Introduction

In recent years, the popularity of social media has led to an increased interest in social media discourse and, more specifically, the hashtag, i.e., the pound sign followed by a word, phrase, or a number, e.g., #314CANNES (Bruns & Burgess, 2011; Cunha et al., 2011; Huang et al., 2010; Wikström, 2014; Zappavigna, 2015). As it is currently used, the hashtag enacts social relations by means of the interpersonal functions it has assumed (Zappavigna & Martin, 2018, p. 11). One of the main functions of the hashtag is to express stance (Laucuka, 2018; Lee, 2018; Wikström, 2014; Zappavigna, 2015), which is a central concept in the present study.

This study examines stance-conveying hashtags in tweets about organic food (see Section 4). We investigate how writers of tweets (tweeters) take stance in their tweets and how they indicate to the readers how to interpret the tweets by means of the hashtags they use. Our aim is to provide a categorisation of functions expressed by hashtags which explicitly convey stance based on a qualitative analysis of tweets, which can also be applied to a different data set. The research question we address is: What functions do stance-conveying hashtags express in tweets about organic food? This is a natural topic for the analysis of stance taking in discourse by virtue of the status of organic food as a credence goodFootnote 1. According to Anisimova et al. (2019), consumers tend to seek information about organic food in sources of information which are independent of organic food companies, such as social media sites, because they (i.e., consumers) trust the feedback of other consumers who have experienced these products. In this sense, the expression of a stance towards organic food by a consumer can impact other consumers.

Twitter is a suitable source of data for examining stance conveying hashtags since the type of discourse is dialogic. Tweeters use hashtags in order to participate in asynchronous online conversations with other Twitter users on a specific topic as highlighted by the hashtags they choose (Huang et al., 2010, p. 175). Being dialogic in character, Twitter hashtagging is expected be a useful instrument for signalling stances taken by the tweeters. Consequently, this dialogic characteristic of Twitter, which has also been noted by other studies, e.g., Mas Manchón and Guerrero-Solé (2019), justifies our analysis of hashtags from a stance-related perspective and adds to work which addresses the need for more research on hashtags from a discourse perspective (Lee, 2018).

We study stance applying Du Bois’ (2007) stance triangle to account for the dialogic nature of hashtagging. Du Bois’ (2007) approach to stance is particularly relevant to our analysis as it highlights the interactive and dialogic nature of stance-taking on social media (Lee, 2018, p. 2). Moreover, we draw on Francis’ (1994) work on labels, and metalinguistic labels in particular, to account for aspects of hashtag use such as the function of a hashtag based on its position in the tweet. Metalinguistic labels are cohesive devices used in written discourse in order to characterise a piece of text as a specific linguistic act, e.g., a statement (Francis, 1994, p. 83). The notion of metalinguistic label is especially relevant to the analysis of epistemic, e.g., #fact, and deontic, e.g., #advice, hashtags to which this study draws attention. Epistemic and deontic hashtags are used in order to characterise the content of the tweet either in terms of its epistemic status (epistemic hashtags), or in terms of a directive expressed in the tweet (deontic hashtags)—more information about epistemic and deontic hashtags will be provided in section 3. To the best of our knowledge, neither epistemic nor deontic hashtags have been explicitly named in the literature despite studies (e.g., Zappavigna, 2015) on this topic. We argue that by using the descriptive terms “epistemic” and “deontic” to refer to such hashtags, we can raise awareness of their systematic use and function. The paper is structured as follows. Section 2 analyses the notion of stance providing an overview of DuBois’ (2007) stance triangle. Section 3 introduces hashtags followed by an overview of the functions of stance-conveying hashtags identified in the literature and moves on to explain the relevance of metalinguistic labels to the analysis of hashtags. Section 4 discusses the data used and the methodology followed, and section "Results and discussion" presents and discusses the results of the study. More specifically, Section "A categorisation scheme of the stance-conveying hashtags". introduces the categorisation scheme of the stance-conveying hashtag functions, and Section "Evaluation of the categorisation scheme" evaluates the hashtag functions we identified. Finally, section "Conclusion" concludes the paper.

Stance Taking in Discourse

Stance is the focus of the present study. It is here defined as “the way speakers position themselves in relation to their own or other people’s beliefs, opinions and statements about things or ideas in ongoing communicative interaction with other speakers” (Simaki et al., 2020, p. 217). Stance taking relates to the evaluation of both concrete objects, such as apples, and abstract entities, such as information. Moreover, it draws attention to the presence of participants in the communicative event. In this sense, stance taking is an interactive phenomenon.

In his analysis, Du Bois (2007, p. 163) defines stance as “a public act by a social actor, achieved dialogically through overt communicative means, of simultaneously evaluating objects, positioning subjects (self and others), and aligning with other subjects, with respect to any salient dimension of the sociocultural field”. The claim about stance being “dialogically achieved” is meant to illustrate that stance-takers build on the utterances of previous stance-takers. The definition states that stance is a type of social action, carried out by speakers through language in a dynamic process. It involves three elements: (1) the stance-taker performing an evaluation act, (2) the object undergoing evaluation and (3) the context of evaluation. Du Bois (2007) points out that stance taking entails the evaluation of an object either explicitly or implicitly.

In order to explain his view of stance-taking, Du Bois (2007) introduced the “stance triangle”, which is shown in Figure 1. Du Bois’ (2007, p. 144) stance triangle shows how the three interrelated facets of stance-taking, i.e., (1) evaluation, (2) positioning and (3) alignment, interact as they co-occur in a stance-taking event. He emphasises the contribution of the context to the interpretation of a stance taken in terms of who the stance-taker is, what the stance object is and what other stances have been expressed about the stance object (Du Bois, 2007).

Fig. 1
figure 1

The Stance Triangle (from Du Bois, 2007, p. 163)

Full size image

As can be seen in Fig. 1, the stance triangle includes two subjects, i.e., two speakers, taking a stance towards a single stance object. The existence of the two subjects (speakers) serves to highlight the interactive and dialogic nature of stance-taking mentioned previously. Moreover, Fig. 1 also shows the dynamic character of stance-taking in the sense that a stance taken towards an object relates to a stance taken previously by another stance-taker. Furthermore, it represents the collaboration between the two stance subjects in the construction of stance. For instance, two stance subjects can take turns to evaluate an object, with a stance subject providing the first stance (stance lead) and the second stance subject taking a new stance responding to the stance provided by the first stance subject (stance follow). While expressing a stance, the stance-taker is attributing a characterisation to the object of stance which is undergoing some kind of evaluation (this is the evaluation facet of stance-taking), while invoking a sociocultural value. For example, in the case of the utterance in (1):

  1. (1)

    The dinner is tasty

The stance-taker is attributing the characterisation tasty to the dinner he/ she had. At the same time, the speaker is positioning himself/ herself on a scale (this is the positioning facet of stance-taking). Hence, stance-taking combines both subjective and objective elements in the sense that it involves positioning, which focuses on the stance subject, and evaluation, which focuses on the stance object (Du Bois, 2007, p. 158). The following examples show that positioning may be affective, as in (2), or epistemic as in (3):

  1. (2)

    I love chocolate

  2. (3)

    I guess

The expression of the stance taker’s stance may involve the object of stance, e.g., chocolate in (2), or not, as in (3). Furthermore, while taking a stance, the speaker can indicate how his/ her stance relates to other stances taken previously by other stance-takers (this is the alignment facet of stance-taking), as in (4):

  1. (4)

    I disagree with you

An example offered by Du Bois (2007, p. 159) is found in (5) where Sam’s utterance is the stance lead, and Angela’s utterance is the stance follow, which builds on Sam’s stance:

  1. (5)

    SAM: I don’t like those

    ANGELA: I don’t either.

The analysis of stance and stance-taking deserves more attention as taking a stance is a basic function of language because by taking a stance, speakers not only express their evaluations of things, but they also indicate how their own stance relates to stances previously taken by others (Du Bois, 2007, p. 139). Moreover, the mere existence of evaluative language renders a text interactive in the sense that the writer, or speaker, invites the reader, or hearer, to accept the assumptions made in the text (Hunston & Thompson, 2000), which can also relate to Twitter discourse given its conversational nature.

Du Bois’ (2007) work is highly relevant to the current analysis by virtue of the emphasis it places on the dialogic nature of stance-taking, which also characterises hashtags on Twitter (Huang et al., 2010) since they are used by users in order to contribute to online conversations. That said, it should be mentioned that stance per se can also be used in monologue in the sense that there is always an addressee. Nevertheless, the conversational character of tweets is also supported by our data as in (6) whereby the tweeter is inviting a response from the tweet readers, as is shown in the following example:

  1. (6)

    Looking for retail locations in Wisconsin and Iowa. Any suggestions? #TexasBowl #retail #Organic #Food #shopping #Wisconsin #Iowa #advice

The tweeter in (6) states that he/ she is searching for retail locations in Wisconsin and Iowa asking the readers for suggestions. He/ She adds a sequence of hashtags which repeats some keywords from the tweet probably in order to attract the attention of users who could reply to the tweet. However, in addition to using the hashtags to increase the visibility of the tweet, the tweeter uses #advice in order to specify the type of contribution he/ she is interested in receiving from other users. In this sense, #advice also indicates the stance of the tweeter and their preference for a specific type of response. This request for a specific response is conversational as the tweeter is inviting the reader to respond in as specific way. To the best of our knowledge, observations about the use of hashtags to indicate the preferred uptake from the reader are absent from the literature. More information about hashtags and their functions is offered in the following section.

Hashtags

This section starts with an overview of the hashtag functions identified in the literature. Next, it introduces Francis’ (1994) notion of metalinguistic labels and explains how it can be applied to the analysis of hashtags.

The Development of Hashtags

Bernard (2019) argues that the hash sign has changed from being a functional element to a meaningful element of the message in which it occurs as a result of its use in social media platforms, such as Twitter, where it was introduced in 2007 (Bernard, 2019, p. 8) in the form of hashtags. The view of hashtags as meaningful message elements is also shared by Lee (2018, p. 2), who argues that hashtags often express stance. Moreover, Zappavigna and Martin (2018, p. 11) maintain that the hashtag has started to serve interpersonal functions related to “enacting social relations” after its introduction into social media platforms.

Studies on hashtags have noted a range of functions served by hashtags beyond their initial function of enabling content searchability (Laucuka, 2018; Matley, 2018; Scott, 2015; Zappavigna, 2015). The hashtag function of expressing stance is common among various studies on hashtags (Laucuka, 2018; Matley, 2018; Wikström, 2014; Zappavigna, 2015), and it is regarded as one of its main functions (Lee, 2018).

The following section offers an overview of the stance-conveying functions identified in the literature.

Stance-Conveying Hashtags

As already mentioned, one of the main functions of hashtags is that of expressing stance. Bruns and Burgess (2011) explain that hashtags can be employed by tweeters for adding emphasis to their tweets, thus replacing more traditional ways of indicating emphasis, such as underscoring text. Hashtags are thus seen as devices for marking a part of a tweet as a piece of information to be especially considered by the reader.

Approaching hashtags from a pragmatic perspective, Wikström (2014) shows how speech act theory can be applied to the analysis of hashtag functions. He identifies a set of eight functions in his data: topic tags (which indicate the topic of a tweet), hashtag games (which enable participation in online wordplay games), meta-comments (which indicate the intended interpretation of a tweet, or its illocutionary point), parenthetical explanations/ additions (which clarify the content of a tweet), emotive usage (whereby hashtags illustrate the reaction of the tweet writer as a result of what is being described in the tweet), emphatic usage (whereby hashtags highlight an element of the tweet replacing more traditional means of emphasis addition), humorous and playful usage, memes and popular culture references (hashtags containing quotes).

Apart from highlighting the multifunctionality of hashtags, Wikström’s (2014) relevance to the present study lies in his observation that hashtags can be employed for softening or boosting the content of a post. In addition, hashtags can be used to disambiguate the intended post interpretation. This use falls under his meta-comment function mentioned above, which involves the clarification of the intended interpretation of a tweet. For instance, he mentions how #opinion can be used for suggesting the subjective nature of the information contained in a post, as in (7):

  1. (7)

    Also think “Webcomics” should only be used to refer to a spec. biz model. It’s an outdated descriptor for a genre or a community. #opinion (Wikström, 2014, p. 136).

In this case, #opinion is used for explicitly stating the nature of the information expressed in the post, namely the tweet writer’s opinion. Furthermore, he also explains that the use of #statement can increase the force of a post presenting it as an assertion which is not to be challenged, as in (8):

  1. (8)

    #statement: I’m cooler than you.

However, as he points out, there is incongruence between the strength of a statement and the petty message communicated by the tweet in the example he provides. Consequently, the tweet is to be interpreted as an instance of irony, as he explains. The contribution of hashtags to the coherence of the tweet by indicating the intended tweet interpretation has also been highlighted by Kunneman et al. (2015). They show how #sarcasm has been used in tweets to indicate to the reader that the tweet containing #sarcasm is to be interpreted as an instance of irony. In addition, they point out that using hashtags for indicating the sarcastic character of a tweet leads to a reduction in the number of linguistic markers of sarcasm. The study illustrates the connection between stance and hashtags since the focus of the study was the use of #sarcasm in evaluative expressions by tweeters for guiding the reader towards the intended interpretation of the tweet.

The potential of hashtags to guide the reader towards the intended tweet interpretation has been analysed more thoroughly by Scott (2015) from a relevance-theoretic perspective. Working on data from Twitter, Scott (2015) argues that since a tweet can be accessed by an infinite number of users in different locations at any point in time, tweeters need to ensure that they provide enough contextual information to the reader. Adequate contextual information is necessary so that readers can arrive at the initially intended interpretation of the tweet regardless of where or when they read the tweet. According to Scott (2015, p. 13), hashtags serve the function of guiding the inferential process of the reader towards the interpretation of a tweet. In this sense, hashtags make accessible to readers the relevant assumptions which are necessary for understanding the tweet correctly (Scott, 2015, p. 14). Furthermore, hashtags function as highlighting devices, which direct the reader’s attention to specific items in a tweet.

Adopting a systemic functional approach, Zappavigna (2015) shows how hashtags perform three simultaneous linguistic meta-functions, i.e., (1) experiential, (2) interpersonal, and (3) textual. The experiential meta-function regards the use of hashtags to indicate the topic of a post as in (9), which is about the series Dr Who:

  1. (9)

    I seriously don’t understand why people like Rose Tyler. Is it because she’s pretty & she has a required love thing with the Doctor? #DrWho (Zappavigna, 2015, p. 6).

The interpersonal meta-function pertains to expressing attitudes and building relationships with other individuals, as in (10):

  1. (10)

    Woot! Second zombie and I’m actually feeling it! #soawesome.

The textual meta-function regards the exploitation of hashtags for organising the text in which the hashtag occurs, whereby the hashtag serves as a punctuation mark indicating that the hashtagged element is a piece of metadata. For example, in the case of (10), the element which is marked as metadata is #soawesome. It should be mentioned that hashtags can be separated from the rest of the tweet as in the #soawesome example, or they may be integrated into the tweet as in the case of #organic in (11) from our data:

  1. (11)

    #Organic popcorn. Every kernel pops! Yes, that is #organic butter and real sea salt

An important contribution of Zappavigna’s work (2015) is the attention it draws to the interactive nature of hashtags showing how they are used for building relationships with others as well as for construing stance. In this sense, insights from work on stance can be highly useful for the analysis of hashtag functions, and especially work by Du Bois’ (2007) who views stance–taking as interactive and dynamic.

Drawing on Zappavigna (2015), Laucuka (2018) identifies ten communicative hashtag functions, namely topic marking, aggregation, socialising, excuse, irony, expressing attitudes, providing metadata, initiating movements, propaganda and brand marketing (Laucuka, 2018, p. 56). As can be seen in the categories, Laucuka’s analysis contains two categories which are very closely related to the expression of stance, namely irony (when a hashtag contradicts the previous message) and expressing attitudes (which includes the opinions, feelings and assessments of the tweeter, e.g., #excited) (Laucuka, 2018, p. 61).

So far, the paper has stressed the multi-functionality of hashtags, their role in guiding readers to the intended interpretation, and their potential for expressing stance. When hashtags are used for instructing the reader as to how to interpret a tweet (e.g., #sarcasm), their function resembles labels (Francis, 1994) in the sense that they characterise tweet aspects, which is a form of stance. The next section thus describes labels and explains their relevance to the current study.

Hashtags as Labels

This section discusses the notion of labels (Francis, 1994), which can help account for certain aspects of hashtag use, such as the function of a hashtag based on its position in a tweet. According to Francis (1994), a label is a noun or noun phrase which refers to a preceding or following piece of discourse. Labels preceding the information they characterise are called advance labels, while those following it are called retrospective labels. Advance labels prepare the reader for an upcoming piece of information, while retrospective labels summarise a preceding message. The meaning of labels is unspecified and becomes more specific by means of the noun to which the label refers.

A relevant type of label is the metalinguistic label (Francis, 1994). Francis (1994, p. 83) explains that a metalinguistic label is a noun or noun phrase with unspecified meaning characterising a piece of discourse as a specific linguistic act. For instance, a writer might describe a message as an argument, or a statement. Moreover, she divides metalinguistic labels into four categories based on their meaning: (1) illocutionary metalinguistic labels (e.g., advice, recommendation), (2) language activity metalinguistic labels (e.g., myth, proof), (3) mental process metalinguistic labels (e.g., analysis, opinion), and (4) text metalinguistic labels (e.g., sentence, word).

As Francis (1994) explains, a metalinguistic label reflects the writer’s evaluation or interpretation of the information to which the label refers. At the same time, such labels invite the reader to accept the writer’s evaluation and interpret the text in the same way. As she points out, metalinguistic labels can also be exploited by writers for presenting two perspectives as equivalent, as in example (12) adapted from Francis (1994, p. 85):

  1. (12)

    The mouse antibodies did not remain in the system long enough to be fully effective. The second generation antibody now under development is an attempt to get round this problem.

The writer of (12) has used the label problem to describe the duration of the presence of antibodies as problematic, which reflects merely his/ her evaluation of this information. Since the label follows the information to which it refers, it is a retrospective label. A characteristic of retrospective labels is that their scope is not specified, which may be exploited by writers for persuasive and communicative purposes as this vagueness in terms of the referent of the label can give rise to numerous interpretations of the same message (Francis, 1994, p. 88).

In terms of the relevance of metalinguistic labels to the analysis of stance-conveying hashtags, metalinguistic labels can help explain certain aspects of hashtag use such as the function of a hashtag based on the position in which it occurs, i.e., whether is precedes the tweet utterance or follows it. More information about the application of the notion of metalinguistic label and hashtags will be provided in section "Results and discussion", but before that, section 4 describes the data used and the method followed in the study.

Data and method

The data used in the current study originate in a corpus compiled for the purposes of the LangTool projectFootnote 2 which examines consumer attitudes towards organic food products. The corpus consists of user generated content extracted from social media platforms (Twitter, Facebook, YouTube, Reddit, and Instagram). The data were extracted using various keywords related to the consumption of organic food (e.g., organic, organic food, eco-friendly, bio, etc.), and they contain user comments on the consumption of organic food. For the extraction process, the platforms’ APIs (when available) and automated web scraping methods were used. The extraction of the data was performed during the period from February to May 2019, and texts in any language (including any of the given keywords) from 2006 to 2019 were extracted. The corpus data were processed and the language of the posts was detected. For this study, we used English data from the Twitter part of the corpus (1,653,224 running words). In terms of the methodology, the tweets were obtained in .txt format and processed using AntConcFootnote 3(Anthony, 2022). First, a frequency word list was created for an overview of the most frequently occurring lexical items in the data set. Next, we searched for the most frequent hashtags in our sample using the command-line utility grep, which was necessitated by AntConc’s inability to process the # sign. We sorted the hashtags in order of frequency from highest, (i.e., 75,124 hits: #organic) to lowest, (i.e., 1 hit: #002). Since the focus of this article is on hashtags conveying stance, we only included hashtags which explicitly expressed a stance. For example, #organic was the most frequent hashtag, but it was not included in the analysis as it does not explicitly express a stance, whereas #healthy and #yummy were included as they do express an explicit stance (see Appendix A for an overview of the most frequent hashtags of the data set). Our decision to focus on hashtags explicitly expressing stance was made based on the assumption that the tweets containing them would constitute clearer cases of stance-taking, as in (13) where #healthy and #yum combine with other positive hashtags to express a positive stance:

  1. (13)

    so so pretty! #cherries #fruit #fresh #food #welcomesummer #vegan #organic #healthy #yum

The cut-off point for inclusion of a stance-conveying hashtag into the study was set at 290 hits, which resulted in the inclusion of sixteen (16) stance-conveying hashtags, shown in Table 1. The analysis involved investigating the first 60 tweets in which each hashtag occurred. The sample thus included 960 tweets (i.e., 16 x 60).

Table 1 The most frequent stance-conveying hashtags of the data set
Full size table

The 16 hashtags were used as the point of departure for our study of the function of stance-conveying hashtags. For example, one of the tweets containing #healthy was (14):

  1. (14)

    Don't worry pea happy! #vegan #plantbased #food #love #fitness #humour #humor #veganfood #healthyfood #healthy #health #organic.

Even though the focus was the use of #healthy, it was noted that #humor and #humour were used in order to draw attention to the word play in the tweet, i.e., don’t worry pea happy, in line with the expression don’t worry be happy. Consequently, such hashtags were included in the analysis of stance-conveying hashtag functions in the current study as it was clear that these hashtags also expressed a stance.

Since AntConc did not allow searching for hashtags, Notepad++Footnote 4 was used for close inspection of the hashtags in their contexts of use. The 960 tweets were manually analysed by one of the authors. This qualitative analysis of stance-taking hashtag functions considered the relationship between the hashtags and the content of the tweet, the interaction among the hashtags and the position of the hashtags in the tweet. In order to account for the function of a hashtag based on its position in the tweet, we applied Francis’ (1994) notion of metalinguistic labels discussed in section "Hashtags as labels". The tweets in (15) and (16) illustrate our reasoning in the analysis of the data. Consider (15).

  1. (15)

    #Tips: When using dried oregano crush it a little in your hand to release the essential oils and flavor. #Green #Organic #Food.

To identify the function of #Tips, attention was paid to the message expressed in the main part of the tweet following the first hashtag. More specifically, #Tips serves the pragmatic function of introducing and drawing attention to the ensuing new information. In this sense, it functions as an advance label as it prepares the reader for the upcoming information. Furthermore, the subsequent hashtags #Green #Organic #Food are not stance-taking hashtags but were retrieved since we made use of those seed words to retrieve the data. Those sequences indicated the topic of the post and have an ideational function in Zappavigna’s (2015) system of categories.

Tweet (16) shows a stance-taking tweet where the function is evoked as a result of the interaction of the various hashtags of the same tweet:

  1. (16)

    #Fake #food?! No, thank you! #nongmo #realfood #organic #healthychoices #yourbodyisatemple…

The arrangement of the hashtags in (16) communicates the message that what is signposted by the first two hashtags is a problem, which is followed by a sequence of contrastive hashtags expressing the solution to that problem. As (15) and (16) indicate, attention was paid to the contribution of hashtags and the text towards the final message of the tweet, and this was also the way we went about analysing the data in this study. After the qualitative analysis of the tweets in our sample by one of the authors, all three authors discussed the results of the analysis and agreed on the hashtag function categorisation. It should be noted that the categorisation is not necessarily exhaustive as it is based on these 960 tweets. Moreover, following both Du Bois’ (2007) approach to stance and Zappavigna’s (2015) analysis of hashtag functions, the hashtag categories proposed in the current study are not mutually exclusive.

Having described the methodology for retrieving the data in the first step of the qualitative analysis, the next section reports the results of the analysis, namely the scheme used for the categorisation of the stance-conveying hashtags and the annotation task that was performed to evaluate the scheme.

Results and discussion

In this section, we present the results of the analysis of the hashtags and suggest a categorisation scheme of stance-taking hashtag functions (Section "A categorisation scheme of the stance-conveying hashtags"). In Section "Evaluation of the categorisation scheme", we describe the annotation process of a sample of our data in order to evaluate the categorisation scheme presented in 5.1.

A Categorisation Scheme of the Stance-Conveying Hashtags

In this section we present the outcome of the analysis of the hashtags, offering a categorisation of stance-taking hashtag functions in the data. Each category is discussed along with examples in order to illustrate each function served by the hashtags. The study resulted in the identification of four stance-conveying functions, which are shown in Table 2.

Table 2 Stance-conveying hashtag functions identified in our data set
Full size table

Table 2 shows the four stance-conveying hashtag functions along with examples. The hashtag functions identified include both known and new hashtag functions. The new hashtag functions are epistemic (e.g., #truth, #fact, #lie) and deontic hashtags (e.g., #tip, #advice). As already mentioned, even though research on hashtag functions has discussed the role of hashtags in guiding the reader towards a given interpretation, i.e., category 4 (e.g., Wikström, 2014; Zappavigna, 2015), we decided to label the two subtypes of category 4 as deontic and epistemic in an effort to raise awareness of their systematic use and function in tweets. The group of stance-conveying hashtag functions identified is not intended to be exhaustive, but the functions are the most prominent and inclusive.

The hashtags included in category 1 involve the tweeter taking a stance and expressing an evaluation towards physical objects and/or abstract elements. In Du Bois’ (2007) terms, hashtag function 1 relates to evaluation. Category 2 hashtags describe the expression of the tweeter’s feelings. In this sense, they relate to positioning in Du Bois’ (2007) terms. Category 3 hashtags involve the tweeter directly addressing the reader by taking into account the reader’s stance, which indicates the dialogic and intersubjective aspect of stance-taking. These hashtags relate to the alignment facet of stance-taking as they foreground the relationship between two stances. Category 4 contains hashtags which direct the reader towards a specific tweet interpretation. In particular, category 4 hashtags indicate the preferred tweet interpretation reflecting the content of the tweet. Category 4.1 hashtags are a subtype of category 4 in the sense that they are deontic hashtags which often modify the strength of a recommendation given in the text accompanying the hashtag. Category 4.2 is the second sub-type and involves epistemic hashtags which often boost the credibility of a tweet. To the extent that hashtags in category 4 are used in order to indicate the intended tweet interpretation, they resemble metalinguistic labels (Francis, 1994). More information will be provided in the section about deontic and epistemic hashtags below. Finally, the categorisation proposed here is based on the assumption that each function foregrounds one aspect of stance-taking, i.e., evaluation, positioning, alignment, without excluding the other aspects, which is in line with Du Bois’ (2007) view of stance-taking.

  • Hashtag function 1: Taking stance towards physical objects/ abstract entities

The first category of stance-conveying hashtags involves hashtags used to express the tweeter’s evaluation of either concrete objects or abstract elements. Since hashtags express an evaluation, the first category can be said to foreground the evaluation facet of stance-taking (Du Bois, 2007), as in (17) and (18):

  1. (17)

    #Banana #latte #organic #smoothie #vegan #food #foodporn #yum #cafe #brunch #breakfast…

  2. (18)

    These tomatoes are blowing my mind! So #delicious #delish #organic #yum #yummy #food #germany

The tweeter in (17) uses hashtags to mark the topic of the tweet, i.e., #Banana #latte #smoothie. In addition, he/ she is also evaluating the taste of the banana latte smoothie mentioned at the beginning of the hashtag sequence by means of #yum. Similarly, the tweeter of (18) uses a number of hashtags, i.e., #delicious, #delish, #yum, #yummy, to positively evaluate the tomatoes mentioned in the first part of the tweet. In both (17) and (18), the tweeters express their stances towards the features of a concrete/ tangible object.

Category 1 stance-conveying hashtags can also be used for evaluating abstract entities, as in (19):

  1. (19)

    #Organic #food is such a #joke.

By using #joke, the tweeter in (19) expresses his/ her negative evaluation of the very concept of organic food. It should be noted that the negative evaluation expressed by #joke arises as a result of the context in which the hashtag is used, and it is supported by the exclamative such-a-construction. The hashtag is integrated into the text of the tweet, and so the tweeter’s attitude is expressed in a way that resembles natural speech. The hashtag has a syntactic function in the tweet and is not an added extension to the tweet as in the case of (18).

The next tweet differs from the previous tweets in expressing a negative evaluation as is shown in (20):

  1. (20)

    Spread the word about the dangers of #GMOs #horrible #Monsanto #food Eat #organic #LDKG

The tweeter in (20) is negatively evaluating genetically modified organisms by means of #horrible, encouraging the reader to eat organic food. The hashtag #horrible is consistent with the preceding part of the tweet which describes GMOs as dangerous. Furthermore, the example demonstrates how hashtags foregrounding the evaluation facet of stance-taking in Du Bois’ (2007) terms can also be employed for evaluating abstract elements since #horrible refers to the “dangers of #GMOs”, which is an abstract concept. Moreover, the tweet also shows the interdependence between hashtags and the tweet in their scope in the sense that they work together to produce the final message.

  • Hashtag function 2: Expressing the tweeter's feelings

Apart from expressing the tweeter’s attitude towards a concrete object or an abstract element, stance-conveying hashtags can be used for expressing the feelings of the tweeter, as in (21):

  1. (21)

    I fucking did it!!! My first #homemade #nutella, entirely #made with #organic #food!! I'm so #happy…

The writer of (21) has used #happy to explicitly state how he/ she feels as a result of the cooking described in the tweet. In this sense, #happy relates to positioning (Du Bois, 2007). The meaning expressed by the hashtag reflects the overall content of the utterance and the stance expressed by the first part of the tweet by means of the vocabulary and the exclamation marks employed. This illustrates how punctuation, the text of the tweet and the hashtags coordinate to create a specific message.

  • Hashtag function 3: Invoking the reader’s stance towards a stance object

Category 3 contains more explicitly dialogic hashtags, directly addressing the reader as in (22):

  1. (22)

    #animalcrackers #nutella #organic #junk #delicious #food #yumm #soooooogood #orgasmic #betyourejealous

The tweeter of (22) is evaluating animal crackers with Nutella using a sequence of hashtags for attributing positive evaluations to them, i.e., #delicious #yumm #soooooogood #orgasmic. The use of the four positive evaluative hashtags conveys the tweeter’s excitement, which is consistent with the use of #betyourejealous, which expresses the tweeter’s certainty about the reader’s stance. The hashtag #betyourejealous expresses the tweeter’s positioning on an epistemic scale. It indicates the tweeter’s high degree of confidence in his/ her prediction about the feelings of the reader by virtue of the verb bet in the hashtag. At the same time, #betyourejealous foregrounds the alignment facet of stance taking. As was discussed in section 2, alignment involves the stance of others, and, more specifically, how the current stance relates to the stance taken by others. In this example, #betyourejealous addresses the reader's stance as it is asserted by means of the hashtag that the reader is jealous probably because he/ she shares a positive stance towards the animal crackers mentioned in the tweet. This example reflects the dialogic nature of hashtags, as well as Hunston and Thompson’s (2000, p. 8) observation that evaluative language is used for spreading the assumptions of the speaker or writer. In this case, the hashtag appears to be inviting the reader to accept the writer’s evaluation about the animal crackers mentioned in the tweet.

Another tweet containing a stance-conveying hashtag foregrounding the alignment facet of stance-taking is tweet (23):

  1. (23)

    "#Mom? Why can't they just grow all #organic fruits & vegetables?" asks the 6yo DudeSter. #goodquestion #food #fiercelovebook

The tweeter is (23) is expressing a positive evaluation of the question asked by the 6 year-old child. He/ She uses #goodquestion to show he/ she finds it a relevant question to ask. In this sense, #goodquestion expresses the converging alignment between the stance of the tweeter and the 6 year-old mentioned in the tweet. This example reflects the complex nature of stance-taking and confirms Du Bois’ (2007) arguments for a view of stance-taking as comprising three interrelated facets.

  • Hashtag function 4: Indicating the preferred tweet interpretation

Category 4 contains hashtags which indicate the preferred interpretation of the tweet as in (24):

  1. (24)

    Lettuce talk about gardening...it's a big dill... #garden #humor #pun #fresh #food #organic #dill…

Example (24) illustrates how hashtags can draw attention to tweet elements which should not be missed by the reader. The tweet begins with a pun involving the use of lettuce instead of the phonetically similar let us for initiating a discussion on the topic of gardening followed by a second instance of wordplay between deal and dill. The text of the tweet is followed by a sequence of hashtags indicating not only the topic of the tweet, but also the humorous attitude of the writer by means of #humor, as well as the more specific #pun, which is used for pointing out the wordplay in the post. It thus draws the reader’s attention to the wordplay and signals to the reader that the different spellings are not errors, but conscious choices meant to create a pun. Additional instances of a hashtag instructing the reader as to how to interpret the tweet are found in examples (25) to (27):

  1. (25)

    Don't Fight w/words, Fight w/the #VOTE! #Warning Watch where you Eat & Drink! Make your Own #Food #Organic! Drink Bottled Water! #justsaying

  2. (26)

    Eat those greens. #food #health #tip #organic #didyouknow

  3. (27)

    #FACT #Wheatgrass helps improve blood circulation and prevents anemia. #health #body #greenblood #fitness #healthy #organic #natural #food

In tweet (25) the tweeter has used #Warning before the directive watch what you Eat & Drink!. The hashtag #Warning, on the one hand, instructs the reader as to how to interpret the ensuing piece of information while also preparing them for the upcoming message. In this sense, it functions as an advance label. Furthermore, the use of exclamation marks is also consistent with the hashtag as the punctuation serves to stress the importance of the message, which again shows how punctuation, text and hashtags combine to produce the final message.

Example (26) contains a directive “Eat those greens” which is characterised as a form of advice by the retrospective metalinguistic label #tip, whereas example (27) contains the advance metalinguistic label #FACT which prepares the reader for the upcoming information encouraging them to interpret the information of the tweet as factual. The use of bold typeface and block capitals also serves to construe the upcoming information as factual. A common feature of the hashtags discussed so far is that they point at elements which are present in the tweet and are consistent with the hashtag, e.g., the punctuation and the meaning of the tweet. However, it is possible that the hashtag used for indicating the intended interpretation of the tweet does not reflect aspects of the tweet but is used more strategically by tweeters. Our analysis revealed two such hashtag functions, i.e., deontic and epistemic hashtags, which are discussed in the remainder of this section.

  • Deontic hashtags

As was mentioned, this category is included in the category of hashtags which guide the reader towards the preferred tweet interpretation. However, it is a more specific type of hashtag in that deontic hashtags are used in order to present directives as being of a specific type. An example is shown in tweet (28):

  1. (28)

    in love with spirit 2.0 at #groenepassage #rotterdam. design & food = spot on! Definitemust go! #modern #organic #vegan #healthy #food #tip

Tweet (28) contains a positive review of a restaurant in Rotterdam along with a comment on the worthiness of a visit. The utterance is followed by a sequence of hashtags specifying the topic of the review, i.e., it is about the modern design of the place and the organic food it serves, e.g., #modern. The hashtag sequence culminates in #tip, which modifies the content of the tweet, thus leading the reader to interpret the tweet as a piece of advice. In this sense, #tip functions as a retrospective label as it attributes a characterisation to the entire preceding message. Moreover, the interpretation of the tweet as a piece of advice can be reached on the basis of the lexical items expressing a positive attitude, i.e., in love with, spot on, and Definitemust go!. Since the meaning expressed by the hashtag is some sort of advice, it can be called a deontic modality hashtag.

However, the use of #tip for describing the content of the tweet seems interesting for an additional reason. If #tip is to be interpreted as referring back to Definitemust go!, it appears that the writer’s commitment to the tweet information Definitemust go! is downplayed by #tip in the sense that the hashtag expresses a more neutral meaning compared to the imperative force of Definitemust go!. This is characterised by categorical modality as expressed by definite and must. In this sense, the deontic hashtag masks the various degrees of force involved in the different types of directives given in the tweet. The reader is thus instructed to interpret the entire tweet as a piece of advice despite the different degrees of force used in the tweet. The use of #tip thus seems to result in the presentation of non-neutral meaning as more neutral (hedged) in this case. The tweeter has strategically exploited #tip as a metalinguistic label in order to attribute a characterisation to the information of the tweet and to guide the reader’s interpretation. However, it should be noted that the tweeter may have used #tip at the end of the message for politeness-related reason. Footnote 5In particular, the use of #tip can render the tweet less face-threatening given the limited typing space offered on Twitter.

Another instance of a deontic hashtag is offered in tweet (29):

  1. (29)

    you have the right to #clean, #healthy #food! so buy #organic! it's SOOO worth it!:-) #change #advice

Specifically, #advice indicates that the tweet contains a piece of advice. The tweeter may have used this hashtag as it is well-known and likely to be recognised by other users.Footnote 6 However, in tweet (29) as well, there is a discrepancy between the style of the tweet and the hashtag used. In particular, the tweet contains a command followed by an exclamation mark, i.e., so buy #organic!, which increases the force of the utterance by adding emphasis to it. However, the retrospective metalinguistic label #advice presents the entire tweet as a more neutral piece of advice. In contrast to hashtags in the previous category, deontic hashtags are stance-conveying hashtags which are often strategically employed by tweeters. The examples we found contained deontic hashtags used in order to present strong recommendations as more neutral information.

  • Epistemic hashtags

Within the context of using hashtags for guiding the reader towards the intended interpretation of a tweet, hashtags can also be used for presenting subjective information as factual, as the following tweet shows:

  1. (30)

    Don't be fooled by fancy labeling, #organic #food can still make you fat. #fact #fitness #fitfampic

In the case of (30), the main utterance in the tweet contains a strong recommendation starting with a negated imperative form, i.e., Don't be fooled by, followed by a tentative statement, i.e., #organic #food can still make you fat. However, the subsequent hashtag sequence contains #fact, which functions as a metalinguistic label characterising the content of the main part of the tweet as a fact. Furthermore, #fact is a retrospective label, so the characterisation it adds to the message covers the text preceding the hashtag. Consequently, the reader is driven towards interpreting the entire tweet as containing objective information, while disguising the recommendation in the first part of the post, which can be misleading.

Similar instances are found in tweets (31) to (33):

  1. (31)

    Conventional Strawberry vs. Organic Strawberry! The difference is scary! #fact #organic #food #CANADA #healthypic

  2. (32)

    "#FACT #Organic #farmers and organic #food companies have nothing to hide. No GMOs or toxic persistent pesticides RT" exactly!

  3. (33)

    #FACT Certified #Organic farmers raise crops & livestock without toxic pesticides, antibiotics, or #GMO seeds

Tweet (31) contains a comparison between Conventional Strawberry and Organic Strawberry, followed by a statement about the difference between the two, modified by the adjective scary and accompanied by exclamation marks to show the affective positioning of the tweeter. However, the subsequent hashtag sequence contains hashtags specifying the tweet topic, i.e., #organic #food, as well as #fact, which describes the information presented in the post. It seems then that #fact has the epistemic function of presenting subjective information as objective. Interestingly, the tweet lacks any hedging, which also contributes to the presentation of subjective information as factual, which is not to be disputed, thus boosting the credibility of the information presented in the tweet. Again, since the hashtag functions as a retrospective metalinguistic label characterising a preceding piece of discourse as factual, the reader is encouraged to interpret the tweet as a fact, rather than as an opinion.

It should be noted that our analysis revealed several posts containing #fact which were also characterised by a lack of hedging and categorical modality markers as is shown in examples (32) and (33). The tweets in these examples begin with #FACT, which serves the pragmatic function of introducing the upcoming information. Its occurrence in utterance-initial position and the use of block capitals can be said to be boosting the credibility of the information presented in the tweet. Since the hashtag precedes the main part of the tweet and attributes a characterisation to it, it is an advance metalinguistic label preparing the reader and instructing him/ her to interpret the upcoming information as a fact.

Evaluation of the Categorisation Scheme

This section presents the evaluation of the stance-conveying hashtag functions that were identified and described in Section "Results and discussion", and it also discusses the results of our study. The exploratory study and the analysis of stance-conveying hashtags resulted in the categorisation scheme in Table 2. In spite of the fact that this is a qualitative study of the functions of hashtags for the expression of stance in tweets, we decided to carry out an annotation task in order to evaluate the validity and accuracy of the suggested categories. This evaluation process was carried out in two annotation rounds (pilot and final round), and we calculated the agreement between the annotators’ decisions. More specifically, Annotator A (first author) and Annotator B (second author) performed a pilot round of annotations (50 randomly selected tweets) in order, firstly, to confirm that the functions presented in Table 2 could be identified in and attributed to randomly selected tweets in the dataset, and, secondly, to discuss annotation strategies and resolve examples of conflicting cases. After the pilot round, the two annotators discussed their experiences and made decisions about the final annotation round: (i) to annotate and attribute a label to each hashtag in a tweet that has one of the functions in Table 2, (ii) to base their decision not only on the function that the hashtag conveys but also on the overall function of the tweet and the hashtag together and, (iii) to ignore the neutral hashtags of the tweet (if any). For this annotation round, 103 tweets (about 10% of the dataset) were randomly extracted from the dataset and annotated by both annotators separately. In this set of tweets, at least one of the annotators attributed a stance label to 201 hashtags, and 29 tweets were considered as neutral by both annotators (e.g., Policy World Champions to-be: find out more about the world’s hottest candidates for this year’s #FuturePolicyAward on #Agroecology! @sekemgrouppic., #colorful #pasta #carrots #organic #glutenfree #coeliac #organicfood #foodpics #food #instafood…). Inter-annotator agreement was then measured, and Gwet’s AC1 coefficient (Gwet, 2002), which is a metric used in our previous work (Simaki et al., 2022) to evaluate annotations of discourse data, was calculated at 0.89. This score represents a very high level of agreement and ensures reliable annotations of high quality (Simaki et al., 2022).

Our annotation results showed that the categories shown in Table 2 can be the basis of the identification of the functions of stance-conveying hashtags. The most frequent functions in the annotated data are functions 1(Taking a stance towards physical objects/abstract entities) and 4.1 (deontic hashtags), while functions 3(Expressing the tweeter’s feelings) and 4.2 (epistemic hashtags) were less frequent. In the randomly selected tweets for the annotation process, no hashtags were identified having function 3 (Guessing/invoking the tweet reader’s stance), which shows that this function is rare in the data. The next section summarises our findings and concludes the paper.

Conclusion

The present study examined tweets about organic food aiming to analyse how tweeters take a stance using hashtags and how they instruct readers to interpret the tweet in a specific way. The analysis was qualitative and forms the first step towards the creation of a taxonomy of stance-conveying hashtags which can subsequently be used in other data sets to investigate how general the taxonomy is. Our study has been based on the acknowledgement that a main function of hashtags is to express stance (Laucuka, 2018; Lee, 2018; Wikström, 2014; Zappavigna, 2015), and we have shown how stance-conveying hashtags foreground different aspects of the stance-taking event, i.e., evaluation, positioning and alignment.

The study used the sixteen most frequent stance-conveying hashtags in a large data set, using data from Twitter for the analysis. The analysis resulted in the identification of four stance-conveying hashtag function categories, namely 1) taking a stance towards physical objects/ abstract entities, 2) expressing the tweeter's feelings, 3) guessing/ invoking the tweet reader's stance, and 4) indicating the intended tweet interpretation, which includes 4.1) presenting a directive as being of a specific type (deontic hashtags) and 4.2) commenting on the epistemic status of a message (epistemic hashtags). Moreover, the analysis highlighted the strategic use of deontic hashtags, which often present directives as neutral information, and epistemic hashtags, which often present subjective information as factual. The use of hashtags which indicate the intended tweet interpretation was found to resemble metalinguistic labels (Francis, 1994). The study emphasised the potential of stance-conveying hashtags to instruct readers to interpret a tweet in a particular way, which may lead them to ignore certain other aspects of the tweet which may have been masked by the hashtag. Explicating the operation of deontic and epistemic hashtags thus draws attention to potential mismatches between the tweet and the hashtag in its scope, which can raise awareness of elements which may be ignored as a result of the operation of such hashtags.

In order to evaluate the stance-conveying hashtag taxonomy, we annotated part of the dataset and calculated the intercoder reliability. The high agreement score (0.89) between the two coders indicates the validity of the suggested functions in the given context. The number of tweets analysed for this study was limited to approximately 1000 occurrences, so future work could address the same question using a larger set of tweets searching for instances of hashtags serving less frequent functions (e.g., guessing the reader’s stance and epistemic hashtags), or it could identify new functions that could possibly enrich our taxonomy. Finally, future studies could apply the current taxonomy to different data sets to evaluate the generalisability of the proposed taxonomy.