Di Wu


“I would rather have questions that can’t be answered than answers that can’t be questioned.”

Richard Feynman

1. About Shared Vocabulary

For NLP systems, tokenization is usually the first step. To overcome the so-called “OOV” problem, BPE was proposed. The core idea behind BPE is merely to compress data, yet its benefits go beyond handling OOV (after all, nothing beats character-level tokenization for OOV). In some multilingual scenarios, such as multilingual machine translation, however, things may change. Suppose a system is trained on extremely imbalanced multilingual data: a language-agnostic tokenizer like BPE allocates most of the vocabulary to high-resource languages, while splitting low-resource ones into character sequences. This is tolerable when methods such as temperature sampling are applied — but is it really an elegant solution? I have been thinking about this from the perspective of information transport, trying to answer: what is a good way to measure the quality of a vocabulary in multilingual scenarios, and how to build one?

2. Quantify Knowledge Transfer

For multilingual/multi-task systems, we generally rely on specific designs — shared vocabulary, a shared backbone, or other priors — to encourage knowledge transfer across languages/tasks. We hope such transfer emerges naturally from our intuition-driven modeling (and it does). I aim to quantify the degree of transfer or interplay, and further, to measure and encourage such interactions explicitly.