CommonGen is a new constrained text generation dataset that requires different kinds of commonsense to generate sentences about everyday scenarios, and thus targets generative commonsense reasoning. This repo is for tracking the latest dataset, some baseline models and our evaluation scripts. Please check http://inklab.usc.edu/CommonGen/ for more details. Note that our arxiv article may contain some outdated statistics and information.
-
dataset/final_datasaves the latest version of the data. We may have updates on the dataset in the future. Please stay tuned. -
methodsshows some baseline methods with many frameworks such as OpenNMT and Fariseq, as well as UniLM. -
evaluationcontains the evaluation scripts for a variety of automatic metrics for testing the performance of system predictions against human-written results.
@article{lin2019comgen,
author = {Bill Yuchen Lin and Ming Shen and Yu Xing and Pei Zhou and Xiang Ren},
title = {CommonGen: A Constrained Text Generation Dataset Towards Generative Commonsense Reasoning},
journal = {CoRR},
volume = {abs/1911.03705},
year = {2019}
}
Feel free to directly email yuchen[dot]lin[at]usc[dot]edu if you have any feedback.