View of MLSSDCNN: Automatic Sentiment Examination Model Creation using Multi Domain Light Semi Supervised Deep Convolution Neural Network

(1)

__________________________________________________________________________________

4613

MLSSDCNN: Automatic Sentiment Examination Model Creation

using Multi Domain Light Semi Supervised Deep Convolution

Neural Network

Manasa K N, M. C. PADMA

Abstract: Type Emotional research deliberated at categorizing positive or negative has fascinated more and more consideration in recent years. Traditional methods of emotional categorizing typically work well with labeled data. Transfer training is a popular method to solve problems where the model cannot be applied straight to the target domain in multi-domain emotional categorizing framework. This article intended a transfer learning system derived from the Light Semi Supervised Deep Convolution Neural Network (LSSDCNN). This LSSDCNN build a convolutional neural network model using extracted features moreover distribute the weights among all layers. In this article, LSSDCNN used five broad categories of Linguistic Inquiry and Word Count (LIWC) that includes linguistic and human psychological characteristics. Intellectual Broad Multi Domain Data Transfer Network employs sentences as well as aspects for extracting unique data. Particularly, this network intended to discover general features of multiple domains after that the retrieved aspects information through general features. To find out the predictive performance of proposed MultiDomain Light Semi Supervised Deep Convolution Neural Network (MDLSSDCNN), it is compared with existing sentiment models.

Keywords: Emotion Categorizing, Convolution Neural Network, Light Semi Supervised Deep Convolution Neural Network, Linguistic Inquiry, Word Count, Twitter, CNN and Intellectual Broad Multi Domain Data Transfer Network 1. Introduction

An examination of emotions associated with the use of ideas, assessments, behaviors, and emotions for individuals, issues, topics, and their characteristics [1]. The task is technically difficult but very useful to perform. For example, companies always want to know what customers think about their products and services [2]. When it is necessary to understand whether a plain text document has a positive, negative or neutral orientationsentiment classification is involved.This method of control learns a sample from a labeled set of files and then applies it to an unlabeled test in which polar lines (e.g., positive, negative, or neutral) are detected [3]. The conventional method of emotion classification assumes that both training tools and test sets relate to a single subject.Suppose the model [4] learns a set of book reviews and applies it to a separate set of reviews, but always for a book. The implementation of this activity, known as the emotional partitioning of the domain, ensures optimal performance where files from the same domain are similar. However, this approach is often not applicable in practice where most documents are unlabeled.Twitter comments on social media may contain comments, but there is no information as to whether they are positive, negative or neutral. Distributing file types by human experts is the only way to solve such problems to learn about the distribution of emotions in a domain. This method becomes impossible when very large sets of text are required to be labeled. It would be useful if a scientific model of the source domain could be used to classify the specificity of documents within a separate target domain.

To learn the emotion of the sentences from unlabeled data, Multi-domain sentiment classification has been proposed as a promising direction[5]. It uses effectual information in the source domain (with sufficient labeled data) to help sentiment classification in the target domain (with few or no labeled data). As it is crucial for reducing the reliance on the massive amount of labeled data and significant for domains which are lack of labels, much research attention has been attracted from both academia and industry[6]. In the literature, many methods have been proposed to solve the Multi-domain sentiment classification problem[7-8]. But these approaches mainly concentrate on extracting common features between domains. Unfortunately, they cannot fully consider the effects of the aspect information of the sentences. To overcome this drawback the Intellectual Broad Multi Domain Data Transfer Network (IBMDDTN) is developed. This IBMDDTN considers the information from both sentences and aspects. Specifically, this network is aimed to find common features a Multiple domains and then extracts information from the aspects with the help of common features. Additionally, it adopts an extensive information rich learning mechanism for the sentiment classification, which combines sentences and aspects together. Additionally, it adopts an extensive information rich learning mechanism for the sentiment classification, which combines sentences and aspects together.

In addition, multi-domain emotion classification can accept well-trained rankings from one source to another target domain, which reduces the time and effort of training new classifiers in the domain. These. Existing methods for multidimensional emotion classification require data or other information in the target domain to

(2)

__________________________________________________________________________________

4614

train their models. Therefore, the cost of labeling each domain is very high as well as time consuming.However, collecting and running new corporations requires a heavy workload. In addition, data in the target domain may be private and not always available for training. To overcome these shortcomings, a development approach has been proposed that extracts and classifies comments from one domain, called the source domain, and provides feedback from another domain, called the target domain, using MultiDomain Light Semi Supervised Deep Convolution Neural Network (MDLSSDCNN).

The remaining articles are organized as follows. Section 2 presents work related to emotional examination. Section 3 presents research methods and Section 4 presents experimental procedures and actual outcomes. Section 5 describes the conclusion.

2. Related Work

There have been three or more labeling studies (e.g., evaluation of thoughts, emotions) [9 - 13], as well as a few learning have approved two-dimensional labels [14–26]. Despite significant improvements in the performance of emotional examination, binary classification of emotions acts as a challenge. For instance, the accuracy values of current learning's varies from 70% -90% in conditions of data uniqueness.

There have been several studies to classify emotions with machine learning methods for instance Support Vector Machine (SVM), Naive Bayes (NB), Maximum Entropy (ME), Stochastic Gradient Descent (SGD), as well as ensemble. N-grams is one of the frequently employed methods of machine learning. Read [14] The unused function employed for binary emotional division also get 88.94% accuracy via SVM [14].Kennedy et al adhered unigram along with bigram features binary emotion classification as well as accomplished 84.4% success rate in movie screening data [15] [27].Van [20] handles binary divisions with large Yugiram and magrams and achieves 86.1% success rate for interpreted data for Amazon products [20]. In [21] Akaichi used a mixture of large sacrifice as well as trigonometric features and achieved 72.78% success rate with SVM[21].

Valakunde et al aim to focus on five classes (for instance strong negative, neutral, positive negative, and strong positive) and achieve 81% success rateby SVM with large features. In a study by Gautamet al [18], they used SVM in conjunction with the semester model for emotional examination to categorize binary texts on Twitter and achieve 89.9% success rate through the unigram features. Tripathy et al. [19] employed SVM among n-g characteristics and achieved 88.94% success rate for classification of emotion binaries. Hasan et al. [25] approved the function of Vigram for binary emotional division and achieved 79% using MB for data translated from Twitter. All these studies using n-gram features normally achieve 70-90% accuracy as well as the majority successful representation is SVM. There are also handmade studies for mood classification [25].

Yassine and Hajj [29] used phonetic effects with misspellings and emoticons as a whole and achieved 87% success ratein terminal classification using SVM. Denecke [22] found three points (i.e., positive, negative, and objective) as characteristic and accomplished a 67% accuracy and a 66% binary rating reminder using logical classification. Jiang et al. [16] dealing with binary classification by SVM in addition to gain 67.8% success ratein articles on Twitter, they employed two types of non-purpose functions (e.g., content, Twitter content, as well as emotional glossary).Bahrainian et al [23] employed the positive as well as negative word counts, which were characteristic and attained 86.7% success rate for binary divisions through articles on Twitter.

Neethuet al [17] combine custom Twitter features along with Twitter features in addition to achieve 90% success rate for binary classification by SVM. Karamibekret al [30] mentioned the nomenclature as a function and pooled the characteristics with the unicode. They accomplished 65.46% success rate for triangular classification with SVM. Antai [24] employed regular word frequency and obtained 84% success rate for binary classification by SVM. Ghiassiet al [31] try to classify emotions into five categories, and they define functions that depend on the Twitter domain. They scored 92.7% of F1 points with SVM. Mensikovaet al [26] used the result of retrieving the characteristic unit name (NE) and acquired a 0.9 false positive rate (FPR). These learning were developed based on the definition of traits rather than employing only traits and were more effective (e.g., a score of 92.7% F1). Of course, it is not good to evaluate the outcomes among earlier learning for the reason that they employed dissimilar data sets.

However, as depicted in [17, 30], it is clear that the mixture of hand as well aslg functions should be better than using only n-g functions. Although successful in using machine learning models with manual and m-gram functions, these learning have a general restriction that their effectiveness varies depending on the specific characteristics that are defined; Different data will require a lot of effort from domain experts to get better outcomes. This limitation is also included in the emotion synthesis method for emotion examination [32], which combines other resources (e.g., science), as this will take considerable time and effort for professionals in the field. . An in-depth study model is a solution to such a constraint because it is known to automatically capture

(3)

__________________________________________________________________________________

4615

random patterns (e.g., traits). In addition, as described in [33], the use of in-depth training models for emotional examination will provide a meta-level characteristic that summarizes recent areas.

3. Proposed Approach

This section illustrates the procedure of data collection, processing, feature retrieving schemes for presenting data sets, prediction schemes, and combining knowledge gathering methods employed in experimental examination. The sketch out of the work is depicted in Figure 1. It consists of the following four modules

1. Gathering of Dataset 2. Data Cleaning 3. Feature Retrieval

4. Sentiment Examination Model Construction

Figure 1. Sketch out of the Proposed Work 3.1. Gathering of Dataset

To evaluate the predictable presentation of the psychological along with linguistic characteristics of emotional examination, this work examined the sequence of various sources such as twitter, facebook, website. It contain positive, negative along with neutral emotions. For database compiling process, this work used the structure summarized in [11]. This work employed Twitter4J. It is used to gather tweets. All tweet is captioned as either positive, negative, or neutral. Later than gathering tweets, automatic filtering is used to eliminate unsuitable and needless tweets. As a result this work get a collection of 6188 negatives, 4891 positives as well as 4,252 tweets. To get a unbiased body, dataset of our work contains 4,200 positive along with negative tweets as well as 4,200 positive tweets.

(4)

__________________________________________________________________________________

4616

3.2. Data Cleaning

Due to the inaccurate and informal nature of pure Twitter posts, it is necessary to follow below steps:

 Exclude mentions and replies from other users' tweets placed by a string starting with "@".

 Eliminate link started with "http: //"

 Clear "#" 3.3. Feature Retrieval

Once the database is run, the next step is to create a matrix of functions. Prior to retrieval, the token function is first applied to the pre-processed data. Tokenization is the process of dividing sentences into words. Then take a break. The word stop is commonly used and it is common that they lose their main meaning. Words like "on, are, that's, are" are just a few examples of words that stop.

This end is followed by the disappearance process continues. The purpose of the original was to reduce the reflection form and sometimes to associate the derivative of the word with a common basic form. This program is an easy way to truncate some characters at the end of a word to their roots, using them just set the principle for truncating some characters at the end of a word and hopefully they will get good results. Most. And then the characteristics of the matrix are established.

In this section, this paper examines the different psychological sets of emotional exploration. In this work, the LIWC is used to extract physical properties from a data set. Table 1 lists the main LIWC types and categories.

Table 1. Main Liwc Sets And Categories

As can be seen from the categories listed in Table 1, which shows the table above, language processes include grammatical information such as word counts, letters, personal sounds, texts, texts, and auxiliary verbs.

3.4 Sentiment Examination Model Construction

In the next step, develop highly efficient model for sentence research with Intellectual Broad Multi Domain Data Transfer Network (IBMDDTN).At this stage, the quoted feature matrix is analyzed as positive, neutral or negative to calculate the entire polar line. Multi Domain Sentiment Examination (MDSA) is used for this purpose. There are three main stages: emotional research, reaction research, and final emotional testing.

3.4.1. Sentiment Examination

Sentiment examination of tweet is performed using Light Semi Supervised Deep Convolution Neural Network (MDLSSDCNN). he test results categorize tweets as positive, neutral or negative and also give them a measure of emotional confidence. Emotional confidence just shows positive, neutral or negative. Emotional evaluation is further used to calculate the final emotions of a tweet.

3.4.2. Reverse sentiment examination

MDSA not only considers how positive / neutral / negative the original tweet is, but also how negative / neutral / positive the reverse tweet is. This solves the problem of polar deformation. Therefore, the first tweet must be inverted for any part of the tweet speech to be identified. Adjectives are extracted from lexical databases such as WordNet. Reverse the mood there are the following steps.

(5)

__________________________________________________________________________________

4617

 Each tweet is checked for negative words such as "No".

 Negative words are removed and the words immediately after them are not changed.

 Other words do not change.

The steps outlined above provide reverse tweaking. The feeling of changing tweets was also obtained with the help of the same LSSDCNN emotional analyst who provided his feelings and confidence.

3.4.3 Final sentiment examination

Taking into account both the original and inverse of Twitter, its final sentiment is calculated. The positive of the original tweet is considered to be the negative of the inverse tweet to calculate the final positive of the tweet. Similarly, the negative of the original tweet is considered with the positive of the inverse tweet to determine the negative of the tweet. Assume (+|x) as the probability that Tweet x is positive and (−|x̄) the probability that the reverse side of the tweet x is negative. Similarly, assume (−|x) ashe probability that the twitter x is negative and (+|x̄ ) the probability that the reverse side of the twitter x is positive. The relationship between these probabilities is then given by:

(+|x, x̄ ) = (1−∝). (+|x) + . (−|x̄ ) (−|x, x̄) = (1−∝). (−|x) + . (+|x̄ )

3.4.3.1. Light Semi Supervised Deep Convolution Neural Network (LSSDCNN)

The structure of Convolutional Neural Networks are similar to the structure of conventional neural networks. Each neural network is composed of neurons with training weights, biased values that are of studying weight along with bias. Generally a neuron obtains input, computes its point product moreover the behavior follows a non-linear characteristic. Each Convolutional Neuron Network is composed of single or additional convolutional layers as well as pooling layers or sub sampling layers. and this is depicted in Figure 2.

Figure 2. Sentiment Examining using CNN

The two types of pooling layers of CNN are Maximum and Average pooling in CNN and when united the biggest pixel values are calculated in max whereas in average pooling, the mean values are in use to explanation. The response of this layer is served as input to the subsequent layer of convolutional neural network. Generally convolutional neuron network is employed to sort the types of cancer cells and its severity. Down sampling is accomplished by the employment of pooling layer of CNN. The time consumption is lessened by diminishing the features extraction in convolution layers. To overcome this drawback of CNN, Light Semi Supervised Convolution Neural Network (LSSCNN) is proposed. This reduces high computation cost and improves speed. The dimension reduction of image space is realized by vector of features that is created by PSO from multidimensional image space to low dimensional feature space. Fig 10 shows the architecture of LSSDCNN.

(6)

__________________________________________________________________________________

4618

Figure 3.Sentiment Examining using LSSDCNN

3.4.3.2. Convolution Layer of LSSDCNN

In this layer, the features of the previous layer are attached to the core, which can be learned and put through the activation function to create a map of the characteristics of the result. Each result features can combine a roll with multiple cards. LSSDCNN usually has this problem.

Where Mj is the option of the input feature and the layer is of the type of boundary valid when implemented in MATLAB.

3.4.3.2.1 Sub-sampling Layers of LSSDCNN

This layer creates a pattern that is not included in the pattern. If there is an N input feature then there will be an actual N output card even though the output card will be smaller. More official

Where down (•) is the model function. This function is usually abbreviated on each n-by-n block in the input image, so the resulting image is smaller than n in both dimensions. Each result map has its own bias, multiplication, and spacing.

4. Result and Examination 4.1 Examination Parameters

To examine the effectiveness of the sentiment identification methods, a number of examination parameters are available. This work considers the Detection Accuracy, Precision Rate, Recall Rate, Sensitivity, Specificity, F-Measure and Error Rate to examine the effectiveness.

4.1.1 Detection Accuracy

Detection Accuracy metric finds the percentage of truthiness between the original sentiments and the predicted sentiments.

Accuracy = TP +TN

TP +FP +TN +FN (4.1)

4.1.2 Error Rate

Error Rate finds the percentage of falseness between the original sentiments and the predicted sentiments. Error Rate = No of Images of Falsely predicted sentiments

Total No of statments (4.2)

4.1.3 Precision Rate

(7)

__________________________________________________________________________________

4619

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝑇𝑃

𝑇𝑃+𝐹𝑃 (4.3)

4.1.4 Recall Rate

The recall is found by using the below formula

Recall = TP

TP +FN (4.4)

4.1.5 Sensitivity

Sensitivity is found by using the below formula

Sensitivity = TP

(TP +FN ) (4.5)

4.1.6 Specificity

Specificity is found by using the below formula

Specificity = TN

(FP +TN ) (4.6)

4.1.7 F-Measure

F-measure is found by using the below formula Fm= 1 + α ∗

Precision ∗Recall

α∗(Precision ∗Recall ) (4.7)

4.2 Experimental Examination

In this section the performance of the proposed method is evaluated in various experiments. To evaluated the efficiency of this sentiment examination scheme, the Detection Accuracy, Precision Rate, Recall Rate, Sensitivity, Specificity, F-Measure and Error Rate measures are used.

4.2.1 Trial No 1: Examination of Sentiment Examination Approaches using Accuracy

To examine the performance of this sentiment examination scheme, it is compared with different techniques using the operating indicators which are mentioned in Section 4.1. The output of these indicators are tabulated in Table 2.

Table 2. Examination of Sentiment Examination Techniques using Accuracy

Sentiment Examination Approaches Examination Parameters KNN NB SVM ELM CNN LSSDCNN LF 77.34 76.72 72.29 72.79 91.11 93.21 PF 83.75 81.85 79.95 80.98 92.24 92.92 PCF 84.66 82.76 80.86 81.89 93.15 93.83

(8)

__________________________________________________________________________________

4620

SCF 82.74 80.84 78.94 79.97 91.23 91.91

All 85.11 83.13 81.38 83.03 94.23 97.23

The influence of Sentiment Examination Techniques that are employed in this test using accuracy are successfully assessed. Table 1 illustrates that the highest accuracy value is 97.23 is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques. The output of these indicators are tabulated in Fig.3.

The influence of Sentiment Examination Techniques that are employed in this test using accuracy are successfully assessed. Fig.3 illustrates that the highest accuracy value is 97.23is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques.

4.2.2 Trial No 2: Examination of Sentiment Examination Approaches using Precision Rate

Table 3. Examination of Sentiment Examination Techniques using Precision Rate

Sentiment Examination Approaches Examination Parameters KNN NB SVM ELM CNN LSSDCNN LF 77.65 77.03 72.6 73.1 91.42 93.52 PF 84.06 82.16 80.26 81.29 92.55 93.23 PCF 84.97 83.07 81.17 82.2 93.46 94.14 0 20 40 60 80 100 120 LF PF PCF SCF All

Accuracy Analysis

KNN NB SVM ELM CNN LSSDCNN

(9)

__________________________________________________________________________________

4621

SCF 83.05 81.15 79.25 80.28 91.54 92.22

All 85.42 83.44 81.69 83.34 94.54 97.54

The influence of Sentiment Examination Techniques that are employed in this test using precision rate are successfully assessed. Table 4 illustrates that the highest precision value is 97.54 is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques. The output of these indicators are tabulated in Fig.4.

The influence of Sentiment Examination Techniques that are employed in this test using precision rate are successfully assessed. Fig.9 illustrates that the highest precision value is 97.54 is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques. The output of these indicators are tabulated in Fig.4.

4.2.3 Trial No 3: Examination of Sentiment Examination Approaches using Recall Rate

Table 4. Examination of Sentiment Examination Techniques using Recall Rate

Sentiment Examination Approaches Examination Parameters KNN NB SVM ELM CNN LSSDCNN LF 77.56 76.94 72.51 73.01 91.33 93.43 0 20 40 60 80 100 120 LF PF PCF SCF All

Precision Rate Analysis

(10)

__________________________________________________________________________________

4622 PF 83.97 82.07 80.17 81.2 92.46 93.14 PCF 84.88 82.98 81.08 82.11 93.37 94.05 SCF 82.96 81.06 79.16 80.19 91.45 92.13 All 85.33 83.35 81.6 83.25 94.45 97.45

The influence of Sentiment Examination Techniques that are employed in this test using recall rate are successfully assessed. Table 3 illustrates that the highest recall value is 97.45 is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques. The output of these indicators are tabulated in Fig.5.

The influence of Sentiment Examination Techniques that are employed in this test using recall rate are successfully assessed. Fig.10 illustrates that the highest recall value is 97.45 is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques.

4.2.4 Trial No 4: Examination of Sentiment Examination Approaches using Sensitivity

Table 5. Examination of Sentiment Examination Techniques using Sensitivity

Recall Rate Anlysis

(11)

__________________________________________________________________________________

4623 PF 84.16 82.26 80.36 81.39 92.65 93.33 PCF 85.07 83.17 81.27 82.3 93.56 94.24 SCF 83.15 81.25 79.35 80.38 91.64 92.32 All 85.52 83.54 81.79 83.44 94.64 97.64

The influence of Sentiment Examination Techniques that are employed in this test using sensitivity are successfully assessed. Table 4 illustrates that the highest sensitivity value is 97.64 is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques. The output of these indicators are tabulated in Fig.7.

The influence of Sentiment Examination Techniques that are employed in this test using sensitivity are successfully assessed. Fig.7 illustrates that the highest sensitivity value is 97.64 is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques.

4.2.5 Trial No 5: Examination of Sentiment Examination Approaches using Specificity

Table 6. Examination of Sentiment Examination Techniques using Specificity

Sensitivity Analysis

(12)

__________________________________________________________________________________

4624 PF 84.08 82.18 80.28 81.31 92.57 93.25 PCF 84.99 83.09 81.19 82.22 93.48 94.16 SCF 83.07 81.17 79.27 80.3 91.56 92.24 All 85.44 83.46 81.71 83.36 94.56 97.56

The influence of Sentiment Examination Techniques that are employed in this test using specificity are successfully assessed. Table 6 illustrates that the highest specificity value is 97.56 is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques. The output of these indicators are tabulated in Fig.8.

The influence of Sentiment Examination Techniques that are employed in this test using specificity are successfully assessed. Table 6 illustrates that the highest specificity value is 97.56 is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques.

4.2.6 Trial No 6: Examination of Sentiment Examination Approaches using F-Measure

Table 7. Examination of Sentiment Examination Techniques using F-Measure

Specificity Analysis

(13)

__________________________________________________________________________________

4625 PF 83.9 82 80.1 81.13 92.39 93.07 PCF 84.81 82.91 81.01 82.04 93.3 93.98 SCF 82.89 80.99 79.09 80.12 91.38 92.06 All 85.26 83.28 81.53 83.18 94.38 97.38

The influence of Sentiment Examination Techniques that are employed in this test using f-measure are successfully assessed. Table 7 illustrates that the highest f-measure value is 97.38 is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques. The output of these indicators are tabulated in Fig.8.

The influence of Sentiment Examination Techniques that are employed in this test using f-measure are successfully assessed. Table 6 illustrates that the highest f-measure value is 97.38 is found for LSSDCNN and it is more powerful as it is the highest value when compared with other techniques.

5. Conclusion

In this article, a way of classifying emotions is introduced using the emotional along with linguistic characteristics that LIWC acquires at what time examining emotions on multilinguals such as twitter, facebook, websites. To this end, there are five types of LIWCs and these mixtures are considered. Experimental examination with ranking solutions depicts that psychiatric sets can offer encouraging outcomes when examining the emotions of multimedia data on Experimental examination depicts that this band has more than one individual. For the examination of emotions in the Internet of Things, the highest predictive efficiency (98.31%) was accomplished by mixing the Linguistic along with Psychological Processes of LSSDCNN. Thus, this proposed approach and the mixture of language procedures, mental procedures are best represented in the examination of emotions.

References

1. Balahur A., Turchi M. (2014) “Comparative experiments using supervised learning and machine translation for multilingual sentiment examination”, Computer Speech Language vol. 28, no. 1, pp. 56–75.

2. Cheng Z., et al. (2010) “You are where you tweet: a contentbased approach to geo-location twitter users”, ACM Inter. Conf. Info. Know. Mana., pp.759-768.

0 20 40 60 80 100 120 LF PF PCF SCF All

F-Measure Analysis

(14)

__________________________________________________________________________________

4626

3. Clematide S., Manfred K. (2010). „Evaluation and extension of a polarity lexicon for German‟, Computer Approaches Subject Sentimental Examination, pp. 7–13.

4. Cotelo J.M., et al, (2015). “A modular approach for lexical normalization applied to Spanish tweets”, Exp. Sys. Appl., Vol. 42, No.10, pp. 4743-4754.

5. Cruz F. L., et al, (2014). „ML-SentiCon: Un lexicón multilingüe de polaridades semánticas a nivel de lemas‟, Proce. Lang. Nat., vol. 53, pp. 113–120.

6. Dehdarbehbahani I., Shakery A., Faili H.,(2014). „Semi-supervised word polarity identification in resource

-lean

languages‟, Neural Networking, vol. 58, pp. 50–59.

7. Esuli A., Sebastiani F., (2006).„SENTIWORDNET: A Publicly Available Lexical Resource for Opinion Mining‟, Lang. Reso. Eval., pp. 417–422..

8. Filho P. P. B., et al, (2013). An evaluation of the Brazilian Portuguese liwc dictionary for sentiment examination‟, Info. Hum. Lang. Tech.,.

9. Pak, A.; Paroubek, P. (2010)Twitter as a corpus for sentiment examination and opinion mining. LREc, 10, 1320–1326.

10. Alm, C.O.; Roth, D.; Sproat, R. (2005) Emotions from text: Machine learning for text-based emotion prediction. In Proceedings of the Conference on Human Language Technology and Empirical Methods in Natural Language Processing, Vancouver, BC, Canada, 6–8 October; pp. 579–586.

11. Bartlett, M.S.; Littlewort, G.; Frank, M.; Lainscsek, C.; Fasel, I.; Movellan, J. (2005) Recognizing facial expression: Machine learning and application to spontaneous behavior. In Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR‟05), San Diego, CA, USA, 20–25 June; Volume 2, pp. 568–573.

12. Heraz, A.; Razaki, R.; Frasson, C. (2007) Using machine learning to predict learner emotional state from brainwaves. In Proceedings of the Seventh IEEE International Conference on Advanced Learning Technologies (ICALT 2007), Niigata, Japan, 18–20 July; pp. 853–857.

13. Yujiao, L.; Fleyeh, H. (2018) Twitter Sentiment Examination of New IKEA Stores Using Machine Learning. In Proceedings of the International Conference on Computer and Applications, Beirut, Lebanon, 25–26 July.

14. Read, J. (2005) Using emoticons to reduce dependency in machine learning techniques for sentiment classification. In Proceedings of the ACL Student Research Workshop, Ann Arbor, Michigan, 27–27 June; pp. 43–48.

15. Kennedy, A.; Inkpen, D. (2006) Sentiment classification of movie reviews using contextual valence shifters. Comput. Intell., 22, 110–125.

16. Jiang, L.; Yu, M.; Zhou, M.; Liu, X.; Zhao, T. (2011) Target-dependent twitter sentiment classification. In Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, Portland, Oregon, 19–24 June; Volume 1, pp. 151–160.

17. Neethu, M.; Rajasree, R. (2013) Sentiment examination in twitter using machine learning techniques. In Proceedings of the 2013 Fourth International Conference on Computing, Communications and Networking Technologies (ICCCNT), Tiruchengode, India, 4–6 July; pp. 1–5.

18. Gautam, G.; Yadav, D. (2014) Sentiment examination of twitter data using machine learning approaches and semantic examination. In Proceedings of the 2014 Seventh International Conference on Contemporary Computing (IC3), Noida, India, 7–9 August; pp. 437–442.

19. Tripathy, A.; Agrawal, A.; Rath, S.K. (2016). Classification of sentiment reviews using n-gram machine learning approach. Expert Syst. Appl., 57, 117–126.

20. Wan, X. (2012). A comparative study of cross-lingual sentiment classification. In Proceedings of the 2012 IEEE/WIC/ACM International Joint Conferences on Web Intelligence and Intelligent Agent Technology, Macau, China, 4–7 December; Volume 1, pp. 24–31.

21. Akaichi, J. (2013). Social networks‟ Facebook‟statutes updates mining for sentiment classification. In Proceedings of the 2013 International Conference on Social Computing, Alexandria, VA, USA, 8–14 September; pp. 886–891.

22. Denecke, K. (2008). Using sentiwordnet for multilingual sentiment examination. In Proceedings of the 2008 IEEE 24th International Conference on Data Engineering Workshop, Cancun, Mexico, 7–12 April; pp. 507– 512.

23. Bahrainian, S.A.; Dengel, A. (2013). Sentiment examination using sentiment features. In Proceedings of the 2013 IEEE/WIC/ACM International Joint Conferences on Web Intelligence (WI) and Intelligent Agent Technologies (IAT), Atlanta, GA, USA, 17–20 November; Volume 3; pp. 26–29.

24. Antai, R. (2014). Sentiment classification using summaries: A comparative investigation of lexical and statistical approaches. In Proceedings of the 2014 6th Computer Science and Electronic Engineering Conference (CEEC), Colchester, UK, 25–26 September; pp. 154–159.

(15)

__________________________________________________________________________________

4627

25. Hasan, A.; Moin, S.; Karim, A.; Shamshirband, S. (2018). Machine learning-based sentiment examination for twitter accounts. Math. Comput. Appl. 23, 11.

26. Mensikova, A.; Mattmann, C.A. (2018). Ensemble sentiment examination to identify human trafficking in web data. Workshop on Graph Techniques for Adversarial Activity Analytics (GTA 2018), Marina Del Rey, CA, USA, February

27. Pang, B.; Lee, L. (2014). A sentimental education: Sentiment examination using subjectivity summarization based on minimum cuts. In Proceedings of the 42nd Annual Meeting on Association for Computational Linguistics, Barcelona, Spain, 21–26 July; pp. 271–278.

28. Valakunde, N.; Patwardhan, M. (2013). Multi-aspect and multi-class based document sentiment examination of educational data catering accreditation process. In Proceedings of the International Conference on Cloud & Ubiquitous Computing & Emerging Technologies, Pune, India, 15–16 November 2013; pp. 188–192. 29. Yassine, M.; Hajj, H. (2010). A framework for emotion mining from text in online social networks. In

Proceedings of the 2010 IEEE International Conference on Data Mining Workshops, Sydney, Australia, 13 December; pp. 1136–1142.

30. Karamibekr, M.; Ghorbani, A.A. (2013). A structure for opinion in social domains. In Proceedings of the 2013 International Conference on Social Computing, Alexandria, VA, USA, 8–14 September; pp. 264–271. 31. Ghiassi, M.; Lee, S. A domain transferable lexicon set for Twitter sentiment examination using a supervised

machine learning approach. Expert Syst. Appl. 2018, 106, 197–216.

32. Balazs, J.A.; Velásquez, J.D. (2016). Opinion mining and information fusion: A survey. Inf. Fusion, 27, 95– 110.

33. Chaturvedi, I.; Cambria, E.; Welsch, R.E.; Herrera, F. Distinguishing between facts and opinions for sentiment examination: Survey and challenges. Inf. Fusion (2018)., 44, 65–77.