UniSumm and SummZoo: Unified Model and Diverse Benchmark for Few-Shot Summarization

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1281–1293 July 9-14, 2023 ©2023 Association for Computational Linguistics TECHS: Temporal Logical Graph Networks for Explainable Extrapolation Reasoning Qika Lin1,2, Jun Liu1,3∗, Rui Mao4, Fangzhi Xu1,2, Erik Cambria4 1School of Computer Science and Technology, Xi’an Jiaotong University 2Shaanxi Provincial Key Laboratory of Big Data Knowledge Engineering 3National Engineering Lab for Big Data Analytics 4School of Computer Science and Engineering, Nanyang Technological University qikalin@foxmail.com, liukeen@xjtu.edu.cn, rui.mao@ntu.edu.sg, Leo981106@stu.xjtu.edu.cn, cambria@ntu.edu.sg Abstract Extrapolation reasoning on temporal knowl- edge graphs (TKGs) aims to forecast future facts based on past counterparts. There are two main challenges: (1) incorporating the complex information, including structural dependencies, temporal dynamics, and hidden logical rules; (2) implementing differentiable logical rule learning and reasoning for explainability. To this end, we propose an explainable extrapola- tion reasoning framework TEemporal logiCal grapH networkS (TECHS), which mainly con- tains a temporal graph encoder and a logical decoder. The former employs a graph convo- lutional network with temporal encoding and heterogeneous attention to embed topological structures and temporal dynamics. The latter in- tegrates propositional reasoning and first-order reasoning by introducing a reasoning graph that iteratively expands to find the answer. A for- ward message-passing mechanism is also pro- posed to update node representations, and their propositional and first-order attention scores. Experimental results demonstrate that it outper- forms state-of-the-art baselines. 1 Introduction Knowledge Graphs (KGs) are widely used in in- telligent systems (Ji et al., 2022; Mao et al., 2022; Zhu et al., 2023), where knowledge is commonly represented by triplets in the form of (s, r, o). The limit of conventional KGs is that real-world knowl- edge usually evolves over time. For example, a fact (Donald Trump, presidentOf, USA) is incorrect now because Joe Biden has been the new president of the USA since 2021. For more comprehensive rep- resentations of knowledge, Temporal Knowledge Graphs (TKGs) (Liang et al., 2022) are proposed by introducing time information (time point or in- terval) via quadruplets, i.e., (s, r, o, t). Then, the former example is defined as (Donald Trump, pres- identOf, USA, 2017/01/20-2021/01/20). ∗ Corresponding author. Figure 1: (a) Illustration of a TKG and extrapolation reasoning. (b) An example of temporal rules that can be applied to answer the query in (a). (c) An example of the reasoning graph that is capable of modeling rule (b). TKGs are usually incomplete (Cai et al., 2022; Liang et al., 2022). Many studies predicted future facts, based on past facts, namely TKG forecast- ing or extrapolation reasoning. Figure 1a shows the task that predicts facts at time ti with the facts at ti−2 and ti−1. A model should not only learn topology dependencies, i.e., the neighbor information of an entity (like Barack Obama at ti−2), but also learn temporal dynamics, i.e., the variations of properties of an entity over time (e.g., Angela Merkel evolves during ti−2 to ti−1). Thus, temporal embedding methods, e.g., TNT- ComplEx (Lacroix et al., 2020) and CyGNet (Zhu et al., …

Content truncated to 500 words