Exploiting Hybrid Semantics of Relation Paths for Multi-hop Question Answering over Knowledge Graphs

Proceedings of the 29th International Conference on Computational Linguistics, pages 1813–1822 October 12–17, 2022. 1813 Exploiting Hybrid Semantics of Relation Paths for Multi-hop Question Answering Over Knowledge Graphs Zile Qiao1, Wei Ye2,†, Tong Zhang1,3, Tong Mo1 Weiping Li1, Shikun Zhang2 1 School of Software and Microelectronics, Peking University 2 National Engineering Research Center for Software Engineering, Peking University 3 TencentMT Oteam, China {zileq, wye, zhangtong17, zhangsk}@pku.edu.cn {motong, wpli}@ss.pku.edu.cn Abstract Answering natural language questions on knowledge graphs (KGQA) remains a great challenge in terms of understanding complex questions via multi-hop reasoning. Previous efforts usually exploit large-scale entity-related text corpora or knowledge graph (KG) embed- dings as auxiliary information to facilitate an- swer selection. However, the rich semantics im- plied in off-the-shelf relation paths between en- tities is far from well explored. This paper pro- poses improving multi-hop KGQA by exploit- ing relation paths’ hybrid semantics. Specifi- cally, we integrate explicit textual information and implicit KG structural features of relation paths based on a novel rotate-and-scale entity link prediction framework. Extensive experi- ments on three existing KGQA datasets demon- strate the superiority of our method, especially in multi-hop scenarios. Further investigation confirms our method’s systematical coordina- tion between questions and relation paths to identify answer entities. 1 Introduction Answering natural language questions on knowl- edge graphs (KGQA) is a challenging task (Bol- lacker et al., 2008; Tanon et al., 2016). Recent works mainly pay attention to a complex scenario, namely multi-hop KGQA (Sun et al., 2018; Hu et al., 2018; Saxena et al., 2020; Atzeni et al., 2021), where sophisticated reasoning over multiple edges (or relations) is required to infer the correct answer in the KG (Chen et al., 2019). The main challenge of multi-hop KGQA is to understand complicated questions and reason un- der incomplete KG, usually without supervision signals at the intermediate reasoning steps (Lan et al., 2021). One common strategy to alleviate this dilemma is to exploit auxiliary information to enrich knowledge representation. For example, †Corresponding authors. researchers have exploited entity-related textual corpus (e.g., from Wikipedia) as additional nodes in graph-based neural models (Sun et al., 2018, 2019a), or directly encoded them into enhanced entity representations (Han et al., 2020). A more recent effort, namely EmbedKGQA (Saxena et al., 2020), leverages implicit yet rich information in KG embeddings to answer complex questions over sparse KG. Unfortunately, the relation paths, which may contain beneficial supplementary information to characterize a candidate target entity for a topic entity in a question, are commonly underutilized. To the best of our knowledge, Yan et al. (2021) is the only effort involving exploiting the off-the- shelf relation path information. They use relation paths as simple coarse-grained input features by concatenating their text descriptions. From a more fine-grained and systematic perspective, given a question as a semantic view for the implied rela- tional fact of a <topic entity, target entity> pair, a relation path can serve as another highly-related yet complementary one. Therefore, we propose coordinating the question view and the relation path view to identify …

Content truncated to 500 words