This contribution focuses on the process of building a corpus for TikTok discourse. Particularly, it aims at describing the choices made during the construction of a corpus of Italian TikTok videos. The corpus It-tok was collected to individuate linguistic functional correlates of digital discourse on TikTok. It-tok includes two subsets of videos: the first one is centered on videos concerning themes of interest for the public debate (e.g. abortion, femicide, racism, internal politics); the second one is made up of videos with no specific theme, intended to constitute the control sample for the observations made for the first sub-corpus.
The major goal of this contribution is to present some of the choices made during the process of building It-tok, an Italian TikTok corpus. The It-tok project was born with three aims: • • • to provide a first assessment of the linguistic functional correlates (LFCs) displayed by TikTok content, and, subsequently, of the modality of communication on this specific social network; to highlight how themes of interest for the public debate are treated on this social network; to compare the LFCs found in general TikTok discourse with those emerging in thematically focused content.
By functional linguistic correlates, we mean the set of features that characterize language across different modalities; consequently, spoken texts exhibit specific correlates compared to written, read, or digitally produced texts (see section 3. for a more in-depth discussion): for instance, some LFCs of spoken language with regard to written language highlighted in previous studies regarded a significantly different distribution of PoS [1], the higher count of deictics [2], or of demonstratives [3]. LFCs describe effects on language uses based on the modality the communicative event takes place in. The focus on functional linguistic correlates poses significant challenges for decisions related to the corpus design, precisely because of the platform’s content multimodality. Indeed, the structure of the final product, i.e., the TikTok video, can take highly diverse forms, and can employ a number of semiotic means for conveying the message. Because of the platform’s multifaceted nature, a number of choices were to be made, to achieve a selection of content for analysis which was both replicable in its methodology and suitable for the achievement of the objective. In fact, beyond the thematic relevance, the selection process is complicated by the inherently heterogeneous nature of TikTok videos. These products differ not only in terms of topic but also from a semiotic point of view: they can include spoken language, text on screen, music, sound effects, gestures, visual editing techniques, or a combination of these means. As a result, choices concerning the corpus are of various kinds and must account for the platform’s multimodal nature, which challenges both linguistic analysis and methodological consistency from the outset. In this work, we chose to focus solely on verbal content, but TikTok would allow for a variety of different levels of interest.
For accomplishing the goals illustrated above, we decided to subdivide the collection stage into two phases, one concerned with general discourse (Gen) and one concerned with political and social discourse 0000-0003-0791-8714 © 2025 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). (PolSo).
During the collection, a number of methodological issues arose, which will be described here, together with the solutions we opted for; but, before getting to the decision-making section, we will briefly introduce TikTok and the reasons it was chosen.
TikTok is a social media platform which allows for the publication of video content. It gained much popularity in the latest years especially among the youngest parts of the population [4][5]. A tiktok is a (usually) short video: in fact, while the maximum duration the platform allows for a tiktok is 10 minutes, the mean duration of tiktoks stands around 50 seconds2. This format can be also found on other social media (e.g., reels on Instagram), but the reasons which led to choosing it are linked to the amount and the kind of popularity it reached lately, rather than the specific format of the content considered.
With the beginning of the post-digital era, and the
intersection and overlap of the online and offline lives,
digital content has begun to have a consistent effect on
our analogic life [6]. This is especially true for public
debate themes, and, in the latest years, this influence has
been especially clear for TikTok content, which is
getting central in the political discourse. To give an idea
of the importance TikTok gained in the public opinion
building process, we can provide some examples.
Consider the ban imposed by Donald Trump: as
neoelected President of the US, at the end of January 2025
he imposed the closure of the platform in the US, since
he held that the Chinese government was receiving
sensitive data about US government and citizens
through TikTok users3. Given the amount of public
disagreement with such decision, also manifested
through many Americans signing up to Xiaohongshu,
another Chinese social [7][8], Trump postponed the
closure, and TikTok went dark for only one day on the
20th January. Conversely, political and social topics are
increasingly present on TikTok discourse: the political
importance of the platform can be seen also in spreading
and testifying major political events and boosting
discussion about major socio-political issues, such as the
Black Lives Matter protests (2020) [
Given the importance social media nowadays hold
in our society, there is no doubt that TikTok constitutes
now a fundamental political mean [
Several studies in the field of linguistics (especially
acquisitional, clinical, and variational) have already
considered data coming from TikTok4. Many studies
regard specific domains, and were built through a
punctual methodology, which usually is not focused on
an open resource. The main application fields for these
studies regard especially language learning and teaching
practices enhanced through TikTok [
As explained in the previous paragraph, TikTok has already undergone a number of investigations in linguistics. Still, anyways, a bottom-up description of the functional features characterizing the platform is missing. Most existing studies tend to adopt a top-down 5 TWITA exploits an extraction method which would have not been much effective for tiktoks, as it is based on the extraction of tweets with Italian most frequent words, but tiktoks cannot be extracted based on words in the video, since automatic subtitles are not searcheable. approach, focusing on specific trends or phenomena, without accounting for the underlying structural features of TikTok communication as shaped by the platform’s multimodal and technologically mediated nature. Addressing this gap is one of the central aims of the It-tok project, which seeks to identify TikTok’s LFCs.
The LFCs of a specific modality of communication
consist in the set of features which primarily describe
that specific modality and characterize it with respect to
others [
It is important to distinguish between modality and channel. While channel refers specifically to the physical means of transmission (e.g., auditory, visual, tactile), modality encompasses the broader communicative framework that includes social conventions, technological constraints, and the multimodal configuration of the medium. In the case of TikTok, the modality is particularly complex and hybrid, since it combines features of spoken interaction (e.g., spontaneous speech, direct address to an audience) with elements of edited visual media (e.g., cuts, overlays, subtitles, background music), thereby creating a composite, dynamic communicative environment.
As noted in the literature on this topic, LFCs do not depend on sociolinguistic features of speakers, but, instead, they stay the same across diastratically and diatopically different speakers. For this very reason, the construction of It-tok could avoid taking into consideration sociolinguistic representativeness issues, focusing instead on capturing the linguistic regularities that emerge specifically from the platform’s multimodal communicative modality. The primary goal was to ensure that the corpus would be suitable for identifying these modality-driven patterns, rather than for mapping speaker-based variation. 6 This was also confirmed by a questionnaire we spread over March 2025, concerned with the use of hashtags on TikTok. 7 TikTok for Business, https://ads.tiktok.com/business/creativecenter/inspiration/popular/ hashtag/pc/en. 8 The same questionnaire cited in note 3 revealed that users of
Some issues with building a corpus from TikTok videos
have already been pointed out in [
To identify LFCs of TikTok discourse (subcorpus Gen), and to compare them to the LFCs of that sub-part of TikTok discourse which concerns themes of interest for the public debate (subcorpus PolSo), we proceeded through a double phased data collection.
TikTok API allows for the extraction of a maximum of 100 videos per extraction, which shall be from a 30 days time period, so the extraction was carried out month by month. The affordances of this research API does not allow for queries of tokens within the automatically generated captions (which would have been the preferred path), but it allows for querying hashtags. TikTok displays several characteristics in common with other platforms. One of these, is the affordance of hashtags. Hashtags are (small strings of) words, which function as hyperlinks, and link a content directly to others which contain the same hashtag. Most hashtags are thematic, in the sense that they describe the topic of that content. But this is not the only function they have on social media. In fact, hashtags can also be exploited to gain followers, or views, and in this case their form is a bit different. While, regarding the first function, hashtags do not display particularities on TikTok, considering the second one, these hashtags usually have a very transparent form on other platforms (#followforfollow, #followme). This is not the case for TikTok hashtags. Here hashtags are exploited by users in a way which, according to them, would boost the algorithm, and make them gain more views6, but their form is by far less transparent, namely we have #foryou, #fyp, #perte, which all refer to the for you page of the app. This type of hashtags is by far the most used on TikTok7, compared to thematic ones8. The so-called for you page (it. per te), so commonly cited in the hashtags, is the main page of the app, where users get the content TikTok suggests them based on what they liked or watched for longer9. What distinguishes the fyp from the other scrollable pages is the fact that in the fyp users are TikTok are well aware there are hashtags specifically useful (or thought to be useful) for TikTok content rather than other social media. 9 An equivalent of the TikTok fyp is the search page on Instagram, or the tl (timeline) on X. reached by content not necessarily published by people they follow. Therefore, to get in other people’s fyp means to get more visibility on the app. For this reason, users tend to exploit hashtags connected to fyp.
Another way to boost the popularity of one’s content consists in using thematic trending or popular hashtags, even for videos which have nothing to do with it.
All these features consistently affected our methodology of retrieving data, which had to consider the peculiarities of the platform.
Because of the peculiarities of TikTok hashtags, we chose to pay some special attention to the hashtags used for the query, and in particular we had to avoid keyword with a scope which was too large and concurrently the ones whose scope was too restricted, since we would have risked ending up with no results. We extracted a minimum of 15 videos per month for each of the subcorpora, selecting them by duration (>60s) and region of publication ("IT"=Italy). The video extracted were all published between October 2024 and January 2025. The extraction was performed during February and March 2025. Among the videos reached, only the ones showing the voice_to_text feature, namely the TikTok automatic transcriptions, were considered viable for It-tok. This way we could avoid video memes (usually shorter than 60s), those videos where the message is carried by the music rather than the speech and the ones in which there is no speech at all. Note that we did not use the automatic transcription as the final transcript: its presence was solely employed as a filter to exclude videos that did not contain or feature any spoken language. This way, we isolated the materials containing spoken language, whether continuous or discontinuous, explicitly excluding content such as memes, images carousels, or other materials lacking spoken language.
Finally, we got a total of 196 viable videos. Those
videos were automatically downloaded, transcribed
through the tool Open-AI Whisper in Python [
Finally the .txt files were automatically tagged
(through spaCy [
To sum up, for each video It-tok provides: • • • • • a .mp4 file; a .txt file; an antr.txt file; a .conllu file; a .eaf file.
The CoNLL-U file PROPN tags were exploited to carry
out the anonymization of the files.
As for now, the CoNLL-U files were checked just for the
PoS and lemma columns. Here we also tagged discourse
markers (DMs), in order to make them easily retrievable.
We chose to tag DMs because we thought they could
provide a measure of the extent to which TikTok
discourse could be compared to spoken language, and
since they are also included in the features which make
up LFCs [
During the extraction, both in the process for PolSo
and Gen, we noticed that the number of minimum
extraction necessary for reaching the minimum of 15
viable videos per month differed sensibly from month to
month, as can be seen in Table 2. We supposed it
depended on the period of the year the videos we were
extracting belonged to. Particularly, we decided to
extract videos from October, November and December
of 2024 and January of 2025. As it is well known, the
amount of posting, and the quality of posts on social
media is very much dependent on the time of posting. In
particular, during the last months of the year more
“seasonal” posting happens [
Gen
Oct Nov Dec Jan Feb Mar Oct Nov Dec Jan Feb Mar Years’ Eve) or the whole period of “end of the year” wrapped. This seasonal posting primarily consists of videos that likely do not meet our extraction criteria, as they are probably shorter than 60 seconds and/or lack spoken language. Consequently, to make sure it was a contingency of the peculiarities of the months considered, we attempted a subsequent extraction of February and March 2025, which showed a piece of evidence favoring our hypothesis, as they show a rate of videos featuring voice_to_text similar to the one displayed by January. This happens because the trends usually developing or spreading at the end of the year are trends that usually do not produce videos that would have been considered viable for our data (i.e., they are usually short, with songs or media carrying the message rather than the words and consequently not featuring voice_to_text). 4.2. PolSo Our thematic section was collected by extracting videos whose description included (at least) one in a list of hashtags. Due to the original thematic nature of hashtags, they usually have a general form, which made us prefer them with respect to keywords in video descriptions, as these last would have needed a broader consideration of their flected and/or derived forms (i.e., femminista ‘feminist.SG’, femministe ‘feminist.PL.F’, femministi ‘feminist.PL.M’, femminismo ‘feminism’).
The selection of the hashtags was carried out based
on our common sense of users, and on the most recent
surveys about what worries Gen Z10 the most (especially
compared to GenX), carried out by IPSOS in 202211 [
Following the questionnaire, we plan on realizing an
expansion of the PolSo section with themes from foreign
politics12.
4.3. Gen
As regards the generalist section, our modus operandi
was completely different. Since we could not find any
TikTok generalist corpus building methodology which
would have been somewhat exhaustive, we opted for a
format-based extraction strategy. Specifically, we
selected three widely used formats on the platform,
which are very common on the platform and that differ
primarily in their varying degree of (perceived)
interactionality: storytimes, answers, stitches.
Particularly, the extraction of these last two types was
based on the external caption TikTok automatically
produces when creating a video in these formats:
namely, risposta a ‘answer to’ and #stitch con ‘stitch
with’. Storytimes were extracted through hashtags.
In order to maintain internal consistency and
comparability, we also aimed to keep the duration of the
videos across the three formats as uniform as possible.
4.3.1. Storytime
The first format we exploited was the storytime,
extracted through the corresponding hashtag. A
storytime video displays a person usually speaking
directly to the camera, telling a story, usually from their
personal life, and unsolicited by anyone. Therefore,
storytimes are strongly monological. They make a
format of their own on a number of platforms13, as they
were also published on YouTube since 2015.
10 GenZ is the most present generation on TikTok [
Category hashtags environmental ecologismo ‘ecologism’, ecoansia and ecological ‘ecoanxiety’, ecoterrorismo ‘ecoterrorism’, crisis overtourism, antispecismo ‘antispeciesism’, specismo ‘speciesism’, ecofemminismo ‘ecofeminism’ national identities and policies social intersectional rights politicians’ names 4.3.2. Answer capitalismo ‘capitalism’, anticapitalismo ‘anticapitalism’, migrante ‘migrant’, migranti ‘migrants’, rifugiati ‘refugees’, antifascista ‘antifascist’, antirazzista ‘antiracist’, razzismo ‘racism’ femminismo ‘feminism’, feminist, metoo, femminicidio ‘femicide’, patriarcato ‘patriarchy’, violenzadigenere ‘genderbased violence’, aborto ‘abortion’, misoginia ‘misogyny’, omofobia ‘homophobia’, transfobia ‘transphobia’, omolesbobitransfobia ‘homo- lesbo- bitrans- phobia’, dirittiLGBT ‘LGBT rights’, abilismo ‘ableism’, grassofobia ‘fatphobia’, femminismointersezionale ‘intersectional feminism’, intersezionale ‘intersectional’, privilegio ‘privilege’, woke giorgiameloni, governomeloni, matteosalvini, giuseppeconte, ellyschlein, antoniotajani The second format we selected is composed by answers. In these videos, the creator selects a comment to one of their videos and answers to the comment through a tiktok, rather than through writing another comment. This affordance is exploited for two reasons: either because writing would have costed too much time, or because a video answer is content in itself, whereas a written answer is not, and TikTok algorithm is said to boost accounts posting frequently. Since answers directly refer to one comment, they can be considered more interactional, compared to storytimes. This is also to be seen in the linguistic features answers they display: more deictic expressions or second person pronouns and verb forms. 4.3.3. Stitch Stitches are somewhere in the middle between answers and storytimes. A stitch is a video which starts with/from another video section, usually lasting no more than 10 seconds, by another creator14. The stitched video (i.e., the original one) can be either used as a base for speaking one’s mind on a subject introduced by said stitched video itself, making it just a clarifier for the context, or can have the same function as the base comment in the answer videos, and in this case, as happens for answer videos, linguistic features include second person pronouns and verb forms. So, because of their very nature, stitches can be seen as formats more interactional than storytimes, but less interactional than answers.
All the processes the videos went through were to be checked manually, as automatized processing turned out to be not always completely reliable. As for the transcriptions, this could be due to the quality of the sound, the presence of dialect or of strong regional markedness, or the (highly variable) speech rate. Regarding the tagging, as we mentioned earlier (4.1), at the moment we did not check the syntactic tagging yet, but the PoS tagging and lemmatization outputs were to be manually checked, as they presented some inaccuracies. Lemmatization was to be corrected especially for cases such as: Regarding PoS tagging, main issues pertained: • • • • • • less frequent verbs or verb forms, e.g. tatuo ‘tattoo.PRS.1SG’ was lemmatized as tatuere instead of tatuare, or future forms like ripeterai ‘repeat.FUT.2SG’ lemmatized as ripeterai instead of ripetere; irregular verb forms, e.g., sai ‘know.PRS.2SG’ got lemmatized as saare instead of sapere; verb forms displaying suppletivism, e.g., vai ‘go.PRS.2SG’ got lemmatized as vai instead of andare. loans, marked as proper names; big numbers, such as years, which in the transcription were written in words (e.g. not 20 but twenty) and got marked as proper names; deverbal nouns, e.g. (il) ritorno ‘(the) come back’, tagged as VERB instead of NOUN.
As for April 2025, we extracted a total of 196 videos for
a total duration of more than 7 hours (see Table 4 for a
deeper insight). We are in the process of analyzing some
14 According to TikTok support: “Stitch allows you to combine a
video on TikTok with one you're creating” [
LFCs, focusing on particular lexicosyntactic traits.
Anyways, we are well aware that a number of
sociophonetic and morphological features could be
analyzed, but we chose to focus only on some levels. In
the meantime, It-tok shall be published online by the
beginning of 2026. Concerning the publication and the
treatment of data, we considered what was done
during other studies based on TikTok data.
Particularly, since the text data cannot be traced back
directly to the original videos, and since the content
we extracted is publicly accessible online, the data is
to be considered of public domain [
Table 6 confirms what we asserted about the way hashtags are used on TikTok: on the one hand, it is well true that some of the most used hashtags, excluding the ones we explicitly searched videos for, are thematic, but this is a very specific way of using hashtags, since the only ones we find are connected to PolSo themes. Videos from PolSo, in fact, display an overabundance of hashtags, compared to the ones in Gen (i.e., 9,37 vs. 3,2 hashtags per video, in average). Nonetheless, the most frequent hashtags remain #perte and #fyp, both very TikTok-specific and directed towards gaining views and/or boosting the content through the algorithm.
Turning to LFCs, we preliminarly looked at the
distribution of PoS in the It-tok corpus. PoS were chosen
for first assessment of some modality characteristics
because it has been shown that they correlate with the
spokenness of texts. In particular, nouns and verbs, and
their respective modifiers, adjectives and adverbs, have
been said to act as pivotal units in the construction of a
text. Their frequency offers significant insights into how
different types of texts are syntactically structured and
how modality influences linguistic composition.
Specifically, nouns tend to be more prominent in written
texts, while the frequency of verbs increases
progressively as one moves towards more natural
spoken language [
Lessico dell’Italiano Parlato (LIP), a corpus of
spoken Italian [
It-tok 17,9% 5,6%
The data for LIP, PTLLI and CS in Table 7 is taken
from [
NOUN ADJ
It must be considered, anyways, that the two subcorpora differ greatly by linguistic features displayed. As an example, Figure 1 shows the difference in the occurrence of DMs, which are strongly associated with spoken modality, with respect to the total of tokens in the different texts15.
As can be seen in Figure 1, even though the two subcorpora are equally spoken, it seems that the thematic one employs a kind of speech which is probably less hesitant. Further research will include a differentiation based on the different themes, within PolSo, and based on different features, between the PolSo and Gen.
Next advancements in It-tok building involve exploring the language features findable on TikTok (e.g., newly imported constructions or neologisms), broading It-tok and its scope, and the building of a treebank of TikTok discourse.
Though It-tok being a still very small corpus, it displays the potentiality to show a number of linguistic 15 p < 0.001 for the Mann-Whitney test. uses hardly findable in traditional corpora, like creative uses or newly registered loans (see 1-3), for which It-tok could also be searched.
(1) venire blastato da meraviglioso (0125_S) mio nipote è stato ‘to be blasted by my nephew was wonderful’ In (1), blastato < blastare < en. to blast stands for ‘getting humiliated’ through words. An adapted loan can be seen also in (2), where flexare sth. < en. to flex stands for ‘show one’s ability in sth.’.
(2) […] possibilità di flexare un po' di statistica (G0125_16) ‘[...] opportunity to flex some statistics’ In (3) it’s the whole passive construction to get borrowed.
(3) Un calciatore della Juventus è stato fatto outing
(1224_F)
‘A Juventus player was outed’
Another set of features, phonetics in nature, that could
be thus investigated regards the so-called “influencer
accent”, which was noticed around the internet but still
never assessed [
Furthermore, due to its informal nature, TikTok
could provide naturalistic data for a number of
linguistics areas of interest, e. g., neologisms and
gendered neologisms [
The expansion we foresee for It-tok regards Gen, but also partially PolSo. Nonetheless, based on the methodology applied for the extraction, Gen could be easily systematically broadened, making it a potential monitor corpus for Italian TikTok discourse through an yearly update. Furthermore, PolSo will be widened to include themes of foreign policy. A further expansion could pertain a set of videos that were systematically excluded with the present methodology: particularly, the video memes, and that could constitute the base for studies on innovative linguistic forms, could be extracted through hashtags such as #memetok. The pseudosuffix -tok can apply to any word X, i.e. X-tok, standing for ‘section of TikTok regarding X’. Examples of usage involve booktok ‘section of TikTok regarding books’, cattok ‘section of TikTok regarding cats’, footballtok ‘section of TikTok regarding football, 16 In fact, some gendered neologisms, such as girl dinner, actually were born from a trend on TikTok, and then spread all over other social networks. feministtok ‘section of TikTok dedicated to feminism’, lefttok ‘section of TikTok filled with leftists. From this pseudosuffix, It-tok takes its name.
Finally, we will be building a treebank of at least 10%
of It-tok, based on the methodology implemented for the
KIPARLA forest project [
With this contribution, we aimed to provide a brief overview of the methods adopted and the decisions made during the construction of an Italian TikTok corpus. Our choices were guided both by the specific communicative dynamics of the platform and by our research objectives, namely, to assess certain LFCs of TikTok discourse and, where applicable, to distinguish between generalist and thematic subtypes.
It-tok is structured to represent the first generalist corpus of spoken Italian on TikTok, and besides its main aims about LFCs and characteristics of political and social discourse online, it can represent a way to open TikTok to linguistic systematic studies, because of its replicable methodology, also applicable to create comparable corpora for other languages.
Due to the aim to describe LFCs in both general TikTok discourse and discourse on political and social topics, we adopted a split extraction strategy. Manual supervision was required at all stages of the automated processing to ensure consistency, accuracy, and compliance with the criteria established for corpus inclusion.
Importantly, the multimodal nature of TikTok, as a
platform where language coexists with visual, auditory,
and gestural elements, means that its texts are
inherently complex and shaped by multiple interacting
variables. These characteristics pose unique challenges
for corpus design and analysis, they also provide
valuable insights into modern digital communication
practices, both in terms of parallel communication
channels and the simultaneous use of multiple semiotic
modes to construct a message (e.g., use of emojis, or
particular visual rendering of verbal language, such as
the SpongeBob Mocking meme to convey derision [
Once completed, It-tok will provide a linguistically annotated corpus of Italian TikTok discourse, featuring transcriptions formatted according to the CLIPS conventions and annotated at multiple levels, including PoS tagging, morphological features, and syntactic 17 Website of the observatory available at URL: https://olindinum.huma-num.fr. dependencies in UD.
The corpus will also include a small but representative treebank, offering structured syntactic analyses of selected texts that reflect the linguistic complexity of this emerging multimodal variety.
I would like to thank prof. Voghera, and the whole linguistics nucleus of the University of Salerno, for the suggestions and the discussions about the project. This project is part of the program of education and sensibilization towards online public discourse of the OLiNDiNUM17 - Observatoire LINguistique du DIscours NUMérique project.
The corpus repository, which documents the corpus and treebank construction processes and the challenges encountered in the syntactic annotation of spoken data, is available on GitHub (https://github.com/cabinsix/ItTok). The repository will host the transcribed and anonymized files, along with their corresponding CoNLL-U formatted versions. The treebank is currently under development and can be accessed via the Arborator platform (https://arborator.grew.fr/?#/projects/It-tok).
Declaration on Generative AI During the preparation of this work, the author(s) used ChatGPT (OpenAI) in order to: Paraphrase and reword and Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.