Category: Research Scenario

Introducing the case of the Spoken Academic Belgium Dutch

The aim of the CLARIAH-VL Open Humanities Service Infrastructure is to advance digitally-enabled research in Humanities and the Arts by, among other things, providing data-level access to digitized and born-digital resources. In this blogpost series, we will communicate on research scenarios leading to and building upon the datasets made available through CLARIAH-VL. This blogpost introduces the research scenario ‘Spoken Academic Belgian Dutch (SABeD)’.

Introduction

In higher education, students are confronted with academic language use, with which they are often not familiar. Since academic language skills are a necessary condition for study success, higher education institutions in Flanders and the Netherlands focus on language support for students. In many institutions, these efforts have evolved into formal, embedded language policies, but research into their implementation is limited (Bonne & Casteleyn, 2022). Research (Deygers, 2017; Deygers et al., 2017) shows that Dutch language learners struggle with academic spoken Dutch, even when they passed the university entrance language tests, ITNA or CNaVT. Although academic listening is part of the test, learners indicate that the listening tasks in the language entrance tests are easier than actual lectures (Deygers et al., 2018). Linguistic features of the listening task in the test have not been empirically validated because of the lack of a corpus of spoken academic Dutch. This is one of the main reasons for building a corpus of spoken academic Belgian Dutch, which consists of (recordings of) academic lectures. Lectures are typical of higher education, and due to the covid pandemic, recorded video lectures are available in abundance.

Dataset

We selected academic lectures, because these constitute the predominant form of instruction in higher education institutions in Flanders, especially in the first bachelor year. Lectures are defined as instructional discourse given before an audience of at least 40 students, in which the lecturer is the dominant speaker and the level of interactivity is modest to low. We chose lectures for first year bachelor students, as both native speakers and foreign learners of Dutch indicate that the language used in lectures is one of the hurdles for comprehension and academic success (Deygers, 2017; Deygers et al., 2017). First year bachelor lectures also constitute the first encounter of the target group (i.e., Flemish first bachelor students and international students commencing university education in Belgian-Dutch) of our corpus with spoken academic Dutch. As such, these lectures make up a solid basis for our corpus compilation, especially considering that we cannot be certain if, and to what degree, the language of lectures in later bachelor and master years differs from that in the first bachelor. It is also important to take into account the primary pedagogical goal of the corpus, i.e., developing learning materials for students entering Flemish higher education.

Research question

One of the main motivations of the project is to create empirically validated vocabulary lists of spoken academic Belgian Dutch, in order to lower the barrier to entering university education by providing learning material for students.

Challenges

1. The first challenge was to obtain an approval of the Social and Societal Ethics Committee of the University of Leuven to allow for publication of the collected data, even after having obtained written permission of the speakers in the data.
2. Another challenge was to actually obtain the videos. Due to the covid pandemic recorded courses were abundantly available, but there was no technically feasible way for bulk downloading. 
3. Transcription of audio is a slow and expensive process.
4. Linguistic processing of spoken language is quite different from processing of written language.
5. The corpus needs to be made searchable and frequency lists need to be made which provide information about the coverage.

Solutions

1. Permission was only obtained for anonymized data, in the sense that all mentions of course and professor names have to be removed or masked in the data. This is doable for the data that is manually transcribed, as these mentions are marked, and automatic audio muting for segments with a mention is applied.
2. The videos had to be manually downloaded one by one.
3. In order to speed up the transcription process, automated speech recognition (Van Dijck et al., 2021) was applied, and the output was manually corrected.
4. In order to obtain word lists, we need to work with lemmas, i.e. the dictionary form of words, and not the surface form of the words, so shallow linguistic processing was performed. A first step consisted of automated punctuation insertion, for which a specific module was developed (Vandeghinste & Guhr 2023), and next, part of speech tagging with Frog (van den Bosch et al. 2007) was applied.
5. The CLARIN Autosearch infrastructure (de Does et al. 2017) allows for uploading a corpus with annotations and makes it searchable with corpus query language. This can be shared with any user with a CLARIN account upon request. Autosearch also allows for frequency list extraction.

Next steps

The following blogposts within this series will detail  the next steps taken as part of this research scenario. This will include:

• A first version of the dataset (only the written data), which has already been deposited and published by the  CLARIN-B Centre INT and can be downloaded at:  https://hdl.handle.net/10032/tm-a2-w4. Metadata has been harvested by CLARIN Virtual Language Observatory.  A version which is publicly accessible for anyone with a CLARIN login is work in progress.  A detailed description of the construction of the corpus can be found in: Jolien Mathysen, Vincent Vandeghinste, Elke Peters and Patrick Wambacq (2024). Constructing SABeD: A Spoken Academic Belgian Dutch Corpus. CLARIN2023: Selected papers. https://doi.org/10.3384/ecp210001
• Creation and publishing of specific Belgian Dutch spoken academic terminology list is ongoing.

References

Bonne, P., & Casteleyn, J. (2022). Taalbeleid en taalondersteuning: Op zoek naar een gedeelde basis en strategie voor implementatie. Tijdschrift voor Onderwijsrecht en Onderwijsbeleid, 4, 279–293. 

de Does, J., Niestadt, J., & Depuydt, K. (2017). Creating Research Environments with BlackLab. In CLARIN in the Low Countries. Ubiquity Press.

Deygers, B. (2017). Validating university entrance policy assumptions. Some inconvenient facts. In E. Gutíerrez Eugenio (Ed.), Learning and Assessment: Making the Connections – Proceedings of the ALTE 6th International Conference (pp. 46–50). Cambridge: ALTE. 

Deygers, B., Van den Branden, K., & Peters, E. (2017). Checking assumed proficiency: comparing L1 and L2 performance on a university entrance test. Assessing Writing, 32, 43–56.

Deygers, B., Van den Branden, K., & Van Gorp, K. (2018). University entrance language tests: A matter of justice. Language Testing, 35, 449–476

Vincent Vandeghinste and Oliver Guhr (2023). FullStop: Punctuation and Segmentation Prediction for Dutch with Transformers. Language Resources and Evaluation. Springer. https://doi.org/10.1007/s10579-023-09676-x. 

Van den Bosch, A., Busser, G., Daelemans, W., & Canisius, S. (2007). An efficient memory-based morphosyntactic tagger and parser for Dutch. In Selected Papers of the 17th Computational Linguistics in the Netherlands Meeting (pp. 99–114).Van Dyck, B., BabaAli, B., & Van Compernolle, D. (2021). A Hybrid ASR System for Southern Dutch. Computational Linguistics in the Netherlands Journal, 11, 27–34