CSS: Contrastive Span Selector for Multi-span Question Answering

This study investigates the task of Multi-span Question Answering (MSQA). Currently, the MSQA task is primarily modeled as a sequence tagging problem, predicting whether each word is a part of an answer. However, this approach independently predicts words without fully utilizing a comprehensive understanding of the complexities in MSQA. In this paper, we propose a novel model, Contrastive Span Selector. Our model utilizes a multi-head biaffine attention mechanism to generate the span representations and employs a CNN block for span-wise interaction. Additionally, we incorporate the question and a global token into the encoding process, projecting all vectors into a shared representation space. To train our model, we employ contrastive learning with a dynamic threshold to control the similarity boundary between answer spans and non-answer spans. Our model outperforms the tagger model by 6.32 in F1 score for exact match on the MultiSpanQA multi-span setting and 5.69 on the expand setting, establishing it as the state-of-the-art model for MSQA. The code is available at: https://github.com/phzh24/Contrastive-Span-Selector .