On Answer Position Bias in Transformers for Question Answering

Extractive Transformer-based models for question answering (QA) are trained to predict the start and end position of the answer in a candidate paragraph. However, the true answer position can bias these models when its distribution in the training data is highly skewed. That is, models trained only with the answer at the beginning of the paragraph will perform poorly on test instances with the answer at the end. Many studies have focused on countering answer position bias but have yet to deepen our understanding of how such bias manifests in the main components of the Transformer. In this paper, we analyze the self-attention and embedding generation components of five Transformer-based models with different architectures and position embedding strategies. Our analysis shows that models tend to map position bias in their attention matrices, generating embeddings that correlate the answer and its biased position, ultimately compromising model generalization.