CICERO: A Dataset for Contextualized Commonsense Inference in Dialogues

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5010 - 5028 May 22-27, 2022 c⃝2022 Association for Computational Linguistics CICERO: A Dataset for Contextualized Commonsense Inference in Dialogues Deepanway Ghosal DeCLaRe Siqi Shen Navonil Majumder DeCLaRe Rada Mihalcea Soujanya Poria DeCLaRe DeCLaReDeCLaRe Lab, Singapore University of Technology and Design, Singapore University of Michigan, USA {deepanway ghosal@mymail.,navonil majumder@,sporia@}sutd.edu.sg {shensq,mihalcea}@umich.edu CICERO is available at: https://declare-lab.github.io/CICERO Abstract This paper addresses the problem of dia- logue reasoning with contextualized common- sense inference. We curate CICERO, a dataset of dyadic conversations with five types of utterance-level reasoning-based inferences: cause, subsequent event, prerequisite, motiva- tion, and emotional reaction. The dataset con- tains 53,105 of such inferences from 5,672 di- alogues. We use this dataset to solve rele- vant generative and discriminative tasks: gen- eration of cause and subsequent event; gen- eration of prerequisite, motivation, and lis- tener’s emotional reaction; and selection of plausible alternatives. Our results ascertain the value of such dialogue-centric commonsense knowledge datasets. It is our hope that CI- CERO will open new research avenues into commonsense-based dialogue reasoning. 1 Introduction Conversational content on the internet is quickly growing, and such content holds valuable knowl- edge about how information exchange takes place among speakers. A key step towards under- standing such dialogues is gaining the ability to reason with the information shared in the dia- logue. To this end, we curate a dataset of dyadic conversations named CICERO (ContextualIzed CommonsEnse InfeRence in dialOgues) , which contains inferences around the utterances in the dialogues. The dataset focuses on five types of reasoning-based inferences for a given utterance in a dialogue: cause, subsequent event, prerequisite, motivation, and emotional reaction. Arguably, making such reasoning-based infer- ences often demands commonsense knowledge, especially when the inference is implicit. Fig. 1a shows such a case where the cause behind the target utterance is not explicit in the context. How- ever, applying the commonsense knowledge worn gloves motivates −−−−−−−→ buy new pair of gloves Can I help you? Yes, I am looking for a pair of gloves. I think I wear size six. The gloves are on this counter. Let's see... How about the blue pair? The color suits you and they are washable, too. The speaker's gloves have worn out. Causes Speculative (a) Dad, what will we have for dinner? I don't know. Maybe fried fish, chicken soup and... Oh, no. We've had such things since Mom's gone. I'm sorry. But that is the only thing I can cook. What do you want to have then? I'd like to go to McDonald's this time. Tom is bored of eating the same dishes and want to try something different in his meal. Causes Contextual Tom prefers fast food over homemade food. Causes Contradictory (b) Figure 1: Illustration of (a) contextualized commonsense speculation and (b) contradictory inferences in dialogues. allowed the annotator to infer a probable cause of the utterance. On the other hand, commonsense can be crucial in sifting relevant information from …

Content truncated to 500 words