Boosting QA Performance Through SA-Net and AA-Net with the Read+Verify Framework

This paper proposes an ensemble model for the Stanford Question Answering Dataset (SQuAD) with the aim of improving performance compared to baseline models such as Albert, and Electra. The proposed ensemble model incorporates Sentence Attention (SA-Net) and Answer Attention (AA-Net) components, which leverage attention mechanisms to emphasize important information in sentences and answers, respectively. Additionally, the model adopts a read+verify architecture. In the Read stage, the model’s focus is on accurately predicting answer text, while in the Verify stage, it emphasizes the ability to determine the presence or absence of an answer, providing a probability for the existence of an answer. To enhance the training data, techniques for data augmentation are utilized, including Synonyms Replacement and Random Insertion. The experiment results demonstrate significant improvements on the Albert and Electra baseline models, highlighting the effectiveness of the proposed ensemble model for SQuAD.