Semantic answer type prediction (SMART) is known to be a useful step towards effective question answering (QA) systems. The SMART task involves predicting the top-$k$ knowledge graph (KG) types for a given natural language question. This is challenging due to the large number of types in KGs. In this paper, we propose use of extreme multi-label classification using Transformer models (XBERT) by clustering KG types using structural and semantic features based on question text. We specifically improve the clustering stage of the XBERT pipeline using textual and structural features derived from KGs. We show that these features can improve end-to-end performance for the SMART task, and yield state-of-the-art results.

翻译：语义答案类型预测（SMART）被视为构建高效问答系统的关键步骤。该任务涉及为给定自然语言问题预测知识图谱中的前$k$个类型，但由于知识图谱中类型数量庞大而极具挑战性。本文提出采用基于Transformer模型的极端多标签分类方法（XBERT），通过结合结构特征与基于问题文本的语义特征对知识图谱类型进行聚类。我们重点改进了XBERT流水线中的聚类阶段，利用从知识图谱中提取的文本与结构特征。实验表明，这些特征能够提升SMART任务的端到端性能，并取得当前最优结果。