With the help of Wessel Stoop of the Humanities Lab of Radboud University, at least 200 of the most recent tweets were collected for each of the terms. With the approval of Twitter, tweets were collected from an initially 30-day period before the start of the investigation (which was October 2021), extended to a longer period if the initial search did not yield enough relevant occurrences. We2 manually eliminated irrelevant occurrences (e.g. occurrences where the word was not used as a noun) until we had 200 relevant uses for each of the terms.

For each term, the 200 tweets were annotated by two research assistants. Occasionally, the wider context of a tweet was taken into account (if the tweet was still available online), when the message itself did not contain enough information. At least 20 tweets per category were annotated together to sharpen the categories and at least 60 tweets were annotated by both annotators independently to calculate the inter annotator agreement. For the terms that turned out to be more difficult to classify, an extra set of occurrences was doubly annotated. All tweets were finally also considered by the author and if disagreement arose, this classification was reconsidered by the author and one of the assistants until they reached an agreement. A small number of tweets was replaced in this last round because on closer inspection they turned out to fall outside the predefined selection criteria. The newly selected tweets were considered by the author and the research assistant.

The purpose of the study was to make an inventory of the ways each of the terms were used and with what frequency to explore whether a clear categorization of terms can be identified in their patterns of usage, such that, e.g. neutral terms can be distinguished from non-neutral terms. As, on the one hand, we aimed for an empirically driven annotation of the words, but on the other hand needed classification criteria in order to compare the various terms, we decided to make use of the categories proposed by Bartlett et al. (2014). Their study on the use of English slurs is one the few empirical and quantitative studies that have been done on the topic, and as they also used Twitter as a source of data, their categories are likely to be applicable for our purposes. Since Bartlett et al. (2014) only investigated slurs and our set of terms also includes other words, however, we had to adapt and extend the categories to a certain degree. In total, we used seven categories. Two of the categories, however, occurred very infrequently and will therefore not be discussed in the remainder of the text.3 The remaining five categories are explained below.

Negative Context (NC): A term was categorized as NC if it was used in a negative context or otherwise in a negative way, as in (4). Our initial annotation included two categories that, based on Bartlett et al. (2014), distinguished between occurrences that expressed a negative stereotypical attitude and occurrences that were used in a derogative manner but without explicit stereotypical references (after Bartlett et al.’s (2014) category Casual Use). The distinction between the two categories seemed too fuzzy for our data in the end. A term was also classified as NC if it was used to insult a third person, as in (5). In the latter case it was often difficult to determine whether the person insulted belonged to the target group or whether the term was used as a general swear word.

Targeted Use (TU): A term was labelled TU if the term was directly addressed to someone, as in (6). Bartlett et al. (2014) used the category Targeted Abuse, but we also categorized uses as such if the term was not necessarily used in a clearly negative manner as in (7).

When a tweet consisted only of the relevant term and was directed at someone (with the @-sign), we annotated the terms as TU unless there was a reason to annotate it otherwise.

Non derogatory (ND): following Bartlett et al. (2014), a term was classified as ND if it was used in a descriptive or otherwise neutral way, or if a stereotype was applied but not in a hurtful or derogatory manner, as in (8).

Mention (ME): A term was classified as ME if the term was not used but mentioned, so the term was used in a meta-linguistic manner, as in (9). This category was not applied by Bartlett et al. (2014).

Negative Trait (NT): As we also included terms that did not (necessarily) refer to a particular (social) group but are rather used to ascribe a negative property to someone as in (10), we also included the category NT. The difference between NC and NT is that NT describes a trait while NC does not.

In this section, the most relevant findings and observations will be discussed.5 Not all numbers will be presented, and the results should be regarded with caution for reasons discussed in the next section, but I will point out a number of interesting first observations.

Let us first look at the so-called neutral words: Oostenrijker, homo and Turk. For reasons of readability, I only included the three most informative categories in Figure 1: Negative Context, Non-Derogatory and Targeted Use. Note that the occurrences were always classified in only one of the categories.

As expected, Oostenrijker is almost exclusively used in a neutral way (97%), as in (11). What was noticeable with respect to Oostenrijker was that quite frequently it was used to refer to Hitler, as in example (12). Those occurrences were also annotated as Non-Derogatory, as the term Oostenrijker itself does not seem to be used in a negative manner.

For homo, we already see that there are more uses in negative contexts (11,5%) and also a number of targeted uses, although there were only three of the latter category. Examples (13) to (15) illustrate the three categories.

For Turk we see more occurrences in a negative context (43%) and more occurrences of targeted use (8%). These are examples of each of the categories.

In Figure 2, the slurs flikker, mietje and geitenneuker, and the swear word eikel are represented. For these terms, it is most informative to look at the categories Negative Context, Negative Trait and Targeted Use.

What seems noticeable is that flikker and geitenneuker have no occurrences in the category Negative Trait. However, this is due to the fact that we did not consider this to be a possible category for these terms as it would be unclear what negative trait they would indicate (see Section 3). What is noticeable is that the supposed slurs mietje and flikker are more frequently used in a targeted manner (32,5% and 36% respectively) than the swear word eikel (26%).

For flikker, Targeted Use, as exemplified in (19) is the most frequent category.

The second most frequent category was Negative Context (32%). Recall that under this category we also counted cases in which a third person is insulted, called a name, as in (20).

This was very frequently the case. We had the impression that most of the time, no reference was made to homosexuality both for the category Negative Context and for the category Targeted Use, but this was typically very difficult to know for sure.

For mietje, Negative Trait was the most frequent category (43,5%). In example (21), Lex refers to the Dutch king Willem-Alexander and the tweet alludes to an incident in which the king and his family flew to their summer house in Greece in the middle of a lockdown and returned to the Netherlands after nationwide criticism.

In such a case the term was not used to indicate that someone is a homosexual but that he has a negative quality, namely that he is cowardly. Only in a very limited amount of cases it was clear that the term did indicate that the person talked about is a homosexual, as in (22). In such cases the term was classified as NC (5%).

If the tweet was about a public figure who is known to be homosexual, we annotated the term as NC when reference was made to their homosexuality, and as NT when this was not the case. A number of tweets for example related to an incident in which an openly homosexual member of parliament pressed charges against a group of farmers who visited him at his home address as an action of protest. The word mietje seemed to be used as qualification of this in the user’s eye cowardly act. At the same time, it is probably not coincidental that precisely this word was chosen, and sometimes explicit reference to his homosexuality was made as well. It shows that a clear distinction between categories is difficult to make, and that the categories are indeed fuzzy and not mutually exclusive. I will come back to this in the next section.

Eikel was more often used to describe someone as in (23) than to insult someone directly as in (24) (65,5% versus 26%). We categorized occurrences of eikel attributed to a third person as negative trait, the trait being something like rude or impertinent. I will also come back to this in Section 3.

What was noticeable with respect to geitenneuker was that it was quite often used to refer specifically to Erdoǧan, the president of Turkey, as in (25). What was furthermore noticeable was that this term, more often than mietje and flikker, occurred in a context that was clearly related to the target group as in (26), and intuitively the tweets with geitenneuker were also the most hateful ones. Both were annotated as Negative Context (58,5%).

Geitenneuker was used in a targeted manner in 21,5% of the cases, as in (27).

When we look at neger, finally, I included the categories Mention, Non-Derogatory, Negative Context and Targeted Use in Figure 3.

What is noticeable is the high frequency of the category Mention (22,5%), so people were frequently talking about the word instead of using it. What is furthermore noticeable is that it is more frequently used in a negative context than in a neutral context (40,5 versus 7,5%) and that it is not so frequently used in a targeted manner (2,5%). In that sense, the word is more like Turk than it is like flikker or mietje. Examples (28) to (31) illustrate the four categories.

It is commonly acknowledged that words may gain expressive or social meaning and lose descriptive meaning through a process of semantic change or pragmaticalization (e.g. Foolen 1997). Gutzmann & Davis (2015) for example argue for a three-stage process in which expressions with mixed content serve as a bridge between originally descriptive expressions and their expressive counterparts. The move from stage one to stage two takes place through the conventionalization of conversational implicatures, thus becoming conventional implicatures, after which the descriptive content may be lost altogether in stage three due to semantic bleaching. It is also typically acknowledged that this process of conventionalization of pragmatic inferences is determined by frequency of use. As Gutzmann & Davis (2015: 203) put it: “Given a sufficiently high frequency, these inferences may be conventionalized and become part of an expression’s lexical content, so that they do not need to be derived on conversational grounds anymore”. Schmid (2015) distinguishes between entrenchment, which takes place in individual speakers, and conventionalization, a process which takes place in communities or social groups. Schmid (2015) argues that repeated processing of identical form meaning pairings lead to routinization of symbolic association or to context-free entrenchment, i.e., the more a word co-occurs with a certain meaning, the less context is needed to activate this meaning. This routinization of association is a gradual process, meaning that expressive content may be entrenched and eventually conventionalized to a certain degree. During this process, “usage affects both entrenchment and conventionalization, while entrenchment and conventionalization in turn influence usage” (Schmid 2015: 7).

In the data presented in Section 2, some terms were much more frequently uttered in a negative context than others. The word Turk, for example, more frequently occurs in a negative than in a neutral context, while Oostenrijker occurs basically only in neutral contexts. What is furthermore interesting, is that it seems that the more often a term is used in a negative context, the more frequent the term is used in a targeted manner as well. Turk, for example, was used in a targeted manner in 8% of the cases. Utterances we attested like Hypocritical Turk! or Sick Turk! or just Turk!, but also, for example, Homo! or F***ing neger, shut up!, are obviously expressive. The adjectives of course add to the expressive meaning of the utterance, and the broader context (if applicable) is typically not very complimentary as well, but the question arises why the writers of such tweets would use or add the terms Turk, homo or neger if those terms did not by themselves convey expressive content, i.e. add to the expressive force of the utterance, as well. It seems therefore that these terms do not always depend on the context to receive an expressive interpretation. On the other end of the scale we see that flikker is mostly used in a targeted manner (more so than the assumed swear word eikel), while for example geitenneuker is more often used in a negative context. Following the abovementioned hypothesized relation between targeted use and expressive content, this might suggest that flikker is in the process of losing its descriptive meaning.

Yet, there is a clear difference between a term like geitenneuker or flikker on the one hand and a term like Turk or homo on the other hand, namely that, unlike geitenneuker, Turk and homo can get a (seemingly) neutral interpretation. Gutzmann (2019: 11) makes a strict distinction between terms that do and terms that do not “have the expressive function hardwired into their lexicon”. A similar strict distinction between conversationally implicated and lexically expressed negative attitude is made by Camp (2013), who argues that a hearer of (32) might suspect that the speaker holds a racist attitude towards Hispanics as he finds this property of the candidate worth mentioning. She furthermore adds that the use of a noun instead of an adjective adds to this suspicion (Camp 2013: 340, original numbering (12)).

Camp stresses, however, that such an attitude is at most implicated and can be consistently denied by the speaker. Deniability or cancellability is often taken to be a crucial characteristic that differentiates between conversational and conventional implicatures (e.g. Potts 2005). Henderson & McCready (2019: 3) argue that the deniability argument “gets at the heart of what it means for content to be conventional”. Often, deniability is motivated by means of a constructed example, as in Henderson & McCready (2019), who illustrate the non-deniability of the pejorative content of the slur Kraut with (33) (Henderson & McCready 2019: 3, original numbering (7)):

The question arises whether deniability is indeed an all or nothing quality of meaning. An increasing degree of conventionalization (or routinization or entrenchment) of expressive meaning may make it harder to plausibly deny a derogative attitude towards the group referred to. A loss of descriptive content—as, for example, seems to hold for flikker and for mietje (whose descriptive content seems to narrow down to a specific property)—on the other hand, may make it easier to deny a negative attitude towards the group (previously) associated with the term. Sternau et al. (2017) experimentally show that the deniability of various types of inferences in other domains (e.g. fragment completion and conjunction enrichments) indeed seems to be a matter of degree. Whether the same holds for the deniability of pejorative content of the terms under consideration is an empirical question. If the deniability and therefore the conventionality of expressive content is indeed best viewed as a gradual property, what would be the consequences for the theoretical analysis of slurs and the related terms discussed in this paper? This will be addressed in the next subsection.

If indeed the expressive meaning of a term, such as Turk, becomes less easily deniable and less dependent on the context, perhaps to a degree in which it is standardly activated and only cancelable or deniable in specific circumstances, how should this aspect of its meaning be characterized? One option would be to qualify it as a generalized conversational implicature. Generalized implicatures arise “in the absence of special circumstances” (Grice 1989: 37). According to Potts (2005), generalized conversational implicatures, like other conversational implicatures, differ from conventional implicatures in that they are calculable based on Grice’s (1975) conversational maxims, and that, consequently, they are not non-detachable, i.e. they are not tied to a specific choice of words. In this respect it is interesting to consider the aforementioned observation by Camp (2013) that nouns may give rise to negative inferences to a stronger degree than adjectives. This suggests that there must be something inherent to their meaning that causes this. In line with Camp’s (2013) suggestion, Carnaghi et al. (2008) experimentally show that people ascribe more stereotypical properties to someone when this person is described with a noun, for example as a homosexual than if a person is described with an adjective, as homosexual. Hogeweg & Neuleman (2022) likewise show that some supposedly (relatively) neutral nouns are considered to be less hurtful than their corresponding slurs while they are considered to be more hurtful than their corresponding adjective noun combinations (while offered in the same, neutral context).

Recanati (2003) describes an additional type of implicature, a default implicature, that on the one hand differs from (generalized) conversational implicatures in that they are not generated through global inferences, but are conventionally associated with certain linguistic elements, and on the other hand differ from conventional implicatures in that they are cancelable. This seems to be an adequate description of the above hypothesized interpretation of a word like Turk. As, like conventional implicatures, these default implicatures are conventionally tied to particular expressions, Recanati (2003) argues that they can be considered semantic rather than pragmatic. This is in line with what could be called ‘rich’ approaches of the lexicon in which words contribute rich semantic representations to the process of interpretation and in which meaning aspects do not necessarily have to be non-defeasible to be part of the lexical, semantic meaning of words (e.g. Hogeweg & Vicente 2020). Zeevat et al. (2017), for example, describe lexical meanings as distributions over bundles of (moderately universal) semantic features. These feature sets can be specified by a number of operations, such that some features may be absolute, while others are e.g. mutually exclusive. A similar approach might also be applicable to the loss of descriptive content, as seems to (partly) be the case for flikker. In this case, it is the descriptive reading that becomes less conventional and more dependent on the context.

Independently of the precise analysis or categorization of the various terms, my main point is that it may not be that easy to make a strict distinction between terms that do and terms that do not ‘have the expressive function hardwired into their lexicon’ and between terms that do or do not (still) have descriptive content. The three stages of pragmaticalization described by Gutzmann & Davis (2015) may therefore best be seen as points in a continuous process. This does not mean that there is no distinguishable class of words that fit all the criteria commonly ascribed to slurs, but that such terms are part of a continuum of (more or less) pejorative terms for (groups of) people.

A gradual view on conventionality in the lexical domain would also pave the way for a less fundamental distinction between expressive and social meaning. Henderson & McCready (2019) argue that unlike slurs, dog whistles (i.e., the phenomenon whereby a speaker sends one message to an outgroup and a second, often controversial message to an ingroup) are not properly analyzed in terms of conventional implicature. An important argument against a conventional implicature account is that the ingroup message can be plausibly denied by a speaker. Henderson & McCready (2019) instead analyze dog whistles as conveying social meaning; they signal a certain persona. Interestingly, Nunberg (2018) provides a similar analysis for slurs. Nunberg (2018) argues against an analysis of slurs in terms of conventional implicature and claims that slurs get their offensive nature through speakers’ exploitation of sociolinguistic variation and their affiliation with speakers who ‘own the word’. The question arises whether there is a need for such a strict opposition between conventional implicature and social meaning. Indeed, Eckert (2019: 769) argues for an integrated view on meaning in which “sociolinguistic variation contributes a purely performative, subtle, speaker-indexical resource, ranging from social category membership to momentary affective states.” As such, social meaning and expressive meaning conveyed through conventional implicature are not fundamentally different types of meaning but they form “a cline of ‘interiority’ from variables that index public social facts about the speaker to more internal, personal affective states.” (Eckert 2019: 751).

Based on the observations with respect to Dutch slurs and related words presented in this paper, I suggest that slurs are part of a larger collection of pejorative or expressive terms for (groups of) people, and that individual terms within this collection can show mixtures and degrees of descriptive, expressive and social meaning. Instead of aiming for ‘a theory of slurs’, it might therefore be more useful to investigate individual or subclasses of terms to find out the range of combinations of meaning they can express and the way they are interdependent and interact. A good example of such an analysis is Burnett’s (2020) analysis of the terms d*** and lesbian. Burnett shows that d*** and lesbian are not descriptively equivalent but that they are associated with different personae; lesbian is associated with a mainstream persona, and d*** is associated with what Burnett describes as an anti-mainstream persona. The appreciation for the different personae differs between people having different ideologies. The interpretation of the terms is then determined by an interaction between the meaning of the term in the listener’s own ideology and their expectations about the speaker’s ideological structure. Burnett (2020: 57) suggests the possibility that a similar analysis would be helpful for other terms that have been analyzed “under the umbrella of the term ‘slur’”, such as f** or slut, but also for ‘non-derogatory’ terms for women as a good woman or a nice young lady.