Predicting Eurovision Song Contest results using sentiment analysis on tweets

(1)

Predicting Eurovision Song Contest results using

sentiment analysis on tweets

Nina Burger 11060239 Bachelor thesis Credits: 18 EC

Bachelor Information Science University of Amsterdam Faculty of Science Science Park 904 1098 XH Amsterdam Supervisor Dr. Ashley Burgoyne Second examiner Dr. Frank Nack 2020-21-06 

(2)

Abstract

This thesis uses sentiment analysis on tweets to predict Eurovision Song Con-test (ESC) results. Sentiment analysis on tweets is a widely used method to analyse the public opinion, not only about voting contests but also about elec-tions or other phenomena. For the years 2015-2019, tweets about the ESC were collected using Twint. After pre-processing, sentiment analysis was done using TextBlob. Leave-one-group-out cross-validated LASSO regression was performed to build three models, predicting the final total points, final televoting points, and final jury points. For all models, the two most impor-tant features were polarity, and the interaction between polarity and subjec-tivity. A cross-validated R2_{of 0.28 was found for the most successful model.}

Based on this R2 _{score, it was concluded that the model could predict ESC}

performance to a large extent. This conﬁrms the idea that Twitter is a good reﬂection of the public opinion and suggests that other voting contest results and perhaps election results can be predicted using sentiment analysis on tweets. 

(3)

Acknowledgements

I would like to express my gratitude to my supervisor, Dr. Ashley Burgoyne, for his guidance and encouragement during the writing of this thesis. I would also like to thank my family and friends for their unconditional support. 

(4)

1 Introduction

In the last few years, data analysis has become a very important part of research in all kinds of fields. New techniques are constantly being developed and one of them is senti-ment analysis. Sentisenti-ment analysis is the classification of emotions in textual data and it can offer a new perspective on public opinion. It is particularly useful for things such as analysing a customer’s experience with a company (Feldman, 2013), or analysing the public view during an election (Wang, Can, Kazemzadeh, Bar, & Narayanan, 2012). One of the biggest sources of opinions on the internet is Twitter and it proves to be use-ful for making predictions. In 2012, Bae & Lee used sentiment analysis on tweets to mea-sure the influence of popular Twitter users on their followers and found that their results could be used to predict real-world phenomena. In 2017, Joyce & Deng found that their sentiment analysis on tweets about Donald Trump during the 2016 US presidential elec-tions correlated highly with their polling data.

The Eurovision Song Contest (ESC) is the largest live music event in the world, in which mostly European countries perform an original song and vote for the best song at the end of the night (European Broadcasting Union, 2020). A big part of the viewer’s experi-ence is not only voting for their favourite, but also betting on who they think is going to win the contest (Royston, 2019). Every year the bookmakers compile their odds for the viewers to place bets on. By doing this they essentially try to predict public opinion. In the last few years, viewers have used ESC related hashtags on social media such as Twit-ter to react to the contest. This provides a clear representation of public opinion on the ESC and its participants. The objective of this thesis is to use sentiment analysis on tweets to predict ESC results.

In order to predict the results of the ESC, a model was built that collects tweets about the ESC from the years 2015-2019 and predicts contest results using sentiment analysis on those tweets. Sentiment analysis on tweets that are speciﬁcally about the ESC is a good way of measuring the public opinion around the ESC entries, and predicting the results in this way could be useful for bookmakers in determining their betting odds. The

(6)

research question for this bachelor thesis is: To what extent can a sentiment analysis of tweets about the ESC predict contest results?

In his bachelor thesis, Xie (2020) found a moderate linear correlation between sentiment in YouTube comments and contestant ranking in the ESC. It is expected that sentiment analysis of tweets about the ESC can predict contest results to a similar extent. Since the ﬁnal results are a combination of viewer votes (televoting) and votes by a profession-al jury, it is expected that the model will be better at predicting the televoting results than the jury results as well as the ﬁnal results.

In the next section, background information is provided about the ESC and the diﬀerent techniques used in this thesis. Section 3 describes the methodology of how the model was built and evaluated. In section 4, the results are described. Section 5 consists of a discus-sion of the results and a concludiscus-sion, as well as limitations and suggestions for further studies.

(7)

2 Theoretical background

This section provides background information on the ESC. Moreover, it discusses several techniques and Python libraries that were used in this thesis, related to Twitter scraping, sentiment analysis, and linear regression.

2.1 The Eurovision Song Contest

The ESC is a yearly contest in which mostly European countries compete to win the prize for best original song. Each year there are two semi-finals in the days leading up to the final round. Around 26 countries get to perform an original song in the final round. At the end of the evening, all countries competing in the contest can vote for each other. Countries cannot vote for themselves. Each country’s final vote is made up of points giv-en by viewers (televoting points) and points givgiv-en by a 5-member professional jury (jury points). The song with the highest number of points wins the contest and the winning country gets to present the contest the next year. Over the years, a total of 52 countries have participated in the contest and with around 200 million viewers each year it is the largest live music event in the world (European Broadcasting Union, 2020).

2.2 Twitter scraping

The ﬁrst step in this thesis is to scrape Twitter. It was found that the Standard Twitter Application Programming Interface (API) can only fetch a limited number of tweets in a short amount of time and that only tweets created in the last week can be fetched. Other Python libraries that scrape Twitter, such as Tweepy, were considered, but this library and most other libraries use the Twitter API and thus the same problem persists.

Twint is a Python library that can be used to scrape Twitter without using the Twitter API. Twint is able to fetch almost all tweets. The usage of Twint is similar to Twitters’s 1

search function, as the same search operators can be used. In 2019, Bonsón, Perea, & Bednárová used Twint to collect data for an empirical study on the way the Andalusian municipalities use Twitter to engage with their citizens. They also used the Twint

https://github.com/twintproject/twint

(8)

tionality of collecting the number of likes, replies, and retweets on a tweet. Since Twint has been used successfully for data collection in existing research and because it does not use the Twitter API, which comes with restrictions, Twint is the best-suited library for this thesis.

2.3 Sentiment analysis

Sentiment analysis can be used to measure the emotion in a text. Sentiment is represent-ed by a polarity score and a subjectivity score. Polarity describes how positive or nega-tive a text is, while subjectivity describes whether a text is subjecnega-tive, like an opinion, or objective, like a fact. Most sentiment analysis tools use a classifications process. Firstly the subjectivity of a text is measured. Usually, objective texts are classified as neutral, while subjective texts go through a new classification process, in which polarity is mea-sured. This way, only relevant texts are considered (Lima, de Castro, & Corchadeo, 2015). In this thesis, there is no classification process. Instead, the polarity and subjec-tivity scores are used to create separate features and interaction features.

Natural Language ToolKit (NLTK) is a well-known Python library that can be used to process natural data. NLTK’s Vader sentiment analysis tool categorises text into either a ‘positive’, ‘neutral’, or ‘negative’ category. Another possible library for sentiment analysis is Pattern. Pattern’s sentiment analysis tool provides two scores; polarity and subjectivi-ty. Polarity runs from -1 to 1, a value of -1 meaning a text is negative and a value of 1 meaning a text is positive. Subjectivity runs from 0 to 1, a value of 0 meaning a text is objective and a value of 1 meaning a text is subjective.Although both libraries are good options, the Python library TextBlob is best suited for this thesis. TextBlob is built on 2

NLTK and Pattern, but it is more accessible as it is simple to use. TextBlob’s sentiment analysis tool works with the same scores as Pattern; polarity and subjectivity, which in this thesis are useful to create interaction features, using a tweet’s sentiment and its number of likes, retweets, and replies. In 2018, Oikonomou & Tjortjis successfully used TextBlob to predict the winner of the US presidential elections in 2016 and emphasised the importance of the use of the subjectivity score in sentiment analysis of tweets.

https://github.com/sloria/textblob

(9)

2.4 Regression and LASSO

In machine learning, regression models can be used to predict a dependent variable (hereafter referred to as the target) based on independent variables (hereafter referred to as features). Linear regression does this using a linear function of the features. LASSO is a form of linear regression that uses L1 regularisation, which performs feature selection by restricting the coeﬃcients of the least important features to be exactly zero (Müller & Guido, 2016, p. 55). The lambda (λ) parameter - referred to as alpha (α) by some authors - is used to control the complexity of the models. A higher value of alpha puts more re-striction on the coeﬃcients of the model, thus fewer features are used. Alpha is a hyper-parameter that has to be tuned to get the best performing model. This can be done in a variety of ways and in this thesis, it is done using cross-validation; by repeatedly split-ting the data into training and test sets to get a model that is more stable and better at generalising to new data (Müller & Guido, 2016, p. 254).

(10)

3 Methodology

This section describes how the data was collected and pre-processed. Next, it describes how sentiment analysis was done, which features were created, and how the regression models were built and evaluated.

3.1 Data collection

English tweets about the ESC in the years 2015-2019 were scraped from Twitter using Twint. In order to target tweets about the ESC in speciﬁc years, a dataset containing contestant names and country names was used. Because this dataset did not include the 3

final televoting points and the final jury points for the year 2015, this data was added manually. Besides the textual content of a tweet, tweet date, number of likes, number of retweets, and number of replies were also selected. Tweets that were posted between and including the days of the semi-final and final rounds (5 days each year), and that contain either a country or the name of a country’s contestant, were scraped. Retweets and tweets containing hyperlinks were ignored. Only tweets about the contestants that made it to the final round were collected. Tweet collection was limited to a maximum of 200 tweets per country per tweet date as not to overload the library. After the collection of tweets for each year, the tweets were saved in pickle files.

3.2 Data preparation

After loading the pickle ﬁles for each year, they were concatenated into one Pandas data frame. The capitalisation of all tweets was changed to lowercase. To make sure that all tweets were ESC related, tweets that did not contain the following terms were removed from the dataset: “#esc”, “eurovisi”, “melodifestivalen”, “songfestival”, “songcontest”, and “melodigrandprix”. These terms were chosen by translating “Eurovision Song Contest” into major European languages and seeing which terms collected the most tweets. A to-tal of 24 204 tweets remained. To avoid usernames aﬀecting the sentiment analysis, they were replaced by the following string: “@username”. Punctuation, emoji, and the newline

eurovisionworld.com, as curated by Janne Spijkervet

(11)

character “\n" were removed from the tweets, as well as the most common words in the English language (stopwords).

After pre-processing, the minimum number of tweets for a country in a year was 35 and the maximum number of tweets for a country in a year was 367. The minimum number of tweets on a tweet date was 261 and the maximum was 1483. In each year, the number of tweets increased as the date got closer to the ﬁnal round.

3.3 Sentiment analysis and feature engineering

In total, twelve features were created from the data. Sentiment analysis was performed on the tweets, giving each tweet a polarity and subjectivity score. The numbers of likes, retweets, and replies on a tweet were also used as features. Using the polarity, subjectivi-ty, and the number of likes, retweets, and replies of a tweet, interaction features were created.

For each tweet, the product of polarity and subjectivity (polarity × subjectivity) was calculated. This created a feature in which the polarity of a more subjective tweet was weighted more heavily than that of a more objective tweet.

In addition, for each tweet, the product of polarity and the number of likes (polarity × likes), the product of polarity and the number of retweets (polarity × retweets), and the product of polarity and the number of replies (polarity × replies) were calculated. The same was done with subjectivity and likes (subjectivity × likes), retweets (subjectivity × retweets), and replies (subjectivity × replies). This created features in which the polarity or subjectivity of a tweet was weighted more heavily if a tweet had more likes, retweets, or replies.

The mean of every feature was taken per country and put into a new data frame for each year. To each new data frame, the final total points per year were added as well as the final televoting points and final jury points. These final points were used as the targets for the regression models. The features and targets were standardised. Standardising the

(12)

targets is important because the diﬀerent number of countries competing each year caus-es the point distribution to vary.

3.4 Regression models and evaluation

Three diﬀerent regression models were built using the Python library Scikit-Learn. 4

Model A was used to predict the final total points, model B was used to predict the final televoting points and model C was used to predict the final jury points.

For model A, leave-one-group-out cross-validation was done on the data from the years 2015-2018. For every fold, one year was used as the test set while the others were used as the training set. Cross-validation was used to tune the alpha parameter in LASSO. An alpha value of 0.02 was chosen.

Model B and model C followed the same process as model A, but a slightly different dataset was used to train the models. Instead of final total points, the target for model B was final televoting points and the target for model C was final jury points. After cross-validation, an alpha of 0.02 was used for model B and an alpha of 0.022 was used for model C.

In this thesis, the success of a model is measured by R-squared (R2_{). R}2_{is the coeﬃcient}

of determination and in linear regression, it represents the variance in the target that is explained by a model. A high R2_{indicates a highly predictive model. The interpretation}

of R2_{depends on the context of the study. In a social study, the R}2_{value should be}

in-terpreted diﬀerently than in an engineering study, as there can be a lot of intervening variables in the social sciences. As a guideline, Cohen (1988) has suggested that an R2_of

0.02 can be interpreted as a small eﬀect, an R2_{of 0.13 can be interpreted as a medium}

eﬀect, and an R2_{of 0.26 can be interpreted as a large eﬀect.}

https://github.com/scikit-learn/scikit-learn

(13)

4 Results

This section describes which coeﬃcients were selected for each model. Next, the cross-validated R2 _{and the predictions for the year 2019 are discussed for each model.}

4.1 Model A (LASSO on ﬁnal total points)

The coeﬃcients of model A are visualised in Figure 1 and Table 1. Seven features were selected by the model and the two features with the highest weight were polarity and po-larity × subjectivity.

The cross-validated R2_{for model A was 0.28. The R}2_{for each fold is visualised in Figure}

2. The model was tested on data from the year 2019. The ﬁnal total points for that year could be predicted by model A with an R2_{of 0.35. The regression line is visualised in}

Figure 3.

Figure 1: coeﬃcients of model A

Figure 2: cross-validated R2_{per model}

(14)

As visualised in Table 2, the predicted top 10 using this model was Italy, France, Serbia, Sweden, Switzerland, the Netherlands, Russia, Australia, Greece, and Azerbaijan, with Italy in 1st place and Azerbaijan in 10th place. The actual top 10 included The Nether-lands, Italy, Russia, Switzerland, Sweden, Norway, North Macedonia, Azerbaijan, Aus-tralia, and Iceland with the Netherlands in 1st place and Iceland in 10th place. This means that 7 countries were correctly predicted to be in the top 10. The full prediction tables for all models can be found in the appendix.

Predicted

rank Actual rank Predicted ﬁnal points Actual ﬁnal points Polarity Subjectivity

Italy 1 2 336 472 0.246 0.445 France 2 16 257 105 0.203 0.417 Serbia 3 18 254 89 0.179 0.452 Sweden 4 5 250 334 0.208 0.513 Switzerland 5 4 250 364 0.181 0.512 Netherlands 6 1 243 498 0.195 0.445 Russia 7 3 233 370 0.197 0.484 Australia 8 9 228 284 0.208 0.451 Greece 9 21 213 74 0.236 0.469 Azarbaijan 10 8 187 302 0.188 0.452

Figure 3: regression line of model A (2019)

(15)

After scaling the target back to its original form, the model gave a negative prediction for the lowest scoring country. In the full prediction table in the appendix and the re-gression plot in Figure 3, any negative scores are replaced by 0.

4.2 Model B (LASSO on ﬁnal televoting points)

The coeﬃcients of model B are visualised in Figure 4 and Table 3. Seven features were selected by the model and polarity and polarity × subjectivity had the highest weights.

The cross-validated R2_{for model B was 0.27. The R}2_{for each fold is visualised in Figure}

2. The model was tested on data from the year 2019. The ﬁnal televoting results for that year could be predicted with an R2_{of 0.29. As visualised in Table 4, the predicted}

top 10 using this model was Italy, Switzerland, France, Sweden, the Netherlands, Serbia, Russia, Australia, and Greece with Italy in 1st place and Greece in 10th place. The ac-tual top 10 included The Netherlands, Italy, Russia, Switzerland, Sweden, Norway, North Macedonia, Azerbaijan, Australia, and Iceland with the Netherlands in 1st place and Ice-land in 10th place. This means that 7 countries were correctly predicted to be in the top 10.

Figure 4: coeﬃcients of model B

(16)

4.3 Model C (LASSO on ﬁnal jury points)

The coeﬃcients of model C are visualised in Figure 5 and Table 5. Seven features were selected by the model and polarity and polarityXsubjectivity were the features with the highest weights.

The cross-validated R2_{for model C was 0.24. The R}2_{for each fold is visualised in Figure}

5. The model was tested on data from the year 2019. The ﬁnal jury results for that year

Table 4: predictions for 2019, model B (top 10)

Figure 5: coeﬃcients of model C

Table 5: coeﬃcients of model C Predicted

rank Actual rank Predicted televoting points Actual televoting points Polarity Subjectivity Italy 1 2 182 253 0.246 0.445 Switzerland 2 4 133 212 0.181 0.512 France 3 16 131 38 0.203 0.417 Sweden 4 5 129 93 0.208 0.513 Netherlands 5 1 128 261 0.195 0.445 Serbia 6 18 124 54 0.179 0.452 Russia 7 3 119 244 0.197 0.484 Australia 8 9 117 131 0.208 0.451 Greece 9 21 106 24 0.236 0.469 Azarbaijan 10 8 97 100 0.188 0.452

(17)

could be predicted with an R2_{of 0.28. As visualised in Table 6, the predicted top 10}

us-ing this model was Italy, France, Switzerland, Sweden, Serbia, the Netherlands, Aus-tralia, Russia, Greece, and Azerbaijan with Italy in 1st place and Azerbaijan in 10th place. The actual top 10 included The Netherlands, Italy, Russia, Switzerland, Sweden, Norway, North Macedonia, Azerbaijan, Australia, and Iceland with the Netherlands in 1st place and Iceland in 10th place. This means that 7 countries were correctly predicted to be in the top 10. 

Table 6: predictions for 2019, model C (top 10) Predicted

rank Actual rank Predicted jury points Actual jury points Polarity Subjectivity

Italy 1 2 171 237 0.246 0.445 France 2 16 130 40 0.203 0.417 Switzerland 3 4 124 202 0.181 0.512 Sweden 4 5 121 126 0.208 0.513 Serbia 5 18 119 67 0.179 0.452 Netherlands 6 1 118 241 0.195 0.445 Australia 7 9 115 152 0.208 0.451 Russia 8 3 111 219 0.197 0.484 Greece 9 21 107 24 0.236 0.469 Azarbaijan 10 8 98 150 0.188 0.452

(18)

5 Discussion

Based on Xie’s (2020) predictions using sentiment analysis on YouTube comments, it was expected that ESC performance could be predicted using sentiment analysis on tweets to a similar extent. Xie was able to predict the ﬁnal results with an R2_{of 0.202, which can}

be interpreted as moderately successful.

In this thesis, the cross-validated R2_{for the best performing model was 0.28 and the}

ﬁ-nal total results of the 2019 ESC could be predicted with an R2_{of 0.35. Following}

Co-hen’s (1988) guidelines, it can be concluded that this model can predict ESC perfor-mance to a large extent.

While it was also expected that model B (final televoting points), would be able to pre-dict the ESC results better than model C (final jury points), and model A (final total points) this was not the case. Each model’s coefficients are similar and all models can predict the results to a large extent. Model A was a little bit better at predicting the re-sults for the year 2019 than the other models, which is the opposite of what was expect-ed. However, no big differences can be seen in the prediction tables.

It was found that in all models, polarity and polarity × subjectivity were the most im-portant features. This means that countries that receive more positive tweets do better in the competition and countries that receive more negative tweets do worse. This effect gets reduced the more subjective a tweet is. This is unusual for sentiment analysis, be-cause often the first step of sentiment analysis is classifying objective texts as neutral, thus ignoring objective texts (Lima et al., 2015), while in the case of this thesis, it seems that objective texts have a bigger effect on the outcome than subjective texts. It is inter-esting that polarity × subjectivity instead of subjectivity corrects for polarity. A reason for this could be that, while both polarity and subjectivity are important for sentiment analysis, polarity says the most about a text’s sentiment.

While enough tweets and features were collected for each country, the amount of obser-vations per year is around 26, because that is the number of countries that participate in the ﬁnal round each year. It is important to remember that this limits how precise

(19)

pre-dictions can be. More years could be included in the dataset, but this might lead to oth-er problems as musical taste and Twittoth-er’s role in society can change ovoth-er the years. As mentioned in section 3.1, tweet collection was limited to a maximum of 200 tweets per country per tweet date. That is suﬃcient for the purpose of this study, but in further research, it could be interesting to see if there is a connection between the number of tweets per country and contestant ranking.

For this thesis, only English tweets were taken into account. TextBlob can detect a text’s language and can do sentiment analysis based on the English translation of a tweet, but since translation can alter sentiment, this was not included in this thesis. However, after recent developments, the use of Machine Translation Systems has become more reliable for sentiment analysis (Balahur & Turchi, 2012). Therefore, as long as a good Machine Translator System is used, translations of non-English tweets could also be included in the dataset in the future, or separate sentiment analysers could be used for tweets in dif-ferent languages.

For further research, it would be interesting to look at what country a tweet is posted

from in addition to what country a tweet is about in order to predict the number of

points given from a country to another. For this to work, it would be necessary to collect even more tweets and the location from which a tweet is posted must also be collected. The language of a tweet could possibly be used to determine which country a user is from as well.

These results suggest that ESC results can be predicted to a large extent using sentiment analysis on tweets and this is a conﬁrmation of the idea that Twitter provides a good ﬂection of the public opinion. It is also a motivation for researchers to try to predict re-sults for other contests or elections using sentiment analysis on tweets. 

(20)

References

Bae, Y., & Lee, H. (2012). Sentiment analysis of twitter audiences: Measuring the posi-tive or negaposi-tive inﬂuence of popular twitterers. Journal of the American Society for

Information Science and Technology, 63(12), 2521–2535. https://doi.org/10.1002/asi.

22768

Balahur, A., & Turchi, M. (2012, July). Multilingual sentiment analysis using machine translation?. In Proceedings of the 3rd workshop in computational approaches to

sub-jectivity and sentiment analysis (pp. 52-60).

Bonsón, E., Perea, D., & Bednárová, M. (2019). Twitter as a tool for citizen engagement: An empirical study of the Andalusian municipalities. Government Information

Quar-terly, 36(3), 480–489. https://doi.org/10.1016/j.giq.2019.03.001

Cohen, J. (1977). F Tests of Variance Proportions in Multiple Regression/Correlation Analysis. Statistical Power Analysis for the Behavioral Sciences, 407–453. https:// doi.org/10.1016/b978-0-12-179060-8.50014-1

European Broadcasting Union. (2020, May 28). What is Eurovision: How to explain it to the world. Retrieved from https://eurovision.tv/story/what-is-eurovision-how-to-ex-plain-it-to-the-world

Feldman, R. (2013). Techniques and applications for sentiment analysis. Communications

of the ACM, 56(4), 82–89. https://doi.org/10.1145/2436256.2436274

Joyce, B., & Deng, J. (2017). Sentiment analysis of tweets for the 2016 US presidential election. 2017 IEEE MIT Undergraduate Research Technology Conference (URTC). https://doi.org/10.1109/urtc.2017.8284176

Lima, A. C. E. S., de Castro, L. N., & Corchado, J. M. (2015). A polarity analysis framework for Twitter messages. Applied Mathematics and Computation, 270, 756– 767. https://doi.org/10.1016/j.amc.2015.08.059

(21)

Müller, A. C., & Guido, S. (2016). Introduction to Machine Learning with Python. Culemborg, Netherlands: Van Duuren Media.

Oikonomou, L., & Tjortjis, C. (2018, September). A Method for Predicting the Winner of the USA Presidential Elections using Data extracted from Twitter. In 2018

South-Eastern European Design Automation, Computer Engineering, Computer Networks and Society Media Conference (SEEDA_CECNSM) (pp. 1-8). IEEE.

Royston, B. (2019, May 18). In it to win it, but what are the odds? Retrieved from https://eurovision.tv/story/in-it-to-win-it-but-what-are-the-odds

Wang, H., Can, D., Kazemzadeh, A., Bar, F., & Narayanan, S. (2012, July). A system for real-time twitter sentiment analysis of 2012 us presidential election cycle. In

Pro-ceedings of the ACL 2012 system demonstrations (pp. 115-120).

Xie, K. (2020). Sentiment analysis, a viable approach towards predicting the Eurovision? (Bachelor thesis)

(22)

Appendix

Prediction table of the ﬁnal total points in 2019 (model A) Predicted

rank Actual rank Predicted ﬁnal points Actual ﬁnal points Polarity Subjectivity

Italy 1 2 336 472 0.246 0.445 France 2 16 257 105 0.203 0.417 Serbia 3 18 254 89 0.179 0.452 Sweden 4 5 250 334 0.208 0.513 Switzerland 5 4 250 364 0.181 0.512 Netherlands 6 1 243 498 0.195 0.445 Russia 7 3 233 370 0.197 0.484 Australia 8 9 228 284 0.208 0.451 Greece 9 21 213 74 0.236 0.469 Azarbaijan 10 8 187 302 0.188 0.452 Spain 11 22 185 54 0.145 0.403 Cyprus 12 13 179 109 0.168 0.442 Iceland 13 10 179 232 0.161 0.449 United Kingdom 14 26 175 11 0.160 0.414 Norway 15 6 172 331 0.182 0.459 Malta 16 14 162 107 0.173 0.494 Israel 17 23 153 35 0.124 0.374 Denmark 18 12 152 120 0.236 0.545 Czech Re-public 19 11 151 157 0.200 0.433 North Macedonia 20 7 137 305 0.154 0.464 Slovenia 21 15 131 105 0.109 0.482 Germany 22 25 119 24 0.102 0.418 San Marino 23 19 114 77 0.091 0.454 Estonia 24 20 106 76 0.157 0.437 Albania 25 17 96 90 0.163 0.462 Belarus 26 24 81 31 0.105 0.440

(23)

Prediction table of the ﬁnal televoting points in 2019 (model B) 

Predicted

rank Actual rank Predicted televoting points Actual televoting points Polarity Subjectivity Italy 1 2 182 253 0.246 0.445 Switzerland 2 4 133 212 0.181 0.512 France 3 16 131 38 0.203 0.417 Sweden 4 5 129 93 0.208 0.513 Netherlands 5 1 128 261 0.195 0.445 Serbia 6 18 124 54 0.179 0.452 Russia 7 3 119 244 0.197 0.484 Australia 8 9 117 131 0.208 0.451 Greece 9 21 106 24 0.236 0.469 Azarbaijan 10 8 97 100 0.188 0.452 Iceland 11 10 95 186 0.161 0.449 Cyprus 12 13 93 32 0.168 0.442 Spain 13 22 90 53 0.145 0.403 Norway 14 6 86 291 0.182 0.459 United Kingdom 15 26 85 3 0.160 0.414 Malta 16 14 78 20 0.173 0.494 Israel 17 23 74 35 0.124 0.374 Denmark 18 12 72 51 0.236 0.545 Czech Re-public 19 11 70 7 0.200 0.433 Germany 20 25 70 0 0.102 0.418 Slovenia 21 15 67 59 0.109 0.482 San Marino 22 19 51 65 0.091 0.454 North Macedonia 23 7 48 58 0.154 0.464 Estonia 24 20 47 48 0.157 0.437 Albania 25 17 38 47 0.163 0.462 Belarus 26 24 35 13 0.105 0.440

(24)

Prediction table of the ﬁnal jury points in 2019 (model C)

Predicted

rank Actual rank Predicted jury points Actual jury points Polarity Subjectivity

Italy 1 2 171 237 0.246 0.445 France 2 16 130 40 0.203 0.417 Switzerland 3 4 124 202 0.181 0.512 Sweden 4 5 121 126 0.208 0.513 Serbia 5 18 119 67 0.179 0.452 Netherlands 6 1 118 241 0.195 0.445 Australia 7 9 115 152 0.208 0.451 Russia 8 3 111 219 0.197 0.484 Greece 9 21 107 24 0.236 0.469 Azarbaijan 10 8 98 150 0.188 0.452 Iceland 11 10 93 153 0.161 0.449 Spain 12 22 91 50 0.145 0.403 Cyprus 13 13 91 28 0.168 0.442 Norway 14 6 88 247 0.182 0.459 United Kingdom 15 26 88 1 0.160 0.414 Denmark 16 12 80 46 0.236 0.545 Israel 17 23 80 35 0.124 0.374 Malta 18 14 78 18 0.173 0.494 Czech Re-public 19 11 76 8 0.200 0.433 Germany 20 25 73 0 0.102 0.418 Slovenia 21 15 64 77 0.109 0.482 Estonia 22 20 57 46 0.157 0.437 North Macedonia 23 7 54 69 0.154 0.464 Albania 24 17 48 43 0.163 0.462 San Marino 25 19 46 87 0.091 0.454 Belarus 26 24 44 12 0.205 0.440

Predicting Eurovision Song Contest results using sentiment analysis on tweets