1. INTRODUCTION
The many new forms of written communication on the Internet provide an enormous resource for linguists interested in language variation and use (Egbert, Biber & Davies Reference Egbert, Biber and Davies2015). There is, however, a need for more systematic large-scale studies of the nature of online data, since ‘[w]ithout a clear understanding of the linguistic variability of Internet texts we are severely limited in our ability to use this powerful resource for linguistic . . . research’ (Egbert et al. Reference Egbert, Biber and Davies2015:1817).
Different forms of written online communication (often generally subsumed under terms such as computer-mediated communication or netspeak – Crystal Reference Crystal2011) have often been described as a cross or hybrid between speech and writing, see for instance Yates (Reference Yates and Herring1996), Crystal (Reference Crystal2006, Reference Crystal2011) and Baron (Reference Baron, Vincent and Fortunati2009); but see also Jensen (Reference Jensen2014) for a critique of hybrid accounts. Likewise, the type of online data focused on in the current short communication, Swedish blog data, has been described as informal and spoken in nature compared to more traditional edited written material (e.g. Ahlberg et al. Reference Ahlberg, Andersson, Forsberg and Tahmasebi2015, Hillbom Reference Hillbom2015).
Whereas the accounts by e.g. Yates (Reference Yates and Herring1996) and Baron (Reference Baron, Vincent and Fortunati2009) are based on systematic comparisons between different types of data, the descriptions of the Swedish blog data often appear more impressionistically based as no reference is made to systematic overall comparisons between blog data and other genres or registers. The current short communication attempts to address this gap by quantitatively investigating how Swedish blog data compares with spoken and written registers in terms of type/token ratio, ratios of words in different frequency bands, and most common words. The purpose is to encourage further, preferably more qualitative, studies of the nature of blog data.
In linguistic variation studies, the terms genre and register are not always clearly distinguished and the choice ‘comes down to personal preference or tradition’ (Egbert et al. Reference Egbert, Biber and Davies2015:1818). Biber & Conrad (Reference Biber and Conrad2009) however make a clear distinction between genre and register. Registers are defined on the basis of situational characteristics (e.g. participants, interactivity, communicative purpose, etc.) and are analysed in terms of pervasive linguistic features functionally connected to the situational context. Genres, on the other hand, are text varieties defined by their conventional structures and features. In work on English blogs, Myers (Reference Myers2010a) appears to make a similar differentiation when he defines blogs as a genre (in part based on overall structuring features such as the use of links, reverse chronological order of blog entries, a column layout, etc.) while also acknowledging that the blog genre is not ‘defined by a particularly [sic.?] linguistic register’ (Myers Reference Myers2010a:19). In work on register classification of web texts, Egbert et al. (Reference Egbert, Biber and Davies2015) categorize blogs into different registers depending on their type, for instance news blogs and travel blogs are in the narrative register whereas opinion blogs and religious blogs are in the opinion register. Further, Biber, Egbert & Davies (Reference Biber, Egbert and Davies2015) claim that especially personal blogs often combine different purposes, such as narration and opinion, and therefore suggest that new hybrid registers are emerging. It is clear that the blog genre displays great variation, and that more work is needed to understand how different register features function in relation to the varying purposes of texts in this genre. The current work will only scratch the surface of this material but hopes to encourage further study in doing so.
The blog data under investigation here comes from the Swedish Blog mix collection, available through the Korp interface (Borin, Forsberg & Roxendal Reference Borin, Forsberg and Roxendal2012) of the Swedish Language Bank (https://spraakbanken.gu.se/eng/). On 9 April 2018, Blog mix contained Swedish blogs from 1998 to 2017 (totalling 615,658,549 tokens). Data from this collection will be compared with interactive spoken and formal written registers, using figures for relevant features presented by Allwood (Reference Allwood1998). Before presenting the quantitative comparison, I will review how current studies using Swedish blog data describe the material.
2. STUDIES USING SWEDISH BLOG DATA
Over the past years, a number of studies have made use of data from the Swedish Blog mix collection. The list of topics investigated include Swedish motion constructions (Olofsson Reference Olofsson2014), passive constructions in Scandinavian languages (Julien & Lødrup Reference Julien and Lødrup2013, Engdahl & Laanemets Reference Engdahl and Laanemets2016), ‘touch’ adjectives in Swedish (Hillbom Reference Hillbom2015), pragmaticalization of Swedish connective så att ‘so that’ (Rawoens Reference Rawoens2015), current usage of the gender-neutral third-person pronoun hen in Swedish (Ledin & Lyngfelt Reference Ledin and Lyngfelt2013) and vocabulary of football commentaries (Bergh & Ohlander Reference Bergh and Ohlander2012); Swedish lexicography work (Sköldberg & Hannesdóttir Reference Sköldberg and Hannesdóttir2016), supervised classification and semi-automatic discovery of Swedish pseudo-coordination constructions (Ahlberg et al. Reference Ahlberg, Andersson, Forsberg and Tahmasebi2015), and the creation of a reference dataset for natural language processing (Eide, Tahmasebi & Borin Reference Eide, Tahmasebi and Borin2016) have also been carried out using the Swedish Blog mix resource.
The Swedish Blog mix studies reviewed here often comment on the nature of the Blog mix data. As these comments shed important light on assumptions about, or observations regarding, the nature of the data, they will be of specific interest here. In general, no reference is made to systematic overall comparisons between different genres or registers which suggests that the descriptions are more loosely based on general observations of the data, or on results related to the specific features under study. The informal and casual nature of the data is often mentioned, as is its readiness to embrace novel language usage (e.g. Ledin & Lyngfelt Reference Ledin and Lyngfelt2013, Hillbom Reference Hillbom2015). As many of the studies investigate novel or informal features, the motivations for the choice of empirical material are often related to the study objects. Ledin & Lyngfelt (Reference Ledin and Lyngfelt2013:148, my translation) state that ‘blogs represent a trendy and casual written language variety, where neologisms like hen [the Swedish gender-neutral third-person pronoun] can be presumed to gain a foothold relatively fast’. Hillbom's (Reference Hillbom2015) description is similar but adds spoken influence to the list of features: ‘Blog mix . . . contains texts of a casual and informal nature, with spoken influences, where novel uses of the adjectives could occur’ (Hillbom Reference Hillbom2015:135, my translation). Also Malin Ahlberg and colleagues comment on the more speech-like quality of the data:
Since blog texts are typically informal and unedited, they contain a high degree of noise, i.e. misspellings and ungrammatical language. However, since the language of blogs is typically closer to spoken language than edited texts, and SPCs [Swedish pseudo-coordination constructions] tend to be more frequent in spoken language, they contain many SPCs as well as new SPC-like constructions. (Ahlberg et al. Reference Ahlberg, Andersson, Forsberg and Tahmasebi2015:14, my emphasis)
In relation to texts such as Blog mix, Sköldberg & Hannesdóttir (Reference Sköldberg and Hannesdóttir2016:329, my translation) characterize them as ‘unedited, spoken-like texts’. Rawoens (Reference Rawoens2015:57) claims that blog data is ‘a hybrid form between written and spoken language’, which, as presented in the introduction, is a common way to describe written online communication. To conclude, blog texts are described as an informal, spoken-like variety of Swedish, quick to adopt new usages.
Since language features described to be more informal or more common in speech are indeed also found in the blog data by these studies, or found to be more common in the blog data than in other types of investigated material, it does appear to be the case that the material satisfies the descriptions. Blog mix does contain informal language and constructions with less clear grammatical status than more edited texts. To what extent and in what way the genre of blog texts as a whole exhibits features more associated with spoken registers than with formal edited written registers is, however, not determined. As has been found for English blog data, the variation within the material could still be quite substantial, both in terms of how well-edited the texts are (Crystal Reference Crystal2006:246) and in terms of what written registers they can be classified into (Biber et al. Reference Biber, Egbert and Davies2015).
3. COMPARISON BETWEEN BLOG MIX AND EARLIER STUDIES OF SPOKEN AND WRITTEN SWEDISH
This section will investigate how Swedish blog data compares with spoken and written Swedish in terms of word frequencies and most commonly used words, i.e. features that have previously been used in genre or register studies (e.g. Yates Reference Yates and Herring1996, Allwood Reference Allwood1998, Biber, Conrad & Reppen Reference Biber, Conrad and Reppen1998). Specifically, I will compare one of the Blog mix corpora, Blog mix 2001, with the Swedish spoken and written corpora investigated by Allwood (Reference Allwood1998). The spoken corpus represents mainly interactive spoken registers such as conversations and interviews, and the written corpus contains formal edited writing from newspapers and novels, i.e. material that falls into the broadly defined informational and narrative registers, respectively (Biber & Conrad Reference Biber and Conrad2009).
The comparison will take the results presented by Allwood as its starting point, and corresponding figures will be determined for Blog mix 2001, by making use of the associated statistics file (https://svn.spraakdata.gu.se/sb-arkiv/pub/frekvens/stats_BLOGGMIX2001.txt). As different principles underlie the figures obtainable from the Swedish Language Bank and Allwood's (Reference Allwood1998) numerical results (for instance regarding word form statistics), manual re-calculations and re-groupings were necessary to arrive at comparable figures. The details of these manipulations are specified below in connection with the figures presented.
Blog mix 2001 was selected because its word count is comparable to the corpora used by Allwood (Reference Allwood1998). Allwood's (Reference Allwood1998) spoken corpus contains 276,391 words and his written corpus, 271,216 words. Blog mix 2001 contains 287,342 words when punctuation is removed from the original token count of 326,659.
3.1 Type/token ratios
In order to explain how and why the statistics file associated with Blog mix 2001 was manipulated we have to understand the underpinnings of the statistics presented by Allwood (Reference Allwood1998). Allwood's token count is simply the total number of words in the corpora. Punctuation is not included in the word count for the written data. In the spoken data, the token count includes all transcribed words, including feedback words such as ja ‘yes’, m ‘m’ and own communication management words such as ä ‘eh’.
The type counts used by Allwood (Reference Allwood1998) are based directly on word form, at least for the written data. This means that no homonym differentiation has taken place to calculate the overall types in the written data. In this count, springa ‘run/crack’ is counted as the same type regardless of whether we are dealing with the verb ‘run’ or the noun ‘crack’. The spoken word forms are, however, disambiguated according to the written word form in cases where the same spoken form corresponds to two (or more) written forms. For example, the spoken form /å/ corresponds to either och ‘and’ or att ‘to’ in writing. In the type count reported for speech, the spoken word forms are disambiguated according to their written counterparts.
The word frequency statistics for Blog mix 2001 is not counted on direct word forms, but rather lists, as unique lines in the file, word forms with the same Part-of-Speech tag (POS-tagFootnote i) and Lemgram label. The Lemgram label gives the lemma and inflectional table for each word form, and includes both unitary lemmas and multiword lemmas (when relevant). This means that the same word form, depending on how it has been analysed in terms of POS and Lemgram, will have a separate count in the file. This holds for the same word form used as an adjective or verb, such as rädda which can mean both ‘afraid’ and ‘save’, see Table 1. It also holds for usages in different complex lemmas, such as for the word form och ‘and’, see Table 2.
Word class disambiguation of rädda in the Blog mix 2001 statistics file.
a The POS label for rädda ‘afraid’ marks the word form as an adjective (JJ) in the positive form of comparison (POS), agreeing with both the non-neuter (UTR) and neuter (NEU) genders in plural (PLU), and with both indefinite (IND) and definite (DEF) definiteness, in the nominative (NOM) case. The lemgram label identifies the lemma (and the associated inflectional table) of the adjective rädd ‘afraid’, and is formed from the lemma (rädd), the part of speech (av = adjective), and a disambiguating numeral.
b The POS label for rädda ‘save’ marks the word form as a verb (VB) in the infinitive (INF), active (AKT) form. The lemgram label identifies the verb (vb) lemma rädda ‘save’.
Complex lemma disambiguation of och ‘and’ in the Blog mix 2001 statistics file.
Note: The POS label KN stands for conjunction. The lemgram label first gives the specific label for the lemma och ‘and’ (kn = conjunction) and, in relevant cases, indicates when och ‘and’ is part of a multiword expression. The multiword expression starts the complex lemgram label, and is followed by a specification of the part of speech of the unit (abm = multiword adverb; knm = multiword conjunction; ppm = multiword preposition; pnm = multiword pronoun).
To obtain type data for Blog mix 2001 comparable to Allwood (Reference Allwood1998), the statistics file from the Swedish Language Bank was re-sorted on word form only, and the numbers were grouped together. So, in the type count, the POS and Lemgram information available was disregarded. Table 3 compares the type/token counts and ratios in the Blog mix data with those reported for speech and writing by Allwood (Reference Allwood1998).
Comparison of word form type/token ratios in Blog mix 2001, speech and writing.
As type–token ratios measure lexical diversity, we can conclude that the blog data does appear to lie between the spoken and written material when it comes to how many different words are used. There are fewer types used than in writing but more than in speech. Thus, the type of written language variety represented by these blog texts appear to employ a slightly narrower range of words than more traditional writing, but it cannot be said to be closer to speech in terms of vocabulary use.
3.2 Frequency-based rank
Another measurement of relevance from Allwood's (Reference Allwood1998) comparison of speech and writing is to what degree words in different frequency ranges are used in the corpus. This measures to what degree a text employs words of different frequencies. A higher degree of the most common words means that these are employed more in the texts. Table 4 compares Allwood's figures with the figures for Blog mix 2001.
Corpus share of words from different frequency bands.
In terms of frequency-based ranks of word forms only (no homonym disambiguation) the Blog mix data appears closer to writing than to speech in terms of the degree of usage of the 10, 50 and 100 most common words. The most infrequent words (above rank 10,000) make up a smaller proportion in the blog data than in the written data, but larger than in the spoken data.
The blog data appears closer to writing here, which should be compared with the overall type/token count presented above, where the blog variety occupied a position mid-way between speech and writing. We can conclude that blog texts do not repeat the most common words to the same degree as interactive speech, but rather display a closer rank distribution to the written data.
3.3 Most frequent word forms
The most frequent words in the different corpora can reveal interesting differences and shed more light on how blog data compares to the spoken and written data. Table 5 lists the 10 most frequent words in the three text types. As the total number of words in the three text collections are roughly comparable, the absolute numbers are included here for direct comparison.
The ten most frequent words in the three corpora.
Unsurprisingly, all the words in these lists are function words. We can also observe that six of the function words (och ‘and’, det ‘it’, att ‘that, to’, är ‘is’, på ‘on’ and som ‘that, which’) are shared between all three lists. Even though there are rank differences, it seems safe to assume that these six words make up some kind of common core repository of function words used in many different types of Swedish constructions, regardless of register.
There are more commonalities in terms of rank of the six common words between the blog data and writing than between blog data and speech. Three of the words – och ‘and’, att ‘that, to’, på ‘on’ – have the same rank in blogs and writing. These are words of a type Allwood (Reference Allwood1998:Section 3) suggests are needed to ‘construct a complex phrase and sentence structure which is more typical of written than of spoken language’. Additional prepositions occur in the written variety (med ‘with’ and av ‘of, by’), in the blog data (för ‘for’), and in both writing and blog data (i ‘in’). As prepositions often have a function in more complex phrase and clause structures, the higher token and type count for these would suggest a more complex structure in the blog and written data compared to the spoken data.
First-person pronouns are found among the register features associated with involved production, that is, production with ‘an involved, non-informational focus, related primarily to a primary interactive or affective purpose and online production circumstances’ (Biber et al. Reference Biber, Conrad and Reppen1998:149). Face-to-face conversations are typical cases of this kind of production and rank high on involvement features (Biber et al. Reference Biber, Conrad and Reppen1998:152). Since Allwood's (Reference Allwood1998) spoken data is of an interactive type, the high frequency of the first-person pronoun jag ‘I’ is not surprising. Worth noting is that the blog data display similarities to the spoken data in this respect. The pronoun jag ‘I’ is found on the same rank position in Blog mix as in interactive speech. Naturally, the frequency of a single pronoun is not enough to determine the involved nature of the blog genre. Further work is needed here, perhaps on features ‘that can be interpreted as reflecting interpersonal interaction and the involved expression of personal feelings and concerns’ (Biber et al. Reference Biber, Conrad and Reppen1998:150) such as first- and second-person pronouns, wh-questions, emphatics, etc. (Biber et al. Reference Biber, Conrad and Reppen1998:149–150).
4. LIMITATIONS
The kinds of overall quantitative measurements compared in the current short communication cannot reveal the full nature of the blog genre. Much more work of a qualitative nature is needed to better understand how bloggers make use of features associated with different spoken and written registers.
5. CONCLUSIONS
Based on the comparisons made above, it cannot be claimed that the blog data in the Blog mix 2001 corpus is closer to speech than to writing. In terms of overall type/token ratio, the blog data rather takes a middle position between the spoken and written registers compared with here. In terms of vocabulary variance, the blog texts do not rely on the most common words to the same extent as interactive speech does, but rather display a distribution between different frequency ranks more similar to formal edited writing. The ten most frequent words in each of the three text types indicate that the blog data is more like writing from a basic structural perspective, but perhaps display certain personal involvement features normally associated with interactive speech.
To sum up, blog texts are not spoken in nature from the perspectives investigated here, but they are not entirely like more formal and edited writing either. They display less lexical richness than formal edited writing, and, as attested by the studies cited in Section 2, contain more informal language and marginal grammar. In conclusion, more work is needed to better understand how the blog genre makes use of features from different registers. Features reflecting interpersonal interaction may be especially relevant to investigate, as the perceived spoken nature may, at least partly, result from bloggers ‘choosing features appropriate to the interpersonal rhetoric of the genre, and thus using features that are also more likely to be associated with face-to-face communication’ (Myers Reference Myers2010b:270).
ACKNOWLEDGEMENTS
I very much appreciate the insightful and supportive comments from three anonymous NJL reviewers, Ewa Jaworska's careful copy-editing and Marit Julien's helpful advice. All remaining flaws are, of course, entirely my own.
