Research questions:
- Which cross-lingual knowledge transfer strategy is the most beneficial for the Slot Filing task in the given domain (emergency calls)? Pros/cons of different approaches.
- Which strategy to filter translated training data seem to be the most advantageous?
- How do cross-lingual knowledge transfer strategies in question and their effects differ for different model architectures?
Data: simulated emergency dialogues in German and English from the NotAs project, publicly available datasets for NER/Slot Filling tasks (e.g. SGD, MultiWOZ etc.)
Paper suggestions and useful links:
- Gaspers, J., Karanasou, P., & Chatterjee, R. (2018). Selecting machine-translated data for quick bootstrapping of a natural language understanding system. arXiv preprint arXiv:1805.09119. Link: https://arxiv.org/pdf/1805.09119.pdf.
- Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B., De Laroussilhe, Q., Gesmundo, A., … & Gelly, S. (2019, May). Parameter-efficient transfer learning for NLP. In International Conference on Machine Learning (pp. 2790-2799). PMLR. Link: https://arxiv.org/pdf/1902.00751.pdf.
- Pfeiffer, J., Vulić, I., Gurevych, I., & Ruder, S. (2020). Mad-x: An adapter-based framework for multi-task cross-lingual transfer. arXiv preprint arXiv:2005.00052. Link: https://arxiv.org/pdf/2005.00052.pdf.
- https://docs.adapterhub.ml/adapters.html
- https://adapterhub.ml/
