DimonGen: Diversified Generative Commonsense Reasoning for Explaining Concept Relationships

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4719–4731 July 9-14, 2023 ©2023 Association for Computational Linguistics DimonGen: Diversified Generative Commonsense Reasoning for Explaining Concept Relationships Chenzhengyi Liu∗ Jie Huang∗† Kerui Zhu Kevin Chen-Chuan Chang University of Illinois at Urbana-Champaign, USA {cl115, jeffhj, keruiz2, kcchang}@illinois.edu Abstract In this paper, we propose DimonGen, which aims to generate diverse sentences describing concept relationships in various everyday sce- narios. To support this, we first create a bench- mark dataset for this task by adapting the exist- ing CommonGen dataset. We then propose a two-stage model called MoREE to generate the target sentences. MoREE consists of a mixture of retrievers model that retrieves diverse con- text sentences related to the given concepts, and a mixture of generators model that generates di- verse sentences based on the retrieved contexts. We conduct experiments on the DimonGen task and show that MoREE outperforms strong base- lines in terms of both the quality and diversity of the generated sentences. Our results demon- strate that MoREE is able to generate diverse sentences that reflect different relationships be- tween concepts, leading to a comprehensive understanding of concept relationships.1 1 Introduction Concepts are mental representations of classes or categories of objects, events, or ideas, distinguished by shared characteristics that set them apart from other things. For instance, the concept of “dog” represents a class of animals that share character- istics such as being four-legged, having fur, and being domesticated. These concepts are crucial in helping us understand and communicate about the world around us. To fully grasp concepts, it is important to un- derstand the relationships between them. Re- searchers have proposed using generated sentences as a means to model these relationships more ef- fectively (Lin et al., 2020; Huang et al., 2022a,c; Huang and Chang, 2022b). For example, Common- Gen (Lin et al., 2020) aims to generate coherent 1Code and data are available at https://github. com/liuchenzhengyi/DimonGen. ∗Equal contribu- tion. †Corresponding author. (dog, sheep) • A dog herds a flock of sheep together. • Dogs guard the sheep on the mountain pasture. • The dogs are attacking a sheep. Figure 1: An example of DimonGen. The input is a pair of concepts and the output is a set of sentences that capture different ways in which these concepts interact. sentences that describe everyday scenarios involv- ing specific sets of common concepts, while Open Relation Modeling (Huang et al., 2022a) generates informative sentences that describe relationships between concepts/entities. However, in real-world scenarios, concepts often refer to broad classes, and their relationships can be complex. This can make it challenging to summa- rize these relationships through a single sentence. For example, “dog” and “sheep” are both animal concepts, but while “dogs” can herd “sheep”, they can also attack them. A single sentence would not accurately convey this complexity, leading to an insufficient understanding. Additionally, this ap- proach can also introduce bias, particularly when concepts are related to sensitive topics such as gen- der or …

Content truncated to 500 words