The voice of Twitter: observable subjective well-being inferred from tweets in Russian

Subjective well-being data

SWB in Russia is measured by a series of research organizations (Almakaeva & Gashenina, 2020), such as Russia Public Opinion Research Center (VCIOM), Levada Center, Public Opinion Foundation, the Russia Longitudinal Monitoring Survey of HSE University, the Institute of Psychology of the Russian Academy of Sciences, Ronald F. Inglehart Laboratory for Comparative Social Research of HSE University, and the Center for Sociological Research of Russian Presidential Academy of National Economy and Public Administration. Following our previous study on OSWB in Russia (Smetanin, 2022a), we decided to use the VCIOM Happiness Index because it is available on an almost monthly basis starting from 1990, as presented at Fig. 1. The VCIOM Happiness Index is built on the basis of the following question: “Life is full of good and bad moments. But in general, are you happy or not?” It is calculated as the difference between the sum of positive answers (“Yes”, “Rather Yes”) and negative answers (“No”, “Rather No”). The index is measured in points and can range from −100 (all respondents are unhappy) to 100 (all respondents are happy). The higher the index value, the happier Russians feel. The survey method is a telephone interview of a stratified random sample of 1,600 respondents with landline and mobile numbers. Positive affect and negative affect indices are not explicitly reported by VCIOM, but they can be easily calculated based on their data. Positive affect is calculated as the share of respondents who answered “Yes” or “Rather Yes” on the survey question, and negative affect is calculated as the share of respondents who answered “No” or “Rather No”.

Twitter data

We selected the Twitter Stream Grab (https://archive.org/details/twitterstream) as a data source of tweets in Russian. Twitter Stream Grab is a publicly available historical collection of JSON content grabbed from the general Twitter “Spritzer” API stream. According to Twitter, this API provides a 1% sample of the complete public tweets and is not tied to specific topics, so we considered it as a good source of general-domain tweets. Additionally, several studies (Wang, Callan & Zheng, 2015; Leetaru, 2019) performed independent validation of the representativeness of this stream. At the time of writing, Twitter Stream Grab included tweets from September 2011 to May 2021 grouped by month; however, some months were missing. VCIOM Happiness was available on a monthly basis from April 1990 to March 2022, but some months were missing too (see Fig. 1). Since our goal was to compare VCIOM Happiness and Twitter-based indices, we selected for further analysis only those months for which both VCIOM Happiness and Twitter data³ was available. Following this selection strategy, we found 20 months: April 2014, November 2015, April 2016, November 2016, July 2017, March 2018, July 2018, November 2019, April 2020, May 2020, June 2020, July 2020, August 2020, September 2020, October 2020, November 2020, December 2020, March 2021, April 2021, and May 2021.

Considering that the share of geotagged tweets in Russian is extremely low in Twitter Stream Grab (see “Geotagged Tweets in Russia”), we decided to focus on Russian-language tweets in general because they seem to predominantly consist of tweets published from Russia (see “Tweets in Russian and tweets from Russia”). Data were filtered as previously described in Smetanin (2022b). Specifically, since Twitter Stream Grab consists of tweets in different languages, our next step was to remove tweets written in non-Russian languages. Each tweet from this data source already contained information about the language of the text automatically detected⁴ by Twitter, so the language filtering procedure was fairly straightforward. After filtering out all non-Russian tweets and retweets, we obtained 10,869,003 tweets in Russian posted by 1,955,827 unique users (5.55 tweets per user on average) for the selected months.

Training data

As the training dataset for the classification model, we chose RuSentiTweet (Smetanin, 2022b), the largest general-domain tweet sentiment analysis dataset in Russian. RuSentiTweet was generated from Twitter Stream Grab—the same data source that we chose to use in our study—so this dataset is ideal for our study. Compared to other available Russian language tweet datasets (Pontiki et al., 2016; Mozetič, Grčar & Smailović, 2016; Lukashevich & Rubtsova, 2016; Loukachevitch et al., 2015; Rubtsova, 2013), RuSentiTweet is the only general-domain dataset of tweets with manual annotation and reported inter-annotator agreement. RuSentiTweet consists of 13,392 tweets annotated into five classes: Positive, Neutral, Negative, Speech Acts, and Skip. Sentiment classes were previously described in Rogers et al. (2018) and Smetanin (2022b). Specifically, Positive and Negative tweets represent positive and negative sentiments or attitudes, respectively. Neutral tweets are tweets that simply describe a situation in a neutral, factual way and do not contain overt positive or negative sentiment. Speech Act tweets perform the functions of various speech acts, such as greeting someone, congratulating someone, or expressing gratitude for something. Although these tweets also express positive sentiment, they are considered a separate sub-category because they can also be made under social pressure or out of a sense of duty (Rogers et al., 2018). Skip tweets represent noisy and vague sentiments or attitudes.

Models

For sentiment classification of Russian texts, approaches based on language models tend to outperform rule-based and basic machine learning–based approaches in terms of classification quality (Smetanin & Komarov, 2021). As reported in the paper presenting RuSentiTweet (Smetanin, 2022b), the SOTA result on RuSentiTweet was achieved by RuBERT (Kuratov & Arkhipov, 2019), a version of multilingual BERT (Devlin et al., 2019) trained on the Russian part of Wikipedia and Russian news. Since RuSentiTweet was published recently (a couple of months ago), so far there are no other works in which classifiers were trained on it—and on which we can rely to select the most effective approach. However, RuSentiTweet was annotated using the same annotation guidelines as RuSentiment (Rogers et al., 2018) (dataset of Russian-language posts from VKontakte), so we can conceivably rely on RuSentiment studies to select potentially efficient models. According to our recent paper (Smetanin, 2022a), RuRoBERTa-Large (Sberbank, 2021) established new SOTA results on RuSentiment and significantly outperformed RuBERT, so we suppose that it may also outperform RuBERT on RuSentiTweet. RuBERT (https://huggingface.co/DeepPavlov/rubert-base-cased) is a version of BERT (Devlin et al., 2019) trained on the Russian part of Wikipedia and Russian news. RuRoBERTa-Large (https://huggingface.co/sberbank-ai/ruRoberta-large) is a version of RoBERTa (Zhuang et al., 2021) model with BERT-Large (Devlin et al., 2019) architecture and the BBPE tokenizer from GPT-2 (Radford et al., 2019) trained on Russian texts. RuRoBERTa-Large is ranked higher that RuBERT in RussianSuperrGLUE (Shavrina et al., 2020) leaderboard (https://russiansuperglue.com/leaderboard/2), so we expected that it will show higher results on RuSentiTweet also.

Sentiment classification

Fine-tuning of RuRoBERTa-Large was performed using the Transformers library (Wolf et al., 2020) on two Tesla V100 SXM2 32GB GPUs with the following hyperparameters: a number of train epochs from 3 to 5; the number of warm-up steps is 0%, 5%, or 10%; a max sequence length of 128 or 256; a batch size of 18 or 32, a learning rate of 5e−5, an Adam optimiser, and a softmax activation function. RuSentiTweet consists of test (20%) and training (80%) subsets, so we used the test subset for computing final classification metrics and the training subset (with additional division to the validation subset) for the model training. We repeated each experiment three times and reported mean values of the measurements.

The best model (four epochs, 10% warm-up steps, 128 max sequence length, 32 batch size) demonstrated macro $F_1=0.7229$ and weighted $F_1=0.7281$, surpassing the existing SOTA results (Smetanin, 2022b) achieved by the RuBERT model (see Table 1). This result was expected because RuRoBERTa-Large is not only higher in the RussianSuperrGLUE (Shavrina et al., 2020) leaderboard (https://russiansuperglue.com/leaderboard/2), but RuRoBERTa-Large also outperformed RuBERT in the sentiment analysis task in our previous study (Smetanin, 2022a). Since RuSentiTweet was released a few months ago, there are currently no other works with which to compare the quality of the classification, but the magnitude of the results generally aligns with the results achieved by other approaches on other five-class sentiment datasets (Smetanin, 2022b).

As can be seen from the confusion matrix (see Fig. 2), Neutral tweets were commonly misclassified as Positive and Negative, and vice versa. Also, the Skip class was commonly misclassified as Neutral. Speech Act was clearly separated from other classes, except for Positive, because it also represents positive sentiment. In general, the confusion matrix is very similar to that of RuBERT (Smetanin, 2022b). However, the biggest difference is observed in the Speech Act class: RuRoBERTa-Large was able to separate it much more efficiently ( $F_{1,speech}=0.8151$) from the rest of the classes than RuBERT ( $F_{1,speech}=0.7444$).

OSWB index formulas

We selected the following OSWB indicators.

Definition 1. $OSW{B}_{PA}$ is the positive affect indicator (experiencing pleasant emotions and moods) relative to all expressed sentiment and is defined as follows:

(1) $$OSW{B}_{PA}={\displaystyle \frac{POS}{POS+NEG+NEU+SA+SKIP}OSW{B}_{PA}\in [0;1],}$$where POS is the number of positive posts, NEG is the number of negative posts, NEU is the number of neutral posts, SA is the number of posts with greetings and speech acts, and SKIP is the number of ambiguous posts that cannot be unambiguously assigned to one of the other classes (Smetanin, 2022a).

Definition 2. $OSW{B}_{PA,Neu}$ is the positive affect indicator (experiencing pleasant emotions and moods) relative to neutral sentiment and is defined as follows:

(2) $$OSW{B}_{PA,Neu}={\displaystyle \frac{POS}{NEU},OSW{B}_{PA,Neu}\in [0;+\mathrm{\infty }]}$$

Definition 3. $OSW{B}_{NA}$ is the negative affect indicator (experiencing unpleasant, distressing emotions and moods) relative to all expressed sentiment and is defined as follows:

(3) $$OSW{B}_{NA}={\displaystyle \frac{NEG}{POS+NEG+NEU+SA+SKIP},OSW{B}_{NA}\in [0;1]}$$

Definition 4. $OSW{B}_{NA,Neu}$ is the negative affect indicator (experiencing unpleasant, distressing emotions and moods) relative to neutral sentiment and is defined as follows:

(4) $$OSW{B}_{NA,Neu}={\displaystyle \frac{NEG}{NEU},OSW{B}_{NA,Neu}\in [0;+\mathrm{\infty }]}$$

Definition 5. $OSW{B}_{Net}$ is the net affect indicator (difference between experiencing pleasant and unpleasant emotions and moods) relative to all expressed sentiment and is defined as follows:

(5) $$OSW{B}_{Net}={\displaystyle \frac{POS-NEG}{POS+NEG+NEU+SA+SKIP},OSW{B}_{NA}\in [-1;1]}$$

Definition 6. $OSW{B}_{Net,Neu}$ is the net affect indicator (difference between experiencing pleasant and unpleasant emotions and moods) relative to neutral sentiment and is defined as follows:

(6) $$OSW{B}_{Net,Neu}={\displaystyle \frac{POS-NEG}{NEU},OSW{B}_{Net,Neu}\in [-\mathrm{\infty };+\mathrm{\infty }]}$$

Demographic groups

Among demographic characteristics, we selected only the gender of the user, whereas in our previous study on Odnoklassniki data both gender and age group were used. Although the author’s age is also an important demographic characteristic that affects SWB in Russia (Rodionova, 2015), we did not extract age from the Twitter data for our study because Twitter is practically unused by the older generations in Russia (Brodovskaya, Dombrovskaya & Sinyakov, 2016). Even if there are some users from the 55+ age group, their share will be extremely small, and as a result, post-stratification by age may negatively affect the index calculation result⁵ .

There was no information about the user’s gender in the Twitter data, so to determine it, we trained a classification model that reached $F_1=0.9835$ in the task of binary gender classification by the user’s full name (see “Gender detection in Russian-speaking Twitter”). We applied this model on Twitter data and found that the distribution of men and women is quite different from the distribution for the Russia’s population. Whereas over the past 10 years Russia’s population was about 46% male and 54% female (Federal State Statistics Service (Russia), 2021), the Twitter users in our data were about 59% male and 41% female.

Misclassification bias

Given that our sentiment classification model is not error-free, it may introduce a bias in our calculated OSWB indicators. To assess the potential impact of misclassification errors on each of the calculated indicators (see Eqs. (1–7)), we applied the simulation approach for misclassification bias assessment in social indicators research (Smetanin & Komarov, 2022). The core idea of this simulation approach is to simulate the true indicator (i.e., real underlying indicator that can be obtained by 100% error-free algorithm); then, on this basis, approximate the results of the classification algorithm using a confusion matrix; and then calculate the quality metrics. We chose Spearman’s correlation coefficients as the main metric and ran 500,000 simulation iterations for each indicator. According to the results of the simulation, the aggregated p-values are higher than 0.95, and both coefficients demonstrated almost perfect aggregated correlation scores. According to our results, we did not confirm that our model’s misclassification errors have more than a negligible impact on calculated indicators.

OSWB measurement

To calculate the indices, we applied the approach described in Smetanin (2022a). The following are summaries of the basic steps for building indices; the full methodology can be found in Smetanin (2022a). When collecting data, we carried out the initial filtering—that is, we left only Russian-language tweets and removed retweets. After filtering out all non-Russian tweets and retweets, we obtained 10,869,003 tweets in Russia posted by 1,955,827 unique users (5.55 tweets per user on average) for the selected months. Next, we transitioned from tweet-level analysis to user-level analysis. To do this, within each analyzed time interval (in our case, within each month), it is necessary to aggregate the sentiment of tweets that each user published. The main motivation for this step is that, first, for post-stratification by demographic characteristics, we need to operate at the level of users, not tweets; second, we need to remove the repeated sharing of emotions associated with the same event, which arises within the phenomenon of social sharing of emotions (Rimé et al., 1991); and thirdly, we need to ensure that tweets published by more active users are not overrepresented in our sample (Németh & Koltai, 2021). We aggregated sentiment on a user level with a majority voting strategy. After that, for each analyzed month, we counted the number of users who published at least one tweet. Although the central idea behind digital data collection for computational social science is collecting as much data as possible (Hox, 2017), it still required verifying that the minimum sample size requirements are fulfilled. The VCIOM Happiness index typically has samples of 1,600 respondents (VCIOM, 2022) for each analysed time period, so we decided to consider this number the minimum sample size for our research. Other research organizations that measure SWB have similar or even smaller sample sizes for each analysed time period and country: for example, FOM The Mood of Others has samples of 1,600 respondents (FOM, 2022), World Values Survey has samples of 1,200 respondents (WWS, 2020), and Standard and Special Eurobarometer surveys typically have samples of 1,000 respondents (GESIS, 2020). We confirmed that for each analysed month, there were more than 1,600 users. Finally, we calculated indices post-stratified by user gender. Information on the gender distribution of the population of Russia was taken from the official statistics of Rosstat.

Reliability

For reliability measurement, we followed the approach from our previous study on Odnoklassniki data (Smetanin, 2022a) and checked the correlation of OSWB indicators with the VCIOM Happiness Index, VCIOM Positive Affect, and VCIOM Negative Affect. According to previous studies (Stock, Okun & Benito, 1994; Krueger & Schkade, 2008; Levin & Currie, 2014; Kapteyn et al., 2015; Lucas, 2018,), the typical reliability of SWB scales is in the range of 0.50 to 0.84. For single-item SWB measures (like VCIOM measures), the typical reliability is between 0.40 and 0.66 (Krueger & Schkade, 2008). However, considering that our users sample also contains Russian speakers who live outside of Russia (see “Tweets in Russian and Tweets from Russia”) and that age groups were not taken into account during indices calculation⁶ , we expected to have a lower correlation than that achieved in our study on Odnoklassniki data (Spearman’s $r_s=0.825$) (Smetanin, 2022a). The results of the reliability measurement are presented in the next section.

Geotagged tweets in Russia

Initially, we planned to use geotagged tweets to select those that were published in Russia but found two major concerns with this approach. First, we approximately estimated the share of geotagged tweets in Russia based on existing data sources and found that this number is too small for conducting OSWB study. Considering that Twitter generates around 200 billion tweets per year (Gao et al., 2022), only around 373.2 million tweets are generated in Russia (Brand Analytics, 2021), approximately 0.85% of tweets are geotagged (Sloan et al., 2013), and Twitter Stream API provides approximately 1% sample of tweets, we can estimate that the share of geotagged tweets from Russia is approximately 0.0000013, and Twitter Stream API can provide around 220 tweets from Russia per month, which is too small for OSWB study. Second, Sloan & Morgan (2015) found that there are significant demographic variations between those who opt in to geoservices and those who geotag their tweets and suggested that Twitter users who publish geographical information are not representative of the wider Twitter population.

Tweets in Russian and tweets from Russia

Since in this study we are comparing OSWB indicators with the VCIOM Happiness Index for Russia but the OSWB indicators are based on Russian-language tweets (and not tweets from Russia), the question arises of what proportion of tweets in Russian are actually published from Russia. According to various estimates (Arefyev, 2013; BusinesStat, 2021; Kaganov, 2021), most of the Russian-speaking population lives in Russia, and this proportion is growing over years. Moreover, whereas almost all (95.76%) of Russian-speaking population in Russia use Russian as a language of communication, less than a half (45.54%) of Russian speakers outside Russia do so (BusinesStat, 2021). Based on these data, we approximately estimated the share of Russian speakers who use Russian as the language of communication starting from 2004 (see Table 3). If we assume that if a person uses Russian as a language of communication then the person also communicates in Russian on Twitter, then we can assume that in 2004 68% of Russian-language tweets were published from Russia, and by 2019 this number increased to 73.7%. However, it is worth noting that this estimate is rather superficial and does not take into account other important factors, such as the level of internet penetration and Twitter penetration in each country. Thus, it seems that filtering by language without filtering by location allows selecting almost all users from Russia; however, about a quarter of Russian-speaking users who do not live in the territory of Russia are also presented in the sample. This estimation indicates that we should not expect high correlation between OSWB indicator based on Russian-language tweets and VCIOM indices.

Gender detection in Russian-speaking Twitter

Previous studies on gender detection of Russian social media users have shown that depending on the source of data from which gender is determined, the quality of the classification changes significantly. For example, one of the tasks of RusProfiling PAN at FIRE Track competition (Litvinova et al., 2017b) was to to classify the gender of tweets in Russian based on their text. The best system (Markov et al., 2017) reached only 68.25% accuracy in binary classification, and the organizers of the track hypothesised (Litvinova et al., 2017b) that the problem was that tweets are shorter and grammatically poorer in comparison with other sources (e.g., Facebook posts or essays). There were other studies on text-based gender detection that reported higher results, such as Sboev et al. (2019), which showed 0.62 accuracy on LiveJournal posts; Litvinova et al. (2017a) achieved accuracy of 0.72 on tweets and 0.71 on Facebook posts; and Bogachev et al. (2020) reported $F_1$ of 0.84 on RusPersonality (Litvinova et al., 2016). Besides linguistic characteristics, some studies also considered information from user profiles as well as social graph. For example, Tolmachev (2019) achieved accuracy of 0.715 by utilising social graph information. At the same time, it seems that the task of gender detection for Russian social media users tends to be much easier if there is access to the user’s full name. Panchenko & Teterin (2014) showed on a dataset of 100,000 Russian full names from Facebook that simple and computationally efficient models (e.g., n-grams with Logistic Regression) yield excellent results and are able to achieve accuracy up to 96%. The high accuracy of classification based on the full name is quite expected because, first, in the Russian language, first names are usually clearly divided into male and female, and second, the surnames can commonly be attributed to one of the genders based on endings. Thus, considering that our data source contains full names of Twitter users, we decided to use an approach that relies only on the full name for gender detection.

Since the datasets used in existing works for gender detection are not publicly available, we created our collection of full names and genders based on data from Russian social networks. It is not possible (at least now) to automatically create this collection from Twitter because the Twitter API does not provide information about a user’s gender. Thus, we decided to focus on data from VKontakte—the largest Russian online social media and social networking service, which is predominantly used by Russian speakers. In VKontakte, we selected the top-1 group (https://vk.com/public27895931) based on Medialogia rating (Medialogia, 2021) and loaded all the users with specified gender via VKontakte API (https://dev.vk.com/method/groups.getMembers). We collected 13,126,794 VKontake profiles, where each profile contains gender and first and last names written in Cyrillic and Latin alphabets. We did not consider users with unknown gender. Among these profiles, we obtained 6,521,854 unique full names in the Cyrillic alphabet and 6,263,813 unique full names in the Latin alphabet. The number of full names in the two alphabets differs due to the fact that different names in the Cyrillic alphabet can have both the same and different transliteration into the Latin alphabet, and even similar names in one alphabet can have different transliteration to another alphabet. Based on the data obtained, we formed the final dataset for training models using the following logic: if the user’s name in Latin and Cyrillic is different, we added both names to the dataset; if they are the same, we added only one name. The final dataset contains 25,101,673 names (46% male and 54% female). Interestingly, the distribution of male and females in our dataset almost perfectly matches the male and female distribution of population in Russia: over the past 10 years, the proportion of males has been around 46% males and 54% females (Federal State Statistics Service (Russia), 2021).

Following the approach by Panchenko & Teterin (2014), we used L2-regularized Logistic Regression with character n-grams to classify gender. In order to identify the best hyper-parameters (e.g., character n-grams type, n-grams range, usage of IDF, TF-IDF normalisation type), we first ran a grid search with 10-fold cross-validation (80%—training subset, 20%—test subset) on a random sample of 100,000 full names. The model with character n-grams inside word boundaries, n-grams range of (2, 7), usage of IDF, L2 TF-IDF normalisation, and ignoring terms that appear in more than 50% of the documents showed the best $F_1$ score of 0.9771, so we used these hyper-parameters to train the final model on the whole dataset. The final model trained in the full dataset demonstrated $F_1=0.9835$ on the test subset (20% of full names). The magnitude of our results is in line with the results achieved by Panchenko & Teterin (2014) (96%) on Russian full names from Facebook, which can be explained by the fact that Russian full names commonly contain grammatical information about the gender. Finally, we applied this model to names from Twitter to classify users’ gender.

Holidays in Russain-speaking Twitter

Considering that positive emotions expressed in social media have stable fluctuations during different time periods, such as weekly (Dzogang, Lightman & Cristianini, 2018; Smetanin, 2022a) and monthly variations, we decided to focus only on speech acts and greeting as the proxy measure of the popularity of holidays because so far there has been no evidence that they also have strong fluctuations.

Definition 7. $OSW{B}_{SA}$ is the the share of greetings and speech act tweets relative to relative to all expressed sentiment and is defined as follows:

(7) $$OSW{B}_{SA}={\displaystyle \frac{SA}{POS+NEG+NEU+SA+SKIP},OSW{B}_{SA}\in [0;1]}$$

Taking into account that there is also some evidence that people of different genders may have different attitudes towards different holidays (VCIOM, 2018), we decided to first calculate the holiday rating for each gender separately, and then build an aggregate rating.

The results of the analysis are presented in Table 4.

• Official public holidays.

• – The New Year is the most popular holiday on Russian-speaking Twitter: the share of greetings and speech acts on December 31 exceeds the average annual value by more than three times and reaches 12.3% of all tweets for that day. The share of greetings and speech acts decreases on January 1 to 5.3%, but is still quite high and ranks 7th among all days of the year. This observation is fully consistent with the VCIOM (2018) survey, in which the new year was also the most popular holiday (celebrated by 96% of people), and with the study by Volovikova (2018), where it is listed as a favorite holiday for all groups.

• – Defender of the Fatherland Day and International Women’s Day share the second and third places in the ranking of the most popular holidays; however, they are significantly inferior to the New Year in the share of posts with speech acts and congratulations. Generally, this corresponds to the data of VCIOM (2018), which reports that these two holidays are celebrated by 84% and 88% of Russians, respectively.

• – Victory Day is one of the most popular holidays to follow after the New Year, Defender of the Fatherland Day, and International Women’s Day. This is partially consistent with the results of the VCIOM (2018) survey, which showed that Victory Day is one of the most popular holidays; however, according to their results, it is more popular than Defender of the Fatherland Day and International Women’s Day.

• – Russia Day is not among the 10 most popular holidays on Russian-speaking Twitter, ranking 11th in the overall ranking. In the VCIOM (2018) survey, this holiday is also one of the least celebrated compared to other official holidays.

• – Union Day is ranked even lower that the Russia Day, which perfectly aligns with the VCIOM (2018) survey.

• – International Workers’ Day is not really popular on Russian-speaking Twitter, and in comparison with other non-holiday days, it has only a slight increase in the share of speech acts and congratulations, ranking only 15 out of 22. Young Russians, who predominate on the analysed Twitter sample, avoid any ideological interpretation of the meaning of this holiday, more often perceiving May 1 as just an additional day off (VCIOM, 2004). Moreover, over a quarter of a century, the number of Russians who plan to celebrate Labor Day has been gradually decreasing (VCIOM, 2019b).

• – The official holiday of Christmas, which is celebrated on January 7, was at the very bottom of the ranking (20 out of 22). At the same time, the VCIOM (2018) survey showed that Christmas in one of the main holidays in Russia which is being celebrated by 77% of population. We hypothesise that part of this can be explained by the fact that the younger age group that dominates Twitter is less religious than the older age group (Kofanova & Mchedlova, 2010). At the same time, Christmas, which is also celebrated by some on December 25, while not being an official holiday, ranked higher in the ranking and took 9th place. However, the high value on this day can also be caused by the approach of the main holiday of the year: the New Year.

• Foreign holidays. Halloween on Russian-speaking Twitter is one of the most popular foreign holidays, and is in 9th place among all days of 2020 in terms of the share of greetings and speech acts. This finding is not consistent with the results of VCIOM (2018), but we assume that this is due to the fact that Twitter is dominated by a younger age group that is more inclined to celebrate Halloween (VCIOM, 2019a), whereas the VCIOM (2018) survey provides a representative sample of the Russia population. Other foreign holidays, such as Chinese New Year and St. Patrick’s Day, are practically not celebrated on Russian-speaking Twitter: the share of greetings and speech act posts in 2020 these days was not only extremely small, but was generally lower than the yearly average (275th and 335th place, respectively). Overall, Chinese New Year and St. Patrick’s Day have not gained popularity on Russian-language Twitter and sit at the bottom of the overall rankings. This finding aligns with the results reported by VCIOM (2018): these holidays have not yet become widespread in Russia and they are celebrated by only 3–5% of the population, while 95–96% of respondents claim that they do not celebrate them.

• Other common holidays. Among other holidays, one of the most popular is Knowledge Day, which occupies the penultimate place in the top 10 most popular holidays.

Interestingly, on average, women are more likely to post greetings and speech acts tweets, but this pattern is observed not only on holidays, but also on ordinary days.