Deepanway Ghosal, Siqi Shen, Navonil Majumder, Rada Mihalcea, Soujanya Poria
This paper addresses the problem of dialogue reasoning with contextualized commonsense inference. We curate CICERO, a dataset of dyadic conversations with five types of utterance-level reasoning-based inferences: cause, subsequent event, prerequisite, motivation, and emotional reaction. The dataset contains 53,105 of such inferences from 5,672 dialogues. We use this dataset to solve relevant generative and discriminative tasks: generation of cause and subsequent event; generation of prerequisite, motivation, and listener'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 CICERO will open new research avenues into commonsense-based dialogue reasoning.
| Task | Dataset | Metric | Value | Model |
|---|---|---|---|---|
| Question Answering | CICERO | Exact Match | 77.68 | T5-large |
| Question Answering | CICERO | Exact Match | 77.51 | Unified QA |
| Question Answering | CICERO | ROUGE | 0.298 | T5-large pre-trained on GLUCOSE |
| Question Answering | CICERO | ROUGE | 0.2946 | T5-large |
| Question Answering | CICERO | ROUGE | 0.2878 | T5-large pre-trained on COMET |
| Question Answering | CICERO | ROUGE | 0.2837 | BART |
| Natural Language Inference | CICERO | ROUGE | 0.298 | T5-large pre-trained on GLUCOSE |
| Natural Language Inference | CICERO | ROUGE | 0.2947 | T5-large |