Dense-ATOMIC: Towards Densely-connected ATOMIC with High Knowledge Coverage and Massive Multi-hop Paths

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 13292–13305 July 9-14, 2023 ©2023 Association for Computational Linguistics Dense-ATOMIC: Towards Densely-connected ATOMIC with High Knowledge Coverage and Massive Multi-hop Paths Xiangqing Shen, Siwei Wu, and Rui Xia∗ School of Computer Science and Engineering, Nanjing University of Science and Technology, China {xiangqing.shen, wusiwei, rxia}@njust.edu.cn Abstract ATOMIC is a large-scale commonsense knowl- edge graph (CSKG) containing everyday if- then knowledge triplets, i.e., {head event, rela- tion, tail event}. The one-hop annotation man- ner made ATOMIC a set of independent bipar- tite graphs, which ignored the numerous links between events in different bipartite graphs and consequently caused shortages in knowledge coverage and multi-hop paths. In this work, we aim to construct Dense-ATOMIC with high knowledge coverage and massive multi-hop paths. The events in ATOMIC are normalized to a consistent pattern at first. We then propose a CSKG completion method called Rel-CSKGC to predict the relation given the head event and the tail event of a triplet, and train a CSKG completion model based on existing triplets in ATOMIC. We finally utilize the model to complete the missing links in ATOMIC and ac- cordingly construct Dense-ATOMIC. Both au- tomatic and human evaluation on an annotated subgraph of ATOMIC demonstrate the advan- tage of Rel-CSKGC over strong baselines. We further conduct extensive evaluations on Dense- ATOMIC in terms of statistics, human evalua- tion, and simple downstream tasks, all prov- ing Dense-ATOMIC’s advantages in Knowledge Coverage and Multi-hop Paths. Both the source code of Rel-CSKGC and Dense-ATOMIC are publicly available on https://github.com/ NUSTM/Dense-ATOMIC. 1 Introduction ATOMIC is a large-scale human-annotated com- monsense knowledge graph focusing on the infer- ential knowledge in social life (Sap et al., 2019). It consists of nine if-then relation types describ- ing the causes, effects, agent, stative, and theme of an event. The research on ATOMIC has drawn more and more attention in recent years. An in- creasing number of downstream tasks, including ∗*Corresponding author commonsense reasoning (Yu et al., 2022), sto- rytelling (Brahman and Chaturvedi, 2020), ques- tion answering (Heo et al., 2022), dialog genera- tion (Wu et al., 2022), etc., have improved their performances by acquiring and utilizing the com- monsense knowledge from ATOMIC. Currently, ATOMIC was constructed under one- hop annotations. It began with 24,000 pre-defined base events and nine relation types. For each base event and each relation, the annotators were asked to write a possible tail event based on one-hop rea- soning. As shown in Figure 1, given the base event “X asks Y to marry”, the annotated tail events can be “loving” under the relation of “xAttr”, “smiles” un- der the relation of “xEffect”, and “says yes” under the relation of “oEffect”. In such a one-hop annotation manner, each base event and its related annotated tail events shape a bipartite graph containing only B-to-A links, where B denotes the Base event and A denotes the Annotated tail event. Thereby, the whole graph of ATOMIC can be viewed as …

Content truncated to 500 words