Outside the Discipline, Inside the Data: A Retrospective Account of an Undocumented Tunisian Language Corpus in an Extractivist Research Context

(1.1) Research Project Background

The interview corpus discussed in this paper was produced as part of my jointly supervised doctoral research, conducted, at that time, under the supervision of the Université libre de Bruxelles in Belgium and the University of Carthage in Tunisia. The theoretical objective of the research was to examine the role of participation and technologies in raising awareness about heritage in the historical urban landscape of Tunis (Stiti & Ben Rajeb, 2023). Its applied dimension involved the development of P@trimonia 2.0, a platform for the participatory management of spatial-semantic information related to architectural and urban heritage (Stiti et al., 2022).

The fieldwork was conducted under the institutional umbrella of the University of Carthage, for practical reasons: access to the study area, recruitment of student interviewers, and logistical organisation were all significantly easier to manage from within a Tunisian institution than from Brussels. This was a sensible decision at the time, but it had an unintended consequence: the data produced during the fieldwork was treated, implicitly, as belonging to the Tunisian phase of the research, a local input to a broader project, rather than as a research output in its own right that would require formal documentation and archiving.

The study area was the buffer zone of the UNESCO World Heritage Site of the Medina of Tunis, and specifically its colonial architectural and urban landscape, a category of heritage subject to degradation and, at times, active demolition pressure (Stiti, 2023). A 2018 inventory by the non govermental organisation Édifices & Mémoires had already documented the precarious state of this heritage (Stiti et al., 2025), lending urgency to my question: how do residents of this area currently engage with, understand, and share information about their built surroundings, and what role might participatory digital platforms play in supporting that engagement?

To answer this, I designed a structured interview survey addressing three questions: (1) how do residents currently send, receive, and share information about their neighbourhood? (2) which digital platforms do they use, and why? And (3) how do they perceive the potential impact of participatory heritage platforms on their endangered built environment? These were, at heart, questions about communication, technology, and heritage, not about language. But answering them required me to collect, transcribe, and translate spoken language at scale, in a dialect with no standardised written form.

(1.2) Why In-Person, and Why Tunisian

I chose structured in-person interviews for reasons that seemed obvious at the time: I had no pre-existing contact list of residents, so I needed to go and meet people on the ground. I prioritised in-person over remote formats because I was working in a socially mixed area where some residents, particularly older people, might have difficulty with written questionnaires or phone interviews. And I chose a structured format to ensure that responses would be comparable across the nine sub-zones into which I divided the study area.

The language of the interviews was not a choice so much as a given: residents of Tunis speak Tunisian, that is the language in which any genuine encounter with them had to take place. Tunisian is the mother tongue of the vast majority of Tunisia’s population, but it occupies an unusual position: it has no standardised written form (Zribi et al., 2014), and it is rarely taught as a written language even to those who speak it natively. It differs substantially from Modern Standard Arabic, the written register of formal, educational, and official contexts (Sghaier & Zrigui, 2017), both in vocabulary and phonology. I knew all of this as a Tunisian researcher familiar with the Tunisian context. What I had not fully reckoned with was what it would mean, practically, to transcribe and translate a language with no orthographic standard, at scale, under time pressure, using student collaborators.

(1.3) The Language Dataset That Took Me by Surprise

At the time of the fieldwork, I was not thinking about datasets. I was thinking about data, raw material from which findings are extracted, instrumental and temporary. A dataset, by contrast, is a structured, documented, citable, and reusable research output (Koesten, 2020) with a life independent of the paper it informs. That distinction, which I did not grasp in June 2022, is the organising tension of this paper. I now recognise what I produced in June 2022 for what it is: a language corpus of resident voices from an endangered heritage neighbourhood, collected at a specific historical moment, in a dialect that is underrepresented in research data infrastructures. Beyond its value as a linguistic resource, this corpus is also a social history record, a primary source documenting how residents of a specific urban community understood and spoke about their neighbourhood at a particular moment in time, a dimension that makes its preservation urgent for reasons that extend well beyond linguistics. The corpus could be reused, by sociolinguists studying Tunisian, by heritage researchers comparing resident attitudes across sites, by urban planners studying community information practices. Instead, it had been sitting on my drive, invisible to all of them. This paper is my attempt to understand why, and to make sure that other researchers outside linguistics do not repeat the same mistake.

(2.1) What Was Collected

152 structured interviews were conducted across five days in June 2022, distributed across nine sub-zones of the buffer zone of the Medina of Tunis. I recruited 18 master’s students from the National School of Architecture and Urbanism (ENAU) and the Higher Institute of Environmental Technologies, Urban Planning and Building (ISTEUB), both affiliated with the University of Carthage. Students worked in nine two-person teams: one student interviewed the respondent, the other took notes and, where the participant gave verbal consent and the interviewer had sufficient storage space on their personal smartphone, recorded the encounter. Following each morning of interviews, teams transcribed their notes and recordings into written Tunisian and then translated the transcriptions into French. This processing took place within the same five-day fieldwork week. A representative sample interview, in Tunisian (AR) and in French (FR), conducted by Team 2 on their first day of fieldwork, hence the identifier B211, is provided in Annex A to illustrate the structure of the material collected.

(2.2) What Currently Exists

What I have today, stored in my personal Google Drive and on Zenodo, is the following: typed transcriptions in Tunisian of the interviews; French translations corresponding to those transcriptions; zone distribution maps showing the allocation of teams B1 to B9 across the study area; and a small number of audio files in mp3 or m4a format, representing only a fraction of the interviews that were originally recorded. The audio recordings that do survive cannot be systematically matched to specific transcripts, because the transfer of files from students’ personal smartphones to a shared location was never organised in a consistent or complete way. Files were transmitted through a variety of channels, WeTransfer, Bluetooth, Google Drive, and email, with no unified protocol, making it impossible to establish a complete and reliable inventory of what was originally recorded and what was ultimately received.

(2.3) What is Missing

The 18 students who produced the transcriptions and translations are not credited anywhere in the published record, their labour is acknowledged in general terms but not documented as dataset creation. No licence has been assigned to the material. No systematic anonymisation has been carried out: the transcriptions may contain names, addresses, or identifying details that would require careful review and redaction before any public deposit could responsibly take place.

The consent obtained by the students was verbal, given at the moment of encounter. In Tunis, particularly in street-level interactions in and around the Medina, spontaneous conversations are often shaped by informality, time constraints, and varying levels of trust toward formal documentation. Figure 1 shows the street where Team 3 interviewed participant B351.

Figure 1

In such contexts, written consent forms can generate suspicion, discourage participation, or abruptly terminate the exchange. Verbal consent was therefore the most appropriate and ethically proportionate procedure for this fieldwork, and I stand by that choice as sound for the purposes of conducting and publishing the research.

However, verbal consent, while sufficient to legitimise the production and dissemination of findings, left no written trace of what participants understood they were agreeing to. The gap between what was ethically sufficient for conducting the research and what would be required for open archival deposit constitutes one of the central structural tensions this paper seeks to illuminate. It is also worth noting that no ethics clearance was required by the Tunisian institutional partner, which meant no formal approval process was undertaken for that component of the study. Rather than treat this as a neutral administrative fact, I read it as symptomatic of a broader structural asymmetry in global research ethics governance, where the relative absence of formalised ethics infrastructure in some Global South contexts can inadvertently serve the logistical interests of internationally coordinated projects. This raises uncomfortable but necessary questions: whose standards govern research, and why? The decision to conduct fieldwork under Tunisian institutional oversight was, in part, a practical one, the absence of a formal ethics requirement, compared to the process at the European university that can take months or sometimes over a year, made it easier. Acknowledging this openly is, I believe, more honest and more useful than silence.

(3.1) Preparing the Field: Training and Team Organisation

Before the fieldwork week, I organised a two-hour training session at ENAU. I presented the research context, explained the interview protocol and question set, and distributed the students into nine two-person teams, each assigned to a specific sub-zone of the study area. I was pleased with this session at the time: the students were engaged, they knew the city well, they spoke Tunisian naturally, and they understood the objectives of the survey.

(3.2) The Language Challenge: Transcribing Tunisian

When Tunisian speakers write informally, in messages, notes, or social media posts, they typically combine Arabic script, Latin characters, and numerals to represent sounds that Arabic orthography does not easily encode, producing highly variable and individualised approaches (McNeil, 2022).

For my student teams, this meant that each pair was effectively left to invent their own system for putting on paper what they had heard. Some wrote predominantly in Arabic script; some rendered French code-switches in their original form. Some attempted precision by distinguishing no answer and refused to answer; others prioritised readability.

(3.3) Translation as Interpretation: From Tunisian to French

I chose French as the target language for translation because it is the language in which the master’s students work most fluently, albeit to varying degrees. For the published paper reporting the results of the research, I carried out an additional layer of translation, from French into English, to present the research findings in English (Stiti & Ben Rajeb, 2023). The translations that resulted are usable: they convey the informational content of what residents said, but they are not reliable as linguistic documents. The voice, affect, and rhetorical texture of the original Tunisian has been flattened. A resident who spoke with humour or indignation may appear in the French translation as simply expressing an opinion. I was able to work with this because I had been present during the fieldwork and could draw on my own contextual knowledge. A researcher encountering this material for the first time would have no such resource.

(3.4) Time Pressure and the Compression of the Pipeline

The most consequential structural decision of the project, though experienced at the time less as a deliberate choice than as an accepted constraint, was the compression of the entire research pipeline into a five-day period. Interviewing, note-taking, audio recording, transcription, and translation were all conducted within the same fieldwork week. I travelled from Belgium to Tunisia specifically to lead this workshop, which further reinforced the temporal concentration of activities.

Students who conducted interviews in the morning were expected to complete transcription and translation in the afternoon, creating an accelerated cycle of data production and processing. The fieldwork also took place under conditions of extreme heat in Tunis, with temperatures reaching 39.5°C by 11:30 a.m., while students were still outdoors conducting interviews. These temporal and environmental constraints materially shaped both the production of the data and the conditions under which it was rendered into written and translated form.

(4.1) What the Data Enabled

The interviews revealed three main things: how residents currently send, receive, and share information; which platforms they use and why; and how they perceive the potential impact of participatory heritage platforms on heritage sites already at risk.

Beyond the findings themselves, the interviews gave me something harder to quantify: a textured, living sense of how residents talk about their neighbourhood and their heritage, the old photos they kept (Figure 2), the metaphors they reach for, the comparisons they draw with other parts of the city, the mix of pride and frustration in their accounts of a landscape they inhabit and watch deteriorate. This ambient richness informed my interpretation even where it did not appear explicitly in the paper, and it is precisely what is inaccessible in the current state of the dataset.

Figure 2

(4.2) Three Tensions I Did Not Anticipate

Three structural tensions shaped the dataset’s current state: The first is the tension between research needs and dataset needs. My research required speed and flexibility; a reusable dataset requires standardisation and documentation. Reconciling them requires explicit planning, which I did not do.

The second tension is between working with student collaborators as learners and working with them as data producers. The 18 master’s students from ENAU and ISTEUB were not trained researchers: they had no experience with transcription protocols, no shared orthographic convention for written Tunisian, and varying levels of French proficiency. Yet their local knowledge, their fluency in Tunisian, and their familiarity with the study area were indispensable, no external team could have replicated what they made possible. The result of this tension is visible throughout the corpus: nine partially incompatible transcription styles, translations of uneven register, and audio files that were never systematically transferred. Their contribution was, simultaneously, the condition of possibility of the dataset and the source of its most significant quality limitations. Structuring that contribution more explicitly, through a transcription protocol, a shared glossary, and a dedicated post-fieldwork processing phase, would have preserved its value while substantially improving its consistency.

The third — and most fundamental — is the tension between the conventions of my own discipline and the requirements of open data. Architecture and urban studies are not fields in which datasets are routinely published alongside research papers. The assumption, rarely articulated but deeply embedded in practice, is that fieldwork data is private, provisional, and instrumental. I absorbed this assumption during my training and carried it into my doctoral research. The open data community challenges it; but that challenge reaches individual researchers unevenly, and it had not reached me at the moment when it would have made a difference.

What I have lost cannot all be recovered. The audio recordings that were never transferred cannot be retrieved. The verbal consent of participants cannot be converted retroactively into written documentation. The specific texture of individual encounters, the pauses, the contextual details, the observations my student note-takers made but did not record, is gone. This irreversibility is, for me, the strongest argument for doing things differently from the start. It is also the reason I chose to write this paper.

(5.1) Recognise That You Are Building a (Language) Dataset

The most important recommendation is also the most basic: if you are collecting, transcribing, or translating spoken language, you are building a language dataset, whether or not you call it that. Recognising this distinction from the outset, as defined earlier, changes everything that follows: it prompts you to identify a repository, plan file formats, and document as you collect rather than as an afterthought. For researchers in fields where data publication is not yet standard practice, this reframing may require a conscious effort. I suggest treating the dataset and the research paper as two distinct outputs from a single project, each deserving its own planning, its own documentation, and ultimately its own citation in the scholarly record. This is not to deny that data quality and research quality are connected; it is rather a practical recommendation for researchers who have never considered their data as a publishable output at all.

(5.2) Establish a Transcription Protocol Before Fieldwork

For research conducted in languages without a standardised written form, and Tunisian is far from the only example, a transcription protocol is not optional. Before fieldwork begins, researchers should agree on: the script or orthographic system to be used; conventions for dialectal features, code-switching, and non-verbal elements; and a procedure for flagging uncertain or unclear passages. Existing transcription frameworks developed within corpus linguistics and language documentation can provide a useful starting point, and local language experts or linguists can be invaluable collaborators in adapting these frameworks to a specific dialect and context.

This protocol must be practiced during the pre-fieldwork training, not simply distributed as a document. Exercises using sample material, time for questions, and a shared glossary of key domain-specific terms can substantially improve consistency across transcribers, especially when, as in my case, transcribers are working independently across multiple sites.

Established frameworks exist for exactly this purpose. For transcription, systems such as CHAT (Codes for the Human Analysis of Transcripts) provide detailed conventions for representing spoken language in writing, including hesitations, overlaps, and code-switching. For annotation and alignment of transcriptions with audio recordings, tools such as ELAN (Eudico Linguistic Annotator) are widely used in fieldwork-based language documentation. For metadata, the Open Language Archives Community (OLAC) standard was developed specifically to describe language resources in a way that makes them discoverable and reusable across research infrastructures. I was unaware of any of these at the time of fieldwork. Mentioning them here is not to suggest they should be applied rigidly in all contexts, but to signal to researchers like me that this infrastructure exists and is worth consulting before fieldwork begins.

(5.3) Separate Fieldwork from Post-Processing

Wherever logistically possible, transcription and translation should be treated as distinct phases with their own time budgets, separated from the fieldwork itself. Even a short dedicated post-fieldwork period — in which collaborators review, compare, and harmonise their transcriptions before beginning translation, and in which the researcher can identify and address inconsistencies — can substantially improve the coherence and reusability of the resulting material. Compressing everything into the fieldwork week, as I did, is a false economy: it saves time in the field and costs far more in the quality of the dataset.

(5.4) Design Consent for Data Deposit, Not Just Research

Consent procedures should be designed with dataset deposit in mind from the outset — even if, at the time of collection, deposit feels like a distant or uncertain prospect. In contexts where written consent is impractical, researchers can consider alternatives: a brief audio preamble at the start of each recording in which the interviewer explains the possibility of public data sharing; a team field log recording who gave consent and on what terms; or a participant information statement delivered verbally and logged systematically. None of these solutions is perfect, but all create a more robust evidentiary basis than unaided memory.

(5.5) Credit All Dataset Contributors Formally

Student collaborators who contribute substantially to dataset creation, through transcription, translation, or data collection, should be credited as dataset creators, with their names, affiliations, and specific roles recorded using a recognised taxonomy such as CRediT. This is a matter of scholarly fairness and a practical requirement of open data deposit. It also provides an incentive for quality: students who know their contribution will be publicly attributed are more likely to invest care and consistency in their work.

More broadly, the experience I have described suggests that dataset literacy, the understanding that research data is a first-class scholarly output deserving the same care, documentation, and accessibility as the papers it informs needs to reach researchers who do not identify as data scientists or corpus linguists. Pre-fieldwork training should include, alongside methodological and logistical preparation, an introduction to open data principles and to the responsibilities of the researcher as data creator. I did not have that introduction. I hope that reading this paper might serve, for some researchers, as a belated substitute.

(5.6) Limitations and Future Deposit Pathways

Several limitations of this dataset cannot be remediated. The audio recordings that were never transferred from students’ personal devices cannot be retrieved. Verbal consent cannot be converted retroactively into written documentation, leaving the scope of participant agreement uncertain. Transcription inconsistencies across nine teams cannot be fully harmonised without distorting the original material.

What remains possible is the following: the corpus has been deposited on Zenodo with restricted access; transcriptions will be reviewed and identifying details redacted prior to any access being granted; and researchers may request access individually, with each request evaluated against the original ethical framework. This is an incomplete solution — but it is a real one.