Contextualized Commonsense Inference in Dialogues (CICERO)

The purpose of this repository is to introduce new dialogue-level commonsense inference datasets and tasks. We chose dialogues as the data source because dialogues are known to be complex and rich in commonsense. At present, we have released two versions of the dataset:

CICERO-v1

CICERO: A Dataset for Contextualized Commonsense Inference in Dialogues (ACL 2022)

Paper Code Dataset

CICERO contains 53,000 inferences for five commonsense dimensions -- cause, subsequent event, prerequisite, motivation, and emotional reaction -- collected from 5600 dialogues. We design two challenging generative and multi-choice alternative selection tasks for the state-of-the-art NLP models to solve.

CICERO-v2

Multiview Contextual Commonsense Inference: A New Dataset and Task

Paper Code Dataset Demo

Depending on a situation, multiple different reasonings are possible each leading to various unique inferences. In constructing CICERO-v2, we asked annotators to write more than one plausible inference for each dialogue context. We call this task — Multiview Contextual Commonsense Inference, a highly challenging task for large language models. CICERO-v2 contains more than 8000 dialogue contexts each manually annotated with more than one plausible inferences for the following relation types: cause, subsequent event, emotional reaction, motivation.

Citation

If these datasets are useful in your research, cite the following papers:

CICERO: A Dataset for Contextualized Commonsense Inference in Dialogues. Deepanway Ghosal and Siqi Shen and Navonil Majumder and Rada Mihalcea and Soujanya Poria. ACL 2022.

Multiview Contextual Commonsense Inference: A New Dataset and Task. Siqi Shen and Deepanway Ghosal and Navonil Majumder and Henry Lim and Rada Mihalcea and Soujanya Poria. Arxiv 2022.