On sisters and zussen: integrating semasiological and onomasiological perspectives on the use of English person-reference nouns in Belgian-Dutch teenage chat messages

4.1.1 Initial identification

PRN are here defined as common nouns that can be used to refer to people in at least one of their meanings (Aleksandrova 2014: 2963; Zenner et al. 2012: 758), see (1)–(4). Even in bleached contexts such as (2), the person-reference function still holds to at least some extent, with ‘man’ addressing the hearer. Given issues with spelling variations in teenage chat language (Verheijen 2015) and the general difficulty of retrieving PRN automatically (Elmiger 2018: 185), the corpus was screened for PRN through manual inspection by three of the four authors.

4.1.2 Saturation

Rather than screening all 456,751 instant messages in the corpus, we searched tranches of 1,000 randomly selected utterances until saturation was neared. These utterances were selected at random with Python scripts, using the library ‘random’ (Python Software Foundation). This saturation approach follows Geeraerts et al. (1994: 32), who state that when a database “is a fairly representative reflection of the actual situation, the relative increase in the number of lexical types that are added to the material with each new portion of records should be low, or rather, it should diminish with the growth of the database.” Figure 1 shows the effective progressive saturation of PRN in our database, showing the number of new items found per tranche of 1,000 posts. In this stage, inflected forms of a PRN already included (e.g. leerlingen ‘pupils’ for leerling ‘pupil’) in the database were not considered ‘new’. Figure 1 shows the anticipated downward trend: where 55 new PRN were found for the first 1,000 posts, only 7 new nouns occurring in the final tranche, at which point we ended the sampling procedure.

Figure 1:

Progressive saturation of newly attested PRN.

4.1.3 Exhaustive retrieval

The method described above yielded an initial set of 335 different PRN, retrieved from 15,000 randomly selected screened chat posts. As a next step, all occurrences of these PRN were retrieved from the entire corpus of 456,751 chat messages, taking the spelling variation so typical of youth language CMC (Verheijen 2015) into account. This second step was manual and sample-determined: the corpus-broad retrieval included all spelling variants (both typos and deliberate [genre-related] spelling manipulations, e.g. guuurl for girl) and derivational alternatives (female forms such as klasgenote ‘female classmate’ or diminutives such as vriendje ‘little friend’) that we attested in the initially screened 15,000 posts. Given the semantic differences between derivational alternatives (vriendje typically does not mean ‘little friend’, but rather ‘boyfriend’), they were considered as separate entries for the next steps of the procedure. As such, the retrieval phase resulted in a set of 583 word forms subject to word-sense disambiguation and language coding. Finally, note that multiword PRN phrases were extracted and thus also counted more than once in cases where they contained multiple PRNs: the phrase sister from another mother was for instance retrieved and counted twice, once as an occurrence of sister, and a second time as an occurrence of mother.

4.1.4 Word sense disambiguation

To arrive at reliable counts we needed to discard noise, viz. (i) tokens that are used in a non-person-reference meaning, compare e.g. (5) that includes lief as person-reference noun (‘a boyfriend/girlfriend’) and (6) that includes lief as adjective (‘nice’); (ii) PRN tokens that are rather used as (part of) a proper name, e.g. (7) below.^[2] Disambiguation was conducted in one of three ways: first, for word forms where no semantic ambiguity was foreseen (N = 392 of 583 word forms) no manual check was conducted (see [1] above); second, for word forms with foreseen semantic ambiguity and with a total frequency under 200 in the entire corpus (N = 173 of 583 word forms), we proceeded with manual disambiguation for all tokens, only retaining true PRN occurrences; third, for word forms with foreseen semantic ambiguity and with a total frequency over 200 (N = 18 of 583 word forms), random samples of 200 cases were disambiguated, using the proportion of true PRN-meanings as a proxy for the token counts (see below).^[3]

4.1.5 Language coding

Crucial for the current purposes is to provide each PRN with a language code. Overall, we adhered to the principle stipulated by Geeraerts and Grondelaers (2000: 56) that “the non-Dutch character of a word can only exert influence on the language user’s behavior when the expression at issue is identifiable as a non-Dutch word”. This principle was operationalized in two ways: first, by considering the graphemes of the word, with for instance word-initial c or vowel-y-combinations as indicators of English origin; second, by considering the extent to which the grapheme/phoneme mapping of the word deviates from Dutch rules (Onysko 2007: 10; Zenner et al. 2012: 759), as is for instance the case for date, which would sound more like /’daːtə/ than /deɪt/ following a naïve Dutch pronunciation. Language coding was performed by three linguists, with discussion to reach agreement where necessary. Overall, the majority of non-Dutch words are English. Input from other languages is limited to 10 word forms, each with a low frequency: familia (N = 1), señor (N = 1), puto (N = 2), and putaa (N = 1) from Spanish; famille (N = 1), putain (N = 4) and ptnnnnnn (N = 1) from French; habiba (N = 1) and biba (N = 1) from Arabic; and senhor (N = 1) from Portuguese.

The remainder of the discussion will hence target the comparison between English and Dutch forms. Here, a total of eight ambiguous cases remains. First, for three word forms (fan, gangster and stalker), both a Belgian-Dutch and English pronunciation are possible, and it could be the case that the English origin is no longer salient for some language users who rely on the Dutch pronunciation. Second, five word forms (maker(s), man, mate, model, racist) are meaningful person-reference nouns in English and in Dutch, and the context of use often does not provide sufficient cues for reliable disambiguation. These eight word forms received the code ‘ambiguous’.

4.1.6 Lemmatization

Before proceeding to counts of the number of English and Dutch person-reference nouns, a fairly conservative lemmatization of the word forms was conducted. Given the importance of gender for PRN (see Section 1.3), forms with male versus female referents (e.g. medewerker resp. medewerkster, ‘collaborator’) were lemmatized separately; plural forms were lemmatized under the male forms (e.g. medewerkers ‘collaborators’, was lemmatized under the single male form medewerker), except if the female form itself was pluralized (e.g. medewerksters was lemmatized under medewerkster ‘female collaborator’);^[4] typos, spelling errors and (deliberate) genre-related spelling manipulations were lemmatized in the base form except in case of alterations or abbreviations that impact pronunciation (e.g. bae was lemmatized separately from baby, but babyyy was not).

4.1.7 Frequency counts

Type and token counts were calculated at the level of the lemmas. This means that the token counts for different word forms were added up per lemma. For word forms with less than 200 occurrences in the total corpus, the token count equals all occurrences (following manual disambiguation where needed). For word forms with over 200 tokens in the corpus, the proportion of ‘correct’ forms in a manually coded sample of 200 randomly selected occurrences was used to estimate the total number of tokens in the corpus. The lemma lief for example encompasses the word forms lief and lieff. The word form lief occurs 1,120 times in the corpus. Our random sample of 200 manually disambiguated tokens contains 45 true positives, where lief is a PRN, meaning ‘lover’, see (5) above, and 155 false hits or false positives, where lief serves as an adjective meaning ‘sweet’, see (6), which means that we find the target PRN meaning in 22.5 % of the cases. Extrapolating this proportion to the full corpus gives us an estimated token count of 22.5 % of 1,120, being 252. Taken together with the token frequency of the other word forms that are part of the lemma (in this case N = 1 for lieff), we arrive at the final token count of 253 for the lemma lief.^[5]

4.1.8 Results

Table 1 provides type and token counts for the Dutch and English person-reference nouns found in our corpus, complemented by the five most frequently attested types per language. The same information is also provided for the eight ambiguous cases.

Table 1 allows us to address our first research question on the overall distribution of English versus Dutch in the semantic field of PRN in teenage chat language in terms of types and tokens. First, it appears that English has a stronger position in terms of the number of types that occur in our chat corpus than in the number of tokens. Where nearly one out of five PRN types is coded as English, these English forms by estimate take up less than one out of ten PRN tokens that occur in the corpus. So, while the Flemish teenagers’ PRN repertoire does include quite a lot of English forms, these English forms are not used very frequently (compared to the Dutch forms). This can also be derived from the type/token ratio, which is highest for English (indicating types that are used relatively less often) and lowest for Dutch (indicating types that are used more frequently). Second, overall, Dutch clearly dominates in this semantic field, both for types and for tokens. Third, the top five most frequent nouns show both similarities and differences: both top five lists include a word to refer to young men (Dutch jongen and English boy), but the strong position of the Dutch kinship term mama ‘mom’ is not mirrored in the English list. By contrast, where the English list contains the negative evaluative term bitch, no at face-value evaluative nouns are found in the Dutch top five. This calls for further scrutiny to assess whether the general distribution of English and Dutch as found in Table 1 might vary when zooming in on subareas of the semantic field of PRN.

4.2.1 Challenges

As was discussed in Section 1.3, it is far from straightforward to arrive at a fit-for-all subclassification of PRN. Issues include (i) the broad range of features that are used to characterize people (Schnedecker 2018: 19); (ii) the overlap of features in family-resemblance patterns (Dahlgren 1985: 381); and (iii) the polysemous nature of many PRN (Kiesling 2004). For instance, boy can be categorized both with respect to the age of the referent and with respect to the gender of the referent, and is in that sense affiliated to other person-reference nouns that refer to people based on their age (like senior), their gender (like man) or both (like lad). Further, words like acteur ‘actor’ can be used to refer literally to someone’s profession, but this profession can also be cued only metaphorically to refer to someone posing or being insincere. When attempting to provide any structure to the broad semantic field of PRN, it is crucial to factor in these challenges when devising coding categories and coding principles. At the same time, the goal of this second level of granularity is to refrain from fine-grained semantic annotation at the level of individual tokens as this would imply a conflation of RQ2 and RQ3. The empirical challenge then becomes to devise a strategy to identify core semantic features of a set of person-reference nouns at the type level, which is only possible (i) when the semantic features under scrutiny are sufficiently broad to be reliably identified without resorting to contextual nuances; (ii) when coding is sufficiently conservative, i.e. when coders only identify features considered intrinsically part of the core meaning of a PRN, here defined as the original person-reference meaning of the noun. Even then, of course, any approach at this level of aggregation will suffer from inescapable drawbacks: coding semantic features at type-level can never capture all semantic nuances, which further supports the necessity of the RQ3-analysis introduced further down.

4.2.2 Semantic features

In tackling this challenge, we generally adhere to the principle of feature annotation as used in lexical semantics (Geeraerts et al. 1994), defining a number of semantic features that can or cannot be indexed by a particular PRN-type. Combining a top-down literature review of existing subcategorizations of PRN (Section 1.3) and a bottom-up scrutiny of the PRN in our corpus, we arrived at nine broad semantic features to be coded for four encompassing categories.

The first two features, ‘Age’ and ‘Gender’, pertain to the perceived general demographic characteristics of the person referred to (compare Braun 1997; Elmiger 2018; Schnedecker and Aleksandrova 2016). The feature ‘Activity’ relates to actions undertaken by the person referred to and can hence be related to Baeskow (2002)’s nomina agentis (see also Schnedecker and Aleksandrova 2016), and to the subfield ‘professions and occupations’ found in Braun (1997), Dahlgren (1985), and Flaux and Van de Velde (2000). Next, the features ‘Kinship’, ‘Collectives’, ‘School’, ‘Hierarchy’ and ‘Friendship’ all capture the way in which the person referred to is seen in relation to (networks of) other people; see e.g. Dalhgren (1985) and Braun (1997) for the general importance of relations in classifications of people, and Elmiger (2018) and Joosten et al. (2007) on collectives. Finally, supported by Gosselin (2018)’s insistence on the importance of subjectivity in person-reference nouns and sociolinguistic work underlining their socio-indexical potential, we identify the feature ‘Evaluation’.

4.2.3 Coding principles

To avoid overrepresentation of low frequent types, only PRN that occur minimally ten times in the corpus are subject to the coding procedure. This leaves us with 214 nouns for subcategorization, which were each coded independently by three coders for the nine categories. Coding was binary, indicating whether a particular semantic feature is indexed by a particular noun (code ‘yes’) or not (code ‘no). To allow for the level of aggregation desired in addressing RQ2, coding was blind to context, coarse-grained, and conservative.

First, given the prevalence of polysemy and additional contextual meanings in the use of person-reference nouns (e.g. the literal or figurative uses of actor), particularly in the context of youth language, and the resulting slippery slope in delineating these (sub)senses, the choice was made to code semantic features at the type-level rather than at the token-level. For each of the 214 types, semantic-feature annotation was then conducted for the core person-reference meaning of the noun from which other possible person-reference uses originated. To determine this core meaning, the standard dictionary for Dutch, Van Dale Groot Woordenboek van de Nederlandse taal, was consulted.^[6] Of the 214 types included in the dataset, 72 (33.6 %) were listed with only one person-reference noun meaning, which was then straightforwardly considered as ‘the core person-reference meaning’.^[7] Another 102 (47.7 %) types occurred with more than one person-reference noun meaning listed under the same lemma. As a general rule, the first, viz. oldest, original meaning was considered as the core meaning. For instance, Van Dale includes one lemma for broer ‘brother’ with two senses. The first, and hence oldest is ‘male person considered in the relation of his consanguinity to other children of the same parents’, with the second being ‘name to address a little boy whose name one does not know’.^[8] Another 7 (3.3 %) types occurred with more than one person-reference noun meaning for two lemmas, e.g. politie ‘police’ with a separate entry for the collective and the individual meaning. These are candidate homonymous items, which then would require separate semantic-feature annotations. A corpus check revealed that each of these items was used monosemously (e.g. mongool which was only used as an insult). The 33 (15.4 %) remaining types were not listed in Van Dale, so coders aligned on the core meaning through a corpus check. Overall, the lexicographical outcomes aligned with the intuitions of the three coders concerning the most central meaning of the PRNs.

Second, as can be seen in Table 2 above, coding was conservative in that a noun is only coded for a semantic feature when the coder considers it to be explicitly part of the core person-reference meaning of the word, viz. that the feature is triggered blind of context of use. Of course, this type of intuitive judgment is prone to inter-rater disagreement. Hence, we thirdly relied on the principle of conservative coding in making final decisions on the items: as soon as at least one of the coders opts for code ‘no’ for a particular semantic feature, we consider this as an indication that the semantic feature is not sufficiently explicitly part of the core semantics of the noun. Consequently, in these cases the code ‘no’ is adhered to.

In one respect, coding was liberal rather than conservative. None of the semantic features were a priori considered to be mutually exclusive, so as to allow for overlapping features and family resemblance structures: a PRN can receive ‘yes’ for more than one feature.

4.2.4 Results per semantic feature

Needless to say, the generally conservative type-based approach followed here has notable shortcomings; it ignores semantic extensions and bleaches. Precisely these will provide pivotal at the third level of analysis. Still, the semantic-feature annotation approach discussed here leads to intuitive results that can indicate general hotspots of English usage. Overall, 86 out of 214 PRN received ‘yes’ for only one of the semantic features under scrutiny, 96 types receive ‘yes’ for two features, and 14 types receive ‘yes’ for three features. For instance, boerin ‘female farmer’ receives ‘yes’ for ‘gender’ and ‘occupation’. Given the conservative principles adhered to at this level of aggregation, ‘evaluation’ does not receive the code ‘yes’, despite the chance that certain tokens of boerin ‘female farmer’ may have been used pejoratively by the teenagers. More contextual activation of semantic features of this type on top of the core meaning is discussed in RQ3. Only 18 types do not receive ‘yes’ for any of the nine semantic features (as can be seen in Table A.1). The occurrence of overlapping features points to family-resemblance structures that we need to take into account to arrive at a final answer to RQ2 on variation in the distribution of English versus Dutch in subcategories of the field of person-reference nouns, in terms of types and tokens. It is nevertheless also revealing to verify whether we find variation in the foothold of English for each of the semantic categories.

Table 3 contains the descriptive results for the nine semantic subfields, both for the number of types and the number of tokens.^[9] Starting off with results on the type-level, we have tallied which of the 214 types contain a specific semantic feature, distinguishing between Dutch types and English types. Additionally, we report the relative presence of English types by dividing the number of English types that are considered to possess a certain feature by the total number of types that possess that feature. The categories in Table 3 are organized from a stronger to a weaker presence of English on the type-level. For the token-counts, we added up the tokens for all types that are seen to possess the particular semantic feature; the 181 English tokens for ‘Collectives’ result from the sum of the token counts for people (N = 27), police (N = 10), squad (N = 36) and team (N = 108).^[10]

Three observations stand out. First, we once again see that Dutch is present and dominant throughout, with English occurring in less than one out of three instances for all features. This holds for both types and tokens. Second, the findings for RQ1 on the stronger presence of English among the types compared to the token counts prevail for each of the nine semantic features. Nevertheless, the difference in the foothold of English in types and tokens does show variation across the nine features. The most notable variation concerns the patterns for: (i) ‘Evaluation’, where English holds a relatively strong position compared to Dutch, both in number of types (25.81 %) and in number of tokens (24.82 %); (ii) ‘School’, where English is absent altogether; and (iii) ‘Kinship’, where we find the most explicitly mixed pattern, with a notably stronger presence of English in the type counts (22.50 %) than in the token counts (3.06 %).

4.2.5 Results for semantic subfields

This asks for closer scrutiny, but of course we need to take into account the way these semantic features overlap within and across person-reference nouns. Table A.2 in the Appendix provides an overview of the combinations of features attested in the database. Again, no predominance of English stands out for any combination, but it is notable that the three semantic features ‘Evaluation’, ‘School’, and ‘Kinship’ mentioned earlier are mutually exclusive in this corpus. As ‘School’ only contains six types and no English elements at all, we will not look into the feature further. However, we have conducted a detailed analysis of ‘Evaluation’ (N = 62) and ‘Kinship’ (N = 40).

In particular, we started off from Table A.2 in the Appendix to assess all combinations of semantic features including ‘Evaluation’ on the one hand and ‘Kinship’ on the other. For person-reference nouns that are considered to possess the feature ‘Evaluation’, the semantic features ‘Friendship’, ‘Age’, ‘Gender’, and/or ‘Occupation’ can also be indexed. For ‘Kinship’, the co-occurring features are ‘Gender’, ‘Age’ and/or ‘Collectives’. A visualization of the resulting family-resemblance structures for ‘Evaluation’ and ‘Kinship’ can be found in Figures 2 and 3 respectively. The dotted boxes in the figures provide a provisional further subclassification within the subfields based on semantic similarities. Again, the same patterns appear: English occurs in the two subfields, with a stronger position in ‘Evaluation’ than in ‘Kinship’, and again with Dutch as the anchor. Some hubs of English can be found in the field of ‘Evaluation’, particularly concerning bbi versus lieverd and alternatives, cutie versus knapperd and alternatives, or bitch versus kutwijf. These are precisely the areas where English has been said to function as a youth-language marker, but even here, Dutch still has a clear role to play.

Figure 2:

Overview of the subfield ‘Evaluation’.

Figure 3:

Overview of the subfield ‘Kinship’.

4.3.1 Address and affect

When analyzing the relevant posts and their conversational context, two parameters immediately stand out: first, the distinction between forms of address versus common nouns that refer to people other than the addressee(s) and, second, the presence versus absence of an affective connotation. All 231 posts were coded for these parameters. We opted for a broad interpretation of ‘address forms’ by categorizing all references to the addressee(s) as such. By far most of these were direct forms of address (e.g. Heeey sister), but in a small number of cases the addressee was referred to indirectly (e.g.: alles goed met mijn sis? ‘everything okay with my sis?’). One-word utterances were considered semantically ambiguous (N = 14 for English, N = 11 for Dutch) and are excluded from Table 4. Tokens were classified as affective if the content of the post suggests affection for the female sibling (or female friend), i.e. people explicitly express their affection or positive support and/or add heart emoji or symbols for kisses (i.e. one or more x’s). Consequently, examples (8)–(10) received a positive score for affective connotation:

It is no coincidence that all examples contain a form of address: except for one token, the positive affective terms are used as forms of address. The reverse does not hold, since not all address forms are used in an affective context. Finally, it should be noted that negative affection was not included in the analysis because it was underrepresented in the relevant posts.

As can be deduced from Table 4, the English terms carry significantly more often an affective connotation than their Dutch counterparts and are used significantly more frequently for address (p for Fisher’s exact < 0.0001). The Dutch terms generally occur in more neutral contexts in terms of emotion and evaluation. This corroborates the importance of the evaluative dimension as a trigger for using English and points to a lexical specialization between Dutch referential nouns and English address forms.^[11] However, the overall percentages conceal differences between the individual lexicalizations for female sibling. The most important note to be made in that respect is that the diminutive zusje scores considerably higher for affection and use as address form than the other Dutch types: it is used in an affective context in 20.75 % of the cases (versus 1.83 % for zus/zuster) and as an address form in 32.65 % of the cases (versus 3.84 % for zus/zuster). Even then, these scores remain far below those of the English terms. For sis versus sister no such differentiation could clearly be discerned, although sis appeared to be accompanied by heart emoji more frequently. Still, the affective dimension is clearly present in the sister-posts too. Therefore, and in view of the limited number of sister-tokens, we refrain from further conclusions with respect to potentially subtle differences in the affective load of sis versus sister.

4.3.2 Friend or sibling

Since most of the English terms are used as forms of address, the relationship between the interlocutors could be checked in the corpus (which is coded for this parameter). It turns out that nearly all English tokens are used to address a close friend rather than a female sibling. Most telling in that respect is the phrase ‘sister from another mother’, which, judging from its presence across the internet, seems to be widely used in close friendship contexts and is actually used as such (i.e. in English, though always misspelled) by two teenagers. For the Dutch variants it is harder to determine whether the person referred to is an actual sibling or not, as most of the tokens refer to people other than the addressee who do not participate in the conversation, as seen in (11) and (12). Still, by far most of the posts with the Dutch terms do seem to refer to siblings. In some cases, there are clear clues, e.g., other family members are referred to in the same post or reference is made to the home context:

As quite a lot of contexts do not allow for disambiguation, we have not quantified the relative frequency of kinship versus non-kinship denotation. Instead, we took a closer look at all relevant posts produced by the 7 teenagers who use both English and Dutch ‘sister-terms’. These posts (extracted from the entire corpus) confirm the preference for Dutch terms when siblings are being addressed or referred to. One 16-year-old girl for instance continuously addresses or refers to her female friend as sis, but uses zus when explaining she could borrow a dress from her female sibling. Similarly, a 17-year-old-girl refers to her friend as sister from a other motheerrr (sic) but to her sister as zusje. A 19-year-old girl produces five posts with sister, all of them addressing a friend, and 13 posts with zus. The latter mainly refer to female siblings, but in two cases that very same friend is addressed/referred to with zus (see [13]). Interestingly, what she intends to say is: you are like a real (sibling) sister to me. These two zus-tokens thus adopt an affective friendship-related meaning that is mainly reserved for the English terms. In other words, while the pattern is clearly there (English for friends, Dutch for siblings), there is no strict rule. The corpus for instance also includes a 17-year-old girl who greets her sibling with sister (see [14]), but then someone else’s sister is being referred to with zus in another post by the same girl (see [15]). The latter also points to the fact that several conditions (+/− address forms, +/− affective, +/− kinship) may interfere with each other.

4.3.3 Gender

A final note concerns the gender of the eight authors of the English tokens, as seven of them are female. The only boy produces just one token, in the following post: meet your sister. Unlike his female peers, he does not produce English address forms but confines himself to a multi-word unit entirely in English. It should be noted though that boys are well represented as authors of the Dutch variants. In view of the limited number of chatters who use the English female sibling terms, we have to be careful jumping to conclusions. Still, this suggests that addressing one’s (female) friend as sister/sis in Flemish teenage talk is a ‘girl thing’.

Although this explorative study has its drawbacks as it is based on just one case involving a limited number of authors for particular types and hence misses the broader scope achieved at the more coarse-grained levels of analysis (RQ2, RQ3), it seems to be illustrative of the division of labor of English and Dutch in Flemish teenage talk. While the overall representation of English is quite modest, its use is subject to specific socio-pragmatic conditions. Finetuning this socio-pragmatic conditioning is definitely a priority for follow-up research.