Sentiment Analysis of Shared Tweets on Global Warming on Twitter with Data Mining Methods: A Case Study on Turkish Language

As the usage of social media has increased, the size of shared data has instantly surged and this has been an important source of research for environmental issues as it has been with popular topics. Sentiment analysis has been used to determine people's sensitivity and behavior in environmental issues. However, the analysis of Turkish texts has not been investigated much in literature. In this article, sentiment analysis of Turkish tweets about global warming and climate change is determined by machine learning methods. In this regard, by using algorithms that are determined by supervised methods (linear classifiers and probabilistic classifiers) with trained thirty thousand randomly selected Turkish tweets, sentiment intensity (positive, negative, and neutral) has been detected and algorithm performance ratios have been compared. This study also provides benchmarking results for future sentiment analysis studies on Turkish texts.


Introduction
Downpour, storm, rising temperatures, sea level, and retreating glaciers are considered as the main headlines among the indicators of climate change [1][2][3][4][5][6]. anks to the popularity of Twitter and easily accessible Application Program Interface (API) [7][8][9], tweets can be stored by topics related to hashtags. In addition to academic researchers, many firms pay attention to Twitter mainly because of commercial purposes. ese firms also use Twitter to interact with their investors and customers. Comparatively to the traditional media, Twitter's impact is obvious. However, to take advantage of Twitter data, firms require to store and analyze these substantial data produced by Twitter daily. In 2018, more than 336 million active users tweet more than 500 million times per day [10].
Social media and especially Twitter are getting more popular, and its domain becomes stronger than traditional media tools. More users in the social media means more data to access.
For this reason, data-based applications like disaster detection, election predictions, information filtering, and influencing opinions make use of this trend. One of these is sentiment analysis [11][12][13], which is one of the most attractive fields [14].
Sentiment analysis [15,16] is based on the language of a text, and modeling is established by a text from the same language. Because of that, in literature text, analysis of Turkish is limited, and mostly, it is emphasized in English. Due to word structure of Turkish being different from English, the analysis is more complicated.
Machine learning methods [12,[17][18][19] have been commonly used in emotion analysis problems in previous studies. Pang et al. work on compared several machine learning methods to determine the characteristics of emotions [20]. Kaur et al. present that support vector machine (SVM) is used as a hybrid method to analyze emotion on English Twitter data [21]. Taboada et al. have been worked on a label assignment process, which reflects a positive or negative emotion has been used by using a dictionary-based approach [22]. In [23], in the use of Turkish-based approach on the study of artificial neural networks, support vector machines, Naive Bayes, and K-NN neighbors using various machine learning methods have been compared by the results. Since the dictionary-based approach studies are not enough for the studies of Turkish sentiment analysis, the dictionary has been formed by various methods. According to the results of studies, it has been seen that the emotion analysis studies conducted on Turkish texts are relatively low and insufficient compared to the studies which were conducted with English texts. Since the structure of the Turkish language is different according to the English language, an approach to the Turkish language needs to be developed in order to achieve a high success in the sentiment analysis. In this study, we aimed to compare the effect of different quality selection methods on the performance of classification in the sentiment analysis on Turkish Twitter posts. Unlike other similar studies, an integrated classification method is recommended. Additionally, Turkish NLP library has been used differently to reduce the number of features.
In this study, emotional textual analysis is implemented regarding the sensibility of society towards climate change, one of the most important environmental threats. In the first part, data collection is processed. Feature selection for modeling is described in Section 3. Sentiment analysis models are described in Section 4. In the final section, classification and conclusive comparisons take place.

Materials and Methods
Generally, classifiers can be categorized in many ways, namely, with being supervised or unsupervised. In order to test different methodologies, different classifiers belonging to relatively different realms of classification are chosen, namely, Naïve Bayesian, K-NN (nearest neighbor), and support vector machine (SVM).

Naive Bayesian.
As mentioned in previous chapters, Weka Software is used in all analyses. Naïve Bayes classifier in Weka uses probabilistic Naïve Bayes classifier, which is used as descriptive and complementary classifier algorithm, mainly making use of Bayes rule, shown as follows: Naïve Bayes is based on learning from data, and it means that, in order to learn model occurrence of every output calculated, it is named as prior (second term of nominator in equation (2)). Likelihood probability (first term of nominator in equation (2)) is then calculated and multiplied and divided by normalization constant (denominator term in equation (2)).

K-NN (Nearest Neighbor).
In pattern recognition, the K-nearest neighbor algorithm (or K-NN for short) is a nonparametric method used for classification and regression. It is based on the idea that instance must be in a close distance when compared to its closest neighbors [24].

SVM (Support Vector Machine)
. SVM algorithm is a supervised learning algorithm and binary classifier [25]. It is mostly used to solve classification problems [26]. Support vector machine (SVM) is used to separate data belonging to two classes in a most suitable way; to implement this procedure, hyperplanes are specified [27].

Proposed System
In this section, data preparation process is explained before classification. Sentiment analysis through texts is classified using Turkish language. erefore, Turkish tweets are taken with hashtag-related global warming on Twitter. In the next section, the word roots of the sentences are found, and data pollution is reduced.

Data Collection. Twitter API (Application Programming
Interface), like the other APIs, is an independent platform gathered to the developers which is separate from the main website accessed by the main users. e platform sends the JSON (Java Script Object Notation) response value. JSON response value consists of tweet object, user information, text of the tweet, upload date, and location data.
As indicated in Figure 1, in the Visual Studio platform by using C# programming language TwitR library, 848 tweets in Turkish with hashtags "#iklimdegisikligi," "#kureselisinma," and "#iklimetkisi" are stored in the Microsoft Sql database.
In our study, Hayran shared 32 thousand data from his work and had classified content, as a train set after a preoperation. e point that should be emphasized here is that Turkish is a head final language. erefore, adverbs of time go to the end of the verb in the sentence. In order to minimize semantic shifts and decrease the number of features that would arise, these data are passed through the data preprocessing phase of the 2 nd stage in Figure 2. us, we have a pure sentence data that is free of punctuation marks and can reduce the semantic shifts. Additionally, since Turkish is a head final language, the adverbs of time are added adjacent to the verb, which will increase our feature number and reduce our chances of successful classification. In order to avoid this, it is aimed to achieve a more effective result by reaching the roots of the words by applying "word stemming" process in Figure 2 and the last step of the 2 nd stage with the Turkish NLP library "Zemberek." 3.1.1. Data Preprocessing. Tweet texts are usually lacking a formal writing standard and because of that each text is purified by implementing the steps in Table 1 to create a sounder model [30,31]. Purpose of the data preprocessing is to achieve more sensible results by decreasing the size of feature [32][33][34].
For word stemming, a Turkish NLP library named Zemberek is used. Because of having an MPL licence, general use is allowed. anks to this library, after the purification of text after first four steps, roots of the words within the text are determined. After the specified procedures, all data stored in Ms-Sql database are imported to Zemberek library for .Net technology, and then, the word stemming process is implemented. erefore, the data preprocessing is concluded to achieve a solid NLP process. In addition, data evaluation progress is presented step by step in Figure 2.  Computational Intelligence and Neuroscience

Feature Selection
In this section, numerical equivalents of processed word data are shown, and then, classification methods for emotional analysis are implemented. For word splitting (tokenizer) and feature removal processes, the N-Gram technique is used. It relies on prediction and probability and is studied based upon two main headlines: word and character. In this study, word-based calculation is used. It is described as the probability of a word's position in the sentence related to the preceding word. Gram expresses the weight of the controlled value [19]. In this study, 1-3 is held as constant. According to Markov chain, certain words follow each other frequently, and because of that based on equation (1), it is multiplication of words' conditional probabilities: If we look at each tweet according to equation (4), P (global warming problem) � P (warming|global) * P (problem | warming), this is how the multiplication of conditioned probabilities is calculated.

Classification
In the phase of sentiment analysis and classification of tweet data, as the first step, 891 tweets that were pulled from certain hashtags are classified based on emotion (positive is 1 and negative is 0) and separated as test data. As the second step, 16000 positive and 16000 negative tweets are produced and classified beforehand in Hayran et al.'s [35] study and are used as training dataset and attributes, as listed in Table 2 and Figure 3.
Naive Bayes, one of the techniques of supervised machine learning, is subjected to K-NN [24,36] and SVM classification algorithms [37,38]. During the procedure of classification, WEKA machine learning tool is used. Used algorithms are explained in the further sections.
Naive Bayes: through the probability procedures implemented within this dataset, classification of the incoming test data is determined, and it is mostly used in word mining classification. Mainly make use of Bayes Rule, P(c | x) is the posterior probability and P(c | x) likelihood [39], as shown in the following equations: K-NN (nearest neighbor): K is used to determine the class of the new data and to store all conditions based upon the distance measure of the nearest neighbor. K-NN is mostly used in pattern recognition and estimation as a nonparametric technique [40]. K value means that how many neighbors should be taken into consideration.
SVM: SVM algorithm is a supervised learning algorithm and binary classifier [25]. It is mostly used to solve classification problems [26]. Support vector machine (SVM) is used to separate data belonging to two classes in a most suitable way, and to implement this procedure, hyperplanes are specified [27].

Comparative Performance Analysis. Hayran et al. choose
the SVM algorithm as the classifier design. ey determine the sentiment classification by labelling the texts as a training data. Labelling process is executed manually through using emoji expressions (:), :(, etc.). SVM model is tested with the k-fold cross validation method. e main reason for the performance value (80.05% accuracy in Table 3) of our study to be lower compared to the work of Hayran et al. is the creation of a training set without removing emotional symbols like smile emotion symbol ":)" and sad emotion symbol ":(" that would significantly affect the classification in their study. If our model has worked hard on our dataset for training in this case, our model starts to memorize. At the same time if our training set is uniform, the risk of overfitting will be high. erefore, in order to avoid overfitting in our study, an integrated classification method is suggested by removing these sentiment expressions and symbols from our training set. us, this is an important factor in model training and successful classification compared to our study. Erdogan et al. have achieved the highest success rate in their study by making a classification without distinction between Turkish and English text. ey used the logistic regression method as a classification tool in their work. Compared to our study, the use of the English dataset and the inclusion of sentimental emoji increased the rate of successful classification. According to similar studies in the table, the logistic regression classifier has been used in four studies. Accuracy results varying between 65% and 94% have  { } @data been achieved in studies by using this method in Table 3. Ecemis et al. who reached the most successful result have been carried out the classification process using the SVM method. ey have performed the classification process by using a manually chosen text set as a training set. eir study presents that, to complete each sentiment class, strong sentimental words in Turkish are used. It has been observed that the selection of sentences containing only adjectives as a training set increases the success rate. Support vector machine classifier used in this study has been preferred as a classification tool in the other three studies in the table. Accuracy results varying between 64% and 80% have been  Computational Intelligence and Neuroscience 5 achieved in studies by using this method. In similar studies, it has been observed that SVM, linear regression, and other deep learning methods are mostly preferred as classification tools. e main factor in achieving different performance rates of studies using the same algorithm is the selection of training sets in different structures. It has been observed that the dataset used as a training set increases the success rate of sentences based on certain conditions (emotional symbols and strong sentimental words). Figures 4-6, performance measure comparison for the dataset subjected to classification algorithms via WEKA is shown in Table 4 according to the evaluation measures. In Table 4, it is shown that K-NN algorithm is more successful than others.

Performance Results. With the established model in
We reached as a result 74.63 percent accuracy on this enhanced algorithm. In the text preprocessing, using "Word to Vector" as "n-gram" algorithm and taking advantage of Zemberek library increased rates of success to find word roots. We use the  evaluations metrics which are precision, sensitivity, F-measure, and accuracy [55,56]. ese metrics are depending on TP (true-positive) and FP (false-positive) ratios: accuracy � (TP + TN) (P + N) .
In the case of reducing the number of variables and achieving more successful results, Zemberek is used as the Natural Language Processing library to find the roots in each Turkish tweet. Since each word in the sentence does not always make sense alone, vectors are created by using the Ngram (2, 3) technique that treats words in dual and triple groups. In order to find out whether this integrated technique is successful, testing has been conducted on three different classification algorithms. e results of the proposed method have been tested on the Turkish tweet dataset, and the highest performance rate has been obtained with the K-NN classification algorithm used with integrated technique in Table 4. Compared to other classification algorithms in Figure 7, the highest success rate (74.63%) has been achieved by using the K-NN classification algorithm with the N-gram technique because the K-NN classification algorithm classifies on the basis of the closest neighbor proximity as the principle of operation. erefore, with the N-gram (2.3) technique that creates vectors by considering the frequency of using words together, more successful results are obtained compared to other classification algorithms in Table 5.

Conclusion
With the growth of social media in recent years, it has become an important research resource for people's ideas on specific issues. Accordingly, emotion analysis on texts with social media data has been the subject of research. It is important for forward-looking plans with measures and actions to make an emotion detection on subject. We summarized the results of the comparative methods for the analysis of social sensitivity and mentioned promising aspects in this field. e dataset used to support the findings of this study has been deposited in the "Sentiment Analysis on Turkish Tweets Dataset" repository on online data library [57,58].
In this study, posted tweets about climate change, which is one of the biggest environmental topics, associated with this attempted to establish automatically emotion analysis. erefore, a way is clear to ascertain public opinion and   Computational Intelligence and Neuroscience precautions about environmental topics by emotion analysis. We observed that using integrated classification methods instead of a single machine learning technique increased the success rate of accuracy. Considering the high rate of double or triple word groups in Turkish language, it is recommended to choose this integrated method in emotion classification studies. Using word splitting (tokenizer) in the phase of data preprocessing, "Zemberek" library for finding word roots and recommended integrated solution as N-gram for the feature removal process by using K-NN classification machine learning algorithm increased success rates of text analysis, in the case of especially texts in Turkish language.

Data Availability
All the raw data will be made available if needed.

Conflicts of Interest
e authors declare that there are no conflicts of interest.