Abstract:

When dealing with complex question answering task over knowledge graph, traditional semantic parsing method for query graphs requires encoding massive candidate query graphs with complex structures in the ranking stage to obtain their respective multi-dimensional feature representations. However, the global maximum or average pooling operation used during the encoding process often suffers from insufficient extracting capability for representative feature. To address the aforementioned problem, an optimal selection method for query graphs based on hierarchical pooling sequence matching is proposed. Meanwhile, sliding window technique based on hierarchical pooling is adopted to hierarchically extract local salient features and global semantic features of question and query graph sequence pairs during the interactive modeling of candidate query graphs, making the resulting feature vectors better used for semantic matching scoring of candidate query graphs. The proposed method is extensively evaluated on two popular complex question answering datasets, MetaQA and WebQuestionsSP. Experiment results show that by introducing hierarchical pooling operation, representative semantic features of complex query graph sequences can be effectively extracted, and the interactive encoding capability of the original ranking model can be enhanced, which helps further improve the performance of complex question answering systems over knowledge graph.

Key words: complex question answering over knowledge graph, semantic parsing for query graph, hierarchical pooling, interactive encoding