Transferability and Stability of Learning With Limited Labelled Data in Multilingual Text Domain

Artificial intelligence, Classification (of information), Text processing

Using the learning with limited labelled data approaches to improve performance in multilingual domains, where small amount of labels are spread spread across languages and tasks, requires knowing the transferability of these approaches to new datasets and tasks. However, the lower data availability makes the learning with limited labelled data unstable, resulting in randomness invalidating the investigation, when it is not taken into consideration. Nevertheless, previous studies that perform benchmarking and investigation of such approaches mostly ignore the effects of randomness. In our work, we want to remedy this by investigating the stability and transferability, for effective use in the multilingual domains with specific characteristics.