面向多跳问答的多视图语义推理网络

Multi-view Semantic Reasoning Networks for Multi-hop Question Answering

由于多跳知识图谱问答任务的复杂性,现有研究大多通过堆叠多层图神经网络以捕捉更大范围的高阶邻居信息。这种做法将多阶信息融合在一起,以损失节点判别性为代价获取更全局的信息,存在过平滑问题;并且,由于离节点越近的邻居置信度越高,将多阶邻居信息融合在一起的做法会忽略邻居的置信度。此外,多跳知识图谱问答存在许多数据集通常没有给定中间路径的监督信息的弱监督问题,会使模型在进行路径推理时缺乏有效的指导信息,导致模型推理能力降低。为了解决以上问题,论文提出了一种多视图语义推理网络,该网络利用全局和局部两种视图的信息共同进行推理。全局视图信息是指节点的多阶邻居信息,能够为推理提供更丰富的证据;局部视图信息则只关注节点的1阶邻居信息,更具有判别性,能够缓解全局视图信息存在的过平滑问题。同时,该网络将问题分解为多个子问题作为中间路径推理的指导信息,并从问题语义构成的均匀性和一致性出发,设计了一种新颖的损失函数以提升问题分解的质量,以提高模型中间路径推理的能力。论文方法在3个真实数据集上进行了大量实验,实验结果表明,多视图的语义信息能够为推理提供更加全面的证据,将问题分解为子问题的做法能够提高中间路径推理的准确性,证明了论文方法的有效性。

Due to the complexity of multi-hop KBQA, most existing works capture a wider range of higher-order neighbour information via multilayer GNNs. This approach combines multi-order information that sacrifices node discriminativeness to obtain more global information. Furthermore, it suffers from the over-smoothing problem, since fusing the multi-order information ignores the confidence of the neighbours, as neighbours closer to the node have the higher confidence. Another problem with multi-hop KBQA is that many datasets commonly do not provide the supervision information of intermediate paths. Thus, the weak-supervision problem usually makes the model lack effectual guidance information when conducting path reasonings, which results in reduced model reasoning ability. Aiming to solve the above problems, an approach of Multiview Semantic Reasoning Networks(Multi-view SRNs) which utilizes information from both global and local views to jointly perform the reasoning was proposed in this paper. Namely, the global view information refers to the multi-order neighbour information of the node, which provides more numbers of crucial evidence for the reasoning, while the local view barely focuses on the first-order neighbour information of the node,which makes the node representation more discriminative, thereby alleviating the over-smoothing problem of the global information. Moreover,the original question was decomposed into multiple sub-questions as the guiding information for intermediate path reasoning. Then, an innovative loss function based on the uniformity and consistency of the semantic composition of the question was designed to improve the question decomposition quality, which promotes the ability of the model for intermediate path reasoning. The extensive experimental results on three benchmark datasets convincingly demonstrate that multi-view semantic information can provide a more comprehensive evidence for the reasoning, and the proposed method of decomposing the question into sub-questions is able to increase the intermediate path reasoning accuracy.