataset RMSE of ANFIS is 0.45668, while other methods have 0.6127 and 0.6223. Table 8 presents the achievement of AFINN lexicon-based ANFIS with AFINN-specific methods. It has been compared with two approaches-fuzzy rule approach (Vashishtha and Susan 2019) and the non-fuzzy approach (Nielsen 2011) . Our ANFIS has the lowest RMSE as compared to both approaches for all datasets. For SemEval 2017 dataset, ANFIS scored 0.47428, while others scored 0.6127 and 0.7046, respectively. Similar observations can be made for SentiWordNet lexicon-based ANFIS with SentiWordNetspecific approaches in Table 9 . It has been compared with three approaches-fuzzy rule approach (Vashishtha and Susan 2019) and two non-fuzzy approaches: Cavalcanti (Cavalcanti et al. 2011 ) and Ortega (Ortega et al. 2013) . Here also, ANFIS surpasses all methods, for example, for the SemEval 2016 dataset, RMSE for ANFIS is 0.45204 and for others: 0.554, 0.7164 and 0.8115. Hence, we can deduce that all single-lexicon-based ANFISs have lowest RMSE for all datasets compared to state-of-the-art SA models. But it is difficult to decide the best membership function (MF) for each lexicon among all datasets. In other words, we cannot conclude which is the best MF among trimf, gbellmf and gaussmf for the VADER, AFINN or SentiWordNet lexicons. In VADER, there is a confusion among trimf or gaussmf; in AFINN, we can choose gbellmf as it performs slightly better, while in SentiWordNet it is hard to choose between trimf or gbellmf. The execution To solve the above confusion of choosing the best membership function (MF) for a given lexicon, the Mul-tiLexANFIS architecture has been proposed. This system integrates all the lexicons into a single unit. Using each lexicon, two scores: TweetPos, positive score, and Tweet-Neg, negative score, are computed as discussed in Sect. 3. Hence, this integrated system has a total of six inputs, with two MFs for each input, 64 rules and one sentiment label as output. Table 10 displays the performance of MultiLexANFIS for all datasets. This system has been compared with single-lexicon-based ANFIS as well; the best (lowest) RMSE-combination of lexicon-membership function for each dataset, is represented in the second column of Table 10 . For instance, for the Apple dataset, among single-lexicon ANFIS, VADER variant has achieved the lowest RMSE of 0.457 using the gbell MF, while Mul-tiLexANFIS has attained 0.448 RMSE using the gauss MF, which is the lowest value. Similar analysis is accomplished for all the datasets. This integrated MultiLexANFIS achieves lower RMSE as compared to the single-lexiconbased ANFIS for almost all datasets, though for some datasets the results are comparable. We also note down that trimf: triangular membership function performs the best in MultiLexANFIS (7 out of 10 datasets) and gaussmf: Gaussian membership function is at second position. Thus, the problem of confusion among the MFs for all datasets has been resolved. Comparison amidst the single-lexicon ANFISs reveals that VADER lexicon performs best. This is observed in the second column of Table 10 ; among ten datasets, VADER lexicon has the lowest RMSE for six datasets. The execution time of MultiLexANFIS is about 30 min. The training error graph for MultiLexANFIS for the Apple dataset with gaussmf is shown in Fig. 4 . The computed test error, 0.448, is depicted in Fig. 5 . The comparative analysis of our proposed MultiLexANFIS with eight state-of-the-art methods that include various fuzzy, non-fuzzy and deep learning models is presented in Table 11 . The following are the state-of-the-art methods: (1) a deep learning model-recurrent neural network (RNN) developed by Nemes and Kiss (2021) for classifying tweets, (2) a deep learning model-long shortterm memory (LSTM) model developed by Rahman et al. (2016) for text classification, (3) an n-gram model with SVM classifier developed by c, (4) an unsupervised fuzzy rule-based system for sentiment analysis which is a published previous work of authors (Vashishtha and Susan 2019a) , (5) Neuro-fuzzy network incorporating multiple lexicons for social sentiment analysis these models is higher than MultiLexANFIS, and this is because the parameter tuning and optimization in ANFIS are better than the tuning of weights during backpropagation in deep learning models. In Table 11 , for instance we can observe for the Airline Twitter dataset, Mul-tiLexANFIS has the lowest RMSE value-0.3843, while RNN has 0.4417 RMSE and LSTM has 0.4383 RMSE. Overall, we can deduce that after MultiLexANFIS, RNN is second best followed by LSTM. The third approach (Tripathy et al. 2016) creates different n-grams-unigrams, bigrams and trigrams from text, and assigns term frequency-inverse document frequency (TF-IDF) values to each n-gram; these numerical vectors are sent as input to SVM classifier. The errors produced by the n-gram model are higher than MultiLexANFIS for all datasets, and this is because of missing fuzzy factor in the extracted features and in the model. The selected n-grams are based on frequency, while in our model we use the polarity and sentiment score of each word derived from lexicons; this way we are able to deal with natural language in a better way for SA. The proposed MultiLexANFIS is also compared to fuzzy rule classifier's (Vashishtha and Susan 2019a) three variants-VADER, AFINN and SentiWordNet. Our neurofuzzy model yields better results by using scores from sentiment lexicon as features for subjectivity classification (neutral non-neutral) compared to the unsupervised fuzzy rule classifier which uses the same scores for polarity classification (positive, negative or neutral). In fuzzy rule classifier, triangular (trimf) fuzzy membership is used and its parameters are fixed, while in our ANFIS model these parameters are tuned during each epoch, thereby producing lower errors. The next four state-of-the-art methods are unsupervised and non-fuzzy algorithms. They produce very high errors compared to all other approaches, and this is because they are simple and naïve in nature. Finally, we can conclude that the proposed MultiLexANFIS produces lower error for all datasets compared to other approaches and also it takes less execution time in contrast to singlelexicon-based ANFIS. In this paper, a novel MultiLexANFIS has been proposed for social sentiment analysis. This architecture is an adaptive neuro-fuzzy inference system (ANFIS) that incorporates inputs from VADER, AFINN and Sen-tiWordNet lexicons to perform sentiment analysis of social media content. The key highlights of this work are: (1) a novel neuro-fuzzy system: MultiLexANFIS that combines sentiment scores from multiple lexicons to classify tweets into neutral and non-neutral tweets, (2) a novel set of 64 rules for Sugeno-type fuzzy inference system-Mul-tiLexANFIS, (3) single-lexicon-based ANFIS variants to classify tweets and (4) comparison of fuzzy approaches, non-fuzzy approaches and deep learning models with MultiLexANFIS that affirms the supremacy of our proposed neuro-fuzzy system. ANFIS has deployed optimization with the help of a hybrid learning algorithm; that uses least square method and gradient descent to tune parameters. We can deduce that MultiLexANFIS has the Neuro-fuzzy network incorporating multiple lexicons for social sentiment analysis lowest RMSE for the ten benchmark datasets, as compared to other approaches. Results reveal that among the singlelexicon ANFISs, VADER lexicon performs best. The best membership in MultiLexANFIS is triangular (trimf), and Gaussian membership function (gaussmf) is at second position. The problem of confusion among the MFs for all datasets has been resolved using MultiLexANFIS. The MultiLexANFIS is computationally faster than the total combination of three single-lexicon-based ANFIS for all lexicons. Our neuro-fuzzy network is targeted for applications that aim to search neutral social media content posted by the users. Our results thus indicate an effective and computationally feasible solution to social sentiment analysis that incorporates multiple lexicons onto the same platform. These proposed novel rules are domain independent. We can apply these rules for any textual data that employs lexicons to extract positive and negative score for each textual document or sentence or review. The proposed system can handle only those twitter datasets which contain neutral tweets. At present, it cannot deal with sarcasm and mixed tweets, i.e., the tweets that contain both positive and negative opinions. In future, we will compute the sentiment of mixed tweets by enhancing our model using advanced fuzzy deep neural networks. Funding This study was not funded by any agency. Conflict of interest Both authors are associated with the Department of Information Technology, Delhi Technological University. Ethical approval This article does not contain any studies with animals performed by any of the authors. Human and animal rights This article does not contain any studies with human participants or animals performed by any of the authors. A hybrid computational intelligence approach for efficiently evaluating customer sentiments in E-commerce reviews Stock market forecast using sentiment analysis Fuzzy logic based hybrid approach for sentiment analysis of Malayalam movie reviews Sentiment analysis of bangla microblogs using adaptive neuro fuzzy system Sentiwordnet 3.0: an enhanced lexical resource for sentiment analysis and opinion mining An adapted Lesk algorithm for word sense disambiguation using WordNet Application of soft computing techniques for estimating emotional states expressed in TwitterÒ time series data Good to be bad? Distinguishing between positive and negative citations in scientific impact Neuro-fuzzy sentiment analysis for customer review rating prediction Twitter brand sentiment analysis: a hybrid system using n-gram analysis and dynamic artificial neural network From bias to opinion: a transfer-learning approach to real-time sentiment analysis A comparative study of ANFIS membership function to predict ERP user satisfaction using ANN and MLRA Vader: a parsimonious rule-based model for sentiment analysis of social media text Fuzzy logic based sentiment analysis of product review documents ANFIS: adaptive-network-based fuzzy inference system Neuro-fuzzy and soft computing-a computational approach to learning and machine intelligence Fuzzy approach for sentiment analysis Adaptive neuro-fuzzy inference system for classification of texts Sentiment analysis by POS and joint sentiment topic features using SVM and ANN A hybrid adaptive neuro-fuzzy interface and support vector machine based sentiment analysis on political twitter data Neuro-fuzzy network incorporating multiple lexicons for social sentiment analysis Predicting stock market trends using machine learning algorithms via public sentiment and political situation analysis Fine-grained sentiment analysis for measuring customer satisfaction using an extended set of fuzzy linguistic Hegdes Neuro fuzzy approach for predicting sales performance of movies considering sentiments in online reviews S-ANFIS: sentiment aware adaptive network-based fuzzy inference system for Predicting Sales Performance using Blogs/Reviews Using neutral examples for learning polarity The importance of neutral examples for learning sentiment Fuzzy rule based systems for interpretable sentiment analysis Optimized fuzzy technique for enhancing sentiment analysis A fuzzy logic approach to predict the popularity of a presidential candidate A novel framework based on bi-objective optimization and LAN 2 FIS for Twitter sentiment analysis Social media sentiment analysis based on COVID-19 A multimodal convolutional neuro-fuzzy network for emotion understanding of movie clips SSA-UO: unsupervised Twitter sentiment analysis Twitter sentiment analysis using adaptive neuro-fuzzy inference system with genetic algorithm Sentiment analysis on twitter data-set using naive Bayes algorithm A method for detecting and analyzing the sentiment of tweets containing fuzzy sentiment phrases A new LSTM model by introducing biological cell state Sentiment analysis using Neuro-Fuzzy and Hidden Markov models of text Sentence-level subjectivity detection using neuro-fuzzy models A hybrid system for subjectivity analysis Sentiment analysis of online product reviews using DLMNN and future prediction of online product using IANFIS Combining a rule-based classifier with weakly supervised learning for twitter sentiment analysis Fuzzy logic applied to opinion mining: a review Structure identification of fuzzy model Fuzzy min-max neural networks for business intelligence An investigation of membership functions on performance of ANFIS for solving classification problems Classification of sentiment reviews using n-gram machine learning approach Fuzzy logic based dynamic plotting of mood swings from tweets Fuzzy rule based unsupervised sentiment analysis from social media posts Sentiment cognition from words shortlisted by fuzzy entropy Inferring sentiments from supervised classification of text and speech cues using fuzzy rules Fuzzy interpretation of word polarity scores for unsupervised sentiment analysis Highlighting keyphrases using sentiscoring and fuzzy entropy for unsupervised sentiment analysis Neutrality in the sentiment analysis problem based on fuzzy majority Consensus vote models for detecting and filtering neutrality in sentiment analysis A fuzzy logic-based text classification method for social media data An improved algorithm for sentiment analysis based on maximum entropy A multigene genetic programming-based fuzzy regression approach for modelling customer satisfaction based on online reviews Decision support system for candidates classification in the employment process based on ANFIS method Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations Neuro-fuzzy network incorporating multiple lexicons for social sentiment analysis