We propose to use clustering of documents based on their 昀椀ne-grained linguistic properties in order to capture and validate text type distinctions such as medium and register. Correlating the bottom-up, linguistic feature driven clustering with text type distinctions (medium and register) enables us to quantify the in昀氀uence of individual author choice and medium/register conventions on variable linguistic phenomena. Our pilot study applies the method to German particles and intensi昀椀ers in a multimedia corpus, annotated for register. We show that German particles and intensi昀椀ers di昀er across both register and medium. The clustering based on the linguistic features most closely corresponds to the medium distinction, while the strati昀椀cation into registers is re昀氀ected to a lesser extent.
In this pilot study we investigate the use of clustering to capture macro-level distinctions between texts. We construct a bottom-up view of textual similarities via clustering based on their speci昀椀c linguistic features. We compare the results with annotations of the medium and register of the texts to see to what extend e昀ects of the used medium and register can be differentiated from individual author’s variability.
It is known that the linguistic phenomena (e.g., word choice, use of tenses, punctuation, etc.)
found in a text are shaped by many factors. In particular, highly variable phenomena such
as discourse particles are known to be in昀氀uenced by a wide range of aspects. Such aspects
of text level variation can be sociolinguistic factors like author demographics and identity,
author persona, or simply individual style, as investigated by sociolingui3s5ts, 2[
So while both the author as well as various external aspects are known to in昀氀uence linguistic variables, it is di昀케cult to pinpoint to what extent each linguistic feature depends on each of the in昀氀uences. The reason for this is that natural corpus data typically only covers a single medium or register, or con昀氀ates all categories: individual authors only contribute in one medium, each medium contains di昀erent registers or wildly di昀erent topics than the others, or the corpus is balanced for genre but it is not possible to track individual auth1ors.
In this paper, we make use of a social media corpus containing data from two di昀erent media (blogs and tweets), but covering the same set of 44 authors, the same topic (parenting and family life), and the same three registers (more detail below). We cluster the texts in our corpus using the relative frequency of two highly variable linguistic features, German modal and intensifying particles, found in each user’s texts, divided by medium and register. We then compare the resulting clustering with the groupings based on register or medium to assess whether the linguistic features re昀氀ect these external aspects of the utterance situation.
We 昀椀nd that both medium and register are positively correlated with the clustering of documents based on linguistic features, where the alignment is better for the medium distinction than for register. We argue that our method makes it possible to tease apart the individual in昀氀uence of medium, register, as well as individual author properties on the linguistic features studied.
The tables and scripts used in this paper can be accessed via the Open Science Framewor2k.
Various concepts have been used to characterize the situational circumstances in which a
discourse is produced, as these directly or indirectly in昀氀uence the way the discourse is shaped:
text type, genre, topic, register, and others4[
The medium is the speci昀椀c communication channel via which an utterance is made and
reaches its addressee, such as television, phone, oral speech, Twitter, or Facebook. This notion
is helpful in distinguishing between di昀erent communication situations speci昀椀cally related to
di昀erent so-called social media, as each medium carries its own a昀ordances. The a昀ordances a
medium o昀ers its users determine in which way the user and medium can interact13[
While the notion medium is based on the technical implementation of a discourse, register takes various aspects of the extralinguistic context into account, such as whether a speci昀椀c discourse is interactive, what the relation of the disourse participants is like, whether it is emotionally charged or its purpose is merely the exchange of information, et4c].. [Due to this interplay of contextual properties, register has, following Bib3]e,rfr[equently been characterized as multidimensional. Some researchers even propose to do away with a language-external inventory of register altogether6][, and want to instead represent registers as combinations of linguistic features present in the text. We do not follow this approach here, since we want to speci昀椀cally investigate the in昀氀uence of register on linguistic features – and therefore the registers themselves must be delineated independently.
In this work, we distinguish the registers Informative, Narrative, and Persuasive, based on situational properties such as the purpose of conversation (passing on information, reporting on life events, argumentation, respectively), the interactivity with the addressee, and the author involvement (both ranging from low for Informative to high for Persuasive). Linguistic features, with the exception of pronouns, were not used to distinguish between register dimensions. All registers are present in each of the two media.
To compare whether register, the medium or individual authorship has the most in昀氀uence on text similarity, it is necessary that we look at texts from the same author in di昀erent media and registers. To this end, we collected a corpus of German language blog posts and tweets from the same 44 individuals, but in a single domain: parenting. The community of parenting bloggers is relatively coherent and writes about similar topics both in their blogs, as well as on Twitter.
Blogs are a long-form text format with limited interactivity, while tweets are short posts (all our tweets are still under 140 characters) which allow direct responses; both media are public. Thus, the two media o昀er di昀erent types of communicative situations, but they are both available for all three registers introduced2i,nd§epending on the individual usage.
We constructed the corpus using the Twitter API and the user’s corresponding blog’s RSS
feed. The initial data collection was carried out in February, 2017, and the data used here
comprises the 500 most recent tweets and the 5 or 10 (depending on availability) most recent
blog posts. A more detailed description of the corpus and data collection can be found in
[
We manually annotated each blog post with one register (Informative, Narrative, or Persuasive). Since the tweets are o昀琀en too short to be assigned a clear register, we grouped them together and assigned one register to the entire tweet collection from one author, capturing the main usage of Twitter by that author. For tweet collections, we additionally allowed the intermediate registers Narrative-Informative and Narrative-Persuasive, denoting a mix between 3http://staff.germanistik.rub.de/digitale-forensisch-leinguistik/forschung/textkorpus-sprachliche-variioant-insozialen-medien/ these registers.
In addition, all modal and intensifying particles were manually identi昀椀ed and disambiguated in the corpus, with the help of word lists, annotation guidelines, and additional trained linguistics students. We de昀椀ne these phenomena in the next section.
German modal and intensifying particles are used di昀erently in di昀erent communicative
situations (spoken vs. written and formal vs. informal communication) and also depending on the
author using them. Both types of particles are nonin昀氀ected and modify the element in their
scope, and both are generally assumed to be more frequent in speech or conceptually spoken
language [
German modal particles are used to express the author’s attitude towards a proposition or
to make assumptions concerning the “common ground”, the shared knowledge of author and
reader [
Wenn ich schreibe, kann ich immerhin nicht einschlafen wie gestern beim Staatsanwalt. Allerdings bekomme ich trotzdem nicht mit was passiert, weil ich ja schreibe. ‘While I’m writing I can not fall asleep, as it happened yesterday at the prosecutor’s o昀케ce. But I still don’t get what’s happening because I’m JA writing.’ (blogposts-5487-3)4 @[USERNAME] Zu denken, nur diejenigen Väter würden nicht zahlen, die es nicht können, ist doch völlig weltfremd. ‘@[USERNAME] It is DOCH naive to think that only fathers who are not able to pay won’t do so.’ (tweets-1123) Due to the possibility of expressing multiple functions with one modal particle, the meaning of one modal particle can vary in di昀erent contexts. Additionally, not all modal particles have an exact match in other languages and they can not be directly translated to Englis1h0][. Kratzer shows two examples of ‘ja’ that include a translation to English. In both cases, there is no word that matches the meaning of ‘ja’ exactly, it is rather the function of the modal particle that is translated: (2)
Ich bin ja ein Einzelkind.
‘As you know, I am an only child.’ 4If not indicated otherwise, examples are from our corpus.
The use of modal particles varies between individuals and linguistic mod3e4s, [
Intensifying particles (for short, ‘intensi昀椀ers’) can be used to boost or tone down the intensity
of a gradable expression or utterance22[]. Similar to modal particles, there is inter-individual
variation in the use of intensi昀椀ers, but intensi昀椀ers are subject to much more rapid change of use
[
@[USERNAME] wieso kann ein Tattoo so brillante Farben haben? Wo hast du das machen lassen? ‘@[USERNAME] How can a tattoo have such brilliant colours? Where did you get it?’ (tweets-4677)
Da gibt es wirklich tolle Sachen - @[USERNAME] ‘There you can 昀椀nd really great things - @[USERNAME]’ (tweets-7846)
An overview of German intensi昀椀ers can be found in [
Based on the previous sociolinguistic results, we expect that the use of intensi昀椀ers in social media varies by individual demographic factors of the authors (as shown for speech), but may also vary by medium and register. Figur2eshows the distribution of the ten most frequent intensi昀椀ers in our corpus divided by our register dimensions. Similar to modal particles, dif2 ferent intensi昀椀ers are used more or less frequently in di昀erent media or registers. With = 1062.2 and p < 0.01, it can be assumed that there is a dependency between intensi昀椀er count and medium as well as register. This indicates that intensi昀椀ers, as well, can be used as a linguistic feature to cluster documents by medium and register.
It is our hypothesis that register in昀氀uences the low-level linguistic choices in addition to the medium or the author style. Starting from this hypothesis, we carry out a pilot study to cluster texts in a data-driven way based on their linguistic features. We want to 昀椀nd out whether these features enable us to distinguish registers from each other, e.g. rather than clustering each user’s tweet collection with their blog posts (as would be expected if the features re昀氀ect only individual linguistic style). The features we use are the per-sentence frequency of the top 10 modal particles and intensi昀椀ers found. We use the relative frequency of every feature to take the di昀erent lengths of the blog posts/tweet collections into account. Each document is represented by a vector containing the relative frequencies of the particles and intensi昀椀ers (see Table1).
Each user’s texts are split by medium (blogs and tweets) as well as register into a minimum of 2 (and a maximum of 4) documents per user. For example, the user 1095 is represented in three documents:b_1095_I (containing all Informative blog postsb)_,1095_N (Narrative blog posts), and t_1095_I (containing all tweets, which were annotated as Informative). Document names re昀氀ect the medium, user id, and register, in order.
We used the agglomerative clustering algorithm implemented in Python’s scikit-learn
package [
Out of nine groups of clusters, three contain mainly data points labelled as using Narrative register (document names shown in blue), two contain mainly data points labelled as Narrative or Informative (yellow) and one cluster contains data points labelled as Narrative or Persuasive (red). The last three groups of clusters contain data points from all of the register dimensions in equal amount. Out of these nine groups of clusters, six contain data points labelled as coming from blog posts and of the remaining three clusters, only one contains more data points labelled as coming from tweet collections than from blog posts.
Out of 64 pairs of documents that were clustered directly together, 22 had the same register label. 18 out of these 22 cases are nodes where blog posts and tweet collections were clustered together, 23 other cases are nodes which contain data points from the same medium. In one node, neither medium and register nor author are the same. There are only two cases where the same user’s blog posts and tweets were clustered together. For one author, both documents were also labelled as Informative (blog post and tweet collection), for the other author, one was labelled as Informative and one as Narrative.
Even though particles and intensi昀椀ers are not enough data to cluster the same register together in all cases, the algorithm still tends to cluster documents from the same register together, as opposed to grouping the same user’s blog posts and tweet collection. This indicates that medium and register in昀氀uence how users write and that writing in a speci昀椀c register has an independent impact on linguistic choice from just the medium in which the user writes, and their personal linguistic style.
For evaluation, we compared the correlation of the clustering by linguistic features with the
register distribution on the one hand, and the medium on the other. As a quality measure, we
used the V-measure [
We proposed clustering documents in a multi-media and multi-register corpus of German parenting bloggers by their usage of modal and intensifying particles. The method can be used to 5A reviewer suggests that Twitter’s length restrictions for tweets lead to authors choosing shorter and, in general, less intensi昀椀ers than in blog posts. This e昀ect can also be seen for modal particles, though it is less strong. Thus, the question arises whether con昀氀ating both types of particles to cluster the document is reasonable. In fact, using only intensi昀椀ers for clustering leads to a slightly better V-measure for grouping by register (V = 0.1293), but does worse for grouping by medium (V = 0.1628). Using only modal particles leads to worse V-measure results than using only intensi昀椀ers. The prevalence of the medium probably arises from this di昀ering use of particles in both media, possibly due to length restrictions in tweets. generate bottom-up clusters of documents (based on linguistic features) and to compare these clusters to groupings of the same documents by medium and register. We show that both the medium and the register dimensions are re昀氀ected in the variation in our linguistic features. For our feature groups, modal and intensifying particles, the medium has a bigger e昀ect than the register. We would like to argue that clustering enables us to determine the relative importance of individual author properties and text level properties (medium and register) on the linguistic expressions found in a text.
A昀琀er having conducted this pilot study, we will apply this method to a di昀erent dataset to
test the reproducibility of our results. Another natural next step would be to integrate other
linguistic phenomena as features in the clustering. On the one hand, one could choose
phenomena that have been argued to vary based on register or medium. On the other hand, linguistic
variation in small-scale features has been used to account for individual author style, for
example in authorship attribution or author pro昀椀lin1g6[
We would like to thank Lesley-Ann Kern for discussion, and the annotators in preparing the corpus data. We are grateful to the anonymous reviewers for their helpful comments. Funded by the Deutsche Forschungsgemeinscha昀琀 (DFG, German Research Foundation), project ID 317633480, SFB 1287. [37] M. Zimmermann. “Discourse Particles”. InS:emantics. Ed. by P. Portner, C. Maienborn, and K. v. Heusinger. Vol. 2. Handbücher zur Sprach- und Kommunikationswissenscha昀琀 HSK. Berlin: Mouton de Gruyter, 2011, pp. 2011–2038.