AlignFreeze: Navigating the Impact of Realignment on the Layers of Multilingual Models Across Diverse Languages
This work addresses performance issues in multilingual NLP for languages underrepresented in training data, but it is incremental as it builds on existing realignment techniques with a layer-freezing approach.
The paper tackles the problem of performance degradation in multilingual language models during realignment for cross-lingual transfer, particularly in languages differing from the source language, by introducing AlignFreeze, which freezes specific layers to prevent this degradation. The result shows that AlignFreeze improves Part-of-Speech tagging accuracy by more than one standard deviation in seven additional languages compared to full realignment, as demonstrated in experiments across 35 languages.
Realignment techniques are often employed to enhance cross-lingual transfer in multilingual language models, still, they can sometimes degrade performance in languages that differ significantly from the fine-tuned source language. This paper introduces AlignFreeze, a method that freezes either the layers' lower half or upper half during realignment. Through controlled experiments on 4 tasks, 3 models, and in 35 languages, we find that realignment affects all the layers but can be the most detrimental to the lower ones. Freezing the lower layers can prevent performance degradation. Particularly, AlignFreeze improves Part-of-Speech (PoS) tagging performances in languages where full realignment fails: with XLM-R, it provides improvements of more than one standard deviation in accuracy in seven more languages than full realignment.