A Study of Cross-Lingual Ability and Language-specific Information in Multilingual BERT
This work provides incremental insights into cross-lingual NLP for researchers, focusing on practical factors and methods to enhance multilingual models.
The study investigated cross-lingual transfer in multilingual BERT, finding that data size and context window size are key factors for transferability, and demonstrated that manipulating latent representations enables unsupervised token translation and improves cross-lingual ability with a computationally cheap method.
Recently, multilingual BERT works remarkably well on cross-lingual transfer tasks, superior to static non-contextualized word embeddings. In this work, we provide an in-depth experimental study to supplement the existing literature of cross-lingual ability. We compare the cross-lingual ability of non-contextualized and contextualized representation model with the same data. We found that datasize and context window size are crucial factors to the transferability. We also observe the language-specific information in multilingual BERT. By manipulating the latent representations, we can control the output languages of multilingual BERT, and achieve unsupervised token translation. We further show that based on the observation, there is a computationally cheap but effective approach to improve the cross-lingual ability of multilingual BERT.