Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 8017 - 8026 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Improving Time Sensitivity for Question Answering over Temporal Knowledge Graphs Chao Shang1, Guangtao Wang1, Peng Qi1, Jing Huang2∗ 1JD AI Research 2Alexa AI, Amazon {chao.shang3, guangtao.wang, peng.qi}@jd.com jhuangz@amazon.com Abstract Question answering over temporal knowledge graphs (KGs) efficiently uses facts contained in a temporal KG, which records entity rela- tions and when they occur in time, to answer natural language questions (e.g., “Who was the president of the US before Obama?”). These questions often involve three time-related chal- lenges that previous work fail to adequately address: 1) questions often do not specify ex- act timestamps of interest (e.g., “Obama” in- stead of 2000); 2) subtle lexical differences in time relations (e.g., “before” vs “after”); 3) off-the-shelf temporal KG embeddings that pre- vious work builds on ignore the temporal order of timestamps, which is crucial for answering temporal-order related questions. In this paper, we propose a time-sensitive question answering (TSQA) framework to tackle these problems. TSQA features a timestamp estimation mod- ule to infer the unwritten timestamp from the question. We also employ a time-sensitive KG encoder to inject ordering information into the temporal KG embeddings that TSQA is based on. With the help of techniques to reduce the search space for potential answers, TSQA sig- nificantly outperforms the previous state of the art on a new benchmark for question answer- ing over temporal KGs, especially achieving a 32% (absolute) error reduction on complex questions that require multiple steps of reason- ing over facts in the temporal KG. 1 Introduction Temporal knowledge graphs (KGs) record the re- lations between entities and the timestamp or time period when such relation hold, e.g., in the form of a quadruple: (Franklin D. Roosevelt, position held, President of USA, [1933, 1945]). This makes them a perfect source of knowledge to answer questions that involve knowledge of when certain events oc- curred as well as how they are related temporally (see Figure 1 for an example). Unlike question ∗Work done at JD AI Research. Figure 1: An example of complex temporal question on a temporal KG. answering (QA) over non-temporal KGs that is mainly concerned with relational inference, a core challenge in temporal KGQA is correctly identi- fying the time of reference mentioned explicitly or implicitly in the question, and locating relevant facts by jointly reasoning over relations and times- tamps. Inspired by work on relational KGQA (Huang et al., 2019; Saxena et al., 2020), where knowl- edge graph embeddings (Dasgupta et al., 2018; García-Durán et al., 2018; Goel et al., 2020; Wu et al., 2020; Lacroix et al., 2020) learned indepen- dently of question answering are used as input to KGQA models, previous work (Saxena et al., 2021) employs temporal KG embeddings to attack the problem of temporal KGQA. Despite its relative success on simple temporal questions that directly queries facts in the KG with one out of …

Content truncated to 500 words