1 Introduction

Users of social media are presented with a choice: post nothing at all; post something original; or re-post (“retweet” in the case of Twitter) an existing post. The simple amplifying mechanism of reposting encodes a unique digital and behavioral aspect of social contagion, with increasingly important ramifications as interactions and conversations on social media platforms such as Twitter tend to mirror the dynamics of major global and local events [1–4].

Previous studies have explored the role of retweeting in the social contagion literature, though the vast majority of this research is limited to either a given language (e.g., English tweets) or a short period [1, 2, 5, 6]. Here, drawing on a 10% random sample from over a decade’s worth of tweets, we track the rate of originally authored messages, retweets, and social amplification for over 100 languages.

We describe distinct usage patterns of retweets for certain populations. For example, Thai, Korean, and Hindi have the highest contagion ratios, while Japanese, Russian, Swedish, and Finish lie at the other end of the spectrum. While there is a wide range of motives and practices associated with retweeting, our object of study is the simple differentiation of observed behavior between the act of replication of anything and the act of de novo generation (i.e., between retweeted and what we will call organic messages).

We acknowledge two important limitations from the start. First, while it may be tempting to naively view ideas spreading as infectious diseases, the analogy falls well short of capturing the full gamut of social contagion mechanisms [7–16], and a full understanding of social contagion remains to be established. And second, while higher contagion ratios are in part due to active social amplification by users, they may also, for example, reflect changes in Twitter’s design of the retweet feature, changes in demographics, or changes in a population’s general familiarity with social media. Future work will shed light on the psychological and behavioral drivers for the use of retweets in each language across geographical and societal markers, including countries and communities.

1.1 Background and motivation

Social contagion has been extensively studied across many disciplines including marketing [17–20], finance [21–24], sociology [25–27], and medicine [28–30]. Because it can be easier to access data on human social behavior from social media outlets than from other sources such as in-person or text-message conversations, social contagion dynamics are often examined in the context of messages posted and subsequently re-posted on social media platforms [31–34]. Indeed, the flow of information in the context of social contagion on digital media outlets, especially Twitter, has been widely studied over the last decade [6, 35], with attention paid to the spreading of certain kinds of messages, such as rumours [36–40], misinformation and “fake news” [41–44]. Several models have also been proposed to predict the spread of information on Twitter [45], while other models have shown the differences in which various topics can propagate throughout social networks [46, 47]. Studies have also investigated the extent to which information spread on Twitter can have an echo chamber effect [48–50].

The body of research shows overwhelming evidence that retweeting is a key instrument of social contagion on Twitter [3, 51]. One of the earliest analysis of Twitter by Kwak et al. [52] suggests that a retweet can reach an average of a thousand users regardless of the social network of its original author, spreading its content instantly across different hubs of the full Twitter social network. While seemingly simple, there are different styles and drivers of retweeting [2]. The practice of retweeting has become a convention on Twitter to spread information, especially for celebrities. Researchers argue celebrities can act as hubs of social contagion by studying the flow of retweets across their focal networks [5]. Recent work shows how retweets of officials can be either alarming or reassuring amid the COVID–19 pandemic [53, 54]. Statistical features of retweets reveal a strong association between links and hashtags in most retweeted messages [1]. Retweeting is not only an act in which users can spread information, but a mechanism for actors to become involved in a conversation without being active participants [2]. The use of retweets empirically alters the visibility of information and how fast messages can spread on the platform [4].

Other studies have quantified language usage on social media [55, 56], particularly on Twitter [57, 58]. While investigators have studied the use of retweets in the context of social contagion using network-based approaches [35, 46, 54, 59], little research has been done regarding the statistical variability of retweets across the vast majority of languages. In this paper, by applying an updated language identification (LID) process to over a decade of Twitter messages, we explore a macroscopic description of social contagion through the use of retweets across languages on Twitter. Our study addresses a unique property of social contagion on Twitter by statistically quantifying the rate of retweets in each language. We show how the practice of retweeting varies across different languages and how retweeting naturally lends itself to micro-level discussions of social contagion on Twitter, which can also be extended to other social media outlets with similar features.

1.2 Overview

We structure our paper as follows. First, we discuss the state-of-the-art tools presently used for language detection of short and informal messages (e.g., tweets). We then describe our dataset and processing pipeline to answer some key questions regarding social contagion through the use of retweets. Based on our considerations, we deploy FastText-LID [60] to identify and explore the evolution of 100+ languages in over 118 billion messages collected via Twitter’s 10% random sample (decahose) from 2009 to 2020 [61].

For messages posted after 2013, we also analyze language labels provided by Twitter’s proprietary LID algorithm and justify using FastText-LID as an alternative LID tool to overcome the challenge of missing language labels in the historical feed from Twitter (see also Hong et al. [62]).

We study the empirical dynamics of replication: The rate at which users choose to retweet instead of generating original content; and how that rate varies across languages temporally. We quantify the ratio of retweets to new messages (contagion ratio) in each language. In most common languages on Twitter, we show that this ratio reveals a growing tendency to retweet.

Finally, we present a detailed comparison with the historical data feed in Appendix A. We conclude with an analytical validation of our contagion ratios (Appendix B), and the impact of tweet-length on language detection (Appendix C). We also provide an online appendix at: http://compstorylab.org/storywrangler/papers/tlid/.

2 Tweet language identification

Twitter is a well-structured streaming source of sociotechnical data, allowing for the study of dynamical linguistics and cultural phenomena [63, 64]. Of course, like many other social platforms, Twitter represents only a subsample of the publicly declared views, utterances, and interactions of millions of individuals, organizations, and automated accounts (e.g., social bots) around the world [65–68]. Researchers have nevertheless shown that Twitter’s collective conversation mirrors the dynamics of local and global events [69] including earthquakes [70], flu and influenza [71, 72], crowdsourcing and disaster relief [73, 74], major political affairs [75], and fame dynamics for political figures and celebrities [76]. Moreover, analyses of social media data and digital text corpora over the last decade have advanced natural language processing (NLP) research [77–79] such as language detection [80–83], sentiment analysis [84–88], word embeddings [89–92], and machine translation [93–95].

LID is often referred to as a solved problem in NLP research [96–100], especially for properly formatted documents, such as books, newspapers, and other long-form digital texts. Language detection for tweets, however, is a challenging task due to the nature of the platform. Every day, millions of text snippets are posted to Twitter and written in many languages along with misspellings, catchphrases, memes, hashtags, and emojis, as well as images, gifs, and videos. Encoding many cultural phenomena semantically, these features contribute to the unique aspects of language usage on Twitter that are distinct from studies of language on longer, edited corpora [101].

A key challenge of LID on Twitter data is the absence of a large, public, annotated corpus of tweets covering most languages for training and evaluation of LID algorithms. Although researchers have compiled a handful of manually labeled datasets of Twitter messages, the proposed datasets were notably small compared to the size of daily messages on Twitter and limited to a few common languages [81–83]. They showed, however, that most off-the-shelf LID methods perform relatively well when tested on annotated tweets.

As of early 2013, Twitter introduced language predictions classified by their internal algorithm in the historical data feed [102]. Since the LID algorithm used by Twitter is proprietary, we can only refer to a simple evaluation of their own model.Footnote 1 Our analysis of Twitter’s language labels indicates Twitter appears to have tested several language detection methods, or perhaps different parameters, between 2013 and 2016.

Given access to additional information about the author of a tweet, the LID task would conceivably be much more accurate. For example, if the training data for prediction included any or all of the self-reported locations found in a user’s ‘bio’, the GPS coordinates of their most recent tweet, the language they prefer to read messages in, the language associated with individuals they follow or who follow them, and their collective tweet history, we expect the predictions would improve considerably. However, for the present investigation, we assume the only available predictors are found in the message itself. Our goal is to use the state-of-the-art language detection tools to get consistent language labels for messages in our data set to enable us to investigate broader questions about linguistic dynamics and the growth of retweets on the platform over time.

2.1 Open-source tools for LID

Several studies have looked closely at language identification and detection for short-text [103–110], particularly on Twitter where users are limited to a few characters per tweet (140 prior to the last few months of 2017, 280 thereafter [111]). Researchers have outlined common challenges specific to this platform [112, 113].

Most studies share a strong consensus that language identification of tweets is an exceptionally difficult task for several reasons. First, language classification models are usually trained over formal and large corpora, while most messages shared on Twitter are informal and composed of 140 characters or fewer [81, 82] (see Appendix C for more details). Second, the informal nature of the content is also a function of linguistic and cultural norms; some languages are used differently over social media compared to the way they are normally used in books and formal documents. Third, users are not forced to choose a single language for each message; indeed messages are often posted with words from several languages found in a single tweet. Therefore, the combination of short, informal, and multilingual posts on Twitter makes language detection much more difficult compared to LID of formal documents [114]. Finally, the lack of large collections of verified ground-truth across most languages is challenging for data scientists seeking to fine-tune language detection models using Twitter data [81, 115, 116].

Researchers have evaluated off-the-shelf LID tools on substantial subsets of Twitter data for a limited number of languages [81, 82, 116]. For example, Google’s Compact Language Detector (versions CLD-1Footnote 2 and CLD-2Footnote 3) offer open-source implementations of the default LID tool in the Chrome browser to detect language used on web pages using a naive Bayes classifier. In 2012, Lui and Baldwin [80] proposed a model called langid that uses an n-gram-based multinomial naive Bayes classifier. They evaluated langid and showed that it outperforms Google’s CLD on multiple datasets. A majority-vote ensemble of LID models is also proposed by Lui et al. [82] that combines both Google’s CLD and langid to improve classification accuracy for Twitter data.

Although using a majority-vote ensemble of LID models may be the best option to maximize accuracy, there are a few critical trade-offs including speed and uncertainty. The first challenge of using an ensemble is weighing the votes of different models. One can propose treating all models equally and taking the majority vote. This becomes evidently complicated in case of a tie, or when models are completely unclear on a given tweet. Treating all models equally is an arguably flawed assumption given that not all models will have the same confidence in each prediction—if any is reported. Unfortunately, most LID models either decline to report a confidence score, or lack a clear and consistent way of measuring their confidence. Finally, running multiple LID classifiers on every tweet is computationally expensive and time-consuming.

Recent advances in word embeddings powered by deep learning demonstrate some of the greatest breakthroughs across many NLP tasks including LID. Unlike previous methodologies, Devlin et al. [90] introduces a new language representation model called BERT. An additional output layer can be added to the pre-trained model to harvest the power of the distributed language representations, which enables the model to carry out various NLP tasks such as LID.

FastText [60] is a recently proposed approach for text classification that uses n-gram features similar to the model described by Mikolov et al. [117]. FastText employs various tricks [91, 92, 118] in order to train a simple neural network using stochastic gradient descent and a linearly decaying learning rate for text classification. While FastText is a language model that can be used for various text mining tasks, it requires an additional step of producing vector language representations to be used for LID. To accomplish that, we use an off-the-shelf language identification tool [119] that uses the word embeddings produced by the model. The proposed tool uses a hierarchical softmax function [60, 117] to efficiently compute the probability distribution over the predefined classes (i.e., languages). For convenience, we will refer to the off-the-shelf LID tool [119] as FastText-LID throughout the paper. The authors show that FastText-LID is on par with deep learning models [120, 121] in terms of accuracy and consistency, yet orders of magnitude faster in terms of inference and training time [60]. They also show that FastText-LID outperforms previously introduced LID tools such as langid.Footnote 4

2.2 Processing pipeline

While there are many tools to consider for LID, it is important for us to ensure that the language classification process stays rather consistent to investigate our key question about the growth of retweets over time. In light of the technical challenges discussed in the previous section, we have confined this work to using FastText-LID [119] due to its consistent and reliable performance in terms of inference time and accuracy.

To avoid biasing our language classification process, we filter out Twitter-specific content prior to passing tweets through the FastText-LID model. This is a simple strategy originally proposed by Tromp et al. [103] to improve language classification [82, 122]. Specifically, we remove the prefix associated with retweets (“RT”), links (e.g., “https://twitter.com”), hashtags (e.g., “#newyear”), handles (e.g., “@username”), html codes (e.g., “&gt”), emojis, and any redundant whitespaces.

Once we filter out all Twitter-specific content, we feed the remaining text through the FastText-LID neural network and select the predicted language with the highest confidence score as our ground-truth language label. If the confidence score of a given prediction is less than 25%, we label that tweet as Undefined (und). Similarly, if no language classification is made by the Twitter-LID model, Twitter flags the language of the message as undefined [123, 124]. We provide a list of all language labels assigned by FastText-LID compared to the ones served by Twitter-LID in Table 1.

Table 1 Language codes for both FastText-LID and Twitter-LID tools
Full size table

We subsequently extract day-scale time series and Zipf distributions for uni-, bi-, and tri-grams and make them available through an analytical instrument entitled Storywrangler. Our tool is publicly available online at: https://storywrangling.org/. See Alshaabi et al. [125] for technical details about our project.

3 Results and discussion

3.1 Temporal and empirical statistics

We have collected a random 10% sample of all public tweets posted on the Twitter platform starting January 1, 2009. Using the steps described in Sect. 2.2, we have implemented a simple pipeline to preprocess messages and obtain language labels using FastText-LID [119]. Our source code along with our documentation is publicly available online on a Gitlab repository.Footnote 5 Here, we evaluate our results by comparing the language labels obtained by FastText-LID to those found in the metadata provided by Twitter’s internal LID algorithm(s). Our initial analysis of the Decahose metadata indicated missing language labels until 2013, when Twitter began offering a language prediction (we offer an approach to detecting corrupted time series in Dodds et al. [126]).

We find that our classification of tweets using FastText-LID notably improves the consistency of language labels when compared to the labels served with the historical streaming feed. In Fig. 1A, we display a weekly rolling average of the daily number of languages detected by each classifier over time. We see that Twitter’s language detection has evolved over time. The number of languages stabilized but continued to fluctuate in a manner that is not consistent, with uncommon languages having zero observations on some given days. By contrast, the FastText-LID time series of the number of languages shows some fluctuations in the earlier years—likely the result of the smaller and less diverse user base in the late 2000s—but stabilizes before Twitter introduces language labels. We note that the fluctuations in the time series during the early years of Twitter (before 2012) and the first week of 2017 are primarily caused by unexpected service outages which resulted in missing data.

Figure 1
figure 1

Language time series for the Twitter historical feed and FastText-LID classified tweets. (A) Number of languages reported by Twitter-LID (red) and classified by FastText-LID (black) since September 2008. Fluctuations in late 2012 and early 2013 for the Twitter language time series are indicative of inconsistent classifications. (B) Rate of usage by language using FastText-LID maintains consistent behavior throughout that period. The change in language distribution when Twitter was relatively immature can be readily seen—for instance, English accounted for an exceedingly high proportion of activity on the platform in 2009, owing to Twitter’s inception in an English-speaking region

Full size image

FastText-LID classifies up to 173 languages, some of which are rare, thus the occasional dropout of a language seen in this time series is expected. On the other hand, Twitter-LID captures up to 73 languages, some of which are experimental and no longer available in recent years. Nonetheless, Fig. 1B shows that the overall rate of usage by language is not impaired by the missing data, and maintained consistent behavior throughout the last decade.

We compute annual confusion matrices to examine the language labels classified by FastText-LID compared to those found in the historical data feed. Upon inspection of the computed confusion matrices, we find disagreement during the first few years of Twitter’s introduction of the LID feature to the platform. As anticipated, the predicted language for the majority of tweets harmonizes across both classifiers for recent years (see Fig. 7). We notice some disagreement between the two classifiers on expected edge-cases such as Italian, Spanish, and Portuguese where the lexical similarity among these languages is very high [127–130]. Overall, our examination of average language usage over time demonstrates that FastText-LID is on par with Twitter’s estimation. We show the corresponding Zipf distribution of language usage for each classifier, and highlight the normalized ratio difference between them for the most used languages on the platform in Figs. 8–9. We point the reader’s attention to Appendix A for further details of our comparison.

Furthermore, we display a heatmap of the number of messages for each language as classified by FastText-LID in our data set (see Fig. 2). We have over 118 billion messages between 2009-01-01 and 2019-12-31 spanning 173 languages. English is the most used language on the platform with a little under 42 billion messages throughout the last decade. Although the number of Japanese speakers is much smaller than the number of English speakers around the globe, Japanese has approximately 21 billion messages. Spanish—the third most prominent language on Twitter—is shy of 11 billion messages. Arabic and Portuguese rank next with about 7 billion messages for each. We note that the top 10 languages comprise 85% of the daily volume of messages posted on the platform.

Figure 2
figure 2

Overall dataset statistics. Number of messages captured in our dataset as classified by the FastText-LID algorithm between 2009-01-01 and 2019-12-31, which sums up to approximately 118 billion messages throughout that period (languages are sorted by popularity). This collection represents roughly 10% of all messages ever posted

Full size image

In Fig. 3, we show the flow of annual rank dynamics of the 15 most used languages on Twitter between 2009 and 2020. For ease of description, we will refer to Undefined as a language class. The top 5 most common languages on Twitter (English, Japanese, Spanish, Undefined, and Portuguese) are consistent, indicating a steady rate of usage of these languages on the platform. The language rankings correspond with worldwide events such as the Arab Spring [131–134], K-pop, and political events [76]. “Undefined” is especially interesting as it covers a wide range of content such as emojis, memes, and other media shared on Twitter but can’t necessarily be associated with a given language. Russian, however, starts to grow on the platform after 2011 until it peaks with a rank of 7 in 2015, then drops down to rank 15 as of the end of 2019. Other languages such as German, Indonesian, and Dutch show a similar trend down in ranking. This shift is not necessarily caused by a drop in the rate of usage of these languages, but it is rather an artifact prompted by the growth of other languages on Twitter.

Figure 3
figure 3

Annual average rank of the most used languages on Twitter between 2009 and 2019. English and Japanese show the most consistent rank time series. Spanish, and Portuguese are also relatively stable over time. Undefined—which covers a wide variety of content such as emojis, links, pictures, and other media—also has a consistent rank time series. The rise of languages on the platform correlates strongly with international events including Arab Spring and K-pop, as evident in both the Arabic and Korean time series, respectively. Russian, German, Indonesian, and Dutch moved down in rank. This shift is not necessarily due to a dramatic drop in the rate of usage of these languages, but is likely an artifact of increasing growth of other languages on Twitter such as Thai, Turkish, Arabic, Korean, etc

Full size image

3.2 Quantifying Twitter’s social contagion: separating organic and retweeted messages

We take a closer look at the flow of information among different languages on the platform, specifically the use of the “retweet” feature as a way of spreading information. Encoding a behavioral feature initially invented by users, Twitter formalized the retweet feature in November 2009 [135]. Changes in platform design and the increasing popularity of mobile apps promoted the RT as a mechanism for spreading. In April 2015, Twitter introduced the ability to comment on a retweet message or “Quote Tweet”(QT) [136] a message, distinct from a message reply [137].

To quantify the rate of usage of each language with respect to these different means by which people communicate on the platform, we categorize messages on Twitter into two types: “Organic Tweets” (OT), and “Retweets” (RT). The former category (OT) encompasses original messages that are explicitly authored by users, while the latter category (RT) captures messages that are shared (i.e. amplified) by users. We break each quote tweet into two separate messages: a comment and a retweet. We exclude retweets while including all added text (comments) found in quote tweets for the OT category.

For each day t and for each language ℓ, we calculate the raw frequency (count) of organic messages \(f^{(\mathrm{OT})}_{\ell ,t}\), and retweets \(f^{(\mathrm{RT})}_{\ell ,t}\). We further determine the frequency of all tweets (AT) such that: \(f^{(\mathrm{AT})}_{\ell ,t} = f^{(\mathrm{OT})}_{\ell ,t} + f^{ (\mathrm{RT})}_{\ell ,t}\). The corresponding rate of usages (normalized frequencies) for these two categories are then:

$$ p_{t,\ell }^{(\mathrm{OT})} = \frac{f_{t,\ell }^{(\mathrm{OT})}}{f_{t,\ell }^{(\mathrm{AT})}}, \quad \text{and} \quad p_{t,\ell }^{(\mathrm{RT})} = \frac{f_{t,\ell }^{(\mathrm{RT})}}{f_{t,\ell }^{(\mathrm{AT})}}. $$

3.3 Measuring social and linguistic wildfire through the growth of retweets

To further investigate the growth of retweets, we use the ratio of retweeted messages to organic messages as an intuitive and interpretable analytical measure to track this social amplification phenomenon. We define the ‘contagion ratio’ as:

$$ R_{\ell ,t} = f^{(\mathrm{RT})}_{\ell ,t} / f^{ (\mathrm{OT})}_{\ell ,t}. $$

In 2018, the contagion ratio exceeded 1, indicating a higher number of retweeted messages than organic messages (Fig. 4). The overall count for organic messages peaked in the last quarter of 2013, after which it declined slowly as the number of retweeted messages climbed to approximately 1.2 retweeted messages for every organic message at the end of 2019. Thereafter, the contagion ratio declined through 2020 with the exception of a surge of retweets in the summer amid the nationwide protests sparked by the murder of George Floyd.Footnote 6

Figure 4
figure 4

Timeseries for organic messages, retweeted messages, and average contagion ratio for all languages. (A) Monthly average rate of usage of organic messages (\(p_{t, \ell }^{(\mathrm{OT})}\), blue), and retweeted messages (\(p_{t,\ell }^{( \mathrm{RT})}\), orange). The solid red line highlights the steady rise of the contagion ratio \(R_{\ell ,t}\). (B) Frequency of organic messages (\(f^{(\mathrm{OT})}_{\ell ,t}\), blue), compared to retweeted messages (\(f^{(\mathrm{RT})}_{\ell ,t}\), orange). The areas shaded in light grey starting in early 2018 highlights an interesting shift on the platform where the number of retweeted messages has exceeded the number of organic messages. An interactive version of the figure for all languages is available in an online appendix: http://compstorylab.org/storywrangler/papers/tlid/files/ratio_timeseries.html

Full size image

In 2020, Twitter’s developers redesigned their retweet mechanism, purposefully prompting users to write their own commentary using the Quote Tweet [138], along with several new policies to counter synthetic and manipulated media [139–141]. While the long upward trend of the contagion ratio is in part due to increasingly active social amplification by users, the recent trend demonstrates how social amplification on Twitter is highly susceptible to systematic changes in the platform design. Twitter has also introduced several features throughout the last decade, such as tweet ranking, and extended tweet length that may intrinsically influence how users receive and share information in their social networks.Footnote 7 We investigate the robustness of our findings regarding contagion ratios in light of some of these changes in Appendix B and Appendix C. Future work will shed light on various aspects of social amplification on Twitter with respect to the evolution of the platform design, and behavioral drivers for the use of retweets in each language across communities.

Finally, we show weekly aggregation of the rate of usage \(p_{t,\ell }\) for the top 30 ranked languages of 2019 in Fig. 5. The time series demonstrate a recent sociolinguistic shift: Several languages including English, Spanish, Thai, Korean, and French have transitioned to having a higher rate of retweeted messages \(p_{t,\ell }^{(\mathrm{RT})}\) than organic messages \(p_{t,\ell }^{(\mathrm{OT})}\). Thai appears to be the first language to have made this transition in late 2013. In Fig. 6, we show a heatmap of the average contagion ratio \(R_{\ell ,t}\) for the top 30 most used languages on Twitter per year. With the exception of Indonesian, which showed a small bump between 2010 and 2013, most other languages began adopting a higher ratio of retweeted content in 2014. Thai has the highest number of retweeted messages, with an average of 7 retweeted messages for every organic message. Other languages, for example, Hindi, Korean, Urdu, Catalan, and Tamil average between 2 to 4 retweeted messages for every organic message. On the other hand, Japanese—the second most used language on the platform—does not exhibit this trend. Similarly, German, Italian, and Russian maintain higher rates of organic tweets. The trend of increasing preference for retweeted messages, though not universal, is evident among most languages on Twitter. We highlight the top 10 languages with the highest and lowest average contagion ratio per year in Table 2 and Table 3, respectively.

Figure 5
figure 5

Weekly rate of usage of the top 30 languages (sorted by popularity). For each language, we show a weekly average rate of usage for organic messages (\(p_{t,\ell }^{(\mathrm{OT})}\), blue) compared to retweeted messages (\(p_{t,\ell }^{(\mathrm{RT})}\), orange). The areas highlighted in light shades of gray represent weeks where the rate of retweeted messages is higher than the rate of organic messages. An interactive version featuring all languages is available in an online appendix: http://compstorylab.org/storywrangler/papers/tlid/files/retweets_timeseries.html

Full size image
Figure 6
figure 6

Timelapse of contagion ratios. The average ratio is plotted against year for the top 30 ranked languages of 2019. Colored cells indicate a ratio higher than 0.5 whereas ratios below 0.5 are colored in white. Table 2 shows the top 10 languages with the highest average contagion ratio per year, while Table 3 shows the bottom 10 languages with the lowest average contagion ratio per year

Full size image
Table 2 Top 10 languages with the highest annual average contagion ratio (sorted by 2019)
Full size table
Table 3 Bottom 10 languages with the lowest annual average contagion ratio (sorted by 2019)
Full size table

4 Concluding remarks

Understanding how stories spread through and persist within populations has always been central to understanding social phenomena. In a time when information can flow instantly and freely online, the study of social contagion has only become more important.

In the sphere of Twitter, the practice of retweeting is complicated from a social and psychological point of view. There is a diverse set of reasons for participants to retweet. For example, scientists and academics can use this elementary feature to share their findings and discoveries with their colleagues. Celebrities and political actors can benefit from other people retweeting their stories for self-promotion. Attackers can also take advantage of this natural feature of social contagion to pursue malicious intents, deploy social bots, and spread fake news.

In this paper, we have analyzed over a hundred billion messages posted on Twitter throughout the last decade. We presented an alternative approach for obtaining language labels using FastText-LID in order to overcome the challenge of missing labels in the Decahose dataset, obtaining consistent language labels for 100+ languages. We acknowledge that shortcomings of language detection for short and informal text (e.g., tweets) are well known in the NLP literature. Using FastText-LID is not necessarily the best approach for language identification. Our analysis may be subject to implicit measurement biases and errors introduced by word embeddings used to train the language detection tool using FastText [60]. We emphasize that we have not intended to reinvent or improve FastText-LID in this work; we have used FastText-LID only as a (well-established and tested) tool to enable the study of social contagion dynamics on Twitter. Nevertheless, we have presented some further analysis of FastText-LID compared to Twitter-LID in Appendix A. Future work will undoubtedly continue to improve language detection for short text, particularly for social media platforms.

Our results comparing language usage over time suggest a systematic shift on Twitter. We found a recent tendency among most languages to increasingly retweet (spread information) rather than generate new content. Understanding the general rise of retweeted messages requires further investigation. Possible partial causes might lie in changes in the design of the platform, increases in bot activity, a fundamental shift in human information processing as social media becomes more familiar to users, and changes in the demographics of users (e.g., younger users joining the platform).

The metrics we have used to compute our contagion ratios are simple but rather limited. We primarily focused on tracking the rate of organic tweets and retweets to quantify social amplification of messages on the platform. While our approach of measuring the statistical properties of contagion ratios is important, it falls short of capturing how retweets propagate throughout the social networks of users. Future work may deploy a network-based approach to investigate the flow of retweets among users and followers. Whether or not the information is differentially propagated across languages, social groups, economic strata, or geographical regions is an important question for future research, and beyond the scope of our present work.

Geolocation information for Twitter is also limited, and here we have only examined contagion ratios at the language level. Language, transcending borders as it does, can nevertheless be used differently across communities. For instance, characterizing the temporal dynamics of contagion ratios for English, which is used all around the globe, is very different from doing so for Thai—a language that is used within a geographically well-defined population. Different social and geographical communities have cultures of communication which will need to be explored in future work.

In particular, it is important to study the relationship between social contagion dynamics, geographical region, and language. It might be the case that contagion dynamics are more homogeneous across geographic regions even when each geographical region displays high language diversity, or vice versa. However, in order to conduct this line of research, it is necessary to have accurate geotagging of tweets, which is currently not the case except for a very small subsample [142]. Future research could focus on implementing accurate geotagging algorithms that assign tweets a probabilistic geographical location based on their text and user metadata, while fully respecting privacy through judicious use of masking algorithms.