Abstract
One of the principal challenges of religious studies is the lack of a widely accepted normative definition of the terms ‘religion’ and ‘spirituality.’ Responding to recent calls for methodological innovation in this field, in this paper we present a “bottom-up” approach to outlining these concepts based on an interdisciplinary, usage-based approach, which combines qualitative analysis with methods of distributional semantics. Based on the Firthian principle that a word may be known “by the company it keeps,” we used a convenience sample of 138,000 Tweets gathered between March and April 2021 to calculate seven clusters of terms related to ‘spirituality’ and ‘religion.’ Three authors then coded these clusters to propose a tentative name for each. An unsupervised sentiment analysis was conducted and presented on a concept map. We suggest that the Twitter (now ‘X’) data at hand echoes the popular perception of ‘spirituality’ as a positively connoted term, which contrasts with a negatively connoted ‘religion.’ However, the data may be interpreted to arrive at more nuanced and, at times, contradictory interpretations.
1 Introduction
The attempt to arrive at a widely accepted normative definition of the term ‘religion’ remains a cause of debate among sociologists, anthropologists, theologians, and other scholars of religion. The closely related term ‘spirituality,’ however, is relatively rarely subject to debate, and is, at times, treated as a function, subset, or synonym to religion. Despite attempts to arrive at a consensus (Puchalski et al., 2014), what is understood by ‘spirituality’ varies widely according to context (Steensland et al., 2021) and, in the field of medicine and clinical psychology, is often conflated with religion (“R/S”), complicating attempts by medical professionals to integrate spiritual care into healthcare provision (Balboni & Balboni 2018; cf. Koenig, 2012; Worthington et al., 2011). In recent years, scholars have begun to explore these terms though new empirical methods, such as smartphone-based experience sampling, field experiments, and ‘big data’ approaches (Wright, 2022). Usage-based methods, in particular, have been suggested as a possible Occam’s razor to the protracted challenge of conceptualising ‘spirituality’ (Neubert, 2016; Kim et al., 2020). Building on this literature, we explore ‘religion’ and ‘spirituality’ not through normative definitions, but through a data-driven investigation of their semantic differentiation in real-world discourse (Neubert, 2016; Hjelm, 2016; Taira, 2022).
Over the past two decades, such usage-based methods, which are premised on the notion that the meaning of concepts arises from their repeated use in everyday discourse, have become commonplace in linguistics (Bybee, 2007; Goldberg, 2006), and have begun to make inroads into the cognitive sciences (Lenci, 2008). Usage-based approaches are typically ‘unsupervised’ (Schwartz & Ungar, 2015), meaning they do not assume pre-defined conceptual categories but deduce categories from ‘raw’ data of word usage in everyday discourse. Drawing on the methods of distributional semantics, we posit that the usage of the term ‘spirituality’ and its relation to ‘religion’ may be usefully circumscribed by their proximity to other words in a given context. To use a well-known phrase coined by John R. Firth (1890–1960), “You shall know a word by the company it keeps” (Firth, 1957, p. 14).
In order to analyse the ‘real world’ usage of these terms, we apply this usage-based method to a large convenience sample of speech fragments gathered on Twitter (now ‘X’), which provides a previously unavailable repository of spontaneous speech directed to an imagined anonymous public sphere – rather than the consensus of editorial, intellectual, expert or other institutionalized gatekeepers which have shaped discourse in the past. Twitter/‘X’ is a well-established site for the investigation of digital religious life (cf. Burge, 2021; Veidlinger, 2021). While a usage-based approach to define the term ‘religion’ has long been theorised in the literature (Saler, 1993; Bergunder, 2014), empirical work remains rare. With recent changes to the Twitter/‘X’ application programming interface (api), future investigations have become highly restricted.
In the following, we employ a usage-based method to test the hypothesis that (1) in ‘real world’ usage, a clear and interpretable distinction between ‘spirituality’ and ‘religion’ is evident, and (2) that the observed differences reflect those made in the literature. In particular, we hypothesise that the differences will reflect the literature’s favourable view of ‘spirituality’ vis-à-vis the relatively negatively connoted ‘religion,’ as suggested by population surveys and a growing number of Europeans who identify as “spiritual but not religious” (Drescher, 2016; Pew Research Center, 2018).
We present the results of an unsupervised sentiment analysis on the data and propose a tentative name for each topic that is generated by the topic model. A concept map is then generated to visualise the statistical affinity between words most closely associated with ‘religion’ and ‘spirituality’ respectively. In the final section, we discuss the topic model and concept map, and offer some preliminary observations regarding our hypothesis.
2 Methodology
Following Firth (1957), subsequent scholars applied computational methods to describe meaning through contextual correlation. Central to this method has been the claim that “the proportion of words common to the contexts of word A and to the contexts of word B is a function of the degree to which A and B are similar in meaning” (Rubenstein & Goodenough 1965, p. 627; Landauer & Dumais, 1997; Schütze & Pedersen, 1995). The Firthian dictum has been used to detect habitual juxtapositions of terms, or “collocations” (Bartsch &Evert, 2014). As Sahlgren (2006) and others have shown, a Firthian approach can also be used to detect other cognitively associated words. Where observation windows are small, these are syntagmatic relations (i.e. words that tend to occur in the same phrase or fixed expression), whereas expanding the window to include additional context delivers results on the paradigmatic axis, such as synonyms, antonyms, associations and related words (i.e. they can be substituted). This phenomenon is exploited by distributional semantics to detect synonyms, but also semantic associations (Fitzmaurice et al., 2017; Baroni & Lenci, 2010). In our case, if ‘religion’ significantly differs from ‘spirituality’ in real-world usage, this method would produce a very wide range of contextual correlation. On the contrary, if ‘religion’ and ‘spirituality’ have similar or the same meaning, the proportion of shared word associations would largely overlap.
2.1 Data Collection and Preprocessing
Since the number of tweets that one can download and analyze on such a topic is enormous, we limited ourselves to English tweets created in 2021.Two authors (Schneider and Goldzycher) collected a convenience sample of tweets containing the keywords ‘religion,’ and ‘spirituality’ using the R library rtweet (Kearney, 2020). The collection period was between March 3, 2021 and April 4, 2021. We gathered a total of 138,000 tweets with the keywords ‘religion’ and ‘spirituality,’ totalling about 4.5 million words. As the internal language-detection algorithm of Twitter/‘X’ is unreliable, we identified English tweets using the Python library ‘langid’ (version 1.1.6). Using this step, we obtained 102,501 tweets with 3.2 million words, counting 58,376 occurrences of ‘religion’ and 47,811 occurrences of ‘spirituality.’ A total of 5,637 tweets contained both keywords.
Raw tweets contain many artifacts that affect automatic content analysis. Therefore, we first cleaned the dataset using the following procedure: lowercasing the text, removing line breaks, reducing multiple consecutive white space characters to a single white space, replacing emojis with textual descriptions, replacing url s with the stand-in word ‘embeddedurl’ and replacing ‘&’ (an artifact of erroneous character decoding) with ‘&.’ We also removed retweets and duplicates (tweets with the same identification number). Lastly, to improve legibility, url s were removed, and contractions were written out in cited example tweets. We used this data to generate a topic model and a concept map including sentiment detection.
2.2 Generating the Topic Model
A topic model was generated using Mallet (McCallum, 2002) with the Latent Dirichlet Allocation (lda) algorithm. The topic model optimizes the Bayesian probability, p(topic|document) * p(word|topic), for all given documents in a collection. It thus combines document classification (p(topic|document)) and keyword generation (p(word|topic)). Documents and words were generated, and conceptual topics were fitted iteratively starting from a random configuration. The only manually set parameter was the number of word clusters generated, with thematic saturation reached at 7 clusters.1 For setting the hyperparameters α and β, we rely on the recommended built-in hyperparameter optimization with the flag optimize-interval set to 10, as recommended by Mallet.2 Three authors, an anthropologist (Winiger), a church historian (Neuhold) and a theologian (Peng-Keller) then separately assessed the clusters generated by Mallet, proposed a code for each, and calculated Krippendorff’s interrater reliability estimate using spss (version 28) with the Kalpha macro (Hayes & Krippendorff, 2007). We used the resultant codes as tentative names for each word cluster.
We did not include sentiment analysis in our Topic Modelling analysis in order not to add a possibly artificial bias. Instead, we detected positive and negative topics (sections 3.1 and 3.2) and then added sentiment analysis to obtain a clearer picture.
2.3 Generating the Concept Map
For the concept map, we conducted a binary sentiment analysis to automatically annotate each tweet with a label indicating if the tweet expresses a positive or negative sentiment and to label each tweet as either “sentipositive” or “sentinegative.” Sentiment detection is well-established in natural language processing. We used a neural model by Hartmann et al. (2022) based on the transformer architecture (Vaswani et al., 2017). This is a version of the language model RoBERTa (Liu et al., 2019) that was additionally fine-tuned on 15 sentiment analysis datasets covering a wide range of text genres and domains. In an evaluation on 19 sentiment analysis datasets conducted by Hartmann et al. this model achieves the highest accuracy compared to other recent models.
We then used a further ‘unsupervised’ method of distributional semantics, namely textplot (McClure, 2015) to calculate Kernel Density Estimation (kde, Zucchini, 2003), and then applied ForceAtlas2 in Gephi (v. 0.9.5) to render a visual concept map of the 200 most frequent terms (except for the listed English stop words provided in textplot) based on the kde results (Jacomy et al., 2014). kde calculated the semantic distances between words by exploiting the phenomenon that semantically related words appear mostly in the same locus; closely related concepts cluster together in the map, while less closely related words appear at increasing distance. Words that are found in most contexts appear towards the centre of the map, those which are restricted to a few specific contexts appear on the fringes of the map. We included the sentiment information as follows: at the end of each positive tweet we added the pseudoword ‘sentipositive,’ at the end of each negative tweet the pseudoword ‘sentinegative (Figure 1).’
Flow chart of methodology used
Citation: Journal of Religion, Media and Digital Culture 14, 1 (2025) ; 10.1163/21659214-bja10123
3 Results
3.1 Mallet Output
We first present the Mallet output for seven topics, which we refer to as semantic clusters in Table 1.
Interpreting the Mallet output firstly involves manually inspecting the keywords, and secondly reading the documents that are most closely associated with a cluster. We investigated these most prototypical documents and the examples provided are derived from documents that are strongly associated with just one topic.
Cluster 1 is the topic with the biggest weight (0.17) and gathers around ‘religion’ and the ethics of ‘good’ and ‘bad’ things and deeds, including injunctions (‘don’t’), references to ‘science,’ ‘faith,’ and ‘life.’ It is also closely related to the term ‘spirituality.’ Examples are:
- (1)I believe in Religion. I know Truth of Scientific Discoveries. I have gained Knowledge of Religion and Science. Combing them simply makes me, A Muslim Scientist. I don’t need to be Atheist Scientist or Religious Flat Earther God created Binary Sex and Science proved Binary Exists!
- (2)… Sure, religion is an effective way to impart ideas, good or bad. It’s the bad ones that’s the problem. Being spiritual is very effective for evolution, but whats good for evolutionary isn’t spiritual necessarily …
- (3)… He’s disrespecting the whole Christian religion and God! He started this mess! I’m so tired of ppl playing the hating Christian card. It’s not our fault if you feel condemned …
- (4)… I was raised in church, Science has disproven a large portion of the Bible and any kind of god like figure. And one of the greatest things religion preaches is money is the root of all evil …
Cluster 2 is the most strongly weighted topic associated with ‘spirituality.’ Its clusters are subjective experiences (‘love,’ ‘healing,’ ‘motivation,’ ‘wisdom,’ etc.) and practices associated with such experiences, such as ‘meditation,’ ‘yoga,’ and ‘mindfulness.’ Moral and behavioural injunctions such as ‘good’ and ‘don’t’ are absent, as they are not strongly associated with ‘religion.’ Many positive connotations such as ‘love,’ ‘motivation,’ and ‘energy’ are central. The cluster also contains a large proportion of advertisements. Examples are:
- (5)TOMORROW SUNDAY 8th March 8PM: Mantras. Music. Meditation. Motivation! Urban Spirituality Presents \The Mantra Therapy Holi Festival Special!\“ SUNDAY 8th March 2021 8–M – 9.15PM (bst) Watch LIVE: Watch Free, Live at the Mantra Therapy Facebook page
- (6)Design a Conscious Home: Bring mindfulness into our lives by decorating our house with energy balancing techniques enriched with ancient spirituality, striving to live more conscious lives and evolving in unison with Mother Earth
Cluster 3 is closely associated with ‘spirituality’ and contains elements of topic 1 (‘god,’ ‘good,’ ‘faith,’ ‘life’) as well as subjective experiences (‘love,’ ‘hope’) and references to activities with spiritual significance (‘ramadan,’ ‘work’) typical of topic 1. Like cluster 2, mainly words with positive connotations (‘good,’ ‘love,’ ‘great’) appear. It clearly clusters around the Muslim religion, but also contains tweets with other confessional backgrounds. Examples are:
- (7)Ladies! don’t Miss the 2021 WINE; Shrine Women’s #Pilgrimage to #Italy in November! @TeresaTomeo & @KellyWahlquist Amazing #Florence #Siena #Orvieto #Assisi #Rome & more! Pray, eat, & learn to love la dolce vita! …
- (8)dahan dahan lang sa pag-iisip. Sa Kanya lang tayo kakapit. I know it’s tough at this time but hey, let’s still share faith, hope and love. We’ll all be fine in Jesus name. Full video link will be on my bio! God bless!
- (9)#RamadanMubarak to all of you. It’s the month of: Fasting -dawn to dusk Quran -guidance for mankind Special prayers @ night Spirituality -closeness to God Night of Power (Qadr) -better than 1000 months Zakat & Charity Increased good deeds Family time #Ramadan2021
Cluster 4 returns to a clearly Christian cluster of associations. It is strongly associated with ‘spirituality,’ ‘god,’ and ‘faith,’ and noticeably also contains direct references to ‘divine spirit’ ‘jesus,’ ‘lord,’ and ‘bible.’ In this case, attributing this topic to a particular confession is complicated by “younusalgohar,” a reference to Younus AlGohar, co-founder of Messiah Foundation International, a Pakistani new religious movement. Examples are:
- (10)‘Bold unflinching tales of the south, the supernatural, modern Christianity, depression, vision quests, healing & miracles. Inspiring tales of love & maturity in a postmodern America.’ An epic story by …
- (11)It is in pardoning that we are pardoned. #Catholic #Christian #religion #Jesus #Christ #Bible #faith #hope #PeacePrayer #peace #Franciscan #Paxetbonum #prayer #meditation #grace #spirituality #conscience #contrition #repent #repentance #forgiveness #clemency #kindness #fellowship
- (12)GOOD NEWS:
the Bible is still relevant.
the tomb is still empty.
the Gospel still works.
the Spirit still empowers.
the Church still lives.
Jesus still reigns.
God is still good.
Cluster 5 lacks a close association with ‘spirituality’ but contains several terms suggesting public controversies, such as ‘race,’ ‘gender,’ denied ‘freedom,’ ‘politics,’ ‘state,’ and ‘church.’ These terms and the negations (‘don’t’) indicate negative connotations. Examples are:
- (13)Can someone explain how Title vii of the Civil Rights Act of 1964.Title vii of the Civil Rights Act, as amended, protects employees and job applicants from employment discrimination based on race, color, religion, sex and national origin.
- (14)Also, the reason that middle-class America feels left behind is because we have allowed the right to redistribute wealth through tax policy that favors corporations and the rich. They use wedge issues like race, religion and guns to distract voters from economic issues.
- (15)In our General Elections this year, it seems the Holy Spirit has decided to participate to stop politicians misusing & misquoting the Bible … Hypocrites will be exposed … The TRUTH will prevail. Jesus said in John 14:6: I AM THE WAY, THE TRUTH & THE LIFE.
Cluster 6 is again closely associated with ‘spirituality’ and, as in topic 1, occurs in the context of practices such as ‘yoga,’ ‘mindfulness,’ and ‘meditation,’ as well as subjective experiences such as ‘happiness,’ ‘inspiration,’ ‘feelings,’ and ‘self-esteem.’ In this cluster, however, these practices and feelings are closely associated with terms suggesting the sale of books, especially in relation to parenting and children. Other coaching and self-help services are also advertised. Notably, ‘religion’ occurs, but in the context of positively connoted words such as ‘self-esteem’ and ‘happiness.’ Examples are:
- (16)The Romance of Ego & Conflict: A Practical-Spiritual Guide For Improving Your #Conflict & #Negotiation Abilities By Dissolving Your Own Ego #ebook … #Ego @EckhartTolle #spirituality #selfhelp #personaldevelopment #personalgrowth …
- (17)I’ve decided to write a self help book called “Life Sucks and Then You Die”. It will be full of watered down aphorisms and spirituality cobbled together from conflicting sources and traditions. It will include no actual help, but will _feel_ like it does. It will be a bestseller.
- (18)A first of it’s kind children’s books focusing solely on African indigenous Spirituality by Imboni Dr. uZwi-Lezwe Radebe is at the top of the charts at Exclusive Books. #ImboniuZwiLezwe #TheRevelationOfImboni #ExclusiveBooks #AfricanRestoration
Cluster 7 gathers tweets concerned with Hinduism and Islam, and are closely associated with both ‘religion’ and ‘spirituality.’ Noticeable are political references such as ‘country,’ particularly in relation to Hinduism (‘bjp,’ ‘caste’). Injunctions to not do something (‘don’t’) are also present. Examples are:
- (19)@dhwaj99 There are many Hindu ministers and mp in Bangladesh. We don’t discriminate anyone on the basis of Religion. Show me if there are any Muslim mp of your Leading Political Party bjp.
- (20)Both Corona and Quran is dangerous for Humanity …. If we know islam is not a Religion it’s Terrorism then why it’s not banned … #Islam #islamic_terrorism
- (21)When Abhrahamic religions and Bhuddism entered India, they provided circumvention of those middle men, we call priests. These new religions democratized spirituality and destroyed the #caste system. The bjp and #Hindutva is bringing the caste system back.
Many of these tweets concern discrimination due to religion and race. We observe that clusters 2 and 3 generally centre on spirituality and convey more positive attitudes than clusters 5 and 7, which center on religion.
3.2 Cluster Coding and Topic Identification
We assessed the coherence of the Mallet output by manually reviewing the most strongly weighted tweets associated with each cluster. Three authors, an anthropologist (Winiger), a church historian (Neuhold) and a theologian (Peng-Keller), then separately evaluated the thematic consistency of words in each cluster and proposed a tentative code for each, which we henceforth use as its name. Cluster 5, for instance, was named ‘politics,’ while cluster 6 was named ‘Lebensführung’ (conduct of life). The Krippendorff Alpha interrater reliability coefficient was 0.6429, which is near, but below the threshold allowing tentative conclusions (α ≥.667, Krippendorff, 2004).
3.3 Concept Map
We now present the concept map with sentiment analysis, where each tweet that is classified as expressing positive sentiment is labelled ‘sentipositive,’ and each tweet that is classified as negative is labelled ‘sentinegative.’ The following figure shows an automatically created concept map with the 200 most frequently used terms including associated sentiments (Figure 2).
Concept map with 200 terms and sentiment information
Citation: Journal of Religion, Media and Digital Culture 14, 1 (2025) ; 10.1163/21659214-bja10123
As this illustrates, words associated with topics 1 (theistic/Western), 4 (Christian/Evangelical), 5 (Politics), and 7 (Indian) cluster into the right, the ‘sentinegative’ side of the map. Words associated with topics 2 (mystical/subjective/practices) and 6 (conduct of life) cluster on the left, the ‘sentipositive’ side. While the center-left clusters are words associated with topic 3 (Islam). In the center of the map two additional clusters are evident, with the upper cluster gravitating around ‘Jesus,’ ‘bible,’ and related terms, and the lower cluster gravitating around ‘understand,’ ‘teach,’ ‘science’, and related terms suggesting the use of reason. Other than three confessional clusters (Indian, Islamic, and Evangelical Christian religiosity), a contrast is evident between a positively connoted pole of mystical and subjective practices associated with ‘spirituality’ and ethical conduct of life, and a negatively connoted pole associated with the topics of race, gender, and nation-state. This association is shown in the horizontal axis, where ‘spirituality’ forms a largely positively connoted cluster on the left, and ‘religion’ a negatively connoted cluster on the right.
4 Discussion
Over the course of the 20th and early-21st centuries, a process of semantic differentiation has occurred by which ‘religion’ and ‘spirituality’ emerged as dichotomous terms. In this development, ‘religion’ has become negatively connoted as authoritarian, hostile, dogmatic, controversial, etc. Whereas ‘spirituality’ – beginning with William James’ healthy mindedness, the Mesmerists, and the Lebensreform-movement – has become associated with health, authenticity, science and psychology, and salutogenic and life-affirming beliefs (Peng-Keller, 2019; Streib & Hood 2016; Bender & McRoberts, 2012). The emergence of ‘spirituality’ as a counter to Western institutional religion, it has been argued, is deeply intertwined with the (post-) colonial and orientalist production of the culturally Other to legitimise a Christian, hierarchical ordering of “world religions” and, more recently, to appropriate and commodify supposedly authentic cultural artefacts such as Yoga, Zen Meditation, or Native American shamanic ritual (King, 1999; Masuzawa, 2005).
At first glance, this dynamic is reflected in the concept map in the contrast between a negatively perceived, institutionalized ‘religion’ (associated with gender, class, race, and similar topics subject to public controversy) on the one end, and, on the other, a positive (mystical, subjective practice) spirituality associated with the conduct of life. This polarisation is illustrated by the following Tweets associated with topic 2:
- (22)Spirituality doesn’t come from religion, it comes from our Soul.
- (23)Religion divides. Spirituality unites.
- (24)Religion teaches us ignorance and selfishness, whereas spirituality teaches us awareness and promotes communalism. Religion says land belongs to no man but God meaning it justifies land dispossession of black people by colonizers. Religion says it’s okay to be poor because riches.
This dichotomy appears to persist when disaggregating the concept map in the topic model. Topic 2, representing ‘mystical,’ ‘subjective,’ and ‘practices,’ is most closely associated with enjoyable and meaningful experiences such as ‘love,’ ‘healing,’ and ‘peace,’ implying that these may be attained through practical activities such as ‘yoga,’ ‘mindfulness,’ and ‘meditation’ that are based on ‘wisdom’ and concern the ‘soul’ and ‘consciousness,’ or work with ‘energy.’ These terms reflect usage in the New Age and the contemporary esoteric milieu including its historical entanglement with (alternative) medicine (Lüddeckens & Schrimpf, 2018). A similar positive usage is also evident in topic 6, which we termed ‘conduct of life.’ This usage overlaps with topic 2 in its association with good feelings (‘happiness,’ ‘inspiration’), and makes more explicit the practical dimension of spirituality in relation to the reading of books, especially on the development of abilities such as parenting, self-esteem, and mindfulness. The prevalence of advertisements for (self-help) books in this topic also speaks to the critique of late-modern (and particularly ‘New Age’) spirituality as marred by rampant commercialism (Roof, 2001; Lipovetsky, 2011).
On the other end of the concept map, associations related to topic 5 (‘politics’) are visible where they neighbour terms related to Evangelical, Islamic, and Indian religiosity (topics 3, 4, and 7) as well as a host of terms which construct religion in relation to history, nation, state, and church. Notably, and again congruent with the contradistinction of topics 2 (spirituality) and 6 (conduct of life) as relating to personal experience, a host of associations occur which present religion as a matter of public order and morality: ‘society,’ ‘education,’ ‘control,’ ‘roles,’ ‘force,’ ‘law,’ ‘wrong,’ ‘stop,’ ‘freedom,’ ‘organized,’ and ‘women.’ Violent terms, including ‘kill’ and ‘hate,’ a selection of vulgarities, and negations such as ‘don’t’ and ‘isn’t’ illustrate the prescriptive and strongly negative sentiments associated with these terms.
In addition to the contradistinction between ‘negative’ religion and ‘positive’ spirituality suggested by sentiment analysis, the topic model and particularly the concept map may be read in terms of a differentiation of spirituality as a private, highly individualised form of religiosity – a notion of spirituality reflective of a “first wave” of scholarship on this topic which emerged in the 1990s (Steensland et al., 2021). In this framing, the rise of spirituality signifies the replacement of a socially responsible and politically engaged religion with the search for personal authenticity and actualisation – as exemplified Robert Bellah’s (1985) ‘Sheilaism’ (p. 221) – or more sceptically, a narcissistic retreat into the self and even the outright abandonment of the moral and emancipatory force of ‘religion’ (Žižek, 2001; Jain, 2020; Carrette & King, 2005).
Yet, while the topic model and concept map seem to confirm widespread popular distinction of spirituality as a ‘good’ and ‘private’ phenomenon, these distinctions are not always possible to sustain on closer analysis. While our method shows average trends, it cannot reflect the wide dispersion of opinions. As noted by José Casanova (1992), “of all social phenomena, none is perhaps as protean, and, consequently, least susceptible to binary classification, as religion” (p. 17). This observation may also apply to the study of spirituality (Ammerman, 2013), where some authors have differentiated between 10 different types of contemporary spirituality (Streib & Hood, 2016; Cour et al., 2012).
Closer inspection of the topic model and concept map allows for several alternative interpretations, which we briefly and qualitatively explore now. The distinction of a ‘private’ spirituality from a ‘public’ religion has come under sustained critique for conceiving of spirituality as an “outgrowth of modernization” located beyond social and communal influence, argued to reflect secularization theory rather than the cultural and socio-political complexity of the phenomenon (Steensland et al., 2021). This ambivalence is evident in the topic model. Topic 6 (conduct of life), for instance, is associated both with ‘spirituality’ and with ‘religion’ – albeit much more strongly with ‘spirituality’ (the list of keywords is sorted in decreasing order). To be sure, historically, ‘conduct of life’-spirituality has not been privy to the marketplace of yoga teachers and mindfulness advice. Rather, for centuries it has been central to the Christian tradition (de Certeau, 2015, 1992), where it has found its most vibrant expressions precisely at the heart of institutional religious life, such as in contemplative prayer or the spiritual exercises of Ignatius of Loyola (1522–1524) and its contemporary followers. This is reflected in the co-occurrence of ‘religion’ in this cluster, although in the context of positively connoted words such as ‘self-esteem’ and ‘happiness.’
Similarly, the opposition of ‘good’ spirituality and ‘bad’ religion appears to contradict that Islamic religion is also closely related – though highly negatively connoted on the concept map – to terms suggesting practiced, daily religion – such as ‘ramadan,’ ‘month,’ ‘work,’ and ‘days.’ The faultline demarcating ‘good’ spirituality and ‘bad’ religion is not reducible to the distinction between contemporary, private New Age-practices and traditional, confessional religion, as suggested by the positively connoted terms ‘prayer,’ ‘listen,’ ‘join,’ or ‘hope’ that are associated with topics 4 (Evangelical, Christian), 7 (Indian), and 3 (Islam). This is also illustrated by the paradoxical association of spirituality both with the (positively connoted) practice of Yoga in the concept map, and, in the topic model, the concurrent association of ‘spirituality’ with terms such as ‘caste’ and ‘bjp’ (the Indian ruling Bharatiya Janata Party) and related terms referring to the conflict between Hindus and Muslims.
In addition to the positive/negative and public/private distinctions raised here, the generally negatively usage of ‘religion’ may be questioned on the grounds of the tendency of Twitter/‘X’ to amplify heated debates through a rolling, off-hand commentary on current events, which is subject to bias towards content representative of the political right (Huszár et al., 2022) and various algorithmic effects such as filter bubbles and echo chambers (Flaxman et al., 2016; Cinelli et al., 2021). The geographic, ethnic, and confessional over- and underrepresentation of users in this sample also ought to be noted. This is illustrated by the (over-)representation of tweets relating to the Indian ruling party, bjp, the Pakistani new religious movement of Younus AlGohar, and entrepreneurs such as Asaf Shani, a self-declared spiritual seeker, “negotiations expert”, and author of children’s books, who has 57,600 followers. This (over-)representation may be related to usage patterns among Twitter/‘X’ users in South Asia – home to very large populations with high levels of internet usage – and the central role played by religion and ‘spirituality’ in political and specifically nationalist discourse (van der Veer, 2001, 2014). At the same time, there is a conspicuous absence of East Asian and South American users in this sample, which ought not to be mistaken for a general absence of users from these geographic backgrounds. The presence of automated bots, advertisers, and special interest groups may be partly responsible for the strong presence of (book) advertisements.
Finally, the principle that the meaning of a word may be deduced from its neighboring terms generates ideal typical use cases, which represent the most common co-occurrences at the cost of tweets which defy categorisation by means of statistical affinity. It also struggles with the pragmatics of a speech act: a proportion of tweets bear implications which only become evident when read in context, or imply a neutral or generally positive understanding of ‘spirituality,’ but lament a particular behavior which compromises its value. Note for instance, the following:
- (25)social media spirituality messing w y’all heads bad. you not losing friends because you “vibrate on a higher frequency” you’re a bad person and owe them $20
- (26)don’t let nobody on spiritual twitter decide for you. I don’t know what is happening here but lately a lot of folks into ‘spirituality’ seem to choose meanings for everyone and it’s not worth your time getting sucked into that, keep your thoughts firm and the assumption will be true
- (27)Religion is for people who fear hell, spirituality is for people who have been there.
- (28)Spirituality isn’t all love & light. At some point you gonna get dragged.
In view of these complexities, we propose that the method presented here ought to be understood as an opportunity for the critical interrogation of the sociological and political circumstances which factor into the lexical salience of ‘religion’ and ‘spirituality’ reflected on Twitter/‘X.’ The platform is able to show average trends, but not the large distribution of opinions. The ambiguities of ‘spirituality’ shown through this method, we argue, indicate a coherent notion of the term as well as the co-presence of diverging and occasionally contradictory usages (cf. Kim et al., 2020). Despite evident continuities with popular perceptions, particularly regarding the ‘good’/’bad’ and ‘private’/’public’ distinction, a closer analysis of the sample presented here suggests the co-existence of clear trends and a surprisingly heterogenous usage. The utility of the method presented here, we suggest, may therefore be found in stimulating critical reflection on the possibilities and limitations of the cross-cultural relevance of computational semantics in the study of spirituality.
5 Summary
In this study, we employed a usage-based approach based on the Firthian principle, central to the field of distributional semantics, that a word may be known “by the company it keeps.” We explored the utility of this principle using an interdisciplinary, mixed-methods approach combining computer linguistics and qualitative analysis. We analysed the degree relatedness of words associated with ‘religion’ and ‘spirituality’ and their attendant sentiments, either positive or negative, as they were used in the Twittersphere. A convenience sample of 138,000 tweets gathered between March and April 2021 was used to calculate seven different word clusters of highly related terms. Three authors, an anthropologist, church historian, and theologian, then independently coded these clusters to propose a tentative topic name for each. The seven topics thus identified were ‘theistic/Western,’ ‘mystical, subjective practices,’ ‘Islam,’ ‘evangelical Christian,’ ‘politics,’ ‘conduct of life,’ and ‘Indian.’ An unsupervised sentiment analysis was then conducted to visualise this data on a concept map, illustrating the statistical affinity between each topic.
Finally, we discussed the topic model and concept map in light of recent academic literature regarding these two terms. The data corroborated our hypothesis that ‘religion’ and ‘spirituality’ constitute two distinct terms in real-world usage. Secondly, the data showed the collocation of negatively connoted, institutionalized ‘religion’ with gender, class, race, and similar topics subject to heated public debate on one end of the concept map. On the other end of the map, a positively connoted, mystical, subjective practice of spirituality collocated with terms related to an ethical conduct of life. We suggest that our data confirms a popular perception of ‘spirituality’ as a generally highly positive term rooted in lived virtues, while ‘religion’ carries strong negative connotations and is associated with public order and moral injunctions. At the same time, however, the data may be interpreted to arrive at more nuanced and, at times, contradictory understandings. Bearing in mind its limitations, the method outlined here may stimulate critical reflection on the possibilities and limitations of computational semantics in the study of religion and spirituality.
Acknowledgements
This research has been supported by the urpp Digital Religion(s) at the University of Zurich. We are grateful to the anonymous reviewers for their valuable comments.
References
Ammerman, N. T. (2013). Spiritual But Not Religious? Beyond Binary Choices in the Study of Religion. Journal for the Scientific Study of Religion, 52 (2), pp. 258–78. https://doi.org/10.1111/jssr.12024.
Balboni, M. J. & Balboni, T. A. (2018). Hostility to Hospitality: Spirituality and Professional Socialization within Medicine. Oxford: Oxford University Press.
Baroni, M. & Lenci, A. (2010). Distributional Memory: A General Framework for Corpus-Based Semantics. Computational Linguistics, 36 (4), pp. 673–721. https://doi.org/10.1162/coli_a_00016.
Bartsch, S. & Evert, S. (2014). Towards a Firthian Notion of Collocation. Vernetzungsstrategien, Zugriffsstrukturen und automatisch ermittelte Angaben in Internetwörterbüchern, 2 (1), pp. 48–61.
Bellah, R. N., Madsen, R., Sullivan, W. M., Swidler, A., & Tipton, S. M. (1985). Habits of the Heart: Individualism and Commitment in American Life. California, CA: University of California Press.
Bender, C. & McRoberts, O. (2012). Mapping a Field: Why and How to Study Spirituality [Working Paper]. Social Science Research Council.
Bergunder, M. (2014). What is Religion?: The Unexplained Subject Matter of Religious Studies. Method & Theory in the Study of Religion, 26 (3), pp. 246–286.
Blei, D., & Lafferty, J. (2006). Correlated topic models. Advances in Neural Information Processing Systems, 18, 147.
Burge, R. (2021). How Twitter Discusses American Religion. In B. Calfano (Ed.), Exploring the Public Effects of Religious Communication on Politics (pp.150-?). Ann Arbor: University of Michigan Press.
Bybee, J. (2007). Frequency of Use and the Organization of Language. Oxford: Oxford University Press. https://doi.org/10.1093/acprof:oso/9780195301571.001.0001.
Carrette, J. & King, R. (2005). Selling Spirituality: The Silent Takeover of Religion. Abingdon, UK: Routledge.
Casanova, J. (1992). Private and Public Religions. Social Research, 59 (1), pp. 17–57.
Chang, J., Boyd-Graber, J., Gerrish, S., Wang, C., & Blei, D. M. (2009). Reading Tea Leaves: How Humans Interpret Topic Models. In Y. Bengio, D. Schuurmans, J. D. Laferty, C. K. I. Williams, & A. Culotta (Eds.), Proceedings of the 22nd International Conference on Neural Information Processing Systems, Vancouver, Canada (pp. 288–296). Curran Associates.
Cinelli, M., De Francisci Morales, G., Galeazzi, A., Quattrociocchi, W., & Starnini, M. (2021). The Echo Chamber Effect on Social Media. Proceedings of the National Academy of Sciences, 118 (9), e2023301118. https://doi.org/10.1073/pnas.2023301118.
Cour, P., Hørdam Ausker, N., & Hvidt, N. C. (2012). Six Understandings of the Word ‘Spirituality’ in a Secular Country. Archive for the Psychology of Religion, 34 (1), pp. 63–81. https://doi.org/10.1163/157361212X649634.
de Certeau, M. (1992). The Mystic Fable, Volume One: The Sixteenth and Seventeenth Centuries (M. B. Smith, Trans.). Chicago: University of Chicago Press.
de Certeau, M. (2015). The Mystic Fable, Volume Two: The Sixteenth and Seventeenth Centuries (M. B. Smith, Trans.). Chicago: University of Chicago Press.
Drescher, E. (2016). Choosing Our Religion: The Spiritual Lives of America’s Nones. Oxford: Oxford University Press.
Firth, J. R. (1957). Papers in Linguistics, 1934–1951. Oxford: Oxford University Press.
Fitzmaurice, S., Robinson, J. A., Alexander, M., Hine, I. C., Mehl, S., & Dallachy, F. (2017). Linguistic dna: Investigating Conceptual Change in Early Modern English Discourse. Studia Neophilologica, 89 (1), pp. 21–38. https://doi.org/10.1080/00393274.2017.1333891.
Flaxman, S., Goel, S., & Rao, J. M. (2016). Filter Bubbles, Echo Chambers, and Online News Consumption. Public Opinion Quarterly, 80 (1), pp. 298–320. https://doi.org/10.1093/poq/nfw006.
Goldberg, A. (2006). Constructions at Work: The Nature of Generalization in Language. Oxford: Oxford University Press.
Hartmann, J., Heitmann, M., Siebert, M., & Schamp, C. (2022). More than a Feeling: Accuracy and Application of Sentiment Analysis. International Journal of Research in Marketing, 40 (1), pp. 75–87. https://doi.org/10.1016/j.ijresmar.2022.05.005.
Hayes, A. F. & Krippendorff, K. (2007). Answering the Call for a Standard Reliability Measure for Coding Data. Communication Methods and Measures, 1 (1), pp. 77–89. https://doi.org/10.1080/19312450709336664.
Hjelm, T. (2016). Theory and Method in Critical Discursive Study of Religion: An Outline. In F. Wijsen & K. von Stuckrad (Eds.), Making Religion: Theory and Practice in the Discursive Study of Religion, Volume 4 (pp. 13–34). Leiden: Brill. https://doi.org/10.1163/9789004309180_003.
Huszár, F., Ktena, S. I., O’Brien, C., Belli, L., Schlaikjer, A., & Hardt, M. (2022). Algorithmic Amplification of Politics on Twitter. Proceedings of the National Academy of Sciences, 119 (1), e2025334119. https://doi.org/10.1073/pnas.2025334119.
Jacomy, M., Venturini, T., Heymann, S., & Bastian, M. (2014). ForceAtlas2, a Continuous Graph Layout Algorithm for Handy Network Visualization Designed for the Gephi Software. plos one, 9 (6), e98679. https://doi.org/10.1371/journal.pone.0098679.
Jain, A. R. (2020). Peace Love Yoga: The Politics of Global Spirituality. Oxford: Oxford University Press.
Kearney, M. W. (2020). Package ‘Rtweet’ [Source Code]. https://cran.r-project.org/web/packages/rtweet/rtweet.pdf.
Kim, S. H., Lee, N., & King, P. E. (2020). Dimensions of Religion and Spirituality: A Longitudinal Topic Modeling Approach. Journal for the Scientific Study of Religion, 59 (1), pp. 62–83. https://doi.org/10.1111/jssr.12639.
King, R. (1999). Orientalism and Religion: Postcolonial Theory, India and ‘the Mystic East.’ London and New York: Routledge.
Koenig, H. G. (2012). Religion, Spirituality, and Health: The Research and Clinical Implications. isrnPsychiatry, pp. 1–33. https://doi.org/10.5402/2012/278730.
Krippendorff, K. (2004). Content Analysis: An Introduction to Its Methodology. Los Angeles: sage Publications.
Landauer, T. K. & Dumais, S. T. (1997). A Solution to Plato’s Problem: The Latent Semantic Analysis Theory of Acquisition, Induction, and Representation of Knowledge. Psychological Review, 104 (2), p. 211. https://doi.org/10.1037/0033-295X.104.2.211.
Lenci, A. (2008). Distributional Semantics in Linguistic and Cognitive Research. Italian Journal of Linguistics, 20 (1), pp. 1–31.
Lipovetsky, G. (2011). The Hyperconsumption Society. In K. M. Ekström & K. Glans (Eds.), Beyond the Consumption Bubble (pp. 25–36). London; New York: Routledge.
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019, July). RoBERTa: A Robustly Optimized bert Pretraining Approach. ArXiv:1907.11692 [Computer Software]. http://arxiv.org/abs/1907.11692.
Lüddeckens, D. & Schrimpf, M. (Eds.). (2018). Medicine – Religion – Spirituality: Global Perspectives on Traditional, Complementary, and Alternative Healing. Bielefeld: Transcript-Verlag.
Masuzawa, T. (2005). The Invention of World Religions: Or, How European Universalism Was Preserved in the Language of Pluralism. Chicago: University Of Chicago Press.
McCallum, A. K. (2002). mallet: A Machine Learning for Language Toolkit [Source Code]. http://mallet.cs.umass.edu.
McClure, D. (2015). Textplot [Source Code]. https://github.com/davidmcclure/textplot.
Neubert, F. (2016). Die diskursive Konstitution von Religion. Wiesbaden: Springer Fachmedien.
Peng-Keller, S. (2019). Genealogies of ‘Spirituality’ – An Historical Analysis of a Travelling Term. Journal for the Study of Spirituality, 9 (2), pp. 86–98. https://doi.org/10/gjpd2q.
Pew Research Center. (2018, May 29). Being Christian in Western Europe. https://www.pewresearch.org/religion/2018/05/29/being-christian-in-western-europe/.
Puchalski, C. M., Vitillo, R., Hull, S. K., & Reller, N. (2014). Improving the Spiritual Dimension of Whole Person Care: Reaching National and International Consensus. Journal of Palliative Medicine, 17 (6), pp. 642–56. https://doi.org/10.1089/jpm.2014.9427.
Roof, W. C. (2001). Spiritual Marketplace: Baby Boomers and the Remaking of American Religion. Princeton: Princeton University Press.
Rubenstein, H. & Goodenough, J. B. (1965). Contextual Correlates of Synonymy. Communications of the acm, 8 (10), pp. 627–33. https://doi.org/10.1145/365628.365657.
Rüdiger, M., Antons, D., Joshi, A. M., & Salge, T. O. (2022). Topic Modeling Revisited: New Evidence on Algorithm Performance and Quality Metrics. PLoS ONE, 17 (4), e0266325. https://doi.org/10.1371/journal.pone.0266325.
Sahlgren, M. (2006). The Word-Space Model: Using Distributional Analysis to Represent Syntagmatic and Paradigmatic Relations between Words in High-Dimensional Vector Spaces [Doctoral dissertation, Stockholm University]. https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=1521ddb27860cc8834f8a82e62665bf983c8ad2c.
Saler, B. (1993). Conceptualizing Religion: Immanent Anthropologists, Transcendent Natives, and Unbounded Categories. New York and Oxford: Berghahn Books.
Schütze, H. & Pedersen, J. O. (1995). Information Retrieval Based on Word Senses. Proceedings of the 4th Annual Symposium on Document Analysis and Information Retrieval, pp. 161–75.
Schwartz, H. A. & Ungar, L. H. (2015). Data-Driven Content Analysis of Social Media: A Systematic Overview of Automated Methods. The annals of the American Academy of Political and Social Science, 659 (1), pp. 78–94. https://doi.org/10.1177/0002716215569197.
Steensland, B., Kucinskas, J., & Sun, A. (2021). Eminently Social Spirituality: Context, Practice, Power. In B. Steensland, J. Kucinskas, & A. Sun (Eds.), Situating Spirituality: Context, Practice, and Power (pp. 3–30). Oxford: Oxford University Press.
Streib, H. & Hood, R. (Eds). (2016). Semantics and Psychology of Spirituality: A Cross-Cultural Analysis. Cham: Springer International Publishing. https://doi.org/10.1007/978-3-319-21245-6.
Taira, T. (2022). Taking ‘Religion’ Seriously: Essays on the Discursive Study of Religion. Leiden: Brill. https://doi.org/10.1163/9789004511682.
van der Veer, P. (2001). Imperial Encounters: Religion and Modernity in India and Britain. Princeton: Princeton University Press.
van der Veer, P. (2014). The Modern Spirit of Asia: The Spiritual and the Secular in China and India. Princeton: Princeton University Press.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention Is All You Need. In U. von Luxburg, et al.., (Eds.), Advances in Neural Information Processing Systems 30: 31st Annual Conference on Neural Information Processing Systems. Curran Associates, Inc. https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html.
Veidlinger, D. (2021). Religion: A Little Birdie Told Me Something about Religion, Religion on Twitter. In H. A. Campbell & R. Tsuria (Eds.), Digital Religion: Understanding Religious Practice in Digital Media (2nd ed., pp.132–40). Lanham: Routledge.
Worthington., E. L., Hook, J. N., Davis, D. E., & McDaniel M. A. (2011). Religion and Spirituality. Journal of Clinical Psychology, 67 (2), pp. 204–14. https://doi.org/10.1002/jclp.20760.
Wright, B. R. E. (2022). Methodological Innovations for the Study of Spirituality. In B. Steensland, J. Kucinskas, & A. Sun (Eds.), Situating Spirituality: Context, Practice, and Power (pp. 113–27). Oxford: Oxford University Press. https://doi.org/10.1093/oso/9780197565001.003.0006.
Žižek, S. (2001). The Fragile Absolute, Or, Why is the Christian Legacy Worth Fighting For? London and New York: Verso.
Zucchini, W. (2003). Applied Smothing Techniques – Part 1: Kernel Density Estimation. http://staff.ustc.edu.cn/~zwp/teach/Math-Stat/kernel.pdf.
A popular measure for setting the optimal number is perplexity (Blei and Lafferty 2006). In a plot of perplexity against number of topics, the local minimum of clearly distinguishable “elbows”, where the trade-off appears advantageous, can be used as indication of the optimum number of topics. But Chang et al. (2009) observe, that “models which achieve better predictive perplexity often have less interpretable latent spaces” (Chang et al. 2009: 2). Also in our data, we did not observe a clear elbow. Instead, we see a regular decrease. Rüdiger et al. (2022) even reject most of the suggested automatic methods. Our choice of 7 topics is thus the result of inspecting the topics and practical considerations: 7 is enough to be distinctive, but not too many to be excessively detailed and possibly difficult to interpret.
Mallet strongly recommends to set the hyperparameters automatically, see e.g. https://senderle.github.io/topic-modeling-tool/documentation/2018/09/27/optional-settings.html.
