Improve Transformer Models with Better Relative Position Embeddings
This work addresses the need for better position embeddings in Transformers to enhance model accuracy and generalization, particularly for NLP tasks, though it is incremental relative to existing methods.
The paper tackles the problem of underutilized position information in Transformer models by proposing new relative position embedding techniques that increase interaction in self-attention, improving results on SQuAD1.1 and demonstrating robustness for long sequences with a small computational budget.
Transformer architectures rely on explicit position encodings in order to preserve a notion of word order. In this paper, we argue that existing work does not fully utilize position information. For example, the initial proposal of a sinusoid embedding is fixed and not learnable. In this paper, we first review absolute position embeddings and existing methods for relative position embeddings. We then propose new techniques that encourage increased interaction between query, key and relative position embeddings in the self-attention mechanism. Our most promising approach is a generalization of the absolute position embedding, improving results on SQuAD1.1 compared to previous position embeddings approaches. In addition, we address the inductive property of whether a position embedding can be robust enough to handle long sequences. We demonstrate empirically that our relative position embedding method is reasonably generalized and robust from the inductive perspective. Finally, we show that our proposed method can be adopted as a near drop-in replacement for improving the accuracy of large models with a small computational budget.