Measuring language transferability for cross-lingual word sense disambiguation

Cross-lingual transfer is an effective technique in improving word sense disambiguation (WSD) in low resource languages by leveraging knowledge from other higher resource languages. However, the impact of the source language selection for the transfer is still a problem that has not been deeply explored. Current cross-lingual WSD methods employ an experimental or intuitive approach to determine which language is most suitable for transfer, based on the practitioner’s field experience and theoretical knowledge. This may lead to poor performance on languages that are dissimilar or unrelated. In this work, we present a method that combines linguistic similarity and relative linguistic entropy for measuring the transferability between two languages. The experimental results demonstrate that our method is capable of more accurately quantifying the transferability of the languages. Furthermore, we also show that there is a significant correlation between language transferability and WSD performance. These findings facilitate cross-lingual WSD model to generalize over both related and unrelated languages, thus achieving generalized zero-shot learning.