Writing support for adult learners of Dutch in the era of generative AI
- Author
- Joni Kruijsbergen (UGent) and Orphée De Clercq (UGent)
- Organization
- Abstract
- Large language models (LLMs) such as those powering ChatGPT are reshaping writing. While traditional approaches in writing instruction are mostly limited to text revision, the current paradigm shift towards generative AI presents novel opportunities and challenges (Kasneci et al. 2023). However, to unlock the full potential of these advancements, it is crucial to first understand these LLMs’ capabilities. In the field of Natural Language Processing the state of the art is to either fine-tune the LLMs to a specific task or directly prompt them, known as zero-shot prompting (Min et al. 2023). The research presented here explores these two approaches in the context of writing problem detection for learners of Dutch. Research on languages other than English is important to uncover the potential of LLMs on less represented and lower-resourced languages (Volodina et al. 2023). To this purpose a writing dataset has been collected from adult Dutch learners at different CEFR levels comprising 6,336 sentences. Preliminary findings indicate that fine-tuning still outperforms zero-shot prompting, highlighting the importance of a critical perspective on, for example, ChatGPT-generated output. Additionally, the fine-tuning results reveal that multilingual models, despite comprising limited Dutch data compared to English, are on par with monolingual models trained exclusively on Dutch. This corroborates findings from the shared task on multilingual L2 writing error detection (Volodina et al. 2023). The presentation will showcase the dataset, its annotations and more detailed results and insights, taking into account the rapid developments within the field.
Downloads
-
(...).txt
- full text (Accepted manuscript)
- |
- UGent only
- |
- Text
- |
- 2.45 KB
Citation
Please use this url to cite or link to this publication: http://hdl.handle.net/1854/LU-01HQJZM0YZDE8ZPP6E1H3KJPSN
- MLA
- Kruijsbergen, Joni, and Orphée De Clercq. “Writing Support for Adult Learners of Dutch in the Era of Generative AI.” SIG Writing 2024, Abstracts, 2024.
- APA
- Kruijsbergen, J., & De Clercq, O. (2024). Writing support for adult learners of Dutch in the era of generative AI. SIG Writing 2024, Abstracts. Presented at the SIG Writing 2024, Paris, France.
- Chicago author-date
- Kruijsbergen, Joni, and Orphée De Clercq. 2024. “Writing Support for Adult Learners of Dutch in the Era of Generative AI.” In SIG Writing 2024, Abstracts.
- Chicago author-date (all authors)
- Kruijsbergen, Joni, and Orphée De Clercq. 2024. “Writing Support for Adult Learners of Dutch in the Era of Generative AI.” In SIG Writing 2024, Abstracts.
- Vancouver
- 1.Kruijsbergen J, De Clercq O. Writing support for adult learners of Dutch in the era of generative AI. In: SIG Writing 2024, Abstracts. 2024.
- IEEE
- [1]J. Kruijsbergen and O. De Clercq, “Writing support for adult learners of Dutch in the era of generative AI,” in SIG Writing 2024, Abstracts, Paris, France, 2024.
@inproceedings{01HQJZM0YZDE8ZPP6E1H3KJPSN, abstract = {{Large language models (LLMs) such as those powering ChatGPT are reshaping writing. While traditional approaches in writing instruction are mostly limited to text revision, the current paradigm shift towards generative AI presents novel opportunities and challenges (Kasneci et al. 2023). However, to unlock the full potential of these advancements, it is crucial to first understand these LLMs’ capabilities. In the field of Natural Language Processing the state of the art is to either fine-tune the LLMs to a specific task or directly prompt them, known as zero-shot prompting (Min et al. 2023). The research presented here explores these two approaches in the context of writing problem detection for learners of Dutch. Research on languages other than English is important to uncover the potential of LLMs on less represented and lower-resourced languages (Volodina et al. 2023). To this purpose a writing dataset has been collected from adult Dutch learners at different CEFR levels comprising 6,336 sentences. Preliminary findings indicate that fine-tuning still outperforms zero-shot prompting, highlighting the importance of a critical perspective on, for example, ChatGPT-generated output. Additionally, the fine-tuning results reveal that multilingual models, despite comprising limited Dutch data compared to English, are on par with monolingual models trained exclusively on Dutch. This corroborates findings from the shared task on multilingual L2 writing error detection (Volodina et al. 2023). The presentation will showcase the dataset, its annotations and more detailed results and insights, taking into account the rapid developments within the field.}}, author = {{Kruijsbergen, Joni and De Clercq, Orphée}}, booktitle = {{SIG Writing 2024, Abstracts}}, language = {{eng}}, location = {{Paris, France}}, title = {{Writing support for adult learners of Dutch in the era of generative AI}}, year = {{2024}}, }