Sentiment analysis of vegan related tweets using mutual information for feature selection

Approaches for sentiment analysis

The unprecedented growth of user-generated content that we are witnessing nowadays on the World Wide Web (WWW) implies that for the first time in humankind’s history, the broad public can express their opinions on various subjects almost effortlessly. Of course, this happened with the help of contemporary technologies and Social Network (SN) phenomena emerging in the early 2000s. And analyzing public opinions on subjects of interest is vital for government and business.

Sentiment analysis (SA) or opinion mining (OM) is identifying and classifying people’s attitudes toward a subject or entity using the quantitative methods of statistics, computational linguistics, and machine learning (ML) (Neethu & Rajasree, 2013). The sentiment analysis topic has been a subject of rigorous research in the last years (Ligthart, Catal & Tekinerdogan, 2021), and the beginning of the research in this area can be traced back to the middle of the 20th century (Sims, 2015).

SA tasks can be divided into the following major categories (Hemmatian & Sohrabi, 2019; Ligthart, Catal & Tekinerdogan, 2021):

1. Opinion extraction is the process of user opinion extraction from WWW to discover their way of thinking.

2. Sentiment mining is the separation of objective or factual sentences from the subjective ones that contain opinions and sentiments.

   Subjectivity classification attributes a sentence to a class of subjectivity it carries.

3. Sentiment classification categorizes the opinions discovered into predefined classes.

   Polarity determination whether it is positive, negative, or lacks the polarity (neutral).

The approaches for sentiment analysis can be classified as follows (Adwan et al., 2020); see also Fig. 2:

1. Lexicon based approaches use dictionaries (Hu & Liu, 2004) of words that are assumed to be positive, negative, or neutral, and then weigh or score the text according to the relative content of these words to infer its sentiment (Akilandeswari & Jothi, 2018; Das et al., 2018; Azizan et al., 2019).

   Dictionary-based approaches use a vocabulary look-up technique to detect sentiment-rich words and their polarity.

   Corpus-based approaches work with the statistical models developed on many texts to leverage the information on the context of word use and the dictionary-based approach.

2. ML approaches can be subdivided into Supervised and Unsupervised Learning classes of tasks:

   Supervised learning methods require a dataset with items labeled for the sentiment to train classification modeling algorithms (SVM, KNN, DTC, LR, and others).

   Unsupervised learning methods search for the structure of information within the presented data using clustering, topic modeling, and mapping algorithms (Mukhamediev et al., 2020; Kirill et al., 2020).

3. Hybrid approaches, as the name implies, use both the Lexicon-based and ML approaches together.

A challenging task in sentiment analysis is the domain specificity of the developed approaches, which come from the fact that sentiment-rich words can carry different polarity in usage across domains. For example, the term “aggressive” can be perceived as positive when related to sports (“Chelsea played aggressively tonight in a match with Barcelona club.”) and negative in the political domain (“Russian troops aggressively attack Mariupol city in Ukraine, thousands of civilians get injured.”). Thus, topic-aware or domain-specific models are advantageous over the general ones (Akhmetov, Gelbukh & Mussabayev, 2022).

Another problem that definitely should be addressed is the possible data imbalance on the sentiment classes, especially in the domains where most publications are negative (war, accident, crime news). Handling the data imbalance can be done with under-over-sampling techniques, data augmentation, and others (Omarkhan, Kissymova & Akhmetov, 2021; Kelsingazin, Akhmetov & Pak, 2021).

One of the classical works on the Twitter micro-blog post (tweet) analysis is VADER (Hutto & Gilbert, 2014), a simple rule-based sentiment analysis model. VADER employs an empirically validated gold-standard lexical feature list with the associated sentiment or opinion polarity and specializes in micro-blog short text messages. The model shows an F1 classification accuracy of 0.96, outperforming human raters with a 0.84 result. Another model trained on Twitter data (Kocaman & Talby, 2021) shows an accuracy of 0.80.

Related works

Analyzing posts in SN can be of particular interest to government administration, business officials from the marketing and sales side, social scientists, and anthropologists, as they contain vast amounts of information characterizing the post writers’ origins, cultural background, political views, and attitudes to a broad spectrum of objects, events and more abstract things as ideology.

One of the most popular SN is Twitter: the micro-blogging platform, where its 1.3 billion users make 500 million tweets every day (Alhgren, 2022). Analyzing Twitter posts or the so-called tweets involves exploring the interest, thoughts, and opinions in various aspects and domains (Adwan et al., 2020).

The application examples of tweet analysis are to understand people’s opinions towards different topics such as COVID-19, disasters, presidential elections, armed conflicts across the Globe (Ukraine, Syria, Palestine, Iraq), sports, and healthy lifestyles. Much research has focused on mining public opinion on various social, cultural, economic, and environmental topics on Twitter (Reyes-Menendez, Saura & Alvarez-Alonso, 2018; Reyes-Menendez, Saura & Filipe, 2020; Saura, Reyes-Menendez & Palos-Sanchez, 2019; Serrano et al., 2021; Das et al., 2018; Pila, Kvasnikov Stanislavsk & Kvasnika, 2021; Islam, 2019; Gore, 2015; Karami et al., 2018; Pila, Kvasnikov Stanislavsk & Kvasnika, 2021; Saura, Palos-Sanchez & Rios Martin, 2018). Reyes-Menendez, Saura & Alvarez-Alonso (2018) investigated the main factors that worry people about the environment and public health. Analysis of health-related tweets by Islam (2019) revealed an interesting correlation between yoga and veganism. Gore (2015) studied geographic issues of obesity rate in America. The main characteristics of public perceptions of obesity, diabetes, nutrition, and exercise were explored by Karami et al. (2018). Reyes-Menendez, Saura & Filipe (2020) identified the main challenges involved in the MeToo movement. In Saura, Reyes-Menendez & Palos-Sanchez (2019) authors explored negative and positive perceptions of consumers related to the Black Friday event. Next, Serrano et al. (2021) examines online reviews of green Airbnb users to find the main aspects among which the term sustainability appears. Pila, Kvasnikov Stanislavsk & Kvasnika (2021) explored all tweets between 2019 and 2020 that included the hashtag “healthyfood”.

Data collection

We study Twitter messages or tweets related to veganism to gain insight into the change in the sentiment of such posts during the COVID-19 world pandemic, before it, and just after it, including publication of studies and documentaries about veganism.

Scraping the twitter

For this study, data was collected from Twitter between January 1st, 2010, and May 19th, 2022, using Python snscrape module (JustAnotherArchivist, 2022). The search was performed using the following keywords: “vegan”, “vegetarian”, “veggie”, “veganism”, “plant-based”, “cruelty-free”, “plant milk”, “tofu”. We refer to a keyword as a word that is present either in the form of a hashtag or anywhere in the body of a tweet. As we have seen in the Kaggle dataset, out of 1.6 million tweets, the tweets containing at least one of the keywords amounted to 1,876, which means a 0.0011725 content ratio. Thus, assuming 500 million tweets daily, we would have approximately 586,250 of our target tweets. Therefore, for the sake of getting the data fast, we limited the collection of data by 100 tweets for every keyword monthly along the 136 months and 19 days period collecting 113,677 tweets in total.¹

Among the collected tweets, 87% were in English, 3% in Japanese, and 2% in Spanish. We have also labeled tweets in other languages with less than 1000 tweets as “other”; additionally, Twitter has a category “und” for undefined languages; see Fig. 3. Besides English, there are numerous Japanese tweets, which signal that the vegan topic is trendy in Japan.

Table 1:

Kaggle Sentiment140 dataset vegan keyword content.

Keyword	Tweet count
vegan	364
plantbased	0
healthy	955
vegetarian	191
veggie	391
veganism	5
cruelty free	1
plant milk	0
beyond meat	0

DOI: 10.7717/peerjcs.1149/table-1

From Table 2, we see that 70% of the tweets do not have a like, reply, retweet, or quote and are entirely unnoticed. As for the most popular tweets, they comprise the top 10% by replies/retweets or having at least one reply or retweet, the top 30% by having at least one like, and the top 1% by having at least one quote.

Table 2:

A number of replies, retweets, likes, and quotes in the scraped tweets.

	Replies	Retweets	Likes	Quotes
count	113,677
mean	0.33	0.53	3.14	0.05
std	3.92	23.02	344.35	4.20
min	0.00	0.00	0.00	0.00
50%	0.00	0.00	0.00	0.00
70%	0.00	0.00	1.00	0.00
80%	0.00	0.00	1.00	0.00
90%	1.00	1.00	3.00	0.00
95%	1.00	2.00	5.00	0.00
97%	2.00	3.00	9.00	0.00
99%	3.00	7.00	28.00	1.00
max	1,119.00	7,492.00	115,499.00	1,398.00

DOI: 10.7717/peerjcs.1149/table-2

Taking into account that the majority of the tweets have no likes, we have decided to analyze only the tweets that have at least one like. Thus, we have 36,959 tweets to explore; see Fig. 4 for the distribution of the liked tweets by the query.

After filtering out the tweets with no likes, we observed in the Fig. 5 that there had been a steady rise in the number of tweets on the “vegan” topic since 2012 from less than 1,000 tweets per year to more than 5,000 tweets in 2021. This can be explained by the fact, that starting from 2011, a large amount of research has begun to appear on the health benefits of a plant-based diet. So, we suppose that the main reason for the rise of the trend in health care, besides the support of animal rights and environmental protection.

Moreover, there is a pattern of having relatively more tweets on the “vegan” subject during the spring months, which plummet in the middle of the summer and start rising in autumn again; see Fig. 6. The pattern can be explained by the fact that in the winter and spring months, people generally have a vitamin deficiency and start focusing on their health more.

Data description

The collected data set consists of 113,677 tweets. In the data set, 2,682 tweets correspond to health topics, containing healthy, wholesome, and full-blooded keywords. Also, there are three levels of vegan topic detailing, namely “vegan-1”, “vegan-4”, and “vegan-7”. The level of “vegan-1” corresponds to just one keyword, namely “vegan”, and there are 32653 tweets of it. The level of “vegan-4” is about “vegan, vegetarian, veggie, veganism,” and there are 53,201 tweets. The level of “vegan-7” is with the following keywords: “vegan, vegetarian, veggie, veganism, plant-based, plant-based, plant based” and there are 65,213 tweets, and 2,075 tweets include keywords about both veganism and health. Also, there are corresponding intersections between the health topic and the levels of the vegan topics: 1,359, 1,791, 2,252 tweets. The methodology of data collecting assumes that tweets are distributed equally in terms of years, months, and weekdays; however, the distribution by hour expects that users wrote the most tweets in the evening.²

A visual representation of the keyword distribution can be seen in Fig. 7 from which it becomes clear which topics and discussions prevail in the presented data set. For example, the following topics: diets, recipes, animal welfare, etc.

We presented the cross-correlations between keywords in Fig. 8. You can see an anti-correlation between the keyword “plant milk” and the other keywords from the image. From the additional analysis, the cause of such phenomena is the insufficiency of plant-milk tweets against other subtopics, which leads to negative correlation values. Also, we should mention that there has been an increasing trend of mentions on Twitter “plant-milk” from 2010 to the present. Also, there is another systematic distortion in the cross-correlation diagram, particularly the high correlation value between vegan and veganism because one of them is a sub-string of another. All other features of the dataset are pretty much inter-correlated.

It is interesting to show the analysis of the relationship between health and vegan topics in a time-based manner. In Fig. 9, there is Spearman’s rank correlation between the number of tweets dedicated to these topics.

There is a peak correlation with all topics in 2012. We should notice that the correlation with the topic of cruelty-free arises to its maximum with a one-year lag. The lowest maximum owns the keyword “vegan.” The highest maximum owns the keyword “tofu.” The next explicit historical maximum was in 2015; the highest value belongs to the keyword “vegetarian,” and the lowest is the same as in 2012. In 2018 there is another peak with a different leader and outsider, namely the highest value owned to “tofu” and the lowest one to a couple of keywords “plant-based, veggie.” And the lowest peak all the time was in 2020; the leaders of the year are the keywords “plant-based, veganism”, and the outsider is the keyword “tofu”.

Another presented metric of discourse analysis of these topics is mutual information. From Fig. 9 we can see that the peak of simultaneous mentions of keywords from both topics is in 2018. Also, Fig. 9 shows an ascending trend from 2010 to 2018.

We strongly believe that such findings can give a clue to understanding the development of the social discourse on a couple of interrelated topics, “vegan” and “health.”

Pre-processing

Twitter is an excellent source for public opinion mining. Yet, we must deal with its informal language, regular misspelling, wrong word order, etc. Therefore, Twitter data cannot be analyzed in its initial raw form. Twitter users type shortened words to maintain the word length limitation (e.g., 140 characters maximum) in a tweet and tend to use digitally mediated language, so-called Slanguage (Turney, 2002), that incorporates slang, abbreviations that are natural in relaxed and unofficial context. In other words, tweets are required to be cleaned before the analysis processes. This process involves five steps: Cleaning, Case Folding, Tokenization, Slang Lookup, and Stopword Removal.

For better understanding, suppose we have a sample tweet in our data—“I’m glad he responded to you. Too bad he thinks veganism is some kind of perfection or is difficult. Seriously. Difficult? C’mon:)”. The pre-processing steps are explained below and demonstrated in Fig. 10.

Sentiment detection

Mutual information feature selection

By deleting uninformative characteristics, feature selection tries to minimize classification calculation time. Mutual information is a machine learning selection approach demonstrating how important a feature is in making an accurate prediction (Bonzanini, 2016).

We use mutual information to determine the strength of a semantic association between two words. We compute the PMI of each pair of terms and then calculate the semantic orientation given two vocabularies for positive and negative terms. If some word is commonly associated with concepts in the positive lexicon, it has a positive semantic orientation. A schematic representation of the method can be seen in Fig. 11.

To begin, we define a word’s semantic orientation (SO) as its positive and negative connections difference (Turney & Littman, 2003a; Turney & Littman, 2003b; Turney, 2002). In practice, we’re looking for a way to determine “how near” a word is to phrases like good and awful. Point-wise mutual information (PMI) is the preferred measure of “closeness,” which is defined in Eq. (3) (Church & Hanks, 2002; Turney & Littman, 2003a; Turney & Littman, 2003b; Turney, 2002). (3)${\displaystyle PMI\left(t_1,t_2\right)=log\left(\frac{P\left(t_1\wedge t_2\right)}{P\left(t_1\right)\times P\left(t_2\right)}\right),}$ where t₁ and t₂ are terms and P designates probability function.

The semantic orientation (SO) of a term t is defined in Eq. (4). (4)${\displaystyle SO\left(t\right)=\sum _{t^{\prime }\in V^{+}}^{}PMI\left(t,t^{\prime }\right)-\sum _{t^{\prime }\in V^{-}}^{}PMI\left(t,t^{\prime }\right),}$ where V⁺ and V⁻ are vocabularies of positive and negative words respectively.

Semantic orientations of some context-specific words from our collected dataset are presented in Fig. 12. As we can see from the figure, many words specifying animal products have negative semantic orientations in the vegan context. In contrast, the words associated with plant-based diets mostly have a positive semantic orientation. It is of interest to note that the word “eggs” has a far more negative score than the word “beef”, for example (−1.716 and −0.226, respectively). Note that these terms may have different scores when discussed in non-vegan contexts. At this step, we can update the general lists of positive and negative words we initially had with context-specific words from our dataset that got very high or very low scores, respectively.

The corresponding Algorithm 1 along with the pseudocode can be seen below. Python code and all supplemented data are also publicly available.³

 
_______________________ 
 Algorithm 1: Mutual information for feature selection                        ____ 
    Data: tweets in csv file 
    Result: obtained PMI for semantic orientation calculation; 
    df_clean[”tokens”] = tokenized tweets by words; 
    remove stopwords and change slang words; 
    Function build_frequencies(tweets): 
        word_count = calculated count of each word; 
   word_pairs = calculated word pairs count; 
   return (word_count, word_pairs); 
    Function build_probabilities(word_count, word_pairs, 
    tweets_count): 
        word_probability = dictionary that contains calculated probability of 
   each word; 
   word_pair_probability = dictionary that contains calculated 
   probability of each pair; 
   return (word_probability, word_pair_probability); 
    Function build_pmi(tweets): 
        word_count,word_pairs = build_frequencies(tweets) 
   word_probability,word_pair_probability = 
   build_probabilities(word_count,word_pairs,tweets_count) 
   pmi ← emptydictionary for term1 in word_probability do 
       for term2 in word_pairs[term1] do 
   denom = word_probability[term1] * word_probability[term2] 
   pmi[term1][term2] calculated by formula 1 
   end 
   end 
   return pmi, word_count; 
    // Load set of positive and negative words 
   Function get_semantic_orientation(word): 
        pmi, word_count = build_pmi(df_clean[”tokens”]) 
   if word in word_count then 
       return value calculated by Formula 2 
   end    

Emotion detection

Emotion recognition is frequently considered a subtask of polarity detection (Cambria, 2016), and it became the next step naturally after the basic sentiment task (Ashraf et al., 2022). A user puts into the tweet not just a positive or negative meaning but often a range of emotions. Other users can easily recognize and experience these emotions in texts. Emotion investigation can help us better understand opinions rooted in tweets. Detection of emotions behind the tweets was done using text2emotion 0.0.5 (Python Software Foundation, 2020). Five different emotion categories were used, such as Happy, Angry, Sad, Surprise and Fear. Text2emotions returns the dictionary with scores for each emotion category.

We could assume that the tweet belongs to a particular emotion if the score for that emotion is high. However, we don’t want to miss emotions of average and weak power. So we use a fuzzy approach (Zadeh, 1965) to manipulate emotions expressiveness.

A class of objects with a continuum of grades of membership is referred to as a fuzzy set (Zadeh, 1965; Zadeh, 1975). Several studies used a fuzzy approach to represent the strength or type of emotions and differentiate between different emotional labels (Matiko, Beeby & Tudor, 2014; Karyotis et al., 2016; Qamar & Ahmad, 2015).

The primary motivation for using fuzzification in our study is that it allows us to judge emotions in a human-consistent style. Since emotions don’t have clearly defined boundaries, fuzzy sets are very useful for dealing with these concepts. By partitioning the range of feasible emotions corresponding to linguistic tags, we get more intuitive and human-consistent output (Shamoi et al., 2016). Likewise, fuzzy sets can easily depict the gradual change from one emotion label to another can (Zadeh, 1975). A language value is more similar to human cognitive processes than a number, despite being less precise (Shamoi & Inoue, 2012).

We represent Emotion power as an ordered list of terms of the fuzzy (linguistic) variable X = “Emotion power” by means of primary linguistic terms L = {Weak, Medium, Strong}, specifying its level. Figure 13 below depicts the fuzzy sets for the Emotion power. For the sake of simplicity, we use triangular membership functions (Eq. (6)) representing the linguistic labels expressing the fuzzy variable Emotion power for each of the emotion categories. It is good enough to catch the ambiguity of the linguistic assessments. Such fuzzy partition was made based on subjective perception —we just ran a few experiments and adjusted the fuzzy sets. In the future, we can improve the partition by conducting a survey based on human emotion categorization (Shamoi, Inoue & Kawanaka, 2014).

where a, b, c represent the x coordinates of the three vertices of μ_a(x) in a fuzzy set A. Although we have five types of emotions, not all of them are present in a tweet. We take into account only those that obtained a non-zero score. For example, some specific tweet may contain weak sadness and strong surprise emotions only.

In the case of composite emotion impressions, formed from atomic ones with the help of various connectives (and, or) (Shamoi & Inoue, 2012) and hedges (e.g., not, very, more-or-less), general formulas from fuzzy theory, like α-cut, reinforcing and weakening modifiers, set operations (Shamoi, Inoue & Kawanaka, 2020) are used. α-cut is a crisp set that includes all the members of the given fuzzy subset whose values are not less than α for 0 < α ≤ 1. The α-cut of A is defined as Zadeh (1965): (7)${\displaystyle {\left[A\right]}_{\alpha }=\left\{x\in X|A\left(x\right)⩾\alpha \right\}.}$

These basic formulas allow us to retrieve some interesting tweets and judge emotions in human-consistent style, like “extremely happy and surprised”, “very very angry but not sad”, etc.

As we see, emotion analysis can help to detect more complex feelings embedded in a tweet. For example, ‘surprise’ is an emotion that can express both positive and negative sentiments (Sailunaz & Alhajj, 2019). The negative form of surprise is when you eat well, yet the tests reveal nutrient deficiency. On the other hand, the positive emotion of surprise is when you are surprised by an extensive choice of vegan options in a restaurant.

Let us take, for example, a tweet “Sorry Vegans, steak taste good.”. Its sentiment score is Neutral (−0.043) due to its balance of positive and negative words. At the same time, emotion analysis results are ‘Happy’: ‘Medium’, ‘Sad’: ‘Weak’, ‘Fear’: ‘Weak’. So, detecting emotions can be particularly useful for balanced tweets that standard SA can neutralize.

Python code and all supplemented data for emotion detection is publicly available.⁴

Dataset annotation on an 11-point scale

We initiated the annotation to evaluate the method because there are no available vegan-specific datasets labeled by humans or by a classifier trained in the vegan context. The trial dataset consisted of 100 tweets and was created using random sampling from our vegan-related unlabeled dataset described before.

Five humans performed the annotation process. There were three vegans and two omnivores among them. Data annotation should be performed by someone with a thorough understanding of the topic, so local vegans with 15, 5, and 2 years of vegan experience were asked to contribute. Vegans were given double weight since they are more familiar with the lexicon of the context. The human raters used separate Google sheets they got online to label the sentiment polarity of the tweets. A screenshot of the form (Google Sheets) is provided in Fig. 14.

Each tweet in a dataset was assigned a category by all five annotators. A test dataset was annotated using an 11-point scale annotation method described in Ghosh et al. (2015). Each annotator was instructed to assign a score to an opinion in a tweet based on a perceived meaning between −5 (representing strong displeasure) and +5 (for strongly positive tweets). Details can be seen in Fig. 14.

We used tactics given in Ghosh et al. (2015)—If 60% or more of annotator labels are assessed to be outliers, then the annotator judgments are dismissed from the task. We use this formula Eq. (8) to identify if a judgement A_i,j is an outlier: (Ghosh et al., 2015): (8)${\displaystyle |A_{i,j}-avg\left(A_{i^{\prime },j}\right)|>st{d}_t\left(t_j\right),}$ where std_t(t_j) is the standard deviation of all scores given for a tweet t_j.

After the weighted average calculation, we revealed one annotator whose proportion of outliers was 71%, so his annotations were excluded from the analysis, and the weighted average for each tweet was computed again.

Eventually, based on weighted average scoring, each tweet in a trial dataset obtained either a positive, neutral, or negative score (see Table 3). The annotation procedure resulted in the following sentiment distribution: positive sentiments (57%), neutral sentiments (0%), and negative sentiments (43%), with classification thresholds set at −0.1 and +0.1.

Sentiment detection results

Some tweets lead to discussions involving a lot of users. Popular tweets are tweets with relatively high responses in the form of retweets, comments, and likes (Zubiaga et al., 2014). For sentiment analysis, we selected popular tweets from our vegan dataset for each year from 2010 to 2022. By popular tweets, we mean those tweets that were retweeted at least once, indicating that the opinion expressed in them was supported by other users.

Obtained sentiment analysis results are presented in Fig. 15 and Table 4. Results indicate that positive sentiment in popular tweets consistently dominates over the years. It can also be seen that all sentiments started to grow, particularly during those years when the associated studies on the health benefits of a plant-based diet appeared. As for 2022, we see a fall in the number of popular tweets, which is associated with the fact that our dataset included only 4.5 months for 2022. However, it is evident that in 2022 the proportion of tweets with negative sentiment relatively decreased compared to previous years and is nearly equal to the proportion of neutral tweets. This could be interpreted in several ways, the most common of which is the impact of COVID-19 and global warming. Plant-based foods are becoming more common in post-pandemic diets. The investigation of the ‘Popular Tweets’ column of Table 4 reveals that levels of interest in plant-based diets have increased from 2010 to 2022. Note that both Fig. 15 and Table 4 focus only on tweets that have been retweeted at least once (RT ≥ 1).

Table 5 shows examples of some positive, neutral, and negative tweets from our dataset along with their computed scores using Eq. (5).

Emotion detection results

Sentiment analysis helped us to identify that positive sentiments were prevailing. However, what are the leading emotions behind this growing interest, and how were they changing throughout the period?

The graphs in Fig. 16 show the comparisons of emotions across years that indicate some interesting facts. The release of a lot of popular documentary movies revealing violence towards animals in agriculture and the health benefits of plant-based diets can explain the spike in all emotions in 2018 (What the Health, 2017 (7.2/10 IMDb); Game Changers, 2018 (7.8/10 IMDb); Dominion, 2018 (9/10 IMDb); Eating You Alive, 2018 (7.9/10 IMDb); 73 cows, 2018 (7.8/10 IMDb); Let Us Be Heroes, 2018 (8.7/10 IMDb)). Moreover, in 2018, UN IPCC reported that climate change catastrophe would occur in the upcoming 12 years (Masson-Delmotte et al., 2019).

The growth of Happy emotion, which we can clearly observe from 2018, is possibly connected with the fact that more and more vegan options, products, and places are starting to open, and people feel joyful about their health, lifestyle, and food. The ‘surprise’ emotion noticeably increased from 2018 as well, which can be connected to the fact that documentaries about the benefits of the vegan diet and the truth about animal abuse premiered this year (What the Health, 2017 (7.2/10 IMDb); Game Changers, 2018 (7.8/10 IMDb); Dominion, 2018 (9/10 IMDb); Eating You Alive, 2018 (7.9/10 IMDb); 73 cows, 2018 (7.8/10 IMDb); Let Us Be Heroes, 2018(8.7/10 IMDb)).⁵ This fact can also be considered as the reason for the rise of the Sad emotion because most people feel heartbroken, frustrated, and sad about realizing the heavy burden that animals inhabit for just one human’s meal. Furthermore, on the graph, we can recognize that emotions of Fear were mostly medium and strong throughout the period, which contrasts sharply with other types of emotions. This can be explained by the fact that many people tend to believe that a vegetarian or vegan diet can lead to nutritional deficiency (Paslakis et al., 2020).

Figure 17 demonstrates how individuals perceived veganism over time. We investigated that the public’s interest in plant-based diets has been growing noticeably since 2010. In 2017-2022, compared to 2011-2016, it can be noticed that the number of weak happy and weak angry tweets doubled. At the same time, there were no significant changes in the amounts of medium happy, angry and strong happy tweets. But it should be noted that there were fewer strong angry tweets by 2017–2022. From this, it can be assumed that by 2017–2022, people’s attitude to plant-based lifestyles was changing for the better.

Table 6 illustrates examples of emotions identified in tweets from our dataset.

We also performed a basic analysis based on the usage frequency of words in a vegan context across 12 years. The bar chart (see Fig. 18) illustrates the most frequently used terms (excluding keywords we used for our dataset). It can be clearly seen that when people discuss veganism, they usually don’t just talk about food (meat, products, recipe, organic, delicious), but also touch on the topics of animals, health, and beauty (perhaps this is related to the topic of testing products on animals).

Our findings support the results of the effects of sentiment on information diffusion, obtained by Ferrara & Yang (2015). According to their findings, positive posts reach a wider audience, implying that people are more likely to retweet positive content, a phenomenon known as positive bias. Discussions with more than 200 tweets are considered active ones (Ferrara & Yang, 2015). We filtered out only active, extremely popular tweets having at least 200 retweets (RT ≥ 200). Next, we fit the data within a polynomial function as depicted in Fig. 19. It illustrates the number of retweets collected by the tweets as a function of the sentiment score expressed therein. As seen from the figure, among such viral tweets, we have primarily neutral and slightly positive tweets, but not extremely positive ones. Nevertheless, the average number of retweets is higher for negative tweets (0.706, 0.503, and 0.369 for negative, positive, and neutral tweets, respectively).

Accuracy evaluation

We use popular sentiment analysis models VADER (Honnibal & Montani, 2017) and spaCy (Honnibal & Montani, 2017) and the human-annotated dataset discussed earlier to evaluate the performance of our approach.

Table 7 presents the Mutual information scoring function against VADER and spaCy scorings along with the sample tweets from the labeled dataset.

Table 8 shows the three-class classification accuracy for each method. From Table 8 we can see that our classifier (0.77) performs better than VADER (0.69) and spaCy (0.67) for the selected context, but still not as well as individual human rater with special context knowledge (0.87). Unsurprisingly, our scoring method obtained much higher accuracy (0.77) than the accuracy on the training data set labeled automatically without accounting for the vegan-tweets-based lexicon. Although we know that VADER is sensitive to social media lexicon, its accuracy decreases since we use it in domain-specific vegan tweets. The dataset we used for training has narrow variability and context-specific vocabulary, so a PMI-based classifier performs better.

Next, let us look at how well different classifiers agree on the sentiments of the tweets. We calculated the intersections of each method’s positive or negative proportions to compute the agreement. The Agreement Score represents the fraction of tweets both methods agree (have the same predicted polarity) (Gonalves et al., 2013). Table 9 shows the percentage of agreement between methods (phenomenon of methods agreement is discussed in Gonalves et al. (2013)).

From Table 9 we can see that our classifier has a much higher agreement with VADER than with spaCy (63% and 55% respectively). Pearson’s rank correlation between our method and VADER came out to be 0.565, which indicates that they are sensibly correlated.